text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
arXiv:quant-ph/0511145v1 15 Nov 2005
Semantics and Simulation
of Communication in
Quantum Programming
Diploma Thesis
Wolfgang Mauerer1
Quantum Information Theory Group
Institute for Theoretical Physics I and
Max Planck Research Group for Optics, Information and Photonics
University Erlangen-Nuremberg, May 2005
Abstract
We present the quantum programming language cQPL which is an extended version of
QPL [Sel04b]. It is capable of quantum communication and it can be used to formulate
all possible quantum algorithms. Additionally, it possesses a denotational semantics based
on a partial order of superoperators and uses fixed points on a generalised Hilbert space
to formalise (in addition to all standard features expected from a quantum programming
language) the exchange of classical and quantum data between an arbitrary number of participants. Additionally, we present the implementation of a cQPL compiler which generates
code for a quantum simulator.
PACS numbers: 03.67.-a, 03.67.Hk, 03.67.Lx, 89.20.Ff
Revision 1.1 (15. September 2005)
1
eMail: [email protected]
Contents
1 Introduction
1
2 Quantum programming with QPL and cQPL
2.1 Programming and quantum physics . . . . . . . . . . .
2.1.1 Computability . . . . . . . . . . . . . . . . . .
2.1.2 Characteristics of quantum computers . . . . .
2.1.3 Static typing, functionality and runtime errors
2.2 Introduction to QPL and cQPL . . . . . . . . . . . . .
2.2.1 Model of computation . . . . . . . . . . . . . .
2.2.2 Language elements . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
3
3
4
5
5
5
6
Identifiers and variables • Arithmetic and logical expressions • Procedures •
Gates • Control flow • Other features.
2.2.3
Modelling quantum communication with cQPL . . . . . . . . . . . . . 10
Quantum channels • Modules and communication primitives.
2.3
Example programs . . . . . . . . .
2.3.1 Random coin tossing . . . .
2.3.2 Distribution of an EPR pair
2.3.3 Quantum teleportation . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
11
12
12
13
3 A compiler for cQPL
3.1 Structure and implementation . . . . . . .
3.1.1 Compiler technique . . . . . . . . .
3.1.2 The implementation . . . . . . . .
3.1.3 Implementation of communication
3.2 Using the compiler . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
15
15
15
17
18
18
4 Mathematical structures
4.1 Algebraic structures . . . . . . . . . .
4.1.1 Fundamentals . . . . . . . . . .
4.2 Linear operators . . . . . . . . . . . .
4.2.1 General . . . . . . . . . . . . .
4.2.2 Hilbert-Schmidt operators . . .
4.3 Connection with quantum mechanics .
4.3.1 States and effects . . . . . . . .
4.3.2 Observables . . . . . . . . . . .
4.3.3 Classical components . . . . . .
4.3.4 Composite and hybrid systems
4.4 Domain theory . . . . . . . . . . . . .
4.4.1 Basic definitions . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
21
21
21
21
21
22
23
23
24
25
25
26
26
i
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4.5
4.4.2 A fixed point theorem . . . . . . . .
4.4.3 Constructions on domains . . . . . .
cp-Maps and their representation . . . . . .
4.5.1 Operator-sum representation . . . .
4.5.2 Equivalence of Kraus operators . . .
4.5.3 A partial order for Kraus operators .
4.5.4 Kraus aggregations . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
27
28
28
29
30
30
31
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
39
39
41
41
41
42
43
45
46
A partial order for Kraus aggregations • Equivalence of Kraus aggregations.
5 Formal denotational semantics
5.1 Fundamentals of denotational semantics . . .
5.2 Survey of QPL . . . . . . . . . . . . . . . . .
5.2.1 Notational conventions . . . . . . . . .
5.2.2 Language elements . . . . . . . . . . .
5.2.3 Semantics . . . . . . . . . . . . . . . .
5.2.4 Limitation: Quantum communication
5.3 Denotational semantics of cQPL . . . . . . .
5.3.1 Formal definitions . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Typing context • Probabilistic environment • Kraus aggregations.
5.3.2
Some examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.3.3
Extension to multipartite systems . . . . . . . . . . . . . . . . . . . . 55
Semantics of sequential programs • Communication with EPR pairs.
Kraus Aggregation • Typing Context • Probabilistic Environment • Quantum
channels.
5.3.4
5.3.5
5.3.6
Existence of fixed points . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Types of interpretational equivalence . . . . . . . . . . . . . . . . . . . 59
Semantics of the language components . . . . . . . . . . . . . . . . . . 60
Some notational remarks • State transformations and fixed points • Classical
operations • Quantum operations.
5.3.7
5.3.8
Explicit transformations of density matrices . . . . . . . . . . . . . . . 78
The type system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Quantum variable tuples • Application of operators • If conditionals and while
loops • Measurements • Communication.
5.4
Avoidance of runtime errors . . . . . . . . . . . . . . .
5.4.1 Unique ownership of qbits . . . . . . . . . . . .
5.4.2 Prevention of cloning and unphysical situations
5.4.3 Unavoidable non-termination conditions . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
81
81
82
82
6 Prospects
85
6.1 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.2 Latest developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
A List of symbols
87
B Glossary
89
C Formal syntax
91
Bibliography
93
ii
Dich sah ich, und die milde Freude
Floß von dem süßen Blick auf mich;
Ganz war mein Herz an deiner Seite
Und jeder Atemzug für dich.
Johann Wolfgang von Goethe, Willkommen und Abschied
1
Introduction
Although there is no formal proof that quantum computers offer greater computational
power than classical ones, there are a few quantum computer algorithms which provide
efficient solutions for problems which are up to now believed to be classically NP-hard, i.e.,
they cannot be solved in polynomial time. One of these problems – computing discrete
logarithms – is the cornerstone of basically every modern classical cryptographic algorithm,
so the increased interest in quantum computing is obvious, both from a fundamental and a
practical point of view.
How can quantum algorithms be described in an efficient, readable and precise manner?
Usually, this is done with quantum circuits which are combined into a quantum network, but
especially for larger programs, this is a very cumbersome and error-prone method. Much
research has been performed on the construction, implementation and conception of classical programming languages; therefore, a great amount of alternatives are available. The
situation is totally different in quantum computing: Only a handful of languages have been
proposed so far [AG04, BSC01, Öme98, SP00, Sel04b, vT04], and only two working implementations based on quantum computer simulators are available [Öme98, BSC01] (recently,
a third, but still rough implementation was presented in [AG05]). Both are based on imperative/object oriented languages (C, Pascal, C++, . . . ), with quantum features as extensions,
whereas QPL [Sel04b] models a basically functional language.1
In this work, we present an extension of QPL with abilities to cover quantum communication, i.e., the transmission of quantum mechanical states and exploitation of their highly
non-classical properties like superpositions and entanglement. To distinguish our approach
from QPL, we call it cQPL for communication capable QPL. In contrast to quantum computers of which only some very elementary parts have been experimentally realised until now,
implementations of quantum communication are not only available in several laboratories
around the world, but can even be obtained commercially.
To provide some orientation where our approach can be located in contrast to work done
by other contributors to the field, Table 1.1 presents a comparison of quantum programming
languages and their features.
In a nutshell, the contribution of our work to the field of quantum programming languages
is twofold:
❏ We provide a compiler which can serve as a testbed for new ideas in quantum programming, to teach concepts of quantum algorithms or act as an aid to the intuition
1 We need to note, though, that the term functional should not be overestimated in this context. Important features like higher order functions, which are considered to be key elements of classical functional
languages, are missing in QPL. The most interesting part of the language from a physicist’s point of view is
the ability to guarantee freedom against runtime errors already at compile time, no matter how this goal is
achieved.
1
Chapter 1. Introduction
Reference
New language
Respects physicsa
Implemented
Formal semantics
Communication
Universal
QCL
[Öme98]
✗
✗
✓
✗
✗
✓
Q Language
[BSC01]
✗
✗
✓
✗
✗
✓
qGCL
[SP00]
✓
✗
✗
✓
✓
✓
QML
[AG04]
✓
✓
✓b
✓
✗
✓
QPL
[Sel04b]
✓
✓
✓c
✓
✗
✓
cQPL
✓
✓
✓
✓
✓
✓
a Respecting physics is meant in the sense that it is not possible to syntactically specify programs which
would create unphysical situations; of course, such a state will force the simulation to abort with an error
and needs thus to be avoided if possible.
b A partially finished implementation was available at the time of writing.
c If we consider our cQPL compiler to be a QPL compiler as well.
Table 1.1: Comparison of quantum programming languages defined in other approaches to the problem, their features and their shortcomings.
of users who want to experiment (in the sense of goal oriented playing, not laboratory)
with quantum protocols.
❏ We present an alternate approach to the compositional semantics of QPL and provide
the possibility to include and formalise quantum communication as part of the language. This is a necessary step towards the formalisation of open-world programs, but
may also prove itself useful in fields like quantum process calculi, automated protocol
analysis or similar – cf. Chapter 6 for further prospects.
The layout of this thesis is as follows: In Chapters 2 and 3, we provide an overview
about quantum programming, present the language cQPL and describe the compiler and
its implementation. Chapter 4 presents the mathematical tools and requisites necessary
for a denotational semantics of cQPL, and Chapter 5 develops the semantical description.
Chapter 6 finally provides some short remarks on possible further directions that may be
pursued based on the results of this work. All chapters are interlaced with short introductions
to topics, tools and techniques which are uncommon in physics, but necessary for our work;
Appendices A and B provide a list of symbols and a glossary, respectively. The formal
syntax is presented in Appendix C. To aid the reader in staying on track, we have provided
short summaries in grey boxes at some points along the way.
Since the topic dealt with does not only contain problems of physical nature, but also
touches the fields of computer science and mathematics, we have tried to make the text as
self-contained as possible for readers with any of these backgrounds.2 Obviously it was not
possible to present everything in as much detail as required without repeating the introductory textbooks, but numerous references to the literature are provided which hopefully
alleviates any arising problems.
2 In this context, it is interesting to note that 48% of all entries in the bibliography are of physical nature
and 36% can be counted to computer science; the remaining 16% belong to mathematics or are of general
interest.
2
When someone says, “I want a programming
language in which I need only say what I
wish done,” give him a lollipop.
Alan Perlis, Epigrams on Programming
2
Quantum programming with
QPL and cQPL
This chapter presents an overview about classical and quantum computers, the underlying
models of computation and some principal remarks on quantum programming languages.
Afterwards, a short introduction to QPL and cQPL is given. Note that this chapter is
intentionally kept as terse as possible to allow a more detailed coverage of other topics dealt
with in this thesis. Therefore, no attempt is made to to provide special rigour in this chapter.
2.1
2.1.1
Programming and quantum physics
Computability
Classical computers can be described by several models which are apt for different purposes
(for a more detailed description, cf., e.g., [AB02, Sch01]):
❏ Turing machines
❏ General recursive functions
❏ Register machines
❏ Lambda calculus
❏ Logical gates
❏ Universal programming languages
They can all be brought to a common denominator by showing that they are able to
compute respectively solve the same class of problems. They are computationally equivalent
because one model can be used to efficiently simulate any other model; a Turing machine is
normally taken to be the normative instance among them. The fact that a turing machine
can compute everything which is computable in principle is captured in the Church-Turing
thesis which is one of the fundamental axioms of computer science:1
Hypothesis 2.1.1 (Church-Turing). Every function which would naturally be regarded
as computable can be computed by a Turing machine.
This definition places computability in a purely abstract setting without regard to the
laws of physics. Deutsch [Deu85] realised that if the laws of (quantum) physics are used
as basis for computation, an improved version of the Turing machine might lead to greater
computational power. This requires an extended version of the Church-Turing thesis as well:
1 To be precise, one would have to note that Church’s thesis states that “a function of positive integers
is effectively calculable only if recursive”. This is equivalent to Turing’s thesis, though, and the name
Church-Turing thesis is conventionally used in the literature.
3
Chapter 2. Quantum programming with QPL and cQPL
Hypothesis 2.1.2 (Church-Turing-Deutsch). Every physical process can be simulated
by a universal computing device.
Greater computational power is meant in the sense that there are some problems which
can be efficiently solved by a universal computing device according to Church-TuringDeutsch but which cannot be efficiently solved by a Turing machine, e.g., the simluation cost
for a universal computing device on a normal turing machine would be at least exponential
for the most general case.2
Until now, it has not been proven if quantum Turing machines have greater computational power than Turing machines or not; cf. Refs. [Cle99, Aha98] for further information
on this and related topics.
2.1.2
Characteristics of quantum computers
The basic building block for quantum computers are quantum bits (qbits), i.e., quantum
mechanical objects which can be represented by two different basis states, usually denoted
by |0i and |1i. This can, e.g., be realised by two-spin systems such as electrons, with the
two different polarisations of a photon etc. In contrast to classical machines and models of
computation, quantum computing can draw from two additional resources:
❏ Superpositions: The state of a quantum bit can be in a superposition α |0i + β |1i
with α, β ∈ C and |α|2 + |β|2 = 1. If every qbit of a quantum register consisting of n
elements is brought into a symmetric superposition with α = β = 2−1/2 , the register
contains all numbers from 0 to 2n − 1 at the same time. Manipulations of the register
thus manipulate all these numbers in one step, whereas classically, the number of
required manipulations would grow exponentially with the register size. This feature
is conventionally referred to as quantum parallelism.
❏ Entanglement: Two parts of a quantum system can be in an entangled state such that
manipulations on one part of the system influence the other system although they may
be spatially separated.
While superpositions are useful for computational problems, entanglement is suitable for
tasks like secret key growing, but is for example also necessary to connect input with output
registers for quantum function application.
In general, there are several equivalent models for quantum computers which are abstracted from their physical realisation:
❏ Quantum Turing machines
❏ Quantum gates
❏ Universal quantum programming languages
All these models are equivalent, cf., e.g., Refs.[NC00, Aha98, Pre99] for more detailed explanations.
2 Note that there is a very prominent problem which illuminates this field: Factorising integers is possible
at polynomial cost on a quantum computer, but the currently best known algorithms require exponential cost
on a classical computer. This does not prove, though, that quantum computers are per se more powerful than
classical ones because there is, for example, no proof that there is no classical polynomial-time factorisation
algorithm.
4
2.2. Introduction to QPL and cQPL
2.1.3
Static typing, functionality and runtime errors
Selinger proves in [Sel04b] that QPL has the ability to avoid runtime errors by detecting
them at compile time (and can thus reject the program) which is not possible in other
language proposals presented at the time of writing. He argues that this must be attributed
to the static type system of the language. Although the proof for this is solely based on
properties of the static type system, the functional style of QPL does have its merits in this
respect in our opinion, too: It restricts the language to elements which allow to express
universal programs, but do not allow constructions that can produce runtime errors that
cannot be detected at compile time.3
When work on this thesis was started, one of the points we wanted to investigate was if
and how principles of functional languages could be advantageous for quantum languages.
This did not reach fruition, though, because we could not find a simple way to transfer any
advanced functional method like higher-order functions, recursively defined data types etc.
to the quantum case.4
From a physical perspective, the absence of runtime-errors is much more interesting than
any computer science related question like the type system or functionality. In this thesis,
we thus concentrated on preserving the possibilities of static checking provided by QPL as
far as possible while enhancing usability of the language and providing means to handle
commuication.
2.2
Introduction to QPL and cQPL
This section presents a very short introduction to cQPL; some of the things stated in the
following also hold for a (regularised version) of block QPL5 as presented in [Sel04b] which
is used as basis for cQPL. Note that cQPL is explicitely meant to be an experimental
compiler. Only very modest effort was made to make the compiler easy to use (the error
messages provided can, for example, often only be understood if the user is familiar with the
inner working of the compiler), and only very few classical operations which are considered
standard components of programming languages (e.g., numerical operations like sin, cos etc.)
were implemented to save time since this is purely routine work. Nevertheless, care has been
taken to design the compiler for easy extensibility.6
2.2.1
Model of computation
Although quantum gates are the most widespread model in the literature on quantum algorithms, the QRAM model suggested by Knill [Kni96] is more apt as basis for quantum
programming languages. The model consists of two components: A classical computer for
conventional tasks like program flow control, classical calculations etc., and a quantum memory controlled by the classical computer that can not only store quantum states, but also
apply any unitary operator and perform measurements. Figure 2.1 visualises the approach.
3A
static type system is in general not enough to ensure this property: Just think of the many possible
ways to generate runtime-errors in C, which is statically typed as well. This is partially caused by the fact
that typing in C is weak; nevertheless, even strongly typed languages as, e.g., Java and C# still cannot
prevail runtime errors. Thus, static typing alone is not sufficient to avoid all possible runtime errors.
4 To our knowledge, no fully satisfying mechanism for any of these problems has been found until now
although the topic is addressed in several papers.
5 The variant of QPL with a textual structure that includes blocks; alternatively, there is a textual variant
without blocks and a flow diagram representation.
6 For example, integration of the n-qbit Fourier gate could be accomplished by adding only 8 lines of code
to the compiler and 10 to the runtime library.
5
Chapter 2. Quantum programming with QPL and cQPL
Figure 2.1: Hybrid architecture for a quantum computer which consists of a classical computer and a quantum memory with the ability to apply unitary
operators and perform measurements at the disposal of the classical
system (image taken from [Öme03]).
cQPL is centred on the assumption that the control flow of a program can be described
by classical means. Although this is considered to be a loss of generality by some authors
(most notably [Öme03]), it does not represent a real restriction because all known quantum
algorithms can be expressed in this framework. Data is, of course, quantum mechanical,
and can be modified by unitary operators which are equivalently called gates in analogy to
the gate model of quantum computation.
2.2.2
Language elements
The following remarks provide an introduction to cQPL for ordinary users; a formal version
of the syntax will be presented in Appendix C. The compiler distribution provides some
example programs which demonstrate all features.
2.2.2.1
Identifiers and variables
Identifiers for variables can be denoted by strings consisting of the characters A–Z, a–z,
and 0–9, where no digit must be at the beginning. Allocation of new variables is done with
the operator new:
new int loop := 10;
new qbit b1 := 1;
Note that it is mandatory to provide an initial value both for classical and quantum data
types. By default, the data types bit, int, float, qbit and qint are available.
Assigning values to classical variables is possible using the operator :=:
loop := loop - 1;
6
2.2. Introduction to QPL and cQPL
2.2.2.2
Arithmetic and logical expressions
Arithmetic expressions can be given as in most programming languages (e.g., 7+3*x-5).
The operators +, -, * and / are available, they are overloaded to work with any classical
numerical data type.
Operators resulting in a logical value (true or false) are <, >, <=, >=, = and !=; logical
negation is available using !, and predicates can be combined using & (and) and | (or).
Operator priority is defined as usual in arithmetic and logic; parentheses can be used to
explicitely modify this.
2.2.2.3
Procedures
Procedures can take an arbitrary number of input parameters (including none) which are
specified as name:type tuples. Declarations take the following form:
proc test: a:int, b:bit, c:float, d:qbit {
...
}
Procedure calls are achieved with the keyword call:
(eins, zwei, drei) := call test(a0, a1, a2, b1);
Note that the parameters are passed by value so that modifications of the classical
variables given to the procedure (in this case a0, a1 and a2) are not visible in the caller’s
scope. This means that no matter what test does, these variables have the same values
before and after the procedure is called. This is not the case with b1 because it is a quantum
variable and cannot be cloned to implement call-by-value; any modifications performed by
test on the state ofb1 are visible for the caller after test returns.
The result of a procedure is a tuple containing the values the classical parameters had at
the end of procedure execution; the example stores these in the local variables eins, zwei
and drei. Note that the result of a procedure call may be ignored as well by the caller, so
the following variant is also possible:
call test(a0, a1, a2, b1);
The example program proc test.qpl which accompanies the compiler further illustrates
the described behaviour, so we refer the reader to it.
2.2.2.4
Gates
Gate application is performed using the operator *= as in the following example:
q *= Not;
A small number of elementary gates is built into the language core (remember that additional
ones can be added with really little effort as we already mentioned before):
❏ H Hadamard transformation on a qbit.
❏ FT(n) Fourier transformation on n qbits, i.e., the n-fold tensor product of Hadamard
transforms.
7
Chapter 2. Quantum programming with QPL and cQPL
❏ NOT Logical negation on one qbit.
❏ CNot Controlled Not on two qbits.
❏ Phase Phase shift gate on one qbit; the desired shift is given as parameter.
The dimension of the gate and the destination must match. Variable tuples where
identifiers are combined with commas (,) can be used to combine several quantum variables
as in the following example:
new qbit test1 := 0;
new qbit test2 := 1;
test1, test2 *= CNot;
test1 *= Phase 0.5;
User-defined gates can be defined by enclosing a list of (complex) numbers in [[ and ]]
as in the following example:
test1,test2 *= [[0.5, 0.5,
0.5, 0.5,
0.5, 0.5i, -0.5, -0.5i,
0.5, -0.5,
0.5, -0.5,
0.5, -0.5i, -0.5, 0.5i]];
2.2.2.5
Control flow
If-then-else and While are available for directing the control flow.
new int loop := 10;
while (loop > 5) do {
print loop;
loop := loop - 1;
};
if (loop = 3) then {
print "3";
}
else {
print "Nicht 3";
}
The meaning of these operations is the same as in classical languages.
2.2.2.6
Other features
Several more features which do not fit into any of the above categories are available:
❏ dump takes one or more quantum variable identifiers as argument and provides a dump
of the current probability spectrum in the canonical basis |0i , |1i:
dump eins, zwei;
To demonstrate the effect of this command, consider the following example:
8
2.2. Introduction to QPL and cQPL
new qbit a := 0;
new qbit b := 0;
print "State before FT:"; dump a, b;
a, b *= FT(2);
print "State after FT:"; dump a, b;
The program fragment produces the following output when run:
State before FT: 1 |00>
State after FT: 0.25 |00>, 0.25 |01>, 0.25 |10>, 0.25 |11>
Note that there is a fundamental difference between dumping the state of quantum
variables and measuring the state and printing the result, as the following example
shows:
measure a then { print "a is |0>"; } else { print "a is |1>"; };
print "State of b:"; dump b;
print "State of (a,b):"; dump a,b;
If the fragment above is supplemented by these lines, running the program either yields
this
a is |0>
State of b: 0.5 |0>, 0.5 |1>
State of (a,b): 0.5 |00>, 0.5 |01>
or that output (the output of the statements before is omitted):
a is |1>
State of b: 0.5 |0>, 0.5 |1>
State of (a,b): 0.5 |10>, 0.5 |11>
The result of the command dump b never changes, no matter how often the program is
run. The result of the measure command, however, will change so that every output
appears with probability 0.5 for a large number of executions.
❏ skip does nothing, but can be used to fulfil syntactic requirements if, for example,
one branch of an if-then-else-statement is supposed to do nothing:
if (condition) then {
skip;
}
else {
...
}
❏ print prints the value of a variable or an arithmetic expression (e.g., print 5+7;
print a;), but can also be used to output text strings enclosed in quotation marks
to the console (e.g., print "Hello, world!";).
9
Chapter 2. Quantum programming with QPL and cQPL
❏ measure measures a quantum variable and returns a classical result. The result is
governed by a probability distribution according to the state the quantum variable
is in; thus, successive program runs will in general return different results when the
function is called. Note that the measurement is always performed in the standard
basis |0i , |1i for each contributing qbit.
2.2.3
Modelling quantum communication with cQPL
Although quantum communication has already been implemented with several physical
schemes at the time of writing, cQPL does not consider any of these solutions. Instead,
the model presented in Section 2.2.1 is extended in such a way that the peculiarities of
communication can be replaced by reasonable simulation alternatives. In a real physical
setting, communication between two parties can be implemented by using whatever kind
of quantum channel which allows to transfer quantum states. Since this is not quite easy
to implement experimentally (quantum states are very fragile objects), diverse effects like
channel loss, decoherence etc. need to be taken into account in a real-world setting. This is
accounted for by a channel model which describes these effects mathematically.
2.2.3.1
Quantum channels
Obviously, no spatial transmission of quantum states is performed in the simulation.7 A
quantum channel is thus replaced with a label on each qbit present in the simulation which
denotes the respective owner. Sending a qbit is thus equivalent to changing the label of
it. Note that this definition differs from the definition of a channel which is used in some
contributions to the literature, e.g. [Key02]. Here, a channel is seen as something that modifies the state, for example by decoherence, loss or the influence of eavesdroppers, whereas
our definition captures the notion of a channel as a means of unambiguous quantum state
transfer. Inclusion of eavesdroppers (or any modification of the quantum state) is possible
if a quantum channel is replaced by two quantum channels which are connected by a third
module (which we E for Eve following the usual convention) which receives quantum states
from Alice, performs appropriate modifications and resends the states to Bob. Obviously,
Eve can collect multiple qbits that pass the channel and manipulate them collectively to pursue different attack strategies, so this is in no way a restriction to intercept-resend attacks.
Figure 2.2 provides a visualisation of the communication model.
2.2.3.2
Modules and communication primitives
Communicating systems are specified in terms of modules which contain the code the participants execute. Every module is identified by a unique label which identifies it among the
participants. Note that module definitions must only occur at the top-level; modules must
not contain other module definitions. Two parties which are named Alice and Bob can be
implemented by this structure:
module Alice {
...
};
module Bob {
7 It would be possible, for example, to consider networked computers between which simulated quantum
states can be transferred. This would add no new physical insights to the problem, but only add technical
difficulties, so we did not implement such a scheme.
10
2.3. Example programs
Figure 2.2: Model of communication used in cQPL. Transmission of quantum data
is replaced by a quantum heap shared by the communicating parties;
all qbits have a unique label that identifies to whom they belong at
the moment. Sending and receiving can be modelled by changing the
label, the channel itself is modelled by a third party.
...
};
Channels for sending and receiving are implicitly opened between all participants. The
send command is provided for sending variables. Two parameters need to be supplied: A
list of variables (which may be classical or quantum) and the identifier of the receiver. For
example, the following code can be placed in module Alice:
new qbit q1 := 0;
new qbit q2 := 1;
...
send q1,q2 to Bob;
Receiving works similar; consider for example the following code which might be located
in module Bob:
...
receive var1:qbit, var2:qbit from Alice;
...
// Do something with var1, var2
Note the a receive commands implicitly introduces new variables into the present frame,
in this case var1 and var2. The data type of the received quantities must be specified after
the variable name where a colon is used as separator.
2.3
Example programs
We present three small, but complete examples together with their output generated by
executing them to allow a look at the cQPL syntax without resorting to program fragments.
11
Chapter 2. Quantum programming with QPL and cQPL
2.3.1
Random coin tossing
This simple algorithm puts a quantum bit into a symmetric superposition and measures it
to simulate the effect of tossing a perfect coin:
new qbit q := 0;
q *= H;
measure q then { print "Tossed head"; } else { print "Tossed tail"; };
The output is with equal probability either Tossed head or Tossed tail, obviously.
2.3.2
Distribution of an EPR pair
Distributing EPR pairs is one possible method to establish a secret key between Alice and
Bob that can, e.g., be used for a Vernam cipher. The following cQPL program shows how
to do this for a single EPR pair:
module Alice {
proc createEPR: a:qbit, b:qbit {
b *= Not;
b *= H;
a,b *= CNot;
} in {
new qbit first := 0;
new qbit second := 0;
call createEPR(first, second);
send second to Bob;
measure first then { print "Alice’s qbit is |1>"; }
else { print "Alice’s qbit is |0>"; };
};
};
module Bob {
receive q:qbit from Alice;
measure q then { print "Bob’s qbit is |1>"; }
else { print "Bob’s qbit is |0>"; };
};
Running the program creates one of the following two outputs with equal probability:
Alice’s qbit is |0>
Bob’s qbit is |0>
Alice’s qbit is |1>
Bob’s qbit is |1>
Note that the order of Alice and Bob’s output may be inverted as well because after
and before the send/receive synchronisation, the execution order of the threads representing
Alice and Bob is indeterministic.8
8 To be precise: The execution order depends on how the computer used for the simulation implements
threads and their parallel execution, so it is effectively indeterminate. On machines with at least two CPUs
and assuming that each thread runs on one of them, the indeterminism is real.
12
2.3. Example programs
2.3.3
Quantum teleportation
Quantum teleportation is an algorithm that enables to transfer an unknown quantum state
between two parties if both of them share one part of an EPR state (in this case, |β00 i) and
can communicate classically. An easy calculation as is, for example, given in [NC00, p. 27]
or any other quantum information text shows how this works, so we will not repeat it here.
The implementation in cQPL is as follows:
module Alice {
proc createEPR: a:qbit, b:qbit {
a *= H;
b,a *= CNot; /* b: Control, a: Target */
} in {
new qbit teleport := 0; /* Apply unitary operations to set the qbit
to any other desired state */
new qbit epr1 := 0;
new qbit epr2 := 0;
call createEPR(epr1, epr2);
send epr2 to Bob;
teleport, epr1 *= CNot;
new bit m1 :=
new bit m2 :=
m1 := measure
m2 := measure
/* teleport: Control, epr1: Target */
0;
0;
teleport;
epr1;
/* Transmit the classical measurement results to Bob */
send m1, m2 to Bob;
};
};
module Bob {
receive q:qbit from Alice;
receive m1:bit, m2:bit from Bob;
if (m1 = 1) then {
q *= [[ 0,1,1,0 ]];
};
if (m2 = 1) then {
q *= [[ 1,0,0,-1 ]];
};
/* Apply sigma_x */
/* Apply sigma_z */
/* The state is now teleported */
print "Teleported state:";
dump q;
};
13
14
Denn was wir tun müssen, nachdem wir es
gelernt haben, das lernen wir, indem wir es
tun.
Aristoteles, Nikomachische Ethik
3
A compiler for cQPL
A compiler for cQPL was implemented as part of this thesis; since no quantum computers
are available yet, it is obviously targeted at simulators for such. This chapter presents a very
short overview about the compiler’s implementation and the limitations which arise from the
experimental nature of it. Note that this chapter is intentionally kept as terse as possible;
it is not supposed to be a detailed description of the techniques used in implementing the
compiler nor is its intention to go down to the source code level.
3.1
Structure and implementation
Since the initial intention when work on the thesis started was to examine the aptness
of functional methods in programming languages for quantum computers, we decided to
implement the compiler for cQPL in a functional language as well (observe the remarks in
Section 2.1.3 why functionality was realised not to be the important factor). The choice after
testing several alternatives fell on Objective Caml (OCaml) which is, for example, described
on http://caml.inria.fr. One of the reasons for this choice was that very good automated
generators for lexers and parsers are available as part of the compiler distribution which
considerably reduce routine work. As simulation backend, we used the routines supplied
with Ömer’s QCL compiler [Öme98] because this library was the most advanced one at that
time. In the meanwhile, a number of other libraries appeared and existing ones matured,
so most likely, we would have chosen a different simulation backend now because the QCL
library comes effectively without any documentation for the programmer (only the QCL
compiler built on top of the library is documented) which resulted in quite a few technical
obstacles.
A simple compiler for a simple classical language was implemented as part of this thesis
to provide an example for explaining compiler technique and to get used to the OCaml
language and the associated tool-chain.
3.1.1
Compiler technique
As usual in compiler technique, the work is separated into several passes which are shown
in Figure 3.1.
Lexical analysis is performed by a lexer generated with OCamllex. The lexer transforms
the input stream of characters into a stream of tokens where tokens are, for example, keywords like int, measure and so on. This simplifies the parsing process because it does not
need to deal with a program representation at the level of single characters, but can already
work with less elementary units. A more interesting part is the syntactical analysis where
the parser determines if the program is valid according to the rules given in Section C. The
15
Chapter 3. A compiler for cQPL
Figure 3.1: Structure of the cQPL compiler
syntax for cQPL is specified as a LALR(1) grammar which is a special kind of an LR(1)
grammar which, in turn, is a variant of a context free grammar as introduced in Definition 5.1.1. We do not want to go into more technical details here, but refer to Appendix B
and the usual literature on the topic, e.g.. Refs. [ASU86, App04, GBJL02, WM95] for the
exact definitions of such grammars.
After the syntactical correctness of a program has been ensured, the compiler can analyse
the generated parse tree to infer the meaning of the program and perform compile time
checks. These try to find as many errors in the program as possible to ensure that they
will not appear at runtime and lead to abortion of the program with an error. Such checks
include, for example:
❏ Making sure that all variables were declared before use.
❏ Checking that procedures are called with proper arguments.
❏ Matching the dimension of quantum gates with the dimension of variables they are
applied to.
❏ Ensuring that tuples of quantum variables are disjoint.
Many more checks of this kind can be found in the source code. Some (e.g., the first
mentioned two) appear in classical programming languages as well, whereas others (e.g., the
last mentioned two) are specific to quantum languages or necessary only for our approach.
The goal of the analysis phase is to provide a representation of the program that is augmented
with everything necessary so that the code-generation backend can produce its result directly
from this representation. An additional advantage of this strategy is that the code generation
can be replaced by a variant targeting a different simulator.1 The structure of the annotated
parse tree which is passed from the parser to the code generation backend is documented in
the source code.
1 For a very early version of the compiler, a code generation backend for the Fraunhofer simulator
[RAMK+ 04] was provided as well, but we dropped support for this to not spend too much time on implementation issues.
16
3.1. Structure and implementation
3.1.2
The implementation
We do not want to go into any implementation details here, but just provide an overview
about the source files which constitute the compiler. Every detail of the implementation can
obviously be found there.
❏ lexer.mll is an input file for OCamllex which generates the lexical analyser for QPL
from the rules given in the file.
❏ parser.mly is an input file for OCamlyacc which is used to transfer the rules (and
code) given in the file into a parser for cQPL syntax.
❏ parser defs.ml defines the data structures used by the parser to build the abstract
syntax tree. For every element of the grammar, a separate data structure is defined.
Most of them can be augmented with additional information which is inferred during
analysis, e.g., lists of type conversions.
❏ cqpl.ml plays a twofold role: On the one hand, it implements the user interaction
part which handles command line processing and directs the different passes of the
compiler. On the other hand, it implements the semantic analysis where for every
grammar production, a corresponding function is available to perform the appropriate
checks.
❏ gen qcl.ml implements the code generation backend for the QCL library. The generated code depends on the C++ code in qpl_runtime.{cc,h} which provides the
runtime environment for programs. This handles, for example, quantum memory management and provides implementations of standard operators. qpl runtime template.h contain the implementation of runtime routines which use templates and thus
cannot be compiled to a static object; the routines to implement the communication operations (i.e., thread handling, data exchange and locking) can be found in
qpl\_runtime\_comm.{cc,h}.
❏ type.ml contains the type checker and routines to perform lossy and non-lossy conversion between different data types. Especially interesting here is the conversion
between classical and quantum variables. The routines provided by this file are much
more general than required by the compiler and would in principle be able to handle
mixed data types as well.
❏ stacked env.ml provides a stackable environment for the semantic analysis.
Since the compiler is only one part of this thesis, we deliberately accepted some limitations which would have to be removed for a production quality compiler. None of them
would present a problem in principle, but would require some tedious effort that does not
justify the gain:
❏ Error messages are not detailed, and no effort is made to continue translation as far
as possible or perform error recovery after a mistake was spotted.
❏ No specific error productions are provided. Syntactical errors are therefore in general
reported by the lexer and not properly by the parser.
❏ There is absolutely no optimisation for execution speed.
17
Chapter 3. A compiler for cQPL
❏ Communication was not too intensively tested because we concentrated more on the
formal aspects of this. Additionally, it uses big locks instead of finer-grained solutions
which reduces performance.
❏ There is no set of standard routines or a standard library of any kind which would be
required for real-world applications.
3.1.3
Implementation of communication
Since we do not use a real network of quantum and classical computers to simulate communicating parties, but only a single system, it is necessary to find an approach that emulates
the characteristic properties of real-world systems in the simulation environment. It has
already been shown that the QRAM model can – together with an appropriate labelling
mechanism for the qbits – provide a suitable replacement for a quantum channel. The remaining issue that needs to be addressed is how the model the independent execution of
the parties together with the synchronisation when send/receive operations take place is to
be implemented. This is, obviously, a problem of the backend, so the solution presented is
specific to the QCL backend.
For every module defined in a program, a separate thread2 is created which has access to
the global memory management routines and communication channels. Quantum memory
management is performed by the main process; appropriate locking techniques ensure that
no race conditions can occur when quantum bits are allocated by the modules. Sending
quantum data can in this scheme be performed by exchanging pointers to the respective
positions on the quantum heap; the static analysis of the compiler guarantees that no more
than one process is in possession of a given qbit at a time. Again, appropriate mechanisms in
the form of mutexes are used to provide the required synchronisation between the modules
which is required to implement send and receive operations. For every pair of modules, a
separate bi-directional queue is provided to handle the exchange of quantum and classical
data.
3.2
Using the compiler
As all well behaved programs, the compiler can provide a list of its options:
wolfgang@meitner> ./cqpl --help
Quantum Programming Language v1.0
Usage: qpl [<input>] [<output>] [--debug] [--nonative] [--norun]
[--backend qcl] [--qheap size]
<input>: Input filename
<output>: Output filename
--debug Print debug messages
--backend Simulation backend (Only qcl is supported at the moment)
--nonative Generate only backend code, don’t create a native executable
--norun Do not execute the generated native code
--qheap Size of quantum heap (default: 200 qbits)
2 A thread is – depending on the implementation by the underlying library and the operating system
kernel – a lightweight process that shares most ressources he has with other threads that execute in parallel,
but has its own control flow.
18
3.2. Using the compiler
-help Display this list of options
--help Display this list of options
❏ <input> and <output> provide the names of the input and output file. If no input file
is specified, the compiler reads from the stdin channel. If no output if specified, the
filename of the input file together with some appropriate extension according to the
output mode is used.
❏ --debug reports lots of debug messages. Console output is really noisy with this
option enabled, but provides some insight into what the compiler does. Obviously
mainly useful for debugging the compiler itself.
❏ --nonative specifies that no native code is generated, i.e., the program is not transformed into machine code by the backend. For the QCL backend, this means that
only a C++ version of the quantum program is generated, but the C++ compiler is
not called to generate a native executable.
❏ --norun does not execute the generated program, but only leaves the executable file
which can be run later.
❏ --qheap size sets the size of the quantum heap to size qbits.
The compiler source code will be made available in the near future on http://kerr.
physik.uni-erlangen.de/qit/qpl.html once the remaining changes to make the source
code ready for public distribution have been performed.
19
20
A mathematician is a device for turning
coffee into theorems.
Pál Erdős
4
Mathematical structures
Because our work involves the theoretical parts of physics and computer science, many different notations and conventions enter the game. To eschew obfuscation, it behoves to define
a consistent notation to be used in this work, which we shall do in the following sections. Additionally, this chapter introduces some mathematical structures and their properties which
are required for the description of the denotational semantics of cQPL.
4.1
4.1.1
Algebraic structures
Fundamentals
The concept of an algebra is convenient to cover the properties of quantum mechanics in an
abstract setting, so we remind the reader of two elementary definitions for terms that are
often used sloppily in physics.
Definition 4.1.1 (Algebra). Let K be a field. An associative K-algebra A over K is a
nonempty set A together with three operations called addition +, multiplication × and scalar
multiplication · (the last two operations are usually denoted by juxtaposition of symbols) for
which the following properties hold:
❏ A is a linear space under addition and scalar multiplication.
❏ A is a ring under addition and multiplication.
❏ If r ∈ K and a, b ∈ A, then r · (a × b) = (r · a) × b = a × (r · b).
Definition 4.1.2 (Subalgebra). S ⊆ A is subalgebra of an algebra A if it has the properties
of an algebra and is closed under operations of A.
4.2
4.2.1
Linear operators
General
Much of our work is based on the grounds of linear operators, so we present some fundamental
definitions and cite a theorem from the literature which will be useful for us later on.
Remark 4.2.1. Note that we use only linear operators in this work, so operator is used as
a synonym for linear operator without mentioning this explicitely in the following chapters.
21
Chapter 4. Mathematical structures
Definition 4.2.1 (Linear operator). A linear operator T from a normed space X to
another normed space Y is a linear map from D(T ) ⊆ X (the domain of T ) to Y with the
following property for x, y ∈ D(T ), α, β ∈ K:
T (αx + βy) = αT (x) + βT (y).
(4.1)
Definition 4.2.2 (Bounded operator). An operator is called bounded if ∃C ≥ 0, C ∈ R
so that
||T x|| ≤ C · ||x||
(4.2)
for all x ∈ D(T ) with ||x|| ≤ 1.
Theorem 4.2.1. Let X, Y be normed spaces. For a linear operator T : X → Y , the
following properties are equivalent:
❏ T is continuous in every point of D(T ).
❏ T is continuous at 0.
❏ T is bounded.
Proof. Cf. Ref. [Wei00, Theorem 2.1].
4.2.2
Hilbert-Schmidt operators
Let X, Y be Hilbert spaces. An operator K ∈ B(X, Y ) is called Hilbert-Schmidt-operator
if there exists an orthonormal basis {eα : α ∈ A} (where A is some index set) with
P
2
†
α∈A ||Keα || < ∞. In a more physical notation, this means that tr K K < ∞. This
is obviously fulfilled if K ∈ B(H) and dim(H) < ∞.
Theorem 4.2.2 (Hilbert space of Hilbert-Schmidt operators). For Hilbert-Schmidt
operators K, L of a Hilbert space X to a Hilbert space Y , ||·||HS is a norm on this space
induced by the scalar product
X
hK, LiHS :=
hKeα , Leα i .
(4.3)
α
Physicists generally write this as:
hK, LiHS = tr K † L.
(4.4)
Proof. If K is a Hilbert-Schmidt operator, aK is a Hilbert-Schmidt operator as well for
every a ∈ K. If K, L are HS operators, then for every orthonormal basis {eα }, the following
equation holds:
X
X
2
2
2
(4.5)
||(K + L)eα || ≤ 2 ·
||Keα || + ||Leα || < ∞,
α
α
i.e., K + L is a Hilbert-Schmidt operator as well. By h·, ·i, we denote the scalar product
1/2
in the space of Hilbert-Schmidt operators, and ||K||HS = hK, KiHS (as usual, the scalar
product induces a metric).
A comparison of both spaces can be found in Table 4.1.1
1 Note that there would be many other different choices how the norm for Hilbert spaces could be defined
which all fulfil the required properties of a norm that can, e.g., be found in [AG81].
22
4.3. Connection with quantum mechanics
State
Operator
Norm
Operator
norm
Hilbert space
|f i ∈ H
H → H:pD̂ |xi = |x′ i
||f i| = hf |f i
||D̂||sup = sup |D̂ |f i |
|f i∈H
|f |≤1
Hilbert space of Hilbert-Schmidt operators
D̂ : H → H
Λ : D̂ → D̂√≡ (H → H) → (H → H)
||D̂||HS = tr D† D
||Λ|| = sup Λ(D̂) = sup tr Λ(D̂)† Λ(D̂)
D̂∈H
||D̂||≤1
D̂∈H
||D̂||≤1
Table 4.1: A comparison between general Hilbert spaces and Hilbert spaces with
Hilbert-Schmidt operators as basis.
4.3
Connection with quantum mechanics
Quantum mechanics is one of the most advanced theories in physics; most research performed
today is centred around it. There are many ways to describe the theory mathematically;
for our purposes, an abstract algebraic formulation based on the principles and structures
introduced in the previous part seems to be most apt. In the following, we present a concise
overview about the central elements of the theory, following the structure of the review given
in Ref. [Key02]. More elaborate introductions can be found in the usual textbooks on the
topic, e.g., [Sak94]. For those especially interested on the impact of quantum mechanics on
information theory, Refs. [NC00, Pre99, Key02] can be especially recommended.
Probabilistic processes lie at the very heart of quantum mechanics: Predictions made by
the theory hold only in the sense that probabilities for outcomes of measurements can be
provided. A large number of repeated experiments with the same preparation parameters
results in a distribution that states how many times a certain outcome will appear. Exactly
this distribution can be predicted by theory. The outcome of a single measurement can in
general never be forecast with certainty.
4.3.1
States and effects
Ideally, an experiment resulting in a probability distribution can be carried out by repeating
the following two processes until a sufficient amount of statistics has been gathered:
❏ Preparation of a quantum mechanical state according to some fixed procedure that
can be repeated a sufficient number of times.
❏ Measurement of some observable quantity, e.g., spin, energy, etc. Effects are a special
class of measurements which can result in either the answer “yes” or “no” according
to some probability distribution.
Note that we do not only deal with purely quantum mechanical states, but may also
encounter a mixture between classical and quantum mechanical properties (which are usually
termed hybrid systems) that our formalism must be able to account for. Measurement results
fall into the classical category since gauges in the macroscopic world are used to infer them
from the quantum system.2
Every quantum system can be completely characterised by its observable quantities which
in turn are characterised by self-adjoint operators. These operators form an algebra A as
2 The problem of how measurements of a quantum system are to be interpreted (or even how the whole
process can be described consistently) has been and still is one of the fundamental philosophical problems
of quantum mechanics. We take a pragmatic point of view here and do not consider the problem in greater
detail, but refer to the literature, e.g., [Aul00].
23
Chapter 4. Mathematical structures
introduced in Section 4.1.1; since we do only deal with finite-dimensional Hilbert spaces
here, we can restrict ourselves to subalgebras of B(H), i.e., A ⊂ B(H). A is called the
observable algebra of the system and is often identified with the system itself because it is
possible to deduce all properties of the system from its observable algebra. The dual algebra
of A is denoted by A∗ and is the algebra defined on the dual space.
To capture the notions of state and effect mathematically, we introduce two sets according to the following definition:
S(A) = { ̺ ∈ A∗ ̺ ≥ 0 ∧ ̺(1) = 1 }
E(A) = { A ∈ A A ≥ 0 ∧ A ≤ 1 }
(4.6)
(4.7)
S represents the set of states, while E contains all effects. For every tuple (̺, A) ∈ S×E, there
exists a map (̺, A) → ̺(A) ∈ [0, 1] which gives the probability p = ̺(A) that measuring
an effect A on a (system prepared in the) state ̺ results in the answer “yes”. Accordingly,
the probability for the answer “no” is given by 1 − p. ̺(A) is called the expectation value
of a state A; states are thus defined as expectation value functionals from an abstract point
of view. These expectation value functionals can by uniquely connected with a normalised
trace-class3 operator ̺ such that ̺(A) = tr(̺A). In principle, it would be necessary to
introduce two different symbols for the expectation value functional and the operator, but
for simplicity, we omit this complication.
We have to distinguish between two different kinds of states: Pure and mixed ones. This
is a consequence of the fact that both S and E are convex spaces: For two states ̺1 , ̺2 ∈ S(A)
and λ ∈ R, 0 ≤ λ ≤ 1, the convex combination λ̺1 + (1 − λ)̺2 is also an element of S(A).
The same statement holds for the elements of E(A). This decomposition provides a very nice
insight into the structure of both spaces. Extremal points in this space cannot be written
as a proper convex decomposition, i.e., x = λy + (1 − λ)z → λ = 1 ∨ λ = 0 ∨ x = y = z.
These can be interpreted as follows:
❏ For S(A), they are pure states with no associated classical uncertainty.
❏ For E(A), they describe measurements which do not allow any fuzziness as is, e.g.,
introduced by a detector which detects some property not with certainty, but only up
to some finite error (alas, all real-world detectors).
4.3.2
Observables
Until now, we have only been talking about effects, i.e., “yes”/“no” measurements, but not
about measurements with a more complicated result range which are necessary to describe
general observables. Although we would have to consider an infinite (even uncountable)
number of possible outcomes for a general description of quantum mechanics, it is sufficient
to consider only observables with a finite range for our purposes.4 Such observables are
represented by maps which connect elements x of a finite set R to some effect Ex ∈ E(A);
this in turn gives rise to a probability distribution px = ̺(Ex ). More formally, we can put
it as in the following:
A family E = {Ex },P
x ∈ R of effects Ex ∈ A if called a positive operator valued measurement (POVM) on R if x∈R Ex = 1.
3 For
trace-class operators, the trace is independent of the basis chosen to evaluate the trace.
is justified because quantum computers process states of the type (|0i , |1i)⊗n . Although quantum
computers can possess an arbitrary number of qbits, it is still a fixed and (which is most important) finite
number; additionally, we do not care for any continuous quantum properties of these objects.
4 This
24
4.3. Connection with quantum mechanics
Note that the Ex need not necessarily be projectors, i.e., Ex2 = Ex . Should this nevertheless be the case ∀x, the measurement is called a projective measurement.
Observables of this kind can be described by self-adjoint operators of the underlying
Hilbert space H which can (without any claim of formal correctness or even a proof) be
seen as follows: Every self-adjoint operator A on a Hilbert space H of finite dimension can
(because
P of the spectral theorem, cf., e.g., [AG81, Wei00]) be decomposed into the form
A = λ∈σ(A) λPλ where σ(A) denotes the spectrum of A and Pλ the projectors onto the
P
corresponding eigenspace. The expectation value λ λ̺(Pλ ) of P for a given state ̺ can
equivalently be calculated by ̺(A) = tr(̺A). Since this is the standard way of formulating
the expectation value of an operator, both points of view coincide.
4.3.3
Classical components
Systems consisting solely of quantum components are generally not to be found: At the
latest after a measurement has been performed, classical probabilities need to be accounted
for. Therefore, we need to pay attention to hybrid systems composed from quantum and
classical parts as well. Obviously, we have to orient ourselves along the lines of Section 4.3.1
to provide proper grounding for both possibilities. Consider a finite set X of elementary
events, i.e., all possible outcomes of an experiment. Again, S(A) and E(A) define the set of
states and effects, respectively, but this time, the observable algebra is given by all complex
valued functions from the set X to C as defined by
A = C(X) = { f : X → C } .
By identifying the function f with the operator fˆ given by
X
fˆ =
fx |xi hx|
(4.8)
(4.9)
x∈X
where |xi denotes a fixed orthonormal basis, the probability distribution can be interpreted
as an operator algebra similar to the quantum mechanical case because fˆ is obviously an
element of B(H). Thus, C(X) can be used as an observable algebra A along any other
quantum mechanical or classical constituent of a multipartite composite system.
4.3.4
Composite and hybrid systems
Since quantum mechanical and classical systems can be described with very similar structures, the presented formalism is obviously well suited for the presentation of composite
systems. Let A ⊂ B(H) and B ⊂ B(K) be systems given in terms of their observable
algebras; the composite system is then given by
A ⊗ B ≡ span { A ⊗ B A ∈ A, B ∈ B } .
(4.10)
Three cases for the choice of H, K can be distinguished:
❏ If both systems are quantum, then A ⊗ B = B(H ⊗ K).
❏ If both systems are classical, then A ⊗ B = C(X × Y ) with C as defined by Eqn. 4.8
❏ If A is classical and B is quantum mechanical, we have a hybrid system; the composite
observable algebra is then given by C(X)⊗B(H) which cannot be simplified any further.
Observables are operator-valued functions in this case, as expected.
25
Chapter 4. Mathematical structures
4.4
Domain theory
The definition of a proper and sound mathematical semantics for a programming language
necessitates apt structures which can be used as a solid ground underlying the work. The
method we use – denotational semantics – is conventionally based on semantic domains,
which in turn rely on partial orders and recursion theory. The purpose of this section is to
introduce the elements required for the semantic description of cQPL which will be given in
Chapter 5; more details are available, e.g., in [GS90, Win93].
4.4.1
Basic definitions
Definition 4.4.1 (Partial order). A partial order (P, ⊑)is a set P on which there is a
binary relation ⊑ for which the following properties hold ∀p, q, r ∈ P :
❏ p ⊑ p (reflexive)
❏ p ⊑ q and q ⊑ r ⇒ p ⊑ r (transitive)
❏ p ⊑ q and q ⊑ p ⇒ p = q (antisymmetric)
Definition 4.4.2 (Upper bound). For a partial order (P, ⊑) and a subset X ⊆ P , p ∈ P
is an upper bound of X if and only if ∀q ∈ X : q ⊑ p.
The element p is a least upper bound if:
❏ p is an upper bound of X
❏ For all upper bounds q of X, p ⊑ q
Remark 4.4.1. Note that it follows from the definition that the least upper bound is unique.
Definition 4.4.3 (ω-chain). Let (D, ⊑D ) be a partial order. An ω-chain of the partial
order is an increasing chain d0 ⊑D d1 ⊑D · · · ⊑D dn ⊑ · · · of elements of the partial order.
Note that ω represents the increasing chain of natural numbers N0 .
Definition 4.4.4 (Complete partial order). The partial order (D, ⊑D ) is a complete
partial order (cpo) if it has least upper bounds of all ω-chains, i.e., any increasing chain
{dn |n ∈ ω} of elements in D has a least upper bound ⊔{dn |n ∈ ω}, written as ⊔n∈ω dn .
(D, ⊑D ) is a cpo with bottom if it is a cpo which has a bottom element (often also called
least element) ⊥D for which ⊥D ⊑ d ∀d ∈ D holds.
Definition 4.4.5 (Directed-complete partial order). A partial order (D, ⊑) in which
every directed subset has a supremum is called directed-complete partial order (dcpo).
Definition 4.4.6 (Monotone function). A function f : D → E between cpos D and E
is monotonic if and only if ∀d, d′ ∈ D:
d ⊑ d′ ⇒ f (d) ⊑ f (d′ ).
(4.11)
Definition 4.4.7 (Continuous function). A function f : D → E between cpos D and E
is continuous if and only if it is monotonic and for all chains do ⊑ d1 · · · ⊑ dn ⊑ · · · in D
there holds
G
f (dn ) = f (⊔n∈ω dn ) .
(4.12)
n∈ω
26
4.4. Domain theory
4.4.2
A fixed point theorem
Definition 4.4.8 (Fixed point). Let f : D → D be a continuous function on a cpo D
with bottom ⊥D . A fixed point of f is an element d ∈ D such that f (d) = d. A prefixed
point of f is an element d ∈ D such that f (d) ⊑ d.
Theorem 4.4.1 (Fixed-point theorem). Let f : D → D be a continuous function on a
cpo with a bottom D. Define
G
fix (f ) =
f n (⊥).
(4.13)
n∈ω
Then fix (f ) is a fixed point of f and the least prefixed point of f , i.e.:
❏ f (fix (f )) = fix (f )
❏ If f (d) ⊑ d then fix (f ) ⊑ d
Consequently, fix (f ) is the least fixed point of f .
Proof. It follows from continuity of f that
f (fix (f )) = f (⊔n∈ω f n (⊥)) =
G
f n+1 (⊥)
(4.14)
n∈ω
=
G
⊥ ∪ f n+1 (⊥)|n ∈ ω
n∈ω
= fix (f ) .
=
G
f n (⊥)
(4.15)
n∈ω
(4.16)
Thus fix (f ) is a fixed point because f (fix (f )) = fix (f ) is exactly the required property
of a fixed point (adding ⊥ in step 4.15 is justified because the least upper bound is not
influenced by this). Suppose d is a prefixed point. Certainly, ⊥⊑ d. By monotonicity,
f (⊥) ⊑ f (d). But d is a prefixed point, i.e., f (d) ⊑ d, so f (⊥) ⊑ d, and by induction
f n (⊥) ⊑ d ∀n ∈ ω. Thus, fix (f ) = ⊔n∈ω f n (⊥) ⊑ d
Remark 4.4.2. Note that it is customary to define
G
YD f ≡
f n (⊥)
(4.17)
n∈ω
such that Y D is obviously a function (D → D) → D which maps functions to their fixed
points; it is thus termed the fixed point combinator.
Definition 4.4.9 (Scott topology5 ). 6 Let D be a dcpo. A subset A is called Scott closed
if it is a lower set7 and is closed under suprema of directed subsets.8 Complements of closed
sets are called Scott open; they are the elements of σD , the Scott topology on D.
Theorem 4.4.2. A function f : Dn → Dm is continuous in the sense of definition 4.4.7
(i.e., Scott continuous) if it is topologically continuous with respect to the Scott topology.
Proof. Cf. Ref. [AJ94, Theorem 2.3.4]
The last definition and theorem are quite technical, but we need them for the proof of
Theorem 4.5.4 later on.
5 This
definition is taken from [AJ94].
topological space is a set X together with a collection T of subsets where the empty set and X are in
T , the union of any collection of sets in T is in T and the intersection of any pair of sets in T is also in T .
7 A lower set is a finite, non-empty downward-closed subset of a partial order, i.e., {x|x ⊑ y}.
8 A subset A of a poset is directed if it is nonempty and each pair of elements has an upper bound in A.
6A
27
Chapter 4. Mathematical structures
4.4.3
Constructions on domains
We will use the following constructions on dcpos to create new dcpos:
❏ D1 × D2 × · · · × Dn denotes n-tuples respectively cartesian domains. The weaker-than
relation is defined such that
(x1 , x2 , . . . , xn ) ⊑D1 ×D2 ×···×Dn (y1 , y2 , . . . , yn ) ⇔ xi ⊑Di yi
(4.18)
for i = 1, . . . , n and xi ∈ Di , yi ∈ Di .
❏ D1 ⊗D2 ⊗· · ·⊗Dn represents the smash product which identifies all tuples that contain
one or more ⊥-elements. Example: (d1 , ⊥2 ), (⊥1 , d2 ) and (⊥1 , ⊥2 ) are all identified
with a new bottom element ⊥D1 ⊗D2 for di ∈ Di Formally, the new domain D ⊗ E is
the set
{(x, y) ∈ D × E|x 6= ⊥ and y 6= ⊥} ∪ {⊥D⊗E }.
(4.19)
❏ D1 + D2 + · · · + Dn is the separated sum domain which consists of all elements in Di
together with a new bottom symbol ⊥D1 +D2 +···+Dn (usually abbreviated to ⊥D ).
❏ The coalesced sum D1 ⊕ D2 ⊕ · · · ⊕ Dn is similar to the separated sum, but the new
bottom element ⊥D is gained by identifying all elements d1 + d2 + · · · + dn (di ∈ Di )
which contain one or more of ⊥Di .
Summary
❏ Lifting is the operation that adds bottom element to a domain D; the result is denoted
by D⊥ ; continuity is not influenced by this.
We have provided the basic framework required to build the denotational semantics of cQPL. This framework is composed of two parts: On the one hand, we
need an abstract representation of quantum mechanics to account for the physical properties of the programming language. On the other hand, the concept
of partial orders builds the basis for defining semantic domains, i.e., the space
which will be used to place the equations describing the semantics of cQPL in.
The choice of partial orders for that is especially justified by the fact that fixed
points can be constructively obtained in them. These in turn are required to solve
recursive domain equations that will be needed to give a denotation for several
language constructs of cQPL, most important the communication features.
4.5
cp-Maps and their representation
In quantum mechanics, time evolution is described by transformations of density matrices
with an operator Λ that is called a superoperator [Pre99, NC00, Key02].
Definition 4.5.1 (Superoperator). A superoperator Λ : B(H) → B(H) has the following
properties for all density operators ̺ ∈ D with ̺′ = Λ(̺):
❏ Λ is linear.
❏ ̺† = ̺ ⇒ ̺′† = ̺′ (hermeticity is preserved).
❏ tr ̺′ = 1 if tr ̺ = 1 (trace preserving).
28
4.5. cp-Maps and their representation
❏ Λ⊗1 is semidefinite positive (∀n ∈ N : Λ⊗1n ≥ 0), i.e., Λ is a completely positive map.
In other words, this means that Λ is not only semidefinite positive (̺′ is nonnegative
if ̺ is nonnegative) on HA , but also on any possible extension HA ⊗ HB .
Note that if dissipative processes (e.g., postselection of observed events) are considered,
the second condition is loosened to tr(̺′ ) ≤ 1.
4.5.1
Operator-sum representation
Kraus [Kra83] proved a result about the decomposability of completely positive maps which
is ubiquitous in quantum information theory:
Theorem 4.5.1 (Kraus representation theorem). A superoperator Λ as defined in
PN
Def. 4.5.1 can be written as a partition of 1 = k=1 A†k Ak where Ak are linear operators
acting on the Hilbert space of the system such that
̺′ = Λ(̺) =
N
X
k=1
Ak ̺A†k ∀̺ ∈ D
(4.20)
for any density matrix ̺ that represents a mixed or a pure state.
Proof. Cf. Ref. [NC00, Pre99, Kra83]
To illustrate this representation, consider the situation that the system under consideration is in contact with a much larger environment, a common situation for physical problems.
Together, both systems form a closed quantum system. State transformations in this combined system can be described by a unitary transformation U ∈ U (dim(H) · dim(Henv ))
where H denotes the Hilbert space of the system under consideration and Henv the Hilbert
space of the environment. Assume that the environment is in a pure state |e0 i he0 |.9 The
density operator of the system under consideration after the unitary operation was applied
to the total system can be recovered by tracing out the environment:
̺′ = Λ(̺) = tr(U ̺ ⊗ |e0 i he0 | U † )
X
=
hek | U (̺ ⊗ |e0 i he0 |)U † |e0 i
(4.21)
(4.22)
k
=
X
k
=
X
hek | U |e0 i ̺ he0 | U † |ek i
(4.23)
Ak ̺A†k .
(4.24)
k
In the last step, Ak is defined by Ak ≡ hek | U |e0 i.
Remark 4.5.1. We say that a set of Kraus operators {Ak } implements a cp-map Λ if
P
∀̺ ∈ D : k Ak ̺A†k = Λ(̺). This simplifies the further description.
Theorem 4.5.2. The operation elements of a given superoperator Λ are not unique: If {Ej }
is a set of Kraus operators, then a different set of Kraus operators {Fk } describes the same
9 This assumption holds without loss of generality because it can be shown that a system can be purified
by introducing extra dimensions which do not have any physical consequences.
29
Chapter 4. Mathematical structures
operation if and only if there exists a unitary matrix U ∈ U (n) with n = card({Ek }) (where
card(X) is the cardinality of the set X) such that
X
Ukj Ej .
(4.25)
Fk =
j
Note that the shorter set may be padded with zero elements until the cardinality of both
matches.
Proof. Cf., e.g., Ref. [NC00, Theorem 8.2] or Ref. [Pre99].
Remark 4.5.2. Let {Ak } be a set of Kraus operators that represents the cp-map Λ. Note
that if any number of elements Ai is taken from {Ak }, the set still remains a completely
positive map, but is not trace preserving any more.
Remark 4.5.3. Note that superoperators are elements of B(H) which makes it possible to
apply many theorems of linear operator algebra to superoperators. In fact, superoperators can
be used as elements of a Hilbert space as defined in Section 4.2.2. The distinction between
operators and superoperators in physics is therefore in general superfluous.
Remark 4.5.4. It can be shown that the number of Kraus elements needed to express any
arbitrary completely positive map T : B(H1 ) → B(H2 ) is bounded by dim(H1 ) · dim(H2 ),
confer, e.g., [Pre99, p. 102]).
4.5.2
Equivalence of Kraus operators
The unitary connection between two sets of Kraus operators defined in Equation 4.25 gives
rise to an equivalence relation between such sets. Two sets {Aj } and {Bk } are members of
the same equivalence class if there is a unitary matrix which connects both representations:
n
X
A∼
= B ⇐⇒ ∃U ∈ U (n) : Ai =
Uij Bj with i = 1, . . . , n.
(4.26)
j=1
The set of all sets of Kraus operators inducing the same map Λ is defined in the obvious
way:
n
o
X
K(Λ) ≡ {Ak }
Ak ̺A†k = Λ(̺) ∀̺ ∈ D .
(4.27)
k
If we talk about a set of Kraus operators or simply Kraus operators in the following, we
always mean an arbitrary set which is an element of the equivalence class inducing the same
cp-map (i.e., an element of K(Λ)), but will not mention this explicitly every time.
4.5.3
A partial order for Kraus operators
The Löwner partial order [Löw34] for two density operators A, B : B(H1 ) → B(H2 ) is given
by
A ⊑ B ⇐⇒ (B − A) > 0.
(4.28)
This partial order can be extended to sets of Kraus operators by defining
{Ai } ⊑ {Bi } ⇐⇒∀̺ ∈ D ∀n ∈ N :
X
i
(Bi ⊗ 1n )̺(Bi ⊗ 1n ) −
†
30
X
k
(Ak ⊗ 1n )̺(Ak ⊗ 1n )
†
!
> 0.
(4.29)
4.5. cp-Maps and their representation
Partial orders are often interpreted as approximations: If an element A is weaker than
B (A ⊑ B), then A is said to approximate B. This point of view will come handy when we
consider solutions of fixed point equations in the denotational description.
It is necessary for our work to see that Kraus operators form a complete partial order.
For this, observe first the following theorem:
Theorem 4.5.3. The partial order on all density operators ̺ ∈ D given by the Löwner
partial order ⊑ is complete.
Proof. Cf. Ref. [Sel04b, Proposition 3.6]
From this, we can deduce the required statement:
Theorem 4.5.4. The partial order for cp-maps defined by the extended Löwner partial order
given by Eqn. 4.29 is complete, i.e., it forms a cpo.
Proof. Let {A1 } ⊑ {A2 }, . . . be an increasing chain of topologically continuous (and therefore
monotone because of 4.4.2, 4.4.7 and 4.4.6) Kraus operators. Because of Definition 4.29, the
relation ̺1 ⊑ ̺2 is preserved by applying {A1 }, {A2 } with {A1 } ⊑ {A2 } to ̺1 , ̺2 . An ωchain of density operators is conserved if an increasing chain of Kraus operators is applied to
it. Because of Theorem 4.5.3, the fact that the previous consideration applies to all density
operators in D and the uniqueness of the least upper bound, the extended Löwner partial
order is complete as well.
4.5.4
Kraus aggregations
We mentioned that superoperators applied to density matrices describe quantum mechanical processes. Operations performed one after another can therefore be described by the
consecutive application of the corresponding superoperators:
If the sets
given by
{A1k }
̺′ = Λ1 (̺), ̺′′ = Λ2 (̺′ ) ⇒ ̺′′ = Λ2 (Λ1 (̺))
and
{A2k }
(4.30)
implement Λ1 and Λ2 , then the same state transformation is
̺′′ =
XX
k
†
†
A2k A1l ̺A1l A2k .
(4.31)
l
We call a collection of sets of Kraus operators that are to be applied subsequently an
aggregation of (sets of ) Kraus operators or simply Kraus aggregation; the Kraus sets involved
are written as a list of the form
Γ = {A1k }, {A2k }, . . . , {Ank }
The list Γ gives rise to the following quantum mechanical operation:
XX X
†
†
Γ(̺) = ̺′ =
···
Ankn · · · A2k2 A1k1 ̺A1k1 A2k2 · · · Ankn †
k1
k2
(4.32)
(4.33)
kn
List concatenation is formally described by the operator ◦:
Γ1 = {A1k }, {A2k }, . . . , {Ank },
(4.34)
Γ2 = {Bk1 }, {Bk2 }, . . . , {Bkm }
⇒ Γ1 ◦ Γ2 ≡ {A1k }, {A2k }, . . . , {Ank }, {Bk1 }, {Bk2 }, . . . , {Bkm }
31
(4.35)
(4.36)
Chapter 4. Mathematical structures
i.e., the effect of Γ1 ◦ Γ2 on a state ̺ is the same as if first Γ1 and then Γ2 would have been
applied. Note (since this is a potential source of confusion) that the list is “processed” from
left to right, not from right to left!
A Kraus aggregation can also consist of multiple sub-aggregations which are prefixed by
some scalar. Formally, we use the operator + to denote this:
Γ′ = p 1 · Γ1 + · · · + p n · Γn .
(4.37)
1
1
· {NOT} + · {1}.
2
2
(4.38)
10
P If the pi ∈′ R are to be interpreted as probabilities, the normalisation condition is
n pn ≤ 1. Γ can thus be seen as a formal combination of lists. The interpretation of such
an aggregation is straightforward: With probability pk , the Kraus aggregation Λk is selected
whenever Λ′ acts on a density operator. Obviously, lists of this form are apt to introduce
mixed states into the Kraus list formalism. Consider, for example, the aggregation
∆=
The effect of it is to apply the unconditional not-operation (which maps |0i → |1i and
|1i → |0i and may, for example, be implemented with σ̂x ) with probability 0.5 and to leave
the state unchanged with the same probability. If this aggregation is applied to, e.g., the
following (pure) density operator
̺ = |0i h0| ,
(4.39)
the resulting state is the impure density operator given by
1
1
{NOT}(̺) + {1}(̺)
2
2
1
1
1
1
= |1i h1| + |0i h0| = {|1i} + {|0i}
2
2
2
2
̺′ = ∆(̺) =
(4.40)
(4.41)
which describes an impure mixture between {|0i} and {|1i}.
Remark 4.5.5. Note that we will use Kraus lists prefixed with probabilities to describe
different measurement outcomes when we provide the semantics of cQPL in Chapter 5. The
physical way to think about such operations is to take a density operator ̺ and apply the
Kraus elements for the projective measurements on it; this results in the state
X
X
̺′ =
Mk ̺Mk† =
Em (̺)
(4.42)
k
k
where Mk are the projection operators and Em (̺) ≡ Mk ̺Mk† . The probability to obtain the
measurement outcome k is given by
p(m) = tr(Em (̺)).
(4.43)
The probability factors in Kraus aggregations can be calculated in exactly this way; both
points of view therefore provide the same information.
Note that we allow the pre-factors of the sub-aggregations to depend on parameters which
make the complete aggregation dependent on the disjoint union of the set of parameters used
for the sub-aggregations. This is necessary to describe Kraus aggregations dependent on
probability distributions which are unknown before a initial state is given or the outcomes
10 The sum can be smaller than 1 to account for the possibility of non-termination which will happen with
P
probability 1 − pi . It also allows to describe non trace-preserving effects.
32
4.5. cp-Maps and their representation
of some measurements are known. The following example shows a Kraus aggregation where
the first sublist depends on the parameters a1 and a2 and the second on a1 and a3 ; the
complete aggregation obviously depends on a1 , a2 and a3 :
Γ(a1 , a2 , a3 ) = p(a1 , a2 ) · Γ1 + p(a1 , a3 ) · Γ2 .
(4.44)
For a Kraus aggregation of the most general form (where {ai } denotes the set of parameters
for the ith sub-aggregation) given by
Γ(∪i {ai }) =
X
i
pi ({ai })Γi ,
(4.45)
the normalisation condition is obviously still given by
X
i
pi ({ai }) ≤ 1
(4.46)
which necessitates that 0 ≤ pi ≤ 1 ∀i (this is supposed to hold for all p used in the following).
It is possible to contract Kraus (sub-)aggregations which consist of more than one element
to a shorter form because two Kraus sets {Ai } and {Bi } can be contracted to a new set
{Ck } which describes the subsequent application of both initial sets, as the following simple
calculation shows:
({Ak }, {Bi })(̺) = {Bi }({Ak }(̺)) =
N X
N
X
B̂k Âi ̺†i B̂k†
(4.47)
k=1 i=1
2
=
N
X
Ĉn ̺Ĉn†
(4.48)
n=1
with
Ĉn ≡ B̂⌈n/N ⌉ Ân mod N .
(4.49)
Recall that different set cardinalities can be compensated by adding an appropriate number
of zero operators to the smaller set. Since the calculation is valid ∀̺ ∈ D, the new single
element aggregation {Ci } is a unique replacement for the aggregation {Ai }, {Bi }.
Based on this contraction, it is possible to define a standard representation for Kraus
aggregations which is easier to handle formally when aggregations must, for example, be
compared.
With
n
o
X
P ≡ {pi ({ai })} aik ∈ R∀i, k, 0 ≤ pi ({ai }) ≤ 1,
pi ({ai }) ≤ 1
(4.50)
i
being the set of all possible parametrised probability distributions and
K ≡ { Λ Λ is a cp-map }
(4.51)
being the set of all unparametrised Kraus aggregations contracted to the normal form given
by Eqn. 4.48, we can finally define the set of all possible Kraus aggregations formally by
nX
o
pi Λi {pi } ∈ P ∧ Λk ∈ K .
A≡
(4.52)
i
33
Chapter 4. Mathematical structures
4.5.4.1
A partial order for Kraus aggregations
For a Kraus aggregation of the contracted normal form Λ = {Ck }, the definition for a
partial order can be directly transferred from Equation 4.28. If the aggregation contains
sub-aggregations, ⊑ is formally a function dependent on the parameters of the aggregation:
For Γ1 = Γ1 (A1 , . . . , An ) and Γ2 = Γ2 (B1 , . . . , Bn ), the partial comparison Γ1 ⊑ Γ2 becomes a function (A1 , . . . , An , B1 , . . . , Bn ) → {true, false}, i.e., the comparison depends on
the parameters of both sets of parameters involved. Note that this does not concern Kraus
aggregations where all coefficients have defined scalar values. Basically, the parametrised
comparision is nothing else than a comparision of all elements of an unfolded Kraus aggregation as defined in Section 5.3.6.2 followed by folding everything back afterwards.
4.5.4.2
Equivalence of Kraus aggregations
One possible task of denotational semantics is to decide wether two programs which look
different perform the same actions, i.e., if their semantics coincide. This question is in
general complicated to answer constructively. Nevertheless, it is possible for some cases. We
will consider this problem in more detail in Chapter 5. At this point, we are interested in
the question when two Kraus aggregations are semantically equivalent, i.e., induce the same
physical operations. The method used for this is almost identical to the method used for
Kraus sets. Consider two aggregations Γ1 and Γ2 given in the contracted normal form, i.e.,
Γ1 =
N
X
p1k Λ1k
(4.53)
p2k Λ2k .
(4.54)
k=1
Γ2 =
N
X
k=1
Let Sym(M ) by the symmetric group over the finite set M . Both lists are equivalent if
(but not only if) the following condition holds:
Γ1 ∼
= Γ2 ⇔∃ϕ ∈ Sym([1, . . . , N ])∀k ∈ [1, N ]∀̺ ∈ D :
p1k (A1k )
=
p2P(k) (A2P(k) )
∧
Λ1k (̺)
=
Λ2P(k) (̺).
(4.55)
(4.56)
Note that this equivalence requires that the same Kraus operators are used in both lists; it is
nevertheless possible that a different set of Kraus operators prefixed by another probability
distribution induces the same action. The criterion given here is thus sufficient, but not
necessary.
The set E of all aggregations that are equivalent in this sense can be defined analogous
to Eqn. 4.27:
(4.57)
E(Λ) ≡ { Λi ∈ A Λi ∼
= Λ }.
This definition is not very satisfying from a constructive point of view: There is no
simple way to systematically decide if the effects of two aggregations coincide. This can be
improved by giving an explicit criterion for the equivalence between two Kraus aggregations.
We consider the special case of two lists which are composed of the same operators, but are
ordered differently. This happens, for example, when statements in a program are reordered.
With the method given below, we can thus get a criterion to decide if such reorderings
preserve the semantics of programs which is a very important case.
Unparametrised Kraus aggregations can always be written in the standard form given
by Eqn. 4.48 and are thus equivalent to a Kraus set; this again is equivalent to some cpmap Λ. Because we have seen in Section 4.2.2 that such cp-maps form a Hilbert space, it
34
4.5. cp-Maps and their representation
is reasonable to define a commutator (analogous to the case of regular operators) for two
Hilbert-Schmidt operators Λ1 , Λ2 by setting:
[Λ1 , Λ2 ] ≡ Λ1 Λ2 − Λ2 Λ1 .
(4.58)
The following theorem provides a condition for the identity between a list of operators and a
permutation of it which is based on elementary commutators of the elements. Unfortunately,
this is not a general solution since the effect of the theorem might just be to rephrase the
problem in different terms if the structure of the commutators is not apt.
Theorem 4.5.5. Let A1 , A2 , . . . , An be operators and let ϕ ∈ Sym(n) be a permutation of
the index set. Then the difference between the commuted product Aϕ(1) · Aϕ(2) · · · Aϕ(n) and
A1 · A2 · · · An can be written as11
Aϕ(1) · Aϕ(2) · · · Aϕ(n) = A1 · A2 · · · An +
X
(s,t)
Xs,t · [As , At ] · Ys,t · Zs
(4.59)
where (s,t) runs over all inversions of ϕ, i.e., 1 ≤ t < s ≤ n and ϕ−1 (s) < ϕ−1 (t) and where
Xs,t =
Y
Ys,t =
Aϕ(i) ,
Y
Zs =
Aϕ(i) ,
Y
Ak .
(4.60)
s<k≤n
1≤i≤n
1≤i≤n
ϕ(i)<s, i<ϕ−1 (t)
ϕ(i)<s, i>ϕ−1 (t)
Proof. We prove this statement by induction on the list length. The cases n = 0 and n = 1
are trivial. The induction step n → n + 1 can be seen as follows. Let j ∈ [0, . . . , n + 1] such
that ϕ(j) = n + 1. Then,
Aϕ(1) · Aϕ(2) · · · Aϕ(j) · · · Aϕ(n+1) =
Aϕ(1) · · · Aϕ(j−1) · Aϕ(j+1) · Aϕ(j) · Aϕ(j+2) · · · Aϕ(n+1) +
Aϕ(1) · · · Aϕ(j−1) · [Aϕ(j) , Aϕ(j+1) ] · Aϕ(j+2) · · · Aϕ(n+1) = ... =
X
Xn+1,t · [An+1 , At ] · Yn+1,t
Aϕ(1) · · · Aϕ(j−1) · Aϕ(j+1) · · · Aϕ(n+1) Aϕ(j) +
{z
} | {z } n+1,t
|
I.H.
(4.61)
A(n+1)
where 1 ≤ t < n + 1, ϕ(n + 1) < ϕ−1 (t),
Xn+1,t =
Y
Yn+1,t =
Aϕ(i) ,
Y
Aϕ(i)
1≤i≤n+1
1≤i≤n+1
ϕ(i)<n+1, i<ϕ−1 (t)
ϕ(i)<n+1, i>ϕ−1 (t)
and n + 1 is obviously fixed. The final resulting equation thus resembles exactly the form
given by Eqn. 4.59, but we have not used the induction hypothesis yet. Now, by using the
induction hypothesis, it follows that
Aϕ(1) · · · Aϕ(j−1) · Aϕ(j+1) · · · Aϕ(n+1) = A1 · · · An +
X
(s′ ,t′ )
Xs′ ,t′ · [As′ , At′ ] · Ys′ ,t′ · Zs′ (4.62)
11 This representation (which is much more elegant than the one derived by the author) was provided by
Volker Strehl.
35
Chapter 4. Mathematical structures
where the primed identifiers are defined by 1 ≤ t′ < s′ ≤ n, and ϕ−1 (s′ ) < ϕ−1 (t′ ). By
placing this into the part of Eqn. 4.61 marked by I.H., we see that
X
Xs′ ,t′ · [As′ , At′ ] · Ys′ ,t′ Zs′ · An+1 +
Aϕ(1) · · · Aϕ(n+1) = A1 · · · An +
(s′ ,t′ )
X
(n+1,t)
Xn+1,t · [An+1 , At ] · Yn+1,t
= A1 · · · An An+1 +
= A1 · · · An+1 +
X
(s′ ,t′ )
X
Xs′ ,t′ · [As′ , At′ ] · Ys′ ,t′ Zs′ An+1 +
| {z }
Zs
X
Xn+1,t · [An+1 , At ] · Yn+1,t
(n+1,t)
Xs,t [As , At ]Ys,t Zs
(4.63)
(s,t)
where the unprimed variables are now given by 1 ≤ t < s ≤ 1 and ϕ−1 (s) < ϕ−1 (t); the
condition for k in Zs is now obviously s < k ≤ n + 1. The resulting Equation 4.63 has thus
the form for n + 1 as required by the statement.
To illustrate this theorem (note, additionally, that a little program to calculate all elements of the commutator sum is available), consider the permutation given by
1 2 3 4 5
.
(4.64)
5 2 3 1 4
The inversions (s, t) are all pairs of elements in the permuted list where a bigger element
is on the left side of a smaller element, in this case: (5, 2), (5, 3), (5, 1), (5, 4), (2, 1), (3, 1).
Note that the inversions characterise the list completely, cf., e.g., [Knu98, Section 5.1.1].
The method defined above is a variant of insertion sort which is a standard sorting method,
covered, e.g., in [SF96]. This can be seen by inspecting the conditions imposed by the
products defining X, Y and Z:
❏ For Xs,t , ϕ(i) < s, i < ϕ−1 (t) selects all i such that the corresponding elements in the
permuted list are smaller than the element s of the inversion and are placed on the
left hand side of the element t in the permuted list. For (5, 1), the condition would
select i = 2, 3.
❏ The conditions for Y make sure that again only elements which are smaller than s are
selected. This time, they additionally have to be on the right hand side of t in the
permuted list.
❏ Z specifies all elements which are on the right hand side of s in the unpermuted list.
By applying these rules, we can calculate the following sets for each inversion:
(5, 1) → X : i = 2, 3; Y : i = 5
(5, 2) → Y : i = 3, 4, 5
(5, 3) → X : i = 2; Y : i = 4, 5
(5, 4) → X : i = 2, 3, 4
(3, 1) → Z : k = 4, 5
(2, 1) → Z : k = 3, 4, 5
36
4.5. cp-Maps and their representation
This leads to the following identity that is provided by Eqn. 4.59 (note that we use i instead
of Ai to simplify the notation):
52314 = 12345 + 231[5, 4] + 2[5, 3]14 + [5, 2]314 + 23[5, 1]4 + 2[3, 1]45 + [2, 1]345
= 12345 + 23154 − 23145 + 25314 − 23514 + 52314 − 25314 + 23145−
23145 + 23145 − 21345 + 21345 − 12345
= 12345 − 12345 + 23154 − 23154 + 25314 − 25314 + 23145 − 23145+
21345 − 21345 + 23514 − 23514 + 52314
= 52314
It it also instructive to observe the following two identies because they illuminate the
induction step:
14532 = 12345 + 14[5, 3]2 + 143[5, 2] + 1[4, 3]25 + 13[4, 2]5 + 1[3, 2]45
1432 = 1234 + 1[4, 3]2 + 13[4, 2] + 1[3, 2]4
Remark 4.5.6. Because the proof has only made use of general properties of permutations
and of the definition of the commutator, it is not only applicable to Hilbert-Schmidt-operators
as we need, but also for any other objects fulfilling the mentioned properties.
37
Summary
We have explained how to represent quantum operations by cp-maps and these in
turn by a sum of Kraus operators. The Löwner partial order defined for density
matrices was generalised to Kraus operators; this order is complete and is therefore
a cpo as introduced in the beginning of this chapter. Since the denotational
semantics of cQPL will require lists of Kraus operators, we have introduced Kraus
aggregations to handle this formally. Since it is one of the problems of denotational
semantics to decide whether two given programs are equal or not, we have also
derived general and specific criteria for the equivalence of Kraus aggregations.
38
A map is not the territory.
Alfred Korzybski, Science and Sanity – An
Introduction to Non-Aristotelian Systems
and General Semantics
5
Formal denotational semantics
In this chapter, we are going to define a denotational semantics for cQPL, the communication
capable version of QPL [Sel04b]. Before we get into the details, we will give an overview
about the ideas of denotational semantics in general, present a survey of the denotational
semantics of QPL (because we reuse some ideas for the semantics of cQPL) and show why
the approach of annotation-based QPL must fail for communicating programs.
5.1
Fundamentals of denotational semantics
Denotational semantics is a well-understood standard method of theoretical computer science which is used to assign precise and mathematically sound and rigorous semantics to syntactically specified programs; introductions are, e.g., given in Refs. [Mos90, Win93, Rey98].
In this section, we will try to present an elementary introduction to the field. We align our
description along the lines of [Mos90, Section 1–3].
Computer programs are (usually) specified in the form of a textual description; this
description must follow certain rules defined by a grammar. Usually, context-free grammars
are used for this purpose because they are the most apt choice for that kind of problem.
They are defined as follows regarding to [AB02, Sch01]:
Definition 5.1.1 (Context-free grammar). A context-free grammar G is a four-tuple
(N, T, P, s0 ) where N is a finite set of nonterminal symbols, T is a finite set of terminal
symbols with T ∩ N = ∅, P ⊆ N × (N ∪ T )∗ is a finite set of productions and s0 ∈ N is the
start symbol.
As a very simple example, consider a grammar for binary strings of the form 0, 01,
100110, . . . which is recursively given by1
B ::= ’0’ | ’1’ | B’0’ | B’1’.
(5.1)
The terminal symbols2 are 0 and 1, the non-terminal3 symbol is B, and the start symbol is
1 In general, one has to distinguish between abstract and concrete syntax respectively grammars defining
these. The latter is used to specify a representation of programs that can be processed with parsers;
for that, some syntactical elements for disambiguation of certain constructions needs to be introduced.
Additionally, the capabilities and, especially, limitations of different parsing techniques need to be considered
when specifying a concrete grammar. Abstract syntax, on the other hand, is a representation of a program
stripped down to the bare minimum that is able to include all available information; additionally, the
structure of the syntax can be chosen such that it is not most suited for parsing, but for further analysis and
processing of the program. Usually, data structures in the form of trees are used to realise abstract syntax.
2 A constant symbol which cannot be resolved any further, cf. Appendix B.
3 A symbol whose definition consists of a chain of terminal and (at least one) non-terminal symbols and
can thus be resolved further.
39
Chapter 5. Formal denotational semantics
obvious because there is only one non-terminal. The productions are defined by Equation 5.1;
explicitely, they are given by{B × 0, B × 1, B × B0, B × B1}.
This grammar defines the syntactical representation of binary numerals. The really
interesting thing, however, is not how numerals look like, but instead what they mean – in
other words, the semantics of numerals. Obviously, the meaning of a binary numeral is some
natural number, so finding semantics for a binary string is equivalent with constructing a
method which assigns the appropriate natural number to a given syntactical representation
of a binary numeral. The constitutional parts of which the grammar is made up of are called
phrases. In our case, these are given by the strings 0 and 1 and the productions B0 and B1.
Denotational semantics assigns a meaning to sentences constructed according to a given
grammar by assigning a meaning to every elementary phrase of a grammar. The meaning of
phrases which are constructed from multiple sub-phrases (e.g., B0 in the example grammar)
is given by the meaning of these sub-phrases. The meaning of a complete program is thus
determined by the meaning of its constituents. The denotational approach is – in a nutshell
– characterised by the following points:
❏ Denotational semantics assigns some appropriate semantic object to every phrase of
the grammar; the object is called the denotation of the phrase.
❏ Valuation functions are used to connect syntactical objects with their semantical counterparts. For example, BIN is a valuation function that maps text strings consisting
of a series of ’0’ and ’1’ to a natural number.
❏ The denotation of compound phrases must only depend on the denotations of the
sub-phrases, i.e., JA1 , . . . , An K = f (JA1 K, . . . , JAn K). This is also known as the compositionality principle.
The valuation functions for binary numerals can be represented by the following equations:
BIN J0K
BIN JB0K
=
=
0
2 · (BIN JBK)
BIN J1K
BIN JB1K
=
=
1
(2 · (BIN JBK)) + 1
The double brackets JK are used to distinguish between the realms of syntax and semantics, while the valuation function BIN is used to map the phrases in these brackets to natural
numbers, their denotations. Thus, the domain of this function is the semantic domain N.
In Chapter 4, the required material for the specification of domains suitable to support the
denotational semantics of cQPL has been presented; it will be put to use in this chapter.
Especially note that the denotations of the composite phrases B0 and B1 are defined only in
terms of the denotations of their sub-phrases as required by the compositionality principle.
To clarify the effect of the denotational definitions, consider how the meaning of the
numeral 101 is denoted; the abstract syntax generates the tree shown in Figure 5.1 as
representation. This leads to the following denotation (observe that the Bs used in the
equations are not identical):
BIN JBK = BIN JB1K = 2 · BIN JBK + 1
= 2 · BIN JB0K + 1 = 2 · 2 · BIN JBK + 1
= 2 · 2 · BIN J1K + 1 = 2 · 2 · 1 + 1 = 5
Since B = 101, the final denotation is given by BIN J101K = 5. This is precisely the
expected result.
40
5.2. Survey of QPL
Figure 5.1: Derivation tree for the binary numeral 101 generated by the abstract
grammar given in Equation 5.1.
5.2
Survey of QPL
The semantics of QPL is based on the idea of annotating a flow graph that represents a
quantum program with density matrices for the quantum mechanical parts and tuples of
probabilities covering the classical components. Additionally, a typing context is used to
keep track of all variables together with their types that are in use at a certain stage of a
program.
Since our work is based on QPL, it seems appropriate to summarise its central concepts.
The original definition of QPL [Sel04b] provides a more detailed description than given
here; an alternative review can be found in [Sch04]. We align our summary on both sources.
Note that it is nevertheless useful to have some familiarity with the paper introducing QPL
because we can obviously not repeat everything here.
5.2.1
Notational conventions
QPL operates on finite-dimensional quantum states represented by vectors over C. The
basis states for qbits are defined as |0i = (1, 0)t and |1i = (0, 1)t . Combination of multiple
qbits are as usual represented by tensor products of these states. Density matrices are used
as basis for any manipulations performed by the language. If a state is defined by some
n
vector u ∈ C2 , the corresponding density matrix is given by ̺ = uu† and may also be
denoted by {u}. Mixed states are represented by linear combinations of pure states, e.g.,
λ1 u1 u†1 + · · · + λn un u†n . Given four matrices A1 , A2 , A3 , A4 of identical dimension, they can
be concatenated horizontally and vertically by
A1 A2
(5.2)
A3 A4
which is used to specify composite density matrices. This notation is used to specify the
semantics of actions possible in QPL which will be introduced in the following sections.
5.2.2
Language elements
QPL programs are given in terms of quantum flow charts 4 where each edge is supplemented
with all the information necessary to unambiguously specify the meaning of a program.
Every edge is augmented with
❏ a typing context, i.e., a mapping from identifiers of variables to the types of these. It
is written as a list of identifiers followed by their type, e.g., a, b, c : bit, d : int. Typing
4 There is also a textual representation for programs, but this is only considered as an aside in the
definition of QPL.
41
Chapter 5. Formal denotational semantics
contexts encapsulating variables which are not related to the present considerations
are denoted by Γ.
❏ an annotation, i.e., a tuple of density matrices which specifies the state of the system.
The annotation of a classical bit is given by (A, B) where A + B = 1 and A represents
the probability that the value of the bit is 0, whereas B is the probability that the value is
1. The annotation for a quantum bit is of the form given by Eqn. 5.2.
All classical operations possible with QPL and their flow graph representations are shown
in Figure 5.2. Figure 5.3 depicts the quantum mechanical parts.
Figure 5.2: Summary of all classical operations of QPL, taken from [Sel04b]. Note
that the symbol “=” is used to separate typing context and the annotation, which can be confusing at times because it is not associated
with Γ alone.
5.2.3
Semantics
The semantics of a QPL program can be directly inferred from the flow graph representation.
The explicit transformation of a density matrix given in the annotation of the edges serves
as a unique representation of the meaning of a program. [Sel04b] proofs that this approach
is indeed well-defined and also works for recursion and loops, which can be included into the
language. Categorical structures that allow to accommodate superoperators and morphisms
42
5.2. Survey of QPL
Figure 5.3: Summary of all quantum mechanical operations of QPL, taken
from [Sel04b].
to manipulate these according to the possibilities of QPL are used as a formal basis for the
definition of the semantics. This categorical superstructure5 is not too interesting for our
purposes. It suffices to know that the valuation functions for the diverse language elements
are defined as shown in Figure 5.4 and that they indeed fulfil everything which is necessary
for a sound and well-defined interpretation. Note that the way how the semantics is specified
in Figures 5.2 and 5.3 is not equivalent to the method of Figure 5.4: While the first one
relies on explicit transformations of density matrices, the second one uses a more abstract
representation in form of superoperators and is almost completely identical to the basis of
our approach (the functions computed by both approaches of Ref. [Sel04b] are nevertheless
identical except for loops and recursion). Especially, the second variant is compositional,
which is a necessary condition to describe multipartite systems in such a way that the
description of one part is independent of other parts.
5.2.4
Limitation: Quantum communication
Before we lay out the denotational semantics of cQPL, it is advisable to sketch in which sense
the different approaches used in QPL do not work for programs dealing with communication.
5 Note that we are only referring to category theory here, not to the compositional semantics presented
by Selinger.
43
Chapter 5. Formal denotational semantics
Figure 5.4: Valuation functions which define the denotational semantics of QPL,
taken from [Sel04b].
First of all, let ̺AB denote the density matrix of the state shared by Alice and Bob.
The information available for each party can be inferred by calculating the partial trace:
̺A = trB (̺AB ) and ̺B = trA (̺AB ). The bipartite density matrix can never be recovered
from these partial density matrices because there are many bipartite density matrices which
give rise to the same partial density matrices.
One of the goals of denotational semantics is to assign sufficient information to every edge
of a quantum flow graph such that the complete semantics of a program can be reconstructed
by combining only the information given by the edges constituting the program. The denotation of a statement composed of several sub-statements must be completely determined
only by a function of the denotations of the sub-statements.
This is impossible in the annotation-based semantics of QPL because transformations
between explicit density matrices are considered. Since a combination of the partial density
matrices ̺A , ̺B which were manipulated by Alice and Bob does not restore the total bipartite state ̺AB , the QPL annotation would obviously not comply with the physical state
afterwards.
A possible solution is the annotation of the complete flow graph, i.e., of both paths
representing the control flow for Alice and Bob. In this case, the operations performed
by Alice and Bob would be written as tensor products of the type  ⊗ 1B and 1A ⊗ B̂
which act on the complete density matrix ̺AB . This way, we could assign semantics to the
program as a whole, but would loose the ability to construct the denotation of a phrase from
the denotations of its subphrases. This means that the semantics of the complete program
could not be constructed from the denotation of Alice’s and Bob’s program alone which is
in contrast to the key idea of denotational semantics.
Therefore, we need to seek a solution that does not characterise quantum operations by
showing transformations of explicit density matrices, but uses something that captures the
44
5.3. Denotational semantics of cQPL
notion of a transformation in a more abstract sense. Completely positive maps represented
by a set of Kraus operators fulfil this need as we will show in Section 5.3.1; this is basically
the same approach as used for the compositional semantics of QPL. Nevertheless, QPL does
not provide any means of parallel composition, communication and other details which are
necessary to describe quantum communication as we will do in the remaining parts of this
thesis.
5.3
Denotational semantics of cQPL
cQPL is an extended variant of QPL with the ability to express and formalise quantum communication, i.e., the ability to describe multiple parallel flow graphs that exchange quantum
and classical information at well-defined points, but do otherwise know nothing about each
other. To achieve this, we have to base the semantic description on three components in
contrast to the two components (typing context and tuples of density matrices) of QPL:
❏ A Kraus aggregation K as defined in Section 4.5.4 which is used to keep track of the
quantum operations performed on the qbits of the system.
❏ A typing context T used to specify which quantum and classical variables are allocated
at a given moment and which data type they have. This is also important to describe
communication because it allows to uniquely determine to which party a variable
belongs at a given stage of a program.
❏ A probabilistic environment mapping identifiers to values. Since the interaction between quantum and classical parts of the system introduces probability, the values of
classical variables are subject to such a distribution. In general, only the range of
possible values together with the fact that it is governed by a probability distribution
is known in the semantical description.
We refer to these three elements as the three-tuple (K, T, E).
A Kraus aggregation K specifies a quantum mechanical operation which has the same
effect for all density matrices ̺ (in the sense that the application of a Hadamard gate will
yield different effects according to the state it was applied to. Nevertheless, it is still a
Hadamard gate in every case, and this is the really important thing). Thus, the operation
is completely characterised without the need to specify any density matrix at all. This is
exactly what is required when spatially separated operations performed by several parties
on multipartite states are to be described, as we will see later in greater detail.
To realise the benefits of this approach, consider how the generation of a new qbit in
state |0i subsequently followed by the application of a Hadamard gate is described in QPL
(we do not show the complete flow graph, but only the relevant parts of the annotation):
create new qbit q
Γ = A −−−−−−−−−−−→ q : qbit, Γ =
A 0
0 0
45
apply H on q
−−−−−−−−→ q : qbit, Γ =
1
2
A A
A A
Summary
We have presented a quick summary of QPL and the associated denotational
semantics which is based on partial orders of density operators. Additionally, we
have shown why this approach is not suitable to describe quantum communication
respectively the interaction of spatially separated systems where the combined
density matrix is not available as whole in the framework of the annotation-based
semantics.
Chapter 5. Formal denotational semantics
Although only the newly created qbit is concerned, the state of the remaining system is
still implicitly present in A. This is more than needed: It suffices to consider the application
of two operations given by the following Kraus sets:
{Ci }#q ; {Hi }#q
(5.3)
where {Ci }#q stands for “create a new qbit with label q” and {Hi }#q for “apply a Hadamard
gate on q”. With these, we can describe the same operation without resorting to a density
matrix or any other part of the system unconcerned by the operation at all.
The typing context T is basically adopted from QPL. An extension to the framework
used by QPL is the probabilistic environment. For every allocated classical variable in the
current frame, it is used to specify a probability distribution that maps the variable name to
the range of possible values. This distribution is parametrised by density operators because
it depends on the initial conditions of the program fragment and on the path taken in the
flow graph (an example explaining this will follow in the next section). The probabilistic
environment could in principle be replaced by the tuples for classical states as used in QPL,
but this works only well for data types with a very low number of bits. Because of this
reason, QPL tries to hide these tuples most of the time, so we eliminate them completely
and replace them by the probabilistic environment.
The probabilistic environment also deals with quantum variables: For every such variable,
the position of the allocated qbits in the global quantum heap is given by the probabilistic
environment. This is necessary because quantum variables cannot be characterised by a
value as it is possible for classical variables because they do not have a state as such. The
state is replaced by the series of operations which have been performed on the variables;
since these operations need some location to act on, every quantum variable needs to have
a unique position on the quantum heap, i.e., where the qbits are stored.
Note that this approach is somewhat contrary to the spirit of functional programming
because it introduces stateful global variables, but a closer examination reveals that QPL
implicitly uses the same model and that compile-time checking (and thus the protection
against runtime errors) is not affected by this.
In addition to allocated variables, branches in programs are also present in the probabilistic environment. They are identified by a unique ID which is assigned to every branching
node.6 This is necessary because the branching conditions – being based on comparisons of
probability distributed quantities – are in general not represented by some fixed values, but
represented by a probability distribution as well.
5.3.1
Formal definitions
In this section, we will present some methods to characterise and describe the semantic
components of cQPL. Note that in the following, we use Dn to denote the set of all density
operators of dimension n, dropping the subscript if the exact dimension is not important or
can be deduced from the context.
5.3.1.1
Typing context
Let σ be a list of numbers ni ∈ N+ , i = 1, . . . , k as given by σ = nτ1 , nτ2 , . . . , nτk . σ is also
called the signature of a data type. An associated Hilbert space Hσ is given by
6 To
Hσ ≡ H1 ⊗ · · · ⊗ Hk
(5.4)
be precise: Which is assigned whenever the branching node is transversed because we need to account,
e.g., for branches in loops where the same branch might be traversed multiple times.
46
5.3. Denotational semantics of cQPL
where Hi is either B(H) for quantum or C(X) for classical data (cf. Section 4.3.3) where
both are distinguished by the index τ : τ = q for quantum variables and τ = c for classical
q
variables. The dimension of the i-th space is given by 2ni for quantum mechanical and nci
for classical variables. Since we restrict ourselfs to finite-dimensional Hilbert spaces, this
means that we can use Hi = Cni for quantum mechanical and X = [0, 1, . . . , ni − 1] for
classical data. To distinguish between both cases, we define the function q : nτi → [0, 1]
given by
(
1 if τ = q
τ
q(ni ) =
(5.5)
0 otherwise
Note that although our formalism allows to define data types which consist of both quantum
mechanical and classical components, we do not exploit this possibility because we could
not find any reasonable application for this in our work. To keep the formalism as general
as possible, we will nevertheless still retain the possibility as long as no noteworthy effort is
necessary to do so.
We can define a function tq : σ → N to compute the total number of qbits necessary for
a given signature (card(σ) denotes the cardinality of σ):
card(σ)
tq (σ) =
X
i=1
q(nki ) · nki
(5.6)
The analogous function tc for the classical components is given by
card(σ)
tc (σ) =
X
i=1
(1 − q(nki )) · nki
(5.7)
Finally, two functions q(σ) and c(σ) to check if a given data type is purely classical or purely
quantum are necessary:7
q(σ) =
(
1 tc (σ) = 0 ∧ tq (σ) > 0
0 otherwise
(5.8)
c(σ) =
(
1 tq (σ) = 0 ∧ tc (σ) ≥ 0
0 otherwise
(5.9)
For simplicity, we label data types required for practical use with special mnemonics; some
examples can be found in Table 5.1. Note that the type void can, for example, be used to
formally describe statements which return no value and thus have no type.
Since two finite-dimensional Hilbert spaces H1 and H2 are isomorphic if the sum over
the dimensions of their subsystems is equal (this is, e.g., proved in [Wei00, Theorem 2.62]),
i.e.,
O
O
X
X
H1k = H1 ∼
H2l ⇐⇒
dim(H1k ) =
dim(H2l ),
(5.10)
= H2 =
k
l
k
l
the description of types is not unique. For example, the types given by (2q , 2q , 2q , 2q ) and
qshort = 8q are identical and provide only different aspects of the same thing. This
7 Note
that we define a data type consisting of 0 quantum and 0 classical bits, i.e., void, as classical.
47
Chapter 5. Formal denotational semantics
Mnemonic
bit
qbit
short
qshort
int
qint
void
Signature
2c
2q
8c
8q
16c
16q
0(c)
Table 5.1: Signatures and the mnemonics commonly used in programming languages for data types supported by cQPL.
equivalence can also be extended to mixed data types:
X
X
nτ1 , . . . , nτK ∼
q(nτk ) · nτk =
q(mτl ) · mτl
= mτ1 , . . . , mτL ⇐⇒
k
∧
X
k
l
(1 −
q(nτk ))
·
nτk
=
X
l
(1 − q(mτl )) · mτl
(5.11)
This creates an equivalence class for data types which will be useful in Section 5.3.8. The
class of all data types equivalent to σ is given by
tc (σ)
q (σ)
tM
M
inϕ(i+tq ) (Ω#i)c
(5.12)
inϕ(i) (Ξ#i)q ⊕
T (σ) ≡
i=1
i=1
such that Ξ ∈ S(tq (σ)), Ω ∈ S(tc (σ)), ϕ ∈ Sym([1, . . . , tq (σ) + tc (σ)])
where M #i denotes the ith element of the ordered set M and S(k) is the decomposition of
the scalar value k into all possible sums given by
n
S(k) ≡ Z is a set with members ∈ N
card(Z)
X
i=1
o
Z#i = k .
(5.13)
If σ2 ∈ T (σ1 ), we write σ1 ∼
= σ2 .
To illustrate the effect of Eqn. 5.12, consider a data type which consists of 3 quantum
and 3 classical bits. Structurally, it does not make any difference how these components
are ordered, e.g., (1q , 1c , 2q , 2c ) is identical with (3q , 3c ) in this sense.8 Eqn. 5.12 is a
generalisation of this idea: The scalar 3 can be decomposed as 1 + 1 + 1, 2 + 1 and 3
as given by S(3), so there is no difference between any of these groupings. Additionally,
it is not interesting how the components are ordered, e.g., (2 + 1) is equivalent to (1 + 2).
Finally, the quantum and classical components can be arbitrarily interchanged, so we have
to consider this as well. The effect of Eqn. 5.12 is to construct all equivalent representations
of a data types following these considerations.
Note that QPL uses a Cartesian product of complex vector spaces given by Cn1 ×n1 ×
· · · × Cnk ×nk for both classical and quantum mechanical signatures (the set of complex d × d
8 Note that there is a difference between these orderings from the compiler’s point of view because the
different components are located at different locations in memory if different orderings are used. The
semantics is nevertheless unconcerned by this.
48
5.3. Denotational semantics of cQPL
matrices is used to represent the complex Hilbert space of dimension d). This does not
reflect the relationship between corresponding quantum and classical objects directly. For
example, the data type for bits is given by bit = (1, 1), whereas for qbits, the definition
is qbit = 2). This leads to appropriate spaces for these objects, but does not present the
relation between them directly.
We thus used a different approach that makes correspondences more clear which is important when, for example, quantum variables are measured and the result is stored in
a classical variable. Besides, it fits better into the more abstract description of quantum
mechanics as introduced in Section 4.3.
Let Σ be the set of finite strings over the alphabet α. With this and the notion of types,
we can define the typing context used in the semantic description of cQPL.
Definition 5.3.1 (Typing context). A typing context τ is a three-tuple τ = (ι, θ, χ) where
ι is a set of identifiers in Σ, θ is a set of types and χ : ι → θ is a surjective mapping which
assigns a type to every identifier. Identifiers starting with # must not be used by programs.
Because cQPL is strongly and statically typed (i.e., the type of an expression is completely determined by the types of its components and the type of an elementary component
cannot be changed after it has been declared, cf. Appendix B), information contained in
the typing context cannot be modified any more once it has been introduced. Note that this
does not hinder the possibility of overshading entries. This happens when, e.g., a variable
declared in an inner block has the same name as a variable declared in an outer block.
Although both have identical names, their types do not need to match because they are
otherwise completely unconnected.
Typing contexts are modified when new variables are declared (and thus added to the
context) or when variables are removed from the scope (and thus have to be removed from
the typing context). Since it is obvious how this influences a given context τ , we only note
that it is easy to define appropriate morphisms τ → τ ′ which perform the desired job.
Formally, we use the notation
τ → τ ′ = τ ⊕ (ξ → qbit)
(5.14)
to introduce some new identifier ξ with type qbit into the context τ . Equivalently, the
notation τ ⊖ ξ is used to remove ξ which is needed to describe sending quantum variables.
5.3.1.2
Probabilistic environment
The probabilistic environment can be defined formally as follows:
Definition 5.3.2 (Probabilistic environment). Let π be a probability
distribution on a
P
finite set X with probabilities pi for every element of X such that i pi = 1. Let ι be a set of
identifiers, P be a set of probability distributions and M be a surjective map M : ι → P ∪ ⊥.
Then E = (ι, P, M ) is a probabilistic environment.
As usual in denotational semantics, the symbol ⊥ is used to denote an undefined identifier, i.e., a variable which is not present in the environment. To specify components of a
probabilistic environment, we use the notation
π
x
range(x)
x −→
(5.15)
to denote an element of the probabilistic environment where x is the identifier, πx the
associated probability distribution and range(x) the set of possible values which obviously
49
Chapter 5. Formal denotational semantics
depends on the data type of x. Adding a new binding to a given environment E is once more
done with the operator ⊕ which is formally a morphism E = (ι, P, M ) → (ι′ , P ′ , M ′ ) = E ′ :
π
x
range(x)
E ′ = E ⊕ x −→
(5.16)
Note that multiple inclusion of variables overrides the previous definition. Thus, the
meaning of
π′
π
x
x
range(x)) ⊕ x −→
range(x)
E ′ = (E ⊕ x −→
(5.17)
is to create a probabilistic environment which contains πx′ as probability distribution for
the variable x. The previous distribution πx can then not be recovered any more in E ′ .
In direct analogy to Kraus aggregations, probabilistic environments can be combined
with +; the summands are prefixed by some constraint that states which one has to be
chosen with which probability:
p1 ({c1 }) · E1 + p2 ({c2 }) · E2 + · · ·
(5.18)
P
where {ci } are the conditions which determine the values of p and i pi = 1. This construction is necessary for the description of, e.g., if-conditions when it is not a priori determined
which path will be selected. If both paths of an if-condition perform a modification on the
same variable that already existed before the branch, then the variable will have different
values after the merge point. The entries of the probabilistic environment which record this
assignment are then prefixed by the branching probability. This is also one of the reasons
why the branching probability needs to be kept even after the branched paths are merged
again.
Definition 5.3.3 (Distributivity of ⊕ over +). The operation ⊕ is defined to be distribuπx
πx
πx
range(x).
range(x) + E2 ⊕ x −→
range(x) = E1 ⊕ x −→
tive over +, i.e., (E1 + E2 ) ⊕ x −→
This ensures that adding a new binding to a sum of environments results in adding the
binding to all contributing environments automatically.
Remark 5.3.1. This formalism is not equivalent with the functionality introduced by stores
(cf. e.g., [Mos90, Rey98]). It does still not make use of stateful variables per se, but rather
updates the binding of a variable, i.e., the value it is associated with.
Observe that the probabilistic environment is only necessary for classical, but not for
quantum variables: The state (or, rather: the history of all operations performed until the
present moment) of the quantum mechanical constituents of the computation can be reconstructed with the aid of the Kraus aggregation. Nevertheless, the probabilistic environment
is necessary to keep track of quantum variables in a different way which will be introduced
in a moment.
Note that the view on quantum variables differs slightly from that on classical ones: It
is not only necessary to keep track of the structure of a variable (as is done by the typing
context), but also of the position within the quantum heap – this is necessitated by the
underlying model of computation as introduced in Section 2.2.3.9 The environment can be
used to provide this kind of information by supplying a map
9 The
# : ι ∋ v → (i1 , i2 )
(5.19)
value of a quantum variable can obviously not be directly stored in an environment because the
state might be in a superposition. The operations performed on the quantum bit are recorded in the Kraus
aggregation and unambiguously specify the state.
50
5.3. Denotational semantics of cQPL
where ik are integer numbers with 0 ≤ ik < Q and Q is the size of the quantum heap. The
tuple i1 , i2 denotes the interval [i1 , i2 ] which contains i2 − i1 + 1 quantum bits. Obviously,
the number of qbits allocated in the quantum heap must agree with the number of qbits
necessary for the type of the variable as given by the typing context.
In theory, it is possible to assume that the quantum heap can always be partitioned into
consecutive intervals; we do not need to take care of issues like fragmentation which does
obviously appear in implementations and simulations. We assume that the hardware of the
quantum memory take care of this issue by acting like a memory management unit.10 Note
that if the user is allowed to directly address the components of the quantum heap, it is
possible to cause run-time errors as in QCL. Therefore, we do not allow this.
Consider a subset M = [0, n] of N. The set of all interval partitions is given by11
I(M ) ≡ {m ⊆ P(M )|∀n ∈ [1, . . . , card(m) − 1] :
((m#n)#0 − (m#(n − 1))#(card(m#(n − 1)) − 1)) = 1}
(5.20)
where we suppose that the contents of all sets m#i is sorted in ascending order. This
can be used to formally define how the probabilistic environment can be adapted to the
requirements for quantum variables:
Definition 5.3.4 (Quantum part of the probabilistic environment). Let (ι, P, M ) be
a probabilistic environment. It can be extended to fulfil the requirements for the description
of quantum variables by the following construction:
❏ P is extended to P ⊕ (I([0, Q − 1]) ∪ Σ∗ ) where Q is the total number of quantum
bits present in a system. The set of intervals is used to represent quantum variables
which reside on the local quantum heap, i.e., which were allocated in the module the
probabilistic environment belongs to. Σ∗ is used to denote the originating module for
variables which were received from some other party.12
❏ Let Q = { v ∈ ι q(χ(v)) = 1 } be the set of all identifiers for variables
T with quantum
data type. Then, M ′ is an injective morphism Q → I for which range(M ′ ) = ∅
(this ensures that quantum variables do not overlap on the quantum heap) must hold.
Then M is replaced by M ⊕ M ′ in the previous definition.
Note that this definition reflects a fundamental difference between classical and quantum
variables: While a classical variable is nothing else than a mapping between an identifier
and a value that can be governed by a probability distribution, such a mapping is in general impossible for quantum variables because they do not have a value per se, but only a
certain quantum state. To describe this quantum state precisely (disregarding the principal
impossibility of implementing a measurement that delivers this information by inspecting
10 This component of a processor creates a view of the available memory such that every application –
roughly – thinks that it would have an own linear address space which is as big as the the one available for
the whole system.
11 An example might illustrate this definition: Consider the set {[0, 1], [2], [3, 4]}. This is a proper partition
since no elements overlap and the boundaries are adjacent. These conditions can be ensured by considering
the last element of the ith set given by (m#i)#(card(m#n) − 1) and the first element of the (i + 1)th set
given by (m#(i + 1))#0. If the difference between these is +1, then both the adjacency and no overlap
conditions are fulfilled. If this holds for all subsets, we have a proper partition.
12 We have to make sure that every quantum bit in the system belongs to exactly one place in a quantum
heap. This is simple for single-party programs, but gets more complicated when communicating programs
are considered because the case of sending the same quantum bit back and forth between participants must
be taken into account.
51
Chapter 5. Formal denotational semantics
a single copy of a quantum system), one would need an infinite amount of classical information because even a simple system such as a qbit takes values in a continuous space as
a consequence of quantum superpositions. This makes the classical approach of mapping
the identifier to a probability distribution of values impossible. Nevertheless, the quantum
variable is completely characterised if its location on the quantum heap together with the
operations performed on its initial state are known.
Remark 5.3.2. Although we retain the name probabilistic environment also for the version
of the environment extended to quantum variables, there are no probabilities involved in the
connection between variable names and the allocated positions on the quantum heap. The
convention just simplifies the notation.
Remark 5.3.3. Support for mixed quantum/classical types would require a little more effort
compared to the case of full separation because with the introduction of such types, the direct
decomposability of the probabilistic environment would not be feasible any more. Nevertheless, no fundamental difficulties would be associated with this.
Adding a new quantum variable with name ν which occupies the quantum heap positions
given by (q1 , . . . , qn ) is written as
E ⊕q ν : (q1 , . . . , qn )
(5.21)
Removing a quantum variable is denoted by ⊖q ; this is required when qbits are transmitted and thus cannot be accessed any more (we drop the index if there is no danger of
confusion). Note that there is no need for a corresponding operation for classical variables
because overlays provide the required functionality, as will be shown later.
5.3.1.3
Kraus aggregations
Summary
We can directly adopt the definition of Kraus operators as given in Section 4.5.4. Nothing
needs to be modified for our purposes (note that the composition of two Kraus aggregations
was denoted by ◦ instead of ⊕ as used for the other elements of the three-tuple (K, T, E) to
avoid confusion with the symbol + used to combine sub-aggregations).
It is possible to show that this semantic framework can be used to formalise standard
QPL, but we omit the details here.
We have introduced the structures which are necessary to specify the denotational semantics of cQPL; they fulfil the required properties as described in Chapter 4. The semantic framework consists of three components: A Kraus aggregation
which is used to store all quantum mechanical operations performed by commands
of cQPL, a probabilistic environment which maps identifiers to values (possibly
governed by a probability distribution) and provides mechanisms to ensure that
quantum variables do not interfere with each other, and a typing context whose
information is the basis for compile-time correctness checks of programs. These
are grouped in a (K, T, E) tuple which will be omnipresent in the following. Additionally, we have derived some criteria for the identity of data types.
5.3.2
Some examples
Before we commence to extend the formal definitions for multipartite systems, we want to
demonstrate the introduced concepts with some examples which should aid the reader to
see the rationale behind their definition.
52
5.3. Denotational semantics of cQPL
5.3.2.1
Semantics of sequential programs
Consider the following program fragment which applies a Hadamard matrix to the quantum
variable p or q depending on the result of a branch based on a comparison of two classical
variables x and y:
if (x > y) {
q *= H;
}
else {
p *= H;
};
The flow graph representation of the fragment is given in Figure 5.5. To shorten the annotation of the edges and to avoid repeating the same information over and over, we introduce
the injection functions ini . ini are 0-based injections into the ith element of an n-tuple. If we
want to add the contribution {U } to the element K of the tuple (K, T, E) = ξ, we can write
this as ξ ⊕ in0 ({U }). The initial (K, T, E) tuple of the example flow graphs is abbreviated
by ξ; modifications derived from this are denoted by ξ ′ , ξ ′′ , . . ..
{{Γ}} ; {τ ; x, y : int; p, q : qbit} ;
o
n
πy
πx
range(x); y −−→ range(y) ≡ ξ
E ⊕ x −−→
x>y
Unique node ID #42
ξ ′′ ≡ ξ ⊕ in2 (#42 → false)
⊕ in1 (#42 : bit)
ξ ′ ≡ ξ ⊕ in2 (#42 → true)
⊕ in1 (#42 : bit)
p∗ =H
q∗ =H
´
`
ξ ′ ⊕ in0 {H}#p
´
`
ξ ′′ ⊕ in0 {H}#q
Implicit merge for branch #42
n
“
”
“
”o
p (val(#42) = true) · {Γ} ; {H}#p + p (val(#42) = false) · {Γ} ; {H}#q
;
{τ
o
n ; x, y : πint; #42 : bit; p, qπy: qbit} ;
π#42
x
range(x); y −−→ range(y); #42 −
−−−→ range(#42)
E ⊕ x −−→
Figure 5.5: Flow graph of a simple branching operation to demonstrate the elements of the semantic framework: A 3-tuple (K, T, E) is used to annotate every edge of the graph; K is a list (or aggregation) of Kraus
operators, T is the typing context and E the probabilistic environment.
Note that the annotation of the graph uses several abbreviations as
defined in the text.
The initial configuration of the 3-tuple (K, T, E) is given by Γ (initial list of Kraus
operators), E (probabilistic environment) and τ (typing context). We don’t care about
their contents in detail, they can represent the semantics of any valid program fragment that
might be placed before our example. Things which are of interest for our code fragment are:
53
Chapter 5. Formal denotational semantics
❏ The two classical variables x and y, both of type int.
❏ The probability distributions πx , πy which map the variables x and y to a value contained in [0, 2bits per int − 1].
The probability distribution for classical variables arises because we work with Kraus operators describing quantum operations instead of density matrix transformations. Consider
the measurement of a single qbit whose result is stored in a classical bit variable: We know
that the range of the measurement outcome is {0, 1}, but we don’t know with which probability the “0” and the “1” will appear because we do not have an explicit density matrix to
describe the qbit. This piece of information can only be gained when the final semantics of
the program (in form of a total set of Kraus operators) is “applied” to a well-defined initial
configuration; only then quantitative statements about the distribution are feasible. All we
know is that the measurement result x will be governed by a probability distribution πx
with a certain well-defined range, so we preserve that information.
Since the values of x and y are given by a probability distribution, the result of the
comparison x > y (with outcome range {true, false}) can only be specified by another
probability distribution which can be deduced from πx and πy . Since we need to reference
the outcome of the comparison at a later point in the flow graph (when the two edges of the
branch are merged), a unique identifier for the node is created (#42 in this case) and the
probabilistic environment is extended accordingly.
Depending on the outcome of the comparison, a Hadamard gate is applied on either p or
q. This does not change the probabilistic environment or the typing context, but is recorded
by placing an appropriate Kraus operator in the Kraus aggregation (#p and #q denote the
position of the quantum bits in the quantum heap).
After every if-then-else construction, an implicit merge operation which unites the two
branches takes place. The probability distribution of the branching condition is preserved
in the probabilistic environment under the label assigned to the branch statement; the
Kraus aggregation is transformed into a sum that formally resembles a mixed state: With
probability val(#42) == true (which is the probability that x > y evaluated to true),
the operation {Γ̂}; {H}#p was performed, while with probability val(#42) == false, the
operation was {Γ̂}; {H}#q .
5.3.2.2
Communication with EPR pairs
Consider the following (pseudo-)code which describes how Alice creates an EPR pair and
transmits half of it to Bob:
module Alice {
new qbit p := 0;
new qbit q := 0;
createEPR(p,q);
send q to Bob;
new bit b := measure(p);
if (b) { ... } else { ... };
};
module Bob {
receive m from Alice;
new bit b := measure(m);
if (b) { ... } else { ... };
};
54
5.3. Denotational semantics of cQPL
The corresponding flow diagram is given in Figure 5.6.
new qbits p, q := 0
createEPR(p, q)
n
{Γ} ; {C}p ; {C}q {EP R}#p,#q
{τ ; p, q : qbit} ; {E}
send q
nn
Γ′
o
;
receive m
oo
nn
;
{τ ; p : qbit} ; {E}
measure p
ID: #23
o
; {R}#m ;
n
o n
o
′
τ ; m : qbit ; E ′
Γ′′
o
ID: #42
measure m
Figure 5.6: Flow diagram which describes the creation of an EPR pair by Alice;
she keeps the first half, while the second half is sent to Bob. Afterwards, both of them measure their qbit. {Γ′ } is a shorthand for
{Γ}; {C}p ; {C}q ; {EPR}#p,#q , {Γ′′ } is the initial Kraus aggregation
of Bob, τ and E respectively τ ′ and E ′ are the initial typing contexts
and probabilistic environments of Alice and Bob.
Note that the labelling formalism is reduced to the basic necessities in this example
in order to emphasise the central elements. Because Kraus operators are used to describe
the quantum mechanical operations, it is possible to perform spatially disjoint actions by
parties who do not know the total state. When the edges are merged together, the operations
performed by Alice and Bob can (as was mentioned before) be factorised as  ⊗ 1B , 1A ⊗ B̂
for the combined semantics of both branches; the total density matrix is not involved in this,
contrary to annotation-based QPL. The framework to generate the semantics of the total
system from the semantics of the components will be introduced in the following section.
5.3.3
Extension to multipartite systems
Since we want to consider the formal semantics of programs which deal with communication,
we need to extend the previous definitions to the multi-party case. For this, observe that
the number of participants can naturally be assumed to be finite which makes it possible to
label each party with a unique identifier of finite length. For formal simplicity, we assume
that the set of labellings for participants Lc is disjoint with the standard labels used for
variables, i.e., Lc ∩ Σ∗ = ∅. Let Lc (i) denote the unique label given by the ith entry of Lc .
Assume that we have n communicating parties which are labelled with l1 , . . . , ln ∈ Lc .
According to the principle of compositionality (we will explain this for the context of communication in more detail in Section 5.3.6.4 on Page 70), there is a three-tuple (K, T, E)
for every participant, i.e., (K1 , T1 , E1 ), . . . , (Kn , Tn , En ). The combined three-tuple for the
complete system is then given by
(⊗ni=1 Ki , ⊗ni=1 Ti , ⊗ni=1 Ei )
55
(5.22)
Chapter 5. Formal denotational semantics
We will consider how the tensor product needs to be defined for each component to
provide a sound basis for our further needs.13
5.3.3.1
Kraus Aggregation
Consider, for simplicity, two Kraus aggregations
Λ1 = {A1k }, {A2k }, . . . , {Ank }
Λ2 =
{Bk1 }, {Bk2 }, . . . , {Bkm }
(5.23)
(5.24)
(in case of n 6= m, the shorter list can be padded with zero elements so that m = n can
be assumed, but note that it is only for notational convenience). If both lists operate on a
disjoint set of qbits, i.e., no send/receive (pseudo-)operations are contained in the lists, then
∀i, j : [{Aik } ⊗ 1B , 1A ⊗ {Bkj }] = 0 holds. The Kraus aggregation for the composite system
can be written as:
Λ1 ⊗ Λ2 = {A1k } ⊗ {Bk1 }, . . . , {Ank } ⊗ {Bkn }
(5.25)
Note that members of type {Ak } ⊗ {Bk } can be written as {A} ⊗ 1B + 1A ⊗ {B}. If this
is done for all list elements, we see that all combinations of A ⊗ 1 with 1 ⊗ B commute;
this induces many equivalent orderings of the lists – in addition to the normal commutative
equivalences as given by Eqn. 4.57 for the sublists – which needs to be considered when
multi-party aggregations are tested for semantic equality in Section 5.3.5.
Formally, the equivalence class of compatible Kraus aggregations for two independent
parallel systems is given by
)
( n
n
M
M
i
i
inϕ(n+i) 1A ⊗ {Bk }
inϕ(i) {Ak } ⊗ 1B ⊕
(5.26)
i=1
i=1
with ϕ ∈ Sym(2n) : ϕ(i) < ϕ(i + 1)∀i ∈ [1, . . . , n] ∧ ∀i ∈ [n + 1, . . . , 2n]
where the additional constraints on the permutation make sure that the order of {Aik }
and {Bki } is preserved. We can generalise this approach to n Kraus aggregations given
in the contracted normal form (padding is applied as usual to compensate for different
cardinalities):
PN
Definition 5.3.5 (Tensor product for Kraus aggregations). Let Λ1 = k1 =1 p1k1 Λ1k1 ,
PN
. . . , Λn = kn =1 pnkn Λnkn be Kraus aggregations in the contracted normal form. The tensor
product of these is given by
n
O
i=1
Λi ≡
N
X
k1 =1
···
N
X
kn =1
p1k1 · · · pnkn · Λ1k1 ⊗ · · · ⊗ Λnkn
(5.27)
Note that the naming of qbits changes when multipartite systems are merged. For Kraus
sets which symbolically refer to qbits, the labels must be updated accordingly (the exact renaming scheme is given in Definition 5.3.6). Two aggregations are equivalent if they are
member of the same equivalence class as given by Eqn. 5.26 or member of the equivalence
class given by the same formula, but induced by a compatible ordering of one or more of Λi
as defined by Eqn. 5.25.
Note that we will show in Section 5.3.6.4 how send and receive operations can be integrated into this formalism.
13 Note that although formally, different tensor products are used for each element of the (K, T, E) tuple,
we use the same symbol for all of them to simplify notation.
56
5.3. Denotational semantics of cQPL
5.3.3.2
Typing Context
Definition 5.3.6 (Tensor product for typing contexts). The tensor product of n typing
contexts (ι1 , θ1 , χ1 ), . . . , (ιn , θn , χn ) is given by
!
n
n
n
n
[
[
[
O
(5.28)
Lc (i)χi
θi ,
Lc (i)ιi ,
(ιi , θi , χi ) =
i=1
i=1
i=1
i=1
where Lc (i)M denotes a set which contains all elements of M prefixed by the unique
identifier Lc (i); Lc (i)χ denotes the morphism χ adapted to the new naming scheme.
5.3.3.3
Probabilistic Environment
The probabilistic environment can be adapted to multipartite systems analogous to the
typing context, i.e., by prefixing the variable names with appropriate labels and adapting
the morphism used to connect the set of identifiers with the set of data types. Obviously,
the prefix for each subsystem must be the same as used for the typing context.
The quantum part needs to be modified as well: The local quantum heaps must be united
to a single one; necessarily, the positions of all variables on the new heap still need to be
disjoint. This is simple to achieve: If E0 uses the range [0, n1 ] and E1 the range [0, n2 ], the
new range is given by [0, n1 + n2 + 1] and the morphism M ′ needs to be adapted such that
the mapping remains unchanged for variables originating from E0 and the constant offset
n1 + 1 is added to its codomain for variables originating from E1 .14 Likewise, the function
# which associates variable names with quantum heap positions needs to be updated such
that the modified variables names are mapped to the modified positions.
We will not consider the renamings any more in the following parts, but just take them
as given; this simplifies the notation considerably.
5.3.3.4
Quantum channels
Quantum channels are used to exchange information between processes.15 Although not
only quantum data, but also classical variables can be sent, we restrict our considerations to
the first case because the second one is not too interesting from a physical point of view and
would only obstruct the view on the central elements. In particular, classical communication
can be seen as a special case of quantum communication (cf., e.g., [Key02, Section 6.2.2]),
hence the generality of the approach does not suffer from this restriction). Besides, the topic
of classical communication has been investigated in classical programming language theory
for a long time, so we can refer the reader to the wealth of existing literature about this
topic, e.g., Ref. [Rey98].
We have already introduced quantum channels informally in Section 2.2.3.1; here, we
consider the concept formally.
Definition 5.3.7 (Quantum channel). A quantum channel is a five-tuple (O, D, S, R, F )
where O is the origin and D the destination for a quantum variable (these can, e.g., be
14 Note that both local quantum heaps could have already used the full number of available quantum
bits (Q); we do not consider this problem any further because it is alway possible to limit the number of
quantum bits for n communicating systems to n·Q because both n and Q are finite. We are not too concerned
about this problem because we always assume that there are enough qbits available. The reason behind
the restriction to a finite, but fixed number of qbits is to ensure the boundedness of all Kraus operation as
explained in Chapter 4.
15 A user in a real-world implementation is nothing else than a process in the simulation. We thus use
both terms interchangeably.
57
Chapter 5. Formal denotational semantics
represented by processes), F is a FIFO16 containing objects of type (̺, σ), S is a morphism
to place two-tuples (̺, σ) in F and R is a morphism to retrieve two-tuples (̺, σ) from F .
As usual, ̺ represents the density matrix of a quantum variable and σ the associated type.
Thus, the quantum channel can be used to make sure that not only typing is guaranteed
to be preserved along communication (for this, confer further Section 5.4), but also that
quantum data does not appear multiple times in a composite system at the same time
which is necessary to avoid unphysical situations in the simulation.
Remark 5.3.4. Note that quantum channels are only necessary when the parallel composition of two or more processes is considered. For single modules, the functions used to deposit
respectively request information from the channel together with an abstract representation of
the channel (e.g., an identifier) are sufficient.
Remark 5.3.5. Also note that most descriptions of quantum communication protocols do
not consider typing of the exchanged data explicitely, it is nevertheless implicitly implied by
the physical realisation of the protocol, e.g., in the measurement process, by the hardware
used to realise the communication channel or in the way the quantum part is implemented
in general.
5.3.4
Existence of fixed points
Because fixed points are important for the denotational description, we need to prove the
following theorem which states a condition for the existence of such. In the following, the
condition can shown to be fulfilled for every element of the semantics.
Theorem 5.3.1. Let T be a linear operator D → D acting on a complete partial order D
with bottom ⊥D . If T is bounded, then a fixed point of T exists.
Proof. Since T is bounded, we can see from Theorem 4.2.1 that it is continuous as well.
Topological continuity implies Scott continuity as was shown in Theorem 4.4.2. The existence of a fixed point is now given by Theorem 4.4.1, as required.
Remark 5.3.6. Note that the same proof could have been deduced at a slightly more abstract
level using the notion of a pointed dcpo, i.e., a dcpo with a least element. For structures
fulfilling this condition, Theorem 2.1.19 in [AJ94] ensures that the desired least fixed points
exist. Ref. [Sel04b] uses a somewhat similar reasoning in the description of recursive procedures where the existence of least fixed point solutions for these is explained by the fact that
for Scott-continuous endofunctions on pointed complete partial orders, these always exist.
Remark 5.3.7. For those who want to be extra sure, the classical fixed point theorem by
Schauder which states that any continuous map with a countably compact image on a compact, convex subset of a Banach space has a fixed point could also be used to derive the
required property of cp-maps.
We will need fixed points to solve recursive equations which occur in denotations specified by recursive equations. These are required for loops and the combined semantics of
communicating systems.
16 First in, first out queue. Informally, this is a queue where objects can be put in on one side and taken
out on the other side. The object which is put in first comes out first, the second one comes out second etc.
58
5.3. Denotational semantics of cQPL
5.3.5
Types of interpretational equivalence
The term “equivalence” is not unambiguous without further specification. Under which
conditions can two programs or, respectively, the denotations of two programs be considered
as equivalent? QPL has to distinguish between two different types of equivalences as noted
in [Sel04b, Section 6.6]; likewise, we can define several types of equivalence:17
❏ Two programs are textually equivalent if there is a bijective mapping between the set
of all variables the programs use and for every constituent of program A, there is a
constituent of program B such that Ai = Bi , i.e., the programs are identical already
at the level of the syntax. The ordering of these constituents must be identical.
For example, the two program fragments new int a; a:=1; and new int b; b :=
1; are equivalent because the sequence of commands is identical if the replacement
a ↔ b is applied to the variables.
❏ Two programs are denotationally equivalent if their denotations are identical.
The second definition only shifts the problem because it leaves the question of how
to identify equivalent denotations. This is problematic for cQPL because we do consider
multiple representations of superoperators which have identical effects; some care needs to be
taken to exactly specify the meaning of “identical ” in this setting. Denotational equivalence
can be refined to the following cases for cQPL:
❏ Direct denotational equivalence: We can distinguish three different scenarios which
exhibit direct denotational equivalence for (K1 , T1 , E1 ) and (K2 , T2 , E2 ) given as denotations of A1 , A2 :
1. card(K1 ) = card(K2 ), ∀i ∈ [0, . . . , card(K)[: K1 i = K2 i and E1 = E2 , T1 = T2 .
This means that both programs have have the same number of Kraus sets as
denotation, the probabilistic environment and the typing context contain the
same information and the denotations of the statements are pairwise identical.
2. E1 = E2 , T1 = T2 , card(K1 ) = card(K2 ), ∃ϕ ∈ Sym(n) : ∀i ∈ [0, . . . , card(K)[:
K1 i = K2 ϕ(i) such that the sum of commutator products calculated according to
the method given in Section 4.5.4.2 vanish identically, i.e., only permutations with
vanishing commutator have been used. This equality holds if only commuting
statements have been exchanged to match the lists, the total denotation is thus
identical.
3. ∀̺ ∈ D : (K1 , T1 , E1 )(̺) = (K2 , T2 , E2 )(̺) where (K, T, E)(̺) means the application of the Kraus set in K on ̺ where the information contained in E is utilised
to construct the proper superoperators because the exact representation of K in
general depends on information given in T and E. Note that the initial state
resolves any symbolic parametrisations which may be present in K.
The first condition obviously implies the second and third condition; the second implies
the third, but the other direction is not true in general, so equivalences of decreasing
strength are defined by this enumeration.
❏ Heap-permutative respectively variable-permutative denotational equivalence is given
if there exists a permutation of the quantum heap positions (respectively the variable
names) such that direct denotational equivalence holds. These permutation schemes
can be used to align different probabilistic environments to each other.
17 Note
that our definitions of equivalence do not coincide with the types of equivalence given by Selinger.
59
Chapter 5. Formal denotational semantics
Summary
Note that textual equivalence implies denotational equivalence, but the converse statement is not valid in general.
A last form of equivalence that needs to be considered concerns the parallel execution of programs. If {Ai } represents a set of n communicating modules, the order in
which the subsystems are given does not make any difference, i.e., JA1 ||A2 || · · · ||An K ∼
=
JAϕ(1) ||Aϕ(2) || · · · ||Aϕ(n) K for any ϕ ∈ Sym(n) and an according update of the references to
the other subsystems Ak , k 6= m in Am for all m. Likewise, relabelling of communicating
modules does not change the meaning of parallel execution if the reference names in all participating modules are updated correspondingly. This type of equivalence can be referred
to as communicative equivalence.
The problem of how to detect equivalence between different representations will emerge
several times in the following remarks and is not easy to solve constructively.
We have augmented the definitions of the semantic basis (K, T, E) with the elements required to represent communicating systems, i.e., cQPL programs which
are generally independent of each other, but can exchange quantum mechanical
and classical data in a well-defined way. Additionally, we have shown that fixed
points exist in this framework; they are necessary to assign semantics to numerous
components of the language as explained in Chapter 4. Criteria for the equivalence of cQPL programs were specified as well; this allows to check if programs
which are specified by different sequences of commands have the same effect.
5.3.6
Semantics of the language components
Chapter 2 gave an informal18 introduction the the language elements of cQPL. In this chapter, we will use the mathematical and semantical formalism introduced in the preceding
sections to give a rigorous mathematical meaning to these statements. By the compositionality of denotational semantics, this means that all cQPL programs (which are, necessarily,
composed of cQPL statements) have a defined semantics. There are two possible representations for cQPL programs in form of textual descriptions and graphical flow charts; we
resort to the particular representation that is more convenient for the desired purpose in
the following remarks. Establishing a formal correspondence between both possible representations of cQPL is obvious and follows exactly the argumentation used in [Sel04b]; we
will thus not bore the reader with details on how to relate both representations uniquely.
Note that we try to keep the purely classical formalism as terse as possible because
most problems related with this are not too interesting from a physical point of view. More
elaborate descriptions or gentle introductions can be found, e.g., in Refs. [Mos90, Rey98,
Win93].
5.3.6.1
Some notational remarks
Some conventions and notations widespread in semantics are uncommon in physics, thus we
want to make two short remarks before proceeding further.
Typed lambda calculus Computer science literature habitually uses the typed lambda
calculus (cf., e.g., [RP02]) to formulate the equations for valuation functions; this is useful to
18 It should be noted that although informal may sound a little fuzzy, such a description is normally the
maximal level of accuracy with which users of programming languages (and in most cases, implementors as
well) are confronted.
60
5.3. Denotational semantics of cQPL
not only clarify which parameters are used, but also to determine their type. We, in contrast,
use a different notation. Observe, for example, the denotation of the dyadic operator + which
adds two natural numbers:
DOJ+K(n1 , n1 ) = sum(n1 , n2 )
(5.29)
It is intuitively clear that we are talking about a function which takes two natural numbers
as arguments and computes another natural number as result. Nevertheless, we did not
formally specify the data types of the arguments nor of the result of the function.
This can be solved by using the typed lambda calculus in which the function would be
written as:
(5.30)
DOJ+K = λn1 ∈ N.λn2 ∈ N.sum(n1 , n2 )
This very simple example already demonstrates that the representation requires a considerable notational effort. Since nearly all valuation functions for the semantics of cQPL defined
in the following require (K, T, E) tuples in addition to the effective parameters, this would
unduly inflate the length of equations which does not really increase lucidity. Thus, we
stick to a simplified notation which follows the algebraic convention for functions as presented above. The domains where parameters originate from are normally clear from the
context; we will mention it explicitely should this not be the case since typing is obviously
not explicitely part of the simplified description.
Currying/Schönfinkeln Another point we want to mention is the insight that functions of
more than one parameter may always be composed by a number of subsequent functions
which take exactly one parameter. Thus, a function f (x1 , x2 , x3 ) = y with xi , y ∈ N which
is an element of (N × N × N) → N can also be seen as a mapping (N → N → N) → N, the
parentheses may also be omitted. The technique of transforming a function with multiple
arguments into a function which takes only one argument, but returns another function
which requires the remaining arguments and returns the result is conventionally termed
currying, although it was first introduced by Schönfinkel [Sch24]. The process can obviously
be repeated so that there are only functions left which take exactly one argument. In the
following, we will use the form which is more apt for the respective purpose.
5.3.6.2
State transformations and fixed points
Valuation functions for top-level elements of cQPL (i.e., those for expressions) work on a
three-tuple (K, T, E) and produce a new three-tuple (K ′ , T ′ , E ′ ) as we will see in the course
of the following remarks. Thus, these tuples form the domain which is the basis of semantics.
Since we will need fixed points as solution of several recursive domain equations, we need to
show how to calculate them for (K, T, E) tuples. We have already shown that fixed points
exist for Kraus aggregations which fulfil certain conditions. Now, we need to transfer this
to (K, T, E) tuples.
For this, note that the typing context is not involved in the calculation of fixed points:
Its purpose is to ensure well-typedness of programs (which will be explained in more detail
in Section 5.3.8) and (as a consequence) to make sure that ownership of qbits is unique.
Otherwise, it has no semantical meaning.
To consider the contribution of the probabilistic environment, observe that we could do
without it in principle by using a different notation as is, for example, the case in QPL
(we did not adopt this approach because it quickly leads to very long annotations which are
cumbersome to handle; additionally, our approach has a greater similarity with the standard
notation commonly used in denotational semantics). For every possible value of a variable
61
Chapter 5. Formal denotational semantics
in the probabilistic environment, a specific Kraus aggregation can be inferred. Consider
a measurement of two qbits whose result is stored in a classical variable of two bits. A
πx :PROJ (q,T )
probability distribution x −−−−−−−−−−→ [0, 1, 2, 3] is stored in the probabilistic environment
where PROJ (q, T ) represents the information that the quantum variable q contained in
the typing context T was measured (this will be covered in more detail in Section 5.3.6.4
on page 69). Since the exact form of the probability distribution depends on the state of
the measured variable about which nothing is known in the worst case (if, e.g., the variable
was received from a remote party which did not characterise it any further), we have to
account for all possible cases, i.e., for all values the variable can have in principle. Let
{Γ}; {M }#q be the contents of the Kraus list K from the (K, T, E) tuple immediately after
the measurement where {M }#q denotes the Kraus set for a projective measurement. This
list can with the help of the probabilistic environment be rewritten into a four-tuple
({Γ}; {P0 }, {Γ}; {P1}, {Γ}; {P2 }, {Γ}; {P3})
(5.31)
where {Pi } = |ii hi| is one of the projection operators which constitute the POVM elements
of the measurement. Note that the Kraus aggregations contained in tuples of this kind are
always unparametrised.19
If there is now an operation performed which is independent of the measured variable, the
contribution to the Kraus set is appended to all list components in this picture. Consider, for
example, the application of some operator U to another quantum variable v. The resulting
four-tuple of Kraus aggregations then looks like:
({Γ}; {P0 }; {U }#v , {Γ}; {P1 }; {U }#v , {Γ}; {P2 }; {U }#v , {Γ}; {P3 }; {U }#v ).
(5.32)
The first entry belongs to the case that x = 0, the second to x = 1, and so on. Obviously,
it is much simpler to write this in our notation as
({Γ}; {M }#q ; {U }#v )
(5.33)
from which the tuple representation can be reconstructed. This also works if operations
are considered that depend on the state of a classical variable governed by a probability
distribution. Consider, for example, the case that an operator U is applied to some quantum
variable v if the value of x is 2 (the program code for such an operation would be if (x = 2)
then v *= U;). In our notation, the branching condition is preserved in the probabilistic
environment as shown in the example given by Figure 5.5. From this information, the
corresponding tuple representation
({Γ}; {P0 }, {Γ}; {P1 }, {Γ}; {P2}; {U }#v , {Γ}; {P3 }; ).
(5.34)
can be constructed. Note that U is only applied in the case x = 2.
The transfer from our representation to tuples of Kraus aggregations is easier when
classical variables with defined values and no associated uncertainty are considered, so we
will not show an explicit example for this.
As the forgoing considerations have demonstrated, the probabilistic environment and
the Kraus aggregation contained in the (K, T, E) tuple can be used to construct a tuple of
Kraus aggregations where the tuple contains one entry for every combination of values the
classical variables can be in. Fixed points of (K, T, E) triples must therefore be calculated
separately for all possible Kraus aggregations that can be constructed from the triple because
19 This kind of split is one of the reasons why it is more convenient to work with Kraus representations of
cp-maps instead of the cp-maps alone which would also be possible in principle.
62
5.3. Denotational semantics of cQPL
each of them represents another possible meaning of the program. This is possible with the
methods introduced before. After the fixed points have been derived, the usual (K, T, E)
representation can be used again. Thus, calculation of fixed points effectively only requires
the properties of Kraus sets as introduced before. In the following, we need thus only make
sure that the conditions for the existence of fixed points of Kraus aggregations as given in
Theorem 5.3.1 are fulfilled to ensure the existence of fixed points for (K, T, E) tuples.
5.3.6.3
Classical operations
The classical subsystem of cQPL consists of the following parts:
❏ Allocation and (implicit) deallocation of classical variables.
❏ If-then-else expressions.
❏ While-loops (this and the previous point require the evaluation of boolean conditions
which must also be accounted for by the semantics).
❏ Sequential composition.
❏ Do-nothing-operation (skip).
❏ Sending and receiving of classical states.
❏ Assignment to classical variables.
❏ Calling procedures which manipulate classical data.
❏ Blocks.
Note that we do not cover sendig and receiving of classical data because this has extensively been covered in the literature. Additionally, it is in principle always possible to
achieve the same effects with the transmission of quantum mechanical information. In the
following, we cover only the valuation functions which are either absolutely indispensable or
are influenced by the quantum properties of our language.
Sequential composition Modifications made to the typing context, the probabilistic environment and the Kraus aggregation made by the first statement must be taken into account
when the semantics of the second statement is calculated:
EX PJS1 ; S2 K(K, T, E) = EX PJS2 K(EX PJS1 K(K, T, E))
|
{z
}
(5.35)
=(K ′ ,T ′ ,E ′ )
This is utilised many times in the denotational equations for quantum communication.
Blocks Blocks are used to introduce multiple levels of scope into programs. This can
happen both implicitly (e.g., in loops) and explicitely (by syntactical specification of blocks),
but there is no need to distinguish between these cases from a denotational point of view.
Superficially, a block looks just like a collection of multiple statements which are executed
one after another; but some additional points need to be taken into account:
❏ New variables (both quantum and classical) may be declared inside blocks, but they
cease to exist once the block’s scope is left.
63
Chapter 5. Formal denotational semantics
❏ New variables do overshade old ones if they share the identifier.
❏ Changed bindings of already existing variables are also visible after the control flow
has left the block’s scope.
Thus, the probabilistic environment is partially affected by a block. The typing context
before and after the block is identical and thus unaffected by the block’s effect,20 and the
Kraus aggregation records everything that has been done inside the block.
Since the mentioned problems appear in every programming language featuring blocks,
standard solutions are available in every textbook (as usual, cf. Refs. [Mos90, Rey98,
Win93]), so we will not explictely present them here to save some formal overhead.
Conditionals and Operators Dyadic operators combine two subexpressions into one result,
as, e.g., all arithmetic operations do. Conditionals are operators which result in a boolean
variable, i.e., they evaluate to one of the values true or false which are represented by 1
and 0. In contrast to most classical languages, the result of both types need not be fixed
with certainty, but can be governed by a probability distribution. Note that conditionals
may not be used as stand-alone expressions, but can only be part of conditional statements.
This is why their semantic valuation does not result in the usual (K, T, E) tuple, but in
a probability distribution for the possible results which is, e.g., apt for inclusion into the
probabilistic environment. The basic valuation functions are given by:
OPJa DO bK(T, E) = DOJDOK(OP (a)(T, E) ⊗ OP(b)(T, E))
OPJMO aK(T, E) = MOJM OK(OP (a)(T, E))
(5.36)
(5.37)
where DO ∈ { +, −, ∧, ∨, . . . } and MO ∈ { ¬, − }. The meaning of the operations is defined
as usual, but the probability distribution nature of the arguments needs to be taken into
account:
M
M
πa (v1 )πb (v2 )DO(v1 , v2 ).
(5.38)
DOJDOK(a, b) =
v1 ∈range(a) v2 ∈range(b)
This expression results in a new probability distribution that can be used by the elements
further up in the evaluation hierarchy.
Note that the denotation of a single variable is given by the corresponding probability
distribution which can be found in the probabilistic environment:
OPJvK(T, E) = E(v)
(5.39)
This definition ensures that chains of expressions using dyadic and monadic operators
(e.g., 42 + 23 + 4) are covered by the semantics because a and b in Eqns. 5.36, 5.37 can either
be values or other operator expressions.
Also note that the eventual action of the respective operators (+, ∧, . . . ) can be seen
intuitively, so we abstain from further formalisation and rely on the reader to use his common
mathematical sense.
Remark 5.3.8. Note that we do neither consider any problems of numerical accuracy nor
of limited ranges of representable numbers for the specific data types. Consequently, we also
do not care for the problem of division by zero. We are aware that such pitfalls exist, but
are not interested in their solution in this context since their nature is purely classical.
20 But
note that the typing context within the block may well be different than the one outside.
64
5.3. Denotational semantics of cQPL
If-Statements The semantic description of the if-statement is simplified by introducing
the following helper function:
f ((K0 ,T0 , E0 ), (K1 , T1 , E1 ), π, ν, (K, T, E)) =
p(ν = 0) · K ◦ (K0 − K) + p(ν = 1) · K ◦ (K1 − K),
T ⊕ ν : bit, p(ν = 0) · E ⊕ ν : π ⊕ E0 + p(ν = 1) · E ⊕ ν : π ⊕ E1
(5.40)
which eases selection of components of (K, T, E) tuples gained by other evaluations and
additionally circumvents repeated semantic evaluations of some components. The tuple
(K0 , T0 , E0 ) is the result of the evaluation of the if-branch, while (K1 , T1 , E1 ) is for the
then branch. π is the probability distribution governing the branch, and ν is the identifier
which is used to represent this distribution in the probabilistic environment. (Ki − K)
represents the Kraus aggregation that contains only the elements that were appended to Ki
in comparision to K; this ensures that only new contributions introduced in the branches
are added to the Kraus aggregation finally. The denotational description for the if-statement
then reads as
Jif c then C0 else C1 K(K, T, E) =
uid , (K, T, E))
f (EX PJC0 K(K, T, E), EX PJC1 K(K, T, E), OPJcK(T, E), |{z}
{z
} |
{z
} |
{z
}
|
(K0 ,T0 ,E0 )
(K1 ,T1 ,E1 )
π
(5.41)
ν
where uid is a unique identifier for the branch which can be chosen at will, but must not be
identical with other identifiers already in use. Such a choice is simple for non-communicating
programs. The extension to communicating systems is possible if every identifier is given a
unique prefix for each partner as described before.
Note that although dyadic operators might syntactically be used to describe arithmetic
operations and not necessary conditionals, this source of mistake is ruled out by the type
system which only allows boolean typed expressions for c.
While-Statements While statements can be solved using the fixed-point theorem given in
Eqn. 4.4.1. For this, note that the denotation of the while function can be rewritten using
the previously considered if-function together with an explicit block (c denotes a boolean
condition and S a statement):
JwK(K, T, E) ≡ Jwhile c do SK(K, T, E)
JwK(K, T, E) = Jif c then {c; w} else skipK(K, T, E)
(5.42)
(5.43)
Eqn. 5.43 is a recursive equation (c appears both on the left hand and the right hand side)
whose solution is a fixed point. The agreement is that the least fixed point is taken to be
the denotational solution, and this is exactly what the fixed point theorem delivers. To
write this formally is now a standard exercise of denotational semantics [Rey98, Mos90],
but we show it nevertheless because it is an instructive example for the technique of solving
recursive equations which will be necessary for the denotation of quantum communication.
Consider the function F given by
F f (K, T, E) = if JcK(K,T,E) then f (JSK(K, T, E)) else (K, T, E).
(5.44)
Then the fixed point solution can be formally written as
Jwhile c do SK = YΣ→Σ⊥ F,
65
(5.45)
Chapter 5. Formal denotational semantics
where Σ is any (K, T, E) tuple as usual; assume that f is defined like
(
⊥
if x = ⊥
f (x) =
f (x) otherwise
(5.46)
since we have to account for the case that the argument of f is ⊥ because of the recursion
(note that we would have to use two different symbols for f to write this überproperly).
Assignments Assignments in cQPL can be seen as a convenience mechanism which extends
the simple binding of identifiers to values; it is not necessary to introduce the concept of
stateful variables to the language to be able to formalise assignments. The description can
be simplified if some syntactical transformations are applied. For this, first consider the
following program fragment:
new int a := 10;
// Do something using a (part 1)
a := a + 4;
// Do something using a (part 2)
The assignment is equivalent to introducing a new identifier a’:
new int a := 10;
// Do something using a (part 1)
new int a’ := a + 4;
// Do something using a’ (and replace all a by a’) (part 2)
This strategy also works when blocks are taken into consideration:
new int a := 10;
if (...) {
new int a := 5;
a := a + 1;
...
}
else {
new int a := 1;
a := a + 1;
...
}
Note that not possible to employ a static renaming scheme in this case because the
identifiers a in both subsequent blocks would then end up with identical names which leads
to problems. We thus have to postulate that new identifiers are always chosen such that
they do not overlap with previous identifiers and must, of course, also not overlap with
identifiers which can be assigned by the user. This is possible if the set of identifier strings
is denoted by Σ∗1 , we introduce a second set of strings Σ∗2 with Σ1 ∩ Σ2 = ∅; each time
an identifier is overshaded, it is replaced in its complete scope with a new one in Σ∗2 that
has not been used before. While this policy is hard to implement in practice,21 it does not
21 This is exactly the reason why mechanisms like call-by-name disappeared as a curiosity some thirty
years ago.
66
5.3. Denotational semantics of cQPL
present any problem in theory (even cases which require an infinite number of replacements
are no problem because there are infinitely many unique identifiers).
Most important, the approach is also valid when loops are introduced because these can
be rewritten using a (possibly infinite) chain of appropriate if-then constructions as is done
in the denotation of them.
In conclusion, we do not need to take care for the obstacles introduced by overshading, but can simply ignore the problem in the denotation of assignments (ν denotes some
identifier):
Jν := ǫK(K, T, E) = (K, T, E ⊕ ν : EQN JǫK(K, T, E))
(5.47)
where ǫ is some arbitrary arithmetic expression (which includes single identifiers); the denotation of this is obvious and will not be considered further. Note that this valuation function
does not cover the case that the result of a quantum variable measurement is stored in a
classical variable; this will be covered later when we describe the denotation of the measure
funtion on Page 69.
Allocating and destroying variables As we have noted in the previous remarks, we do not
need to take the problem of overshading into account when dealing with assignments; this
obviously also applies to allocations. We refrain from a more detailled description of this
topic because everything necessary for the solution can be readily found in the literature, e.g.,
[Win93, Rey98, Mos90]. Note that allocating new variables does not present any problems
for the boundedness of the Kraus aggregation because the number of qbits is limited by Q,
an arbitrary, but finite quantity.
Procedure handling Procedures in cQPL follow the standard scheme of classical languages
for the non-quantum part. The denotation of such can therefore be directly taken from the
usual textbooks [Rey98, Mos90, Win93], so we will not repeat this here. The interesting
problem is given by recursive procedures, especially when they act on quantum parameters.
For simplicity, we consider directly recursive procedures with quantum parameters; the case
of indirect recursion is in princple identical, but necessitates more formal effort, so we skip
it here. The solution is an adaption of the method presented in [Sel04b, Sections 5.5 and
6.5] for our purposes.
Consider a procedure Y which depends on itself, e.g.,
Y = X(Y )
(5.48)
If Y is given as a flow chart, this can be interpreted as shown in Figure 5.7: A “hole” in the
representation of X is replaced by Y with another hole, this is again replaced by the same,
...
proc rec: test:qbit {
A
if (cond) call rec(test’);
else { ... }
B
}
Formally, we can thus define an approximation relation given by
Yi+1 = X(Yi ).
67
(5.49)
Chapter 5. Formal denotational semantics
Figure 5.7: Unwinding a recursive flow chart is done by placing the flow chart with
a hole (given by the white rectangle) into the hole and repeating the
process again and again. The limit of this sequence is given by a fixed
point as explained in the text.
where Yi are approximations of Y which get better with increasing i. Y0 = ⊥ is the crudest
approximation which simply does not terminate. The solution of Equation 5.49 is another
case for the fixed-point theorem. Once a solution for Y has been found, it can be applied to
any (K, T, E) tuple to calculate the required state transformation. Note that although the
procedure depends only on quantum variables, there nevertheless needs to be a termination
condition. In the example, this condition is given by the if-then-else statement in which
the recursive call is wrapped; since the procedure only takes a quantum parameter, the
condition is either trivial or depends on a measurement of quantum states. In the first case,
the condition depends only on classical variables which were allocated within the procedure,
the condition could thus be determined at compile-time and the recursion unrolled because
the recursion depth is already known. In the second case, we really need to unwind the
procedure.
5.3.6.4
Quantum operations
The quantum mechanically relevant operations of cQPL are :
68
5.3. Denotational semantics of cQPL
❏ Application of (unitary) operators.
❏ Calling procedures which manipulate quantum data.
❏ Sending and receiving quantum states.
❏ Measurements.
❏ Creating new and destroying existing quantum states, where the last operation is not
explicitely, but only implicitly possible when quantum variables drop out of the present
frame. Sending quantum variables makes them disappear from the scope so that they
can not be accessed any more, but does not destroy respectively deallocate them.
We will give formal denotations for these constructs in the remaining part of this section.
Obviously, the description of communication is our foremost concern, so we elaborate this
in most detail.
Unitary operators Unitary operators induce isometries, so the corresponding Kraus set is
obviously bounded and filfills the requirement for fixed points. Every unitary transformation can be expressed by a Kraus set with only one element. The semantics of a unitary
transformation acting on the qbits q1 , . . . , qk is given by
EX PJ(q1 ,...,qk ) *= UK(K, T, E) = (K ◦ {U }(#q1 ,...,#qk ) , T, E).
(5.50)
Note that the type system ensures that the variable/operator dimension on both sides of
the expression matches as required, i.e., the classical variable has the proper size to hold all
potential measurement outcomes.
Measurements Measurements are described by the following semantic equation:
πa :PROJ (q,T )
EX PJa := measure qK(K, T, E) = K ◦ {M }#q , T, E ⊕ a −−−−−−−−−−→ range(a) .
(5.51)
Hereby, PROJ (q, T ) denotes the function to generate the set of projectors for the type of
q which can be resolved from the typing context T . The required basis for the projection
t (χ(q))
(which is actually nothing else than the basis for F2q
in Dirac notation, also named the
standard basis) is given by
(χ(q))
tq O
B=
|in i ∀in ∈ {0, 1} .
(5.52)
n=0
The corresponding Kraus elements are obvious. Accordingly, πa : PROJ (q, T ) denotes the
probability distribution which connects the possible values of a with a probability distribution induced by the projectors. Since the measurement operators work on discrete states in
a finite-dimensional Hilbert space, they are obviously bounded.
Remark 5.3.9. Note that although we restrict measurements to projections onto the standard basis, projective measurements in arbitrary bases can be realised by applying appropriate
unitary transformations prior to the measurement.
69
Chapter 5. Formal denotational semantics
Sending and receiving qbits To consider sending quantum variables, we first split commands which send lists of quantum variables into a list of commands that send one quantum
variable each:
EX PJsend q1 , . . . , qn to moduleK(K, T, E) =
EX PJsend q1 to module; send q2 to module; . . . ; send qn to moduleK(K, T, E).
(5.53)
The denotation of a single send command is given by
EX PJsend q to moduleK(K, T, E) = K ◦ {S}#q , T ⊖ q, E ⊖ q .
(5.54)
{S}#q is not a real physical map as other cp-maps are, but only a “placeholder” to note that
qbits have been sent. This will become important when the semantics of parallel execution
is considered further below. The interesting thing here is that the sent qbit is removed from
the typing context and from the probabilistic environment. Thus, it is not visible any more
in the remaining statements. Further access to it can be detected as erroneous at compile
time; the typing context gives the formal basis for this.
Receiving qbits is described in a similar manner. First, a receive operation with multiple
quantum variables is split into a sequence of single-variable receive operations:
EX PJreceive q1 : qtype1 , . . . , qn : qtypen from moduleK(K, T, E) =
(5.55)
EX PJreceive q1 : qtype1 from module; . . . ; receive qn : qtypen from moduleK(K, T, E).
The denotation of a single-variable receive command is given by:
EX PJreceive q : qtype from moduleK(K, T, E) =
K ◦ {R}#q , T ⊕ q : qtype, E ⊕ q : module .
(5.56)
Again, {R}#q is a placeholder required to denote the semantics of parallel composition.
Parallel composition The compositionality principle of denotational semantics implies that
the formal denotation of sending and receiving qbits must be independent of the conjugate
action, i.e., sending must be independent from reception and reception must be independent from sending. This is fulfilled in the formalism of cQPL as we have shown on page
70. Nevertheless, the denotation of the communication as a whole requires (at last) two
communicating partners. It needs thus make use of both of them to assign semantics to
communication.
Before we start with the formal details, we want to motivate why parallel composition
is necessary at all. For this, consider the case of two processes where A sends a quantum
bit (which we call a1 ) to B and, later on, receives a quantum bit from B (which we call a2 ).
We must distinguish two different cases (note that for more difficult cases with an arbitrary
number of send and receive statements, the conditions become more complicated because
we need to account for more general schemes of mutual influence. These conditions will be
developed stepwise in the following):
❏ The returned quantum bit was not identical with the sent one, a1 6= a2 . The inequality
refers to the positions occupied by the qbits on the combined quantum heap, it is not
related with the names of the qbits.
❏ B returned the quantum bit it got from A, a1 = a2 .
70
5.3. Denotational semantics of cQPL
Consider the consequences for the semantics of parallel composition when the interaction
of both processes is considered (both cases can be treated identically if only the separated
semantics is taken into account): While in the first case (a1 6= a2 ), operations on a2 and a1
end up on different physical locations, the same operations must in the second case (a1 = a2 )
be applied to the same physical location. Therefore, we must be able to construct enough
information from the composed systems such that it is possible to distinguish between both
cases.
Additionally, the distinctness of qbits influences denotational equivalence; many different
orderings of the actions performed by A and B lead to the same semantics (the exact
conditions for this will be formulated later on), but the class of possible reorderings is
usually bigger if there was no interaction on the same physical location in communicating
modules. Figure 5.8 presents a visualisation of this fact using a pseudo flow diagram.
Figure 5.8: Flow graph for two communicating processes. A sends a qbit to B
and B sends one to A; if there are no operations on the same physical
locations, the shaded regions commute. This is obviously not the case
if one or more identical quantum heap positions are modified in both
paths: Since the shaded regions operate on the same physical qbit
then, their actions need not necessarily be interchangeable because
non-commuting operations may have been performed on the same
physical location.
There is no explicit statement in the syntax of cQPL which describes the combination of
two processes. As explained in Chapter 2, communication is realised implicitly using modules
which interact among each other. These modules must thus be given some semantical
meaning; this will be developed in the following. Note that we restrict the description to
two communicating processes which we call A and B (this might stand as an abbreviation
for the omnipresent Alice and Bob) at first. Consider the following cQPL fragment:
71
Chapter 5. Formal denotational semantics
module A {
...
};
module B {
...
};
Any valid cQPL code (except the definition of new modules) may be contained in the
bodies of module A and module B; both entities thus constitute regular cQPL programs
whose meaning is made up by the meanings of all statements they consist of. Thus, we
can write PROGJAK and PROGJBK for the denotation of the code as if the modules were
regular, uncommunicating cQPL programs.
To consider the communicative interactions between both, we introduce the operation
PROGJA||BK ≡ COMM(PROGJAK, PROGJBK)(K∅ , T∅ , E∅ )
(5.57)
where (K∅ , T∅ , E∅ ) denotes the inital (K, T, E) tuple without content. The valuation function COMM is used to compute the combined denotation of A and B which we also call
parallel execution. It must obviously only depend on the denotations JAK and JBK, i.e.,
(K, T, E) tuples. Since we do not want the semantics of communication to depend on the
order in which the communicating partners are specified, the operator || must necessarily
!
be commutative: PROGJA||BK = PROGJB||AK. This equivalence can easily be achieved:
It suffices to define the operator in such a way that the modules are ordered lexicographically, then the order in which they are specified does not influence the denotation; the
commutation relation is thus automatically fulfilled.
PROGJAK is used to evaluate a semantic equation with empty initial context; this is
obviously the case when the top-level of a program is considered (as is the case for modules)
where no definitions can have been made yet. Thus, PROGJAK ≡ EX PJAK(K∅ , T∅ , E∅ ).
Consider the evaluation of JAK and JBK which both do as usual depend on (K, T, E)
tuples. COMM results in the creation of a valuation function which depends on the tensor
product of these tuples, i.e.,
VALJAK(K0 , T0 , E0 )
VALJBK(K1 , T1 , E1 )
⇒ COMM(VALJAK, VALJBK) (K, T, E) .
| {z }
(5.58)
(K0 ⊗ K1 , T0 ⊗ T1 , E0 ⊗ E1 )
Both representations convey the same information. This is immediately obvious if the expressions are written as direct λ-abstractions as defined in Section 5.3.6.1 together with
an appropriate “untensoring” function which separates the tensor product of the combined
(K, T, E) tuple into two components. We will not show this explicitely to avoid introducing
even more symbols.
Note that termination of A and B alone does not imply termination of PROGJA||BK.22
This can be seen by considering the following simple program:
22 Here, the question arises how termination should be defined for communicating processes if only one
part (A) of a total system is considered. We could, for instance, define a demonic partner which always
supplies the required number of qbits the process needs and absorbs any number of qbits sent, but this will
not necessary lead to termination of A. The exact solution of the problem depends on the context in which
it is considered; we will not deal with it any further here, but define a process as terminating if there exists
a demonic partner that behaves in such a way that the process terminates.
72
5.3. Denotational semantics of cQPL
module A {
new qbit q1;
send q1 to B;
};
module B {
receive a1:qbit from A;
receive a2:qbit from A;
};
While both processes represent valid cQPL programs, they will obviously result in some
non-terminating program (which should thus denote ⊥) because B will wait forever for the
second qbit (a2 ) to be sent. Thus, COMM(x, y) = ⊥ is possible although both x 6= ⊥ and
y 6= ⊥.
Extensions of the semantical components (K, T, E) to the multi-party case were defined
in Section 5.3.3; they provide the basis for the denotation of parallel execution which will be
developed stepwise by considering bipartite systems without communication, bipartite systems with single send/receive pairs, bipartite systems with arbitrary send/receive statements
and finally, arbitrary multipartite systems.
Thus, let A and B be two programs which do not use any communication. The semantics
of their parallel execution can be readily denoted:
PROGJAK = (K0 , T0 , E0 ), PROGJBK = (K1 , T1 , E1 )
⇒
(5.59)
PROGJA||BK = (K0 ⊗ K1 , T0 ⊗ T1 , E0 ⊗ E1 ).
Note that the following equivalences hold if Ai , Bi are parts of a program which do not
contain any send/receive operations (we omit the required EX Ps in the second line to
simplify the notation):
EX PJA1 ||B1 ; A2 ||B2 K = EX PJA1 ; A2 ||B1 ; B2 K
{z
} |
{z
}
|
(COMM (JA1 K, JB1 K) ; COMM (JA2 K, JB2 K)) = COMM(JA1 ; A2 K, JB1 ; B2 K).
(5.60)
(5.61)
This ensures that it does not make any difference if we consider the sequential combination
of two parallel executions or the parallel execution of two sequential combinations as shown
in the formula; we will make use of this later on.
The situation gets more complicated if a single send/receive pair is allowed, i.e., if a single
quantum variable can be transferred from A to B (the case B to A is nearly identical, so
we restrict ourselfs to the first case). Let {S} and {R} denote the Kraus pseudo-operations
for sending and receiving. To see how these operations influence the possible equivalent
compositions of a program, observe the following two schematic Kraus aggregations:
ΛA = {A1 }, . . . , {An }, {S}, {An+1}, . . . , {Am }
ΛB = {B1 }, . . . , {Bn′ }, {R}, {Bn′+1 }, . . . , {Bm′ }.
(5.62)
(5.63)
To consider how these aggregations can be rearranged, we define the following function:
(
1 if A and B are operations on disjoint qbits
D(A, B) =
.
(5.64)
0 otherwise
73
Chapter 5. Formal denotational semantics
Remark 5.3.10. The same effect could have been achieved with the standard commutator
in principle, but we want the notation to emphasise additionally that not the simultaneous diagonalisability, but the fact that the operations work on disjoint bases is responsible
for the exchangeability. Additionally, using the standard commutator to show equivalence
between permutations of communicating systems would have interferred with the standard
permutation rules that would then have become more complicated.
Since the two processes are necessarily totally uncorrelated before the transmission takes
place, the parts given by the sub-aggregations
Λ1A = {A1 }, . . . , {An }
Λ1B
(5.65)
= {B1 }, . . . , {Bn′ }
(5.66)
Λ2A = {An+1 }, . . . , {Am }
(5.67)
can be composed as in Eqn. 5.59 because D(Λ1A , Λ1B ) = 1. Since the parts after sending
respectively receiving the quantum variable are uncorrelated as well (they both work on
disjoint sets of qbits; this property remains valid in the combined semantics because the
combination of the probabilistic environments creates an appropriate combined quantum
heap as described in Section 5.3.3), the sub-aggregations
Λ2B
= {Bn′ +1 }, . . . , {Bm′ }
(5.68)
can likewise be parallel composed as in Eqn. 5.59. The only thing which needs to be taken
into account is that references to the position of the received qbit must be replaced by the
position of the qbit on the combined quantum heap in the combined probabilistic environment. This can formally be achieved by the following function (here, S denotes a send and
R a receive statement):
′
⊗
⊗
SR(COMM(JSK, JRK))(K ⊗ , T ⊗ , E ⊗ ) = (K ⊗ , T ⊗ , E ⊗ ).
(5.69)
⊗
Here, K , T , E denotes the usual (K, T, E) parameter tuple with the additional requirement that it must have the structure which is gained by combining two (K, T, E) tuples
with a tensor product as given in Section 5.3.3. SR itself is responsible for two things: On
the one hand, it applies the effect of COMM(JSK, JRK) to the parameter tuple, and on the
other hand, it replaces the portion of E which contains the information about the received
quantum variable so that it now points to the position of the sent quantum variable on the
combined quantum heap; the resulting probabilistic environment is denoted by E ′ . This
is obviously possible since SR does have access to the information provided both by send
and receive. To illustrate the effect of SR, consider the following example: Let the sent
quantum variable be denoted by q and the received one by r. The names of these variables
will have thus been changed to Lc (1)q and Lc (2)r. The receiving module does not have
any information about the received quantum variable except its type and its local name;
the position on the combined quantum heap is unknown. This can now be changed by SR;
for that, it inserts the position of Lc (1)q on the combined quantum heap into the combined probabilistic environment such that Lc (2)r points to it. Nothing more is necessary to
identify the received quantum variable with the sent one.
Note that RS defines the analogous function for the receive/send case. This is necessary
when bidirectional communication between processes is considered. Except the inverted
direction of data flow, the function is completely equivalent to SR.
Thus, the semantics of parallel execution for processes with a single send/receive operation can be reformulated as
PROGJA||BK = PROGJA1 ||B1 ; S||R; A2 ||B2 K
74
(5.70)
5.3. Denotational semantics of cQPL
where JAi K and JBi K denote the parts of the program which induce the operations described
by ΛiA,B as given by Eqns. 5.65–5.68. It is already known how to compute the semantics of
JA1 ||B1 K. To compute the denotation of the complete statement, we first consider how to
include the send/receive pair into the description:
PROGJA1 ||B1 ; S||RK = SR(COMM(JSK, JRK))
⊗
⊗
(COMM(PROGJA1 K, PROGJB1 K)(K∅
, T∅⊗ , E∅
)).
(5.71)
Note that although PROG with argument JA1 K, JB1 K appears both on the left and right hand
side of this equation, it is not truly recursive, but can here be seen as just a breakdown into
simpler cases. A general recursive formula will be derived in the following.
By using Eqn. 5.70, we can now give the denotation of the complete parallel composition
as defined in Eqn. 5.61; for this, we define the righthand side of Eqn. 5.71 to be denoted by
ξ to increase clarity:
PROGJA1 ||B1 ; S||R; A2 ||B2 K = COMM(PROGJA2 K, PROGJB2 K)(ξ).
(5.72)
Note that establishing the semantics of a given description is one thing we need to do;
finding equivalent descriptions for a given communication is another task. For this, we need
to consider all rearrangements that preserve semantics. This allows us to decide if two
communicating programs are identical because we can check if they can be brought to the
same form.
For the case of a single send/receive pair, the operations can be shifted in the Kraus
aggregation if certain conditions hold:
❏ The send operation can be postponed to the end of the aggregation (at least in the case
where no more send/receive operations take place) or brought forward by k positions
in the Kraus aggregation if D({An−l }, S) = 1 ∀l = 0, . . . , k − 1.
❏ The receive operation can be brought forward to the first position of the Kraus aggregation (again, this relies on the fact that only a single send/receive operations takes
place) or postponed by k positions if D({Bn+1+l }, R) = 1 ∀l = 0, . . . , k − 1.
Performing shifts characterised by these operations together with reorderings as given
by Eqn. 5.26 generates the equivalence class of all programs with a single send/receive pair
that posses the same semantics.
The next step is to include arbitrary send/receive operations into the communication
between A and B. As in the case of a single send/receive combination, the send/receive
statements act as synchronisation points; the denotation needs thus be aligned along them.
The problem to solve is now given by
PROGJA1 ||B1 ; S1 ||R1 ; A2 ||B2 ; S3 ||R3 ; · · · ; Sn−1 ||Rn−1 ; An ||Bn K.
(5.73)
First, we will establish the semantics for the given ordering; semantics-preserving permutations will be considered afterwards.
Note that we only consider the denotation of data flow in one direction for simplicity;
the semantics of the case R||S which may appear mixed with the other form is gained by
replacing SR with RS at the appropriate places). To find a solution for this equation, define
A′ ||B ′ ≡ A2 ||B2 ; Sn ||Rn ; · · · ; Sn−1 ||Rn−1 ; An ||Bn .
(5.74)
Eqn. 5.73 then has the form
PROGJA1 ||B1 ; S1 ||R1 ; A′ ||B ′ K.
75
(5.75)
Chapter 5. Formal denotational semantics
According to Eqn. 5.72, the solution of Eqn. 5.75 is given by
EX PJA′ ||B ′ K(SR(COMM(JS1 K, JR1 K))
⊗
⊗
(COMM(PROGJA1 K, PROGJB1 K))(K∅
, T∅⊗ , E∅
)).
(5.76)
By introducing ξ1 as abbreviation for the part following EX PJA′ ||B ′ K and expanding A′ ||B ′
to A2 ||B2 ; S2 ||R2 ; A′′ ||B ′′ , the formula reads as
EX PJA2 ||B2 ; S2 ||R2 ; A′′ ||B ′′ K(ξ1 ).
(5.77)
This type of equation is already well-known; it can be further resolved to
EX PJA′′ ||B ′′ K(SR(COMM(JS2 K, JR2 K))
(COMM(PROGJA2 K, PROGJB2 K))(ξ1 )).
(5.78)
By recursively defining
⊗
⊗
ξ0 = (K⊥
, T⊥⊗ , E⊥
)
(5.79)
ξi = (SR(COMM(JSi K, JRi K))(COMM(PROGJAi K, PROGJBi K))(ξi−1 ))
(5.80)
we see that the final solution of Eqn. 5.73 is given by
EX PJAn ||Bn K(ξn )
(5.81)
where ξn needs to be expanded as defined above.
The whole process thus leads to a recursive valuation function given by
EX PJA||B; S||R; C||DK(K, T, E) =
EX PJC||DK(SR(COMM(JSK, JRK))(COMM(EX PJAK, EX PJBK))(K, T, E)
(5.82)
whose solution can be found by resolving the recursion in the usual way.
Now, consider which alternative orderings of the Kraus aggregations preserve semantics
in the multi send/receive and receive/send case. For this, observe the following symbolic
representation of two Kraus aggregations (to save some notational effort and to increase
lucidity, we represent the Kraus sets which are not concerned with communication by boxes.
Although the boxes have identical widths, they do not need to contain the same number of
Kraus sets):
A1
S1A
B1
R1B
A2
R1A
A3
B2
S1B
B3
(5.83)
.
(5.84)
As in the case of a single send/receive pair, we know that the blocks A1 and B1 (considering again the compatible displacements of S1A and R1B ) can be arbitrarily combined
because they work on disjoint subsets of the quantum heap; the same holds for A2 , B2 and
A3 , B3 . In addition to the previously given rules, the following restrictions hold for shifting
send and receive operations in multi send/receive scenarios:
❏ Two consecutive send statements can only be interchanged if the corresponding receive
statements are interchanged, and vice versa.
76
5.3. Denotational semantics of cQPL
❏ If D(Ai , Bi+1 ) = D(Bi , Ai+1 ) = 1 the blocks Ai , Ai+1 and Bi , Bi+1 can be taken like
single blocks when possible orderings according to Eqn. 5.26 are considered. Note that
the validity of this condition can only be detected if the combined quantum heap is
considered. If it holds, then the sender does nothing to the sent and the receive does
nothing to the received quantum bit; thus, the statements contained in the blocks can
be executed in arbitrary order.
Two systems are identical if their denotations can be unified by performing rearrangements
according to these rules.
Another possibility that needs to be considered is the case where the number of send/receive pairs in the parallel processes does not match. As we have described before, this leads
to a non-terminating process because either the sending process wants to ship a quantum
variable, but cannot deliver it and thus blocks or the receiver wants to get a variable, but
blocks indefinitely because there is no sender for one. The semantics should thus be given
by ⊥ for both cases.
For the case of two parties (where we do not need to consider the case that a balanced
number of send/receive statements is present, but the distribution among the parties is
unmatched), this is covered by the following definition:
EX PJA||BK = ⊥ if R(JAK) 6= S(JAK) ∨ S(JAK) 6= R(JBK)
(5.85)
where S denotes the number of send and R the number of receive statements. The denotations JAK and JBK are supposed to be those derived for the parallel composition. The
extension to higher-dimensional systems is obvious, so we omit it here.
Finally, we can describe the semantics for multi-party communication with arbitrary
send/receive statements which is the most general case and therefore includes everything
considered before. The problem which needs to be evaluated is given by
PROGJA1 ||A2 || · · · ||An K,
(5.86)
where (note the different notation compared to before!) Ai represents all statements given
in module i; this may contain any number of send/receive statements that are now denoted
by Ski and Rki where i is the receiver for S and destination for R and k the sequence number
within the other send/receive statements of the communication channel the statement works
in (if we consider for example three parties A1 ,A2 and A3 , then there are the channels A1 –A2 ,
A1 –A3 , A2 –A3 ).
The semantic context the evaluation is based on is given by the tensor product of the
semantic contexts of the subsystems, i.e., (K ⊗ , T ⊗ , E ⊗ ) = (K1 ⊗ · · · ⊗ Kn , T1 ⊗ · · · ⊗
Tn , E1 ⊗ · · · ⊗ En ) and equivalent for the initial context. The valuation function COMM
given by Eqn. 5.57 can be extended from the two-party case to the n-party case without any
problems, we denote this by COMM⊗n . Since communication still takes place between two
partners (although there are now many choices for such two-partner subsystems), there is
always a pair of corresponding send/receive respectively receive/send statements. To take
the n-dimensional semantical context into account, the definition of SR (the version for
two parties is given in Eqn. 5.69) needs to be adjusted as follows when sending a quantum
variable from system m to system m′ is to be covered:
′
′
′
SRm,m (COMM⊗n (JS m K, JRm K))(K ⊗ , T ⊗, E ⊗ ) = (K ⊗ , T ⊗ , E ⊗ ).
(5.87)
SR is again responsible to apply the effect of COMM⊗n (JSK, JRK) to the parameter
tuple; note that in this case, do-nothing-operations 1 are used for all systems except m and
77
Chapter 5. Formal denotational semantics
m′ because these are not concerned with the communication. This is to ensure that the
dimensionality matches.
Additionally, SR replaces the portion of E which contains the information about the
received quantum variable so that it now points to the position of the sent quantum variable
on the combined quantum heap. This is identical to the effect in the simplified case for two
systems. If in this case the sent quantum variable is denoted by q in system m and the
received one by r in system m′ , then the names of these variables will have been changed to
Lc (m)q and Lc (m′ )r. SR simply inserts the position of Lc (m)q on the combined quantum
heap into the combined probabilistic environment such that Lc (m′ )r points to it.
With this, we can generalise the recursive definition of Eqn. 5.82 to the case with an
arbitrary number of participants:
′
EX PJA1 || · · · ||An ; S m ||Rm ; B1 || · · · ||Bn K(K, T, E) =
′
′
EX PJB1 || · · · ||Bn K(SRm,m (COMM⊗n (JS m K, JRm K))
(COMM(EX PJA1 K, . . . , EX PJAn K))(K, T, E).
(5.88)
Finally, this is the solution to the most general case of communication which can be
expressed in cQPL.
5.3.7
Explicit transformations of density matrices
Although the abstract view on quantum operations which we have presented in this work
is quite suitable for reasoning about general formal properties of quantum systems, the demands of practical work are usually of a different nature: Here, one is interested in the
calculation of explicit states and probabilities which determine a system and allow predictions about its past, present and future behaviour. This goal is usually achieved by specifying
the initial state of the system, subjecting this to diverse transformations and measuring the
required properties which give rise to the desired explicit probability distributions.
The semantics of a cQPL program can be used to generate exactly this information:
The abstract transformation given by the semantical denotation of a program is a function
which maps the density matrix of the input state to the density matrix of the output state.
Obviously, the election of a certain density matrix as initial state implies loss of generality,
but in turn allows to infer real-world information, not just abstract properties of generalised
systems.
5.3.8
The type system
Typing judgements make statements about the connection between expressions and their
types; cf., e.g.,[Car97] for an introduction. For our purposes, the following two building
blocks are necessary to describe the properties of cQPL:
E ⊢ e : T ⇔ Expression e has type T in E
E ⊢ F ⇔ F is well-typed in E.
(5.89)
(5.90)
Proper typing is necessary to eliminate certain runtime errors by applying appropriate compile time checks (cf. Section 5.4). Additionally, it is the key to showing that our formalism
ensures that quantum bits can – especially in communicating systems – be only manipulated
by one party at a time (a similar line of reasoning, albeit for a quite different formalism,
was used in [GN05]). The typing context provided by T in the (K, T, E) tuple is the basis
for this.
78
5.3. Denotational semantics of cQPL
Properties of the type system are customary expressed with judgements of the following
general form:
P1 · · · Pn
(5.91)
C
where the Pi are called the premises and C the conclusion. If all premises are true, the
conclusion is fulfilled. Such judgements can be used to deduce the type of a given composite
expression in an automated, formal manner. The following elementary typing judgements
hold for cQPL:23
Scalars
new t n := v
C1 ; C2
Conditionals
Arithmetic
i∈N
(analogous for bits, floats etc.)
E ⊢ i : int
E⊢v:t
E⊢n:t
E ⊢ C1
E ⊢ C2
(Composition preserves well-typedness)
E ⊢ (C1 ; C2 ) : void
E ⊢ v1 : t1 ∧ E ⊢ v2 : t2 c(t1 ) = c(t2 ) = 1
(op ∈ { <, >, =, . . . })
E ⊢ op(v1 , v2 ) : bit
E ⊢ v1 : t1 ∧ E ⊢ v2 : t2 c(t1 ) = c(t2 ) = 1
(op ∈ { +, −, ·, :, . . . })
E ⊢ op(v1 , v2 ) : max(t1 , t2 )
(5.92)
(5.93)
(5.94)
(5.95)
(5.96)
Again, we do not consider division by zero or overflows; up- and downcasting of data
types and procedure handling is also skipped. An equivalence relation between two types
was given by Eqn. 5.11 in Section 5.3.1.1; this can be immediately carried forward to typing
judgements:
E ⊢ x : σ1
.
(5.97)
σ1 ∼
= σ2 ⇒
E ⊢ x : σ2
Note that we do not consider these equivalences explicitely in the following to simplify the
notation, all statements are automatically supposed to hold for all equivalent types as well
without further noting this.
Also note that subtyping (i.e., considering one type as a subtype of another and allowing
appropriate conversions) is not explicitely taken into account because this is also a problem
which is specific to the classical data types of cQPL and thus not of too much interest here.
5.3.8.1
Quantum variable tuples
Tupling of variables must make sure that no component appears more than once in the list
because this would allow to write programs which violate the no-cloning principle and thus
lead to runtime errors. Formally, the requirement is given by24
∀k = 1, . . . , n : q(σk ) = 1 ∧
#xk ∩ (#x1 ∪ · · · ∪ #xk−1 ∪ #xk+1 ∪ · · · ∪ #xn ) = ∅
P
.
E ⊢ (x1 , . . . , xn ) : i σi
(5.98)
The meaning of this is as follows: E ⊢ xi : σi formulates the requirement that all variables
are well-defined. The condition ∀k = 1, . . . , n : q(σk ) = 1 requires that all components are
E ⊢ x1 : σ1 · · · E ⊢ xn : σn
23 Remember:
c(x) = 1 ensures that the data type of x is purely classical.
q(k) = 1 ensures that the data type does not contain any classical components, #q denotes
the positions in the quantum heap occupied by a quantum variable.
24 Remember:
79
Chapter 5. Formal denotational semantics
quantum data types; the tuple thus has no classical components which is justified by our
abdication of mixed types. The condition #xk ∩(#x1 ∪· · ·∪#xk−1 ∪#xk+1 ∪· · ·∪#xn ) = ∅
is a formal version of the requirement that no variable may appear more than once in the
list of variables. The conclusion which can be drawn from these premises is that (x1 , . . . , xn )
is a proper quantum variable tuple, i.e., a well-typed expression in the current environment.
5.3.8.2
Application of operators
The application of unitary operators requires that the dimension of the operator matches
the dimension of the variable or variables it is applied to. Formally, this is written as
E ⊢ (x1 , . . . , xn ) : q tq (q) = dim(U ) ∧ U ∈ U (n)
.
∀k : E ⊢ xk : qk ∧ ((x1 , . . . , xn )*=U ) : void
(5.99)
The statement additionally ensures that the typing of the qbits involved is not influenced
by the operator application. Note that we do not explicitely specify a formal condition for the
unitarity of an operator U given in terms of a function of its components. The membership
in U (n) is sufficient for our purposes. The distinctness of the destination variables for the
transformation is already ensured by the tupling requirements given above, so it does not
need to be checked explicitely.
5.3.8.3
If conditionals and while loops
The condition for this construction must have type bit, whereas both possible paths must
be well-typed:
E ⊢ c : bit
E ⊢P ∧E ⊢Q
.
(5.100)
E ⊢ (if c then P else Q) : void
A similar condition holds for the while loop:
E ⊢ c : bit
E⊢P
.
E ⊢ (while(c) do P ) : void
5.3.8.4
(5.101)
Measurements
The classical data type used to store the result of a measurement must have the same number
of bits as there are qbits in the quantum variable. This is represented by the condition
E ⊢ a : σ1 , c(σ1 ) = 1 ∧ E ⊢ b : σ2 , q(σ2 ) = 1 tq (σ2 ) = tc (σ1 )
E ⊢ (a := measure b) : void
5.3.8.5
(5.102)
Communication
Sending qbits When qbits are sent, the type system has to make sure that no qbit is
sent twice because this would result in the same effects as if operators could be applied to
multiple copies of the same qbit; using a tuple to combine the sent qbits automatically solves
this problem:
E ⊢ (x1 , . . . , xn ) : σ, q(σ) = 1
.
(5.103)
E ⊢ (send q1 , . . . , qn ):void
Note that the type system is not concerned with the actual receiver of the qbits; this
information is only required for the denotation of the expression, but not to ensure welltypedness.
80
5.4. Avoidance of runtime errors
Receiving qbits When qbits are received, the type system must make sure that the destination variables are not yet defined in the receiver’s context, i.e., they must not be well-typed
expressions. Afterwards, the variables used in the receive statement are well-defined in the
typing context and have the data type required by the statement. This can be formally
written as
∀i : ¬(E ⊢ xi ), q(σi ) = 1
.
(5.104)
E ⊢ (receive x1 : σ1 , . . . , xn : σn ):void ∧ ∀i : E ⊢ xi : σi
As in the case of sending, it is not interesting for the type system from which communication
partner the qbits originate.
5.4
Avoidance of runtime errors
QPL is a functional language with a static type system which guarantees the absence of
runtime errors (note that functionality is subject to a precise definition of the term; it is
certain that classical languages need to have additional properties – most important higherorder functions – to be called fully functional. But this is not really relevant from a physicist’s
point of view, as we have discussed before). It is very desirable that runtime errors can be
avoided as far as in principle possible, from a physicists point of view, it does not matter
how this is achieved. This desire was brought forward into cQPL and manifests itself in two
points: Cloning is (as in QPL) prevented already at the syntactical level, and communication
does not allow different processes to access qbits concurrently.
5.4.1
Unique ownership of qbits
Observation 5.4.1. No part of the quantum heap is accessible to two or more parties at
the same time during parallel execution of arbitrary cQPL programs.
Rationale: When a module is considered stand-alone, it is obvious that all qbits present
in the system are owned by one party. Uniqueness of the ownership is guaranteed by the
quantum part of the probabilistic environment which ensures (as described in Section 5.3.1)
that it is impossible for two or more names to refer to overlapping sets of qbits.
Parallel composition of systems is performed by always considering pairs of send/receive
statements; while sending removes the sent qbit from the typing context of the originating
system, receiving adds it to the typing context of the destination. Since both commands are
always considered in pairs and denoted atomically,25 a quantum variable may not be in two
typing contexts at the same time. Access to quantum variables is only possible for a user
when the variable is present in his typing context, this ensures (together with the fact that
quantum heaps of communicating systems cannot overlap by virtue of Definition 5.3.3) that
it is impossible for two or more modules to access identical qbits at a time.
It is possible to prove this statement formally based on the observations in Section 5.3.8.
Since this is on the one hand a general problem of semantics theory and on the other
hand burdened with many technical difficulties, we omit a precise proof here, but refer to
25 This
means that nothing can happen in between sending and receiving the quantum bit.
81
Summary
We have given the denotational semantics of all language components of cQPL excluding some standard cases that are readily available in the literature. Together
with the definition of the type system (by intentional omission of all technical
details), this completes the effort of assigning a precise meaning to quantum programs written in cQPL.
Chapter 5. Formal denotational semantics
[WF94] where the exact details can be found. [GN05] is one source where the proofs of
the aforementioned reference have been adapted to a quantum system which fulfils exactly
the same properties as ours (basically, not too much except notational details needs to be
changed).
X
5.4.2
Prevention of cloning and unphysical situations
One of the fundamental consequences of quantum mechanics is that it is impossible to
define a unitary operator that can duplicate arbitrary quantum states with perfect fidelity; a
straightforward calculation shown in nearly every text on quantum mechanics (e.g., [NC00,
Pre99]) proves this. Obviously, quantum programming languages must make sure that
cloning is forbidden because otherwise, processes contrary to the laws of physics could be
simulated. Since most other approaches to quantum programming (e.g., [Öme98, BSC01,
Kni96, SP00]) require the possibility to address quantum bits via references or pointers, they
cannot ensure at compile-time that two distinct variables do not share the same quantum bit;
they must provide appropriate checks at runtime which ensure this condition. Aside from
efficiency considerations, this is unsatisfying because especially for long-running programs,
termination with an error which was caused by a programming mistake is undesirable.
The static typing of QPL allows together with some syntactical checks to ensure that
once a program was approved to be correct by the static syntactical and semantical analysis
of the compiler, no runtime errors caused by unphysical cloning of quantum states can
happen. A similar statement can be observed for cQPL:
Observation 5.4.2. cQPL programs which do not use communication primitives can be
guaranteed to execute without runtime errors if the syntactic and semantic analysis deems
them correct.
Rationale: No quantum bit can be referred to by multiple identifiers in cQPL, as was
shown in the previous section. Thus, the distinctness of the quantum components of a list
of identifiers (which is used to specify the list of qbits an operator works on or given as
parameter to a procedure) can be guaranteed by ensuring that the same identifier does not
appear multiple times in the list. Therefore, the same line of reasoning for the impossibility
of cloning or generating unphysical situations as in [Sel04b, Section 4.8] applies.
X
Remark 5.4.1. Note that non-termination is something different than a runtime error.
Remark 5.4.2. Note that static typing can also prevent the possibility for some runtime
errors which originate from the classical parts of the language; this is well-known in programming language theory (cf., e.g., Refs. [RP02, WM95, App04] for details) so that we will
not dwell into this any further here.
5.4.3
Unavoidable non-termination conditions
Albeit cQPL tries to prevent runtime errors as good as possible, the introduction of communication opens the possibility of writing programs that cannot be checked at compile time
if they will terminate at runtime although nothing would hinder the separate modules to
terminate. Nevertheless, by restricting the code to a certain subset of cQPL,26 it is still
possible to produce programs which will execute guaranteed without termination problems
and without runtime errors. Note that non-termination is not considered as a runtime error.
26 Which can, in principle, solve all problems that might arise in quantum programming, but is not a very
practical.
82
5.4. Avoidance of runtime errors
If a program of the form while (1) do skip is provided, then executing the skip command forever (and thus doing nothing forever) is exactly the intention of the program and
therefore the correct behaviour which should be reflected by the denotation.
In Section 5.3.6.4, we have already considered an example of a non-terminating program.
The culprit here was the different number of sent versus received qbits, but since the number
of sent and received qbits is fixed at compile time on both sides, this error can obviously
be detected by the semantic analysis; the program can be rejected. Unfortunately, this
possibility is not always the case because the exact number of how many qbits will be
sent and how many will be received can not be decided in general. Consider the following
example:
module A {
new qword nq;
nq *= H(8);
new word n := measure nq;
while (n >= 0) {
new qbit q;
send q to B;
n := n-1;
}
};
module B {
receive q1:qbit, q2;qbit, q3:qbit;
};
Since n in module A may contain (with equal probability) any value in [0, 28 − 1], the
number of sent qbits cannot be determined with certainty, but is governed by the probability
distribution of n. It may be the case that the program terminates (namely, if exactly three
qbits are sent by A), but it may also be the case that less or more than three qbits are sent.
This results in either a blocking process A which cannot find a receiver for the qbits it wants
to transmit, or in a blocking process B which is not satisfied with a proper number of qbits
and blocks to wait for the missing ones.
Fortunately, there are only three commands in cQPL that allow to execute a sequence
of communication commands for which it is not possible to determine at runtime how many
there will be, so we can make the following observation:
Observation 5.4.3. cQPL programs using communication can be guaranteed to execute
without runtime errors if the syntactic and semantic analysis deems them correct and the
following possibilities of the language are not used:
❏ While-loops with a termination condition that contains a probabilistic variable.
❏ If-conditionals that are based on a probabilistic variable.
❏ Recursive procedures whose recursion depth cannot be determined at compile time.
Rationale: All send and receive statements which are given as a sequence of commands
(which may include the use of blocks) can be counted at compile time; their order is obviously
also known. If the if-statement is used with a condition that can be computed at compile
time, one path can be eliminated. Thus, the statement is nothing else than a regular
contribution to the list of statements. While-loops with a compile-time computable number
83
Chapter 5. Formal denotational semantics
of iterations can be replaced by inlining the loop body the appropriate number of times, so
they also become only a regular contribution to a sequence of commands. If the recursion
depth of a procedure can be calculated, it be converted to an iteration where the number
of steps and thus the number of communication commands are known. Therefore, it is also
just a regular contribution to a sequence of commands.
X
Remark 5.4.3. Note that the checks required to determine the number of loop iterations
etc. at compile-time are based on well-understood analysis techniques in computer science;
nevertheless, we did not actually implement these checks in the cQPL compiler because it is
nothing else than a routine task with little benefit and no gain of any valuable insight, but
just a technical problem.
84
I ain’t no physicist, but I know what matters.
Popeye the sailor
6
6.1
Prospects
Outlook
The field of quantum programming languages is – as everything connected with quantum
information – still a young one, and many things that are standard in classical programming
languages still need to be adapted for these. Some ideas which were tried to be realised
during the work on this thesis, but did not reach fruition are:
❏ Integration of higher-order functions. Everything we tried ended up in requiring closures for an implementation, but this is (to our knowledge) impossible to achieve
because of the no-cloning theorem. Having them would be quite desirable for many
applications.
❏ The ability to describe the complete loss of quantum bits caused by imperfect channels
or eavesdroppers. This is obviously hard to integrate into a programming language,1
but should be possible by heading for a protocol specification variant of cQPL.
❏ Consideration of more general eavesdropping models where the strategy needs not be
fixed, but can be one of multiple independent alternatives. The work provided in
[dH02] would possibly provide a suitable basis with demonic choices.
❏ Most texts about quantum programming languages extensively use categories to describe the underlying structures. We found that this does not really add any substantial
points, but merely more notation and nomenclature, so we did not follow this style
although some effort was made in the beginning to become familiar with the field.
Nevertheless, the material provided here could serve as starting point for the following
possible extensions:
❏ Quantum instead of classical control, i.e., allowing conditions to be based on quantum
and not classical logic. It would be possible to simulate this with QCL, but the benefit
is questionable because no known quantum algorithm makes use of such a feature.
❏ The method presented here could provide a basis to formulate quantum process algebras as, e.g., presented in [GN05, AM05].
❏ An extension from discrete to continuous systems would allow the simulation of general quantum systems and could thus be useful for a much wider range of quantum
information applications.
1 Just think about the situation that would arise if variables in classical programming languages could
randomly disappear. . .
85
Chapter 6. Prospects
❏ Faulty hardware models could be integrated at the simulation layer, but this would
presumably be very challenging at the semantic level.
Initially, it was planned to also investigate the possibility of integrating the semantic
framework into a theorem prover which could possibly facilitate automated analysis techniques for quantum protocols. Some preliminary experiments were performed by describing
the BB84 protocol in a classical protocol simulator, but this has only shown that the gap
between the requirements for such an automatisation and what is currently available is still
very wide for all approaches to quantum programming.
6.2
Latest developments
After this thesis was finished, another QPL compiler written by D. Williams was presented in
a joint work by Nagarajan, Papanikolaou and Williams [NPW05]. Since both efforts work on
closely related fields, it seems apt to sketch similarities and differences of them (to distinguish
it from our implementation, we call William’s compiler sqrQPL2 in the following):
❏ Both compilers use the quantum computer model defined by Knill [Kni96] as basic
architecture.
❏ sqrQPL provides an own quantum simulator which is called sequential quantum random
access machine. Code generated for this architecture resembles machine language quite
closely.
❏ sqrQPL provides support for a smaller subset of QPL than cQPL.
❏ sqrQPL has the ability to automatically decompose arbitrary unitary matrices into a
set of standard gates. As a result (and in addition to theoretical elegance) of this, the
simulator needs only provide support for very few different elementary gates.
❏ The semantics of sqrQPL is fully covered by the one given for QPL, while cQPL needs
to provide additional semantics for the added features.
❏ cQPL already includes support for communication and concurrency, whereas work to
bring these abilities to sqrQPL will be started in the future according to [NPW05].
It would be interesting (and should be possible without too much effort) to provide an
SQRAM-backend for cQPL; since Ref. [NPW05] states that support for communication and
concurrency is (at least in preliminary form) already present in their simulator, no major
obstacle does seem to exist to hinder such an endeavour. In summary (and, obviously, seen
from the author’s subjective point of view) sqrQPL is a straight implementation of QPL
where the ability to decompose complicated gates into a set of simpler ones is the essential feature. The focus of cQPL is mainly on the semantics of (quantum) communication;
although the cQPL compiler seems (at the time of writing) to provide a bigger and more
versatile language core than sqrQPL, it is more or less a by-product of the actual work.
2 Because their compiler targets a virtual machine which is termed sequential quantum random access
machine.
86
. . . und Lasse sagte, die Sprache der Jungen
sei sowieso die einzig wahre.
Astrid Lindgren, Wir Kinder aus Bullerbü
A
List of symbols
The following presents a list of symbols used in this work. Note that the meanings given
here are not necessarily the only ones with which they were used.
A . . . . . . . . . . . Observable algebra
#number . . . Unique node id
A . . . . . . . . . . . Finite ordered set
#string . . . . . Position(s) occupied by
quantum variable variable
on the quantum heap
JK . . . . . . . . . . . Separate syntax and
semantics
|| . . . . . . . . . . . . Parallel composition
$ . . . . . . . . . . . . Superoperator on B(H)
F
. . . . . . . . . . . Least upper bound
A#b . . . . . . . . The bth element of the
ordered set A
B(H) . . . . . . . . Set of all bounded
operators on Hilbert
space H
COMM . . . . Valuation function for
parallel execution
C(X) . . . . . . . . Complex-valued functions
X →C
c(σ) . . . . . . . . . Check if a given data
type is purely classical
card(X) . . . . . Cardinality of X
⊥ . . . . . . . . . . . Least element of a partial
order
∼
= . . . . . . . . . . . Equivalence, reflexive and
transitive
⊑ . . . . . . . . . . . Binary partial order
(K∅ , T∅ , E∅ ) Initial (K, T, E) tuple
Γ . . . . . . . . . . . List of Kraus sets
Λ . . . . . . . . . . . Completely positive map
Dn . . . . . . . . . . Set of all density
operators of dimension n
S(k) . . . . . . . . Set of all decompositions
of k ∈ N
D(A, B) . . . . . Determine if A and B
operate on disjoint qbits
DO . . . . . . . . . Valuation function for
dyadic operators
E . . . . . . . . . . . Environment
χ(v) . . . . . . . . Type associated with a
variable v
ω . . . . . . . . . . . Increasing chain of
natural numbers
ϕ . . . . . . . . . . . A permutation
π : B . . . . . . . . Probability distribution
obtained by applying a
projective measurement
defined by the basis B
̺ . . . . . . . . . . . . A density operator
E ⊢ x : T . . . . x has type T is valid in
environment E
E ⊢ F . . . . . . . F is well-typed in
environment E
E(A) . . . . . . . . Effects of A
Σ . . . . . . . . . . . State in form of a
(K, T, E) tuple
σ . . . . . . . . . . . Signature for types
EQN . . . . . . . Valuation function for
arithmetic expression
EX P . . . . . . . . Valuation function for
expressions
A . . . . . . . . . . . Set of all possible Kraus
aggregations
87
Appendix A. List of symbols
F . . . . . . . . . . . Set of all fixed points of a
permutation
F2 . . . . . . . . . . Binary group/ring/field
fix . . . . . . . . . . Fixed point
in . . . . . . . . . . . Injection
I(M ) . . . . . . . Set of all intervals in M
K . . . . . . . . . . . Kraus aggregation
q(σ) . . . . . . . . . Check if a given data
type is purely quantum
qtype . . . . . . Arbitrary quantum data
type
R . . . . . . . . . . . Number of receive
statements in a Kraus
aggregation
K . . . . . . . . . . . Set of all unparametrised
Kraus agregations
Lc . . . . . . . . . . Set of labels for
communication partners
M . . . . . . . . . . Finite set
RS . . . . . . . . . Valuation function for a
receive/send pair
MO . . . . . . . . Valuation function for
monadic operators
OP . . . . . . . . . Valuation function for
operators
P(M ) . . . . . . . Powerset of M
S(A) . . . . . . . . States of A
S . . . . . . . . . . . Number of send
statements in a Kraus
aggregation
smash . . . . . . . Smash product (with a
single bottom element)
SR . . . . . . . . . Valuation function for a
send/receive pair
nc . . . . . . . . . . . Classical data type with
n bits
nq . . . . . . . . . . Quantum data type with
n qbits
pos(x, L) . . . . Position of x in the list L
Sym(M ) . . . . Symmetric group over M
T (σ) . . . . . . . . Set of data types
equivalent to σ
T . . . . . . . . . . . Typing context
PROG . . . . . . Valuation function for
programs
PROJ . . . . . Generate Kraus set with
projection operators for a
quantum type
Q . . . . . . . . . . . Size of the quantum heap
(global constant!)
Q . . . . . . . . . . . Set of all quantum
variables in a typing
context
q(nτ ) . . . . . . . Distinguish between
classical and quantum
components of a data
type
tq (σ) . . . . . . . . Number of quantum bits
contained in a data type
σ
tc (σ) . . . . . . . . Number of bits contained
in a data type σ
U (n) . . . . . . . . Unitary group of degree n
VAL . . . . . . . . Arbitrary valuation
function
X . . . . . . . . . . . Finite set
x : t . . . . . . . . . Variable x with type t
Y D . . . . . . . . . Fixed point combinator
on cpo D
88
Die Bedeutung eines Wortes ist das, was die
Erklärung der Bedeutung erklärt.
Ludwig Wittgenstein, Philosophische
Grammatik
B
Glossary
Some terms used in this work are not too commonplace in physics, so we collected the most
important definitions to remind the reader of their meaning if it cannot be immediately
recollected. Some of the definitions were inspired by [Wik05].
Abstract syntax Grammar used to specify the possible shapes of the parse
tree.
EBNF Extended Backus-Naur Form
Environment Structure which provides a
mapping between identifiers of variables and the values associated with
them.
Backus-Naur Form Metasyntax with a
standardised set of symbols and notations which is used to express contextfree grammars.
FIFO Queue with first-in, first-out behaviour, i.e., the output of the queue
is in the same order as the input.
Compiler-Compiler A program used to
generate a parser which can perform
syntax analysis on programs that follow a given grammar.
Functional languages do not work on explicit states, but use transformations
that map input to output parameters
(→ referential transparency). Assignment to variables is not possible since
they represent immutable bindings for
values. In our notation, functionality
is exploited as far as it is necessary for
the ability to guarantee freedom from
runtime errors. Classical examples of
functional languages include Lisp, ML,
and Haskell.
Compile time refers to all actions which
are performed by the compiler before
the program is executed, e.g., syntactical and semantical analysis, scoping
rule enforcement, type analysis, optimisation, code generation etc.
Concrete syntax Syntax in which textual
representations of programs must be
specified.
Identifier Name of a variable in a program.
Context free grammar Formal grammar
in which every production rule needs
to be of the form V → w where V is a
non-terminal and w a list of terminal
and non-terminal symbols.
Imperative languages work on a global
state that is modified during runtime.
The most widespread languages (C,
C++, Pascal etc.) follow this approach. It is normally impossible to
decide if a program written in an imperative language will terminate or
produce errors without executing it.
Data type A data type is a name or label
for a set of values and some operations
which can be performed on that set of
values.
89
Appendix B. Glossary
Lexer A lexer is the part of a compiler
which takes the source code of a program (in textual form) and disseminates it into a stream of tokens which
is fed to the parser.
Scope Rules used to determine what, if
any, entity a given occurrence of an
identifier in a program refers to.
Semantic analysis is the part of a compiler that adds semantic information
to the parse tree (for example, the required space in memory for variables)
and performs sanity checks which may
detect errors in the code before it is
executed.
Lexicographic order Two strings x, y ∈
Σ∗ can be ordered such that x > y
if x#i − y#i > 0 for the first i ∈ N for
which x#i 6= y#i.
LALR(1) Certain class of context-free
grammars that needs to be specified
subject to some constraints on its
form, but can be handled by Yacc-style
parsers.
Static typing means that once a type has
been assigned to an object, it cannot
be changed any more.
Strong typing means that not only values,
but also identifiers are typed.
Mutex Mutual exclusion. A technique realised with the aid of special variables
which ensures that only one component of a parallel program can be in
the region protected by the mutex at
a time.
Terminal symbol A symbol of a grammar
that represents a constant.
Token Tokens are the smallest elementary
parts of a program form the parser’s
point of view. While int is a threeletter word respectively a string of
characters in the source code, the
parser regards it as a single entity
which describes the data type of integers.
Non-terminal symbol A symbol that is
composed of terminal symbols and
possibly other non-terminal symbols.
Parser The parser is the part of a compiler
which analyses the grammatical structure of a program (which is fed to him
in the form of tokens produced by the
lexer ). This process is also known as
syntactical analysis.
Type Also called data type. It is a label for
a set of values together with some operations that can be performed on the
set.
Type system Set of rules that determines
which type a given object has, how
types can be combined etc.
Parse tree Representation of a program
which is generated by the parser. Since
tree-based data structures are used to
represent the information, the abstract
syntax of the language (which is easier
to analyse) can be utilised.
Type checking is the pass of a compiler
which ensures that all operations of
a program are applied to variables of
proper type; it can, e.g., ensure that
string concatenation is not tried to be
performed on integers.
Runtime refers to the time when a program is executed and the compiler has
no more influence on what happens.
Alternatively, it may denote a library
with helper functions supplied by the
compiler which are necessary for the
generated code to work (this may be
also referred to as runtime library).
Yacc Yet another compiler compiler. One
of the early approaches to automated
parser generation.
Most modern
parser generators follow the concept of
this program.
90
C
Der Satz ist der sprachliche Ausdruck dafür,
dass sich die Verbindung mehrerer Vorstellungen in der Seele des Sprechenden vollzogen
hat, und das Mittel dazu, die nämliche
Verbindung der nämlichen Vorstellungen in
der Seele des Hörenden zu erzeugen.
H. Paul, Prinzipien der Sprachgeschichte
Formal syntax
The formal syntax for cQPL is defined by the following rules which are used to generate the
parser. Words in typewriter face denote tokens recognised by the lexer, whereas slanted
text is used for productions. identifiers are given by a letter followed by an arbitrary number
of letters, digits and underscores. The empty production is denoted by ǫ.
program: stmt list EOF
| module list EOF
stmt list: statement;
| stmt list statement;
module list: module def ;
| module list module def ;
module def : module identifier { stmt list }
proc decl: proc identifier:context -> context block in statement
| proc identifier:context block in statement
context: identifier:var type more context | ǫ
nonempty context: identifier:var type more context
more context: , identifier:var type more context | ǫ
block: { stmt list }
var type: bit | qbit | qint | int | float
send stmt: send args to identifier
receive stmt: receive nonempty context from identifier
allocate stmt: new var type identifier := arith expr
arith expr: int value
| float value
| true
| false
| identifier
| (arith expr)
| arith expr + arith expr
| arith expr - arith expr
| arith expr * arith expr
| arith expr / arith expr
| arith expr < arith expr
| arith expr > arith expr
| arith expr <= arith expr
| arith expr >= arith expr
| arith expr == arith expr
| arith expr != arith expr
91
Appendix C. Formal syntax
| arith expr & arith expr
| arith expr | arith expr
| - arith expr
| ! arith expr
proc call: call identifier (args)
| (var list) := call identifier (args)
args: identifier more args | ǫ
more args: , identifier more args | ǫ
if stmt: if arith expr then statement
| if arith expr then statement else statement
measure stmt: measure identifier then statement else statement
assign stmt: identifier := arith expr
assign measure stmt: identifier := measure identifier
while stmt: while arith expr do statement
gate stmt: var list *= gate
gate: H | CNot | Not | Phase float value
| FT (int value)
| [[ number list ]]
number list: sign float value
| sign int value
| sign float value PLUS sign imaginary value
| sign int value PLUS sign imaginary value
| sign imaginary value
| number list, sign float value
| number list, sign int value
| number list, sign imaginary value
| number list, sign float value + sign imaginary value
| number list, sign int value + sign imaginary value
sign: - | + | ǫ
var list: identifier | var list, identifier
skip stmt: skip
print stmt: print "string"
| print arith expr
| dump var list
statement: proc call
| proc decl
| while stmt
| allocate stmt
| if stmt
| print stmt
| assign stmt
| assign measure stmt
| measure stmt
| skip stmt
| block
| gate stmt
| send stmt
| receive stmt
92
Bibliography
[AB02]
Alexander Asteroth and Christel Baier. Theoretische Informatik. Pearson
Studium, 2002.
[AC04a]
Samson Abramsky and Bob Coecke. A categorial semantics of quantum protocols. arXiv:quant-ph/0402130, pages 1–20, 2004.
[AC04b]
Samson Abramsky and Bob Coecke. A categorical semantics of quantum protocols. In LICS ’04: Proceedings of the 19th Annual IEEE Symposium on Logic
in Computer Science (LICS’04), pages 415–425, Washington, DC, USA, 2004.
IEEE Computer Society.
[AG81]
N.I. Achieser and I.M. Glasmann. Theorie der linearen Operatoren im HilbertRaum. Verlag Harri Deutsch, 1981.
[AG04]
T. Altenkirch and J. Grattage. A functional quantum programming language.
arXiv:quant-ph/0409065, 2004.
[AG05]
Thorsten Altenkirch and Jonathan Grattage. QML: Quantum data and control. Submitted for publication, Febuary 2005.
[Aha98]
Dorit Aharonov. Quantum computation. arXiv:quant-ph/9812037, 1998.
[AJ94]
Samson Abramsky and Achim Jung. Handbook for Logic in Computer Science,
volume 3, chapter Domain Theory. Clarendon Press, Oxford, 1994.
[AM05]
P. Adão and P. Mateus. A process algebra for reasoning about quantum security. In Electronic Notes in Theoretical Computer Science. Springer, 2005.
Preliminary version to be presented at 3rd International Workshop on Quantum Programming Languages.
[App04]
Andrew W. Appel. Modern Compiler Implementation in ML. Cambridge
University Press, New York, NY, USA, 2004.
[ASU86]
Alfred V. Aho, Ravi Sethi, and Jeffrey D. Ullman. Compilers: principles,
techniques, and tools. Addison-Wesley Longman Publishing Co., Inc., Boston,
MA, USA, 1986.
[Aul00]
Gennaro Auletta. Foundations and Interpretation of Quantum Mechanics.
World Scientific, 2000.
[BSC01]
S. Betelli, L. Serafini, and T. Calarcoet. Toward an architecture for quantum
programming. arXiv:cs.pl/0103009, 2001.
93
Bibliography
[BW]
Björn Butscher and Hendrik Weimer. Simulation eines Quantencomputers.
Universität Stuttgart.
[Car97]
Luca Cardelli. Type Systems, chapter 103. Handbook of computer science and
engineering. CRC Press, 1997.
[Cle99]
Richard Cleve. An introduction to quantum complexity theory. arXiv:quantph/9906111, 1999.
[Deu85]
David Deutsch. Quantum theory, the Church-Turing principle and the universal quantum computer. Proceedings of the Royal Society of London Ser. A,
A400:97–117, 1985.
[dH02]
J. den Hartog. Probabilistic extension of semantical models. PhD thesis, Vrije
Universiteit Amsterdam, 2002.
[DJ92]
David Deutsch and Richard Jozsa. Rapid solutions of problems by quantum
computation. Proceedings of the Royal Society of London, pages 553– 558,
1992.
[DS63]
Nelson Dunford and Jacob T. Schwartz. Linear Operators. Interscience Publishers, 1963.
[GA05]
Jonathan Grattage and Thorsten Altenkirch. A compiler for a functional quantum programming language. submitted for publication, January 2005.
[GBJL02]
Dick Grune, Henri E. Bal, Ceriel J. H. Jacobs, and Koen Langendoen. Modern
Compiler Design. John Wiley, 2002.
[GN04]
S.J. Gay and R. Nagarajan. Communicating quantum processes. Proceedings
of the conference for quantum programming languages, pages 91–107, 2004.
[GN05]
Simon J. Gay and Rajagopal Nagarajan. Communicating quantum processes.
In POPL ’05: Proceedings of the 32nd ACM SIGPLAN-SIGACT sysposium
on Principles of programming languages, pages 145–157, New York, NY, USA,
2005. ACM Press.
[Gro96]
Lov K. Grover. A fast quantum mechanical algorithm for database search.
Proceedings of the Twenty-Eighth Annual ACM Symposium on Theory of Computing, pages 212–219, 1996.
[Gru99]
Jozef Gruska. Quantum Computing. McGraw–Hill International, 1999.
[GS90]
C.A. Gunther and D.S. Scott. Semantic Domains, chapter 12, pages 635–674.
Elsevier Science Publishers, 1990.
[Hol82]
Alexander S. Holevo. Probabilistic and Statistical Aspects of Quantum Theory,
volume 1 of North-Holland series in statistics and probability. North-Holland,
Amsterdam, 1982. First publ. in Russian in 1980.
[JL04]
Philippe Jorrand and Marie Lalire. Toward a quantum process algebra. In
CF’04: Proceedings of the first conference on computing frontiers, pages 111–
119, New York, NY, USA, 2004. ACM Press.
[Key02]
Michael Keyl. Fundamentals of quantum information theory. arXiv:quantph/0202122, 369(5):431–548, 2002.
94
Bibliography
[KN00]
E.H. Knill and M.A. Nielsen. Encyclopedia of Mathematics, Supplement III,
chapter Theory of quantum computation. Kluwer Academic Publishers, 2000.
[Kni96]
E. Knill. Conventions for quantum pseudocode. Technical Report LAUR-962724, 1996.
[Knu98]
Donald E. Knuth. Art of Computer Programming, Volume 3: Sorting and
Searching (2nd Edition). Addison-Wesley Professional, April 1998.
[Kra83]
Karl Kraus. States, Effects and Operations. Fundamental Notions of Quantum
Theory. Academic Press, Berlin, 1983.
[Lou03]
Kenneth C. Louden. Programming Languages: Principles and Practice. Thomson, Pacific Grove, second edition, 2003.
[Löw34]
K. Löwner. Über monotone Matrixfunktionen. Mathematische Zeitschrift,
38:177–216, 1934.
[MB01]
S-C. Mu and R. S. Bird. Quantum functional programming. 2nd Asian Workshop on Programming Languages and Systems, 2001.
[Mer98]
Eugen Merzbacher. Quantum Mechanics. Wiley, John & Sons, 3 edition, 1998.
[Mos90]
Peter D. Mosses. Denotational semantics, chapter 11, pages 577–629. Elsevier
scientific publishers, 1990.
[NC00]
Michael L. Nielsen and Isaac L. Chuang. Quantum computation and quantum
information. Cambridge University Press, New York, NY, USA, 2000.
[NPW05]
Rajagopal Nagarajan, Nikolaos Papanikolaou, and David Williams. Simulating
and compiling code for the sequential quantum random access machine. In
Selinger [Sel05].
[Öme98]
Bernhard Ömer. A procedural formalism for quantum computing. Master’s
thesis, TU Vienna, 1998.
[Öme00]
Bernhard Ömer. Quantum Programming in QCL. Master’s thesis, TU Vienna,
2000.
[Öme03]
Bernhard Ömer. Structured quantum programming. PhD thesis, TU Vienna,
2003.
[Pre99]
John Preskill. Lecture notes for the course quantum computation (physics
229). www.theory.caltech.edu/people/preskill/ph229, 1999.
[RAMK+ 04] Helge Rosé, Torsten Asselmeyer-Maluga, Matthias Kolbe, Falk Niehoerster, and Andreas Schramm. The fraunhofer quantum computing portal.
arXiv:quant-ph/0406089, 2004.
[Rey98]
John C. Reynolds. Theories of programming languages. Cambridge University
Press, 1998.
[RP02]
Peter Rechenberg and Gustav Pomberger. Informatik-Handbuch. Hanser Fachbuch, 2002.
[Sak94]
Jun John Sakurai. Modern Quantum Mechanics. Addison-Wesley, 1994.
95
Bibliography
[Sch24]
Moses Schönfinkel. Über die Bausteine mathematischer Logik. Math. Ann. 92,
pages 305–316, 1924.
[Sch01]
Uwe Schöning. Theoretische Informatik - kurzgefasst. Spektrum, Akad. Verl.,
4 edition, 2001.
[Sch04]
Andreas Schroeder. Quantenflussdiagramme und die Quantenprogrammiersprache QPL. Seminar der Lehr- und Forschungseinheit für theoretische Informatik, LMU München, 2004.
[Sel04a]
Peter Selinger. A brief survey of quantum programming languages. In Lecture
Notes in Computer Science 2998. Springer, 2004.
[Sel04b]
Peter Selinger. Towards a quantum programming language. Mathematical.
Structures in Comp. Sci., 14(4):527–586, 2004.
[Sel05]
Peter Selinger. Proceedings of the 3rd international workshop on quantum
programming languages. In Peter Selinger, editor, Proceedings of the 3rd International Workshop on Quantum Programming Languages, Electronic Notes
in Theoretical Computer Science. Elsevier Science, 2005.
[SF96]
Robert Sedgewick and Philippe Flajolet. An Introduction to the Analysis of
Algorithms. Addison-Wesley, 1996.
[Sho94]
Peter W. Shor. Algorithms for quantum computation: Discrete logarithms
and factoring. IEEE Symposium on Foundations of Computer Science, pages
124–134, 1994.
[SP00]
J.W. Sanders and P.Zuliani. Quantum programming. Lecture notes in computer science, 1837, 2000.
[Sto87]
Joseph E. Stoy. Denotational semantics. MIT Press, 4 edition, 1987.
[Str00]
Christopher Strachey. Fundamental concepts in programming languages.
Higher Order Symbol. Comput., 13(1-2):11–49, 2000.
[vT04]
Andre van Tonder. A lambda calculus for quantum computation. SIAM Journal on Computing, 33:1109–1135, 2004.
[Wei00]
Joachim Weidmann. Lineare Operatoren in Hilberträumen, volume 1. B.G.
Teubner, 2000.
[WF94]
Andrew K. Wright and Matthias Felleisen. A syntactic approach to type soundness. Information and Computation, 1994.
[Wik05]
Wikipedia community.
2005.
[Win93]
Glynn Winskel. The formal semantics of programming languages: an introduction. MIT Press, Cambridge, MA, USA, 1993.
[WM95]
Reinhard Wilhelm and Dieter Maurer. Compiler Design. Addison Wesley
Longman Publishing Co., Inc., Redwood City, CA, USA, 1995.
Wikipedia online dictionary, www.wikipedia.net,
96
Thanks
❏ To PD Dr. Norbert Lütkenhaus for taking the peril of a journey into the strange and
unaccustomed, his support by discussions and suggestions and for giving me complete
freedom in deciding what to work on.
❏ To Prof. Dr. Dr. Volker Strehl for many pointers into the right direction and for
sacrificing time for a student from another faculty.
❏ To Tobias Moroder, Hans Loehr, Martin Trini, Johannes Rigas, Volkher Scholz and
Markus Diefenthaler for proofreading and many valuable corrections and suggestions.
❏ To the guys in my office (Tobi Moroder, Johannes Rigas and Dr. Matthias Jakob)
for providing a pleasant environment to work in, many inspiring level eights and our
shared pleasure of working under illumination provided by the moon.
❏ To an anonymous reviewer for encouraging comments.
❏ To dict.leo.org for countless suggestions on english vocabulary; quick answers to
many questions were given by www.wikipedia.net.
❏ To Dr. Peter Selinger for detailled explanations regarding his work.
❏ To the red and the green forrest fairy because they are way too mythical to not be
thanked; without any doubt, the same holds for HM Queen Elizabeth II.
❏ To all members of the QIT group (Tobi, Johannes, Philippe, Matthias 1, Matthias
2, Joe, Geir Ove, Marcos, Ivan) for their help, valuable discussions and the pleasant
working environment, not to forget the shared fun among some of us in chasing little
bouncing objects on diverse courts.
❏ To my parents and my family for their overall and ubiquitous love and support in any
aspect of life.
97
| 6cs.PL
|
APPROXIMATING (UNWEIGHTED) TREE AUGMENTATION VIA
LIFT-AND-PROJECT,
PART II
arXiv:1507.01309v2 [cs.DS] 29 Aug 2015
JOSEPH CHERIYAN AND ZHIHAN GAO
Abstract. In Part II, we study the unweighted Tree Augmentation Problem (TAP) via the Lasserre
(Sum of Squares) system. We prove that the integrality ratio of an SDP relaxation (the Lasserre
tightening of an LP relaxation) is ≤ 23 + , where > 0 can be any small constant. We obtain this
result by designing a polynomial-time algorithm for TAP that achieves an approximation guarantee
of ( 32 + ) relative to the SDP relaxation. The algorithm is combinatorial and does not solve the SDP
relaxation, but our analysis relies on the SDP relaxation. We generalize the combinatorial analysis of
integral solutions from the previous literature to fractional solutions by identifying some properties
of fractional solutions of the Lasserre system via the decomposition result of Karlin, Mathieu and
Nguyen (IPCO 2011).
1. Introduction
In the weighted Tree Augmentation problem we are given a connected, undirected multigraph G
bT ) of G; the goal
with non-negative costs on the edges, together with a spanning tree T = (V (G), E
bT , of minimum cost such that (V, E
bT ∪ F ) is 2-edge connected.
is to find a set of edges, F ⊆ E(G)−E
b
By a link we mean an element of E(G)−ET ; thus, a link is an edge of G that can be used to augment
bT (a tree-edge) if T + uw − ê = (V, E
bT ∪ {uw}−{ê})
T . We say that a link uw covers an edge ê ∈ E
is connected. We say that a set of links F covers the tree T if every edge of T is covered by at least
bT ∪ F ) is 2-edge-connected. Thus, the goal is
one link of F ; it can be seen that F covers T iff (V, E
to compute a set of links of minimum cost that covers T .
The weighted Tree Augmentation problem was first studied by Frederickson and Jaja in 1981
[4]. They showed that the problem is NP-hard, and they presented a 2-approximation algorithm.
Subsequently, it has been proved that even the unweighted Tree Augmentation problem is APX-hard,
see [9, Section 4]; thus the problem has no PTAS assuming P6=NP.
There have been some important advances on this problem for the corresponding unweighted
(i.e., uniform weight) problems. Following [5], we use the abbreviation TAP for the unweighted Tree
Augmentation problem. Nagamochi, see [6], presented the first algorithm for TAP that improved on
the approximation guarantee of 2; the approximation guarantee is ≈ 1.875. Subsequently, Even, et
al., [5] built on the ideas and techniques initiated by Nagamochi and presented an elegant algorithm
and analysis that achieves an approximation guarantee of 1.8. In a conference publication from
2001, Even, et al., reported a 1.5 approximation algorithm for TAP, see the extended abstract [3],
and recently, Kortsarz & Nutov [10] presented the journal version of this result.
We prove that the integrality ratio of an SDP relaxation (the Lasserre tightening of an LP relaxation) is ≤ 32 + , where > 0 can be any small constant. We obtain this result by designing
a polynomial-time algorithm for TAP that achieves an approximation guarantee of ( 32 + ) relative
to the SDP relaxation. The algorithm is combinatorial and does not solve the SDP relaxation, but
our analysis relies on the SDP relaxation. Observe that our integrality ratio (and approximation
guarantee) is proved relative to a “weaker lower bound” than the optimum value (since the feasible
Date: August 29, 2015.
Based on one chapter of the second author’s Ph.D. Thesis.
1
region of a relaxation is a superset of the convex full of integer solutions). Moreover, the approximation guarantee of 1.5 follows as a corollary of our main result (but, we cannot eliminate the “+”
in our bound on the integrality ratio by this corollary).
Linear programming relaxations for the weighted version of TAP have been studied for many
years. (We use the standard abbreviation LP to mean a linear programming relaxation or a linear
programming problem.) There is an obvious “covering” LP: we have a variable xe for each edge
bT and we have a covering constraint for each edge of T . It is well known that the
e ∈ E(G) − E
integrality ratio of this LP is ≤ 2; this can be deduced from Jain’s result [7]. A lower bound of
1.5 on the integrality ratio is known [1]; in fact, the construction for the lower bound uses uniform
bT , hence, the lower bound applies for TAP. We formulate an LP
weights for the edges in E(G) − E
that is a tightening of the obvious “covering” LP for TAP; see (LP0 ) in Section 3.
The Lasserre system applies to an initial LP, and it derives a sequence of tightenings of the initial
LP; these tightened relaxations are indexed by a number t = 0, 1, . . . called the level, where the level 0
tightening means the initial LP. A key “decomposition theorem” (see Theorem 4.1, [11, 8]) asserts
that a feasible solution at level t can be written as a convex combination of feasible solutions at a
lower level such that all of these lower-level solutions y are “locally integral.” Here, “locally integral”
bT , such that the solution y takes only zero or one
means that there is a specified subset J ⊆ E(G)− E
values on this subset (i.e., ye ∈ {0, 1}, ∀e ∈ J). A key point is that the difference in levels (between
the level t of the given feasible solution and the level of the lower-level “locally integral” solutions)
does not depend on the size of J, rather, it depends on the following “combinatorial parameter”
determined by J. Suppose that there exists a constant k such that every feasible solution x of the
initial LP has ≤ k entries in J that have value one, i.e., |{e ∈ J : xe = 1}| ≤ k for every feasible
solution x of the initial LP. Then for any t > k, a feasible solution at level t can be written as a convex
combination of “locally integral” feasible solutions at level (t − k). This property does not hold for
other weaker Lift-and-Project systems such as the Lovász-Schrijver system or the Sherali-Adams
system.
The analysis of our algorithm is based on a potential function. Our potential function is derived
from the Lasserre tightening of the initial LP.
Our algorithm is “combinatorial” and we do not need to solve the initial LP nor its Lasserre tightening to run the algorithm; but, the analysis of the algorithm relies on the Lasserre tightening. Our
algorithm essentially follows the “algorithmic scheme” of [5, Section 3.4], see Section 7 for details.
The algorithm is a greedy-type iterative algorithm that makes a leaves-to-root scan over the tree T
and (incrementally) constructs a set of links F that covers T . The algorithm starts with F := ∅, at
each major step it adds one or more links to F (it never removes links from F ), and at termination,
it outputs a set of links F that covers T such that |F | ≤ the potential function. The algorithm
incurs a cost of one unit for each link added to F . The key to the analysis is to show that for each
major step, the cost incurred (i.e., one plus the number of links added to F ) is compensated by a
part of the potential function.
Informally speaking, our analysis in Section 8 asserts the following:
if the naive algorithm gets “stuck” then the instance contains a small combinatorial
obstruction, a so-called deficient tree, see Theorems 8.7, 8.8.
This assertion is the key to this paper; it turns out that the algorithmic aspects as well as the
analysis of the approximation guarantee are straightforward consequences. Our analysis in Section
8 makes essential use of the Lasserre system and the decomposition theorem, see Figure 1.
For ease of exposition and for the sake of motivation, this paper has several discussions that
contain forward references. But, these forward references are not relevant for the correctness of our
arguments; indeed, all discussions containing forward references could be deleted, and this will have
2
Decomposition theorem
for the Lasserre system
Lemma 8.5
Lemma 7.2
Lemma 8.2
Potential function
of Lemma 5.2
Lemma 8.6
Lemma 8.3
Theorem 8.7
shows deficient trees
under appropriate conditions
Lemma 8.4
Figure 1. Illustration of the role of the decomposition theorem for the Lasserre
system (Theorem 4.1). Our analysis consists of three blocks of results: low-level,
middle-level, and high-level; these are shown from left to right in the figure. Note
that Lemmas 7.2, 8.2 are often used in the proofs of subsequent results.
no effect on the validity of our proofs. Also, Part II repeats some definitions, figures, and discussions
from Part I, but proofs from Part I are not repeated.
An outline of Part II is as follows. Section 2 has definitions and notation; several items are new
to Part II (and not needed in Part I). We adopt the notation and terms of Even, et al., [5], where
possible; this will aid readers familiar with that paper. Section 3 presents the initial LP; this is
the same LP as in Part I. Section 4 discusses the Lasserre tightening of the initial LP, and proves
some basic properties and inequalities; most of these results are from Part I, but there are two new
results. Section 5 derives our potential function, based on a solution y of the Lasserre tightening;
this potential function differs significantly from the potential function of Part I. Section 6 starts the
presentation of the algorithm (and credits) by elaborating on two preprocessing steps. Section 7
completes the discussion of the algorithm (and credits) by presenting the main loop of the algorithm.
The most important component of this paper is Section 8; this section presents the analysis of the
algorithm by first proving some low-level properties, then builds on this to prove some intermediatelevel lemmas, and then proves the key theorem on deficient trees (Theorem 8.7); this section also
presents and proves the last piece of the algorithm, namely, the handling of deficient trees.
2. Preliminaries and notation
This section presents definitions and notation.
Standard notation including tree T , link set E. Let G = (V, E(G)) be a connected, undirected
bT ) be a spanning tree of G. We assume that |V | ≥ 2. By a tree-edge
multigraph, and let T = (V, E
bT ; we call E the link set and we call an
we mean an edge of T . Let E denote the edge-set E(G)−E
element ` ∈ E a link ; thus, a link is an edge of G that can be used to augment T . An instance of
TAP consists of G and T . We assume that all instances of interest have feasible solutions, that is,
bT ∪ E) is 2-edge connected. The goal is to find a minimum-size subset F of E
we assume that (V, E
bT ∪ F ) is 2-edge
such that augmenting T by F results in a 2-edge connected multigraph, i.e., (V, E
connected.
For two nodes u, v ∈ V , we use Pu,v = Pv,u to denote the unique path of the tree T between u
and v.
For a node v ∈ V , we denote the number of tree-edges incident to v by degT (v). For any node
v ∈ V , we use δE (v) to denote the set of links incident to v.
3
For any U ⊆ V , we denote the set of links with both ends in U by E(U ), and for any two subsets
U, W of V , we denote the set of links with one end in U and the other end in W by E(U, W );
thus, E(U, W ) := {uw ∈ E : u ∈ U, w ∈ W }. We use similar notation for some subsets of
E; for example, E reg denotes a particular subset of E (defined below), and E reg (U, W ) denotes
{uw ∈ E reg : u ∈ U, w ∈ W }.
We say that a link uw covers a tree-edge ê if Pu,w 3 ê. Similarly, we say that a subtree of T is
covered by a set of links J ⊆ E if each tree-edge of the subtree is covered by some link of J. For any
bT , we use δE (ê) to denote the set of links that cover ê, thus, δE (ê) = {uw ∈ E : ê ∈
tree-edge ê ∈ E
Pu,w }.
Shadows and the shadow-closed property. For two links u1 v1 and u2 v2 , if Pu1 ,v1 ⊆ Pu2 ,v2 ,
then we call u1 v1 a shadow (or, sublink ) of u2 v2 .
For any link uv ∈ E, if all sublinks of uv also exist in E, then we call E shadow-closed. Clearly,
if E is not shadow-closed, then we can make it shadow-closed by adding all sublinks of each of the
original links. It can be seen that this preserves the optimal value of any instance of TAP.
Following Even, at al., [5], we make the next assumption (see Assumption 2.2 of [5]).
Assumption: E is shadow-closed.
Root, ancestor, descendant and rooted subtrees. One of the nodes r of T is designated as
the root; thus, we have a rooted tree (T, r).
Let v be a node of T . If a node w belongs to the path Pv,r , then w is called an ancestor of v, and
v is called a descendant of w. If a descendant w of v is adjacent to v (thus, w 6= v), then w is called
a child of v, and v is called a parent of w. Clearly, every node in V −{r} has a unique parent. If a
node v has no child, then we call v a leaf ; if v has no child, then degT (v) = 1. Note that r has at
least one child (since |V | ≥ 2), thus, r is not a leaf, even if degT (r) = 1. Throughout, we use L to
denote the set of leaves; thus, L = {v ∈ V : v is a leaf of T }.
For any leaf v of T , up(v) denotes a node q in Pv,r that is nearest to the root and adjacent to v
via a link. Since E is shadow-closed, it can be seen that up(v) is an ancestor of v, for each leaf v.
For any node v, we use Tv to denote the rooted subtree of (T, r) induced by v and its descendants.
We use L(Tv ) to denote the set of leaves of the subtree Tv . Throughout, the terms tree or subtree
refer to a rooted subtree of (T, r) that is “closed w.r.t. descendants” (if a node w is in the subtree,
then all descendants of w are in the subtree).
Property 2.1. Suppose that T̄ is a rooted tree. Let T̄v1 and T̄v2 be two (rooted) subtrees of T̄ . Then
T̄v1 and T̄v2 either share no node or one is contained in the other.
¯ etc. denote an arbitrary item/set (rather than
Remark 2.2. Throughout, symbols such as T̄ , J,
the complement of another set).
Stems and twin links. We call a node s of T a stem if s 6= r, s has exactly two children, s has
exactly two descendants that are leaves, and there exists a link in E between the two leaves of Ts ;
we call the link between the two leaves of Ts a twin link, and denote it by twinlk(s). (Our definition
of stem differs slightly from [5, Definition 3.1].) Let E twin denote the set of twin links. Observe that
there is a one-to-one correspondence between twin links and stems. Throughout, we use S to denote
the set of stems; thus, S = {v ∈ V : v is a stem of T }. Moreover, we use R to denote the set of
nodes that are neither stems nor leaves; thus R = V −(S ∪ L). (The notion of stems and twin links
is due to [6].)
up (s) = {vs ∈ E : v ∈
dn (s) =
For any stem node s, we define δE
/ V (Ts )}. Similarly, we define δE
{vs ∈ E : v ∈ V (Ts )}.
4
Contraction and compound nodes. We use the standard notion of contracting a link or a set
of links, see [5], [2, Chapter 1]. We use T 0 := T /F to denote the tree obtained by contracting
bT ∪ F ) to a single node. Each of the
each of the 2-edge connected components of T + F = (V, E
0
contracted nodes of T = T /F is called a compound node, see [5, Section 3.2]; thus, each compound
node corresponds to a set of two or more nodes of V (T ). Each of the other nodes of T /F is called
an original node.
We also call a node of T an original node. We call a link in E an original link. When we say that
` is a link w.r.t. T we mean that ` is an original link; when we say that `0 is a link w.r.t. T 0 = T /F ,
we mean that there exists an original link ` whose image in T /F is `0 .
Buds and buddy links. Besides stems and their associated subtrees, one other type of subtree
(and some of the incident links) plays an important role in this paper.
We call a leaf b0 a bud (see Figure 2) if there exists a (rooted) subtree Tv with exactly three
leaves b0 , b1 , b2 such that (i) up(b0 ) is a descendant of v (possibly, up(b0 ) = v), (ii) Tup(b0 ) contains a
stem s such that b0 , b1 are the leaves of Ts , and (iii) the link b1 b2 exists. We call b1 b2 the buddy link
of b0 and denote it by buddylk(b0 ). Observe that there exists an ancestor q of s in Tv such that
q is the least common ancestor of s and b2 ; possibly, q = v, and possibly, q = v = r; moreover,
L(Tup(b0 ) ) = {b0 , b1 } or L(Tup(b0 ) ) = {b0 , b1 , b2 }.
Note that b1 may be a bud as well. In that case, each leaf of Tup(b1 ) is in {b0 , b1 , b2 }, there exists
a link between b0 , b2 , and buddylk(b1 ) = b0 b2 .
v = up(b0 )
q
v=q
up(b0 )
s
b2
b1
s
b2
b0
b1
b0
Figure 2. Illustration of a bud b0 and the subtree Tv . The figure on the left shows
the case when up(b0 ) is an ancestor of all three leaves in Tv . The figure on the right
shows the other case.
Consider a bud b0 , and let s, q be as above; we define
Rspecial (b0 ) := V (Pb0 ,up(b0 ) )−V (Ps,q )−{b0 }.
Thus Rspecial (b0 ) contains the internal nodes in the tree-path between b0 and s; moreover, if up(b0 )
is an ancestor of all three leaves b0 , b1 , b2 , then Rspecial (b0 ) also contains all nodes on the tree-path
between the parent of q and up(b0 ); there are no other nodes in Rspecial (b0 ) (see Figure 3).
Fact 2.3. Rspecial (b0 ) is the set of nodes w such that there exists a link b0 w in E, and w is not on
the tree-path between the two ends of buddylk(b0 ). Every node in Rspecial (b0 ) has a unique child,
and is an ancestor of b0 . For two buds b, b̄ that are not the leaves of the same subtree Ts , where s
is a stem, the sets Rspecial (b) and Rspecial (b̄) are disjoint. (Thus, Rspecial (b) ∩ Rspecial (b̄) = ∅, unless
the least common ancestor (in T ) of b, b̄ is a stem s.)
We denote the set of buds and the set of buddy links by Lbud and E buddy , respectively. Observe
that there is a unique buddy link for each bud; thus, there is a bijection between Lbud and E buddy .
5
up(b0 )
q
q
up(b0 )
s
b2
b1
s
b2
b0
b1
b0
Figure 3. Illustration of Rspecial (b0 ) for a bud b0 . The nodes in Rspecial (b0 ) are
indicated by solid circles. The figure on the left shows the case when Tup(b0 ) has
three leaves. The figure on the right shows the other case.
For any node w of the tree T , we denote by Lbud (w) the set of buds in the subtree Tw . For a
stem s, note that Lbud (s) may contain zero, one, or two nodes.
Let Rspecial = ∪b∈Lbud Rspecial (b) and Rnonspcl = R−Rspecial . Thus, we partition the set R (of nodes
that are neither stems nor leafs) into two subsets, the “special” subset Rspecial and the “normal”
subset Rnonspcl .
We use E reg to denote the set E−(E twin ∪ E buddy ), namely, the set of links that are neither twin
links nor buddy links.
Vectors and convex combinations. For any vector x ∈ RE , let ones(x) denote the set of links
of x-value one, thus ones(x) = {uv ∈ E : xuv = 1}.
P
For a vector x ∈ RE and any subset J of E, x(J) denotes
e∈J xe . Given several vectors
P
i ; thus, Z is a set of indices, and
λ
v
v 1 , v 2 , . . . , we write one of their
convex
combinations
as
i∈Z i
P
we have λi ≥ 0, ∀i ∈ Z, and i∈Z λi = 1.
3. Initial LP relaxation
This section presents our initial LP relaxation of shadow-closed instances of TAP, denoted by
(LP0 ); this is the same LP as in Part I (Section 3).
Let u1 v1 and u2 v2 be a pair of links. We call it an overlapping pair of links if (i) Pu1 ,v1 , Pu2 ,v2
have one or more tree-edges in common, and (ii) either an end of u1 v1 is in Pu2 ,v2 , or an end of u2 v2
is in Pu1 ,v1 . We call a set of links J an overlapping clique if every pair of links in J is an overlapping
pair. We use the notion of overlapping pairs for stating (LP0 ).
(LP0 ) minimize :
X
xuv
uv∈E
subject to :
X
uv∈δE (ê)
bT
∀ê ∈ E
xuv ≥ 1,
xu1 v1 + xu2 v2 ≤ 1, ∀ overlapping pairs u1 v1 , u2 v2 ∈ E
0≤x≤1
4. Lasserre tightening and its properties
This section presents the Lasserre tightening of the initial LP and proves some basic properties
and inequalities; most of these results are from Part I, but there are two new results. The proofs of
the results from Part I are omitted.
6
Let Last (LP0 ) denote the level t tightening of (LP0 ) by the Lasserre system. 1 Recall that ones(x)
denotes the set of links of x-value one, where x ∈ RE is a feasible solution of (LP0 ). Rothvoß, see
[11, Theorem 2], formulated the following decomposition theorem for the Lasserre system, based
on an earlier decomposition theorem due to Karlin-Mathieu-Nguyen [8]. (We use this particular
formulation and not the original statement of [8]; hence, we reference both [8] and [11].)
Theorem 4.1. Let J ⊆ E. Let k be a positive integer such that |ones(x) ∩ J| ≤ k for every feasible
solution x of (LP0 ). Then for every P
feasible solution y ∈ Last (LP0 ), where t ≥ k + 1, y can be
written as a convex combination y = i∈Z λi xi such that xi is in Last−k (LP0 ) and xi |J is integral
(i.e., xiuv is integral for each uv ∈ J), for all i ∈ Z.
Lemma 4.2. Let J ⊆ E be an overlapping clique. For every feasible solution x of (LP0 ), we have
|ones(x) ∩ J| ≤ 1. Furthermore, for every level t ≥ 2 and every feasible solution y of Last (LP0 ), we
have y(J) ≤ 1.
Lemma 4.3. Let w be a leaf of T , and let u be an ancestor of w such that every internal node
(if any) of Pw,u has exactly one child. Let ê = vu be the tree-edge of Pw,u that is incident to u
(possibly, v = w). Then, δE (ê) is an overlapping clique. In particular, δE (w) is an overlapping
clique. Moreover, we have y(δE (w)) = 1 for any feasible solution y ∈ Last (LP0 ) where t ≥ 2.
Recall that the matching polytope of the subgraph induced by the leaves, G(L) = (L, E(L)) is
given by the following constraints:
x(δE(L) (v)) ≤ 1,
x(E(W )) ≤
∀v ∈ L
|W | − 1
, ∀W ⊆ L, |W | odd
2
x≥0
The next result is essentially the result on the matching polytope from the survey of Rothvoß,
see [12, Lemma 13, Sec 3.3], translated to our setting.
Lemma 4.4. Let > 0, and let y ∈ Las 1 +1 (LP0 ). Then,
2
G(L) = (L, E(L)).
y|E(L)
1+
is in the matching polytope of
Lemma 4.5. Let x be a feasible solution of (LP0 ) and let s be a stem. Then |ones(x) ∩ δE (s)| ≤ 3.
Proof. Since s is a stem, s is incident to three tree-edges. Let ê1 , ê2 , ê3 be the tree-edges incident with
s. Clearly, Ji = δE (êi ) ∩ δE (s) is an overlapping clique, for each 1 ≤ i ≤ 3, and δE (s) = ∪1≤i≤3 Ji .
Then the result follows from Lemma 4.2.
Lemma 4.6. Let t ≥ 3, and let y be a feasible solution of Last (LP0 ). Suppose that s is a stem.
Then:
up (s)).
(1) y(twinlk(s)) ≤ y(δE
(2) For each bud b in Ts (if it exists), we have y(buddylk(b)) ≤ y(bs) + y(E(b, Rspecial (b))).
Proof. Let the leaves of Ts be b0 and b1 . Let J = δE (b0 ) ∪ δE (b1 ). Since b0 , b1 are leaves, by
Lemma 4.3, δE (b0 ) and δE (b1 ) are overlapping cliques. Thus, by Lemma 4.2, |ones(x) ∩ J| ≤ 2
for
solution x of (LP0 ). By Theorem 4.1, y can be written as a convex combination
P any feasible
i such that xi ∈ Las (LP ) and xi | is integral, ∀i ∈ Z.
λ
x
1
0
J
i∈Z i
1Formally speaking, Las (LP ) is a set of vectors in R2[|E|] . In what follows, we abuse the notation and take
t
0
Last (LP0 ) to be the projection on the subspace indexed by the singleton sets; thus, we take Last (LP0 ) to be a set of
vectors in R|E| .
7
(1): Consider statement (1) of the lemma. Observe that twinlk(s) = b0 b1 . Since b0 b1 ∈ J, either
xi (b0 b1 ) = 0 or xi (b0 b1 ) = 1, for all i ∈ Z. Let Z(1) = {i : xi (b0 b1 ) = 1}. Then
P
y(b0 b1 ) = i∈Z(1) λi xi (b0 b1 ). Consider xi , where i ∈ Z(1) . Let ês be the tree-edge between
s and its parent. Then, xi (δE (ês )) ≥ 1. Notice that every link in δE (ês ) with positive
xi -value must have s as its end in Ts ; otherwise, if such a link has an end at some other
node of Ts , then it will be overlapping with the link b0 b1 , and we would have a violation
of the overlapping constraints of (LP0 ). Thus, xi (δE (ês )) = xi (δE (ês ) ∩ δE (s)). Hence,
up (s)) = xi (δ (ê ) ∩ δ (s)) = xi (δ (ê )) ≥ 1 = xi (b b ), ∀i ∈ Z . Consequently,
xi (δE
0 1
E s
E
E s
(1)
P
P
P
i
i
up
i
up
up (s)).
y(b0 b1 ) = i∈Z(1) λi x (b0 b1 ) ≤ i∈Z(1) λi x (δE (s)) ≤ i∈Z λi x (δE (s)) = y(δE
(2): Consider statement (2) of the lemma. W.l.o.g., let b = b0 (the same argument applies for
i
i
b = b1 ). Let buddylk(b0 ) = b1 b2 . Since b1 b2 ∈ J, either
Px (b1 b2 ) =i 0 or x (b1 b2 ) = 1, for all
i
i ∈ Z. Let Z(2) = {i : x (b1 b2 ) = 1}. Then y(b1 b2 ) = i∈Z(2) λi x (b1 b2 ).
Consider xi , where i ∈ Z(2) . Let ` = b0 w be a link such that xi (`) = 1; such a link must
exist. If w 6= s and w is in Pb1 ,b2 , then ` is overlapping with b1 b2 , and we would have a
violation of the overlapping constraints of (LP0 ). Hence, by Fact 2.3, w ∈ Rspecial (b0 ) ∪ {s}.
special (b ))).
Then, xi (buddylk(b0 )) = 1 ≤ xi (`) P
≤ xi (b0 s) + xi (E(b0 , R
0
P
i
Consequently, y(buddylk(b0 )) = i∈Z(2) λi x (b1 b2 ) ≤ i∈Z(2) λi (xi (b0 s)+xi (E(b0 , Rspecial (b0 )))) ≤
P
i
i
special (b )))) = y(b s) + y(E(b , Rspecial (b ))).
0
0
0
0
i∈Z λi (x (b0 s) + x (E(b0 , R
For any stem s and for any x ∈ RE , we define slackx (s) to be
1
X
1 dn
up
x(δE (s)) +
(x(E(b, Rspecial (b))) − x(buddylk(b))) +
x(δE
(s)) − x(twinlk(s)) .
2
2
bud
b∈L
(s)
Lemma 4.6 implies the following fact.
Fact 4.7. For any stem s and for any feasible solution y to Last (LP0 ) where t ≥ 3, we have
slacky (s) ≥ 0.
Remark 4.8. Although the above fact assumes that the level t is ≥ 3, observe that slackx (s) is
well defined for all levels ≥ 0.
5. Potential function
This section presents the potential function used by our algorithm. The potential function is based
on a feasible solution y to the Lasserre tightening of (LP0 ). It is essential to tighten (LP0 ) because
the potential function is not valid for feasible solutions of (LP0 ). Possibly, our potential function
may be obtained via weaker lift-and-project systems (e.g., Lovász-Schrijver, or Sherali-Adams). We
prefer to apply the Lasserre tightening because our analysis in Sections 7–8 makes essential use of
the decomposition theorem (Theorem 4.1), and this theorem is not known to hold for any weaker
lift-and-project system; see Part I (Section 10).
Our potential function is defined via a subset of the leaves that is denoted by Λ. This subset is
determined by the instance of TAP. Informally speaking, it consists of all the leaves of all occurrences
of a particular type of subtree, called a bad 2-stem tree, see Section 6. Our potential function consists
of two parts, a “preprocessing” part and a “normal” part. Our algorithm applies two preprocessing
steps; the first preprocessing step contracts all occurrences of maximal bad 2-stem trees, and for
this we have to “charge” the “preprocessing” part of our potential function.
We say that a set of leaves Λ is compatible if the following holds:
• For every twin link and for every buddy link, either both ends of the link are in Λ or no end
of the link is in Λ; in other words, no twin link and no buddy link is present in E(Λ, L − Λ).
8
In what follows, let Λ ⊆ L denote a compatible set.
We denote the set of stems with both leaves in Λ by SΛ . Similarly, let Lbud
denote the set of
Λ
bud
buds in Λ, and let L(L−Λ) denote the set of buds in L−Λ. By the definition of a compatible set, the
following fact holds.
Fact 5.1.
bud
Lbud
(s),
(L−Λ) = ∪s∈S −SΛ L
E twinlk (L−Λ) = {twinlk(s) : s ∈ S−SΛ },
E buddylk (L−Λ) = ∪s∈S −SΛ {buddylk(b) : b ∈ Lbud (s)}.
creg
b(L−Λ) deLet M
denote a maximum matching of the subgraph (L−Λ, E reg (L−Λ)), and let U
(L−Λ)
creg ; thus U
b(L−Λ) = (L−Λ)−{v ∈
note the set of nodes of this subgraph exposed by the matching M
(L−Λ)
creg }.
V : v is an end of some link ` ∈ M
(L−Λ)
We mention that our potential function (the right-hand side of the inequality in Lemma 5.2 below)
creg , U
b(L−Λ) . Thus, when we use our potential function, we have to ensure
refers to the terms Λ, M
(L−Λ)
that these terms have been defined already; we will “fix” our potential function by appropriately
creg , U
b(L−Λ) in Section 6.
defining Λ, M
(L−Λ)
Given Λ and y ∈ RE , we use lbd y (Λ) to denote the quantity
1
1 X
3
y(E(Λ)) + y(E(Λ, L−Λ)) + y(E(Λ, V −L)) +
y(δE (s));
2
2
2
s∈SΛ
this is one of the terms in our potential function; informally speaking, this is the main component
of the “preprocessing” part of the potential function.
1
+ 1, 3}. Let y ∈ Last (LP0 ). Then,
Lemma 5.2. Let > 0 be a constant, and let t ≥ max{ 2
3
( + )y(E)
≥
2
1
+ y(E(V −L)) +
2
3 creg
b(L−Λ) | + lbdy (Λ)
|M
| + |U
2 (L−Λ)
1 X
1 X
y(δE (v)) +
y(E(v, V −Lbud
(L−Λ) (v))) +
2
2
nonspcl
special
v∈R
v∈R
Moreover, every item on the right-hand side of the inequality is nonnegative.
X
slacky (s).
s∈S −SΛ
Proof. Observe that y is feasible for Last (LP0 ) for t ≥ 3, hence, y ≥ 0 and (by Fact 4.7) slacky (s) ≥
0, ∀s ∈ S. Hence, every item on the right-hand side of the inequality is nonnegative.
For each link uw, we distribute the value 32 yuw as follows:
– if both u, w ∈ L, then uw keeps the value 32 yuw ;
– if both u, w ∈ V −L, then uw keeps the value 21 yuw and each of the ends u and w gets value
1
2 yuw ;
– if only one of u or w is in V −L, then uw keeps the value yuw and the end in V −L gets value
1
2 yuw .
(In other words, each node in V −L borrows value 21 yuw from each link uw incident to it, and the
link keeps the remaining value; links uw that are not incident to V −L keep all of the value 32 yuw .)
9
Thus, we have
3
3
3
3
y(E) =
y(E(Λ)) + y(E(L−Λ)) + y(E(Λ, L−Λ))
2
2
2
2
1
1 X
+ y(E(V −L)) + y(E(V −L, Λ)) + y(E(V −L, L−Λ)) +
y(δE (v)).
2
2
v∈V −L
Then we increase the coefficients of the twin links and buddy links in E(L − Λ) from 32 to 2 by
borrowing the value 21 y(E twin (L−Λ)) + 12 y(E buddy (L−Λ)) from the last term above, and adding it to
the term 32 y(E(L−Λ)). Thus, we replace the term 32 y(E(L−Λ)) by 2y(E(L−Λ)) − 12 y(E reg (L−Λ)),
and we replace the last term by
1 X
1
1
y(δE (v)) − y(E twin (L−Λ)) − y(E buddy (L−Λ))
2
2
2
v∈V −L
1 X
1 X
1 X
1 X
dn
up
y(δE (v)) +
y(δE (v)) +
y(δE (s)) +
=
(y(δE
(s)) + y(δE
(s)))
2
2
2
2
nonspcl
special
s∈SΛ
s∈S −SΛ
v∈R
v∈R
1
1
− y(E twin (L−Λ)) − y(E buddy (L−Λ))
2
2
1 X
1 X
1 X
1 X
1
up
y(δE (v)) +
y(δE (v)) +
=
y(δE (s)) +
y(δE
(s)) − y(E twin (L−Λ))
2
2
2
2
2
s∈SΛ
s∈S −SΛ
v∈Rnonspcl
v∈Rspecial
1
1 X
dn
y(δE
(s)) − y(E buddy (L−Λ))
+
2
2
s∈S −SΛ
1 X
X
1 X
(1) 1 X
special
(v)))
+
y(E(b,
R
(b)))
=
y(δE (s)) +
y(δE (v)) +
y(E(v, V −Lbud
(L−Λ)
2
2
2
s∈SΛ
v∈Rnonspcl
v∈Rspecial
b∈Lbud
(L−Λ)
X
X
1
1
1
1
up
dn
+
y(δE
(s)) − y(E twin (L−Λ)) +
y(δE
(s)) − y(E buddy (L−Λ))
2
2
2
2
s∈S −SΛ
s∈S −SΛ
1 X
1 X
(2) 1 X
=
y(δE (s)) +
y(δE (v)) +
y(E(v, V −Lbud
(L−Λ) (v)))
2
2
2
s∈SΛ
v∈Rnonspcl
v∈Rspecial
X
1 X dn
y(δE (s)) +
+
(y(E(b, Rspecial (b))) − y(buddylk(b)))
2
s∈S −SΛ
b∈Lbud (s)
1 X
up
+
y(δE
(s)) − y(twinlk(s))
2
s∈S −SΛ
X
1 X
1 X
1 X
=
y(δE (s)) +
y(δE (v)) +
y(E(v, V −Lbud
(v)))
+
slacky (s),
(L−Λ)
2
2
2
nonspcl
special
s∈SΛ
s∈S −SΛ
v∈R
v∈R
where (2) follows from Fact 5.1 due to the assumption that Λ is compatible, and (1) follows from
the following equation:
[
[
E(w, Lbud
=
E(b, Rspecial (b)).
(L−Λ) (w))
w∈Rspecial
b∈Lbud
(L
−Λ)
This equation is based on some observations. First, the set on the right-hand side is clearly a subset
of the set on the left-hand side. Conversely, consider a link wa in the set on the left-hand side,
where w ∈ Rspecial and a ∈ Lbud
(w). Since w is an ancestor of a and w is not on the tree-path of
(L−Λ)
10
the link buddylk(a), by Fact 2.3, we have w is in Rspecial (a), i.e., wa ∈ E(a, Rspecial (a)). Hence, the
link wa belongs to the set on the right-hand side as well.
Thus, the expression for 23 y(E) can be written as
3
3
1
= y(E(Λ)) + y(E(Λ, L−Λ)) + 2y(E(L−Λ)) − y(E reg (L−Λ))
2
2
2
1
+ y(E(V −L)) + y(E(V −L, Λ)) + y(E(V −L, L−Λ))
2
X
1 X
1 X
1 X
y(δE (v)) +
y(E(v, V −Lbud
(v)))
+
+
y(δE (s)) +
slacky (s)
(L−Λ)
2
2
2
nonspcl
special
s∈SΛ
s∈S −SΛ
v∈R
v∈R
3
1
1 X
=
y(E(Λ)) + y(E(Λ, L−Λ)) + y(E(Λ, V −L)) +
y(δE (s))
2
2
2
s∈SΛ
+ y(E(L−Λ, Λ)) + 2y(E(L−Λ)) + y(E(L−Λ, V −L))
X
1
1 X
1 X
slacky (s)
+ y(E(V −L)) +
y(δE (v)) +
y(E(v, V −Lbud
(v))) +
(L
−
Λ)
2
2
2
nonspcl
special
s∈S −SΛ
v∈R
v∈R
1
− y(E reg (L−Λ))
2
X
1
= lbd y (Λ) +
y(δE (v)) + y(E(V −L))
2
v∈L−Λ
X
1 X
1 X
(v))) +
slacky (s)
+
y(δE (v)) +
y(E(v, V −Lbud
(L
−
Λ)
2
2
nonspcl
special
s∈S −SΛ
v∈R
v∈R
1
− y(E reg (L−Λ))
2
P
By Lemma 4.3, we have y(δE (v)) = 1 for any v ∈ L. Hence, we replace the term v∈L−Λ y(δE (v))
(the 2nd term in the displayed equation above) by |L−Λ|. Moreover, by Lemma 4.4, we have
y| reg
y|E(L)
is in the matching polytope of (L, E(L)). Thus, E (L−Λ) is in the matching polytope of
1+
1+
creg |. We derive the inequality (stated
(L−Λ, E reg (L−Λ)), which implies y(E reg (L−Λ)) ≤ (1+)|M
(L−Λ)
in the lemma) by replacing the term − 21 y(E reg (L−Λ)) (the last term in the displayed equation above)
creg |. Now, observe that |L−Λ| − 1 (1 + )|M
creg | = |U
b(L−Λ) | + 3 |M
creg | −
by − 12 (1 + )|M
2
2
(L−Λ)
(L−Λ)
(L−Λ)
reg
reg
1
c
b
c
2 ()|M(L−Λ) |, because U(L−Λ) = (L−Λ)−{v ∈ V : v is an end of some link ` ∈ M(L−Λ) }. Note
creg
creg | ≤ 1 |L| = 1 P
that |M
v∈L y(δE (v)) ≤ y(E), hence, − 2 |M(L−Λ) | ≥ −y(E). Thus, we get our
2
2
(L−Λ)
potential function:
3
( + )y(E)
2
≥
3 creg
b(L−Λ) | + lbd y (Λ) + 1 y(E(V −L))
|M(L−Λ) | + |U
2
2
1 X
1 X
+
y(δE (v)) +
y(E(v, V −Lbud
(L−Λ) (v))) +
2
2
nonspcl
special
v∈R
v∈R
6. Algorithm and credits I: Preprocessing steps
We state our main result for (unweighted) TAP:
11
X
s∈S −SΛ
slacky (s).
1
+ 1}. Let y denote an
Theorem 6.1. Let > 0 be any (small) constant, and let t ≥ max{17, 2
3
optimal solution of Last (LP0 ). The integrality ratio of Last (LP0 ) is ≤ 2 + . Moreover, there is a
polynomial-time algorithm for TAP that finds a feasible solution (a set of links that covers the tree
T ) of size ≤ ( 23 + )y(E).
For the rest of this paper, we let y ∈ RE denote an optimal solution of Last (LP0 ), where t ≥
1
max{17, 1 + 2
}, where > 0 is any constant. Our goal is to show that our algorithm finds a set of
links that covers T of size ≤ ( 32 + )y(E). We achieve this goal using our potential function (this is
the right-hand side of the inequality in Lemma 5.2); we will “fix” the potential function below by
creg , U
b(L−Λ) ). Recall from Part I (Section 6) that the
defining the relevant terms (namely, Λ, M
(L−Λ)
potential function provides credit to the algorithm.
Also, recall from Part I (Section 6) that the combinatorial algorithm is a greedy-type iterative
algorithm that makes a leaves-to-root scan over the tree T and (incrementally) constructs a set of
links F that covers T . The algorithm starts with F := ∅, at each step it adds one or more links to
F (it never removes links from F ), and at termination, it outputs a cover F of T whose size is ≤
the potential function.
The algorithm incurs a cost of one unit for every link that it picks, and it incurs a cost of one unit
for each new compound node that it creates in the execution. The key to the analysis is to show
that for each step, the cost incurred is compensated by a part of the credit.
We mention that (with one exception, see Lemma 8.3) the nodes or links that get contracted into
a compound node are no longer relevant for the algorithm or the analysis. In particular, the credit
(if any) of such nodes or links may be used at the step when they get contracted into a compound
node, but after that step, any remaining credit of such nodes or links is not used at all.
0
denote the path of T 0 between u
For the current tree T 0 = T /F and nodes u, w of T 0 , let Pu,w
and w. By a fitting cover of a rooted tree Tv0 ⊆ T 0 we mean a set of links J that covers all of the
0
= Tv0 .
tree-edges of Tv0 but does not cover any other tree-edge; thus, we have ∪uw∈J Pu,w
6.1. Semiclosed trees. We recall the notion of a semiclosed tree w.r.t. an arbitrary matching. This
notion is due to Even, et al., based on earlier work by Nagamochi [6]; also, see [5, Definition 2.3].
Let Tv0 be a rooted subtree of the current tree T 0 = T /F . Let M̄ be an arbitrary matching of the
leaf-to-leaf links. Tv0 is called semiclosed w.r.t. M̄ if the following conditions hold:
(i) Each link in M̄ either has both ends in Tv0 or has no end in Tv0 .
(ii) Every link incident to an M̄ -exposed leaf of Tv0 has both ends in Tv0 .
Let M̄ (Tv0 ) denote the set of links in M̄ that have both ends in Tv0 .
We define
[
Γ(M̄ , Tv0 ) := M̄ (Tv0 ) {up(w)w : w is an M̄ -exposed leaf in Tv0 };
thus, we associate a “basic link set” with the pair M̄ , Tv0 . In general, the “basic link set” may not
be a cover of Tv0 .
By a minimally semiclosed tree Tv0 we mean that Tv0 is semiclosed but none of the proper rooted
subtrees of Tv0 is semiclosed.
Lemma 6.2 (Even, et al., [5]). Let Tv0 be a minimally semiclosed tree w.r.t M̄ . Then Γ(M̄ , Tv0 ) is
a fitting cover of Tv0 .
6.2. Maximum matching. Our algorithm and analysis are based on a maximum matching of the
leaf-to-leaf links that are neither twin links nor buddy links. Let M denote one such matching; thus,
M is a maximum matching of the subgraph (L, E reg (L)). By an M -link we mean a link that is in
M . We denote the set of M -exposed leaf nodes by U .
12
The image of M w.r.t. T 0 is {u0 w0 : uw ∈ M, u0 6= w0 }, where u0 , w0 denote the images (w.r.t.
of the original nodes u, w of T (i.e., we have u0 = u if u is an original node of T 0 , otherwise, u0
denotes the compound node of T 0 that contains u; w0 is defined similarly). We abuse the notation
and use M to denote both M and its image w.r.t. T 0 , and by an M -link of T 0 we mean the image
(w.r.t. T 0 ) of an original M -link.
For the rest of the paper, unless mentioned otherwise, a semiclosed tree means a subtree that is
semiclosed w.r.t. the matching M .
T 0)
6.3. Bad 2-stem trees. Let Tv be a semiclosed tree rooted at v that has exactly two stems s1 , s2
and four leaves, where we denote the leaves of the tree Tsi by ui , wi for i = 1, 2.
By a leafy 3-cover of Tv we mean a set of three links J such that J covers Tv , one of the links in
J has one end in Tv and one end in L−L(Tv ), and the other two links in J have both ends in Tv .
We call Tv a bad 2-stem tree if (i) one of the links in E({u1 , w1 }, {u2 , w2 }) is in M , (ii) two of
the leaves are M -exposed, (iii) one of the leaves is incident to all the links in E({u1 , w1 }, {u2 , w2 })
(thus E({u1 , w1 }, {u2 , w2 }) has one or two links), (iv) there exists a cover of Tv of size three, and
(v) there exists no leafy 3-cover of Tv . Let us fix the notation such that w1 w2 is the unique M -link
in E(L(Tv )); thus, u1 and u2 are M -exposed (see Figure 4).
By a maximal bad 2-stem tree Tv we mean a bad 2-stem tree that is not a proper subtree of
another bad 2-stem tree. (Thus, any tree Tq rooted at a proper ancestor q of v (if q exists) must
violate one of the conditions for being a bad 2-stem tree.)
v
s1
u1
s2
w1
w2
u2
Figure 4. Illustration of a bad 2-stem tree. The dashed lines indicate links and the
thick dashed line indicates an M -link.
Let F prep = {Tv1 , . . . , Tvk } denote the set of maximal bad 2-stem trees
S of T ; clearly, any two
treesSin F prep are disjoint, by Property 2.1. We use V (F prep ) to denote Tv ∈F prep V (Tv ). We define
Λ = Tv ∈F prep L(Tv ), that is, Λ consists of all the leaves of all the trees in F prep .
Since each bad 2-stem tree has a cover of size 3, the shadow-closed property implies that each
bad 2-stem tree has a fitting cover of size ≤ 3. Our algorithm applies a preprocessing step that
contracts each tree Tv ∈ F prep by a fitting cover of size ≤ 3.
Preprocessing step 1 (Λ-contraction): For every tree Tv ∈ F prep , add a fitting cover of Tv of
size ≤ 3 to F and contract Tv to a compound node.
The cost incurred for this step is ≤ 4|F prep |, since each tree in F prep incurs a cost of ≤ 3 for its
fitting cover and a cost of 1 for the resulting compound node.
This cost is charged to one part of our potential function, namely, it is charged to
X
1
y(δE (v)).
lbd y (Λ) +
2
nonspcl
prep
∩V (F
v∈R
Lemma 6.6 below shows that this quantity is ≥ 4|F prep |.
13
)
Remark 6.3. The results in this section show that Λ-contraction is valid, in the sense that the
algorithm has sufficient credits to pay for the cost of this step. Moreover, Λ-contraction is essential for the overall analysis (i.e., the algorithm cannot skip Λ-contraction), because the proof of
Theorem 8.7 (Subcase 2.2) relies on Λ-contraction.
Lemma 6.4. Let F prep and Λ be as defined above. Then, we have
lbdy (Λ) +
≥
X
Tv ∈F prep
1
2
X
y(δE (v))
v∈Rnonspcl ∩V
3
(F prep )
1
1
y(E(L(Tv ))) + y(E(L(Tv ), L−L(Tv ))) + y(E(L(Tv ), V −L)) +
2
2
2
X
u∈V (Tv )−L
y(δE (u))
P
Proof. Recall that lbd y (Λ) denotes 32 y(E(Λ))+ 21 y(E(Λ, L−Λ))+y(E(Λ, V −L))+ 21 s∈SΛ y(δE (s)).
Observe that a bad 2-stem tree has no buds, so every node of such a tree is either a leaf, or a
stem, or a node of Rnonspcl . Hence,
(Rnonspcl ∩ V (F prep )) ∪ SΛ =
[
(V (Tv )−L)
Tv ∈F prep
and so we have
1
2
X
v∈Rnonspcl ∩V
y(δE (v)) +
1 X
y(δE (s)) =
2
s∈SΛ
(F prep )
X
Tv ∈F prep
1
2
X
y(δE (u)).
u∈V (Tv )−L
We need the following claim to prove the lemma.
Claim 6.5. Let Λ1 , Λ2 , . . . , Λk be a partition of Λ. Then, we have
3
1
y(E(Λ))+ y(E(Λ, L−Λ)) + y(E(Λ, V −L))
2
2
k
k
k
X
3X
1X
≥
y(E(Λi )) +
y(E(Λi , L−Λi )) +
y(E(Λi , V −L)).
2
2
i=1
i=1
i=1
We partition Λ into the sets Λi = L(Tvi ), where Tv1 , . . . , Tvk are the maximal bad 2-stem trees in
F prep . Then, we have
lbd y (Λ) +
1
2
X
v∈Rnonspcl ∩V
y(δE (v))
(F prep )
X 1
X
3
1
= y(E(Λ)) + y(E(Λ, L−Λ)) + y(E(Λ, V −L)) +
y(δE (u))
2
2
2
prep
Tv ∈F
u∈V (Tv )−L
X 3
X
1
1
y(E(L(Tv ))) + y(E(L(Tv ), L−L(Tv ))) + y(E(L(Tv ), V −L)) +
≥
2
2
2
prep
Tv ∈F
14
u∈V (Tv )−L
y(δE (u)) .
To prove the claim, observe that
3
1
y(E(Λ)) + y(E(Λ, L−Λ)) + y(E(Λ, V −L))
2
2
k
k
k
k
3 X
1X
X
1 XX
≥
y(E(Λi , Λi )) +
y(E(Λi , Λj )) +
y(E(Λi , L−Λ)) +
y(E(Λi , V −L))
2
2
2
i=1
=
i=1 j6=i
i=1
k
k
k
i=1
i=1
i=1
i=1
X
1X
3X
y(E(Λi )) +
y(E(Λi , L−Λi )) +
y(E(Λi , V −L)).
2
2
Lemma 6.6. Let Tv be a bad 2-stem tree. Then we have
1
1
3
y E(L(Tv )) + y E(L(Tv ), L−L(Tv )) + y E(L(Tv ), V −L) +
2
2
2
Hence, we have
lbdy (Λ) +
1
2
X
v∈Rnonspcl ∩V (F prep )
X
u∈V (Tv )−L
y(δE (u)) ≥ 4.
y(δE (v)) ≥ 4|F prep |.
Proof. The second statement follows immediately from the first statement and Lemma 6.4. We
focus on the first statement.
Let s1 , s2 denote the two stems in Tv , and let u1 , w1 (respectively, u2 , w2 ) denote the two leaves
in Ts1 (respectively, Ts2 ). Thus, L(Tv ) = {u1 , w1 , u2 , w2 }. W.l.o.g. let w1 w2 denote the unique
M -link in E(L(Tv )); the two twin links u1 w1 , u2 w2 are also in E(L(Tv )), and E(L(Tv )) may contain
one other link incident to w1 or w2 ; there is no link between u1 and u2 (see Figure 4). Note that
every link incident to u1 or u2 (the M -exposed leaves of Tv ) must have both ends in Tv , since Tv is
semiclosed w.r.t. M .
S
Let J denote the set of links incident to the leaves of Tv , thus, J = i=1,2 δE (ui ) ∪ δE (wi ) . For
any feasible solution x of (LP0 ), Lemmas 4.3
Pimpliesithat |ones(x)i ∩ J| ≤ 4. Hence, byi Theorem 4.1,
y can be written as a convex combination i∈Z λi x such that x ∈ Las3 (LP0 ) and x |J is integral,
∀i ∈ Z. (In what follows, we have to apply Lemma 4.6 to xi , hence, xi must be feasible for level 3
(or higher) of the Lasserre tightening.)
Thus, it suffices to show that for any i ∈ Z, we have
1
1
X
3 i
x E(L(Tv )) + xi E(L(Tv ), L−L(Tv )) + xi E(L(Tv ), V −L) +
xi (δE (u)) ≥ 4;
2
2
2
u∈V (Tv )−L
let α denote the left-hand side of the above inequality.
Observe that every link in J with positive xi -value must have xi -value one. Suppose that one of
the links ` ∈ E(L(Tv )) has positive xi -value, thus xi (`) = 1; then, we have 32 xi (E(L(Tv ))) ≥ 32 , thus `
contributes value 23 to α. Also, if one of the links ` with one end in L(Tv ) and the other end at a nonP
leaf node of Tv has positive xi -value, then we have xi (E(L(Tv ), V −L))+ 21 u∈(V (Tv )−L) xi (δE (u)) ≥
1 + 21 = 23 , thus, ` contributes value 32 to α.
We complete the proof by case analysis, by considering the number of links with positive xi -value
such that one end is in L(Tv ) and the other end is not in Tv .
Case 1.: Suppose that there are no links with positive xi -value such that one end is in L(Tv ) and
the other end is not in Tv . Then, we focus on the number of links in E(L(Tv )) that have
positive xi -value; this number is zero, one, or two, because every link incident to a leaf
15
has xi -value zero or one, and moreover, xi (δE (u)) = 1 for each leaf u of Tv , by (LP0 ) and
Lemma 4.3. Thus, we have three subcases.
Subcase 1.1.: xi (E(L(Tv ))) = 0. Every link of xi -value one incident to a leaf of Tv has its
other end at a non-leaf node of Tv , hence, each such link contributes 23 to α; we have
four such links, so α ≥ 6.
Subcase 1.2.: xi (E(L(Tv ))) = 1. The link ` in E(L(Tv )) with xi -value one contributes 23
to α. The two leaves in L(Tv ) that are not incident to ` each contribute 32 to α, because
each is incident to a link of xi -value one that has its other end at a non-leaf node of Tv ,
hence, α ≥ 29 .
Subcase 1.3.: xi (E(L(Tv ))) = 2. Since xi (δE (u)) = 1 for each leaf u of Tv , the links in
E(L(Tv )) with xi -value one share no ends. By the definition of bad 2-stem trees, all the
links in E({u1 , w1 }, {u2 , w2 }) are incident to one of the four leaves. It follows that the
twin links u1 w1 and u2 w2 both have xi -value one, and each contributes 23 to α. Next,
up (s )) ≥ xi (u w ) = 1 for j = 1, 2. Thus,
focus on the stems in Tv ; by Lemma 4.6, xi (δE
j
P j j
each stem contributes at least 12 to α via the term 21 u∈(V (Tv )−L) xi (δE (u)). Hence,
α ≥ 2( 32 ) + 2( 12 ) = 4.
Case 2.: Suppose that there is at least one link with positive xi -value such that one end is in L(Tv )
and the other end is not in Tv . Again, we focus on the number of links in E(L(Tv )) that
have positive xi -value; it can be seen that this number is zero or one. (Note that every link
incident to a leaf has xi -value zero or one, and moreover, for every leaf u of Tv , we have
xi (δE (u)) = 1.) Thus, we have two subcases.
Subcase 2.1.: xi (E(L(Tv ))) = 0. Every link of xi -value one incident to u1 or u2 (the M exposed leaves of Tv ) has its other end at a non-leaf node of Tv , hence, each such link
contributes 23 to α; we have two such links. Every link of xi -value one incident to w1
or w2 (the M -covered leaves of Tv ) has its other end either in V −L, and so contributes
1 to xi (E(L(Tv ), V −L)), or else has its other end in L−L(Tv ), and so contributes 21 to
1 i
3
1
2 x (E(L(Tv ), L−L(Tv ))); again, we have two such links. Hence, α ≥ 2( 2 ) + 2( 2 ) = 4.
i
i
Subcase 2.2.: x (E(L(Tv ))) = 1. Let `v denote a link of positive x -value such that one
end is in L(Tv ) and the other end, denoted by q, is not in Tv . Note that xi (`v ) = 1.
The end of `v in L(Tv ) must be M -covered; w.l.o.g. assume that this node is w1 .
Note that xi (δE (u)) = 1 for each leaf u of Tv . Since Tv is a bad 2-stem tree, the
links in E(L(Tv )) that are not incident to w1 are the twin link u2 w2 and possibly the
link u1 w2 (there is no link between u1 , u2 ). Thus, either xi (u2 w2 ) = 1 or xi (u1 w2 ) = 1,
and this contributes 32 to α via the term 23 xi (E(L(Tv ))). In either case, one of the
M -exposed leaves u1 or u2 is incident to a link of xi -value one that has its other end at
a non-leaf node of Tv , and this contributes 32 to α via the term xi (E(L(Tv ), V −L)) +
1P
i
u∈(V (Tv )−L) x (δE (u)).
2
We have two subcases, depending on whether q (the end of `v in V −V (Tv )) is a leaf
or not.
Subcase 2.2.1.: q ∈
/ L. Then, the link `v = w1 q contributes 1 to α via the term
xi (E(L(Tv ), V −L)). Then, we have α ≥ 23 + 32 + 1 = 4.
Subcase 2.2.2.: q ∈ L. Then, the link `v = w1 q contributes 21 to α via the term
1 i
2 x (E(L(Tv ), L−L(Tv ))).
Moreover, observe that we cannot have xi (u1 w2 ) = 1, otherwise the three links
`v = w1 q, u1 w2 , u2 w2 form a leafy 3-cover of Tv (see Figure 5(a)); recall that a
bad 2-stem tree cannot have a leafy 3-cover.
16
(a)
q
v
s1
u1
w1
(b)
s2
w2
q
v
s1
u2
u1
w1
s2
w2
u2
Figure 5. Illustration of leafy 3-covers of Tv for Subcase 2.2.2 in the proof of Lemma
6.6. The dashed lines indicate links and the thick dashed lines indicate M -links.
Thus, u2 w2 is the unique link in E(L(Tv )) with xi -value one. Now, focus on the
up (s )) ≥ xi (u w ) = 1. Thus, δ up (s ) contributes at
stem s2 ; by Lemma 4.6, xi (δE
2
2
2 2
E
1
1P
least 2 to α via the term 2 u∈(V (Tv )−L) xi (δE (u)).
Finally, note that u1 s2 is not present; otherwise, we would have a leafy 3cover consisting of the (hypothetical) link u1 s2 and the links `v = w1 q, u2 w2 (see
up (s ).
Figure 5(b)). Hence, the link of xi -value one incident with u1 is not in δE
2
3
1
1
3
Thus, we have α ≥ 2 + 2 + 2 + 2 = 4.
This completes the case analysis, and completes the proof.
6.4. Credit assignment for the algorithm and the preprocessing. Recall that the algorithm starts with a number of credits equal to the potential function, namely, the right-hand side
of the inequality in Lemma 5.2. In order to specify the potential function, we need to specify
creg , U
b(L−Λ) . We take M
creg
M
to be M ∩ E(L−Λ), i.e., the restriction of M to the subgraph
(L−Λ)
(L−Λ)
(L−Λ, E reg (L−Λ)). It can be seen that M ∩ E(L−Λ) = M −E(Λ) is a maximum matching of the
creg
b(L−Λ)
subgraph (L−Λ, E reg (L−Λ)), as required by the definition of M
in Section 5. We take U
(L−Λ)
b(L−Λ) in Section 5.
to be U ∩ (L−Λ); again, this agrees with the definition of U
The cost of Λ-contraction (Preprocessing step 1) is
≤ lbd y (Λ) +
1
2
X
u∈Rnonspcl ∩V
y(δE (u)),
(F prep )
where F prep is defined above. We subtract this quantity from our potential function, and then add
the credit of the compound nodes formed by contracting the trees Tv ∈ F prep , to get the remaining
potential function (i.e., total credits available) for the rest of the execution. Clearly, the remaining
potential function is
3
|M ∩ E(L−Λ)| + |U ∩ (L−Λ)| + |F prep |+
2
X
X
1
1
1 X
y(E(V −L)) +
y(δE (u)) +
y(E(u, V −Lbud
(u))) +
slacky (s)
(L
−
Λ)
2
2
2
nonspcl
prep
special
s∈S −SΛ
u∈(R
−V (F ))
u∈R
17
By the integral potential function we mean the sum of the first three terms above (namely,
∩ E(L−Λ)| + |U ∩ (L−Λ)| + |F prep |), and by the fractional potential function we mean the sum
of the remaining terms (namely, the sum of the four terms that use y).
The following observation is useful for simplifying our notation.
3
2 |M
Fact 6.7. For the current tree T 0 (after Λ-contraction), U ∩ (L−Λ) is the same as the set of M exposed original leaves, and M ∩ E(L−Λ) = M (T 0 ).
We start with the credit given by the integral potential function, and we maintain the following
assignment of credits to the nodes of T 0 and the links of M (T 0 ) = M ∩ E(L−Λ):
•
•
•
•
every M -exposed original leaf has one credit,
every compound node has one credit,
every link of M (T 0 ) has 23 credit, and
the root r has one credit.
It can be seen that the integral potential function suffices for assigning credits to the tree T 0 that
results from Λ-contraction. (See Part I (Section 6) for a discussion on the unit credit for the root
r.)
We define the integral credit S
of a set of links J (w.r.t. T 0 ) to be the sum of the credits of the
0 ), plus the sum of the (integral) credits of the nodes
0
M -links
pq such that V (Pp,q ) ⊆ uw∈J V (Pu,w
S
S
0 .
0
in uw∈J V (Pu,w ), plus one if r occurs as an original node in uw∈J Pu,w
In otherSwords, the integral credit of J is the sum of 23 times theSnumber of M -links pq such that
0
0
0 )⊆
the number
V (Pp,q
uw∈J V (Pu,w ), plusS
uw∈J V (Pu,w ), plus the number of compound nodes in
0 .
of M -exposed original leaves in the same set, plus one if r occurs as an original node in uw∈J Pu,w
Informally speaking, when a step of the algorithm contracts all the links in J, then this amount of
integral credit is available for this step (but not for subsequent steps).
Now, consider the fractional potential function. We use it to maintain an assignment of fractional credits to the (rooted) subtrees of T 0 . For any subtree Tv0 of T 0 , observe that V (Tv0 ) ∩
(Rnonspcl −V (F prep )) = V (Tv0 ) ∩ Rnonspcl , because none of the original nodes in V (F prep ) is present
in T 0 after the Λ-contraction; similarly, we have V (Tv0 ) ∩ (S−SΛ ) = V (Tv0 ) ∩ S, because none of the
stems in SΛ is present in T 0 after the Λ-contraction. For any subtree Tv0 of T 0 and for any vector
x ∈ RE , we use Φ(x, Tv0 ) to denote
1
x(E((V −L) ∩ V (Tv0 ), (V −L) ∩ (V (T 0 )−V (Tv0 ))))+
2
X
X
1
1
x(δE (u)) +
x(E(u, V −Lbud
(L−Λ) (u))) +
2
2
0
nonspcl
0
special
u∈V (Tv )∩R
u∈V (Tv )∩R
X
s∈V
slackx (s).
(Tv0 )∩S
The first term is defined on links that have both ends at original non-leaf nodes of T 0 , and
moreover, have exactly one end in Tv0 . Similarly, the last three terms denote credits that are assigned
to original nodes in Tv0 . We mention that the credits given by these four terms cannot be used two
or more times by the algorithm. This holds because all credit (integral or fractional) associated
with an original node of T 0 can be used by a step that first contracts the original node to a new
compound node; after that, none of the credit associated with this original node is available (any
original node contained in a compound node is essentially “invisible” to the algorithm).
Informally speaking, Φ(y, Tv0 ) denotes the fractional credit of Tv0 , that is, the part of the fractional
potential function that is “owned” by Tv0 . The fractional credit of Tv0 will be used together with its
integral credit for contracting Tv0 .
18
6.5. Second preprocessing step. We apply a second preprocessing step after the Λ-contraction
and before the main loop of the algorithm. It turns out that the set of nodes contracted by this step
is disjoint from the set of nodes contracted by the Λ-contraction.
Let b0 be a bud, let b1 b2 = buddylk(b0 ), and let s denote the stem such that b0 is a leaf of Ts ;
w.l.o.g. let b1 be the other leaf of Ts . Moreover, if b1 is also a bud, then we assume w.l.o.g. that
up(b0 ) is an ancestor of up(b1 ).
Preprocessing step 2: For each bud b0 and the two ends b1 , b2 of buddylk(b0 ) such that all three
of b0 , b1 , b2 are M -exposed, we contract the two links up(b0 )b0 and b1 b2 .
Fact 6.8. Preprocessing step 2 has sufficient credits.
Proof. Consider one “triple” consisting of a bud b0 and the two ends b1 , b2 of buddylk(b0 ) such that
b0 , b1 , b2 are M -exposed. Note that each of the M -exposed leaves b0 , b1 , b2 has 1 unit of credit. Thus,
we have sufficient credit for contracting two links and forming one compound node.
Remark 6.9. The results in this section show that Preprocessing step 2 is valid, in the sense that
the algorithm has sufficient credits to pay for the cost of this step. Moreover, Preprocessing step 2
is essential for the overall analysis (i.e., the algorithm cannot skip this step), because the proof of
Lemma 8.3 relies on this step.
7. Algorithm and credits II: Overall algorithm
We present pseudo-code for the overall algorithm, after discussing some preliminaries. Also, we
state and prove several assertions, i.e., we prove some basic properties maintained by the algorithm.
These assertions are critical for the analysis in the next section.
7.1. (Up-to-5) greedy contractions and assertions on M . Recall that
credit of a
Sthe integral
0 ), plus the
0 )⊆
V
(P
set of links J is given by the credits of the MS-links pq such that V (Pp,q
u,w
uw∈J
0 ), plus one if r occurs as an original node
sum
of
the
(integral)
credits
of
the
nodes
in
V
(P
u,w
uw∈J
S
0 .
in uw∈J Pu,w
We define an (up-to-5) greedy contraction to be a contraction of a set of links J such that
(i) |J| ≤ 5;
S
0
forms a connected
(ii) the contraction of J results in a single compound node, i.e., uw∈J Pu,w
graph;
(iii) the integral credit of J is ≥ |J| + 1.
Clearly, if an (up-to-5) greedy contraction is applicable, then, in polynomial time, the algorithm
can find a set of links J that is eligible for this step (by examining all link sets of size ≤ 5).
Remark 7.1. Clearly, the algorithm has sufficient credit for every application of (up-to-5) greedy
contraction. The overall analysis relies on repeatedly applying (up-to-5) greedy contractions until
no such contractions are applicable. In fact, every lemma/theorem in Section 8 assumes that no
(up-to-5) greedy contractions are applicable.
The following assertion is similar to Lemma 6.3 of Part I.
Lemma 7.2 (Assertions on M ). Suppose that no (up-to-5) greedy contractions are applicable. Then:
0
(1) For every M -link uw, every node in Pu,w
is an original node. In particular, w.r.t. T 0 , both
ends of each M -link are original leaf nodes.
(2) There exist no links between M -exposed leaves of T 0 .
19
7.2. Good semiclosed trees. Let Tv0 be a (rooted) subtree of T 0 . Let U (Tv0 ) denote the set of
M -exposed leaves of Tv0 (including both original nodes and compound nodes). Let C(Tv0 ) denote the
set of compound non-leaf nodes of Tv0 . Recall thatSM (Tv0 ) denotes the set of M -links that have both
ends in Tv0 . Note that every node in (V (Tv0 ) ∩ R) (V (Tv0 ) ∩ S) is an original node.
Recall that a semiclosed tree means a tree that is semiclosed w.r.t. the matching M , unless
mentioned otherwise. After the preprocessing steps, whenever we mention semiclosed trees, we
assume that no (up-to-5) greedy contractions are applicable (see Section 7.1). Then, Lemma 7.2(1)
implies that M is a set of leaf-to-leaf links w.r.t. the current tree T 0 . Hence, semiclosed trees (of T 0
w.r.t. M (T 0 )) are well defined.
We define the credit of a (rooted) subtree Tv0 of the current tree T 0 to be the sum of the fractional
credit of Tv0 , namely, Φ(y, Tv0 ), and the integral credit of Tv0 . The latter is given by the sum of the
following terms: 23 |M (Tv0 )|, the number of compound nodes in Tv0 , the number of M -exposed original
leaves in Tv0 , and an additional one if the root r is in V (Tv0 ) ∩ R. We call a semiclosed tree Tv0 good
if its credit is ≥ |Γ(M, Tv0 )| + 1.
The next result can be proved using arguments similar to the arguments used for proving Lemma 6.4
of Part I (although the potential function and credits in Parts I and II are different).
Lemma 7.3. Let Tv0 be a semiclosed tree. If at least one of the following conditions is satisfied, then
Tv0 is good.
•
•
•
•
•
Tv0 = T 0
C(Tv0 ) 6= ∅
|M (Tv0 )| ≥ 2
Φ(y, Tv0 ) ≥ 1
|M (Tv0 )| = 1 and Φ(y, Tv0 ) ≥ 21 .
7.3. Summary of the algorithm. We start with F := ∅ (F is the set of links picked by the
algorithm) and T 0 := T (T 0 is the current tree T /F ).
apply Preprocessing step 1 (Λ-contraction);
apply Preprocessing step 2;
while T 0 is not a single node do
repeatedly apply (up-to-5) greedy contractions until no such contractions are applicable;
find a good semiclosed tree Tv0 with a fitting cover J of size |Γ(M, Tv0 )| (Algorithm 2 in
Section 8 presents the details for finding such a semiclosed tree);
add J to F , contract Tv0 to a new compound node, update T 0 ;
end
Algorithm 1: Find an approximately optimal solution for TAP.
7.4. Stem assertion of the algorithm. This section presents a basic assertion that is important
for our analysis.
Recall that the algorithm iteratively contracts a set of links such that the tree-paths associated
with these links form a connected graph; the set of chosen links and their associated tree-paths are
0
contracted to form a new compound node. When we say that a contraction hits
S a node v0 of T , we
mean that a step of the execution contracts a set of links J such that v is in uw∈J V (Pu,w ); thus,
at least one of the tree-edges incident with v is covered by one of the links in J.
Stem assertion: Let s be a stem. The first contraction that hits a node of Ts (during the execution
of the algorithm) must hit s.
Property 7.4. The algorithm maintains the stem assertion.
20
Proof. Let s be a stem. First, consider the two preprocessing steps. Either all of the tree Ts is
contracted or none of the tree-edges of Ts is contracted by Preprocessing step 1 (Λ-contraction).
The same statement holds for Preprocessing step 2. Hence, the stem assertion is maintained by the
two preprocessing steps.
Next, suppose that the first contraction (in the execution) that hits a node in Ts is an (up-to5) greedy contraction that contracts a set of links J. If one of the links uw ∈ J has one end
0
in Ts and the other end not in Ts , then the tree path Pu,w
(in the current tree w.r.t. the first
contraction that hits a node of Ts ) must contain s, hence, the stem assertion is maintained. The
remaining possibility is that all links of J have both ends in Ts . Then observe that the number of
integral credits available in Ts is either zero, one, or two, and, in the last case, both leaves of Ts are
M -exposed. (By definition of stem, the root r is not in Ts , since r 6= s.) The greedy contraction of
J requires |J| + 1 integral credits. Thus, |J| = 1 and J contains the twin link of s. Contracting the
twin link clearly maintains the stem assertion.
Finally, suppose that the first contraction (in the execution) that hits a node in Ts is the contraction of a good semiclosed tree Tv0 . By Property 2.1, one of Ts , Tv0 is contained in the other. It is
easily seen that no proper subtree of Ts is a semiclosed tree. (Note that there is a sublink between
each leaf of Ts and s via the twin link of s, hence, any semiclosed tree containing a leaf of Ts that
is M -exposed will contain s too. Also, any semiclosed tree containing a leaf of Ts that is M -covered
will contain s too.) The only remaining possibility is that Tv0 contains Ts ; then, the contraction of
Tv0 maintains the stem assertion.
8. Analysis of the algorithm and deficient trees
This section has our main results. Informally speaking, the key result (Theorem 8.7) asserts the
following: if a semiclosed tree Tv0 is not good, then either Tv0 is a deficient tree (defined below) or Tv0
is a particular type of tree that is easily bypassed by our analysis.
The analysis consists of two parts. In Section 8.2, using local integrality of feasible solutions to the
Lasserre system, we show that all semiclosed trees are good, except for a few cases. The nontrivial
cases give deficient trees. Section 8.3 shows how to handle deficient trees. This leads to an efficient
algorithm for finding a good semiclosed tree together with a fitting cover of appropriate size.
Deficient 3-leaf tree: Suppose that Tv0 is a semiclosed tree with exactly three leaves a, b1 , b2 .
Clearly, among the nodes w of Tv0 either there is exactly one node with degT 0 (w) = 4 or there are
two nodes with degree 3 in T 0 . In the latter case, we denote these two nodes by u and q; moreover,
we fix the notation such that u is an ancestor of q, and the leaf b1 is not a descendant of q; thus,
a, b2 (but not b1 ) are descendants of q. In the former case, we denote by u the unique node that
is incident to four tree-edges. We call Tv0 a deficient 3-leaf tree (see Figure 6) if (i) the link b1 b2 is
present and it is in M (T 0 ), (ii) the link ab1 is present, and (iii) there exists a link b2 w such that
w ∈ V (T 0 )−V (Tv0 ).
Moreover, in the first case (with a unique node u in Tv0 with degT 0 (u) = 4), if conditions (i)–(iii)
hold with both labelings (b1 , b2 ) and (b2 , b1 ) of the M -link, then we fix the notation such that up(b2 )
is an ancestor of up(b1 ). For both cases, we call b2 the ceiling leaf of Tv0 .
Deficient 4-leaf tree: Let Tv0 be a semi-closed tree with 4 leaves and exactly one stem node s
such that all nodes in Ts0 are original, |M (Tv0 )| = 1, and the link in M (Tv0 ) has exactly one end in Ts0 .
Let the four leaves of Tv0 be a, b1 , b2 , c, let the two leaves of Ts0 be a, b1 , and let the link in M (Tv0 ) be
b1 b2 . Note that c is the M -exposed leaf of Tv0 that is not in Ts0 . Let p be the least common ancestor
of s and c. Then, Tv0 is called a deficient 4-leaf tree (see Figure 7) if (i) Tp0 contains all leaves of Tv0 ,
(ii) the link cs is present, and (iii) there exists a link b2 w such that w ∈ V (T 0 )−V (Tv0 ). We call the
link cs the latch of Tv0 .
21
v
(a)
v
(b)
u
u
b2
q
b1
a
b1
a
b2
Figure 6. Illustration of deficient 3-leaf tree. The dashed lines indicate links and
the thick dashed lines indicate M -links.
The contraction of the latch cs in a deficient 4-leaf tree results in a deficient 3-leaf tree due to the
presence of the links ab1 , up(b2 )b2 (see Figure 6(a)). Let b be the ceiling leaf of the resulting tree.
Clearly, up(b) is an ancestor of up(b2 ).
v
p
v
c
q
q=p
u
s
b2
s
b2
b1
a
c
b1
a
Figure 7. Illustration of deficient 4-leaf trees. The dashed lines indicate links and
the thick dashed lines indicate M -links.
8.1. Properties from assertions. We start with an observation on compound leaf nodes. The
next result (Lemma 8.2) states some properties pertaining to stem nodes and semiclosed trees; these
properties are often used in the analysis of the algorithm. The proof relies on the stem assertion.
After that we present a key lemma that pertains to some buds (Lemma 8.3); see the discussion
preceding that lemma.
Fact 8.1. If a compound leaf node contains a node u ∈ V (T ), then that compound node contains
Tu .
Lemma 8.2. Suppose that no (up-to-5) greedy contractions are applicable. Let Tv0 be a semiclosed
tree with C(Tv0 ) = ∅.
(1) If Tv0 contains a stem node s, i.e., s ∈ S ∩ V (Tv0 ), then every node in Ts0 is original.
(2) In the original tree T , suppose that s is a stem node, and w is a leaf of Ts . If w is contained
in a compound node hci that is an M -exposed leaf of Tv0 , then all nodes of Ts are contained
in hci.
(3) In the original tree T , suppose that s is a stem node, and w is a leaf of Ts . If w is an original
M -exposed leaf of Tv0 , then s is an original node of Tv0 .
22
Proof. The first part follows from the stem assertion.
Consider the second part. By the stem assertion, and the fact that w is contained in hci, it can
be seen that s is contained in some compound node. If s is contained in hci, then the proof is done
by Fact 8.1. Now, suppose that s is contained in a different compound node hai. Then, there exists
a link between hci and hai (w.r.t. T 0 ), because the link ws is present in E (the input) since ws
is a sublink of twinlk(s). It can be seen that Tv0 contains hai as a leaf, because Tv0 is semiclosed,
C(Tv0 ) = ∅, hci is an M -exposed leaf of Tv0 , and the link between hci and hai is present. This gives a
contradiction because the compound leaf node hai contains s so it contains Ts (by Fact 8.1), hence,
hci cannot contain w.
The third part follows from arguments similar to that used for the second part; we give a sketch.
Suppose that s is contained in a compound node hai. If hai is a leaf of T 0 , then hai would contain
the subtree Ts (by Fact 8.1), and this would contradict the fact that the leaf w is an original node.
Since C(Tv0 ) = ∅, hai cannot be a non-leaf node of Tv0 . Thus, Tv0 contains w, but it does not contain
hai. This contradicts the fact that Tv0 is semiclosed, because there is a link between the M -exposed
leaf w of Tv0 and hai (due to the sublink ws of twinlk(s)).
The next lemma addresses a subtle point in our analysis. Consider a semiclosed tree Tv0 and
its fractional credit Φ(y, Tv0 ), and observe that the links between a bud b in Tv0 and a node in
V (Tv0 ) ∩ Rspecial (b) do not contribute to the fractional credit; see the third term in the definition of
Φ(y, Tv0 ) (the last displayed equation in Section 6.4). Nevertheless, under appropriate conditions,
we can exhibit fractional credits even when such links are present (see Lemma 8.5 and its proof).
In order to do this, we have to focus on buds contained in M -exposed compound leaves such that
there exists a link between the bud and a node “outside” the compound leaf (this is one part of
the analysis that examines an original node even after it has been contracted into a compound
node). The next lemma gives a tool for “counting” such buds, and the lemma is used twice in the
subsequent analysis. The first use is in the proof of Lemma 8.4 to show that a semiclosed tree Tv0
with |M (Tv0 )| ≤ 1 has O(1) such buds, and the second use is in the proof of Lemma 8.6 to show that
a semiclosed tree Tv0 with M (Tv0 ) = ∅ has no such buds.
Lemma 8.3. Suppose that no (up-to-5) greedy contractions are applicable. Let Tv0 be a semiclosed
tree with C(Tv0 ) = ∅. Let b0 be a bud that is contained in an M -exposed compound leaf hci of Tv0
such that there exists a link ` incident to b0 that has exactly one end (namely, b0 ) in hci. Let
buddylk(b0 ) = b1 b2 , and let b1 b0 be the twin link (w.r.t. T ) incident to b0 . Then, b2 is an original
leaf node of Tv0 and b2 is incident to a link of M (Tv0 ).
Proof. Let the original ends of ` be b0 , w; thus, w is not contained in hci. Let s be the stem associated
with b0 ; note that the leaves of Ts are b0 , b1 . By Lemma 8.2(2), Ts is contained in hci. Since b0 is
contained in an M -exposed leaf hci of Tv0 , C(Tv0 ) = ∅, and Tv0 is semiclosed, either up(b0 ) is contained
in hci, or up(b0 ) is an original non-leaf node of Tv0 . Clearly, hci cannot contain up(b0 ), otherwise, hci
would contain Tup(b0 ) (by Fact 8.1), hence, the link b0 w would be contained in hci. Thus, up(b0 ) is
an original non-leaf node of Tv0 .
Consider the node b2 . Suppose that b2 is an original leaf node of T 0 . Then, b2 is in Tv0 (since Tv0 is
semiclosed), and b2 is M -covered (by Lemma 7.2(2) and the existence of the link b1 b2 ). Moreover,
both ends of the M -link incident to b2 are in Tv0 (since Tv0 is semiclosed). Thus, the lemma holds if
b2 is an original leaf node of T 0 .
Now, we may assume that b2 is contained in a compound node hai. The hypotheses of the
lemma imply that hai is a leaf in Tv0 . We must have hai = hci. Otherwise, either hai is M -covered,
contradicting Lemma 7.2(1), or hai is M -exposed, contradicting Lemma 7.2(2) because the link b1 b2
implies a link between the M -exposed leaves hai, hci. We conclude that hci contains all three of
b0 , b1 , b2 . In what follows, we use this to derive a contradiction to the fact that up(b0 ) is an original
non-leaf node in Tv0 .
23
By Preprocessing step 2, one of b0 , b1 , b2 is M -covered w.r.t. T , otherwise, the preprocessing step
contracts two links to form a compound node that contains up(b0 ). We have two cases, depending
on the M -links incident to b0 , b1 , b2 w.r.t. T .
Case 1: There exists an M -link w.r.t. T that has one end b ∈ {b0 , b1 , b2 } and the other end u ∈
L−{b0 , b1 , b2 }. If hci does not contain u, then we get a contradiction to Lemma 7.2(1) (M link incident to compound node). Otherwise, hci contains u, hence, hci contains Pb,u ; note
that up(b0 ) is in Pb,u , so hci contains up(b0 ). This is a contradiction.
Case 2: Each M -link w.r.t. T incident to one of b0 , b1 , b2 has both ends in {b0 , b1 , b2 }. There is only
one such M -link, namely, b0 b2 , since the twin link b0 b1 and the buddy link b1 b2 cannot be
in M . Then b0 is an M -covered bud w.r.t. T (see Figure 8).
up(b0 )
q
s
b2
b1
b0
Figure 8. Illustration of an M -covered bud w.r.t. T .
We may assume that b0 b2 ∈ M and b1 is M -exposed. Suppose that up(b1 ) is a descendent
of up(b0 ); then, b1 is a bud; this follows from the definition of bud (see Section 2). Then, we
get a contradiction since b0 b2 ∈ M and b0 b2 = buddylk(b1 ) (M has no buddy links). Hence,
up(b1 ) is a proper ancestor of up(b0 ). Thus, up(b0 ) 6= r, and no (rooted) subtree of Tup(b0 )
is semiclosed since b1 is M -exposed.
The following claim gives a contradiction and completes the proof of this lemma.
Claim. Suppose that b0 b2 is an M -link and b1 is M -exposed. Then, the first
contraction that hits a node of Tup(b0 ) (during the execution of the algorithm)
must hit up(b0 ).
First, consider the two preprocessing steps. Preprocessing step 1 contracts all bad 2-stem
trees. Note that no leaf of a bad 2-stem tree is a bud, hence, by Property 2.1, Tup(b0 ) is
disjoint from every bad 2-stem tree. Preprocessing step 2 applies when a bud b and the two
ends of buddylk(b) are all M -exposed. Since b0 is M -covered, Preprocessing step 2 does not
apply to b0 , b1 , b2 . Thus, the contractions in the preprocessing steps do not hit any node of
Tup(b0 )
Next, suppose that the first contraction (in the execution) that hits a node in Tup(b0 ) is
an (up-to-5) greedy contraction that contracts a set of links J. If one of the links uw ∈ J
has one end in Tup(b0 ) and the other end not in Tup(b0 ) , then the contraction must hit up(b0 ),
0
because the tree path Pu,w
(in the current tree at that contraction) contains up(b0 ). Thus,
the claim holds. The remaining possibility is that all links of J have both ends in Tup(b0 ) .
But observe that the number of integral credits available in Tup(b0 ) (in the current tree at
that contraction) is exactly 25 (due to one M -link and one M -exposed node), and this does
not suffice for contracting two or more links; moreover, no single-link contraction applies
because no singleton set J has integral credit > 32 .
24
Finally, suppose that the first contraction (in the execution) that hits a node in Tup(b0 ) is
the contraction of a good semiclosed tree Tw0 . By Property 2.1, one of Tup(b0 ) , Tw0 is contained
in the other. Since no (rooted) subtree of Tup(b0 ) is semiclosed, it follows that Tw0 contains
Tup(b0 ) . Then, the contraction of Tw0 hits up(b0 ). This proves the claim, and completes the
proof of the lemma.
8.2. Most semiclosed trees are good. Let Tv0 be a semiclosed tree. Let Lmatched (Tv0 ) denote the
set of M -covered leaves of Tv0 .
We call a compound node hci of Tv0 open if it is an M -exposed leaf of Tv0 , and moreover, hci
contains a bud b0 such that there exists an original link between b0 and an original node that is
not contained in hci. We call such a bud an open bud (it must be contained in an open compound
node and one of the original links incident to the bud is not contained in the compound node). Let
B comp (Tv0 ) denote the set of open buds of Tv0 . Let B orig (Tv0 ) denote the set of original nodes of Tv0
that are M -exposed buds. Let B(Tv0 ) = B orig (Tv0 ) ∪ B comp (Tv0 ).
Lemma 8.4. Suppose that no (up-to-5) greedy contractions are applicable. Let Tv0 be a semiclosed
tree with C(Tv0 ) = ∅, |M (Tv0 )| ≤ 1 and |S ∩ V (Tv0 )| ≤P2.
Then, y can be written as a convex combination i∈Z λi xi such that xi ∈ Las3 (LP0 ) and xi |J is
integral, where J = {` ∈ δE (u) : u ∈ (S ∩ V (Tv0 )) ∪ Lmatched (Tv0 ) ∪ B(Tv0 )}.
Proof. We claim that |ones(x) ∩ J| ≤ 14 for any feasible solution x of (LP0 ). The conclusion (of the
lemma) follows easily from Theorem 4.1 and this claim.
To prove the claim, observe that |S ∩ V (Tv0 )| ≤ 2 and |Lmatched (Tv0 )| = 2|M (Tv0 )| ≤ 2, hence, by
Lemmas 4.3, 4.5, we have
|ones(x) ∩ {` ∈ δE (u) : u ∈ S ∩ V (Tv0 )}| ≤ 6,
|ones(x) ∩ {` ∈ δE (u) : u ∈ Lmatched (Tv0 )}| ≤ 2;
moreover, for each u ∈ B(Tv0 ), we have |ones(x) ∩ δE (u)| ≤ 1. Thus, to complete the proof of the
claim, we have to show that |B(Tv0 )| ≤ 6.
First, consider any bud b0 ∈ B orig (Tv0 ) and its associated stem s. By Lemma 8.2(3), s is an original
node of Tv0 , and so, s ∈ S ∩ V (Tv0 ). By Lemma 8.2(1), each leaf of Ts0 is an original node in T 0 . Let
b1 be the leaf of Ts0 other than b0 . By Lemma 7.2(2), b1 is M -covered (otherwise, the twin link b0 b1
connects two M -exposed leaves). Hence, b1 ∈
/ B orig (Tv0 ). It follows that each stem in S ∩ V (Tv0 ) has
at most one leaf in B orig (Tv0 ). Since |S ∩ V (Tv0 )| ≤ 2, we have |B orig (Tv0 )| ≤ 2.
Now, consider one of the buds b0 ∈ B comp (Tv0 ); b0 is contained in some open compound node and
one of the original links incident to b0 has its other end “outside” this compound node. Then, by
Lemma 8.3, the buddy link buddylk(b0 ) shares an end with an M -link in M (Tv0 ). Since |M (Tv0 )| ≤ 1
and each end of an M -link can be incident to ≤ 2 buddy links, we have |B comp (Tv0 )| ≤ 4. Therefore,
|B(Tv0 )| = |B orig (Tv0 )| + |B comp (Tv0 )| ≤ 6. This proves the claim.
Lemma 8.5. Suppose that no (up-to-5) greedy contractions are applicable. Let Tv0 be a semiclosed
tree with C(Tv0 ) = ∅. Let J be a set of links that each have at least one end in Tv0 and no end in
Lmatched (Tv0 ) ∪ (S ∩ V (Tv0 )). Let x be a feasible solution of Las3 (LP0 ) such that x is integral on the
links {δE (u) : u ∈ B(Tv0 )}. Then,
1 1
Φ(x, Tv0 ) ≥ min{ , x(J)}.
2 2
P
1P
0
Proof. Let g(x, Tv ) denote 2 u∈Rnonspcl ∩V (Tv0 ) x(δE (u))+ 12 u∈Rspecial ∩V (Tv0 ) x(E(u, V −Lbud (u))). By
P
Fact 4.7, we have Φ(x, Tv0 ) ≥ g(x, Tv0 ) + s∈V (Tv0 )∩S slackx (s) ≥ g(x, Tv0 ). We will show that either
each link ` ∈ J contributes 12 x(`) to g(x, Tv0 ), thereby ensuring g(x, Tv0 ) ≥ 21 x(J), or there exists a
set of links that contribute 21 to g(x, Tv0 ).
25
Consider any link ` ∈ J with x(`) > 0. By Lemma 7.2(2), ` cannot have both ends at M -exposed
leaves. Thus, ` has at least one end in V (Tv0 ) ∩ R, since it has no end in Lmatched (Tv0 ) ∪ (S ∩ V (Tv0 )),
Tv0 is semiclosed, and C(Tv0 ) = ∅. It is easily seen that, except for one case, ` contributes ≥ 12 x(`) to
g(x, Tv0 ).
The exceptional case occurs when ` has original ends b0 and w, where b0 ∈ Lbud (w) and w ∈
special
R
(b0 ) ∩ V (Tv0 ), see Fact 2.3. Since J has no links incident to Lmatched (Tv0 ), it follows that
b0 ∈ B(Tv0 ) because either b0 is an M -exposed original leaf of Tv0 , or b0 is contained in an M -exposed
compound leaf of Tv0 ; in the latter case, b0 is an open bud (w.r.t. Tv0 ). Then, we have x(`) = 1
because x is integral on the links incident with nodes in B(Tv0 ). We claim that g(x, Tv0 ) ≥ 21 . Let ê
denote the tree-edge between w (the end of ` in V (Tv0 ) ∩ R) and its parent. Let δ + (ê) denote the
set of links with positive x-value that cover ê. Clearly, x(δ + (ê)) ≥ 1. Consider any link `q = pq
that is in δ + (ê), where q is a descendent of w (possibly, q = w) and p is not in Tw0 . If q is a
proper descendant of w, then (by the definition of Rspecial (b0 )) q is a descendant of the unique
child of w (see Figure 3), hence, ` and `q form an overlapping pair such that x(`) + x(`q ) > 1;
this contradicts the overlapping constraints of (LP0 ). Hence, we have q = w. Clearly, p (the other
end of `q ) cannot belong to Lbud (q) = Lbud (w) since p is not a descendant of q. Hence, we have
g(x, Tv0 ) ≥ 21 x(E(w, V −Lbud (w))) ≥ 21 x(δ + (ê)) ≥ 21 . This completes the proof.
Let Tv0 be a semiclosed tree. We construct an auxiliary graph in order to analyze the credits
available in Tv0 . We denote the auxiliary graph by AG(Tv0 ). This is a bipartite multigraph, and the
two sets in the node bipartition are denoted by AS(Tv0 ) and AU(Tv0 ). The first set consists of the
M -covered leaves Lmatched (Tv0 ) and the stems S ∩ V (Tv0 ). The second set contains an auxiliary node
v̄ (informally speaking, v̄ represents the node set V (T 0 )−V (Tv0 )), as well as all the M -exposed leaves
of Tv0 , thus, AU(Tv0 ) = {v̄} ∪ U (Tv0 ).
We define the edge set of AG(Tv0 ) as follows: for every link pq w.r.t. T 0 with p ∈ (S ∩ V (Tv0 )) ∪
matched
L
(Tv0 ), q ∈ U (Tv0 ), the edge pq is in AG(Tv0 ), and for every link pq w.r.t. T 0 such that p ∈
(S∩V (Tv0 ))∪Lmatched (Tv0 ), q ∈ V (T 0 )−V (Tv0 ), the edge pv̄ is in AG(Tv0 ). Thus, AG(Tv0 ) is a multigraph
(multiple copies of an edge may be present), and every edge in AG(Tv0 ) corresponds to a link w.r.t.
T 0 (see Figure 9).
AG(Tv0 )
Tv0
v
c
v̄
b
a
s
d
d
s
c
b
a
Figure 9. The left figure shows a semiclosed tree Tv0 where dashed lines indicate
links and the thick dashed line indicates an M -link. The right figure shows the
auxiliary graph AG(Tv0 ) and its node bipartition {s, b, c}, {v̄, a, d}.
In what follows, we may abuse the notation by not distinguishing between edges (sets of edges)
of AG(Tv0 ) and the corresponding links (sets of links) w.r.t. T 0 .
26
Lemma 8.6. Suppose that no (up-to-5) greedy contractions are applicable. Let Tv0 be a semiclosed
tree such that Tv0 6= T 0 , C(Tv0 ) = ∅, |M (Tv0 )| ≤ 1 and |S ∩ V (Tv0 )| ≤ 2.
(1) Suppose that M (Tv0 ) = ∅. Then for any feasible solution x for Las3 (LP0 ), we have Φ(x, Tv0 ) ≥
1. Furthermore, Tv0 is good.
(2) Suppose that |M (Tv0 )| = 1, and |U (Tv0 )| ≥ |S ∩ V (Tv0 )| + 1. Moreover, suppose that x is a
feasible solution for Las3 (LP0 ) such that x|J is integral and Φ(x, Tv0 ) < 21 , where J = {` ∈
δE (u) : u ∈ (S ∩ V (Tv0 )) ∪ Lmatched (Tv0 ) ∪ B(Tv0 )}.
Then, |U (Tv0 )| = |S ∩ V (Tv0 )| + 1. Moreover, the auxiliary graph has a perfect matching
AM(Tv0 ) such that the following conditions hold:
(i) x(`) = 1 for each link ` ∈ AM(Tv0 ),
(ii) the links of AM(Tv0 ) cover Tv0 ,
(iii) for each stem node s ∈ S ∩ V (Tv0 ), twinlk(s) is in AM(Tv0 ),
(iv) AM(Tv0 ) has no links of the form v̄s, where s ∈ S ∩ V (Tv0 ).
0
0
Proof. Let êv denote the tree-edge between v and its parent;
P êv is well defined since Tv 0 6= T .
¯
¯
Let J = δE (êv ) ∪ (∪u∈U (Tv0 ) δE (u)). Then, x(J) = x(δE (êv )) + u∈U (Tv0 ) x(δE (u)) ≥ 1 + |U (Tv )|; the
equation holds because (i) Tv0 is semiclosed so none of the links in δE (êv ) is incident to an M -exposed
leaf of Tv0 , and (ii) by Lemma 7.2(2), no link has both ends at M -exposed leaves; the inequality holds
because x(δE (ê)) ≥ 1 forP
every tree-edge ê.
P
Let g(x, Tv0 ) denote 12 u∈Rnonspcl ∩V (Tv0 ) x(δE (u)) + 12 u∈Rspecial ∩V (Tv0 ) x(E(u, V −Lbud (u))).
P
By Fact 4.7, we have Φ(x, Tv0 ) ≥ g(x, Tv0 ) + s∈V (Tv0 )∩S slackx (s) ≥ g(x, Tv0 ).
Case (1): M (Tv0 ) = ∅. Clearly, |U (Tv0 )| ≥ 1, since Tv0 has at least one leaf.
¯ ≥ 1 (1 + |U (Tv0 )|) ≥ 1. We will show that every link ` ∈ J¯ contributes
Observe that 21 x(J)
2
≥ 12 x(`) to g(x, Tv0 ).
First, we show that S∩V (Tv0 ) = ∅. Otherwise, consider any s ∈ S∩V (Tv0 ); by Lemma 8.2(1),
every node in Ts0 is original, and so the twin link of s has both ends at M -exposed nodes
(since M (Tv0 ) = ∅); this contradicts Lemma 7.2(2).
Next, we show that B(Tv0 ) = ∅. Since S∩V (Tv0 ) = ∅, Lemma 8.2(3) implies that B orig (Tv0 ) =
∅ (if a bud b0 ∈ B orig (Tv0 ) is present as an original M -exposed leaf of Tv0 , then its stem is
present as an original node of Tv0 ). Also, we have B comp (Tv0 ) = ∅, otherwise, by Lemma 8.3,
there exists a link in M (Tv0 ) (this is impossible since M (Tv0 ) = ∅). Thus, B(Tv0 ) = ∅.
Now, observe that every link ` ∈ J¯ has an end in V (Tv0 ) ∩ R, and hence, it contributes
1
0
0
2 x(`) to g(x, Tv ). Therefore, Φ(x, Tv ) ≥ 1. Since this inequality holds for every feasible
solution x of Las3 (LP0 ), we have Φ(y, Tv0 ) ≥ 1. Hence, by Lemma 7.3, Tv0 is good.
Case (2): In this case, |M (Tv0 )| = 1 and |U (Tv0 )| ≥ |S ∩ V (Tv0 )| + 1. Moreover, x|J is integral, where
J = {` ∈ δE (u) : u ∈ (S ∩ V (Tv0 )) ∪ Lmatched (Tv0 ) ∪ B(Tv0 )}.
First, we show that x(δE (s)) ≤ 1 for each stem s ∈ S ∩ V (Tv0 ). We use a contradiction
argument. Suppose that x(δE (s)) > 1. Let δ + (s) denote the set of links of positive x-value
incident to s. Note that every link ` in δ + (s) has x(`) = 1 because x|J is integral. Since
x(δE (s)) > 1, we have |δ + (s)| ≥ 2. This implies that each of the buddy links associated
with the stem s has x-value zero. To see this, note that the set of links in δ + (s) covering
each of the three tree-edges incident to s is an overlapping clique, hence, at most one link of
δ + (s) belongs to one of these overlapping cliques (by the overlapping constraints of (LP0 ));
moreover, each buddy link associated with s is overlapping with the links that belongs to
two of these overlapping cliques (see Figure 10); hence, if a buddy link ` associated with s
has x(`) > 0, then we get a violation (of an overlapping constraint of (LP0 )) for one of the
three overlapping cliques. Since the buddy links associated with s (if any) have x-value zero,
27
we have Φ(x, Tv0 ) ≥ slackx (s) ≥
contradiction.
1
2
x(δE (s)) − x(twinlk(s)) ≥ 21 , and this gives the desired
q
p
s
b2
`
b3
b1
b0
Figure 10. Illustration of a buddy link ` in the proof of Lemma 8.6. Note that `
up (s) and δ (ê) ∩ δ (s), where ê is the tree-edge
is overlapping with links in both δE
E
E
between b3 and s. For example, ` is overlapping with the links sp and sb3 .
For any M -covered leaf w in Tv0 , by Lemma 7.2(1), w is an original node, and moreover,
we have x(δE (w)) ≤ 1, by Lemma 4.3. Thus, we have x(δE (w)) ≤ 1 for each node w ∈
Lmatched (Tv0 ) ∪ (S ∩ V (Tv0 )) = AS(Tv0 ).
0
¯
¯ S
Let Je = J−
w∈AS(Tv0 ) δE (w); this is the set of links in J but not in AG(Tv ). We have
e ≥ 1 , which
e < 1; otherwise, by Lemma 8.5, we would have Φ(x, Tv0 ) ≥ min{ 1 , 1 x(J)}
x(J)
2 2
2
0
is a contradiction. Then, by the conditions in the hypothesis (|M (Tv )| = 1 and |U (Tv0 )| ≥
|S ∩ V (Tv0 )| + 1) and theX
fact that |AU(Tv0 )| = |U (Tv0 )| + 1, we have
e = x(J)
¯ −
1 > x(J)
x(J¯ ∩ δE (w)) ≥ |AU(Tv0 )| − |AS(Tv0 )|
w∈AS(Tv0 )
= |U (Tv0 )| + 1 − 2|M (Tv0 )| − |S ∩ V (Tv0 )|
= |U (Tv0 )| − (1 + |S ∩ V (Tv0 )|) ≥ 0.
Hence, we have |U (Tv0 )| = |S ∩ V (Tv0 )| + 1, and |AS(Tv0 )| = |AU(Tv0 )|. For each w ∈ AS(Tv0 ),
we have x(J¯ ∩ δE (w)) = 1 because x(`) is integral for each link ` ∈ J¯ ∩ δE (w) (otherwise,
P
0
0
0
¯
e
w∈AS(Tv0 ) x(J ∩ δE (w)) ≤ |AS(Tv )| − 1 implies that x(J) ≥ |AU(Tv )| − |AS(Tv )| + 1 ≥ 1).
0
For each w ∈ AS(Tv ), define `w to be the link in δE (w) ∩ J¯ with x(`w ) = 1. We claim
that these links form a perfect matching of the auxiliary graph, thus, AM(Tv0 ) = {lw :
w ∈ AS(Tv0 )}. Otherwise, there exist two links `w1 , `w2 (where w1 , w2 ∈ AS(Tv0 )) that are
¯ ≥ |AU(Tv0 )| + 1, which implies that
incident to the same node u ∈ AU(Tv0 ). Then, x(J)
P
0
0
e = x(J)
¯ −
¯
x(J)
w∈AS(Tv0 ) x(J ∩ δE (w)) ≥ |AU(Tv )| + 1 − |AS(Tv )| ≥ 1. This contradicts the
e < 1. Our claim follows.
fact that x(J)
We claim that AM(Tv0 ) covers Tv0 . Otherwise, there exists a tree-edge ê0 in Tv0 that is not
covered by AM(Tv0 ). Let δ + (ê0 ) denote the set of links of positive x-value in δE (ê0 ). Then, we
have x(δ + (ê0 )) ≥ 1, and none of the links in δ + (ê0 ) is incident to (S ∩ V (Tv0 )) ∪ Lmatched (Tv0 );
the latter assertion holds because x(`w ) = 1 and x(δE (w)) ≤ 1 for every w ∈ AS(Tv0 ), i.e., the
nodes in AS(Tv0 ) are already “saturated” by AM(Tv0 ). Thus, Lemma 8.5 applies to δ + (ê0 ),
and we have Φ(x, Tv0 ) ≥ min{ 21 , 12 δ + (ê0 )} ≥ 12 , which is a contradiction. Our claim follows:
AM(Tv0 ) covers Tv0 .
Additionally, we claim that AM(Tv0 ) has no links between v̄ and S ∩ V (Tv0 ). By way of
contradiction, assume that there exists a link ` = sq ∈ AM(Tv0 ) such that s ∈ S ∩ V (Tv0 )
and q ∈
/ V (Tv0 ). Suppose that q is an original non-leaf node. Then sq is a link between two
28
original non-leaf nodes with x(sq) = 1, hence, we have Φ(x, Tv0 ) ≥ 21 due to Fact 4.7 and the
following term in Φ(x, Tv0 ):
1
x(E((V −L) ∩ V (Tv0 ), (V −L) ∩ (V (T 0 )−V (Tv0 )))).
2
This is a contradiction. Otherwise, if q is a compound node or an original leaf, then we
claim that an (up-to-5) greedy contraction applies, and this too gives a contradiction. Note
that |S ∩ V (Tv0 )| ≤ 2 and |M (Tv0 )| = 1, hence, |AM(Tv0 )| ≤ 4. The credit assigned to the
leaves and matching links in Tv0 is 23 + |U (Tv0 )| = 32 + |S ∩ V (Tv0 )| + 1 = |AM(Tv0 )| + 21 . If
q is a compound node or an original M -exposed leaf, then q provides one additional credit,
and this suffices for an (up-to-5) greedy contraction of the links of AM(Tv0 ); otherwise, q is
an M -covered leaf, and the M -link `q incident to q provides additional 23 credit, and this
suffices for an (up-to-5) greedy contraction of `q together with the links of AM(Tv0 ). Our
claim follows: AM(Tv0 ) has no link of the form v̄s, s ∈ S ∩ V (Tv0 ).
Finally, we prove that twinlk(s) is in AM(Tv0 ) for each stem s ∈ S ∩ V (Tv0 ). Consider any
stem s ∈ S ∩ V (Tv0 ). By Lemma 8.2(1), Ts0 is contained in Tv0 and every node in Ts0 is original.
Let the leaves of Ts0 be a, b0 ; thus, twinlk(s) = ab0 .
Lemma 7.2(2) implies that one end of the twin link ab0 is M -covered. Since x|J is integral,
0
and J contains ab0 (each link incident to Lmatched (Tv0 ) is in
PJ), either ab0 ∈ AM(Tv ) or
x(ab0 ) = 0. We may assume that x(ab0 ) = 0. Suppose that b∈Lbud (s) x(buddylk(b)) = 0,
i.e., there does not exist a buddy link incident to Ts0 that has positive x-value. Consider the
link `s in AM(Tv0 ) incident with s; clearly, x(`s ) = 1. Thus, we have Φ(x, Tv0 ) ≥ slackx (s) ≥
1
1
2 x(`s ) = 2 . This is a contradiction.
q
s
b2
è
`s
a
b0
Figure 11. Illustration of è and `s in the proof of Lemma 8.6.
S
Now, we may assume that there exists a buddy link è ∈ b∈Lbud (s) {buddylk(b)} such that
x( è) > 0. W.l.o.g., let the original ends of è be a and b2 . Let q be the least common ancestor
of a and b2 w.r.t. T ; see Figure 11 and also see Figure 2 in Section 2. Observe that Tv0
contains q (either as an original node or within a compound node), because Tv0 has at least
3 leaves, whereas every proper subtree of Tq has ≤ 2 leaves. Thus, Tv0 contains both ends of
è. Note that b2 (the end of è not in T 0 ) may be an original leaf of T 0 or it may be contained
s
v
in a compound leaf of Tv0 (since C(Tv0 ) = ∅, b2 cannot be contained in a non-leaf compound
node of Tv0 ).
up (s), otherwise, ` , è form an overlapping pair (since x( è) > 0, x(` ) =
Observe that `s 6∈ δE
s
s
0 . Thus, ` = sb (` has an end
1). By similar arguments, `s cannot have both ends in Pa,s
s
0
s
dn (s) and ` ∈ AM(T 0 )). Let `¯ denote the link in
at an M -exposed leaf of Ts0 because `s ∈ δE
s
q
v
0
AM(Tv ) that covers the tree-edge between q and its parent (possibly, q = v, and in this case,
`¯q is the link incident with v̄ in AM(Tv0 )). Note that `¯q has an end at a leaf in Tq0 . Then,
29
`¯q forms an overlapping pair with either `s or è (since x(`¯q ) = 1, x(`s ) = 1, x( è) > 0). This
contradicts the overlapping constraints of (LP0 ). This completes the proof of the lemma.
The following theorem is the key result of this paper. It is used only once; see the last paragraph
of the proof of Theorem 8.8. Note that the last alternative in Theorem 8.7 (item (3)) is addressed
in the last paragraph of the proof of Theorem 8.8.
Theorem 8.7. Suppose that no (up-to-5) greedy contractions are applicable. Let Tv0 be a semiclosed
tree that is not good. Then one of the following holds for Tv0 .
(1) Tv0 is a deficient 3-leaf tree.
(2) Tv0 is a deficient 4-leaf tree.
(3) Tv0 has 4 leaves with |M (Tv0 )| = 1, and moreover Tv0 has no cover of size 3.
Proof. Since Tv0 is not good, Lemma 7.3 and Lemma 8.6(1) imply that Tv0 6= T 0 , C(Tv0 ) = ∅, and
|M (Tv0 )| = 1.
Observe that Tv0 has at least one M -exposed leaf. Otherwise, since |M (Tv0 )| = 1, Tv0 has exactly
two leaves and there exists an M -link between these two leaves; moreover, by Lemma 7.2(1), every
node on the path of T 0 between these two leaves is original; it follows that the link in M (Tv0 ) is a
twin link; this contradicts the definition of M . Thus, Tv0 has an M -exposed leaf and exactly two
M -covered leaves. Hence, Tv0 has at least three leaves.
Also, observe that |S ∩ V (Tv0 )| ≤ 2. Otherwise, suppose that |S ∩ V (Tv0 )| ≥ 3. Then, by
Lemma 8.2(1), for every stem s ∈ S ∩ V (Tv0 ), every node in Ts0 is original; hence, there exists a
stem s∗ ∈ S ∩ V (Tv0 ) such that both the leaves of Ts0∗ are M -exposed and there exists a twin link
between these two leaves; this contradicts Lemma 7.2(2).
Let ê denote the tree-edge between v and its parent. Let J denote {` ∈ δE (u) : u ∈ (S ∩ V (Tv0 )) ∪
Lmatched (Tv0 ) ∪ B(Tv0 )}, where B(Tv0 ) = B orig (Tv0 ) ∪ B comp (Tv0 ) (see the discussion before Lemma 8.4
for the definitions of B orig (Tv0 ) and B comp (Tv0 )).
P
By Lemma 8.4, y can be written as a convex combination i∈Z λi xi such that xi ∈ Las3 (LP0 )
and xi |J is integral. Since Tv0 is not good and |M (Tv0 )| = 1, Lemma 7.3 implies that Φ(y, Tv0 ) < 21 .
Thus, there exists an i0 ∈ Z such that Φ(xi0 , Tv0 ) < 12 .
We claim that |U (Tv0 )| ≤ |S ∩V (Tv0 )|+1. To see this, suppose that |U (Tv0 )| ≥ |S ∩V (Tv0 )|+1. Then,
all of the conditions of Lemma 8.6(2) apply, hence, the lemma implies that |U (Tv0 )| = |S ∩V (Tv0 )|+1.
Our claim follows.
Now, we analyze a few cases, depending on the number of leaves of Tv0 . By Lemma 8.2(1), for
every stem s ∈ S ∩ V (Tv0 ), every node in Ts0 is original; thus, each stem s ∈ S ∩ V (Tv0 ) contributes
two original leaves to Tv0 . Therefore, |S ∩ V (Tv0 )| ≤ |L(Tv0 )|/2.
Case 1: Tv0 has exactly three leaves. Let the three leaves be a, b1 , b2 , where a is M -exposed and
b1 , b2 are M -covered; thus, b1 b2 is the unique link in M (Tv0 ), and b1 , b2 are original nodes by
Lemma 7.2(1). Clearly, |S ∩ V (Tv0 )| ≤ 1. We have two subcases, depending on |S ∩ V (Tv0 )|.
Subcase 1.1: |S ∩ V (Tv0 )| = 1. Let s be the stem in Tv0 . The M -link b1 b2 cannot have both
ends in Ts0 (since M contains no twin links). Thus, the M -exposed leaf a is a leaf of Ts0 .
Then, it can be seen that a is a bud and buddylk(a) = b1 b2 (see the definition of bud
in Section 2). This is a contradiction, since M contains no buddy links.
Subcase 1.2: |S ∩ V (Tv0 )| = 0. Then we have |U (Tv0 )| = 1 = 1 + |S ∩ V (Tv0 )|. By
Lemma 8.6(2)(i–iv), there exist two links `v ∈ δE (ê) and `a ∈ δE (a) such that xi0 (`v ) =
xi0 (`a ) = 1, these two links cover Tv0 , and moreover, each of b1 , b2 is incident to exactly
one of these two links (since the auxiliary graph has a perfect matching formed by these
two links).
30
If there is only one non-leaf node with degree other than 2 in Tv0 (see Figure 6(a)),
then Tv0 is a deficient 3-leaf tree. We are done. Otherwise, we have exactly two non-leaf
nodes u, q in Tv0 with degree other than 2. In fact, both these nodes have degree 3 since
Tv0 has exactly 3 leaves. W.l.o.g., let u be a proper ancestor of q. Then, Tq0 has only two
leaves. By the argument at the beginning of the proof, the M -link in Tv0 cannot connect
both leaves in Tq0 . This implies that one leaf of Tq0 is M -exposed. Thus, a is a leaf of
Tq0 . W.l.o.g., let the other leaf of Tq0 be b2 . Then, b1 is the third leaf, which is not in Tq0 .
Suppose that `v is incident to b1 and `a is incident to b2 . Then, the tree-edge between q and its parent is not covered by these two links (see Figure 12(a)). This is a
contradiction. Hence, `v is incident to b2 and `a is incident to b1 (see Figure 12(b)).
Therefore, Tv0 satisfies all the conditions of a deficient 3-leaf tree.
(a)
v
`v
u
(b)
`v
u
q
b1
v
q
b1
b2
`a
a
b2
`a
a
Figure 12. The links `v and `a in Subcase 1.2 of the proof of Theorem 8.7. The
dashed lines indicate links and the thick dashed lines indicate M -links.
Case 2: Tv0 has exactly four leaves. Let the four leaves be a1 , a2 , b1 , b2 , where a1 , a2 are M -exposed
and b1 , b2 are M -covered; thus, b1 b2 is the unique link in M (Tv0 ), and b1 , b2 are original nodes
by Lemma 7.2(1). We have 1 = |U (Tv0 )| − 1 ≤ |S ∩ V (Tv0 )| ≤ 2. As above, we have two
subcases, depending on |S ∩ V (Tv0 )|.
Subcase 2.1: |S ∩ V (Tv0 )| = 1. Let s be the stem in Tv0 . By Lemma 7.2(2), it can be seen
that the twin link of s is incident to one M -exposed leaf, say a1 , and to one M -covered
leaf, say b1 (see Subcase 1.1).
Note that |U (Tv0 )| = 2 = 1 + |S ∩ V (Tv0 )|. By Lemma 8.6(2)(i–iv), there exist three
links a1 b1 (the twin link of s), `v ∈ δE (ê) and `a2 ∈ δE (a2 ) such that xi0 (a1 b1 ) =
xi0 (`v ) = xi0 (`a2 ) = 1, these three links cover Tv0 , and moreover, each of s, b2 is incident
to exactly one of the two links `v , `a2 (since the auxiliary graph has a perfect matching
formed by the three links); moreover, s cannot be incident to `v (by Lemma 8.6(2)(iv)).
Thus `v is incident to b2 , and `a2 is incident to s, i.e., `a2 = a2 s.
Let p be the least common ancestor of s and a2 (in Tv0 ). If Tp0 does not contain all
the leaves of Tv0 , then it can be seen that the tree-edge between p and its parent is not
covered by the three links a1 b1 , a2 s, `v (see Figure 13(a)). This is a contradiction (since
Tv0 is covered by {a1 b1 , a2 s, `v }). It follows that Tp0 contains all the leaves of Tv0 . Then,
it can be seen that Tv0 is a deficient 4-leaf tree (see Figure 13(b)).
Subcase 2.2: |S ∩ V (Tv0 )| = 2. If Tv0 has no cover of size 3, then item (3) in the statement
of the theorem holds. Thus, we may assume that Tv0 has a cover of size 3.
Let s1 , s2 be the two stems in Tv0 . Let a1 , b1 be the original leaves of Ts01 , and let
a2 , b2 be the original leaves of Ts02 . We may assume that a1 , a2 are M -exposed and b1 , b2
are M -covered, because each twin link is incident to one M -exposed leaf and to one
31
(a)
v
(b)
`v
v
p
a2
`v
p
b2
a2
`a2
`a2
b1
s
s
b2
a1
b1
a1
Figure 13. The links `v , `a2 and a1 b1 in subcase 2.1 of the proof of Theorem 8.7.
The dashed lines indicate links and the thick dashed lines indicate M -links. Another
case could occur, but it is not shown here (see Figure 7).
M -covered leaf by Lemma 7.2(2) (see Subcase 2.1). Thus, b1 b2 is the unique link in
M (Tv0 ).
By Lemma 8.2(1) and C(Tv0 ) = ∅, Tv0 has no compound nodes, i.e., all its nodes are
original.
Observe that Tv0 has at least 3 21 credits from the two M -exposed leaves and the
M -link.
Consider the possible leaf-to-leaf links of Tv0 . The link a1 a2 does not exist, by
Lemma 7.2(2), since a1 , a2 are both M -exposed. If both the links a1 b2 and a2 b1 exist, then an (up-to-5) greedy contraction applies (we contract these two links and we
have 3 21 credits). Thus, at most one of the links a1 b2 or a2 b1 is present.
If Tv has no leafy 3-cover w.r.t. T , then it satisfies all the conditions for a bad 2-stem
tree, hence, it would have been contracted in Preprocessing step 1 (Λ-contraction). This
is a contradiction.
Thus, we may assume that Tv has a leafy 3-cover Je = {`0 , `1 , `2 } w.r.t. T , where
`0 = uw is a link such that one end u is in Tv and the other end w is a leaf in L−L(Tv ).
Let `00 = uw0 be the link w.r.t. T 0 that corresponds to `0 , where w0 = w if w is still an
original node in T 0 , and otherwise, w0 is the compound node containing w. Note that
in either case, w0 is not in V (Tv0 ) since every node in Tv0 is original. If w0 is a compound
node or an M -exposed original leaf, then an (up-to-5) greedy contraction applies (we
contract the 3 links in Je and we have 4 21 credits). Otherwise, if w0 is an M -covered
original leaf, then again an (up-to-5) greedy contraction applies (we contract 4 links,
namely, the links in Je and the M -link incident to w, and we have 3 12 + 32 = 5 credits).
Thus, the existence of a leafy 3-cover gives a contradiction.
Case 3: Tv0 has at least five leaves. Then, Tv0 has at least three M -exposed leaves and two M -covered
leaves. Thus, we have 3 ≤ |U (Tv0 )| ≤ |S ∩ V (Tv0 )| + 1 ≤ 3, since |S ∩ V (Tv0 )| ≤ 2; it follows
that |U (Tv0 )| = 3, |S ∩ V (Tv0 )| = 2, and Tv0 has exactly 5 leaves.
Let s1 , s2 be the two stems in Tv0 . Let a1 , b1 be the original leaves of Ts01 , and let a2 , b2
be the original leaves of Ts02 . We may assume that a1 , a2 are M -exposed and b1 , b2 are M covered, because each twin link is incident to one M -exposed leaf and to one M -covered leaf
by Lemma 7.2(2) (see Subcase 2.1). Thus, b1 b2 is the unique link in M (Tv0 ). Let u be the
fifth leaf of Tv0 , where u is not in Ts01 nor in Ts02 .
Since |U (Tv0 )| = 3 = 1 + |S ∩ V (Tv0 )|, Lemma 8.6(2)(i–iv) implies that there exist four links
`v ∈ δE (ê), `u ∈ δE (u), a1 b1 , a2 b2 (the twin links of s1 , s2 ) such that each of s1 , s2 is incident
32
to exactly one of the two links `v , `u (since the auxiliary graph has a perfect matching formed
by these four links); moreover, `v cannot be incident to {s1 , s2 } (by Lemma 8.6(2)(iv)). This
is impossible. Thus, Tv0 cannot have more than four leaves.
The theorem follows from the above case analysis.
8.3. Addressing deficient trees. This section discusses the missing piece of the algorithm, namely,
the handling of deficient trees. Even, et al., see [5, Section 4.3], presented an elegant method for
addressing deficient 3-leaf trees. We use essentially the same method in Part I (Section 7.2). The
key point is to compute another (auxiliary) matching of the leaf-to-leaf links of T 0 , and then to find
a minimally semiclosed tree Tv0 w.r.t. this auxiliary matching; moreover, we prove (in Theorem 7.4
of Part I) that Tv0 is a good semiclosed tree w.r.t. the “original matching,” and Tv0 has a fitting cover
of appropriate size. In this section, we extend this method to address deficient 3-leaf trees as well as
deficient 4-leaf trees. The extension is straightforward (but non-trivial). We start by “transforming”
the current tree T 0 to another tree Te; this is achieved by applying “latch contractions” to the latches
of deficient 4-leaf trees in an appropriate way (the details are discussed below). Then, we apply the
method from Part I (Section 7.2) to Te. Thus, the main goal in this section is to find a semiclosed
tree Tv0 w.r.t. M of the current tree T 0 such that Tv0 is good (i.e., it has ≥ |Γ(M, Tv0 )| + 1 credits)
and Tv0 has a fitting cover of size |Γ(M, Tv0 )|.
For a deficient 3-leaf tree Tv0 , if Tv0 is not a proper subtree of another deficient 3-leaf tree, then
we call Tv0 a maximal deficient 3-leaf tree. Similarly, we define a maximal deficient 4-leaf tree. By
Property 2.1, any two different maximal deficient 3-leaf trees are disjoint; similarly, any two different
maximal deficient 4-leaf trees are disjoint.
Let E latch denote the set of latches of all maximal deficient 4-leaf trees of T 0 . Let Te = T 0 /E latch .
We mention that Te is an auxiliary graph that is used for achieving the main goal of this section
(find a semiclosed tree Tv0 with desired properties), and other than that, the algorithm does not
refer to Te at all; thus, the algorithm does not add any latch to F while constructing Te, nor does it
incur any credits while constructing Te. By a latched-node we mean a node of Te that results from
the contraction of a latch (note that these are not compound nodes of T 0 , and the algorithm does
not assign any credits to a latched-node).
Since two different maximal deficient 4-leaf trees are disjoint, each “latch contraction” results in
a distinct latched-node; thus, there is a bijection between the latches of T 0 and the latched-nodes of
Te. For a subtree Tew of Te, we use latch(Tew ) to denote the set consisting of latches that correspond
to latched-nodes of Tew .
For a node w of T 0 , let w
e denote the node of Te that corresponds to w. If there exists no latch cs of
0
0
T such that w is in Pc,s (the path of T 0 between c, s), then w
e = w, otherwise, w
e is the latched-node
0 (in this case, the latch cs is unique
formed by contracting the (unique) latch cs such that w is in Pc,s
because two different maximal deficient 4-leaf trees are disjoint).
Then, as in Part I (Section 7.2), we construct an auxiliary matching of the leaf-to-leaf links of
Te denoted by M new . For ease of exposition, we denote the image (in Te) of M (T 0 ) by M (Te), and
we refer to the image (in Te) of a link in M (T 0 ) as an M -link w.r.t. Te (this is justified because any
M -link w.r.t. T 0 has no ends in common with the nodes of T 0 that get contracted into latchednodes). To construct M new , we start with M new := M , then we examine each maximal deficient
3-leaf tree Tewe of Te and we replace the unique link of M (Tewe ) in M new by another leaf-to-leaf link.
In more detail, consider any maximal deficient 3-leaf tree Tewe , and let the three leaves be a, b, d,
where a is M -exposed, b is the ceiling leaf, and bd is the unique link in M (Tewe ); we remove the
M -link bd from M new and instead we add the link ad to M new (see Algorithm 2). Since any two
different maximal deficient 3-leaf trees are disjoint, this replacement takes place independently for
each maximal deficient 3-leaf tree. We mention that M new is an auxiliary matching that is used for
33
achieving the main goal of this section (find a semiclosed tree Tv0 with desired properties), and other
than that, the algorithm does not refer to M new at all; whereas, the matching M (and its image
M (T 0 )) are used throughout the algorithm and its analysis.
Theorem 8.8. Suppose that no (up-to-5) greedy contractions are applicable. Let Te = T 0 /E latch ,
and let Teve be a minimally semiclosed tree w.r.t. M new . Then, ve cannot be a latched-node of Te; thus,
ve = v, where v is a node of T 0 . Moreover, Tv0 is a good semiclosed tree w.r.t. M and Tv0 has a fitting
cover Γ(M new , Tev ) ∪ latch(Tev ) of size |Γ(M, Tv0 )|.
Proof. We start by addressing the first statement. By way of contradiction, suppose that ve is a
latched-node. Observe that every latched-node is contained in a deficient 3-leaf tree of Te (a latchednode is formed by a “latch contraction” that results in a deficient 3-leaf tree). Hence, Teve is a subtree
of a maximal deficient 3-leaf tree of Te. But, by our construction of M new , the ceiling leaf of every
maximal deficient 3-leaf tree is M new -exposed, hence, no subtree of a maximal deficient 3-leaf tree
is semiclosed w.r.t. M new . Thus, Teve is not semiclosed w.r.t. M new . This is a contradiction.
It follows that ve cannot be a latched-node. Thus, there exists a node v in T 0 such that v = ve;
moreover, Teve = Tev , and Tv0 is well defined.
Claim 8.9. (1) If Tv0 has a node in a maximal deficient 4-leaf tree of T 0 , then Tv0
properly contains this maximal deficient 4-leaf tree.
(2) If Tv0 has a node in a maximal deficient 3-leaf tree of T 0 , then Tv0 properly contains
this maximal deficient 3-leaf tree.
We address statement (1) of the claim. Suppose that Tv0 has a node in a maximal deficient 4-leaf
tree Tw0 . By “latch contraction,” Tw0 is transformed to a deficient 3-leaf tree Tewe of Te. Thus, Tev has
a node in a maximal deficient 3-leaf tree Teue , where u
e is an ancestor of w
e in Te. But, (as discussed
new
e
e
above) no subtree of Tue is semiclosed w.r.t. M . Hence, Tv cannot be a subtree of Teue . Then
Property 2.1 implies that Tev properly contains Teue . Hence, v is a proper ancestor of w in T 0 , and Tv0
properly contains Tw0 .
Next, we address statement (2) of the claim. Suppose that Tv0 has a node in a maximal deficient
3-leaf tree Tw0 . If this maximal deficient 3-leaf tree is contained in a maximal deficient 4-leaf tree,
then we are done by the previous discussion. Hence, by Property 2.1, we may assume that Tw0 has
no node in a maximal deficient 4-leaf tree. Then Tewe contains no latched-node, and Tewe = Tew = Tw0 is
a maximal deficient 3-leaf tree of Te. Since Tev has a node in Tew , and no subtree of Tew is semiclosed
w.r.t. M new , Property 2.1 implies that Tev properly contains Tew . It follows that Tv0 properly contains
Tw0 . This proves the claim.
Claim 8.9 implies that Tv0 is neither a deficient 3-leaf tree nor a deficient 4-leaf tree.
Next, we address the second statement in the theorem. To show that Tv0 is semiclosed w.r.t. M ,
we must show that every link incident to an M -exposed leaf of Tv0 has both ends in Tv0 , and every
M -link incident to Tv0 has both ends in Tv0 . Let u be an M -exposed leaf of Tv0 . Suppose that u is
contained in Tw0 , where Tw0 is either a deficient 3-leaf tree or a deficient 4-leaf tree (w.r.t. T 0 ). Then
Tv0 contains Tw0 , by Claim 8.9. Since Tw0 is a semiclosed tree w.r.t. M (by definition of deficient trees),
all links incident with u have both ends in Tw0 , hence, also in Tv0 . Next, suppose that u does not
belong to any deficient 3-leaf tree or any deficient 4-leaf tree (w.r.t. T 0 ). Such an M -exposed node
of Tv0 is an M new -exposed node of Te. Since Tev is semiclosed w.r.t. M new , it follows that Tev contains
both ends of each link incident with u
e = u; hence, Tv0 contains both ends of each link incident with
u. Next, consider an M -link uw incident to Tv0 , where u is a node of Tv0 . If u is contained in Tw0 ,
where Tw0 is a deficient 3-leaf tree or a deficient 4-leaf tree, then note that Tw0 is semiclosed, and
34
moreover, Tw0 is properly contained in Tv0 , hence, both ends of uw are in Tv0 . Otherwise, uw has no
end in a deficient 3-leaf tree or in a deficient 4-leaf tree; then, by our construction of M new , we have
uw ∈ M new (M and M new are the same, except “inside” some deficient tree). Since Tev is semiclosed
w.r.t. M new , it follows that uw has both ends in Tev , hence, uw has both ends in Tv0 . Thus, Tv0 is
semiclosed w.r.t. M .
By the construction of Te and M new , Claim 8.9, and the fact that Tev is semiclosed, we have
|M (Tv0 )| = |M new (Tev )|; moreover, the number of M -exposed nodes of Tv0 is equal to the sum of
the number of M new -exposed nodes of Tev and |latch(Tev )| (since each latched-node of Tev corresponds
to an M -exposed leaf of a maximal deficient 4-leaf tree of Tv0 ). Hence, we have |Γ(M, Tv0 )| =
|Γ(M new , Tev )| + |latch(Tev )|. Clearly, Γ(M new , Tev ) is a fitting cover of Tev , because Tev is a minimally
semiclosed tree w.r.t. M new (by Lemma 6.2). It follows that Γ(M new , Tev ) ∪ latch(Tev ) is a fitting
cover of Tv0 of size |Γ(M, Tv0 )|.
Finally, we apply Theorem 8.7 to prove that Tv0 is good. Suppose that Tv0 is not good. By
Claim 8.9, Tv0 is neither a deficient 3-leaf tree nor a deficient 4-leaf tree. Then, Theorem 8.7(3)
implies that Tv0 has 4 leaves, |M (Tv0 )| = 1, and Tv0 has no cover of size 3. Since |M (Tv0 )| = 1 and
|U (Tv0 )| = 2, we have |Γ(M, Tv0 )| = 3. The previous paragraph shows that Tv0 has a fitting cover of
size |Γ(M, Tv0 )| = 3. Thus, we have a contradiction. It follows that Tv0 is good. This completes the
proof.
The procedure described above to find a good semiclosed tree is summarized in the following
pseudocode.
let E latch denote the set of latches of all maximal deficient 4-leaf trees of T 0 ;
let Te := T 0 /E latch ;
start with M new := M ;
for each maximal deficient 3-leaf tree Tewe in Te do
let b be the ceiling leaf, a be the M -exposed leaf and db be the M -link in Tewe ;
update M new by replacing db by da (M new := M new −{db} ∪ {da});
end
find a minimally semiclosed tree Tev w.r.t. M new ;
Tv0 is a good semiclosed tree w.r.t. M with a fitting cover Γ(M new , Tev ) ∪ latch(Tev ) of size
|Γ(M, Tv0 )| (by Theorem 8.8);
return Tv0 and its fitting cover Γ(M new , Tev ) ∪ latch(Tev );
Algorithm 2: Find a good semiclosed tree by addressing deficient trees.
Algorithm 2 shows how to find a good semiclosed tree Tv0 together with a fitting cover of size
|Γ(M, Tv0 )| for the main loop in Algorithm 1 in Section 7. (The first step of Algorithm 2 applies
“latch contractions” to construct the auxiliary tree Te from T 0 ; “latch contractions” have no effect
on F or on the credits of the algorithm.)
In conclusion, Algorithm 1 runs in polynomial time and returns a solution for TAP of size ≤
3
( 2 + )y(E). This proves Theorem 6.1.
Although our integrality ratio (and approximation guarantee) is proved relative to a “weaker
lower bound” than the optimum value (since the feasible region of a relaxation is a superset of the
convex full of integer solutions), we loose an additive term of in the approximation guarantee,
compared to the 23 approximation guarantee of [10]. But, we can show that our methods achieve an
approximation guarantee of 32 relative to the (integral) optimum value. Observe that the algorithm
is combinatorial and does not refer to the SDP relaxation at all (our analysis does rely on the SDP
relaxation and properties of its solutions). Thus, we can apply the algorithm and analysis for the
35
level t = |E| tightening of (LP0 ), and at this level, the set of (feasible) solutions of Last (LP0 )
corresponds to the convex hull of integer solutions. Consequently, for level t = |E|, Lemma 4.4 can
be restated as follows: if y ∈ Last (LP0 ), where t = |E|, then y|E(L) is in the matching polytope
of G(L) = (L, E(L)). Hence, for level t = |E|, the upperbound on the potential function stated in
Lemma 5.2 can be improved from ( 23 + )y(E) to 23 y(E). In other words, for y ∈ Last (LP0 ), where
t = |E|, the potential function is ≤ 23 y(E). Sections 6–8 show that Algorithm 1 returns a solution
for TAP whose size is ≤ the potential function. Thus, we have the following result.
Corollary 8.10. Let y denote an optimal (integral) solution of TAP. Algorithm 1 runs in polynomial
time, and returns a solution for TAP of size ≤ 32 y(E).
It can be seen that Algorithm 1 can be implemented to run in time O(|V |3 + |V |2 m50 ), where m0
denotes the number of (multi) links in the input and we assume that m0 = O(|V |2 ). This can be
achieved using simple data structures such as adjacency matrices and adjacency lists; the details are
not straightforward, but we skip them. To handle the (up-to-5) greedy contractions, every subset
of size ≤ 5 of the set of maximal links has to be examined (for each link uw in a subset, Pu,w has to
be examined), and there are O(|V |) such contractions, hence, these contractions contribute a total
of O(|V |2 m50 ) to the running time.
Theorem 8.11. Algorithm 1 can be implemented to run in time O(|V |3 + |V |2 m50 ).
Acknowledgments. We thank André Linhares for many discussions. We thank several other
colleagues who read preliminary drafts and gave us insightful comments.
References
[1] J. Cheriyan, H. J. Karloff, R. Khandekar, and J. Könemann. On the integrality ratio for tree augmentation. Oper.
Res. Lett., 36(4):399–401, 2008.
[2] R. Diestel. Graph Theory (4th ed.). Graduate Texts in Mathematics, Volume 173. Springer-Verlag, Heidelberg,
2010.
[3] G. Even, J. Feldman, G. Kortsarz, and Z. Nutov. A 3/2-approximation algorithm for augmenting the edgeconnectivity of a graph from 1 to 2 using a subset of a given edge set. In Approximation, Randomization and
Combinatorial Optimization: Algorithms and Techniques, APPROX-RANDOM 2001, Proceedings, volume 2129
of Lecture Notes in Computer Science, pages 90–101, 2001.
[4] G. N. Frederickson and J. Ja’Ja’. Approximation algorithms for several graph augmentation problems. SIAM J.
Comput., 10(2):270–283, 1981.
[5] G.Even, J.Feldman, G.Kortsarz, and Z.Nutov. A 1.8 approximation algorithm for augmenting edge-connectivity
of a graph from 1 to 2. ACM Transactions on Algorithms, 5:21:6–21:17, 2009.
[6] H.Nagamochi. An approximation for finding a smallest 2-edge connected subgraph containing a specified spanning
tree. Discrete Applied Mathematics, 126:83–113, 2003.
[7] K. Jain. A factor 2 approximation algorithm for the generalized Steiner network problem. Combinatorica, 21(1):39–
60, 2001.
[8] A. R. Karlin, C. Mathieu, and C. T. Nguyen. Integrality gaps of linear and semi-definite programming relaxations
for Knapsack. In Integer Programming and Combinatoral Optimization, volume 6655 of Lecture Notes in Computer
Science, pages 301–314. Springer Berlin / Heidelberg, 2011.
[9] G. Kortsarz, R. Krauthgamer, and J. R. Lee. Hardness of approximation for vertex-connectivity network design
problems. SIAM J. Comput., 33(3):704–720, 2004.
[10] G. Kortsarz and Z. Nutov. A simplified 1.5-approximation algorithm for augmenting edge-connectivity of a graph
from 1 to 2. CoRR, abs/1507.02799, 2015.
[11] T. Rothvoß. Directed Steiner tree and the Lasserre hierarchy. CoRR, abs/1111.5473, 2011.
[12] T.Rothvoß. The Lasserre hierarchy in approximation algorithms. MAPSP Tutorial: Lecture notes http: // www.
math. washington. edu/ ~ rothvoss/ lecturenotes/ lasserresurvey. pdf , 2013.
Dept. of Comb. & Opt., University of Waterloo, Waterloo, Ontario N2L3G1, Canada.
E-mail address: [email protected]
E-mail address: [email protected]
36
| 8cs.DS
|
arXiv:1603.07051v1 [cs.AI] 23 Mar 2016
Cosolver2B: An Efficient Local Search Heuristic for
the Travelling Thief Problem
Mohamed El Yafrani
Belaı̈d Ahiod
LRIT, associated unit to CNRST (URAC 29)
Faculty of Science, Mohammed V University of Rabat
B.P. 1014 Rabat, Morocco
Email: [email protected]
LRIT, associated unit to CNRST (URAC 29)
Faculty of Science, Mohammed V University of Rabat
B.P. 1014 Rabat, Morocco
Email: [email protected]
Abstract—Real-world problems are very difficult to optimize.
However, many researchers have been solving benchmark problems that have been extensively investigated for the last decades
even if they have very few direct applications. The Traveling Thief
Problem (TTP) is a NP-hard optimization problem that aims to
provide a more realistic model. TTP targets particularly routing
problem under packing/loading constraints which can be found in
supply chain management and transportation. In this paper, TTP
is presented and formulated mathematically. A combined local
search algorithm is proposed and compared with Random Local
Search (RLS) and Evolutionary Algorithm (EA). The obtained
results are quite promising since new better solutions were found.
I. I NTRODUCTION
The travelling thief problem (TTP) is a novel NP-hard
problem introduced in [1] to provide a model that better
represents real-world problems. The particularity of TTP is
that it is composed of two other NP-hard problems, namely
the travelling salesman problem and the knapsack problem,
which are interdependent.
The problem can be introduced with the following simplified statement:
”Given n cities and m items scattered among these cities, a
thief with his rented knapsack should visit all n cities, once
and only once each, and pick up some items. The more the
knapsack gets heavier, the more the thief becomes slower. What
is the best path and picking plan to adopt to achieve the best
benefit ?”.
Two particular properties can be extracted from the statement above. First, the overall problem is composed of two
sub-problems: choosing the best path and picking up the most
profitable items. Second, the sub-problems are interdependent:
when the knapsack gets heavier, the speed of the thief decreases. The composition increases the number of possible
solutions, and the interdependence makes it impossible to
isolate sub-problems.
This kind of multiple interdependent components is widely
found in logistics and supply chain management. Nevertheless, few attempts have been made to study and solve these
problems as a whole, and a lot of effort have been made to
solve the components independently. In [3], the authors explain
why there is a gap between the work done by the Evolutionary Computation researchers and real-world applications. As
far as we know, these observations can be extended to all
metaheuristics community.
Examples of other realistic problems with multiple interdependent sub-problems include most supply chain optimization
problems [7], [8], routing problems with loading constraints
such as 2L-CVRP and 3L-CVRP [9], [10], [3], and water tank
delivery [14].
Since TTP was introduced, some algorithms were proposed
to solve it. An Evolutionary Algorithm and a Random Local
Search were proposed in [13] to provide a starting point to
other researchers. In [2], an approach named Cosolver was
introduced to solve TTP by separating the sub-problems and
managing a communication between them. Lastly, an interesting work on TTP introduces many complexity reduction
techniques in order to solve very large instances in a time
budget of 10 minutes [11].
In this paper, we follow the ideas proposed in [2] and [11]
to propose further improvements and introduce an algorithm
based on the Cosolver framework that can solve TTP instances
efficiently.
This paper is organized as follows. In section II TTP is
mathematically defined and investigated. Section III is dedicated to introduce our approach for solving TTP. The tests and
results are presented in section IV. Section V concludes the
paper and outlines areas for future research.
II. BACKGROUND
In this section, we present some background information
about TTP. The problem is formulated and an abstract algorithm is presented.
A. The Travelling Thief Problem
Herein we formulate the Travelling Thief Problem (TTP)
which combines two other well known benchmark problems,
namely the Travelling Salesman Problem (TSP) and the Knapsack Problem (KP).
In TTP, we consider n cities and the associated distance
matrix {dij }. There are m items scattered in these cities, each
item k have a profit pk and a weight wk . A thief with his
knapsack is going to visit all these cities (once and only once
978-1-5090-0478-2/15/$31.00 c 2015 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any
current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or
redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
each), and pick up some items to fill his knapsack. We note
W the maximum capacity of the knapsack, vmin and vmax are
the minimum and maximum possible velocity respectively.
Also, we consider also the following constraints and parameters:
•
•
•
Each item is available in only one city. We note {Ai } the
availability vector, Ai ∈ {1, ..., n} contains the reference
to the city that contains the item i.
We suppose that the knapsack is rented. We note R the
renting price per time unit.
The velocity of the thief changes accordingly to the
knapsack weight. We note vxi the velocity of the thief
at city xi (see equation 1).
vxi = vmax − C ∗ wxi
(1)
−vmin
where C = vmaxW
is a constant value, and wxi
represents the weight of the knapsack at city xi .
The goal of the problem is to find the tour x and the picking
plan z that optimize the total travel gain defined in equation
2.
G(x, z) = g(z) − R ∗ f (x, z)
(2)
P
Where g(z) = m pm ∗ zm is the total value of the items
Pn−1
P
subject to m wm ∗ zm ≤ W , and f (x, z) = i=1 txi ,xi+1 +
txn ,x1 is the total travel time.
The solutions could be naturally coded as follows. The tour
x = (x1 , ..., xn ) is a vector containing the ordered list of cities,
and the picking plan z = (z1 , ..., zn ) is a binary vector such
as zi is equal to 1 if the item i is picked, 0 otherwise.
The interdependence between KP and TSP have been investigated. As shown in [1], [11], optimizing the sub-problems
in isolation (even to optimality) does not guarantee finding
good solutions for the overall problem. Therefore, finding good
global solutions requires an algorithm that takes interdependence of components in consideration, which makes the design
of such an algorithm quite difficult.
B. Cosolver
Cosolver is an abstract algorithm proposed in [2] to solve
TTP. The idea behind the algorithm is simple. Decompose
the overall problem and solve components separately using a
fitness function that takes in consideration the interdependence
between the two components.
Thus, given an initial solution (x0 , z0 ), TTP is decomposed
into the following problems.
1) The Travelling Salesman with Knapsack Problem
(TSKP): consists of finding the best tour x′ that, combined
with the last found picking plan z ′ , optimizes the TTP objective function G. Also, because the total profit function g
does not depend on the tour, instead of maximizing G, we
can consider minimizing the total travel cost T (see equation
3).
T (x, z) = R ∗ f (x, z)
(3)
2) The Knapsack on the Route Problem (KRP): consists of
finding the best picking plan z ′ that maximizes the total gain
G when combined with the last found tour x′ .
Figure 1 represents a flowchart of the Cosolver framework.
III. P ROPOSED
APPROACH
In this section, we propose an algorithm that implements
the Cosolver framework and uses some techniques proposed
in [11].
A. Solution initialization
In TTP, the initial solution has a big impact on the search
algorithm. Thus, the initialization strategy should be chosen
carefully. In our implementation, we use a heuristic approach
to initialize both the tour and the picking plan.
Firstly, the tour is generated using a good TSP algorithm
such as the Lin-Kernighan heuristic [6] or the Ant Colony
Optimization algorithm [5].
Then, the picking plan is initialized using the following
approach.
1) The insertion heuristic proposed in [11] is used to
fill the knapsack with items according to three fitness
approximations.
2) A simple bit-flip search on inserted items is performed
to eliminate some useless items from the picking plan
using the objective function.
We will refer to this heuristic as the insertion & elimination
heuristic.
B. Complexity reduction techniques
Our algorithm uses the following complexity reduction
techniques.
1) TSKP neighborhood reduction: The Delaunay triangulation [4] is used as a candidate generator for the 2-OPT
heuristic. This strategy was also proposed in [11].
2) Objective value recovery: Instead of using the objective
function to calculate the objective value of neighbors, this
value can be recovered by keeping track of time and weight
information at each tour’s city. The following vectors are used
to perform such operation.
acc
• Time accumulator (t
): a vector that contains the current
time at each city of the tour.
acc
• Weight accumulator (w
): a vector that contains the
current weight at each city of the tour.
reg
• Time register (t
): a vector that contains the added time
at each city of the tour.
reg
• Weight register (w
): a vector that contains the added
weight at each city of the tour.
Figure 2 shows the effect (in gray) of mutating a TTP
solution on the total travel time. Two mutations are shown
(in black): the one-bit-flip on the picking plan, and a 2-OPT
exchange on the tour.
A similar technique named incremental evaluation was
proposed in [11]. In our implementation, we go further and
use these techniques also to recover the vectors.
Fig. 1.
bit- ip
x(1)
A simplified Cosolver flowchart
x(i)
x(n)
t(1)
t(i-1)
x(1)
2-opt
t(n)
x(j)
x(i)
x(n)
t(1)
t(n)
t(i)
t(j)
Fig. 2. Example of changes that can be applied to a TTP solution (in black)
and its effect on the total traveling time function (in gray)
3) Objective function calculation: The objective function
has a complexity of O(m ∗ n) as proposed in [12], where
m is the number of items and n is the number of cities. The
complexity could be reduced to O(k∗n) where k is the number
of items per city. In the CEC ’2015 competition website 1 ,
the Java implementations already has a reduced complexity
but it uses a technique highly dependent on the TTP instance
generator. We believe that a better way is to class items per
city.
Figure 3 explains this technique using a simple example. In
the example, 4 cities are considered: city 1 contains 4 items (1,
2, 3 and 4), city 2 contains 2 items (5 and 6), city 3 contains
no items, city 4 contains 3 items (7, 8 and 9).
Fig. 4.
Fig. 3.
Data structure for classing items by cities
C. Proposed algorithm
Our heuristic, namely Cosolver2B, is based on local search
algorithms, and it implements the Cosolver framework. In
addition to the techniques presented above, we use the 2-OPT
heuristic to solve TSKP and a bit-flip search to solve the KRP.
The algorithm is summarized in the flowchart in figure 4.
1 http://cs.adelaide.edu.au/
optlog/CEC2015Comp/
Cosolver2B: Cosolver with 2-opt and bit-flip
The algorithm starts with a tour generated using the LinKernighan heuristic. The insertion & elimination heuristic is
then used to initialize the picking plan. The obtained tour
and picking plan are then given to the TSKP optimizer that
tries to improve the current tour using a 2-OPT heuristic.
The resulting tour is combined with the current picking plan
and given to the KRP optimizer that uses a simple bit-flip to
find a decent picking plan. If there is improvement, the new
picking plan is combined with the current tour and given to
the TSKP optimizer. This process is repeated until there is no
improvement.
IV. E XPERIMENTAL RESULTS
We tested our algorithm on 15 TTP instances of various
sizes [13]. The obtained results are compared with RLS and
TABLE I
R ESULTS OF C OSOLVER 2B COMPARED TO EA AND RLS
instance
berlin52 n255 uncorr-similar-weights 05
kroA100 n495 uncorr-similar-weights 05
ch150 n745 uncorr-similar-weights 05
ts225 n1120 uncorr-similar-weights 05
a280 n1395 uncorr-similar-weights 05.ttp
lin318 n1585 uncorr-similar-weights 05
u574 n2865 uncorr-similar-weights 05
dsj1000 n4995 uncorr-similar-weights 05
rl1304 n6515 uncorr-similar-weights 05
fl1577 n7880 uncorr-similar-weights 05
d2103 n10510 uncorr-similar-weights 05
pcb3038 n15185 uncorr-similar-weights 05
fnl4461 n22300 uncorr-similar-weights 05
rl11849 n59240 uncorr-similar-weights 05
usa13509 n67540 uncorr-similar-weights 05
pla33810 n169045 uncorr-similar-weights 05
EA
Cosolver2B
firstfit
Cosolver2B
bestfit
objective
time
objective
RLS
time
objective
time
objective
time
20591
38864
40922
84907
104365
75387
223165
320165
573667
589756
801110
1119168
1477673
2152954
2226676
-9929644
1.28
1.92
2.69
3.82
5.64
6.87
21
47
95
127
214
427
600
600
600
600
20591
38864
40922
84907
104365
75387
223165
320165
573667
589756
801110
1119169
1478963
4328939
5077892
-3158915
0.89
1.06
1.38
1.69
1.84
3.81
5.42
12
21
28
44
101
265
600
600
600
25440
39599
52225
88925
107233
114915
243571
332916
571634
665866
869584
1117659
1220469
1785671
4420300
14887314
0.13
0.21
0.38
0.61
0.88
1.25
5.36
5.95
68
139
357
600
600
600
600
600
26658
39597
53075
89079
112003
114915
253770
332917
586576
684731
869584
1086092
696650
478884
4082169
15085752
0.11
0.27
0.43
0.92
1.09
1.54
6.51
8.44
83
148
389
600
600
600
600
600
EA proposed in the same paper 2 .
We have implemented two versions of Cosolver2B. The first
is best fit (also known as best improvement) which searches
the entire neighborhood on a given iteration and selects the
best neighbor. The second is first fit (also known as first
improvement) which stops the neighborhood search once a
better solution is found.
Since our algorithms are deterministic, one run per instance
is sufficient. The algorithms have a maximum runtime limit
of 600 seconds. Note that our implementation uses the Java
platform, and the tests are performed on a core i3-2370M CPU
2.40GHz machine with 4 GB of RAM, running Linux.
On the other hand, due to their random behavior, 30 runs
per instance were performed using RLS and EA. The tests on
EA and RLS were performed on a different machine (Matlab
2014, core i7 2600 CPU 3.4 GHz, with 4GB of RAM, running
Windows 7). Thus, The runtimes for these algorithms are given
for guidance. Furthermore, in our comparison, we only select
the best found solution for these two algorithms.
The results are reported in table I. Note that the runtimes
are measured with seconds and the best objective values are
made bolder.
The results show that our algorithm was able to find new
better solutions for most tested instances only by combining
local search algorithms. Furthermore, we have made the following observations:
• Cosolver2B surpasses EA and RLS on various TTP sizes.
• The runtime of Cosolver2B is very decent for small and
mid-size instances. However, for very large instances the
runtime is quite high compared to RLS and EA.
2 for
instance files and raw results of RLS and EA, the reader is referred to
https://sites.google.com/site/mohammadrezabonyadi/standarddatabases/travellingthief-problem-data-bases-and-raw-results
•
•
•
Both EA and RLS have an unpredictable behavior, while
Cosolver2B is deterministic and garantees decent solutions for most instances.
For many instances, most optimization runtime and gain
improvement is done on the KRP sub-problem.
The best fit strategy performed better than first fit for
most small and mid-size instances.
V. C ONCLUSION
In this paper, a combined local search heuristic has been
proposed. Our approach is based on the Cosolver framework
and complexity reduction techniques. The experimental results
have shown promising results both in terms of solution quality
and runtime. The good performance of Cosolver2B and the
memetic algorithm MALTS on very large scale instances
[11] proves the importance of making strong assumptions
about the problem domain when dealing with multi-component
problems.
Furthermore, we engage to improve our algorithm in terms
of space exploration and runtime by using more sophisticated
search heuristics.
ACKNOWLEDGMENT
The authors express their gratefulness to the following
persons who have contributed to this work with their useful
comments: Asmae El Ghazi, Aziz Ouaarab, and Mehdi El
Krari.
R EFERENCES
[1] Mohammad Reza Bonyadi, Zbigniew Michalewicz, and Luigi Barone.
The travelling thief problem: the first step in the transition from
theoretical problems to realistic problems. In Evolutionary Computation
(CEC), 2013 IEEE Congress on, pages 1037–1044. IEEE, 2013.
[2] Mohammad Reza Bonyadi, Zbigniew Michalewicz, Michaŏ Roman Przybyŏek, and Adam Wierzbicki. Socially inspired algorithms
for the travelling thief problem. In Proceedings of the 2014 conference
on Genetic and evolutionary computation, pages 421–428. ACM, 2014.
[3] Andreas Bortfeldt. A hybrid algorithm for the capacitated vehicle
routing problem with three-dimensional loading constraints. Computers
& Operations Research, 39(9):2248–2257, 2012.
[4] Boris Delaunay. Sur la sphere vide. Izv. Akad. Nauk SSSR, Otdelenie
Matematicheskii i Estestvennyka Nauk, 7(793-800):1–2, 1934.
[5] Marco Dorigo and Luca Maria Gambardella. Ant colonies for the
travelling salesman problem. BioSystems, 43(2):73–81, 1997.
[6] Keld Helsgaun. An effective implementation of the lin–kernighan
traveling salesman heuristic. European Journal of Operational Research,
126(1):106–130, 2000.
[7] Maksud Ibrahimov, Arvind Mohais, Sven Schellenberg, and Zbigniew
Michalewicz. Evolutionary approaches for supply chain optimisation:
part i: single and two-component supply chains. International Journal
of Intelligent Computing and Cybernetics, 5(4):444–472, 2012.
[8] Maksud Ibrahimov, Arvind Mohais, Sven Schellenberg, and Zbigniew
Michalewicz. Evolutionary approaches for supply chain optimisation.
part ii: multi-silo supply chains. International Journal of Intelligent
Computing and Cybernetics, 5(4):473–499, 2012.
[9] Manuel Iori and Silvano Martello. Routing problems with loading
constraints. Top, 18(1):4–27, 2010.
[10] Stephen CH Leung, Zhenzhen Zhang, Defu Zhang, Xian Hua, and
Ming K Lim. A meta-heuristic algorithm for heterogeneous fleet vehicle
routing problems with two-dimensional loading constraints. European
Journal of Operational Research, 225(2):199–210, 2013.
[11] Yi Mei, Xiaodong Li, and Xin Yao. Improving efficiency of heuristics
for the large scale traveling thief problem. In Simulated Evolution and
Learning, pages 631–643. Springer, 2014.
[12] Yi Mei, Xiaodong Li, and Xin Yao. On investigation of interdependence
between sub-problems of the travelling thief problem. Soft Computing,
pages 1–16, 2014.
[13] Sergey Polyakovskiy, Mohammad Reza, Markus Wagner, Zbigniew
Michalewicz, and Frank Neumann. A comprehensive benchmark set
and heuristics for the traveling thief problem. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO), Vancouver,
Canada, 2014.
[14] Jacob Stolk, Isaac Mann, Arvind Mohais, and Zbigniew Michalewicz.
Combining vehicle routing and packing for optimal delivery schedules
of water tanks. OR Insight, 26(3):167–190, 2013.
| 9cs.NE
|
State and Parameter Estimation Based on Filtered Transformation
for a Class of Second-Order Systems
Mehdi Tavan, Kamel Sabahi, and Saeid Hoseinzadeh
Abstract— This paper addresses the problem of state and
parameter estimation for a class of second-order systems with
single output. A new filtered transformation is proposed for the
system via dynamic vector and matrix. In this method, the
dynamics of the vector and matrix are derived by immersion
and invariance technique such that the state estimation
condition is guaranteed. Compared to the classical approaches
that persistency of excitation (PE) condition is required for
parameter convergence, the proposed method needs a weaker
one, so called non-square-integrability condition, in the
transformation via dynamic matrix. Simulation results are
concluded for a class of regressors which are not PE but satisfy
the new condition.
I. INTRODUCTION
The persistency of excitation (PE) condition on the
regressor is one of the main tools to ensure convergence in
parameter estimators for identification purposes. Recently, a
new procedure has introduced in [1] to design the parameter
estimators in which the PE condition is replaced with a
weaker one, called non-square-integrability condition. In this
procedure some dynamic regressors are designed to
guarantee the new condition for the determinant of an
extended regressor. Applications of this method for some
electrical examples show better transient and robust
performance properties [2, 3]. The use of dynamic regressors
can be found for adaptive observer design via filtered
transformation in [4] and the references cited therein. Besides
the PE condition for parameter convergence, an additional
disadvantage is overparameterisation in such a design. The
immersion and invariance (I&I) based filtered transformation
introduced in [5] reduces the order of the observer with
elimination of output injection error term in update law.
Application of this method for DC-DC boost converter is
evaluated via experiments in [6]. The filtered transformations
in [4, 5] are formed via a dynamic vector which yields to a
descent algorithm with singular descent rate that is at the
core of convergence problem in parameter estimation [7].
In this paper a filtered transformation is introduced via
dynamic matrix for a class of second-order systems with
single-output. The main contribution of this paper is to
propose an estimator based on the new transformation, which
asymptotically estimates unavailable state and unknown
parameters. The parameter convergence is ensured under the
non-square-integrability condition. Fulfilling this condition is
discussed for the estimator. Moreover, it is shown that the
estimator provides an arbitrary convergence speed. Also
M. Tavan and Saeid Hoseinzadeh are with the Department of Electrical
Engineering, Mahmudabad Branch, Islamic Azad University, Mahmudabad,
Iran (e-mail: [email protected]).
K. Sabahi is with the Department of Electrical Engineering, Mamaghan
Branch, Islamic Azad University, Mamaghan, Iran.
simulation results are concluded for a class of regressors
which are not PE but satisfy the new condition.
II. PROBLEM STATEMENT AND PRELIMINARY
Consider a class of single-output nonlinear systems that
can be described in suitable co-ordinates by equations of the
form
̇
(
)
(
̇
)
(
)
(1)
(
)
(2)
where
is the measurable state,
is the
unmeasurable state,
is a constant vector of unknown
( ) are known
parameters,
( ) ( ) ( ) and
mappings. As mentioned in Introduction, our objective
consists of estimating and . As shown in [6], the average
model of some popular switch converters, such as DC-DC
boost, buck and buck-boost, can be represented in the form of
above equations.
For the system, the unmeasurable state can be assumed
to be generated by a parameter dependent change of
coordinate
[4]
[
(3)
]
where ( )
is an auxiliary dynamic vector, which
[
] is the
its dynamics are to be defined, and
unavailable vector. From (2) and (3) the dynamics of can
be written as
̇
[
]
(4)
with
(
[
̇)
]
where
and
are the zero matrix and zero
vector, respectively. Although using the vector filtered
transformation ( ) for such systems like in [5, 6] facilitate
the state estimator design, the limitation imposed by the PE
condition is still remained in the parameter estimation. To
overcome this problem the matrix filtered transformation is
introduced as
[
where
()
(5)
]
, in which
(
is identity matrix,
)
( )
,
,
is
an auxiliary dynamic matrix, which its dynamics are to be
[
] is the new unavailable vector. In
defined, and
turn, the dynamics of can be obtained as
̇
[ ]
(6)
̇
[
(
(
where
( )
)
̇)
(
]
)
.
and
(
)
As will become clear in the next section, the derivation of
the ( )-dynamics and the ( )-dynamics follows the idea
in [5, 6]. In the proposed approach, the advantages of I&I
technique with input-output filtered transformation is
combined to construct a proper immersion and an auxiliary
dynamic filter for the estimation purpose.
In order that the state be observable from the output ,
the following assumption is imposed.
Assumption 1: The system (1), (2) is forward complete,
i.e. trajectories exist for all
. The function ( ) and
its time derivative are bounded and | |
for all
and
.
A. Filtered Transformation via Dynamic Vector
This subsection is devoted to summarize the method
proposed in [5, 6] with some modifications. The change of
coordinate (3) and the dynamic equation (4) are considered
for design methodology in this subsection. Following the I&I
procedure [8] and [5, 6], define the estimation error as
(
)
(7)
where
is the estimator state, the dynamics of which
is to be defined, and the mapping
yet to
be specified. To obtain the dynamics of , differentiate (7),
yielding
̇
̇
̇
̇
̇
̇
Let
(
)[
]
(
)
(8)
]
]
)
(10)
[
], the
From (7) and the definition of
[
] with
estimation error can be decomposed into
and
. Also, with respect
to (3), the state estimation error can be expressed as ̅
. The following proposition establishes that there exists
the mapping and the dynamics of the filtered vector such
that the state estimation error ̅ converges to zero.
Proposition 1:
Assumption 1 with
(
)
Consider
( )
[
(
)
| |
̇
the
system
(10)
under
]
( )[
(11)
] ( )
(
(
)
)
(12)
where
and
are constant and
( )
positive definite matrices,
is a strictly
increasing function with bounded derivative. Then the system
(10) has a uniformly globally stable equilibrium at the origin,
and ̅
. If, in addition, ( ) is bounded, then
( ) is bounded and ̅ converges to zero.
Proof 1: Notice that the derivation of the signum
( ) is zero, due to the observation restriction
function
()
. Substituting (11), (12) into (10) yields the error
dynamics
̇
( ̇
[
)
]
(13)
| |
[
]
( ) has been used to get the last equality.
where | |
Consider now the following Lyapunov function candidate
( )
(
)
(14)
whose time-derivative along the trajectories of (13) satisfies
̇
where ̇ is given in (4), and the output dynamics ̇ can be
rewritten in term of , as
[
(9)
[
(
III. ESTIMATOR DESIGN
In this section, two classes of estimators are designed for
the system (1), (2). Although, the first one facilitates
estimator design using a filtered transformation via dynamic
vector, it requires the regressor persistency of excitation for
the parameter estimation. It is worth pointing out that the
second one ensures parameter convergence with PE
dismissal.
)
where ̇ is to be defined. Replacing (9) in (7), yields
̇
(
)(
̇
with
[
]
| |
(
| |
(
)
(
) )
(15)
for some constant
, where the positive definite
property of has been used to get the bound. The equations
(14) and (15) imply that the system (13) has a uniformly
globally stable equilibrium at the origin,
,
,
and then ̅
. Note that the above holds for
any . It is clear that, (12) is input-to-state stable dynamic,
then ( ) and its time derivative remains bounded for any
⁄
bounded . Consequently, , , ̇, and
being
̈
bounded, ̇ in (13) and then
in (15) are bounded.
Therefore, the convergence of ̅
to zero follows
directly from Lemma 1 in [9].
] converges to zero, an
Remark 1: If ̅ [
](
).
asymptotic estimate of is given by ̂ [
As is well known ([10], Lemma B.2.3), the parameter
estimation error
converges to zero if the regressor vector
( ) and its time derivative are uniformly bounded and the
regressor satisfies PE condition. Notice that, the bounded
condition is satisfied here for any bounded .
Remark 2: In comparison with the estimator proposed in
[5, 6], the gain matrix is added to filter dynamics in (12)
which can help to make the regressor exciting over a time
sequence arbitrary.
Remark 3: In comparison with the mapping in [5, 6],
the gain matrix is added in (11) and then in error dynamics
(13). The system (13) is in the form of damped nonlinear
oscillator in which the transient time can be improved
arbitrary by reducing the overshot via tuning.
B. Filtered Transformation via Dynamic Matrix
To dismissal the PE condition in parameter estimation,
the design methodology in this subsection is based on the
change of coordinate (5) and the dynamic equation (6). In
the light of the proposed design procedure in the previous
subsection, a general expression of ( ) can be expressed
by
(
(
)
( )[
)
( )
( (
)
( (
)[
(
| |
( )(
(
̇
| |
(
)
[
)
̇
(
(18)
))
.
](
[ ]
(21)
[
], the
From (16) and the definition of
estimation error can be decomposed into
and
. Also, with respect to (5), the state
estimation error of
can be stated as ̅
. A
choice of and
which achieves the design objective is
given in the following proposition.
Proposition 2: Consider the system (21) verifying
Assumption 1, with
( )
( )
( )
(22)
( )
(23)
where ,
and ( ) are chosen as in Proposition 1, in
addition, suppose that the eigenvalues of
are different.
Then the system (21) has a uniformly globally stable
equilibrium at the origin,
and ̅
. Moreover, if
( ) is bounded, then ( ) is bounded and ̅ converges to
zero. If in addition
( )
(resp.
( )
), then
converges to zero asymptotically (resp. exponentially).
Proof 2: Consider the following Lyapunov function
candidate
( )
(
)
(24)
whose time-derivative along the trajectories of (21), is
̇
| |
(17)
))
]
[
]
[
]
(25)
Replacing (23) and (24) in (25), we obtain
] ( )
( )
( )
where
choosing the update laws as
̇
[
(16)
where ( )
is to be defined, and ( )
( )
is a vector function in which, ( ) is defined
in Proposition 1. Notice that, the dynamics of the vector
function ( ) can be obtained in term of , as
̇
| |
̇
Now,
by
[ ]
] ( ))
)
(19)
̇
| |
(
)
(26)
which implies that the system (21) has a uniformly globally
stable equilibrium at the origin,
,
,
and then ̅
. The above holds for any
. Inserting
(23) and (24) into (20) yields
̇
| |
(
)
(
)
(27)
which is input-to-state stable with respect to perturbation
( ). Therefore, the dynamic equation (26) with
̇
induces
. On the other hand, for bounded ,
⁄ , boundedness of ̇ and ̈ can be obtained
, ̇, and
from (21)-(23) and (26) respectively. As a result, the
convergence of
and then ̅ to zero is guaranteed by
Lemma 1 in [9]. Note now that
(28)
(20)
and with respect to (6), (16)-(18), the resulting error
dynamics can be obtained as
for some
, where use has been made of positive
definite property of . For any ( )
, the expression
( ) ( ) is symmetric, and its eigenvalues are real and
nonnegative [11]. Also, if ( ) are the eigenvalues of ( ),
then ( ) are the eigenvalue of ( ) ( ), for all
. Let
, then
( ) be the minimal value of
( )
( )
( )
( ) for any
(29)
On the other hand, it is well known the determinant of a
square matrix is the product of all its eigenvalues. As a
( )
result, if
( )
then
. Now using (26)(29) and doing some basic bounding, yields
̇
| |
(
Remark 4: Replacing (23) and (24) in (21) yields to the
following error dynamics
[
IV. SIMULATION EXAMPLE
(30)
(
). It is clear that
for some
and ( )
exponential convergence of , and then can be directly
obtained from (30) due to the positiveness of ( ). Note
( ), if
now that for bounded
( )
then all
( ) ( )
eigenvalues ( )
[1]. As a result,
and from (30) asymptotic stability of
and then
is
guaranteed.
| |
Remark 8: Notice that, a sudden change in the system
parameters distorts the output signal in (1), (2). This can be
transferred to the
( )-dynamics in (27) by an output
⁄ . Then, the
feedback which appears in the term
dynamics are actuated to adapt ( ) with the changes, even
if the regressor ( ) is not dependent on the output signal,
i.e. ( ).
)
( )
̇
Generally, a necessary condition for any positive definite
matrix in (27) is that its eigenvalues be different.
]
in which the effect of overshoot induced by
be reduced arbitrarily by tuning .
Consider the system (1), (2) with
[
] and
( )]. Assume that
[
( )
is derived from the following equations
(
Remark 5: If
converges to zero, an
asymptotic estimate of is given by ̂
. Also, If
̅ [
] converges to zero, an asymptotic estimate of
). In turn, with respect to
is given by ̂ [
](
the definition of
, an estimate of unmeasurable state is
given by ̂
̂.
C. Discussion
The main question arises at this method is that, for a
( ) how to fulfill the condition
given regressor
( )
. The answer lies in the ( )-behavior tuning.
With respect to the first order differential equation (27), the
behaviour of ( ) consists of transient component and
steady state component. These components can be affected
by arbitrarily chosen of a possible output feedback, induced
by the term ⁄ , and the gain . In addition, the transient
component also depends on the initial value of ( ), i.e.
( ). Hence, some notes are remarked following.
Remark 6: In order to fulfil the condition
( )
, it
(
))
is reasonable to choose ( ) such that
(
.
However, the effect of this choice fades away over time.
Remark 7: The equation (27) becomes a simpler
differential equation if the positive definite matrix
becomes diagonal. In this case, (27) can be decomposed to
individual differential equations. Now, to satisfy the state
( )
, a necessary condition is that the entries on the
main diagonal of be different. This yields different steady
state components for each individual differential equation.
̇ ( )
( )
̇( )
(
)
(
( )
) ( )
(32)
(33)
where
are positive constant and
, and
( ) belongs to the following set of differentiable function
{
(31)
can
)
( )
| ̇( )
()
( )
()
( )
(34)
The set
is borrowed from [1] with some modifications.
( )
For any
, the condition ( ) ( )
insures
and then ( ) remains bounded. Also, the condition
()
( )
( ) and then ( ) is not PE
yields
because it converges to zero as
.
⁄
Let us assume
in (12) and (27). Then, the
steady state component of (12)-solution for
and
can be obtained as
[
[
(
)
(
)
( )]
(35)
( )
( )]
(36)
respectively. With respect to Remark 7, choosing
(
) for (27) yields to
[
where
]
(37)
is the steady state component of (27)-solution.
Remark 9: The regressors
because
( )
( )
(
)
, and
(
)
due to positive semi definite property of the
matrixes. Notice that, the matrix
, but
(
)
because
(
in which, by (34), ̇ ( )
)
̇( )
(
)(
)
.
A. Simulation Results
Consider the system introduced in this section with
(38)
( )
√
(39)
[3]
which
[1]. Simulations are run to evaluate the
performance of estimators introduced in Proposition 1 and 2.
All initial values correspond to the system and the estimators
are set to zero.
The time histories of the parameters estimation errors
̂ and the state estimation errors
̂ are shown for
estimator b1 (dashed line), estimator b2 (dashed-dot line),
and estimator B (solid-line) in Fig. 1. The estimator bi is
formed by replacing
in Proposition 1 and the
estimator B is formed by replacing
(
) in
Proposition 2, while the mapping ( )
and the gain
are chosen for all of them. The graphs in Fig. 1 show
)
the performance of the estimators for the set (
(
).
[4]
[5]
[6]
[7]
[8]
From the graphs in Fig. 1 it can be seen that, for
,
a bigger value of
has a better response in the parameter
estimation. In the all graphs, the estimator B has a faster
response in parameter and state estimation. The last graph of
Fig. 1 shows the absolute value of the ( )-determinant.
From (38) and (39) the determinant of ( ) is not square
integrable which is the core of convergence property of the
estimator B.
[9]
[10]
[11]
regressor extension and mixing," IFAC-PapersOnLine, vol. 49,
pp. 99-104, 2016.
A. Bobtsov, D. Bazylev, A. Pyrkin, S. Aranovskiy, and R.
Ortega, "A robust nonlinear position observer for synchronous
motors with relaxed excitation conditions," International
Journal of Control, vol. 90, pp. 813-824, 2017.
R. Marino and P. Tomei, "Global adaptive observers for
nonlinear systems via filtered transformations," IEEE
Transactions on Automatic Control, vol. 37, pp. 1239-1245,
1992.
M. Tavan, A. Khaki-Sedigh, M.-R. Arvan, and A.-R. Vali,
"Immersion and invariance adaptive velocity observer for a class
of Euler–Lagrange mechanical systems," Nonlinear Dynamics,
vol. 85, pp. 425-437, 2016.
M. Malekzadeh, A. Khosravi, and M. Tavan, "Immersion and
invariance Based Filtered Transformation with Application to
Estimator Design for a Class of DC-DC Converters,"
Transactions of the Institute of Measurement and Control, vol.
(Accepted), 2018.
R. Brockett, "The rate of descent for degenerate gradient flows,"
in Proceedings of the 2000 MTNS, 2000.
A. Astolfi, D. Karagiannis, and R. Ortega, Nonlinear and
adaptive control with applications: Springer Science & Business
Media, 2007.
G. Tao, "A simple alternative to the Barbalat lemma," IEEE
Transactions on Automatic Control, vol. 42, p. 698, 1997.
R. Marino and P. Tomei, Nonlinear control design: geometric,
adaptive and robust: Prentice Hall International (UK) Ltd.,
1996.
C. D. Meyer, Matrix analysis and applied linear algebra vol. 71:
Siam, 2000.
The next simulation is run to show the influence of the
gain
on the performance of the estimator B. All
conditions and gains are the same as previous simulation.
The time histories of the parameters and the state estimation
errors are shown for
(dashed line),
(dasheddot line), and
(solid-line) in Fig. 2. As it is
evident in Fig. 2, the overshot is decreased by increasing the
value of , and as a result, faster responses in the both
parameter and state estimation are obtained.
V. CONCLUSION
A new I&I based filtered transformation is introduced via
dynamic matrix for a class of second-order systems with
single-output. The main contribution of this paper is to
propose an estimator based on the new transformation,
which asymptotically estimates unavailable state and
unknown parameters. Parameter convergence is guaranteed
under non-square-integrability assumption for the matrix.
Moreover, it is shown that the estimator provides an
arbitrary convergence speed. Also simulation results are
concluded for a class of regressors which are not PE but
satisfy the new condition. Some practical results on
electrical systems are under way to show effectiveness of the
proposed design.
Figure 1. Transient performance of the estimator bi and estimator B.
REFERENCES
[1]
[2]
S. Aranovskiy, A. Bobtsov, R. Ortega, and A. Pyrkin,
"Performance enhancement of parameter estimators via dynamic
regressor extension and mixing," IEEE Transactions on
Automatic Control, vol. 62, pp. 3546-3550, 2017.
S. Aranovskiy, A. Bobtsov, R. Ortega, and A. Pyrkin, "Improved
transients in multiple frequencies estimation via dynamic
Figure 2. Transient responces of the estimator B for changs in .
| 3cs.SY
|
Towards Data Quality Assessment in Online Advertising
Sahin Cem Geyik 1,+,∗ , Jianqiang Shen 2,∗ , Shahriar Shariat
Ali Dasdan 4,+ , Santanu Kolay 2
arXiv:1711.11175v1 [cs.AI] 30 Nov 2017
2
3,+
,
1
LinkedIn Corporation, [email protected]
Turn Inc., {jianqiang.shen, santanu.kolay}@turn.com
3
Uber, [email protected]
4
Vida Health, [email protected]
ABSTRACT
1. INTRODUCTION
In online advertising , our aim is to match the advertisers
with the most relevant users to optimize the campaign performance. In the pursuit of achieving this goal, multiple
data sources provided by the advertisers or third-party data
providers are utilized to choose the set of users according to
the advertisers’ targeting criteria. In this paper, we present
a framework that can be applied to assess the quality of such
data sources in large scale. This framework efficiently evaluates the similarity of a specific data source categorization
to that of the ground truth, especially for those cases when
the ground truth is accessible only in aggregate, and the
user-level information is anonymized or unavailable due to
privacy reasons. We propose multiple methodologies within
this framework, present some preliminary assessment results, and evaluate how the methodologies compare to each
other. We also present two use cases where we can utilize the data quality assessment results: the first use case is
targeting specific user categories, and the second one is forecasting the desirable audiences we can reach for an online
advertising campaign with pre-set targeting criteria.
Online advertising strives to serve the most beneficial advertisement (ad) to the most relevant online users in the
appropriate context (a specific website, mobile application,
etc.). This typically results in attaining higher return-oninvestment (ROI) for the advertisers [10], where the value
is generated either from a direct response such as a click or
conversion (e.g. the purchase of a product, subscription to
a newsletter, etc.), or through delivering a branding message. For this purpose, advertisers receive help from multiple entities in the domain. Supply-side platforms (SSP)
provide ad-space (inventory) on websites or mobile apps, to
serve ad impressions to users. Ad-exchanges run auctions
on available inventory from SSPs. Demand-side platforms
(DSP) act on behalf of the advertisers and aim to bid on
the most valuable inventory. Advertisers often get performance reports from an independent evaluation agency # .
For privacy reasons, these reports, in most cases, only contain aggregate metrics (e.g. click-through rate, percentage
of female audiences).
In order to reach the right audience usually defined by the
advertiser, which in general would improve direct response
and branding metrics, the advertisers need to utilize various data sources to label the users in the most accurate way
possible. Data management platforms (DMP) have been
emerging as a central hub to seamlessly collect, integrate
and manage large volumes of user data [6]. Such user data
could be first-party (i.e. historical user data collected by advertisers in their private customer relationship management
systems), or third-party (i.e. data provided by third-party
data partners, typically each specializing in a specific domain, e.g., demographics, credit scores, buying intentions).
While first-party data is proprietary to the advertiser and
free to utilize, third-party data often carries a pre-negotiated
cost per impression (ad served to a user in a website or application). In both cases, it is important for the advertiser
to know how accurate a data source is. That is, if a data
source has tagged a user to be in category ci (user property,
e.g. gender, age, income), how likely it is for the user to
actually be in that category.
In this paper, we are investigating the above problem
which we call data quality assessment in online advertising.
The main issue in evaluating the accuracy of a data source
Categories and Subject Descriptors
H.3.5 [Information Storage and Retrieval]: Online Information Services; I.2.1 [Artificial Intelligence]: Applications and Expert Systems
General Terms
Algorithms, Application
Keywords
Online Advertising; Data Quality; Targeting; Forecasting
∗
The authors contributed to this work equally.
+
This work was completed when the authors were at Turn.
#
This work has been presented in the KDD 2016
Workshop on Enterprise Intelligence.
There are certain independent evaluation agencies in online advertising domain, whose names we cannot list here
to comply with the company policy. Advertisers trust these
organizations to collect the ground truth.
Table 1: Confusion matrix of data source s for tagging users with category c.
Ground Truth
Predicted by data source s
c
not c
Unknown
c
n+,+
n+,−
n+,
not c
n−,+
n−,−
n−,
n
n
n
Unknown
,+
,−
,
is the lack of ground truth in the user-level granularity. For
example, the advertisers, in reality, never have access to the
confusion matrix (Table 1) of a data source in either first or
third-party cases. Therefore, the only way for an advertiser
to evaluate the quality of a data source is to run an advertising campaign on a set of users and then evaluate the performance in hindsight. Even in those cases, the post-campaign
data is often constrained (mostly due to privacy concerns)
and in aggregate, that is, only the total number of users in
different categories is provided, and not a granular user-tocategory assignment. If it were possible to have the granular
data, it would then be trivial to just use the ground truth
data source to come up with the accuracy metrics, e.g. filling
in the entries of the confusion matrix. Therefore, utilizing
the aggregate performance statistics makes the data quality evaluation task quite challenging, and somewhat similar
to aggregate learning tasks in machine learning [5], few of
which are also directly applicable to this problem.
The main contributions of this work are as follows:
• formal definition of data quality assessment problem,
and the challenges of solving it in online advertising
domain,
• multiple approaches for evaluating the quality of a data
source, which also take into account the efficiency requirements due to the large number of possible data
sources∗ to be evaluated,
• several use cases where data quality assessment comes
in handy for online advertising, and,
• initial evaluation of our methodology utilizing simulated
data and real-world advertising campaigns.
Rest of the paper is organized as follows. In Section 2, we
give a more formal definition of the data quality assessment
problem. Next, we discuss the literature that deals with either data quality assessment, or aggregate learning (which,
as aforementioned, is relevant to our problem) in Section 3.
We present our two proposed assessment methodologies in
Section 4 and Section 5, and later give some use cases on
how we can utilize our data quality assessment output in
Section 6. Finally, we present some initial results in Section 7 and conclude the paper along with some potential
future work in Section 8.
2.
RESEARCH PROBLEMS
As we have explained in the previous section, we seek to
evaluate first or third-party data sources available for online
advertising using multiple accuracy metrics.
* While we cannot list the exact number due to the company
policy, there are currently over 200k active data sources in
Turn’s system.
Definition 1. A sound data source tags each virtual user
(cookie ID that might be specific to a browser and device)
with one and only one of the 3 labels – {Positive, Negative,
Unknown}.
The tagging process could be explicit or deductive, but cannot be self-contradictory. For example, a user can have a
positive tag – Age25, or a negative tag – NotAge25, but
cannot be tagged as both Age25 and NotAge25. The data
source can also simply indicate that it has no knowledge on
a user by tagging it as Unknown. In real-time bidding, the
positive tags are the most important, as advertisers usually
utilize them to target the desirable audiences.
The problem of data quality assessment is defined as the
following:
Definition 2. Given a sound data source S, its data
quality assessment is defined as a measurement mS that has
error no more than ǫ over user examples drawn from user
set U , with probability of at least δ,
P ( Eui ∈U Ω(ui , Si ) − mS ≤ ǫ) ≥ δ,
where Ω(ui , Si ) is a metric to measure the granular targeting
performance when this data source tag user ui with Si .
As an example, suppose we have a data source which we
utilize to tag a user as Male (positive example) or Not Male
(negative example). Consider two evaluation metrics, which
are Accuracy (percentage of correct taggings by our data
source), and True Positive Rate (percentage of positive examples, i.e. Males, that our data source also tags as males).
For the accuracy metric, we have the following Ω:
Ω(ui , S) = I(GT (ui ) == Si ),
which does an exact comparison of the ground truth tagging
of user ui (GT (ui )) against the tagging by data source S
(Si ). On the other hand, if we were to calculate true positive
rate, then we would have the following Ω:
Ω(ui , Si ) = I(GT (ui ) == Si == Male),
which counts only those cases where both ground truth and
the data source tag the user as Male.
Note that the above problem definition is a very general
formulation, which is typically used in evaluating Machine
Learning models [7, 8, 12]. As long as both the ground truth
category of a user and that of the data source are available,
one can
come up with a perfect data quality assessment, i.e.
Eui ∈U Ω(ui , Si ) ≈ mS from Def. 2. The problem occurs
when we don’t have direct access to the ground truth category of every single user. Typically, Ω(ui , Si ) is unknown,
but rather the category distribution of groups of users is provided. The main reason is to protect the privacy of users [1].
In these kind of situations, especially in online advertising,
we may utilize a specific data source to make smart advertising decisions to choose the most appropriate set of users,
and in the end, we can receive an aggregated report from
a third-party evaluator, which is considered as the ground
truth and provides a non-granular distribution of the audience we have reached over many categories of interest. As
an example, the report may provide that over all users we
have 20% Male, and 80% Not Male. When this occurs, we
can no longer do a one-to-one comparison between ground
truth and data source in the user granularity, but rather
need to come up with alternative methods that can deal
with aggregated data, which is our main focus in this paper.
In many cases, we need to select the best data source
from a large set of candidates with the same semantic goal
and adopt it for targeting. For example, given a set of data
sources that tag users as male, female, or unknown, we may
care more about their relative performance and less about
their absolute measurements. The data quality assessment
can then be simplified as a ranking problem:
Definition 3. Given two sound data sources S 1 and S 2 ,
and an accuracy metric Ω, a data quality ranking system
outputs a rank measurement r 1 for S 1 and r 2 for S 2 such
that
r 1 > r 2 ⇐⇒ Eui ∈U Ω(ui , Si1 ) > Eui ∈U Ω(ui , Si2 ) .
Once we have the rank measurements for each sound data
source, we can order them and select the best one.
3.
PREVIOUS WORK
As we have explained in the problem definition, evaluation
of a data source can be taken as any other machine learning
model evaluation task, provided that we have the ground
truth information in the user granularity. A detailed evaluation of 18 performance metrics for classification problems
is given in [7]. These 18 metrics can be listed as accuracy,
kappa statistic, mean F-measure, macro average arithmetic,
macro average geometric, AUC of each class against the rest
(two variants), AUC of each class couples (two variants),
scored AUC, probabilistic AUC, macro average mean probability rate, mean probability rate, mean absolute error, mean
squared error, LogLoss, calibration loss, and calibration by
bins. The paper provides a detailed correlation analysis and
noise sensitivity analysis . Also, the survey by Gunawardana
et al. [8] discusses both the evaluation settings and proper
evaluation metrics for different classes of recommendation
problems, of which online advertising is a sub-problem.
When we only have access to aggregated ground truth
data, evaluation of a data source is much harder. There has
been significant work in aggregate learning tasks which utilize aggregate assignments of classes to groups of samples
to train a model. Our aim in this paper is significantly different from such works, since we already have a model (i.e.
data source), and we are trying to evaluate its performance
utilizing many campaigns and multiple aggregates of ground
truth data. Cheplygina et al. [5] provides an overview of aggregate learning methodologies, which may utilize granular
response variables/feature vectors (single instance) or aggregate response variables/feature vectors for groups (multiple
instance) to train their models, and later, testing them. Musicant et al. [11] utilizes aggregate outputs for the response
variables to specialize the training process of k-nearest neighbors, decision trees, and support vector machines. In [3], the
authors utilize aggregate views of data, which consist of a
choice of different combinations of features, response variables, and combining machine learning models learned from
these views. Another interesting work is presented in [16],
which gives error bounds on how a model learned from aggregate data can perform. They assert that a machine learning
model should minimize empirical proportion risk, and prove
that under certain assumptions for the class distributions,
learning in the aggregate setting can actually improve individual classification performance.
Finally, specific to the online advertising domain, we can
list [14] as being a relevant work to ours. In this paper,
similar to aggregate learning techniques, the authors aim
to learn a predictive model to decide whether a user is in
a specific ground truth category using the aggregate data
over many campaigns, by assigning the most likely label to
all users in the aggregate, or assigning a probabilistic single
label. They utilize logistic regression with L2 -norm regularization, where the response variables are the artificially
generated labels.
4. BRUTE FORCE EVALUATION
In this section we will present our first proposal for data
quality assessment, which includes setting up specialized
campaigns for a data source and utilizing the targeting results directly for evaluation.
Note that we typically rely on the independent survey
agencies to collect the ground truth analysis data on our
audience population. Such agencies use offline data (such as
credit card information) and online data (such as information filled in social networking websites) to profile an Internet user. Reports from these survey agencies are generally
considered as the ground truth by advertisers. Such reports
are aggregated statistics and contain no user-level information due to privacy reasons.
4.1 Performance Campaign for Data Source
An intuitive and straightforward way to evaluate a data
source is to set up a campaign that only targets certain
users which are tagged by data source S to be in category
c. This way we can calculate the quality of the data source
N(c )
as p(cg |cs ) = N(cgs ) , where N (cg ) is the number of users
in category c reached by this campaign as reported by the
ground truth and N (cs ) is the total number of users reached
by the campaign via at least one impression. Note that we
can put a limit on the number of impressions to be served
to a user so that we can increase the unique user reach and
have more reliable results.
We applied this methodology to evaluate age and gender
categorizations of some well-established data providers in
online advertising. Table 2 demonstrates some results for
one of the better performing such data providers in age categorization. We anonymize the name of the data provider
and exact age ranges listed in the table to comply with the
company’s regulations. In Table 2, R1,· · · · R10,· represent
the age ranges such that they are mutually exclusive and
sorted in ascending order, e.g. R5 is the range that is the
immediate higher range after R4 (i.e. minimum age in range
R5 is one larger than maximum age in range R4 ), and the immediate lower range before R6 . Ri,p represents the predicted
results by the data source for age range i, while Ri,g stands
for ground truth data for the same age range. For example,
we can observe from the table that p(R4,g |R4,p ) = 0.345,
that is, when our data source classifies the user to be in age
range R4 , 34.5% of the time it is correct, or in other words,
34.5% of the users reached by campaign that targets category R4 , as predicted by the data source, were actually in
category R4 , as provided by ground truth. Note that for
this particular data source, the exact match, highlighted in
bold, is quite high compared to random.
Data providers utilize various models and online/offline
information to tag users. Please note that data sources from
Table 2: Ground truth distributions of anonymous data provider’s category taggings.
Age Ranges
R1,p
R2,p
R3,p
R4,p
R5,p
R6,p
R7,p
R8,p
R9,p
R10,p
R1,g
0.414
0.058
0.031
0.031
0.037
0.056
0.08
0.065
0.05
0.035
R2,g
0.245
0.297
0.053
0.041
0.057
0.049
0.067
0.074
0.044
0.036
R3,g
0.033
0.246
0.298
0.089
0.056
0.061
0.054
0.082
0.066
0.055
R4,g
0.018
0.037
0.227
0.345
0.063
0.041
0.035
0.043
0.065
0.048
the same data provider may have quite diverse accuracy values. For example, the accuracy results in Table 2 range
from 0.29 to 0.49. Thus we cannot evaluate one data source
and assume its sibling data sources have similar predictive
power. Each data source needs to be evaluated individually. This causes a significant disadvantage with the above
methodology, which is the fact that we need to set up a separate campaign for each category so that we can gather the
accuracy statistics. To remedy this problem, we propose
a second methodology in Section 5, which solves an optimization problem to come up with the best fitting accuracy
probabilities based on the aggregate reports.
4.2 Cost Analysis
In this subsection, we further analyze the cost of the brute
force method discussed in Section 4.1. It must be noted that
obtaining the ground truth, that is the aggregated labeled
data, is costly on its own right. However, in the following
lemma, we focus only on the ad serving costs to underline
the utility and benefit of our approach.
Lemma 1. Given a data source that can tag the user by
one of the possible c categories (e.g. the data source gives a
positive/negative output on one age group of possible c age
groups), then to observe a significant difference between the
calculated accuracy of one category versus others, we need
at least ⌈ 4202.969
(1 − 1c )⌉ impressions.
c
Proof. Assuming a uniform distribution, we assume the
average on-target rate is 1c for each data source (although the
intention of a data source is to increase this value). Each impression can be considered as a Bernoulli trial with 1c probability of success. The sample variance is, thus, 1c (1 − 1c ).
We would like to detect a significant difference between the
prediction accuracy of the correct category versus the rest
of possible tags. The industry standard accepts a 5% error.
Then, for a significance level of α = 5% for a two tailed
hypothesis
test and to attain at least 90% power, we have
√
n
√0.05
>
(z0.975 + z0.9 ), where n stands for the number
1
1
c
(1− c )
of users, z0.975 and z0.9 are the values of the quantile function of the standard normal distribution for 97.5% and 90%,
respectively [2]. Therefore, the number of users that receive
the ad impressions must be more than ⌈ 4202.969
(1 − 1c )⌉ for
c
each data source.
Based on Lemma 1, for d data sources that provide information on one of the possible c categories, we need at least
d⌈ 4202.969
(1 − 1c )⌉ impressions. As we discussed before, it is
c
necessary to evaluate each data source individually, considering the diversity of their predictive power. Given the very
large number of data sources, this causes the brute-force
approach to incur a very significant cost.
R5,g
0.012
0.021
0.049
0.182
0.337
0.053
0.03
0.03
0.048
0.039
R6,g
0.032
0.043
0.042
0.057
0.212
0.339
0.041
0.04
0.033
0.048
R7,g
0.043
0.051
0.042
0.037
0.047
0.23
0.332
0.048
0.048
0.061
R8,g
0.064
0.089
0.049
0.039
0.038
0.041
0.204
0.339
0.044
0.06
R9,g
0.075
0.091
0.132
0.116
0.082
0.065
0.077
0.203
0.488
0.106
R10,g
0.03
0.046
0.054
0.041
0.05
0.046
0.048
0.052
0.101
0.492
5. ACCURACY INFERENCE
High-quality data sources can enable advertisers to reach
the right audience at the right moment. Because they have
become an important component of online advertising, more
and more online/offline data are being ingested into Turn’s
data management platform. As we have mentioned previously, there are currently over 200k active data sources in our
system. Lemma 1 established that explicitly evaluating each
of these data sources by running performance campaigns is
overwhelmingly costly: not only a large amount of money is
required to run the performance campaigns, but also enormous manual effort to set up and manage those campaigns is
essential as well. We need an efficient way to simultaneously
infer the accuracy of multiple data sources.
As we have presented in Section 2, our focus in this paper
is to calculate the accuracy metrics of a data source for single
or multiple categories. In essence, we are trying to calculate
a set of probabilities, which represent the likelihood of a data
source predicting correctly/incorrectly that a user belongs
to a category ci . In Figure 1, we have shown the set of
probabilities that we aim to predict. For representational
purposes, we have shown the accuracy probabilities of a data
source which denote its capabilities to tag a user as Male or
not, though the same logic follows for any category. The
probabilities in the figure can be listed as follows:
• α1 : The probability of the user actually being Male
when the data source tags it as Male. This value can
also be called precision or positive predictive value.
• α2 : The probability of the user actually being Not Male
when the data source tags it as Male. This value can
also be called false discovery rate.
• α3 : The probability of the user being Unknown (i.e.
ground truth does not exist) when the data source tags
it as Male.
• β1 : The probability of the user actually being Male
when the data source tags it as Not Male.
• β2 : The probability of the user actually being Not Male
when the data source tags it as Not Male. This value
can also be called negative predictive value.
• β3 : The probability of the user being Unknown (i.e.
ground truth does not exist) when the data source tags
it as Not Male.
• γ1 : The probability of the user actually being Male
when the data source tags it as Unknown.
• γ2 : The probability of the user actually being Not Male
when the data source tags it as Unknown.
• γ3 : The probability of the user being Unknown (i.e.
ground truth does not exist) when the data source tags
it as Unknown.
As it can be seen from the figure, and trivial from the defini-
α1
Male
D1 by Data Source S
α2
Male
by Ground Truth
N1
Not Male
by Ground Truth
N2
α3
β1
Not Male
D2 by Data Source S
β2
β3
γ1
D3
Unknown
by Data Source S
γ2
γ3
Unknown
by Ground Truth
N3
Figure 1: Objective probabilities of a data source S
for a specific category (is user Male).
We log the received report data along with the first-party
campaign data into our in-house data warehousing system
called Cheetah [4], which is built on top of the Hadoop
framework [13]. Cheetah is designed specifically for our online advertising application to allow various simplifications
and custom optimizations. Campaign facts are stored within
nested relational data tables. Fast MapReduce jobs are designed to extract key features of the performance campaigns,
compare them with the ground truth and infer the accuracy
of a data source. Utilizing the collected campaign information, we present two approaches in this section to efficiently infer the quality of data sources: one that ranks data
sources, and another which directly deduces precision (α1 )
of a data source.
5.2 Ranking Based Assessment
tions, we have α1 +α2 +α3 = β1 +β2 +β3 = γ1 +γ2 +γ3 = 1.
Among these nine variables, α1 , i.e. precision, is often the
most important value for the advertisers, since it denotes
the goodness of a data source to be used for their advertising purposes. To calculate this value, we presented the
methodology, which is based on creating specific campaigns
to evaluate a singular data source for a specific category in
Section 4. In this section we are proposing an optimization
scheme, which utilizes the aggregated category distributions
over multiple campaigns, for both the data source we want
to evaluate, and the ground truth. In the following subsections, we call the above nine variables predictive values of a
data source.
5.1 Setup for Inference
We propose to set up multiple performance campaigns
without using any data source for targeting, so the audience will not be explicitly skewed by any data source. We
compare the ground truth of each campaign against the hypothesis of each data source, and infer the quality of the
data source.
We follow a set of rules to set up a performance campaign.
First, the targeting criteria should be minimal and cannot
be biased by any third-party data. For example, targeting
the online users in U.S. is fine, since this is purely based on
IP address and not biased; however, using a data source to
limit audience to middle-aged men is not acceptable as the
quality of this data source is what we want to assess. In general, only geographical location should be used as targeting
criteria. Second, the targeted websites must be discriminative, that is, the population of the visiting users should be
largely skewed towards one of our possible tags. This way,
we will not mistakenly estimate a data source to be accurate,
when in fact it is predicting the label in a random manner.
For example, a website is beneficial for such an experiment
if 70% of visitors are female and 30% of visitors are male,
and not necessarily if the distribution is 50% to 50% (since,
in such a case, a random prediction of female vs. male will
closely fit the overall audience in aggregate). One should
note that obtaining such knowledge is not always feasible
before running the campaign. Therefore, in our system, we
target only the websites that are known, based on our domain experience and verified by independent reports, to be
popular among a certain group of audience. After creating
such a campaign, we run it for a certain period to collect
data. The ground truth is collected through independent
agencies as described in Section 4.
Ranking Data Sources. In many instances, we only
need to choose the best data source from a large set of candidates with similar semantic purposes. If this is the case,
then data quality assessment becomes a ranking problem.
A data source’s absolute precision α1 is of less importance
then, and rather its rank among others is critical.
Since the independent evaluation agency sends us the aggregate statistics on a campaign, we can similarly construct
such statistics using a data source. Note that this approach
will only represent the view of the data source, and not
the ground truth, unlike the independent agency case. We
can then evaluate this data source based on how close the
constructed statistics are to that of the ground truth, and
therefore rank data sources based on such closeness measure.
This logic is presented in Algorithm 1.
Algorithm 1: Measurement calculation for ranking.
Input: ψ: aggregation function, f : closeness function
Input: S: data source
Output: µ̄: score for quality of data source S
1 foreach performance campaign C do
2
U ← Retrieve audience of C;
3
R̂ ← ψ(U, S);
4
R ← Retrieve the ground truth report;
5
µC ← f (R̂, R);
6
7
µ̄ ← Calculate the average value of µC ;
return µ̄;
There are multiple ways to design the closeness function f
to compare two aggregated statistics. Since the positive taggings are the most valuable for online advertising purposes,
we propose to compare the positive population distributions
between the ground truth and the data source. We define
the percentage of population marked as positive by a data
source as
count+ (S, U)
R̂ =
,
(1)
count+ (S, U) + count− (S, U)
where count+ counts the number of users in U marked as
positive by S, and count− counts negatives. Given that R is
the ground truth ratio of positive population, a simple way
to calculate the closeness can be defined as:
R̂ − R .
(2)
However, this does not consider the scale of R̂ or R, which
are usually quite small for a rare positive group. To make
the measurements more comparable, instead we propose to
calculate the relative error as the closeness:
ˆ + (S, U)
count+ (U) − count
,
RelativeErr =
count+ (U)
two predictive values. Since the data sources are unbiased,
the order of metrics for negatives are also preserved. Averaging is monotonic, therefore we can expand the previous
statements to micro and macro averaging cases as well.
where count+ (U) is the number of positives reported by the
ground truth, count− (U) is the number of ground truth negˆ + (S, U) is the scaled number of positives
atives, and count
marked by the given data source. We want to scale the number of positives of the data source, because the population
recognized by S might be quite different from the one recognized by the ground truth. For example, the independent
evaluation agency might have data on 10000 users, while S
might only have data on 1000 users. Therefore, we need to
extrapolate the population unrecognized by S to scale up
the populations:
5.3 Precision Inference Approach
RelativeErr =
count+ (U) − R̂ · (count+ (U) + count− (U))
count+ (U)
= 1−
R̂
count+ (U)/(count+ (U) + count− (U))
= 1−
R − R̂
R̂
=
R
R
(3)
The above value reflects the potential error rate if we scale
the data source’s recognizable population to the size of the
ground truth population. Per Algorithm 1, we calculate
average relative error (RelativeErr) across all performance
campaigns for each data source. We can then rank data
sources based on their average relative errors.
Soundness Analysis. A ranking algorithm needs to be
sound to ensure the optimal assessment: Given two data
sources S 1 and S 2 , a ranking algorithm is sound if it outputs
measurements r 1 for S 1 and r 2 for S 2 , such that r 1 < r 2 if
and only if S 1 is more likely to perform better than S 2 , as
we defined in Def. 3. We will show that the RelativeErr
based ranking algorithm is sound in many cases. First, let
us define the notion of unbiasedness for a data source:
Definition 4. A data source S is unbiased if and only if
its positive predictive value equals to its negative predictive
value: α1 ≈ β2 .
Our experience suggests that many data sources we utilize
are on demographics and can be considered as unbiased. For
example, the accuracy that a data source claims someone
as male, in general, is close to the accuracy that it claims
someone as female (Not Male as in Figure 1). In real-time
bidding, better on-target metrics, i.e. improving the ratio of
audience that really have the data source’s claimed characteristics, is the endeavor of any data provider. We can show
that the RelativeErr based ranking algorithm is sound:
1
k
Lemma 2. Given a set of sound data sources hS , .., S i,
the RelativeErr based ranking algorithm is sound for precisions and orders the data sources based on their expected
performance. In addition, if the data sources are all unbiased, the algorithm is sound for any on-target metrics.
Proof. Per definition, R̂ of the better data source is
closer to the reality, thus R − R̂ is smaller. Since R is constant, the order of RelativeErr is preserved for precisions.
On-target metrics can be the precision, or the negative predictive value, or simply the micro or macro average of the
Although the ranking methodology is able to pinpoint the
highest performing data sources, the output ranking measurement is only a surrogate of precision. It correlates with
the underlying precision, but is inherently different. As we
will show later, in online advertising, it is often necessary
to forecast the campaign performance as well as evaluate
whether a third-party data source is worth the extra amount
of money that an advertiser has to pay in order to utilize
it. In such cases we need an accurate estimation of a data
source’s precision.
Direct Inference of Data Source Precision. We propose an efficient way to directly estimate the predictive values of a data source. As shown in Figure 1, a data source’s
hypothesis on the audience population can be mapped into
the ground truth using its predictive values. Given a performance campaign C i , let the size of Positive, Negative and
i
i
Unknown audiences identified by data source S be D+
, D−
and Di correspondingly (these are scaled values to cover
the whole population of C i ), and the size of ground truth
Positive, Negative and Unknown audiences be Gi+ , Gi− and
Gi correspondingly. When the audience population size is
large, it is clear that we have
i
i
Gi+ ≈ D+
· α1 + D−
· β1 + Di · γ1
i
i
Gi− ≈ D+
· α2 + D−
· β2 + Di · γ2
i
i
Gi ≈ D+
· α3 + D−
· β3 + Di · γ3
Combining with the probability simplex constraint and the
unbiasedness constraint, we can estimate a data source’s predictive values by solving a quadratic optimization problem:
given n performance campaigns hC 1 , C 2 , .., C n i, we search for
predictive values α1 , α2 , α3 , β1 , β2 , β3 , γ1 , γ2 , γ3 so that
min
α,β,γ
n
X
i=1
i
i
(D+
· α1 + D−
· β1 + Di · γ1 − Gi+ )2
i
i
+(D+
· α2 + D−
· β2 + Di · γ2 − Gi− )2
i
i
+(D+
· α3 + D−
· β3 + Di · γ3 − Gi )2
s.t. 0 ≤ αj , βj , γj ≤ 1
3
X
j=1
αj = 1,
3
X
∀j = 1, 2, 3
βj = 1,
j=1
− ξ ≤ α1 − β2 ≤ ξ .
3
X
γj = 1
(4)
(5)
(6)
j=1
(7)
Here, (4) is our optimization objective which aims to find the
best mapping between the data source’s hypothesis and the
ground truth. Note that here we assume the size of audiences
of campaigns C i to be similar, which can be controlled at
campaign set up time. Otherwise we need to normalize by
the audience size of each campaign. (5) and (6) enforce
the probability simplex. (7) attempts to help us find the
unbiased solution, and predefined constant ξ controls our
confidence on the unbiasedness of the predictive values.
We, therefore, can run a few performance campaigns, extract each data source’s hypothesis on those campaigns, compare with the ground truth and solve the above optimization
problem. As we will show, this will efficiently give us the estimated predictive values of data sources in batch (among
those, precision is the most valuable for online advertising).
Performance Analysis. The proposed inference approach is efficient, in terms of both computation complexity
and money. First, it is straightforward to show that the
quadratic programming problem has a semi-definite Hessian
with a bowl shape. The optimization problem is convex
and can be solved efficiently with polynomial time complexity. Additionally, we only need to run a limited number
of performance campaigns to simultaneously estimate the
predictive values of multiple data sources. In practice, it
is possible that a data source’s predictive values are slightly
different in different performance campaigns due to variance.
Given a campaign C i , it is natural to assume a data source’s
predictive values for this specific campaign are
αij = αj + ε1j , ∀j = 1, 2, 3,
βji = βj + ε2j , ∀j = 1, 2, 3,
γji = γj + ε3j , ∀j = 1, 2, 3,
where ε is normally distributed with zero mean. In such
cases, we can get the unbiased estimate of a data source’s
predictive values by running a limited number of performance campaigns:
Theorem 3. Given k ≥ 3 performance campaigns, our
direct inference method can get the unique and unbiased estimate of a data source’s predictive values. Furthermore,
given any predictive value αi and its estimation α̂i , we have
P ( αi − α̂i ≤ si · Tδ/2,2k−6 ) ≥ 1 − δ,
where 0 < δ < 1 is a constant, si is the standard error of the
estimation, and Tδ/2,2k−6 is (1 − δ/2)-th quantile of Student
Distribution with 2k − 6 degrees of freedom.
Proof. The optimization can be converted into a linear regression problem within a simplex search space. This
regression problem contains 6 free regressors and each campaign provides 2 points in the space. When we have k ≥ 3
campaigns, the quadratic matrix is positive-definite and we
will have a unique global optimal solution. A Bias-VarianceNoise decomposition shows the solution is unbiased.
Since the errors are normally distributed, the sum of the
regression residuals is then distributed proportional to Student Distribution with 2k − 6 degrees of freedom:
t = (αi − α̂i )/si ∼ T2k−6
We then construct the confidence levels for the estimated
regressors.
By running more campaigns, we can quickly reduce the estimation errors and get highly reliable predictive value estimations of multiple data sources. Given its computational and
economic efficiency, we adopted the direct inference method
and utilize it continuously to generate the quality report on
data sources.
6.
USE CASES
In this section, we will discuss some use cases where the
quality assessment of first or third-party data sources can
be useful. First we will talk about targeting in online advertising, and the amount that an advertiser should be willing
to pay for a data source. Then we will give a very general
use case in campaign forecasting, i.e. to predict, before an
online advertising campaign starts, what category of users
will actually be reached by a pre-set targeting criteria.
6.1 Targeting in Online Advertising
Advertisers aim to reach the best audiences to promote
their products, so that they can increase the likelihood of a
click or an action happening. The automated way of grouping users into beneficial and non-beneficial subsets is often
called audience segmentation. For an informative work on
how this kind of audience segmentation can improve click
rates, refer to [15]. In this paper, however, we focus on a different kind of targeting where the advertisers already have a
pre-defined set of users they want to target. As an example,
suppose that an advertiser wants to reach only female audiences within the age range 21-35. There are multiple data
sources this advertiser can utilize to reach this group, but
as discussed in this paper, none of these data sources gives
a definitive classification. Intuitively, an accurate prediction
of the quality of a data source is essential for advertisers to
choose it over others. Also, note that here we mostly care
about the precision or positive predictive value of a data
source (i.e. α1 , when a data source suggests that a user is in
category c, the likelihood that this user actually belongs to
category c), since this is the signal that the advertiser uses
to bid on a user.
Here, we would also like to discuss the consideration of
data cost. In general, when an advertiser wants to utilize a
third-party data source for bidding purposes, it should pay
the third-party provider a certain amount of money. This
cost is generally per impression served using this data source,
hence can have significant effect on the ROI of an advertiser
(i.e. advertiser needs to pick up extra clicks/conversions to
make up for the money paid to the third-party for the targeting information it provides). An important point for an
advertiser to consider is if using a data contract is “worth”
its price. We will give a simple calculation here for the case
when the advertiser utilizes no data sources to reach a specific audience (i.e. free targeting), and whether adding the
data source and paying for it makes sense. Our main argument is that, by paying for the data source to target a
specific audience, the reduced cost of the mis-targeted impressions (i.e. those impressions that are served to the audience that are out of our desired audience) should make up
for the data cost. In other words, for the same amount of
money we should get more of the desired impressions, although our total number of impressions is less due to data
cost. Please note that below we assume the effective cost
per impression (cpi) to be the same for both free targeting
and data source assisted targeting, just that the data source
has the additional data cost per impression (cpidata ):
totalSpend
(1 − errorRate(dataSource)) ≥
cpi + cpidata
totalSpend
(1−errorRate(freeTargeting))
cpi
α1,freeTargeting
α1,data
≥
.
cpi + cpidata
cpi
Above, totalSpend is the amount of money that the campaign spends, cpi is effective cost per impression, and cpi data
totalSpend
is data cost per impression (hence cpi+cpi
is the numdata
ber of impressions picked up by data source assisted targetis the number of impressions that can be
ing, and totalSpend
cpi
picked via free targeting, i.e. no data cost). errorRate is indeed the inaccuracy of free targeting (percentage of audience
that is not desired), and percentage of the cases when the
data source predicts a user to be in desired audience, while,
in fact, it is not. In the next inequality, we actually translated (1 - errorRate) into α1 from Section 5. After further
reorganizing the above inequality, we get the following:
α1,data
−1
.
(8)
cpidata ≤ cpi ×
α1,freeTargeting
This means that for a data source to be beneficial for a
campaign,
its data cost per
impression should be less than
α1,data
cpi × α1,freeTargeting − 1 . Please note that we have the
assumption here that effective cost per impression would be
the same for free targeting vs. data source which is not
always valid, i.e. we may have to pay more to show ads
(impression) to those users that the data source tagged to
be desirable. Also, it can be seen that the benefit of the
data source often depends on how expensive the impressions
are for a campaign, hence is campaign specific. Finally, the
above calculations do not take into account the cost of data
evaluation utilizing our proposed two methodologies, which
was also mentioned in Section 4.2. However, this evaluation
can be performed once for each data source, and hence is
not of significance for each campaign that utilizes it.
6.2 Forecasting
Forecasting the performance (return-on-investment), reach
(unique users we can show an ad to), and delivery (amount
of money we can spend on advertising given the targeting
criteria) of a campaign is a significant problem that has to
be dealt in online advertising [9]. Here we show that by utilizing the accuracy metrics (i.e. α1→3 , β1→3 , and γ1→3 from
Section 5) over multiple data sources that may tag a user,
we can actually predict the expected number of users that
will fall into a specific audience/category, on top of the total
spend/reach as in the traditional forecasting problem, once
the advertising campaign goes live.
Here is how the forecasting process in online advertising
works. Once the advertiser sets some targeting criteria (filtering of users to show ads according to anonymous user
properties) and goals (in terms of clicks and conversions)
for a campaign, we can utilize our system as explained in [9]
to find out which users this campaign is likely to reach. We
can already calculate expected number of unique users and
delivery for the campaign from this information alone. Furthermore, one problem that we can solve for the advertisers
is the prediction of what percentage of these users will fall
into a specific user category/class c. Note that the approach
mentioned in [14] does work on this problem of predicting
likelihood of a user belonging to a specific category via utilizing many features, but their focus is on targeting rather
than forecasting. Here, we suggest that rather than training
a simple model for predicting the membership in a category,
we can utilize multiple data sources and their estimated predictive values to forecast the expected number of users that
will fall into a category. This information is not used in
bidding time, hence contrary to the targeting use case we
explained in Section 6.1, there is no data cost.
Figure 2 summarizes the overall idea. In the first step,
Targeting
Criteria T
Users which obey
the Targeting Criteria T
Figure 2: Forecasting Example.
we communicate the targeting criteria set by the advertiser
to our set of users. For real-time forecasting, we often need
to use a sampled set of users [9]. Once we filter the users
that are appropriate for the targeting criteria, we can go
through each of these users and see whether these users are
tagged by a first or third-party category cd . Once we see
these taggings, we can calculate the probability of this user
belonging to a desired ground-truth category cg by the probability p(cg |cd ) which is the α1 from Section 5, i.e. precision.
If we assume that each user is tagged by one and only one
ci , we can forecast the expected number of users that will
belong to category cg as:
X X
X
I(u has ci ) p(cg |ci ) . (9)
p(cg |u) =
E[ |cg | ] =
u∈T
u∈T ci ∈C
In (9), T is the set of users that belong to a certain set of
targeting criteria T. cg is the category that the advertiser desires to forecast how many of their targeted users will belong
to. ci is a first/third-party category from a list of categories
C for which we have the prediction values. I is an indicator
to see whether a user has category ci , and above formula is
valid only because we assume that each user has only one of
possible ci s. If each user can have multiple first/third-party
categories (as is the case in real situations), we need to aggregate multiple p(cg |ci )s, where we can utilize combination
methods such as getting the maximum, minimum, average
or median of the precision values.
7. EXPERIMENTAL RESULTS
In this section we will give some preliminary results for
our optimization-based evaluation technique. We have already given some preliminary results for our first methodology (Section 5.1) in Table 2. As aforementioned, we do aim
to calculate all nine prediction values of a data source for
a category, but for purposes of online advertising, the most
important one is the precision (α1 ). Only if this value is
high we can reliably use this data source to reach a certain
category of users.
Simulation Results. To evaluate our methodology we
ran several simulated campaigns, where for each campaign
we create an audience of 100 users and assign them to predicted categories as follows:
• Random, disjoint sets of 20 users each are assigned to
category c, not c, and unknown,
• The rest 40 users are assigned to either category c, not
Difference
0.10 0.20 0.30
0.00
High Quality
Low Quality
1
2
3 4 5 6 7 8 9 10
Number of Campaigns
0.12
Figure 3: Number of campaigns versus difference
between predicted and actual precision values.
0.00
Difference
0.04
0.08
High Quality
Low Quality
0.00
0.10
0.20
Noise Level
0.30
Figure 4: Noise versus difference between predicted
and actual precision values for six campaigns scenario.
c, and unknown in a uniform manner.
Then, we generate the ground truth categories for two types
of data sources with the following actual probability values:
• High Quality: This data set has the following underlying nine probability values: α1→3 = (0.8, 0.15, 0.05),
β
0.7, 0.1), γP
1→3 = (0.2,P
1→3 = (0.4, 0.5, 0.1), note that
P
α
=
β
=
i
i
1→3
1→3
1→3 γi = 1,
• Low Quality: This data set has the following underlying nine probability values: α1→3 = (0.4, 0.5, 0.1),
β1→3 = (0.3, 0.6, 0.1), and γ1→3 = (0.5, 0.4, 0.1).
Please note that this kind of synthetic data generation is
quite counter-intuitive. We first create the predicted values using some pre-set distribution, and then generate these
users’ actual categories using the predictive values of the
two data sources. For example, if the user in our synthetically generated audience has a predicted category of not c by
High Quality data source, then we assign it to ground truth
category of c by probability β1 = 0.2, not c by probability
β2 = 0.7, and unknown by probability β3 = 0.1.
Once we generate the dataset, we actually have aggregated values of category counts for each data source. Using
these category counts, we can utilize our data quality assessment method we described in Section 5.3, and look into the
difference between our computed predictive values, and the
actual predictive values as given above.
The results of the above described simulations are given in
Figures 3 and 4. In Figure 3, we estimate the nine predictive
values for each of the data sources using our methodology, by
utilizing multiple campaigns (the results are averaged over
100 trials at each value in x-axis). As we have proven in
Section 5.3, we do need at least three campaigns to get a
Figure 5: Pearson correlations between the estimated positive population and the ground truth.
unique solution for all nine probabilities. We can observe
that starting with four campaigns, the difference of the real
values versus our predicted values fall to zero. We have
plotted the difference between α1 s (precision values, since
α1 = p(actually positive | predicted positive)), but the results are similar for other α2→3 , β1→3 , and γ1→3 .
Next, we performed another experiment where we introduced a uniform noise between ±ζ (where we changed ζ
between 0 and 0.35) to the above nine real predictive values, and then generated the ground truth assignments. We
tried to recover the nine predictive values using six campaigns and present the difference (averaged over 100 trials)
between real and predicted α1 values in Figure 4. We can
see that even under significant noise levels, our methodology
can recover the precision values accurately. Because α1 for
the high quality data source is higher, we can also observe
that the noise effect is slightly less.
Real-World Results. Following the methods we discussed in Section 5.1, we ran 156 performance campaigns,
each of which targeted a specific website. We used half of
the campaigns to calculate RelativeErr and predictive values for around 100 data sources. Then, we tried to estimate
the positive population in the rest of campaigns using these
100 data sources. We utilized the average of positives predicted by the sources, and calculated the correlation with the
ground truth positive sizes. For the direct inference method,
it is clear that for each campaign C i , its estimated positive
i
i
population is Gi+ ≈ D+
· α1 + D−
· β1 + Di · γ1 (for a single data source). For the RelativeErr method, by deducing
(3), we can roughly estimate Gi+ = M · τ̂ R̂/(1 − µ̄), where
M is the population size, τ̂ is the percentage of population
recognized by the ground truth in the training set, and µ̄ is
the average R − R̂/R (i.e. RelativeErr from Eq. 3) of the
training set for a single data source.
Each method’s Pearson correlation coefficient is shown in
Figure 5. The direct inference method gives a significantly
more accurate estimate of the positive population (p < .001)
and it correlates well with the ground truth.
Next, we utilized two popular data sources S1 and S2 as
the targeting criterion and ran test campaigns individually.
The reported positive rates from the independent evaluation
agency can be treated as the ground truth of their precisions.
The ground-truth precisions and our estimated values are
listed in Figure 6. The direct inference method yields a
much closer estimation to the ground truth (p < .001), while
the ranking method preserves the orders but the values are
substantially different from the ground truth.
Our proposed approach was deployed into Turn’s data
management platform and generates weekly reports on the
of combining multiple data sources and their quality assessments to come up with a better model.
Acknowledgments
We thank many talented scientists and engineers at Turn for
their help and feedback in this work.
9. REFERENCES
Figure 6: Reported measurements on the positive
population from different methods.
Figure 7: Inferred precision of a data source over a
three month period.
quality of our top data sources. We have received positive
feedback from our campaign optimization managers in the
field, commenting that the reported precisions are close to
the real campaign results. Interestingly, by evaluating our
data sources periodically, we are forming a positive reinforcement loop over their data quality: feeling the pressure,
data providers work consistently to improve their data quality. For example, the estimated precision of one data source
over a three month period is plotted in Figure 7. It is clear
that the data source’s quality has been improved over this
time period.
8.
CONCLUSIONS AND FUTURE WORK
In this paper, we have presented a novel framework to
evaluate first or third-party data sources on user properties
for online advertising, which is a particularly challenging
task when the ground truth is reported in aggregate form.
We call this problem data quality assessment, and presented
two solutions, one utilizing the data sources directly in a
campaign, and another one, which utilizes outputs from multiple online advertising campaigns to optimize a set of probabilities which represent the “goodness” of a data source.
We have also presented some use cases on how these evaluations can be utilized in online advertising domain, mainly in
targeting, assessing the amount of money that an advertiser
should pay for a data source, and forecasting. Some preliminary simulation and real-world results were also presented
that show the effectiveness of our methodology, as well as
some results on the performance of a well-established actual
data provider for age categorization of users on multiple realworld advertising campaigns.
Possible future work mainly lies on the use cases of the
evaluation output of our methodologies, as given in Section 6. Our current focus is on accurate targeting of online
users, which also needs to take into account the problem
[1] N. R. Adam and J. Wortmann. Security-control
methods for statistical databases: A comparative
study. ACM Computing Surveys, 21:515–556, 1989.
[2] G. Casella and R. L. Berger. Statistical inference,
volume 2. Duxbury Pacific Grove, CA, 2002.
[3] B. Chen, L. Chen, R. Ramakrishnan, and D. R.
Musicant. Learning from aggregate views. In Proc.
IEEE ICDE, 2006.
[4] S. Chen. Cheetah: a high performance, custom data
warehouse on top of mapreduce. Proc. VLDB
Endowment, 3(1-2):1459–1468, 2010.
[5] V. Cheplygina, D. M. J. Tax, and M. Loog. On
classification with bags, groups and sets. Pattern
Recognition Letters, 59:11–17, 2015.
[6] H. Elmeleegy, Y. Li, Y. Qi, P. Wilmot, M. Wu,
S. Kolay, A. Dasdan, and S. Chen. Overview of turn
data management platform for digital advertising.
Proc. VLDB Endowment, 6(11):1138–1149, 2013.
[7] C. Ferri, J. Hernandez-Orallo, and R. Modroiu. An
experimental comparison of performance measures for
classification. Pattern Recognition Letters, 30:27–38,
2009.
[8] A. Gunawardana and G. Shani. A survey of accuracy
evaluation metrics of recommendation tasks. Journal
of Machine Learning Research, 10:2935–2962, 2009.
[9] A. Jalali, S. Kolay, P. Foldes, and A. Dasdan. Scalable
audience reach estimation in real-time online
advertising. In Proc. ICDMW, pages 629–637, 2013.
[10] K.-C. Lee, B. Orten, A. Dasdan, and W. Li.
Estimating conversion rate in display advertising from
past performance data. In Proc. ACM KDD, pages
768–776, 2012.
[11] D. R. Musicant, J. M. Christensen, and J. F. Olson.
Supervised learning by training on aggregate outputs.
In Proc. IEEE ICDM, pages 252–261, 2007.
[12] C. Parker. An analysis of performance measures for
binary classifiers. In Proc. IEEE ICDM, pages
517–526, 2011.
[13] T. White. Hadoop: The Definitive Guide. O’Reilly
Media, Sebastopol, CA, 2012.
[14] M. H. Williams, C. Perlich, B. Dalessandro, and
F. Provost. Pleasing the advertising oracle:
Probabilistic prediction from sampled, aggregated
ground truth. In Proc. ACM ADKDD, 2014.
[15] J. Yan, N. Liu, G. Wang, W. Zhang, Y. Jiang, and
Z. Chen. How much can behavioral targeting help
online advertising? In Proc. ACM WWW, pages
261–270, 2009.
[16] F. X. Yu, K. Choromanski, S. Kumar, T. Jebara, and
S. Chang. On learning from label proportions.
arXiv:1402.5902v2, 2015.
| 2cs.AI
|
arXiv:1605.04160v1 [cs.DS] 13 May 2016
Searching Lattice Data Structures of Varying
Degrees of Sortedness ∗
Mohammad Obiedat
Department of Science, Technology, and Mathematics,
Gallaudet University, 800 Florida Avenue NE, Washington,
DC 20002-3695, USA†
December 25, 2017
Abstract
Lattice data structures are space efficient and cache-suitable data
structures. The basic searching,
insertion, and deletion operations are
√
of time complexity O( N ). We give a jump searching algorithm of
time complexity O(J(L) log(N )), where J(L) is the jump factor of the
lattice. J(L)√approaches 4 when the degree of sortedness of the
√ lattice
approaches N . A sorting procedure of time complexity O( N ) that
can be used, during the system idle time, to increase the degree of
sortedness of the lattice is given.
Keywords. Lattice data structure, cache-suitable, jump searching,
jump factor, degree of sortedness, incremental sorting.
1
Introduction
Searching is one of the most time-consuming operation of many software
applications. Basically, a searching procedure takes a target key and it has to
∗
†
Published in the International Journal of Data Structures, 1(1) 64-76 (2015)
Email address: [email protected]
1
decide whether that key is present or absent. This seemingly simple operation
becomes very challenging when the set of keys is dynamically changing, by
insertion and deletion, or when using minimum space and energy are very
crucial. Therefore, any data structure that is designed to organize pieces
of information in a dynamic environment should be evaluated by the space
required to hold the keys, by the time required to perform the searching,
insertion, and deletion operations, and by the suitability of the structure for
caching or more generally for memory hierarchies.
In Bjarne Stroustrup’s recent paper [12], he emphasizes the importance of
compactness and predictable memory access patterns for efficiency. Stroustrup points out that power consumption is roughly proportional to the number of memory accesses. Consequently, the less data we store and move the
better for efficiency, especially in infrastructure software and applications for
hand-held devices like smartphones. Stroustrup makes a strong case in favor
of using arrays instead of linked lists or more generally pointer-based structures. There are two important differences between arrays and linked lists
that affect the efficiency of their operations. First, arrays are more space
efficient than linked lists. Second, arrays are more suitable for caching than
linked lists, because an array’s elements are stored in contiguous memory
locations, while a linked list’s nodes are scattered in memory.
Most of the current index structures are either pointer-based or use extra storage space in order to obtain the optimal O(log (N)) time complexity
of their operations. Obviously, such data structures are not suitable for
memory-constrained systems. Sequential search of unsorted arrays is space
efficient and very suitable for caching, but unfortunately it is of time complexity O(N), which is too slow for most applications.
The lattice data structure (LDS) was novelly introduced by Berkovich in
1992, see [2]. The space required to hold a lattice with N keys is equal to
the space required to hold an array with N keys. Additionally, it is possible
to devise more efficient algorithms using the LDS than using the array. This
makes the LDS a good contribution to the current trends of developing data
structures and algorithms for memory-constrained applications. Also, the
LDS operations have good temporal and spatial locality which makes it a
rich environment for the ongoing research on cache-conscious data structures
and cache-oblivious algorithms. Consequently, in addition to the theoretical
importance of the LDS, we argue that this structure is the ideal choice for
memory-constrained systems, for real-time systems, for systems where power
consumption is crucial, and for systems where incremental adjustment is
2
feasible.
The basic searching,
√ insertion, and deletion operations of the LDS are of
time complexity O( N). In this paper, we give a general jump searching
algorithm of time complexity O(J(L) log(N)), where J(L) is the jump factor
of the lattice. The jump factor of the√lattice approaches 4 when the degree of
sortedness of the lattice approaches N. In order to keep the jump factor
√ of
the lattice small, we provide a sorting procedure of time complexity O( N )
that can be used, during the system idle time (e.g., by giving it the least
priority when scheduling the lattice operations), to increase the degree of
sortedness of the lattice.
Experimental evaluation of randomly built LDSs shows that the performance of the basic searching procedure of the LDS is similar to skip list
search when the structures have up to 1000 keys, and the performance of the
jump searching procedure is similar to skip list search when the degree of
sortedness of the lattice is close to 0.9h, where h is the height of the lattice.
When the ratio of the number of times the sorting procedure is called to the
number of times the insertion and deletion operation are called is close to 5,
the performance of the jump searching procedure is similar to its performance
on a lattice with degree of sortedness close to 0.9h.
To obtain a good performance of the searching operation, a hybrid searching algorithm can be implemented, together with the sorting procedure, by
searching the lattice using the basic searching procedure when the degree
of the sortedness of the lattice is less than or equal to 0.9h, and the jump
searching procedure when the degree of the sortedness of the lattice is greater
than 0.9h.
The layout of this paper is as follows. In Section 2, we formalize the
definition of the LDS, present the basic searching, insertion, and deletion algorithms, and then analyze the space and time complexities of implementing
these algorithms. In Section 3, we devise a general jump searching algorithm
that can be used instead of the basic searching algorithm to obtain a better performance and time complexity when the jump factor of the lattice is
small. Then, we provide a detailed analysis of the jump factor and devise a
sorting algorithm that can be used to reduce the jump factor of the lattice
during the system idle time. In Section 4, we compare the performance of
the jump searching algorithm with skip lists search. Then, we examine the
effect of using the sorting algorithm on the jump factor of the lattice during
the system idle time. Conclusion and recommendations for further research
are given in Section 5.
3
2
Lattice Data Structures
In this section, we formally define the lattice data structure and discuss its
basic algorithms and properties.
In order to define the lattice data structure, we first recall the definition
of Ferrers diagrams, see [13]. Let λ = (n1 > · · · > nm > 0) be an integer
partition of a positive integer N. A Ferrers diagram of shape λ consists of m
rows, where the top row contains n1 equisized square cells (hereafter referred
to as cells), the second row from the top contains n2 cells, etc. Each row is
left-justified. For a given non-negative integer number h, we are interested in
up-side-down Ferrers diagrams of shape (h + 3, h + 2, . . . , 2, 1), see Figure 1
for an up-side-down Ferrers diagram of shape (7, 6, 5, 4, 3, 2, 1). The following
numbering schema and terminologies will be used for such diagrams. Rows
and columns are numbered from bottom to top and from left to right, starting
with one, respectively. Cells in each row and each column are numbered
from left to right and from bottom to top, starting with one, respectively.
By diagonal k, we mean the cells that lie on the line segment that can be
formed by connecting the upper-left corner of the first cell of row k and the
lower-right corner of the first cell of column k where 1 6 k 6 h+ 3. The head
of diagonal k is the first cell of row k and the tail is the first cell of column
k. Cells in each diagonal are numbered from head to tail, starting with one.
Using up-side-down Ferrers diagrams, instead of the usual Ferrers diagrams,
is only a matter of convenience and preference.
Definition 2.1 A lattice data structure (LDS) of height h is an arrangement
of a finite set of distinct positive integers, referred to as proper keys, in an
up-side-down Ferrers diagram of shape (h + 3, h + 2, . . . , 2, 1) such that:
1. Each cell in row one contains 0.
2. Each cell in column one contains 0.
3. Each cell in diagonal h + 3 contains ∞, except the head and the trail
which contain 0.
4. All other cells contain proper keys, except the last k non-zero cells of
diagonal h + 2 which contain ∞ for some 0 6 k 6 h − 1.
5. The proper keys in each row, column, and diagonal are in increasing
order from left to right, from bottom to top, and from head to tail,
respectively.
4
Remark 2.2 In the original definition of the LDS given by Berkovich [2],
diagonals were not required to be sorted. As we will see in the next section, by requiring diagonals to be sorted one can apply a variety of searching
techniques to improve the time complexity of the basic searching algorithm.
Unless otherwise indicated, we will not distinguish between a cell’s location and its content, so we will say that “cell two in row four is less than cell
three in row four” instead of “the content of cell two in row four is less than
the content of cell three in row four”. A LDS of height 4 and proper keys
{3, 5, 6, 9, 12, 20, 30, 31} is shown in Figure 1.
0
0
0
0
∞
12 ∞
9
30 ∞
0
5
0
3
6
0
0
0
20 ∞ ∞
31 ∞ ∞
0
0
0
0
Figure 1: A LDS of height 4
Remark 2.3 The purpose of the improper keys in the boundary cells, row
one, column one, and diagonal h+3, is to serve as a wall; the LDS procedures
can not pass these cells and move down, to the left, or to the right. Row one
cells will serve as failure cells; if a searching procedure reaches these cells
then it means that the key is absent. Choosing the keys to be positive integers
and the relation 6 to be the usual less than or equal relationship is only a
matter of convenience; in practice the keys could belong to any set with a
partial order relationship.
In Figure 2, we illustrate the eight possible movements from a given cell
C to an adjacent cell. These movements will be used in devising the LDS
algorithms.
5
UL
L
DL
U UR
C
R
D DR
U = Up, D = Down, L = Left, R = Right
UL = Up Left, UR = Up Right
DL = Down Left, DR = Down Right
Figure 2: Movements in a LDS from a given cell C
Remark 2.4 If the cells adjacent to C preserve the lattice structure, i.e., satisfy condition (5) of Definition 2.1, then C also preserves the lattice structure
provided that C > D, C > UL, C < U, and C < DR or DR = 0.
Next, we describe the basic searching, insertion, and deletion algorithms
of the LDS. We assume that we are given a LDS of height h > 1 and a key
K in the search space of the lattice.
Searching. The basic searching algorithm of the LDS, SearchLDS, works
as follows. First, we set C to be the second cell of diagonal h + 2. Then we
perform the following movements until K is declared present or absent: If
K > C, then we move diagonally to DR by setting C = DR. If K < C,
then we move down to D by setting C = D. We declare K is present if
C = K, and declare K is absent if C = 0. This basic searching algorithm,
SearchLDS, is given as Algorithm 1.
Insertion. To insert a key K into a LDS, we first search the lattice for K.
If K is present then it cannot be inserted. If K is absent, then we find a
temporary cell for K as follows: If diagonal h + 2 has cells containing ∞,
then we insert K at the first cell of diagonal h + 2 that contains ∞. On the
other hand, if diagonal h + 2 does not have any cell that contains ∞, then we
expand the lattice by increasing its height by one and then insert K in the
second cell of diagonal h + 3. After K is inserted in the temporary cell, say
cell C, we maintain the lattice structure by applying the inward algorithm,
Inward, given as Algorithm 2.
6
Algorithm 1 Basic Searching Algorithm of the LDS
SearchLDS (LDS, h, K)
// Given a LDS of height h, and a key K.
// Determine whether K is present or absent.
C = second cell of diagonal h + 2;
while (C 6= K and C 6= 0) do
if (K > C) then
C = DR; // a DR movement
else
C = D; // a D movement
end if
end while
if C = K then
K is present at C;
else
K is absent;
end if
Deletion. To delete a key K, we first find its location, say K is located at
cell C. If C is the last proper key of diagonal h + 2, then we place ∞ in
cell C. If C is not the last proper key of diagonal h + 2, then we place the
last proper key in diagonal h + 2, say F , in cell C and place ∞ in cell F .
Finally, we maintain the lattice structure by applying the Inward algorithm
if (C < D or C < UL), and the Outward algorithm, given as Algorithm 3,
if (C > U or C > DR and DR 6= 0). In all cases, we reduce the height of
lattice by one if diagonal h + 2 does not have any proper key.
In the following theorem, we give the number of proper and improper
keys of a LDS of height h.
Theorem 2.5 Let L be a lattice of height h, where h > 1, and suppose that
diagonal h + 2 contains k proper keys where 1 6 k 6 h. Then
1. The number of proper keys in L is h(h − 1)/2 + k.
2. The number of improper keys in L is 4h − k + 6.
Proof: Straightforward.
7
Algorithm 2 Inward Algorithm of the LDS
Inward (LDS, C)
while (C < D or C < UL) do
if (UL > D) then
swap UL and C; // a UL swap
else
swap D and C; // a D swap
end if
end while
Algorithm 3 Outward Algorithm of the LDS
Outward (LDS, C)
while (C > U or (C > DR and DR 6= 0) ) do
if (DR < U and DR 6= 0) then
swap DR and C; // a DR swap
else
swap U and C; // a U swap
end if
end while
Corollary 2.6 The
√ hold N proper keys is of height
√smallest lattice that can
h where h = ⌊(1 + 8N − 7)/2⌋ = ⌈(−1 + 8N + 1)/2⌉.
In the following lemma, we give upper bounds on the number of comparisons to search for a key, and the average number of comparisons to search
for a present key when a LDS is searched using SearchLDS.
Lemma 2.7 Let L be a lattice of height h and suppose that diagonal h + 2
contains k proper keys where 1 6 k 6 h. If we search L using SearchLDS,
then
1. The number of comparisons to search for any absent key is h + 1.
2. The number of comparisons to search for any present key is less than
or equal to h.
3. The average number of comparisons to search for a present key is
2h
2h3 − 2h + 3k 2 + 3k
≈
.
2
3h − 3h + 6k
3
8
Proof: Straightforward.
Theorem 2.8 The time complexity of SearchLDS,
Inward, and Outward al√
gorithms of a LDS with N proper keys is O( N).
Proof:
√ From Lemma 2.7 and Corollary 2.6, the time complexity of SearchLDS
is O( N). To find the time complexity of the Inward algorithm, suppose C
is located on diagonal s and column t. Then with each D swap, s is reduced
by 1 and t is not changed, while with each UL swap, t is reduced by 1 and s
is not changed. So with each swap, s + t is reduced by 1 . Hence
√ the Inward
algorithm makes up to 2h swaps. Consequently, it is an O( √
N ) algorithm.
Similarly, the time complexity of the Ouward algorithm is O( N).
Corollary 2.9 If SearchLDS is used to locate a key or determine its absence,
then the time complexity of the insertion and deletion
√ procedures, described
in this section, of a LDS with N proper keys is O( N).
One of the main advantages of LDS over other searching structures such
as balanced trees (e.g., AVL trees [11] and self-adjusting trees [10]), hash
tables [4], skip lists [6, 7, 8], Jumplists [3], and J-lists [1] is the space required to hold and maintain the lattice. A lattice can be easily represented,
either explicitly or implicitly, by a one-dimensional array where each cell is
represented by one array element. In explicit representation of LDSs, a lattice is mapped into a one-dimensional array by mapping diagonal one to the
first array element, diagonal two to the second and third array elements, etc.
Implicit representation of LDSs is similar to explicit representation, except
that only cells with proper keys are mapped into array elements. While implicit mapping is a little more space efficient that explicit mapping, explicit
mapping is the natural choice, because of its simplicity and its superiority in
term of time.
In Table 1, we compare the extra space required to hold a proper key by
various searching data structures.
3
Jump Searching of Lattice Data Structures
Lattice data structures outperform other data structures such as balanced
trees, self-adjusting trees, and skip lists in terms of space, time bounds of
some operations, and suitability for memory hierarchies. However, the basic
9
Table 1: Required extra space per proper key by various searching structures
Structure
Extra space per proper key
LDSs-Explicit mapping approaches 0 as h → ∞
LDSs-Implicit mapping 0
J-lists
1 integer
Skip lists
an average of 4/3 pointers
Jumplists
2 pointers and 1 integer
AVL trees
2 pointers and 2 bits
Red-black trees
2 pointers and 1 bit
Splay trees
2 pointers
Hash tables
variable
√
searching algorithm of the LDS, of time complexity O( N), cannot compete
with the searching algorithms of these data structures in terms of time, especially when the size of the search space is large. In this section, using the fact
that diagonals and columns of LDSs are sorted, we devise a jump searching
algorithm that can be used instead of SearchLDS to speed up the search.
The basic idea of the jump searching algorithm is to make wide jumps on
diagonals and columns instead of the sequential movements to adjacent cells
in the basic searching algorithm. We also devise a sorting algorithm that can
be used, during the system idle time, to incrementally increase the degree of
sortedness of the lattice, and hence to improve the performance of the jump
searching algorithm.
Jump searching algorithms over portions of a sorted array are studied
in detail in [9]. In Algorithm 4, JumpSearchLDS, we given a general jump
searching algorithm of the LDS that can be used with a variety of jump
searching procedures of sorted arrays (e.g., binary and interpolation) to obtain a better performance and time complexity than SearchLDS.
Unless otherwise indicated, from now on, we will assume that L is a LDS
of height h > 1, and K is a key in the search space of L. For brevity, let d
stand for a DR movement. We define the search path of K, with respect to
L, to be the sequence of movements (d or D) until SearchLDS terminates.
Definition 3.1 The jump factor of K, with respect to L, is the number of
blocks of identical movements in the search path of K.
We denote the jump factor of K with respect to L by J(K; L), or J(K),
when L is understood. For example, if the search path of K is ddddDDDDDdd
10
Algorithm 4 Jump Searching Algorithm of the LDS
JumpSearchLDS (LDS, h, K)
// Given a LDS of height h, and a key K.
// Determine whether K is present or absent.
C = second cell of diagonal h + 2;
while (C 6= K and C 6= 0) do
if (K < C) then
// a downward jump
Let C be the first cell in the current column such that K > C;
else
// a diagonal jump
Let C be the first cell in the current diagonal such that K 6 C, set
C = 0 if there is no such cell;
end if
end while
if C = K then
K is present at C;
else
K is absent;
end if
then J(K) = 3. We define the jump factor of L by
J(L) = max{J(K) : K is a key in the search space of L}.
Let P be the procedure that performs the jump steps in JumpSearchLDS.
So, P will take a sorted array, a column (respectively, a diagonal) with m
elements and for a given key K will return the first element F such that
K > F (respectively, the first element F such that such that K 6 F or set
C = 0 if there is no such F ). In the following theorem, we give the time
complexity of JumpSearchLDS where the time complexity of P is O(tP (m)).
Theorem 3.2 Let L be a lattice of height h. If the time complexity of P is
O(tP (m)), then the time complexity of JumpSearchLDS is O(J(L)tP (h)).
Proof: Given a lattice of height h. One can form a triangle by connecting
the following three points: the lower-left corner of the first cell of row one, the
lower-right corner of the last cell of row one, and the upper-left corner of the
11
last cell of column one. In Algorithm 4, when a downward jump is performed
the keys in the area outside triangle T1 in Figure 3-A are eliminated from
the search, and when a diagonal jump is performed the keys in area outside
triangle T2 in Figure 3-B are eliminated from the search. Since the time
complexity of each jump is O(tP (h)) and each search requires up to J(L)
jumps, then the time complexity of JumpSearchLDS is O(J(L)tP (h)).
❅
❅
❅
❅
❄
❅
❅
❅
❅
❅
❅
❅
T1 ❅
❅
❘
❅
❅
❅
❅
❅
❅
❅
❘
❅
❅
❅
T2 ❅
❄
❅
A. a downward jump
B. a diagonal jump
Figure 3: Jumps of the LDS
Corollary 3.3 Let L be a lattice with N proper keys. If the time complexity of P is O(log(m)), then the time complexity of JumpSearchLDS is
O(J(L) log(N)).
The time complexity of JumpSearchLDS depends on the jump factor of
the lattice and on the procedure used to perform the jumps. From Theorems 2.8 and 3.2, if a binary-search-like procedure is used to perform the
jump steps, then the time complexity of JumpSearchLDS is less than or equal
the time complexity of SearchLDS
provided that the jump factor of the lat√
tice is less than or equal to N / log(N). In the following lemma, we give
upper bounds on the jump factors of keys.
Lemma 3.4 Let L be a lattice of height h, and K be a key in the search space
of L. Suppose the last movement in the search path of K ends on diagonal s
and column t. Let u = t − 2 and v = h − s + 2, then
1. The number of d movements in the search path of K is u and the
number of D movements is v.
12
u + v if u = v
2u + 1 if u < v
2. J(K) 6
.
2v + 1 if u > v
h − 1 if K is present
.
3. J(K) 6
h
if K is absent
Proof: Straightforward.
Theorem 3.5 Let L be a lattice of height h. Then
1. The average of the jump factors of the present keys is less than or equal
to h/3 + 5/6 ≈ h/3.
2. The average of the jump factors of the absent keys is less than or equal
to h/2 + 1/2 ≈ h/2, provided that the probability the search for a randomly chosen absent key will terminate in a cell on row 1 and column
t for each 2 6 t 6 h + 2 is 1/(h + 1).
Proof: Follows from Lemma 3.4 and by using basic algebraic summations.
Definition 3.6 Suppose α ∈ {3, . . . , h}. A LDS is said to be sorted of degree
α (for brevity, α-sorted) if the first proper key of diagonal s is greater than
the last proper key of diagonal s − 1 for each 4 6 s 6 α + 2.
For example, the lattice in Figure 1 is 3-sorted but it is not 4-sorted.
Remark 3.7 If L is α-sorted, then the proper keys in the first α+2 diagonals
are completely sorted. Also, if β = α/h then the percentage of the proper
keys in the first α + 2 diagonals is β 2 + (β − β 2 )/(h + 1) ≈ β 2 . For example,
if α = h/2 then approximately 25% of the proper keys in the lattice are
completely sorted.
Lemma 3.8 Suppose L is α-sorted and K is a key in the search space of L.
1. If K is present and the last movement in the search path of K ends on
diagonal s, where s < α + 2, then J(K) 6 2.
2. If K is absent and the last movement in the search path of K ends on
diagonal s, where s < α + 1, then J(K) 6 4.
13
Proof: (1) Suppose K is present and s < α + 2. Since the first proper key in
diagonal α + 2 is greater than the last proper key in diagonal α + 1, then K
is less than the first proper key in diagonal α + 2. So, by a diagram chasing,
the search path of K is either D · · · D or D · · · Dd · · · d. Hence, J(K) 6 2.
(2) Suppose K is absent and s < α+1. By a diagram chasing, the search path
of K is either D · · · D, D · · · Dd · · · d, D · · · Dd · · · dD, or D · · · Dd · · · dDd · · · d.
Hence, J(K) 6 4.
Theorem 3.9 If L is α-sorted, then J(L) 6 max{4, 2h − 2α + 2}.
Proof: Follows directly from Lemma 3.8.
Corollary 3.10 Let L be an α-sorted lattice of height h. If the time complexity of the procedure that performs the jump steps in JumpSearchLDS is
O(tP (m)), then the time complexity of JumpSearchLDS is O(max{4, 2h −
2α + 2}tP (h)).
Corollary 3.11 If L is h-sorted, then
1. J(K) 6 2 for any present key K, and
2. J(K) 6 4 for any absent key K.
Corollary 3.12 Let L be an h-sorted lattice with N proper keys. If the time
complexity of the procedure that performs the jump steps in JumpSearchLDS
is O(log(m)), then the time complexity of JumpSearchLDS is O(log(N)).
In the following theorem, we give upper bounds on the average of the
jump factors of the present and absent keys for α-sorted lattices.
Theorem 3.13 Suppose L is α-sorted and β = α/h. If 0.5 6 β 6 1, then
1. The average of the jump factors for the present keys is less than or
equal to
(2β 3 − 4β 2 + 2β)h2 − (3β 2 − 6β + 1)h + β − 3
h+1
≈ (2β 3 − 4β 2 + 2β)h − (3β 2 − 6β + 1).
14
2. The average of the jump factors for the absent keys is less than or equal
to
(β 2 − 2β + 1)h2 + 4h − 4
h+1
≈ (β 2 − 2β + 1)h + 4,
provided that the probability the search for a randomly chosen absent
key will terminate in a cell on row 1 and column t for each 2 6 t 6 h+2
is 1/(h + 1).
Proof: To prove this theorem, we only need to use Lemmas 3.4, 3.8, and
basic algebraic summations, then substitute α = βh.
Remark 3.14 The functions f (β) = 2β 3 − 4β 2 + 2β and g(β) = β 2 − 2β + 1
are decreasing on [0.5, 1] with f (0.5) = g(0.5) = 0.25 and f (1) = g(1) = 0.
So, from Theorem 3.13, when β approaches 1, the average of the jump factors
for the present keys approaches 2 and the average of jump factors for the
absent keys approaches 4. Also, from Theorem 3.9, J(L) 6 max{4, (2 −
2β)h + 2}.
From Theorem 3.13, the average of the jump factors of a lattice data
structure varies inversely with the degree of sortedness of the lattice. So,
in order for JumpSearchLDS to compete with searching structures of time
complexity O(log(N)), we need to maintain LDS with high degrees of sortedness. In Algorithm 5, SortLDS, we provide a procedure for incrementally
increasing the degrees of sortedness of LDS. SortLDS can be used during the
system idle time to increase the degree of sortedness of the lattice.
In the following theorem we give the time complexity of SortLDS.
Theorem 3.15 The√time complexity of SortLDS algorithm of a LDS with
N proper keys is O( N ).
Proof: Straightforward.
15
Algorithm 5 Sorting Algorithm of the LDS
SortLDS(T , h)
// Given a LDS T of height h.
// Swap the keys in the first two cells that prevent T from being h-sorted.
i = 4;
while (i 6 h + 1 and the last proper key on diagonal i is less than the
first proper key on diagonal i + 1) do
i = i + 1;
end while
if i = h + 2 then
break; // The lattice is h-sorted
else
let C be the second cell of diagonal i + 1
let F be the first cell of diagonal i with key larger than the key in C;
swap the keys in F and C;
Outward(T , C);
end if
4
Experimental Evaluation of LDSs and Skip
Lists
In this section, we examine the performance of the LDS searching algorithms
for lattices with varying jump factors. We compare the time performance
values of SearchLDS, JumpSearchLDS, and skip lists search. At the end of
the section, we examine the effect of using SortLDS on the jump factor of
the lattice during the system idle time.
A skip list is a probabilistic searching data structure, invented by William
Pugh in the late 1980’s [6, 8]. Skip lists have most of the desirable properties
of balanced trees such as AVL trees and self-adjusting trees with O(log(N))
average time for most operations, yet they are more simple and space efficient.
The rationale for comparing LDSs with skip lists is twofold. First, skip
lists are among the newest searching data structures, they are compared
against AVL trees and self-adjusting trees in [8], so by comparing LDSs
with skip lists we gain insight into the relationship between these important
structures and LDSs. Second, the most important advantages of LDSs are
their simplicity and space efficiency, but these are claimed to be the most
important advantages of skip lists too.
16
The LDS algorithms are implemented in C programming language using
Microsoft Visual Studio 2010 Compiler. Skip lists source C code is obtained
from William Pugh by anonymous ftp to ftp.cs.umd.edu. The proper keys
are randomly chosen from the set [1, RAND MAX] = [1, 2147483647] and
stored in a file. Then the file is used to build and search lattices of varying
heights and jump factors. Same keys are used to build and search skip
lists. A binary-search-like procedure is used to perform the jump steps in
JumpSearchLDS. The time performance values are the CPU time, measured
in microseconds, on Intel(R) Core(TM) i5-2400 CPU @ 3.10GHz with 4.0GB
RAM computer. Carrying out the experiments using the GNU C compiler
on a Unix environment shows similar comparable results.
From Lemma 3.4, the number of each of the d and D movements in the
search path for any key depends only on the location where the search path
terminates. So, the performance of SearchLDS is independent of the jump
factor of the lattice. In Table 2, we compare the performance of SearchLDS,
skip lists search, and JumpSearchLDS for lattices with varying jump factors. The time is the average searching time of the present keys, h is height
of the lattice, and N is the number of keys. As we see from the table,
the performance of JumpSearchLDS is similar to SearchLDS when β = 0.8,
JumpSearchLDS starts to perform better than SearchLDS when β > 0.9,
that is when more than 81% of the keys are completely sorted. On the other
hand, the performance of JumpSearchLDS is similar to skip lists search when
β = 0.9, JumpSearchLDS starts to perform better than skip lists search
when β > 0.95, that is when more than 90% of the keys are completely
sorted. When β = 1 the lattice is completely sorted and searching using
JumpSearchLDS is similar to searching a sorted array using binary search.
Table 2: Timings of SearchLDS, skip lists search, and JumpSearchLDS
h
N
SearchLDS Skip List
JumpSearchLDS
β = 0.8 β = 0.9 β = 0.95 β = 1.0
10
55
0.048
0.059
0.064
0.061
0.061
0.060
50
1275
0.190
0.144
0.215
0.155
0.134
0.128
100 5050
0.347
0.207
0.370
0.223
0.176
0.157
500 125250 1.840
0.481
1.686
0.678
0.357
0.228
1000 500500 3.990
0.780
3.600
1.284
0.557
0.267
If the lattice is not h-sorted, then
√
N calls to SortLDS increase the degree
17
of sortedness of the lattice by at least 1, and hence reduce the average of the
jump factors of the present keys of the lattice. So, using SortLDS during the
system idle time will incrementally adjust the lattice, by increasing its degree
of sortedness, and hence improves the performance of JumpSearchLDS.
In order to simulate the performance of SortLDS during the system idle
time, we will assume the following scenario: the system starts with an hsorted lattice with N proper keys, when the system is not idle the operations
search, insert, and delete are performed, when the system is idle SortLDS
is repeatedly called, the system goes through idle time when 0.1N new keys
are inserted, and the ratio of the insert operations to the delete operations
is equal to 2. The purpose of this hypothetical scenario is to give us a rough
idea about the impact of using SortLDS, during the system idle time, on the
average of the jump factors of the present keys of the lattice. Let γ to be
the ratio of the number of times SortLDS is called to the number of times
the insert and delete operations are called. √Since the time complexity of the
insert, delete, and SortLDS procedures is N, then, in Table 3, instead of
calculating the average of the jump factors of the present keys of the lattice
with specific ratios of the idle time to the time of insert and delete operations,
we will calculate the average of the jump factors of the present keys for the
lattice with varying values of γ.
Table 3: Average of the jump factors of the present keys for incrementally
adjusted lattices using SortLDS
h
Average jump factor with varying γ
γ=0 γ=3 γ=4 γ=5 γ=6
10
2.2
1.82
1.77
1.77
1.77
50
5.16
3.86
2.95
2.11
1.96
100 8.03
5.81
4.14
2.75
1.97
500 29.11 20.47 14.27 8.02
2.43
1000 52.47 37.46 25.66 13.87 3.11
In Table 4, we calculate the average of the jump factors of lattices of
varying heights and degrees of sortedness.
From Tables 3, 4, the average of the jump factors of the present keys for
incrementally adjusted lattices using SortLDS with γ = 0, respectively γ = 5,
is close to the average of the jump factors of the present keys for sorted lattices
with β = 0.8, respectively β = 0.90. So, using the above scenario, when the
18
Table 4: Average of the jump factors of the present keys for lattices with
varying degrees of sortedness
h
Average jump factor with varying β
β = 0.8 β = 0.9 β = 0.95 β = 1.0
10
1.93
1.75
1.64
1.64
50
4.45
2.70
2.05
1.92
100 7.26
3.53
2.38
1.96
500 29.53
10.22
4.26
1.99
1000 57.37
18.57
6.54
2.00
system starts β = 1.0, and searching using JumpSearchLDS is similar to
searching a sorted array using binary search. As the operations insert and
delete are performed, β deceases from 1.0 to 0.8 and JumpSearchLDS slows
down and becomes similar to SearchLDS. Using SortLDS with γ = 5 increases
β from 0.8 to 0.90, and searching using JumpSearchLDS becomes similar to
skip lists search.
5
Conclusions
In this paper, we have formalized the definition of the lattice data structure
and presented its basic operations. We have shown that the LDS is as space
efficient as the array and it is very suitable for caching. The worst case
time complexity of the LDS basic operations is better than the corresponding worst case time complexity of most of other data structures operations.
In order to make this structure more attractive in term of the average case
time complexity of the searching procedure, we have devised a general jump
searching algorithm of time complexity O(J(L) log(N)).
We have also pro√
vided a sorting algorithm of time complexity O( N) that can be used to
reduce J(L) during the system idle time. One possible direction for further research is to explore using randomized searching procedures where the
search starts at different locations of the lattice instead of the second cell
of diagonal h + 2. Another possible direction is to perform a comprehensive
experimental comparison of the LDS with B + -Trees and T -Trees to find out
the best suitable structure for indexing in main memory. In [5], we showed
that the LDS is a very robust structure for the order-statistic operations.
19
Acknowledgements. I would like to thank Simon Berkovich for his helpful
comments and suggestions. I am also thankful to Regina Nuzzo who carefully
read the manuscript of this paper.
References
[1] J. Bentley, F. Leighton, M. Lepley, D. Stanat, and J. Steele. A Randomized Data Structure for Ordered Sets. Advances in Computing Research,
5:413-428, 1989.
[2] S. Berkovich. Organization of Associative Memory Operations With Lattice Structures. Proceedings of the 35th IEEE Midwest Symposium On
Circuits and Systems, 2:887-890, 1992.
[3] H. Bronnimann, F. Cazals, and M. Durand. Randomized Jumplists:
A Jump-and-Walk Dictionary Data Structure. Proceedings of the 20th
Annual Symposium on Theoretical Aspects of Computer Science, LNCS,
Springer, 2607:283-294, 2003.
[4] D. Knuth. The Art of Computer Programming Vol. 3: Sorting and
Searching. Adison-Wesley Publishing Company, 1973.
[5] M. Obiedat. Time-efficient Algorithms for the Order-statistic Operations
in Lattice Data Structures. To be published.
[6] W. Pugh. A Skip List Cookbook. Computer Science Technical Report
Series, CS-TR-2286.1, 1-12, 1990.
[7] W. Pugh. Concurrent Maintenance of Skip Lists. Computer Science
Technical Report Series, CS-TR-2222.1, 1-13, 1990.
[8] W. Pugh. Skip Lists: A Probabilistic Alternative to Balanced Trees,
LNCS 382 Springer, 437-449 1989.
[9] B. Shneiderman. Jump Searching: A Fast Sequential Search Technique.
Comm. of the ACM, 21(10):831-834, 1978.
[10] D. Sleator and R. Tarjan. Self-adjusting Binary Search Trees. Comm. of
the ACM, 32(3):652-666, 1985.
20
[11] S. Štrbac-Savić and M. Tomašević. Comparative Performance Evaluation of the AVL and Red-Black Trees, Proceedings of the Fifth Balkan
Conference in Informatics, 14-19, 2012.
[12] B. Stroustrup. Software Development for Infrastructure. Computer,
45(1):47-58, 2012.
[13] A. Yong. What is a Young Tableau?. Notice of the AMS, 54(2):240-241,
2007.
21
| 8cs.DS
|
Adaptive Estimation in Structured Factor Models with
Applications to Overlapping Clustering
arXiv:1704.06977v3 [stat.ME] 22 Mar 2018
Xin Bing∗
Florentina Bunea†
Yang Ning‡
Marten Wegkamp§
Abstract
This work introduces a novel estimation method, called LOVE, of the entries and structure of a
loading matrix A in a sparse latent factor model X = AZ + E, for an observable random vector
X in Rp , with correlated unobservable factors Z ∈ RK , with K unknown, and independent
noise E. Each row of A is scaled and sparse. In order to identify the loading matrix A, we
require the existence of pure variables, which are components of X that are associated, via A,
with one and only one latent factor. Despite the fact that the number of factors K, the number
of the pure variables, and their location are all unknown, we only require a mild condition on
the covariance matrix of Z, and a minimum of only two pure variables per latent factor to show
that A is uniquely defined, up to signed permutations. Our proofs for model identifiability are
constructive, and lead to our novel estimation method of the number of factors and of the set of
pure variables, from a sample of size n of observations on X. This is the first step of our LOVE
algorithm, which is optimization-free, and has low computational complexity of order p2 . The
second step of LOVE is an easily implementable linear program that estimates A. We prove
that the resulting estimator is minimax rate optimal up to logarithmic factors in p.
The model structure is motivated by the problem of overlapping variable clustering, ubiquitous in data science. We define the population level clusters as groups of those components of
X that are associated, via the sparse matrix A, with the same unobservable latent factor, and
multi-factor association is allowed. Clusters are respectively anchored by the pure variables,
and form overlapping sub-groups of the p-dimensional random vector X. The Latent model
approach to OVErlapping clustering is reflected in the name of our algorithm, LOVE.
The third step of LOVE estimates the clusters from the support of the columns of the
estimated A. We further guarantee cluster recovery with zero false positive proportion, and
with false negative proportion control. The practical relevance of LOVE is illustrated through
the analysis of an RNA-seq data set, devoted to determining the functional annotation of genes
with unknown function.
Keywords:
Overlapping clustering, latent model, identification, high dimensional estimation, minimax
estimation, pure variables, group recovery, support recovery, sparse loading matrix, matrix factorization.
∗
Department of Statistical Science, Cornell University, Ithaca, NY. E-mail: [email protected].
Department of Statistical Science, Cornell University, Ithaca, NY. E-mail: [email protected].
‡
Department of Statistical Science, Cornell University, Ithaca, NY. E-mail: [email protected].
§
Department of Mathematics and Department of Statistical Science, Cornell University, Ithaca, NY. E-mail:
[email protected].
†
1
1
Introduction
In this work we consider the problem of estimating the p × K sparse loading matrix A that
parametrizes the factorization of a zero-mean observable random vector, X ∈ Rp as
X = AZ + E
(1)
from n i.i.d. realizations of X. The zero mean random vector Z ∈ RK is unobservable, and can be
viewed as a latent factor vector. E ∈ Rp is a zero-mean, unobservable random noise vector, with
independent entries. The number of factors K is not known, and both p and K are allowed to grow,
and be larger, than n. Factor models have been used as dimension reduction devices in virtually
any scientific discipline for nearly a century, and generated an enormous amount of literature. We
refer to the classical monographs of Bollen [1989] and Anderson [2003] for earlier work, and to
Izenman [2008] for a more recent survey and applications.
In this work, we revisit some of the open problems in factor model definition and estimation,
and also consider one of their much less explored usage, to overlapping clustering. For the latter,
we deem two components Xi and Xj of X similar if they have non-zero association, via the matrix
A, with the same latent factor Za . Similar variables are placed in the same cluster, Ga :
Ga := j ∈ {1, . . . , p} : Aja 6= 0 ,
for each a ∈ {1, . . . , K}.
(2)
Since each Xj can be associated with multiple latent factors, the clusters will overlap. The problem
of overlapping clustering is of wide-spread interest in virtually any scientific area, for instance in
neuroscience [Craddock et al., 2012, 2013] and genetics [Jiang et al., 2004; Wiwie et al., 2015], to
give a very limited number of examples. The solutions are typically algorithmic in nature, and their
quality is assessed against a ground scientific truth or via extensive simulation studies, for instance
Krishnapuram et al. [2001]; Bezdek [2013], among many others. These problems have not received
a systematic analysis in the statistical literature and, in particular, the problem of estimating
overlapping clusters of variables, with theoretical guarantees, remains largely unexplored.
We propose model-based clustering via A. However, A cannot be uniquely defined in (1),
without further restrictions, a phenomenon well understood over six decades ago. Most notably,
Anderson and Rubin [1956] provided an in-depth analysis of this problem, and proved that in the
absence of conditions on A and C := Cov(Z), A is not identifiable in model (1). We revisit some
of these conditions here, with a view towards our application to overlapping clustering.
Using the latter as motivation, we formalize our first modeling assumption on A. We consider
models (1) in which each row of A is scaled, to avoid scale ambiguities, and sparse, to avoid the
trivial situation in which each component Xj is associated with all latent factors. Specifically, we
assume that:
PK
(i)
a=1 |Aja | ≤ 1, and each row Aj· := (Aj1 , . . . , AjK ) is sparse, for j ∈ {1, . . . , p}.
P
The inequality in (i) allows for a |Aja | = 0, which renders more flexibility to model (1), relative
P
to the more commonly used equality conditions. If a |Aja | = 0, then Xj = Ej , and Xj is not
associated with any of the latent factors, via this model. The interpretation for clustering is that
2
the corresponding Xj = Ej does not belong to any cluster given by this model, which is a desired
feature in many practical applications, including the one presented in this paper in Section 6.
Condition (i) alone cannot ensure that A in model (1) is uniquely defined, as one can still
construct an invertible matrix Q such that AZ = AQQ−1 Z, with both A and AQ satisfying (i).
Moreover, A and AQ may not have the same sparsity pattern, creating ambiguity in the cluster
definition. We introduce below two additional conditions that allow us to show, in Section 2 below,
that A is identifiable. We defer a detailed discussion of related conditions to Section 4.4.
We call the following condition the pure variable assumption. Informally, it requires the existence
of at least two pure variables Xj , which are components of X associated with one and only one
latent factor. In Section 2 we provide examples that show that if pure variables do no exist, A in
(1) is not uniquely defined.
(ii) For every a ∈ {1, . . . , K}, there exist at least two indices j ∈ {1, . . . , p} such that |Aja | = 1
and Ajb = 0 for all b 6= a.
This assumption has an immediate practical implication for variable clustering. Clusters given by
(2) are defined relative to the unobservable factor Za . A pure variable Xj is an observable proxy
of Za , and helps explain the nature of Ga . For future reference, we let I denote the index set
corresponding to pure variables. In psychology, these variables are called factorially simple items
(McDonald [1999]). A similar condition can be traced back to the econometrics literature, and an
early reference is Koopmans and Reiersol [1950], further discussed in Anderson and Rubin [1956],
who called it “zero elements in specified positions”. These works prove that (ii) corresponding to a
known set I is a sufficient condition for identifying A, for latent factors with arbitrary correlations.
However, full generality on the positive definite covariance matrix C of the latent factors comes at
the steep price of knowing I a priori, which is often unrealistic in practice. Appropriate conditions
on C that guarantee identifiability of I in (ii), in the general case when I is not known, have not
been investigated for the general model (1), to the best of our knowledge.
However, the problem of identifying A under (ii), with I unknown, has been revived more
recently, in the particular case of random vectors X and Z with only non-negative values, and when
A also has only non-negative entries. This set-up corresponds to the area of non-negative matrix
factorization (NMF), in which one studies positive matrix factorizations of the type X = AZ + E,
where the observed data X is a p × n matrix, Z is the K × n unobservable matrix of the latent
vectors values, and E is the p × n noise matrix. In this context, when E = 0, and conditioning
on Z, Donoho and Stodden [2004] was among the first to propose a condition similar to (ii), with
I unknown, coupled with appropriate conditions on Z, leading to an NMF decomposition with
unique factors. Moreover, the unique determination of I in NMF under (ii) was solved in Bittorf
et al. [2012], for known K, and for scaled NMF models, in which the columns of X, Z and A sum
up to 1, under the condition that no row of a scaled version of Z is a convex combination of the
other rows. This condition implies, in particular, that there are no duplicate (scaled) latent factors,
and similar results have been established, in the context of topic models in Arora et al. [2013].
We study here model (1) under (ii), with I and K unknown, under the restriction on the row
sums of A given in (i), but with X, Z and A allowed to have arbitrary signs. This makes the
3
direct import of conditions on Z employed in the NMF literature difficult, and we provide below
a condition on C that is appropriate for our model, and is a key ingredient in the identification
of I in Section 2 below. In this work we depart from the widely used independence assumption
between the latent factors, see for instance, Anderson and Rubin [1956]; Anderson [2003]; Bollen
[1989]; Everitt [1984]; Izenman [2008] and our discussion in Section 4.4. We allow C to be far from
a diagonal, or even a sparse matrix. Instead, we require
(iii) ∆(C) := mina6=b (Caa ∧ Cbb − |Cab |) > 0,
where a ∧ b := min(a, b). If (iii) holds, then Cov(Za ± Zb ) = Var(Za ) + Var(Zb ) ± 2 · Cov(Za , Zb ) ≥
Caa + Cbb − 2|Cab | > 0, which implies that the latent factors are different, up to signs, that is
|Za | =
6 |Zb | a.s. for any a 6= b. Thus, (iii) can be viewed as a minimal natural requirement for the
study of this problem.
We note that under the pure variable assumption (ii), if I is known in advance, the arguments
employed in the proof of Theorem 2 of Section 2 show that (iii) is not required, consistent with
the existing literature on general latent models. When I is not known, our condition (iii) implies
that we do not have duplicate latent factors, up to signs, consistent with similar requirements in
NMF, as explained above. Therefore, condition (iii) can be viewed as bridging the gap between the
classical literature in general latent models and the more recent literature devoted to the specialized
NMF-type models.
Summarizing, this work is devoted to estimation in model (1) with A satisfying (i) - (ii), and
with the positive definite matrix C satisfying (iii). The number of factors K is not known, and both
K and p are allowed to grow and be larger than n. In Section 1.1 below we present our contributions
and the structure of this paper. A detailed contrast with existing literature is presented in Section
4.4.
1.1
Our contributions
1. Identifiability of the allocation matrix A in sparse latent models with pure variables.
We show, in Proposition 2 of Section 2, that the allocation matrix A, which is allowed to have entries
of arbitrary signs, is uniquely defined, up to trivial orthogonal transformations, namely signed
permutation matrices. This is a consequence of one of our main results, Theorem 1 of Section 2.
In this result we highlight and resolve the main difficulty in this problem, that of distinguishing
between the pure variables and the non-pure variables. Both proofs are constructive, and show that
the pure variable set I and allocation matrix A can be determined uniquely from Σ := Cov(X).
Moreover, the number of factors K is not assumed to be known, and its determination is also a
consequence of Theorem 1. To the best of our knowledge, these are new results in both the latent
factors literature and other related matrix factorization literature. We comment on connections to
related results in Section 4.4.
2. Estimation of the allocation matrix A and of the overlapping clusters. The LOVE
b of the sparse and structured matrix A that is tailored to our
algorithm. We provide an estimator A
model specifications. Our approach follows the constructive techniques used in our identifiability
b an estimator of the pure variable set I, and K,
b an estimator of the
proofs. We first construct I,
4
number of clusters, K. These are used to estimate the rows in A corresponding to pure variables.
The remaining rows of A are estimated via an easily implementable linear program that is tailored
to this problem. As part of our procedure, we also develop a novel estimator (12) and (13) of a
precision matrix, C −1 . Our procedure is presented in Sections 3.1, 3.2 and 3.3, respectively. To the
best of our knowledge, our estimation strategy is new, and complements the large body of literature
in factor models. In particular, we do not resort to optimizing a complicated quasi likelihood
function via computationally demanding EM algorithms. These algorithms require, in addition, a
notoriously delicate initialization, especially in high dimensions, and typically only convergence to
a stationary point can be guaranteed, see Rubin and Thayer [1982]. Moreover, as our procedure is
not Bayesian, we do not employ distributional assumptions to construct our estimator. In Section
b using the estimated allocation matrix A.
b The
3.4, we build a collection of overlapping clusters G,
combined procedure is summarized in a new algorithm, LOVE, highlighting our Latent model
approach to OVErlapping clustering.
3. Statistical guarantees. Our estimation procedure does not depend on distributional assumptions, but for the purpose of our statistical analysis, and in particular our minimax analysis, we
assume that X has a sub-Gaussian distribution. LOVE, for appropriate choices of tuning parameters, recovers the population level clusters with a zero false positive proportion and generally low
false negative proportion, with high probability, and under a mild condition on the cluster separation as measured by the quantity ∆(C). This is a direct consequence of a number of results
regarding estimation of identifiable loading matrices in factor models satisfying (i) - (iii) and, to
the best of our knowledge, are all new.
(1) Consistent estimation of the number of factors K;
(2) Control of the relationship between Ib and I for A with entries of arbitrary strength. In
particular, we show I ⊆ Ib ⊆ I ∪ J1 , where we carefully define and characterize J1 as the set
of quasi-pure variables.
b A), defined below, for all q ≥ 1, in particular for
(3) Minimax lower bounds on the norms Lq (A,
q = +∞, for A given by model (1) under (i) - (iii).
(4) Attainment of these bounds, showing that our procedure is minimax optimal and adaptive.
b
(5) Control of the relationship between the support of A and the support of A.
(6) Control of cluster recovery.
The details are given in Sections 4.2 and 4.3. In particular, we emphasize that (2) above, proved
in Theorem 3 of Section 4.1, guarantees recovery of I with minimal mistakes. This result does not
require the necessary, yet unpleasant, signal strength restrictions encountered in the typical exact
support recovery literature. However, under such restrictions, we also obtain Ib = I, with high
probability, in Remark 2 of Section 4.1. Since placing restrictions on the entries in A reduces the
number of configurations of interest, the more general result (2) is a new and practically relevant
result for pure variable recovery.
Results (3) and (4) are given in Theorems 4 and 5, respectively, of Section 4.2. We consider the
5
loss function
b − Ak∞,q ,
b A) := min kAP
Lq (A,
P
1 ≤ q ≤ ∞,
with the minimum taken over all K × K signed permutation matrices P and
kAk∞,q := max kAi· kq = max
1≤i≤p
1≤i≤p
K
X
j=1
1/q
|Aij |q
,
is the maximum `q norm of the rows of A. We let s = maxi∈[p] kAi· k0 be the row-sparsity index.
As explained in Section 3.3, our estimation procedure is closely connected to sparse estimation
in errors-in-variable regression models with design matrix C, and the error rates we obtain can
be interpreted in that context. We show that the error of estimation with respect to the Lq loss
function, for each q, is proportional to s1/q n−1/2 , multiplied by kC −1 k∞,1 . This is consistent with
the most recent results regarding error rates expressed in terms of the `q -sensitivity of C in Belloni
et al. [2017], as discussed in Section 4.2. The results hold up to logarithmic factors in p and s.
Results (5) and (6) are presented in Theorem 6 of Section 4.3. Moreover, we can further partition
the variables in each cluster into two signed sub-groups consistently. In our model formulation,
A is allowed to have positive and negative entries. Since A can only be identified up to signed
b However, we can identify consistently
permutations, one cannot expect sign consistency for A.
the two sub-groups of each cluster that contain variables that are associated with the common
latent factor in the same direction, although the direction itself is not identifiable. These results
are presented in Section 4.3.
We conduct an extensive simulation study in Section 5 to assess the numerical performance of
our proposed strategy. The study confirms our theoretical findings. We conclude the validation
of our approach with a data analysis, devoted to determining the functional annotation of genes
with unknown function. Our analysis confirms existing biological ground truths, as our procedure
tends to cluster together genes with the same Gene Ontology (GO) biological process, molecular
function, or cellular component terms.
We summarize our contributions in the table below, restricting attention to estimation in general
latent models (1) under (i) - (iii), without any further restrictions on the signs or scales of X and
Z.
Model (1) under (i) - (iii)
Identifiability Conditions
Estimation: I
Estimation: A
Guarantees: I
Guarantees: A
Cluster recovery
Our results
Existence of I; I and K unknown; mild conditions on C.
Optimization-free. Runs in O(p2 ) time
Not MLE-based. Uses linear program. Runs in O(p2 + pK).
Recovered
Finite sample lower bounds for k k∞,q recovery
Adaptive finite sample upper bounds
p and K can grow with n
Guaranteed
Existing results in factor models
Existence of known I; general C
×
E-M type, with known I. Computationally involved. Not guaranteed to estimate A
×
×
Row-wise asymptotic normality of MLE-type estimators
p can grow with n; K fixed.
×
In Section 4.4 we discuss our results further, and provide a detailed comparison between our
work and related contributions. All proofs are deferred to Section A of the supplementary material.
6
1.2
Notation
We use the following notation throughout this paper. For the n consecutive integer set starting
from 1, we write [n] = {1, . . . , n}. The sign of any generic number N is denoted by sign(N ). For
any m × d matrix M and index sets I ⊆ {1, . . . , m} and J ⊆ {1, . . . , d}, we write MI to denote
the |I| × d submatrix (Mij )i∈I,1≤j≤d of M consisting of the rows in the index set I, while we denote by MIJ the |I| × |J| submatrix with entries Mij , i ∈ I and j ∈ J. The ith row of M is
denoted by Mi· , and the jth column of M is denoted by M·j . Let kM k∞ = max1≤j≤m,1≤k≤d |Mjk |,
Pd
Pd
P
P
2 )1/2 , kM k
Mjk
kM k1 = 1≤j≤m,1≤k≤d |Mjk |, kM kF = ( m
∞,1 = max1≤j≤m
k=1 |Mjk |
k=1
j=1
Pm
and kM k1,∞ = max1≤k≤d j=1 |Mjk | denote the matrix max norm, matrix `1 norm, matrix Frobenius norm, matrix 1 norm and matrix ∞ norm. We denote by h · i the Frobenius scalar product.
P
For a vector v ∈ Rd , define kvkq = ( di=1 |vj |q )1/q for 1 ≤ q < ∞, kvk∞ = max1≤j≤d |vj | and
kvk0 = |supp(v)|, where supp(v) = {j : vj 6= 0} and |A| is the cardinality of the set A. We write
M T for the transpose of M and diag(m1 , . . . , md ) for the d × d diagonal matrix with elements
m1 , . . . , md on its diagonal, while diag(M ) is the diagonal matrix obtained from the diagonal elements of a square matrix M . We use c0 , c1 , . . . to denote generic constants. Finally, a signed
permutation matrix is an orthogonal matrix that permutes the index and switches the sign within
each column. We write HK as the hyperoctahedral group of K × K signed permutation matrices.
2
Identifiability
In this section we show that the allocation matrix A given by Model (1) and (i) - (iii) is identifiable,
up to multiplication with a signed permutation matrix.
For any A ∈ Rp×K which satisfies Model (1), we can partition the set [p] = {1, . . . , p} into two
disjoint parts: I and its complement J := [p] \ I such that for each row Ai· of AI , there exists only
one a ∈ [K] such that |Aia | = 1. We name I the pure variable set and J the non-pure variable set.
Specifically, for any given A, the pure variable set I is defined as
[
I(A) :=
Ia ,
Ia := {i ∈ [p] : |Aia | = 1, Aib = 0, for any b 6= a} .
(3)
a∈[K]
We write I(A) in (3) to emphasize that the pure variable set is defined relative to A. In the
following, we will not write this explicitly when there is no confusion. We also note that the sets
{Ia }1≤a≤K form a partition of I.
To show the identifiability of A, it suffices to show that AI and AJ are identifiable, respectively,
up to signed permutation matrices. By the definition of AI , this matrix is identifiable provided the
partition of the pure variable set I is. The identifiability of I, and thus the problem of distinguishing
between the sets I and J, on the basis of the distribution of X alone, is the central challenge in
this problem. We meet this challenge in Theorem 1 below: part (a) offers a necessary and sufficient
characterization of I; part (b) shows that, as a consequence, I and its partition I := {Ia }1≤a≤K
are identifiable. Let
Mi := max |Σij |
(4)
j∈[p]\{i}
7
be the largest absolute value of the entries of row i of Σ excluding |Σii |. Let Si be the set of indices
for which Mi is attained:
Si := j ∈ [p] \ {i} : |Σij | = Mi .
(5)
Theorem 1. Assume that model (1) and (i) - (iii) hold. Then:
(a) i ∈ I
⇐⇒
Mi = Mj for all j ∈ Si .
(b) The pure variable set I can be determined uniquely from Σ := Cov(X). Moreover, its partition
I := {Ia }1≤a≤K is unique and can be determined from Σ up to label permutations.
The identifiability of the allocation matrix A and that of the collection of clusters G = {G1 , . . . , Gk }
in (2) use the results from Theorem 1 in crucial ways. We state the result in Theorem 2 below.
Theorem 2. Assume that Model (1) with (i) - (iii) holds. Then, there exists a unique matrix A,
up to a signed permutation, such that X = AZ + E. This implies that the associated overlapping
clusters Ga , for 1 ≤ a ≤ K, are identifiable, up to label switching.
Remark 1. We show below that the pure variable assumption (ii) is needed for the identifiability
of A, up to a signed permutation. Assume that X = AZ + E satisfies (i) and (iii), but not (ii). We
construct an example in which X can also be written as X = ÃZ̃ + E, where à and Z̃ satisfy the
same conditions (i) and (iii), respectively, but à 6= AP for any K × K signed permutation matrix
P and à may have a sparsity pattern different from A. To this end, we construct à and Z̃ such
that ÃZ̃ = AZ. Let à = AQ and Z̃ = Q−1 Z, for some K × K invertible matrix Q to be chosen. By
the definition of C, Q needs to satisfy C = QCQT . In addition, we need to guarantee that à = AQ
satisfies (i). For simplicity, we set K = 3. The following example satisfies all our requirements:
p
1 0 0
1/2
3/8 0
p
C = 0 2 0 , Q = −2 3/8 1/2 0 .
0 0 3
0
0
1
It is easy to verify that C = QCQT holds. For any 1 ≤ j ≤ p, consider
r
r
1 1 2 3 1 2
Aj· = ( ,
, −
)
4 8 3 4 8 3
then
1
Ãj· = Aj· Q = (0,
8
r
3
1
+
2 16
r
2 3 1
, −
3 4 8
r
2
),
3
which also satisfies condition (i). However, Aj· and Ãj· have different sparsity patterns. Thus, if
the matrix A does not satisfy (ii), A is generally not identifiable.
3
Estimation
We develop estimators from the observed data, which is assumed to be a sample of n i.i.d. copies
X (1) , . . . , X (n) of X ∈ Rp , where p is allowed to be larger than n. Our estimation procedure consists
8
of the following four steps:
(1) Estimate the pure variable set I, the number of clusters K and the partition I;
(2) Estimate AI , the submatrix of A with rows Ai· that correspond to i ∈ I;
(3) Estimate AJ , the submatrix of A with rows Aj· that correspond to j ∈ J;
(4) Estimate the overlapping clusters G = {G1 , . . . , GK }.
3.1
Estimation of I and I
Given the different nature of their entries, we estimate the submatrices AI and AJ separately. For
the former, we first estimate I and its partition I = {I1 , . . . , IK }, which can be both uniquely
constructed from Σ, as shown by Theorem 1. We use the constructive proof of Theorem 1 for this
step, replacing the unknown Σ by the sample covariance matrix
n
X
b= 1
Σ
X (i) (X (i) )T .
n
i=1
Specifically, we iterate through the index set {1, 2, . . . , p}, and use the sample version of part (a)
of Theorem 1 to decide whether an index i is pure. If it is not deemed to be pure, we add it to the
set that estimates J. Otherwise, we retain the estimated index set Sbi of Si defined in (5), which
corresponds to an estimator of Mi given by (4). We then use the constructive proof of part (b) of
Theorem 1 to declare Sbi ∪ {i} := Ib(i) as an estimator of one of the partition sets of I. The resulting
procedure has complexity O(p2 ), and we give all the specifics in Algorithm 1 of Section 3.5. The
algorithm requires the specification of a tuning parameter δ, which will be discussed in Section 5.1.
3.2
Estimation of the allocation submatrix AI
b K
b and Ib = {Ib1 , . . . , Ibb } from Algorithm 1, we estimate the matrix AI by
Given the estimators I,
K
b ×K
b matrix with rows i ∈ Ib consisting of K
b − 1 zeros and one entry equal to either +1 or −1
a |I|
b
as follows. For each a ∈ [K],
bia = 1. Note that A
bia can only be +1 or −1 by
(1) Pick an element i ∈ Iba at random, and set A
the definition of a pure variable.
bja = sign(Σ
b ij ).
(2) For the remaining j ∈ Iba \ {i}, we set A
This procedure induces a partition of Iba = Iba1 ∪ Iba2 , where Iba1 and Iba2 are defined below:
(
bka = A
bla , for k, l ∈ Iba1 or k, l ∈ Iba2
A
.
bka 6= A
bla , for k ∈ Ib1 and l ∈ Ib2
A
a
3.3
(6)
a
Estimation of the allocation submatrix AJ
We continue by estimating the matrix AJ , row by row. To motivate our procedure, we begin by
highlighting the structure of each row Aj· of AJ , for j ∈ J. We recall that Aj· is sparse, with
kAj· k1 ≤ 1, for each j ∈ J, as specified by assumption (i). In addition, model (1) subsumes a
9
further constraint on each row Aj· of A, as explained below. To facilitate notation, we rearrange Σ
and A as follows:
"
#
" #
ΣII ΣIJ
AI
Σ=
and A =
.
ΣJI ΣJJ
AJ
Model (1) implies the following decomposition of the covariance matrix of X:
"
# "
# "
#
ΣII ΣIJ
AI CATI AI CATJ
ΓII
0
Σ=
=
+
.
ΣJI ΣJJ
AJ CATI AJ CATJ
0 ΓJJ
In particular, ΣIJ = AI CATJ . Thus, for each i ∈ Ia with some a ∈ [K] and j ∈ J, we have
Aia Σij =
A2ia
K
X
Ajb Cab =
b=1
K
X
T
Ajb Cab = Ca·
Aj· .
(7)
b=1
Averaging display (7) over all i ∈ Ia yields
1 X
T
Aia Σij = Ca·
Aj· ,
|Ia |
i∈Ia
for each a ∈ [K].
(8)
For each j ∈ J, we let β j = Aj· and
T
X
X
1
1
θj =
AiK Σij .
Ai1 Σij , . . . ,
|I1 |
|IK |
i∈I1
(9)
i∈IK
Since Aia ∈ {−1, 1}, for each i ∈ Ia and a ∈ [K], the entries of θj are respective averages of the
sign corrected entries of Σ corresponding to the partition of the pure variable set. Summarizing,
modeling assumption (i) and equation (8) above show that the estimation of AJ reduces to estimating, for each j ∈ J, a K-dimensional vector β j that is sparse, with norm kβ j k1 ≤ 1, and that
satisfies the equation θj = Cβ j . Both C and θj , for each j ∈ J, can be estimated directly from the
b we estimate the a-th entry of θj by
data as follows. For each j ∈ J,
1 X b b
b
Aia Σij , a ∈ [K],
θbaj =
|Iba |
(10)
i∈Iba
and compute
baa =
C
1
|Iba |(|Iba | − 1)
X
i,j∈Iba ,i6=j
b ij |,
|Σ
bab =
C
1
|Iba ||Ibb |
X
i∈Iba ,j∈Ibb
bia A
bib Σ
b ij ,
A
(11)
b and b ∈ [K]
b \ {a} to form the estimator C
b of C. The estimates (10) and (11) rely
for each a ∈ [K]
crucially on having first estimated the pure variables and their partition, according to the steps
described in Sections 3.1 and 3.2 above.
We could cast the estimation of β j as a sparse, errors in variables regression problem by viewing
b as observed quantities in the model given by θbj = Cβ j +ε and C
b = C +η, with “unobserved
θbj and C
10
b − C. Although the “errors” ε and η are heavily correlated in this
errors” ε = θbj − θj and η = C
context, unlike the typical assumptions on the errors in such models, it is nevertheless tempting
to use an appropriate extension of the optimal estimation procedure developed for sparse errors in
variables regression in Belloni et al. [2017]. This would essentially boil down to minimizing kβ j k1
b j − θbj k∞ . While Belloni et al. [2017] showed that their conic
under a suitable constraint on kCβ
programming estimator is optimal in a minimax sense, conic programming for each row j ∈ Jb would
be computationally costly and involve too many tuning parameters. Moreover, our final goal of
support recovery of β j would require additional thresholding for each row. The latter drawback is
explicitly mentioned in the discussion following Theorem 1 in Belloni et al. [2017].
For these reasons, we do not pursue this line of estimation here. Instead, we have developed a
more computationally efficient method, that only involves two tuning parameters, but shares the
same rate of convergence of the resulting estimators as the aforementioned procedure in Belloni
et al. [2017]. We exploit the fact that the square matrix C is invertible and take the equation
b θbj , based
β j = C −1 θj as our starting point. The idea is to first construct a pre-estimator β̄ j = Ω
−1
b of the precision matrix Ω := C , followed by a sparse projection
on an appropriate estimator Ω
j
of β̄ . Alternatively, we could use a simple hard threshold in the second step.
We first motivate our proposed estimator of Ω. From the decomposition
β̄ j − β j
we immediately have
b θbj − θj ) + (Ω
b − Ω)θj
= Ω(
b θbj − θj ) + (ΩC
b − I)β j ,
= Ω(
b ∞,1 kθbj − θj k∞ + kΩC
b − Ik∞ kβ j k1 .
kβ̄ j − β j k∞ ≤ kΩk
(12)
(13)
Since we can show, in Lemma 11, that kθbj −θj k∞ has optimal convergence rate, and since kβ j k1 ≤ 1
b should ideally render values for kΩk
b ∞,1 and kΩC
b − Ik∞ that are
under our model, our estimator Ω
as small as possible. With this in mind, we propose the linear program
subject to
b t̂) = arg
(Ω,
Ω = ΩT ,
min
b K
b
t∈R+ , Ω∈RK×
b − Ik∞ ≤ λt,
kΩC
t
kΩk∞,1 ≤ t,
(14)
(15)
with tuning parameter λ. This linear programming problem clearly is tailored to our purpose, and
b adds a novel estimator for C −1 to the rich literature of precision matrix
its optimal solution Ω
estimation (Meinshausen and Bühlmann [2006]; Yuan and Lin [2007]; Friedman et al. [2008]; Cai
et al. [2011, 2016], to name a few). Its novelty consists in (a) the usage of the matrix k · k∞,1 norm,
rather than the commonly used matrix k · k1 norm, and (b) the fact that this norm appears in the
upper bound of the restriction (15). Our analysis does not require any additional assumptions on
C, i.e., no sparsity assumption on C or C −1 is imposed. In particular, we avoid the common, albeit
highly undesirable, assumption that the condition number of C is bounded [Cai et al., 2016]. After
b θbj , for each j ∈ J,
b we solve the following optimization problem
we compute β̄ j = Ω
βbj = arg min kβk1
b
β∈RK
11
(16)
subject to
kβ − β̄ j k∞ ≤ µ,
(17)
for some tuning parameter µ that is constant in j, to obtain our final estimate βbj as the optimal
solution of this linear program. This solution is sparse and properly scaled, in accordance to our
b we obtain A
b b, and merging A
bb with
model specification (i). By stacking βbj over all rows j ∈ J,
J
I
b b produces our final estimator A
b of A. Its statistical property is analyzed in Section 4 along with
A
J
precise forms of the tuning parameters.
3.4
Estimation of the overlapping groups
Recalling the definition of groups in (2), the overlapping groups are estimated by
b1 , . . . , G
bb , G
b a = i ∈ [p] : A
bia 6= 0 , for each a ∈ [K].
b
Gb = G
K
(18)
as the pure noise cluster. We can estimate G0 in (19) by
b 0 = i ∈ [p] : A
bia = 0, for all a ∈ [K]
b .
G
(20)
b with the same latent factor Za are therefore placed in the
Variables Xi that are associated (via A)
b a . To accommodate potential pure noise variables, we further define
same group G
G0 := j ∈ {1, . . . , p} : Aja = 0, for all a ∈ {1, . . . , K}
(19)
However, our main focus is on G because it completely determines G0 .
In many applications, it may be of interest to identify the sub-groups of variables that are all
either positively or negatively associated with the same latent factor. To this end, we define
n
o
G s := Gs1 , . . . , GsK , Gsa := G1a , G2a :=
(21)
i ∈ Ga : Aia > 0 , i ∈ Ga : Aia < 0 ,
and they are estimated by
s
b1 , . . . , G
bs ,
Gbs = G
b
K
n
o
s
ba =
b
ba : A
bia > 0 , i ∈ G
ba : A
bia < 0 , for each a ∈ [K].
G
i∈G
(22)
The fact that A is only identifiable up to a signed permutation matrix, has the repercussion that
the labels of the two sub-groups in Gsa are not identifiable. Thus, variables placed in the subgroups
G1a and G2a are, respectively, associated with Za in the same direction. The directions between two
sub-groups, henceforth called direction sub-groups, are opposite. This can be identified, although
the direction itself cannot. We show in Section 4 that the direction sub-groups can be identified,
and well estimated.
3.5
LOVE: A Latent variable model approach for OVErlapping clustering.
We give below the specifics of Algorithm 1, motivated in Section 3.1, and summarize our final
algorithm, LOVE in Algorithm 2.
12
Algorithm 1 Estimate the partition of the pure variables I by Ib
b δ)
1: procedure PureVar(Σ,
b
2:
I ← ∅.
3:
for all i ∈ [p] do
b ij | ≤ |Σ
b il | + 2δ
4:
Ib(i) ← l ∈ [p] \ {i} : maxj∈[p]\{i} |Σ
5:
P ure(i) ← T rue.
6:
for all j ∈ Ib(i) do
b ij | − maxk∈[p]\{j} |Σ
b jk | > 2δ then
7:
if |Σ
8:
P ure(i) ← F alse,
9:
break
10:
end if
11:
end for
12:
if P ure(i) then
13:
Ib(i) ← Ib(i) ∪ {i}
b
14:
Ib ← Merge(Ib(i) , I)
15:
end if
16:
end for
b
17:
return I.
18: end procedure
b
19: function Merge(Ib(i) , I)
20:
for all G ∈ Ib do
21:
if G ∩ Ib(i) 6= ∅ then
22:
G ← G ∩ Ib(i)
23:
return Ib
24:
end if
25:
end for
26:
Ib(i) ∈ Ib
27:
return Ib
28: end function
4
. Ib is a collection of sets
. Replace G ∈ Ib by G ∩ Ib(i)
. add Ib(i) in Ib
Statistical guarantees
We provide in this section statistical guarantees for:
b
(1a) The estimated number of clusters K;
b
(1b) The estimated pure variable set Ib and its estimated partition I;
(2)
(3)
b and its adaptation to the unknown row sparsity of A.
The estimated allocation matrix A
The individual Group False Positive Proportion (GFPP), the individual Group False Negative
Proportion(GFNP), the Total False Positive Proportion (TFPP) and the Total False Negative
Proportion (TFNP) for the estimated overlapping groups.
13
Algorithm 2 The LOVE procedure for overlapping clustering.
b from I.I.D. data (X (1) , ..., X (n) ), the tuning parameters δ, λ and µ.
Require: Σ
b the estimated set of pure variables Ib and its
1: Apply Algorithm 1 to obtain the number of clusters K,
b
partition of I.
bb from (6).
2: Estimate AI by A
I
−1
b from (14) and β̄ j for each j ∈ J.
b
3: Estimate C
by Ω
b
b
b
b
4: Estimate AJ by AJb from (16). Combine AIb with AJb to obtain A.
b 1 , ..., G
b } and its direction subgroups Gbs = {G
b s , ..., G
b s } from (18)
5: Estimate overlapping groups Gb = {G
1
K̂
b
- (22) by using A.
b Gb and Gbs .
6: Output A,
K̂
We make the blanket assumption for the remainder of this paper that X is sub-Gaussian, that
is, the Orlicz norm kXj kψ2 of each Xj is bounded by a common constant σ.1 The sub-Gaussian
condition implies maxj∈[p] Σjj ≤ 2σ and kCk∞ ≤ 2σ. Let
E = E(δ) :=
max
1≤i<j≤p
b
Σij − Σij ≤ δ .
(23)
p
We assume throughout that δ = c0 log(p ∨ n)/n, for some constant c0 , and log p = o(n), so that
δ = o(1), for n large enough, where a ∨ b = max(a, b). Taking c0 = c0 (σ) > 0 large enough, Lemma
2 in Bien et al. [2016] guarantees that E holds with high probability:
P(E) ≥ 1 − c1 (p ∨ n)−c2
(24)
for some positive, finite constants c1 and c2 . Apart from δ, the quantity
∆(C) := ν > 0,
(25)
plays an important role in our analysis. Indeed, assumption (iii) requires that ν > 0 in order to
guarantee that the latent factors are distinguishable from one another. We can view ν as a measure
of their separation, and naturally therefore, the size of ν impacts the quality of all our estimators,
in addition to the magnitude of δ.
4.1
b Ib and Ib
Statistical guarantees for K,
We first analyze the performance of our estimator Ib of I, and its corresponding partition. This
problem belongs to the general class of pattern recovery problems, and it is well understood that
under strong enough signal conditions one can expect Ib = I, with high probability. This turns out
to be indeed the case for our problem, but we obtain this as a corollary of a more general result. We
set out to quantify when our estimated set contains the least taxing type of errors, under minimal
assumptions. To make this precise, we introduce the concept of quasi-pure variables. A quasi-pure
variable Xi has very strong association with only one latent factor, say Za , in that |Aia | ≈ 1, and
1
The Orlicz norm of Xj is defined as kXj kψ2 = inf {c > 0 : E [ψ2 (|Xj |/c)] < 1} , based on the Young function
ψ2 (x) = exp(x2 ) − 1.
14
very low association with the rest: |Aib | ≈ 0, for all b 6= a. Formally, we define the set of quasi-pure
variables as:
J1 := {j ∈ J : there exists a ∈ [K], such that |Aja | ≥ 1 − 4δ/ν}.
(26)
For each a ∈ [K] we further define the set of quasi-pure variables associated with the same factor:
J1a := {j ∈ J1 : |Aja | ≥ 1 − 4δ/ν}.
(27)
When ν is a strictly positive constant, := 4δ/ν = o(1). The lower bound |Aja | ≥ 1 − in (27)
implies, under condition (ii), that |Ajb | ≤ , for any b 6= a and j ∈ J1a , justifying the name quasipure variables for those components of X with indices in J1 . We observe, for future reference, that
{J11 , . . . , J1K } forms a partition of J1 .
We show in Theorem 3 that, with very high probability, the estimated Ib contains the pure
variable set I, and is in turn contained in a set that includes all pure variables and quasi-pure
variables. Importantly, Ib will not include indices of variables Xj that are associated with multiple
latent factors at a level higher than . Equally importantly, if a quasi-pure variable Xi is included
b then this variable will have the corresponding |Aia | ≈ 1, and it will be placed together with
in I,
the pure variables associated with the same factor Za , for some a, and not in a new cluster. This is
crucial for ensuring that the number of clusters K is consistently estimated, and also for establishing
the cluster misclassification proportion in Section 4.3 below.
Theorem 3. Assume Model (1) with (i) - (iii), and
p
ν > 2 max 2δ, 2kCk∞ δ .
(28)
Then:
b = K;
(a) K
(b) I ⊆ Ib ⊆ I ∪ J1 .
Moreover, there exists a label permutation π of the set {1, . . . , K}, such that the output Ib =
Iba a∈[K] from Algorithm 1 satisfies:
π(a)
(c) Iπ(a) ⊆ Iba ⊆ Iπ(a) ∪ J1 .
All results hold with probability larger than 1 − c1 (n ∨ p)−c2 , for c1 , c2 positive constants defined in
(24).
The conclusion of Theorem 3 holds only under condition (28), which stipulates that the separation
between the latent factors, as measured by ν, is not only strictly positive, which was needed for
identifiability, but slightly above a quantity that depends on the estimation error δ, and which
becomes o(1) for n large enough.
15
Remark 2. Let e1 = (1, 0, . . . , 0)T and HK be the hyperoctahedral group of signed permutation
matrices. If AI and AJ are well separated in the sense that
min
j∈J, P ∈HK
Aj· − P e1
1
> 4δ/ν,
then J1 = ∅, and Theorem 3 yields exact recovery of the pure variable set and of its partition:
Ib = I and Ib = I, with high probability. However, we expect J1 6= ∅, as we expect quasi-pure
variables to be present in a high dimensional model, which is the context for which Theorem 3 has
been established.
4.2
b
Statistical guarantee for A
b obtained in
In this section we state, and comment on, the statistical properties of the estimate A
p
Sections 3.2 and 3.3. Recall that δ = O( log(p ∨ n)/n) was given in (23) above, and the estimation
of AJ made use of two tuning parameters: λ, in (15), and µ, in (17). Theorem 4 establishes the
properties of our estimates relative to the theoretically optimal values of these tuning parameters,
both of which are functions of δ, while their data adaptive calibration is discussed in Section 5.1
below. We let λ = 2δ 0 and µ = 5kC −1 k∞,1 δ 0 , with
8
δ0 =
kCk∞ − 3 δ,
(29)
ν
for ν defined in (25) above. When ν and kCk∞ are strictly positive constants we thus have
p
p
λ = O( log(p ∨ n)/n) and µ = O(kC −1 k∞,1 log(p ∨ n)/n). We consider the loss function for two
p × K matrices A, A0 as
Lq (A, A0 ) := min kAP − A0 k∞,q ,
1 ≤ q ≤ ∞.
P ∈HK
(30)
Here HK is the hyperoctahedral group of all K × K signed permutation matrices and
kAk∞,q := max kAi· kq = max
1≤i≤p
1≤i≤p
K
X
j=1
for a generic matrix A ∈ Rp×K .
1/q
|Aij |q
,
Theorem 4. Assume the conditions in Theorem 3 hold. Let λ and µ be as defined above, and set
s = maxi∈[p] kAi· k0 . Then,
b A) ≤ 10s1/q kC −1 k∞,1 δ 0 ,
Lq (A,
1 ≤ q ≤ ∞,
with probability larger than 1 − c1 (n ∨ p)−c2 , for c1 , c2 positive constants defined in (24), provided
that (2µ + 4δ/ν) < 1. We use the convention that s1/q = 1 for q = +∞.
Remark 3.
16
1. In fact, we prove the stronger result
bi· − (AP )i·
min A
P ∈HK
q
≤ 10(si )1/q kC −1 k∞,1 δ 0 ,
1 ≤ q ≤ ∞,
with sparsity index si = kAi· k0 for each row Ai· , i ∈ [p] of A. The signed permutation matrix
P that achieves the minimum is determined by the alignment of the pure variables and is the
same for each i ∈ [p].
bi· k1 ≤ 1, for each i ∈ [p], with high
2. Inspection of the proof of this result quickly reveals that kA
probability, in accordance with our model requirement (i).
3. Our primary focus is the bound for q = +∞, as this leads to inference on support recovery
of A. More generally, for any q ≥ 1, it is well understood that the quality of estimating a
sparse vector in high-dimensional regression-type models depends on the interplay between its
sparsity and the behavior of the appropriate Gram matrix associated with the model, which
reduces to C = E[ZZ T ] in our case. The concept of `q -sensitivity, introduced by Gautier and
Tsybakov [2011]; Belloni et al. [2017], is the most general characterization of this interplay
to date. It facilitates a link between the `q -norm of sparse vectors β and the `∞ -norm of the
product between the Gram matrix and β, uniformly over vectors β of sparsity s, ranging over
a collection of cones. In our context, that of a square, invertible matrix C, the reciprocal of
the `∞ -sensitivity of C becomes essentially kC −1 k∞,1 , which indeed links kβk∞ to kCβk∞ .
Similarly, the quantities s1/q kC −1 k∞,1 provide concrete substitutes of the reciprocals of the
`q -sensitivities of C, and all of our rates in Theorem 4 match the lower bounds in Theorem
5, up to a logarithmic factor in p/s and the quantity kC −1 k∞,1 .
The size of kC −1 k∞,1 ranges from the constant kC −1 k∞ , when all latent factors are independent, to the fully general case of kC −1 k∞,1 = O(K). In the latter case the bounds become
p
meaningful when K < O( n/ log p). However, if C −1 is sparse, then kC −1 k∞,1 may be considerably smaller than K. In particular, if Z has a multivariate normal distribution and many
factors Zi are conditionally independent, then kC −1 k∞,1 is small. We do not make any of
these assumptions here, and regardless of the situation, Theorem 4 shows that our estimation
procedure adapts automatically to it.
We now show that the rates of convergence in Theorem 4 are optimal (up to a logarithmic
factor in p/s and provided kC −1 k∞,1 remains bounded) in a minimax sense for all estimators over
the parameter space
As := A ∈ [−1, 1]p×K : A satisfies conditions (i) and (ii) and max kAi· k0 ≤ s .
1≤i≤p
For our purpose of establishing a minimax lower bound, it suffices to consider a particular subGaussian distribution of (Z, E). We choose to take the multivariate Gaussian with E[(Z, E)] = 0 ∈
RK+p , Cov(Z, E) = 0 ∈ RK×p and Cov(Z) = IK , the K × K identity matrix, and Cov(E) = Γ =
diag(σ12 , . . . , σp2 ).
17
Theorem 5. Assume Z ∼ N (0, IK ) and E ∼ N (0, Γ) are independent. Let K ≥ 2, p ≥ 2K + 1,
1 ≤ s ≤ 4K/5 and
r
log(K/s)
s
≤ C1 ,
(31)
n
for some universal constant C1 > 0. Then, for all 1 ≤ q ≤ ∞,
)
(
r
log(K/s)
1/q
b A) ≥ C2 s
≥ C3 ,
(32)
inf sup PA Lq (A,
n
b A∈A
A
2
for some positive constants C2 , C3 which only depend on σmin
:= mini∈[p] Γii . The infimum is taken
1/q
b
over all estimators A of A and we use the convention s = 1 for q = +∞.
Since for C = I, the quantity kC −1 k∞,1 becomes 1, we see that the rates of Theorems 4 and
5 coincide up to the logarithmic factors log(K/s) and log(p ∨ n). This shows that our procedure
is not only adaptive, but nearly minimax optimal up to the logarithmic factor provided kC −1 k∞,1
remains bounded.
4.3
Statistical guarantee for Gb and Gbs .
For easy of presentation, and without loss of generality, throughout this section, we continue to
write A for its orthonormal transformation AP that uses the optimal signed permutation matrix
b
P ∈ HK from Theorem 4 to align the columns and signs of A with that of A.
b = K. The latter
We define two criteria to evaluate the estimated clusters Gb on the event K
holds with high probability by Theorem 3. We first define the individual Group False Positive
Proportion (GFPP) and the individual Group False Negative Proportion (GFNP) as,
b a ) :=
GFPP(G
ba |
|(Ga )c ∩ G
,
c
|(Ga ) |
b a ) :=
GFNP(G
b a )c |
|Ga ∩ (G
,
|Ga |
for each a ∈ [K],
(33)
b a )c := [p] \ G
b a , with the convention GFPP(G
b a ) = 0 if |(Ga )c | = ∅.
where (Ga )c := [p] \ Ga and (G
b a . Furthermore,
GFPP and GFNP quantify the misclassification proportion within each group G
with the same convention, we can define the Total False Positive Proportion (TFPP) and Total
b
False Negative Proportion (TFNP) to quantify the overal misclassification proportion of G.
b :=
TFPP(G)
PK
c
b
a=1 |(Ga ) ∩ Ga |
,
PK
c
a=1 |(Ga ) |
b :=
TFNP(G)
Finally, given µ = 5kΩk∞,1 δ 0 with δ 0 specified in (29), we define
PK
b c
a=1 |Ga ∩ (Ga ) |
.
PK
a=1 |Ga |
J2 := {i ∈ J : for any a with Aia 6= 0, |Aia | > (2µ) ∨ (4δ/ν)}.
(34)
(35)
and J3 := J \ (J1 ∪ J2 ). J2 can be viewed as the set where every non-zero entry of Aj· is separated
away from 0 for each j ∈ J2 . The following theorem shows that J2 plays a critical role in quantifying
b and the misclassification proportion of G.
b Let Sb := supp(A).
b
both the support recovery of A
18
Theorem 6. Under the conditions of Theorem 4, with probability greater than 1 − c1 (n ∨ p)−c2 for
some positve constant c1 and c2 defined in (24), we have:
(a)
b ⊆ supp (A) ,
supp AJ2 ⊆ supp(A)
b b) = sign A b .
sign(A
S
S
(b) Let saj = 1{|Aja | =
6 0} and taj = 1{|Aja | ≤ (2µ) ∨ (4δ/ν)}, for each j ∈ J and a ∈ [K].
P
a
a t
b a ) = 0; GFNP(G
b a ) ≤ Pj∈J1 ∪J3 \J1 j .
GFPP(G
a
j∈J sj + |Ia |
(c) Let sj =
Remark 4.
PK
a=1 1{|Aja |
=
6 0} and tj =
b = 0;
TFPP(G)
PK
≤ (2µ) ∨ (4δ/ν)}, for each j ∈ J.
P
j∈J1 ∪J3 tj
b
TFNP(G) ≤ P
.
j∈J sj + |I|
(36)
a=1 1{|Aja |
(37)
1. From our proof of Theorem 6, it is easy to verify that the expression of TFNP in (37) continues to hold for the Direction False Positive Proportion (DFPP) and the Direction False
P
Negative Proportion (DFNP) defined in (40) below with sj replaced by K
a=1 1{Aja < 0} or
P
PK
PK
K
a=1 1{0 < Aja ≤
a=1 1{−(2µ) ∨ (4δ/ν) ≤ Aja < 0} or
a=1 1{Aja > 0}, tj replaced by
+
−
±
(2µ) ∨ (4δ/ν)} and I replaced by I or I , where I := ∪a∈[K] {i ∈ Ia : Aia = ±1}.
b a ) will be small if either ta is small
2. According to display (36), it is easy to see that GFNP(G
j
P
for j ∈ J1 ∪ J3 or |J1 | + |J3 | − |J1a | is dominated by |Ia | + j∈J saj . Moreover, from display
(37), TFNP will be small in the following two cases:
- |J1 | + |J3 | is dominated by |I| + |J2 |;
- tj is small relative to sj , for j ∈ J1 ∪ J3 .
To illustrate this, consider tj ≡ t and sj ≡ s, for each j ∈ J, to simplify the expressions a
bit, and assume |J1 | + |J3 | = α(|I| + |J2 |), for some α ≥ 0. We show in the supplementary
material that
1
(s − 1)|J2 |
b ≤t
TFNP(G)
s+
1+
,
α
|I| + |J2 |
Thus, when either t or α is small, that is, when |J1 | + |J3 | is dominated by |I| + |J2 |, then
TFNP will be small. Note that even when t itself is large but bounded by some constant, TFNP
p
might also be small since s can be close to K which is allowed to grow as O( n/ log p).
3. If J2 = J with µ = 3kC −1 k∞,1 δ, from noting that J2 ⊆ J \ J1 , Remark 2 in Section 4.1 yields
Ib = I. We can choose λ = δ in (15) and µ = 3kC −1 k∞,1 δ in (17), and follow the proof of
Theorems 4 and 6 to arrive at the following conclusions:
b = supp (A) ,
supp(A)
Moreover, we get exact cluster recovery:
19
b = sign(A).
sign(A)
b a ) = GFNP(G
b a ) = 0, for each a ∈ [K].
(a) GFPP(G
b = TFNP(G)
b = 0.
(b) TFPP(G)
b 0 = G0 . Again, all statements hold with probability greater than
This immediately yields G
1 − c1 (n ∨ p)−c2 .
e in which we combine
4. We prove that Theorem 6 also holds for the hard threshold estimator A
j
j
j
j
bb with A
e b. Each row of A
e b is estimated by βea = β̄a 1{|β̄a | > µ} of βa = Aja , a ∈ [K],
b using
A
I
J
J
−1
0
the same µ = 5kC k∞,1 δ as before for the threshold µ. However, we cannot guarantee that
the scaling restriction of condition (i) holds for this estimator.
4.4
Discussion and related work
To the best of our knowledge, optimal estimation of identifiable sparse loading matrices A in model
(1) satisfying (i) - (iii), when both I and K are unknown, and when the entries in X, Z and A are
allowed to have arbitrary signs, has not be considered elsewhere and our results bridge this gap.
There exists, however, a very large body of literature on related problems. We review the most
closely related results below, and explain the differences with our results.
Results regarding the identifiability of A in general latent models, typically without sparsity,
are scattered throughout over more than six decades of literature. They all involve conditions on
both A and C, and there is typically a trade-off between the restrictions on A versus those on C,
as first summarized and proved in Anderson and Rubin [1956], reviewed in Lawley and Maxwell
[1971] and later in Anderson and Amemiya [1988]. We recall them briefly here for the convenience
of the reader.
By far the most commonly used assumption is that the latent factors are uncorrelated, so that
C is either the identity or a diagonal matrix. In this case, it is typically further assumed that the
scaled columns of A are orthogonal, see, for instance, the literature review in Izenman [2008]. An
alternative requirement is that A contain a K × K lower diagonal matrix, see, e.g., Geweke and
Zhou [1996] and, moreover, that the placement of this matrix within A is known, which requires
careful justification [Carvalho et al., 2008], and may be problematic from a practical perspective
[Bhattacharya and Dunson, 2011].
In general, latent factors are correlated, which is our point of view in this work. Then, starting
with Anderson and Rubin [1956], one places on the structure of A constraints that are different than
those made when C is diagonal. The most common of those assumptions involves the existence of
a pure variable set I, similar to our assumption (ii). If I is known, classical results in Anderson
and Rubin [1956] and the proof of our Theorem 2 show that C can be an arbitrary positive definite
matrix. When I is unknown, conditions on the latent factors also need to be imposed. Sufficient
conditions on Z, with provable guarantees for the identification of I, are only known, to the best
of our knowledge, in the NMF literature: the uniqueness of I follows from the uniqueness of the
solution of an appropriate linear program, applied to population quantities, and tailored to matrices
with non-negative entries, see Bittorf et al. [2012]. In contrast, the arguments of Section 2 above
are optimization-free and can be used for matrices that have entries of arbitrary sign. Therefore,
20
we provide a new addition to the literature on pure-variable and loading matrix identification, in
general latent models, and also in the particular case of NMF.
A related, but different, identifiability question regards the covariance matrix Σ of X which,
under (1), can be written as the sum between a rank K matrix and a diagonal matrix:
Σ = ACAT + Γ,
(38)
and Γ := Cov(E) is a diagonal matrix with possibly different entries. In these models, the identifiability question is whether Σ can be decomposed uniquely as the sum between ACAT and Γ.
Answers to this question generated a large amount of literature. We refer the reader to Ledermann
[1937]; Anderson and Rubin [1956]; Shapiro [1982, 1985]; Bekker and ten Berge [1997] for earlier
results, and to Bai and Ng [2002]; Chandrasekaran et al. [2011, 2012]; Candès et al. [2011]; Hsu
et al. [2011], Fan et al. [2013], Wegkamp and Zhao [2016] for more recent works, that also address
the problems of rank estimation and optimal estimation of high dimensional covariance matrices. It
is noteworthy that these works, relative to one another, give different types of sufficient conditions
under which one can separate the low rank matrix ACAT from Γ. However, since we always have
ACAT = (AQ)(QT CQ)(QT AT ), for any orthonormal Q, they do not guarantee the identifiability
of A itself. Conversely, we show in Theorem 2 in Section 2 that under conditions (i) - (iii), C
and Γ are identified, and A is identified up to signed permutations. Therefore, we also identify
uniquely the decomposition of Σ. Our conditions are not always comparable to those employed for
the unique decomposition of Σ, but in special cases they imply them. Although the uniqueness of
the decomposition of Σ is a by-product of our results, we do not pursue the covariance estimation
problem in this work, but we included the above discussion for completeness.
Furthermore, we do not view the problem of estimating the number of factors K as that of
estimating the rank of a matrix. This approach is taken in Bai and Ng [2002], via penalized
least squares, but provided that either C = I or AAT = I and that K is bounded by a fixed
integer. Alternatively, we could adapt the criteria in Bunea et al. [2011]; Bing and Wegkamp
[2018]; Wegkamp and Zhao [2016] to (1) to allow for K → ∞ in the rank estimation problem.
However, proving that such an estimator is consistent would ultimately require an unnecessary
lower bound restriction on the K-th largest eigenvalue of ACAT . In contrast, our Theorem 3 shows
that such conditions can indeed be avoided. We estimate directly the set I and its partition via
LOVE, and as a byproduct K, at a low computational cost of order p2 .
Estimation of A in identifiable factor models is typically based on iterative alternating least
squares procedures or the EM algorithm, see for instance Rubin and Thayer [1982]; Bai and Li
[2012] and the references therein. As discussed in these works, the resulting algorithms are not
suitable for large data sets due to their notoriously slow convergence to a solution that is typically
not the global optimum. Bayesian estimation, see, e.g. Carvalho et al. [2008] and the references
therein, offers an alternative approach which may become computationally very demanding in high
dimensions, requires a likelihood framework, and careful prior specification. Moreover, existing
procedures do not estimate A under our model specifications (i) - (iii), and any adaptation would
still require the challenging estimation of I. Our procedure offers a solution to the computational
problem, as LOVE does not require a likelihood or other prior distributional specifications, is
tailored to our model with unknown I, and has provable low computational complexity.
21
The statistical properties of estimators of A in model (1) (i) - (iii) have not been studied,
and even particular cases of the model have received a very limited amount of attention, from
a theoretical perspective. When I is known and K is fixed, Bai and Li [2012] established the
asymptotic normality of the MLE in a model similar to ours, although the estimator they ultimately
construct is not necessarily the MLE under this model, but rather an appropriate transformation
of the stationary point of a quasi-likelihood for a different factor model. We give the specific details
of their construction in Supplement C.1. If I is unknown, but K is known, and moreover, the
columns of X, A and Z have non-negative entries that sum up to 1, Arora et al. [2013] provide a
practical algorithm for the estimation of A and offer bounds on the `1 matrix norm loss of their
estimator. The extra restrictions on this model are motivated by a specific model, the topic model,
appropriate for vectors with discrete distributions, for instance multinomial. The construction and
analysis of these estimates are not transferable to our general framework, as they depend heavily
on these restrictions. Our results of Section 4.2 bridge this gap in the literature and offer lower and
upper bounds for the performance of estimators of A in model (1) (i) - (iii).
Finally, to the best of our knowledge, overlapping clustering based on model (1) has not been
analyzed. A particular case of this model, corresponding to a matrix A with binary entries, has
been considered in Bunea et al. [2016a,b] for non-overlapping clustering. According to their model,
all p variables are pure variables, as the model assume that Xj = Zk + Ej , for all j ∈ Gk and
k ∈ {1, . . . , K}, {Gk }1≤k≤K form a partition of {1, . . . , p}. When C is positive definite, the nonoverlapping clusters are shown to be identifiable, and the work of Bunea et al. [2016a,b] is devoted
to exact recovery of clusters with minimax optimal cluster separation, a very different problem
than the one considered here.
5
Simulation Studies
In this section, we first discuss our procedure for selecting the tuning parameters, then evaluate the
performance of LOVE based on estimation error and overall clustering misclassification proportion.
In the supplementary materials, we compare LOVE with existing overlapping clustering algorithms
and study the performance of LOVE for the non-overlapping clustering problem.
5.1
Data driven choice of the tuning parameters
Tuning parameter δ. Proposition 3 specifies the theoretical rate of δ, but only up to constants
that depend on the underlying data generating mechanism. We propose below a data-dependent
way to select δ, based on data splitting. Specifically, we split the data set into two independent
b (1) . On the
parts, of equal sizes. On the first set, we calculate the sample covariance matrix Σ
p
second set, we choose a fine grid of values δ` = c` log p/n, with 1 ≤ ` ≤ M , for δ, by varying the
b
proportionality constants c` . For each δ` , we obtain the estimated number of clusters K(`)
and
b
b
b
b
the pure variable set I(`) with its partition I(`). Then we construct the |I(`)| × K(`) submatrix
bb of A,
b and estimate C(`)
b
b
b
A
via formula (11). Finally, we calculate the |I(`)|
× |I(`)|
matrix
I(`)
bb C(`)
b A
bT . In the end, we have constructed a family F = {W1 , . . . , WM } of the fitted
W` = A
I(`)
b
I(`)
22
b
matrices W` , each corresponding to different I(`)
that depend in turn on δ` , for ` ∈ {1, . . . , M }.
Define
1
b (1)
b
,
(39)
CV (I(`))
:= q
b I(`)
b − W`
ΣI(`)
F-off
b
b
|I(`)|
|I(`)|
−1
where kBkF-off := kB − diag(B)kF denotes the Frobenius norm over the off-diagonal elements
b
of a square matrix B. We choose δ cv as the value δ` that minimizes CV (I(`))
over the grid
` ∈ [M ]. To illustrate how the selection procedure works, we provide an example in Section B of
the supplementary material.
Tuning parameters λ and µ. The tuning parameter λ in the linear program (15) for estimating
Ω = C −1 is specified by λ = 2δ 0 with δ 0 defined in (29). Since δ 0 is proportional to δ, we use λ = c0 δ cv
where c0 is some constant and could be tuned by a cross-validation strategy used in the related
work on the precision matrix estimation, for instance Cai et al. [2011]. More precisely, we randomly
b on the first dataset for each
split the data into two parts. For a given grid of λ, we compute Ω
value in the grid. Then we choose the one which gives the smallest likelihood loss from the second
dataset, where the likelihood loss is defined by
L(Ω, C) = hΩ, Ci − log det(Ω).
From Remark 4 (3) in Section 4.3, when J2 = J, we can choose λ = δ which is the smallest λ
we should consider. Therefore, we set the grid of λ equal to [δ cv , 3δ cv ]. From our simulation, the
selected λ is δ cv in most cases. Hence we recommend to use λ = δ cv and our simulations are based
on this choice.
b estimates C −1 . Our extensive
Recall that µ = c1 kC −1 k∞,1 δ for some constant c1 , and that Ω
b ∞,1 δ cv yields stable performance, with Ω
b solved from
simulations show that the choice of µ = kΩk
(14) and δ cv selected via cross-validation.
5.2
Estimation error and cluster recovery with LOVE
In this section, we study the numerical performance of LOVE in terms of clustering and estimation
accuracy. To the best of our knowledge, there is no comparable algorithm with provable guarantees
developed for our framework, especially if the set I is unknown, as explained in detail in Section
4.4 above, and further re-visited in Section C.1 of the supplementary material.
We generate the data in the following way. We set the number of clusters K to be 20 and
simulate the latent variables Z = (Z1 , . . . , ZK ) from N (0, C). The diagonal elements of C is given
by Cii = 2 + (i − 1)/19 for i = 1, . . . , 20, and the off-diagonal elements are generated as Cij =
(−1)(i+j) 0.3|i−j| (Cii ∧ Cjj ) for any i 6= j. In addition, the error terms E1 , ..., Ep are independently
sampled from N (0, σp2 ), where σp2 itself is sampled from a uniform distribution on [1, 3]. Since the
rows of A corresponding to pure variables in the same cluster are allowed to have different signs, we
consider the following configuration of signs for pure variables in each cluster: (3, 2), (4, 1), (2, 3),
(1, 4) and (5, 0), with the convention that the first number denotes the number of positive pure
variables in that group and the second one denotes the number of negative pure variables. Among
23
the 20 groups, each sign pattern is repeated 4 times. To generate AJ , for any j ∈ J, we randomly
assign the cardinality sj of the support of Aj· to a number in {2, 3, 4, 5}, with equal probability.
Then, we randomly select the support from {1, 2, . . . , K} with cardinality equal to sj . For Ajk
which is nonzero, we set it as Ajk = sign ·(1/sj ) with sign randomly sampled from {−1, 1}. Thus,
we can generate X according to the model X = AZ + E. In the simulation studies, we vary p from
200 to 1000 and n from 300 to 1000. Each simulation is repeated 50 times.
b are not directly comparable, since
Recall that the true allocation matrix A and our estimator A
they may differ by a permutation matrix. To evaluate the performance of our method, we consider
b have the same dimension, we first
the following mapping approach [Wiwie et al., 2015]. If A and A
b kF is minimized.
find the mapping (i.e., the signed permutation matrix P ∈ HK ) such that kA − AP
q = AP
b with A to evaluate the estimation and
Thus, we can compare the permuted estimator A
recovery error. Under this mapping approach, we can evaluate TFPP and TFNP defined in (34).
Moreover, in order to account for the direction sub-groups defined in (21), we can define Direction
False Positive Proportion (DFPP) and Direction False Negative Proportion (DFNP) as follows:
DFPP =
PK
1
b2
a=1 |Ga ∩ Ga |
,
PK
1
a=1 |Ga |
DFNP =
PK
2
b1
a=1 |Ga ∩ Ga |
.
PK
2
a=1 |Ga |
(40)
Figure 1 shows the percentage of exact recovery of number of clusters K, TFPP, TFNP, DFPP
b = K, we can
and DFNP of LOVE. Since the last four measures are well defined only if rank(A)
compute them when the number of clusters is correctly identified. We can see that the proposed
method correctly selects K and as long as the number of clusters is correctly selected, TFPP, TFNP,
DFPP and DFNP of our method are very close to 0, which implies that the sign and sparsity pattern
b as measured by the matrix
of A can be correctly recovered. We present the estimation error of A
√
`1 norm scaled by pK and the Frobenius norm scaled by pK in Table 1.
As expected, the estimation error decreases when the sample size increases from 300 to 1000,
which is in line with our theoretical results. The simulations are conducted on an macOS Sierra
system version 10.12.6 with 2.2 GHz Intel Core i7 CPU and 16 GB memory. Even with p = 1000
and n = 1000, the computing time of our method for each simulation is around 1 minute.
Moreover, we evaluated the performance of the LOVE procedure for K varying in a wide range,
from 3 to 30, and when AJ contains many very small entries. The results are consistent with what
we observed in this section and deliver the same message. The GFPP and GFNP are similar as
e defined in Remark 4
TFPP and TFNP and the performance of the hard thresholding estimator A,
b To save space, we have omitted those results.
of Section 4.3, is similar to A.
We also compared the performance of LOVE with other off-the-shelf algorithms for overlapping
clustering, and tested LOVE for non-overlapping clustering. We included these results in sections
C.2 and C.3 of the supplementary material.
6
Application
To benchmark LOVE, we used a publicly available RNA-seq dataset of 285 blood platelet samples
from patients with different malignant tumors [Best et al., 2015]. We extracted a small subset of 500
24
b as measured by the matrix `1 norm (`1 ) (divided by
Table 1: The average estimation error of A
√
pK) and the Frobenius norm (`2 ) (divided by pK). Numbers in parentheses are the simulation
standard errors.
p
200
400
600
800
1000
n = 300
n = 500
n = 700
`1
`2
`1
`2
`1
`2
`1
`2
0.018
(0.001)
0.026
(0.002)
0.029
(0.002)
0.031
(0.002)
0.032
(0.002)
0.062
(0.005)
0.075
(0.007)
0.079
(0.006)
0.083
(0.006)
0.083
(0.006)
0.015
(0.001)
0.023
(0.001)
0.025
(0.001)
0.026
(0.001)
0.027
(0.001)
0.053
(0.003)
0.064
(0.003)
0.067
(0.003)
0.068
(0.004)
0.069
(0.003)
0.013
(0.001)
0.021
(0.001)
0.023
(0.001)
0.024
(0.001)
0.025
(0.001)
0.048
(0.008)
0.059
(0.006)
0.063
(0.003)
0.064
(0.004)
0.065
(0.004)
0.012
(0.001)
0.018
(0.001)
0.020
(0.001)
0.022
(0.001)
0.022
(0.001)
0.041
(0.002)
0.051
(0.003)
0.055
(0.003)
0.057
(0.004)
0.057
(0.004)
0.8
0.6
0.4
0.2
0.0
0.0
0.2
cluster
TFPP
TFNP
DFPP
DFNP
recovery and error rates
0.4
0.6
0.8
1.0
p = 1000
1.0
p = 400
recovery and error rates
n = 1000
300 400 500 600 700 800 900
300 400 500 600 700 800 900
sample size
sample size
Figure 1: Percentage of exact recovery of number of clusters K (cluster), total false positive proportion (TFPP), total false negative proportion (TFNP), direction false positive proportion (DFPP)
and direction false negative proportion (DFNP) for LOVE.
Ensembl genes to test the method. The goal of the benchmarking was to test whether (i) clusters
corresponded to biological knowledge, specifically Gene Ontology (GO) functional annotation of the
genes [Ashburner et al., 2000], (ii) overlapping clusters corresponded to pleiotropic gene function.
LOVE produced twelve overlapping clusters (Table 2) which aligned well with a-priori expectation.
25
Table 2 lists the number of pure genes and the total number of genes in twelve overlapping clusters.
Figure 2 shows that each cluster overlaps with the other and also gives us a clear picture on how two
clusters possibly overlap. For example, 18 genes belong to both cluster 3 and cluster 11, whereas
cluster 2 and cluster 3 have only one common gene. The genes with the same GO biological
process, molecular function or cellular component terms tended to be assigned to the same cluster.
For example, ENSG00000273906 and ENSG00000273328 are both RNA genes. They were both
assigned to the same cluster (cluster 6, Figure 2). However, they were also assigned to other
clusters, suggesting they have pleiotropic functions. This suggests that the latent variables used
for clustering are likely to have biological significance and can potentially be used for functional
discovery for genes with under-explored functions. We found 308 genes with zero expression across
all samples. None of them were assigned to any of the 12 estimated clusters, as desired. Indeed,
our model not only allows for the existence of pure noise variables Xj = Ej , but variables with
structural zero values as well, as Γjj = Var(Ej ) = 0 is permitted. Formally we place them in the
pure noise cluster G0 , for further scientific scrutiny.
Table 2: Number of pure genes and total number of genes in each group.
G1
G2
G3
G4
G5
G6
G7
G8
G9
G10
G11
G12
2
58
2
35
2
67
4
105
2
80
10
104
2
28
2
43
2
44
4
74
2
94
15
108
Number of pure genes
Total number of genes
14
8
G67
G5
G3
10
11
6
9
7
6
15
Group size
13
28
18
35
10
5
43
1
6
G4
9
8
44
7
7
3
8
4
8
4
G2
11 G8
2
14 7
4
4
11
3
102 4
8
7
14
9
G7
4
1
80
94
3 10
3
1
74
6
14
7
2
67
9
5
4
58
G11
16
6
7
3
104
105
1
3
3
3
9
G9
G10 6
108
7
4
G1
5
G12
Figure 2: Left panel: Number of genes overlapped in different groups. The nodes represent 12
groups with the same labels and sizes as those in Table 2. The number shown on the edge between
two nodes represents the number of genes shared by the two groups, which corresponds to the width
of that edge. Right panel: Illustration of three genes ENSG00000273906, ENSG00000273328 and
ENSG00000273113 and their allocation matrix relative to 12 groups. For instance, the jth gene
bj6 = 0.04, A
bj9 = 0.37, A
bj11 = −0.02.
ENSG00000273906 belongs to groups 6, 9 and 11 with A
26
Acknowledgements
We are grateful to Jishnu Das for help with the interpretation of our data analysis results. Bunea
and Wegkamp are supported in part by NSF grant DMS 1712709. Bing was supported in part by
NSF grant DMS-1407600.
Supplementary Material
The supplementary materials contain the proofs and additional numerical results.
References
Anderson, T. (2003). An Introduction to Multivariate Statistical Analysis. Wiley Series in Probability and Statistics, Wiley.
Anderson, T. W. and Amemiya, Y. (1988). The asymptotic normal distribution of estimators
in factor analysis under general conditions. Ann. Statist. 16 759–771.
Anderson, T. W. and Rubin, H. (1956). Statistical inference in factor analysis. In Proceedings
of the Third Berkeley Symposium on Mathematical Statistics and Probability, Volume 5: Contributions to Econometrics, Industrial Research, and Psychometry. University of California Press,
Berkeley, Calif.
Arora, S., Ge, R., Halpern, Y., Mimno, D. M., Moitra, A., Sontag, D., Wu, Y. and Zhu,
M. (2013). A practical algorithm for topic modeling with provable guarantees. In ICML (2).
Ashburner, M., Ball, C. A., Blake, J. A., Botstein, D., Butler, H., Cherry, J. M.,
Davis, A. P., Dolinski, K., Dwight, S. S. and Eppig, J. T. (2000). Gene ontology: tool for
the unification of biology. Nature genetics 25 25–29.
Bai, J. and Li, K. (2012). Statistical analysis of factor models of high dimension. Ann. Statist.
40 436–465.
Bai, J. and Ng, S. (2002). Determining the number of factors in approximate factor models.
Econometrica 70 191–221.
Bekker, P. A. and ten Berge, J. M. (1997). Generic global identification in factor analysis.
Linear Algebra and its Applications 264 255 – 263. Sixth Special Issue on Linear Algebra and
Statistics.
Belloni, A., Rosenbaum, M. and Tsybakov, A. B. (2017). Linear and conic programming
estimators in high dimensional errors-in-variables models. Journal of the Royal Statistical Society:
Series B (Statistical Methodology) 79 939–956.
Best, M. G., Sol, N., Kooi, I., Tannous, J., Westerman, B. A., Rustenburg, F.,
Schellen, P., Verschueren, H., Post, E., Koster, J. et al. (2015). Rna-seq of tumoreducated platelets enables blood-based pan-cancer, multiclass, and molecular pathway cancer
diagnostics. Cancer cell 28 666–676.
Bezdek, J. C. (2013). Pattern recognition with fuzzy objective function algorithms. Springer
Science & Business Media.
27
Bhattacharya, A. and Dunson, D. (2011). Sparse bayesian infinite factor models. Biometrika
98 291–306.
Bien, J., Bunea, F. and Xiao, L. (2016). Convex banding of the covariance matrix. Journal of
the American Statistical Association 111 834–845.
Bing, X. and Wegkamp, M. H. (2018). Adaptive estimation of the rank of the coefficient matrix
in high dimensional multivariate response regression models. ArXiv .
Bittorf, V., Recht, B., Re, C. and Tropp, J. A. (2012). Factoring nonnegative matrices with
linear programs. arXiv:1206.1270 .
Bollen, K. A. (1989). Structural Equations with Latent Variables. Wiley.
Bunea, F., Giraud, C. and Luo, X. (2016a). Minimax optimal variable clustering in g-models
via cord. arXiv preprint arXiv:1508.01939 .
Bunea, F., Giraud, C., Royer, M. and Verzelen, N. (2016b). Pecok: a convex optimization
approach to variable clustering. arXiv preprint arXiv:1606.05100 .
Bunea, F., She, Y. and Wegkamp, M. H. (2011). Optimal selection of reduced rank estimators
of high-dimensional matrices. Ann. Statist. 39 1282–1309.
Cai, T., Liu, W. and Luo, X. (2011). A constrained 1 minimization approach to sparse precision
matrix estimation. Journal of the American Statistical Association 106 594–607.
Cai, T. T., Liu, W. and Zhou, H. H. (2016). Estimating sparse precision matrix: Optimal rates
of convergence and adaptive estimation. Annals of Statistics 44 455–488.
Candès, E. J., Li, X., Ma, Y. and Wright, J. (2011). Robust principal component analysis?
Journal of the ACM (JACM) 58 11.
Carvalho, C. M., Chang, J., Lucas, J. E., Nevins, J. R., Wang, Q. and West, M. (2008).
High-dimensional sparse factor modeling: Applications in gene expression genomics. Journal of
the American Statistical Association 103 14381456.
Chandrasekaran, V., Parrilo, P. A. and Willsky, A. S. (2012). Latent variable graphical
model selection via convex optimization. Ann. Statist. 40 1935–1967.
Chandrasekaran, V., Sanghavi, S., Parrilo, P. A. and Willsky, A. S. (2011). Ranksparsity incoherence for matrix decomposition. SIAM J. Optim. 21 572–596.
Craddock, R. C., James, G. A., Holtzheimer, P. E., Hu, X. P. and Mayberg, H. S.
(2012). A whole brain fmri atlas generated via spatially constrained spectral clustering. Human
brain mapping 33 1914–1928.
Craddock, R. C., Jbabdi, S., Yan, C.-G., Vogelstein, J. T., Castellanos, F. X., Di Martino, A., Kelly, C., Heberlein, K., Colcombe, S. and Milham, M. P. (2013). Imaging
human connectomes at the macroscale. Nature methods 10 524–539.
Donoho, D. and Stodden, V. (2004). When does non-negative matrix factorization give a correct
decomposition into parts? In Advances in Neural Information Processing Systems 16 (S. Thrun,
L. K. Saul and P. B. Schölkopf, eds.). MIT Press, 1141–1148.
Everitt, B. (1984). An Introduction to Latent Variable Models. Monographs on Statistics and
Applied Probability. Springer.
Fan, J., Liao, Y. and Mincheva, M. (2013). Large covariance estimation by thresholding prin-
28
cipal orthogonal complements. Journal of the Royal Statistical Society: Series B (Statistical
Methodology) 75 603–680.
Friedman, J., Hastie, T. and Tibshirani, R. (2008). Sparse inverse covariance estimation with
the graphical lasso. Biostatistics 9 432–441.
Gautier, E. and Tsybakov, A. B. (2011). High-dimensional instrumental variables regression
and confidence sets. arXiv preprint arXiv:1105.2454v4 1105.2454.
Geweke, J. and Zhou, G. (1996). Measuring the pricing error of the arbitrage pricing theory.
The review of financial studies 9 557–587.
Hsu, D., Kakade, S. M. and Zhang, T. (2011). Robust matrix decomposition with sparse
corruptions. IEEE Trans. Inform. Theory 57 7221–7234.
Izenman, A. J. (2008). Modern Multivariate Statistical Techniques: Regression, Classification,
and Manifold Learning. Series: Springer Texts in Statistics.
Jiang, D., Tang, C. and Zhang, A. (2004). Cluster analysis for gene expression data: A survey.
IEEE Transactions on knowledge and data engineering 16 1370–1386.
Koopmans, T. C. and Reiersol, O. (1950). The identification of structural characteristics. Ann.
Math. Statist. 21 165–181.
Krishnapuram, R., Joshi, A., Nasraoui, O. and Yi, L. (2001). Low-complexity fuzzy relational
clustering algorithms for web mining. IEEE transactions on Fuzzy Systems 9 595–607.
Lawley, D. N. and Maxwell, A. E. (1971). Factor analysis as a statistical method. 2nd ed.
American Elsevier Publishing Co., Inc., New York.
Ledermann, W. (1937). On the rank of the reduced correlational matrix in multiple-factor analysis. Psychometrika 2 85–93.
McDonald, R. P. (1999). Test theory: a unified treatment. Taylor and Francis.
Meinshausen, N. and Bühlmann, P. (2006). High-dimensional graphs and variable selection
with the lasso. Annals of Statistics 1436–1462.
Rigollet, P. and Tsybakov, A. (2011). Exponential screening and optimal rates of sparse
estimation. Ann. Statist. 39 731–771.
Rubin, D. B. and Thayer, D. T. (1982). Em algorithms for ml factor analysis. Psychometrika
47 69–76.
Shapiro, A. (1982). Rank-reducibility of a symmetric matrix and sampling theory of minimum
trace factor analysis. Psychometrika 47 187–199.
Shapiro, A. (1985). Identifiability of factor analysis: some results and open problems. Linear
Algebra and Its Applications 70 1–7.
Tsybakov, A. B. (2009). Introduction to Nonparametric Estimation. Springer, New York.
Wegkamp, M. and Zhao, Y. (2016). Adaptive estimation of the copula correlation matrix for
semiparametric elliptical copulas. Bernoulli 22 1184–1226.
Wiwie, C., Baumbach, J. and Röttger, R. (2015). Comparing the performance of biomedical
clustering methods. Nature methods 12 1033–1038.
Yuan, M. and Lin, Y. (2007). Model selection and estimation in the gaussian graphical model.
Biometrika 94 19–35.
29
A
A.1
Appendix
Proofs of the results from Section 2
We begin by stating and proving two lemmata that are crucial for the main results of this section.
All results are proved under the condition that model 1 and (i) - (iii) hold.
Lemma 7. For any a ∈ [K] and i ∈ Ia , we have
(a) |Σij | = Caa for all j ∈ Ia ,
(b) |Σij | < Caa for all j 6∈ Ia .
Proof. For given i ∈ [p], we define the set s(i) := {1 ≤ a ≤ K : Aia 6= 0}. For any i ∈ Ia and j 6= i,
we have
|Σij | =
=
X
Aia
a∈s(i)
X
X
Ajb Cab
b∈s(j)
Ajb Cab from the definition of Ia
b∈s(j)
≤
X
b∈s(j)
|Ajb | · max |Cab |
b∈s(i)
≤ Caa using conditions (i) and (iii).
Furthermore, using conditions (i) and (iii), we observe that we have equality in the above display
for j ∈ Ia , and strict inequality for j 6∈ Ia , which proves the lemma.
Lemma 8. We have
(a) Si ∩ I 6= ∅, for any i ∈ [p],
(b) Si ∪ {i} = Ia and Mi = Caa , for any i ∈ Ia and a ∈ [K],
where Mi and Si are defined in (4) and (5), respectively.
Proof. Lemma 7 implies that, for any i ∈ Ia , Mi = Caa and Si = Ia \ {i}, which proves part (b).
From the result of part (b), it remains to show Si ∩ I 6= ∅ for any i ∈
/ I. Let i ∈
/ I be fixed. We
have
!
X
X
(41)
Ajb
Aia Cab
Mi = max |Σij | = max
j6=i
≤ max max
j6=i b∈s(j)
j6=i
X
b∈s(j)
a∈s(i)
Aia Cab = max
j6=i
a∈s(i)
P
X
Aia Cab∗
a∈s(i)
for some b∗ ∈ [K]. A direct computation yields |Σij | = | a∈s(i) Aia Cab∗ | for any j ∈ Ib∗ , that is,
the maximum Mi of |Σij | is achieved at all j ∈ Ib∗ . Since Ib∗ 6= ∅ by condition (ii), this completes
the proof of claim (a).
30
Proof of Theorem 1. We have all the necessary ingredients to proceed with the proof of the
main result of this section.
Proof of (a). We first show the sufficiency part. Consider any i ∈ [p] with Mi = Mj for all j ∈ Si .
Part (a) of Lemma 8 states that there exists a j ∈ Ia ∩ Si for some a ∈ [K]. For this j ∈ Ia , we
have Mj = Caa from part (b) of Lemma 8. Invoking our premise Mj = Mi as j ∈ Si , we conclude
that Mi = Caa , that is, maxk6=i |Σik | = Caa . By Lemma 7, the maximum is achieved for any pair
i, k ∈ Ia . However, if i 6∈ Ia , we have that |Σik | < Caa for all k 6= i. Hence i ∈ Ia and this concludes
the proof of the sufficiency part.
It remains to prove the necessity part. Let i ∈ Ia for some a ∈ [K] and j ∈ Si . Lemma
8 implies that j ∈ Ia and Mi = Caa . Since j ∈ Si , we have |Σij | = Mi = Caa , while j ∈ Ia
yields |Σjk | ≤ Caa for all k 6= j, and |Σjk | = Caa for k ∈ Ia , as a result of Lemma 7. Hence,
Mj = maxk6=j |Σjk | = Caa = Mi for any j ∈ Si , which proves our claim.
Proof of (b). We start with the following constructive approach. Let N = [p] be the set of all
variable indices and O = ∅. Let Mi and Si be defined in (4) and (5), respectively.
(1) Choose i ∈ N and calculate Si and Mi .
(a) If Mi = Mj , for all j ∈ Si , set I (i) := Si ∪ {i}, O = O ∪ {i} and N = N \ I (i) .
(b) Otherwise, replace N by N \ {i}.
(2) Repeat step (1) until N = ∅.
We show that {I (i) : i ∈ O} = I. Let i ∈ O be arbitrary fixed. By (a), we have i ∈ I. Thus, there
exists a ∈ [K] such that i ∈ Ia . By Lemma 8, i ∈ Ia implies Ia = Si ∪ {i} = I (i) . On the other
hand, let a ∈ [K] be arbitrary fixed. By condition (ii), there exists at least one j ∈ Ia . Once again,
by part (b) of Lemma 8, if j ∈ Ia , then Sj ∪ {j} = Ia , that is, I (j) = Ia .
Proof of Theorem 2
Theorem 1 shows that Σ uniquely defines I and its partition I, up to permutation of labels. Given
I and its partition I = {I1 , . . . , IK }, for any i ∈ I, there exists a unique 1 ≤ a ≤ K such that
i ∈ Ia . Then we set |Ai· | = ea , the canonical basis vector in RK that contains 1 in position a and is
zero otherwise. Thus, the |I| × K matrix AI with rows Ai· is uniquely defined up to multiplication
with a signed permutation matrix P .
We show below that AJ is also identifiable up to a signed permutation matrix. We begin by
observing that, for each i ∈ Ik , for some k ∈ [K], and any j ∈ J, Model 1 implies
X X
X
Σij =
Aia Ajb Ckb = Aik
Ajb Ckb
a∈s(i) b∈s(j)
and since A2ik = 1, we obtain
b∈s(j)
T
Aik Σij = Ck·
Aj·
31
and, after averaging over all i ∈ Ik ,
T
Ck·
Aj· =
1 X
Aik Σij .
|Ik |
i∈Ik
Repeating this for every k ∈ [K], we obtain the formula
T
X
X
1
1
CAj· =
Ai1 Σij , . . . ,
AiK Σij := θj .
|I1 |
|IK |
i∈I1
i∈IK
The covariance matrix C can be uniquely constructed from Σ via
1
|Ia |(|Ia | − 1)
Caa =
for any a ∈ [K], and
1
|Ia ||Ib |
Cab =
X
X
i,j∈Ia ,i6=j
|Σij |
Aia Ajb Σij
i∈Ia ,j∈Ib
for a, b ∈ [K] with a 6= b. Notice that mina∈[K] |Ia | ≥ 2, which is part of our model requirement
(ii), is needed for the construction of Caa . Since the covariance matrix C is assumed to be positive
definite, Aj· = C −1 θj , for each j ∈ J, which shows that AJ can be determined uniquely from Σ up
to a signed permutation. Therefore, AJ is identifiable which concludes the proof.
A.2
Proofs of the results from Section 4.1
The proof of Theorem 3 will repeatedly use Lemma 9, stated and proved below. Let
ci := max |Σ
b ij |.
M
j∈[p]\{i}
(42)
Lemma 9. Under the conditions in Theorem 3, for any i ∈ Ia with some a ∈ [K], the following
inequalities hold on the event E:
b ij | − |Σ
b ik | ≤ 2δ,
|Σ
b ij | − |Σ
b ik | > 2δ,
|Σ
b ij | − |Σ
b ik | < 2δ,
|Σ
For any i ∈ J1a , we have
for all j, k ∈ Ia \ {i} and j 6= k;
(43)
for all j ∈ Ia \ {i}, k ∈
/ (Ia ∪ J1a );
(44)
for all j ∈ J1a and k ∈ Ia \ {i}.
ci − |Σ
b ij | ≤ 2δ,
M
for any j ∈ Ia .
(45)
(46)
Proof of Lemma 9. For the entire proof, we work on the event E defined in (23). To prove (43), we
observe that, for any i, j, k ∈ Ia , Σij = Σik = Caa by Lemma 7, whence
b ij | − |Σ
b ik |
|Σ
≤
|Σij | − |Σik | + 2δ = 2δ.
32
To prove (44), we first observe that, for any j ∈ Ia , |Σij | = Caa by Lemma 7, whence
E
b ij | ≥ Caa − δ.
|Σ
Next, we notice that, for any ` ∈ [p],
|Σi` | =
K
X
A`b Cab = A`a Caa +
b=1
(iii)
X
(47)
A`b Cab
b6=a
≤ |A`a |Caa + (1 − |A`a |)(Caa − ν) = Caa − (1 − |A`a |)ν.
(48)
For any j ∈ Ia and k ∈ [p] \ (Ia ∪ J1a ), the definition of J1 implies |Aka | ≤ 4δ/ν, hence
(48)
E
b ik | ≤ |Σik | + δ ≤ Caa − (1 − |Aka |)ν + δ ≤ Caa − ν + 5δ,
|Σ
so that
E
b ij | − |Σ
b ik | ≥ |Σij | − δ − |Σ
b ik | ≥ |Σij | − Caa + ν − 6δ > 2δ,
|Σ
by using ν > 8δ · (kCk∞ /ν) ≥ 8δ. To prove (45), observe that, for any j ∈ J1a and k ∈ Ia \ {i},
(48)
E
b ij | ≤ Caa − (1 − |Aja |)ν + δ < Caa + δ = |Σik | + δ ≤ |Σ
b ik | + 2δ.
|Σ
So far, we have proved (43) - (45) and it remains to show (46). For any i ∈ J1a , we have, for some
c ∈ [K],
E
ci ≤ max |Σik | + δ (41)
M
=
k∈[p]\i
K
X
b=1
(∗)
Aib Cbc + δ ≤
K
X
b=1
E
b ij | + 2δ.
Aib Cba + δ = |Σij | + δ ≤ |Σ
It remains to show that inequality (∗) holds, for any c 6= a. On the one hand, we have
K
X
b=1
(iii)
Aib Cbc ≤ |Aia ||Cac | + (1 − |Aia |)Ccc ≤ |Aia |(Caa − ν) + (1 − |Aia |)Ccc ,
while on the other hand, we find
K
X
b=1
(iii)
Aib Cab ≥ |Aia ||Caa | − (1 − |Aia |)(Caa − ν).
Combining the preceding two display yields
K
X
b=1
Aib Cab −
K
X
b=1
Aib Cbc ≥ ν − (1 − |Aia |)(Caa + Ccc ).
The term on the right is positive, since condition (28) guarantees that
4δ
(Caa + Ccc ) ≥ (1 − |Aia |)(Caa + Ccc ),
ν
where the last inequality is due to the definition of J1 . This concludes the proof.
ν>
33
Lemma 9 remains valid under the conditions of Remark 2 in which case J1 = ∅ and we only
need ν > 4δ to prove (44).
Proof of Theorem 3. We work on the event E throughout the proof. Without loss of generality,
we assume that the label permutation π is the identity. We start by pointing out that the following
three claims are sufficient to prove (a) - (c). Let Ib(i) be defined in step 4 of Algorithm 1.
(1) For any i ∈ J \ J1 , we have P ure(i) = F alse.
(2) For any i ∈ Ia and a ∈ [K], we have P ure(i) = T rue, Ia ⊆ Ib(i) and Ib(i) \ Ia ⊆ J1a .
(3) For any i ∈ J1a and a ∈ [K], we have Ia ⊆ Ib(i) .
If we can prove these claims, then (1) implies that none of variables in J \ J1 will be selected in
any set of Ib via i ∈ J \ J1 . (2) implies that for any a ∈ [K], there exists Iba such that Ia ⊆ Iba and
Iba \ Ia ⊆ J1a . Moreover, this together with Merge in Algorithm 1 prevents Iba from selecting any
variable from [p] \ (Ia ∪ J1a ). Finally, (3) guarantees that none of pure variables will be excluded
b = K and Ib = {Ib1 , . . . , IbK } is the desired partition.
by any i ∈ J1 in the Merge step. Thus, K
Therefore, in the following we proceed to prove (1) - (3).
To prove (1), let i ∈ J \ J1 be fixed. We first prove that P ure(i) = F alse when Ib(i) ∩ I 6= ∅. It
suffices to show that, there exists j ∈ Ib(i) such that the following does not hold
cj − |Σ
b ij | ≤ 2δ.
M
(49)
Let Ib(i) ∩ I 6= ∅, so there exists j ∈ Ib ∩ Ib(i) for some b ∈ [K]. For such j, we have |Σij | =
P
| K
a=1 Aia Cab | and
E
b ij | ≤
|Σ
K
X
a=1
(iii)
Aia Cab + δ ≤ |Aib |Cbb + (1 − |Aib |)(Cbb − ν) + δ < Cbb − 3δ,
(50)
using the definition of J1 to justify the last inequality. On the other hand, since j ∈ Ib , part (b) of
Lemma 8 implies
E
cj = max |Σ
b jk | ≥ max |Σjk | − δ = Cbb − δ.
M
(51)
k∈[p]\{i}
k∈[p]\{i}
cj − |Σ
b ij | > 2δ. This shows that for any i ∈ J \ J1 , if Ib(i) ∩ I 6= ∅,
Combining (50) with (51) gives M
then P ure(i) = F alse. Therefore, to complete the proof of (1), we show Ib(i) ∩ I = ∅ is impossible
when i ∈ J \ J1 under our assumptions. If Ib(i) ∩ I = ∅, then there exists some j ∈ J ∩ Ib(i) and
|Σij | =
K X
K
X
b=1 a=1
Aia Ajb Cab ≤ max
1≤b≤K
K
X
Aia Cab =
a=1
K
X
a=1
Aia Cab∗ = |Σik |
for some b∗ ∈ [K] and any k ∈ Ib∗ (the set Ib∗ is non-empty by condition (ii)). Therefore,
E
b ij | − |Σ
b ik | ≤ |Σij | − |Σik | + 2δ
|Σ
34
≤
2δ
However, since Ib(i) ∩ I = ∅ and k ∈ Ib∗ , we know k ∈
/ Ib(i) , which implies
b ij | − |Σ
b ik | > 2δ,
|Σ
from Step 4 of Algorithm 1. The last two displays contradict each other, and we conclude that, for
any i ∈ J \ J1 , Ib(i) ∩ I 6= ∅.
To prove (2), let i ∈ Ia be arbitrarily fixed with some a ∈ [K]. We first show that P ure(i) =
T rue. From steps 7 - 8 of Algorithm 1, it suffices to show that, for any j ∈ Ib(i) , (49) holds. From
(44) in Lemma 9, given Step 4 of Algorithm 1, we know that, for any j ∈ Ib(i) , j ∈ Ia ∪ J1a . Thus, we
cj is achieved by
write Ib(i) = (Ib(i) ∩ Ia ) ∪ (Ib(i) ∩ J1a ). For any j ∈ Ib(i) ∩ Ia , by the same reasoning, M
a
some element in either Ia or J1 . For both cases, since i, j ∈ Ia and i 6= j, (43) and (45) in Lemma
9 guarantee that (49) holds. On the other hand, for any j ∈ Ib(i) ∩ J1a , (46) in Lemma 9 implies
that (49) still holds. Thus, we have shown that, for any i ∈ Ia , P ure(i) = T rue. To show Ia ⊆ Ib(i) ,
ci can only be achieved by indices in Ia ∪ J a . In both cases,
let any j ∈ Ia \ {i} and observe that M
1
(i)
(43) and (45) imply j ∈ Ib . Thus, Ia ⊆ Ib(i) . Finally, Ib(i) \ Ia ⊆ J1a follows immediately from (44).
We conclude the proof by noting that (3) immediately follows from (46).
A.3
Proofs of the results from Section 4.2
We divide the proof of Theorem 4 into three steps:
bI aligns
Step 1. We show that there exists a signed permutation Pb such that the columns of A
with those of AI in terms of label and sign, as detailed in Lemma 10;
bb − Ā b;
Step 2. We write Ā = APb, and prove first the error bounds for A
I
I
b
b
b
Step 3. We prove the error bounds for A − Ā = A − AP , with the same Pb, which further implies
b and A.
that Pb aligns the columns of A
Lemma 10. Under conditions of Theorem 4, there exists a signed permutation matrix Q such that
bia ) for any i ∈ Iba with each a ∈ [K].
Ā = AQ satisfies that sign(Āia ) = sign(A
b = K, I ⊆ Ib ⊆ I ∪ J1 and Iπ(a) ⊆ Iba ⊆ Iπ(a) ∪ J π(a) ,
Proof of Lemma 10. Theorem 3 guarantees K
1
with high probability, for any a ∈ [K] and some label permutation π. Let us write Q = Q1 Q2 , with
bI ,
the unsigned permutation matrix Q1 which relabels the columns of AI according to those of A
and with Q2 = diag(q1 , . . . , qK ) with qa ∈ {+1, −1} for each a ∈ [K].
q = AQ1 , we proceed to show that, for each a ∈ [K], sign(A
qia ) · qa holds
bia ) = sign(A
Denoting A
π(a)
b
b
for any i ∈ Ia , in which case each qa can be uniquely constructed. Since Ia ⊆ Iπ(a) ∪ J1 , it suffices
to prove that, for any a ∈ [K],
bja )
bia )
sign(A
sign(A
=
,
qia )
qja )
sign(A
sign(A
π(a)
for any i, j ∈ Iπ(a) or i, j ∈ J1
with i 6= j.
(52)
bI , for any i, j ∈ Iπ(a) or i, j ∈ J π(a) , we
From the definition of AI and the way we construct A
1
consider the following two cases:
qia ) = sign(A
qja ). To show A
bia = A
bja , from (6),
If sign(Aiπ(a) ) = sign(Ajπ(a) ), this implies sign(A
1
2
b
b
b
we need to show i, j ∈ Ia or i, j ∈ Ia which is equivalent to show Σij > 0. For any i, j ∈ Iπ(a) or
35
P
π(a)
i, j ∈ J1
with i 6= j, display (26) gives |Akπ(a) | ≥ 1 − 4δ/ν and b6=π(a) |Akb | ≤ 4δ/ν, for k = i, j.
Thus, using sign(Aiπ(a) ) = sign(Ajπ(a) ), we have
X
X
X
Σij = Aiπ(a) Ajπ(a) Cπ(a)π(a) + Aiπ(a)
Ajc Cπ(a)c + Ajπ(a)
Aib Ajc Cbc
Aib Cπ(a)b +
c6=a
b6=a
b,c6=π(a)
(iii)
≥ Aiπ(a) Ajπ(a) Cπ(a)π(a) − |Aiπ(a) |(1 − |Ajπ(a) |)(Cπ(a)π(a) − ν)
X
Aib Ajc Cbc
− |Ajπ(a) |(1 − |Aiπ(a) |)(Cπ(a)π(a) − ν) −
b,c6=π(a)
2
8δ
16δ 2
4δ
4δ
≥ 1−
Cπ(a)π(a) −
Cπ(a)π(a) − 2 Cb∗ b∗ + 8δ,
· 1−
ν
ν
ν
ν
2
2
16δ
16δ
48δ
≥ 1+ 2 −
Cπ(a)π(a) − 2 Cb∗ b∗ + 8δ.
ν
ν
ν
for some b∗ 6= π(a)
b ij ≥ Σij − δ > 3δ > 0.
Since (28) implies 8δCb∗ b∗ < ν 2 and ν > 8δ, on the event E, we have Σ
qia ) 6= sign(A
qja ). Similarly, to show A
bia 6= A
bja ,
If sign(Aiπ(a) ) 6= sign(Ajπ(a) ), this gives sign(A
b ij < 0. Using the same arguments yields
we prove Σ
E
b kl ≤ Σkl + δ < −3δ < 0.
Σ
Therefore, given Ib = {Iba }a∈[K] , we can construct the signed permutation Pb = Q which alligns the
bI .
columns of AI with those of A
For ease of notation and without loss of generality, we make the blanket assumption that the
signed permutation Pb is the identity so that Ā = A for the remainder of the proof. We note that
the signed permutation Pb will be the same when estimating each row Aj· for j ∈ J.
bb and parts (a) - (c) in Theorem 3, we can write, for
Proof of step 2: From the construction of A
I
each a ∈ [K], Iba = Ia ∪ La with La := Iba ∩ J1a . For any i ∈ Iba , the definitions of I and J1a imply
bia ), we have
|Aia | ≥ 1 − 4δ/ν. Since Lemma 10 guarantees that sign(Aia ) = sign(A
bb − A bk∞ = max kA
bi. − Ai. k∞ ≤ 4 δ.
kA
I
I
ν
b
i∈I
b we have
Let si = kAi· k0 for i ∈ [p]. Then, for any i ∈ I,
bj· − Āj·
A
q
≤
4 1/q
s δ,
ν i
1 ≤ q ≤ ∞.
For Step 3 of the proof of Theorem 4, we will make use of the results of Lemmas 11 and 12, stated
here first and proved at the end of this section, in order to preserve the flow of the presentation.
Lemma 11. Under the conditions of Theorem 4, on the event E, we have
where δ 0 is given in (29).
b − Ck∞ ≤ 2δ 0 ,
kC
max kθbj − θj k∞ ≤ δ 0 ,
j∈Jb
36
(53)
Lemma 12. Under the conditions of Theorem 4, on the event E, we have βaj = 0 implies βbaj = 0,
b
for any j ∈ Jb and a ∈ [K].
b recall that β j = C −1 θj = Ωθj since C is invertible. Also recall
Proof of Step 3. For each j ∈ J,
b θbj . We first show kβ̄ j − β j k∞ ≤ 5kΩk∞,1 δ 0 . For notational convenience, we remove all
that β̄ j = Ω
the super indices. From Lemma 11, the following event
(
)
b − Ck∞ ≤ 2δ 0 , max kθbj − θj k∞ ≤ δ 0 ,
E 0 = kC
j∈Jb
is implied by the event E = Eδ . On the event E 0 , the true Ω := C −1 satisfies the constraint since
b − Ik∞ = kΩ(C
b − C)k∞ ≤ kC
b − Ck∞ kΩk∞,1 ≤ 2δ 0 kΩk∞,1 .
kΩC
b
Then the pair (kΩk∞,1 , Ω) of (t, Ω) is feasible. Consequently, the optimality and feasibility of (b
t, Ω)
imply
b ∞,1 ≤ b
bC
b − Ik∞ ≤ 2δ 0 b
t ≤ 2δ 0 kΩk∞,1 .
(54)
kΩk
t ≤ kΩk∞,1 ,
kΩ
Then, on the event E 0 , we obtain
b θb − Ωθ
b + Ωθ
b − βk∞
kβ̄ − βk∞ = kΩ
b ∞,1 kθb − θk∞ + kΩθ
b − βk∞
≤ kΩk
b ∞,1 + kΩCβ
b
≤ δ 0 kΩk
− βk∞
0
b − Ik∞ kβk1
≤ δ kΩk∞,1 + kΩC
bC
b − Ik∞ + kΩ
bC
b − ΩCk
b ∞
≤ δ 0 kΩk∞,1 + kΩ
b ∞,1 kC
b − Ck∞
≤ 3δ 0 kΩk∞,1 + kΩk
(since kβk1 ≤ 1)
≤ 5δ 0 kΩk∞,1 .
The feasibility of βbj implies that kβbj − β̄ j k∞ ≤ µ. By the triangle inequality, we obtain
kβbj − β j k∞ ≤ kβbj − β̄ j k∞ + kβ̄ j − β j k∞ ≤ 2µ,
b = K on the event E gives
since µ = 5δ 0 kΩk∞,1 . Then following from Lemma 12 and using K
bj· − Aj· kq =
kA
K
X
a=1
|βbaj − βaj |q
!1/q
=
X
a∈sj
1/q
|βbaj − βaj |q
1/q
≤ 2sj µ,
for any 1 ≤ q ≤ ∞. This completes the proof of the last step and of Theorem 4.
To conclude this section we give below the proofs of the intermediary results used in the proof.
37
b = K. Then, from the definition of C
baa ,
Proof of Lemma 11. On the event E, we showed that K
we have
X
1
baa − Caa | ≤ max
b ij | − Caa
max |C
|Σ
1≤a≤K
1≤a≤K |Iba |(|Iba | − 1)
b
i,j∈Ia ,i6=j
E
≤ δ+
1
|Iba |(|Iba | − 1)
X
i,j∈Iba ,i6=j
|Σij | − Caa .
Theorem 3 states that, on the event E, Iba = Ia ∪ La where La = Iba ∩ J1a , for any a ∈ [K]. Therefore,
we consider the following three cases:
(1) For any i, j ∈ Ia and i 6= j, Lemma 7 implies |Σij | − Caa = 0.
(2) For any i ∈ Ia and j ∈ La , the definition of J1a gives
8δ
|Σij | − Caa ≤ 1 − |Aja | (2Caa − ν) ≤ kCk∞ − 4δ.
ν
P
(3) For any i, j ∈ La and i 6= j, since i, j ∈ J1a , we know |Aka | ≥ 1 − 4δ/ν and b6=a |Akb | ≤ 4δ/ν,
for k = i, j. Thus,
X
X
X
|Σij | − Caa ≤ 1 − |Aia ||Aja | Caa + |Aia |
|Ajc ||Cac | + |Aja |
|Aib ||Cab | +
|Aib Ajc |Cbc
c6=a
b6=a
b,c6=a
≤ 1 − |Aia ||Aja | Caa + |Aia |(1 − |Aja |)(Caa − ν) + |Aja |(1 − |Aib |)(Caa − ν)
+ (1 − |Aib |)(1 − |Ajc |)Cb∗ b∗
(for some b∗ 6= a)
"
#
4δ 2
8δ
16δ 2
≤ 1− 1−
Caa + (Caa − ν) + 2 Cb∗ b∗
ν
ν
ν
≤
16δ
kCk∞ − 8δ,
ν
(by the definition of J1 )
(by (28)).
Therefore, by combining cases (1) - (3), we have
baa − Caa | ≤ δ + |Ia ||La | + |La |(|La | − 1) ·
max |C
1≤a≤K
|Iba |(|Iba | − 1)
16
≤
kCk∞ − 7 δ.
ν
16δkCk∞
− 8δ
ν
where the last inequality comes from that |La | + |Ia | = |Iba |. For the off-diagonal entries, since
bia ) = sign(Aia ), for any i ∈ Ib and a ∈ [K], we have
sign(A
max
1≤a,b≤K,a6=b
bab − Cab | ≤ δ +
|C
we consider the following three cases:
38
1
|Iba ||Ibb |
X
i∈Iba ,j∈Ibb
|Σij | − |Cab | ,
(1) For any i ∈ Ia , j ∈ Ib , we have |Σij | − |Cab | = 0.
(2) For any i ∈ Ia , j ∈ J1b , we have
|Σij | − |Cab | ≤ (1 − |Ajb |)|Cab | +
(3) For any i ∈ J1a , j ∈ J1b , we obtain
Σij
= Aia Ajb Cab + Aia
X
X
c6=b
|Ajc ||Cac | ≤
Ajd Cad +
d6=b
X
8δ
kCk∞ − 4δ.
ν
Aic
c6=a
X
Ajd Ccd .
d∈s(j)
Thus,
≤ (1 − |Aia ||Ajb |)|Cab | + |Aia |(1 − |Ajb |)kCk∞ + (1 − |Aia |)kCk∞
8δ 16δ 2
8δ 16δ 2
≤
− 2
(Caa − ν) +
− 2
kCk∞
ν
ν
ν
ν
16δ
≤
kCk∞ − 8δ.
(by ν < kCk∞ )
ν
|Σij | − |Cab |
Therefore, combining the three cases gives
bab − Cab | ≤ δ + |Ia ||Lb | + |La ||Ib | + 2|La ||Lb | ·
max
|C
1≤a,b≤K,a6=b
2|Iba ||Ibb |
16
kCk∞ − 7 δ.
≤
ν
16δkCk∞
− 8δ
ν
Combining the diagonal and off-diagonal cases yields
16
b
kC − Ck∞ ≤
kCk∞ − 7 δ ≤ 2δ 0
ν
bia ) for any i ∈ I,
b we obtain
We now proceed to bound maxj∈Jb kθbj − θj k∞ . From sign(Aia ) = sign(A
max kθbj − θj k∞ ≤ δ +
j∈Jb
X
1 X
Σij −
Ajb Cab .
a∈[K],j∈Jb |Iba |
max
i∈Iba
P
b∈s(j)
Since for any i ∈ Ia and any j ∈ J, Σij = b∈s(j) Ajb Cab , we focus on the case when i ∈ La . For
any i ∈ La and j ∈ J, (41) yields
X
X
X
X
Ajb Cab − Σij = (1 − Aia )
Ajb Cab −
Aic
Ajb Cbc ,
b∈s(j)
b∈s(j)
c6=a
b∈s(j)
which, by the definition of J1 , implies
X
b∈s(j)
Ajb Cab − Σij
≤ (1 − |Aia |)|Ajd ||Cad | + (1 − |Aia |)|Ajd0 ||Cad0 | (for some d, d0 ∈ [K])
≤
4δ
8δ
(2kCk∞ − ν) ≤
kCk∞ − 4δ.
ν
ν
39
Since we have Jb ⊆ J, we have
|La |
·
max kθbj − θj k∞ ≤ δ + max
a
|Iba |
j∈Jb
8
8
kCk∞ − 4 δ ≤
kCk∞ − 3 δ = δ 0 ,
ν
ν
which concludes the proof of Lemma 11.
Proof of Lemma 12. Let j ∈ Jb be arbitrarily fixed and βbj be the optimal solution of (16) with
µ = 5kΩk∞,1 δ 0 . For simplicity, we remove the super indices. Starting with the following KarushKuhn-Tucker condition:
subject to
we obtain
sign(βba ) + λa sign(βba − β̄a ) = 0,
λa (|βba − β̄a | − µ) = 0,
λa ≥ 0,
for a = {1, . . . , K},
(56)
b
b
b
b
b
0 = sign(βa ) βa − β̄a + λa βa − β̄a = sign(βa ) βa − β̄a + λa µ,
(55)
(56)
(57)
by multiplying both sides of (55) by βba − β̄a . In what follows we prove that if βa = 0, for some a,
then βba = 0. Since this is true when λa = 0 from (55), we only consider when λa 6= 0. Note this
implies |βba − β̄a | = µ from (56). If we assume βba > 0, then (57) gives
β̄a − βba = λa µ.
Since |βba − β̄a | = µ, we further obtain λa = 1 and
β̄a = µ + βba > µ.
(58)
Recall that kβ − β̄k∞ ≤ µ. This implies β̄a ≤ µ + |βa | = µ, which contradicts (58), so βba cannot
be strictly positive. Similarly, βba < 0 cannot hold based on similar arguments. Thus, βba = 0 from
b
which we conclude supp(βbj ) ⊆ supp(β j ) for any j ∈ J.
Proof of Theorem 5. First we construct a set of “hypotheses” of A. Let
M := {v ∈ {0, 1}K : dH (0, v) = s}
where dH (·) denotes the Hamming distance between two binary vectors. Following Lemma A.3 in
Rigollet and Tsybakov [2011] when s ≤ 4K/5, there exists M0 ⊂ M such that, for any w(i) 6=
w(j) ∈ M 0 ,
dH w(i) , w(j) > s/16,
(59)
and
log |M0 | ≥ c0 s log(K/s),
40
(60)
for some constant c0 > 0. We let w(0) = (0, . . . , 0) ∈ RK . Then, we choose
#
"
B
T ∈ Rp×K , for each j = 0, 1, . . . , |M0 |,
A(j) =
η w(j)
where
and
B1
B2
(p−1)×K
B= .
,
∈R
..
BK
η=
r
eT
kT
ek
|Ik |×K
Bk =
, for k ∈ [K],
.. ∈ R
.
eTk
c0 c1
8
r
log(1 + K/s)
.
n
(61)
(62)
(63)
2
with some constant c1 = σmin
∧ 1. We use ek to denote the canonical basis of K dimensional space
and 0 to denote the zero vector. Note that, for each Bk , the only non-zero values are at the kth
P
(j) with only one
column. By specifying as above, we choose K
k=1 |Ik | = p − 1 and consider the A
non-pure row. It is easy to verify that A(j) ∈ A for each j = 0, 1, . . . , |M0 | under (31).
We denote by KL(P, Q) the Kullback-Leibler divergence between two probability distributions
P and Q. In order to apply Theorem 2.5 in Tsybakov [2009] to prove (32), for fixed α ∈ (0, 1/8),
we need to check the following three conditions:
(a) KL(PA(i) , PA(0) ) ≤ α log |M0 |, for each i = 1, . . . , |M0 |.
(b) For any 0 ≤ i < j ≤ |M0 |, with some constant c0 > 0,
r
log(K/s)
0 1/q
(i)
(j)
≥cs
.
Lq A , A
n
(c) Lq ( · ) satisfies the triangle inequality.
To show (a), note that X ∼ N (0, ACAT + Γ) from Z ∼ N (0, C = IK ). Invoking Lemma 14
gives
nη 2 s
1
≤
KL PA(i) , PA(0) ≤
log |M 0 |, for any i = 1, . . . , |M 0 |,
(64)
2
16
2σmin
2
by using (60), (63) and c1 = σmin
∧ 1.
To prove (b), for any i = 1, . . . , |M0 |, observe that
Lq A(i) , A(0) = ηkw(i) kq = s1/q η
and, for any i 6= j different from 0,
Lq A(i) , A(j) = ηkw(i) − w(j) kq ≥ (s/16)1/q η ≥ (s1/q η)/16,
by using (59). Combining these two and using the expression of η yield
r
log(K/s)
(i)
(j)
0 1/q
,
Lq A , A
≥cs
n
41
(65)
for 0 ≤ i < j ≤ |M0 |.
e A)
b
Finally, we verify (c) by showing that Lq (·) satisfies the triangle inequality. Consider (A, A,
and observe that
e =
Lq (A, A)
=
≤
=
e ∞,q
min kAP − Ak
P ∈HK
e ∞,q
min kAP − AQk
b ∞,q + kA
b − AQk
e ∞,q
min
kAP − Ak
P,Q∈HK
P,Q∈HK
b ∞,q + min kA
b − AQk
e ∞,q
min kAP − Ak
P ∈HK
Q∈HK
b + Lq (A,
e A).
b
= Lq (A, A)
Therefore, we conclude the proof of (32) by invoking the Theorem 2.5 in Tsybakov [2009].
Lemma 13. (Sherman-Morrison formula) Suppose M ∈ Rn×n is an invertible square matrix and
u, v ∈ Rn are column vectors. Then M + uv T is invertible if and only if 1 + v T M −1 u 6= 0. If
M + uv T is invertible, then its inverse is given by
(M + uv T )−1 = M −1 −
M −1 uv T M −1
.
1 + v T M −1 u
Lemma 14. Assume model (1) and (Z, E) ∼ N (0, diag(IK , Γ)). Let A(0) and A(i) be constructed
as (61) and (62), for any 1 ≤ i ≤ M 0 with M 0 satisfying (60). Let PA(0) and PA(i) be the probability
densities of X parametrized by A(0) and A(i) , respectively. Then we have
KL (PA(i) , PA(0) ) ≤
nη 2 s
2 ,
2σmin
(66)
2
where σmin
= min1≤i≤p Γii .
Proof of Lemma 14. Notationally, we denote by I1 , . . . , IK the set of indices, by Id the d×d diagonal
matrix, by 1|Ik | the |Ik |-dimensional vector and by ΓIk the Ik ×Ik submatrix of Γ. Finally, 0 denotes
the zero-vector and its length might vary line by line.
From the property of Kullback-Leibler divergence, we only need to verify the case when n = 1.
We consider arbitrary A(i) constructed as (61) and (62) for some 0 ≤ i ≤ M 0 . For notational
simplicity, we write A = A(i) = (B T , ξ)T where ξ = ηw(i) ∈ RK . For this A ∈ Rp×K , under
assumptions (a) - (c), observe that
"
#
"
#
T +Γ
BCB
Bξ
Σ
Σ
11
12
B
Σ = ACAT + Γ =
:=
ξT BT
kξk2 + γ
Σ21 Σ22
e 6= A constructed in
by decomposing Γ := diag(ΓB , γ) := diag(ΓI1 , . . . , ΓIK , γ). Similarly, for any A
the same way, we have
"
#
"
#
T +Γ
e
e
BCB
B
ξ
Σ
Σ
11
12
B
e = AC
e A
eT + Γ =
Σ
e 2 + γ := Σ
e 21 Σ
e 22
ξeT B T
kξk
42
Since the Kullback-Leibler divergence between two p-dimensional multivariate normal distributions
N0 := Np (µ0 , Σ0 ) and N1 := Np (µ1 , Σ1 ) is given by
det Σ1
1
> −1
−1
tr Σ1 Σ0 + (µ1 − µ0 ) Σ1 (µ1 − µ0 ) − p + log
,
KL(N0 , N1 ) =
2
det Σ0
we have
1
det Σ
−1 e
KL(PAe, PA ) =
tr Σ Σ − p + log
,
e
2
det Σ
(67)
from the fact that X has mean zero. By using the formula of the inverse of a block matrix
"
#−1 "
#
A B
A−1 + A−1 B(D − CA−1 B)−1 CA−1 −A−1 B(D − CA−1 B)−1
=
,
C D
−(D − CA−1 B)−1 CA−1
(D − CA−1 B)−1
for square matrices A and D and non-singular matrices A and D − CA−1 B, we have
#
"
#
"
−1
−1
−1
−1
Ω11 Ω12
Σ−1
−Σ−1
−1
11 + Σ11 Σ12 Σ22·1 Σ21 Σ11
11 Σ12 Σ22·1
:=
,
Σ =
−1
Σ−1
Ω21 Ω22
−Σ−1
22·1
22·1 Σ21 Σ11
with Σ22·1 = Σ22 − Σ21 Σ−1
11 Σ12 . This gives
e 2 + γ) .
e T B T + Ω21 B ξe + Ω22 (kξk
e = tr Ω11 Σ11 + Ω12 (ξ)
tr Σ−1 Σ
|
{z
}
{z
}
|
T
2
T1
We first calculate T1 . Observe that
e T BT
Ω11 Σ11 + Ω12 (ξ)
−1 e T T
−1
−1
= Ip−1 + Σ−1
11 Σ12 Σ22·1 Σ21 − Σ11 Σ12 Σ22·1 (ξ) B
−1
T T
Σ21 = ξ T B T
= Ip−1 + Σ−1
11 Σ12 Σ22·1 ∆ B ,
where ∆ := ξ − ξe ∈ RK . Recall that C = IK . It implies
D1
..
Σ11 = BCB T + ΓB =
, with
.
DK
(68)
Dk = 1|Ik | 1T|Ik | + ΓIk .
Thus,
Σ−1
11
−1
D1
=
..
.
−1
DK
,
with
Dk−1
=
Γ−1
Ik
−
−1
T
Γ−1
Ik 1|Ik | 1|Ik | ΓIk
1 + 1T|Ik | Γ−1
Ik 1|Ik |
, for each k ∈ [K],
P
where the second part uses the Sherman-Morrison formula in Lemma 13. Let αk := i∈Ik 1/σi2 .
Then,
eT
! kT "
#
−1
−1
T
−1
Γ
1
1
Γ
e
Γ
1
|I
|
|I
|
I
I
k
|I
|
k
k
k
k
. = 0 · · · Ik k · · · 0 ∈ R|Ik |×K
Dk−1 Bk = Γ−1
(69)
Ik −
.
1 + αk
1 + αk
.
eTk
43
where only the kth column are non-zero. This further implies
−1
ξ1 −1
D1 B1
1+α1 ΓI1 1|I1 |
..
..
p−1
Σ−1
ξ =
∈R .
.
.
11 Σ12 =
(70)
ξK
−1
1+αK ΓIK 1|IK |
−1
DK
BK
On the other hand, note that
h
i
∆T B T = ∆1 1T|I1 | · · · ∆K 1T|IK | ∈ R1×(p−1) .
(71)
Since Σ22·1 is a scalar, combining (68), (70) and (71) yields
−1
T T
T1 = p − 1 + tr Σ−1
11 Σ12 Σ22·1 ∆ B
ξ1 −1
1+α1 ΓI1 1|I1 |
h
i
..
= p − 1 + Σ−1
∆1 1T|I1 | · · · ∆K 1T|IK |
.
22·1 · tr
ξK
−1
1+αK ΓIK 1|IK |
= p−1+
Σ−1
22·1
= p−1+
Σ−1
22·1
K
X
ξk ∆k −1
tr ΓIk 1|Ik | 1T|Ik |
1 + αk
k=1
K
X
k=1
αk
ξk ∆ k .
1 + αk
by using tr
Γ−1
Ik
= αk
Then we proceed to calculate T2 . Similar as (70) and (71), we have
i ξ
h
ξK
1
−1
−1
T
T
T
−1
−1
T
T
=
∈ R1×(p−1)
=
ξ
1
Γ
·
·
·
1
Γ
Σ21 Σ−1
· · · BK DK
B1 D1
11
1 + α1 |I1 | I1
1 + αK |IK | IK
and
1|I1 | ξe1
B ξe = ... ∈ Rp−1 .
1|I | ξeK
K
These two immediately yield
−1 e
−1
e2
T2 = −Σ−1
22·1 Σ21 Σ11 B ξ + Σ22·1 (kξk + γ)
=
Σ−1
22·1
= Σ−1
22·1
K
X
ξk ξek T −1
e +γ−
1 Γ 1
kξk
1 + αk |Ik | Ik |Ik |
k=1
!
K
X
α
k
e 2+γ−
kξk
ξk ξek .
1 + αk
2
!
k=1
Adding up T1 and T2 gives
e = p − 1 + Σ−1
tr Σ−1 Σ
22·1
e 2+γ+
kξk
44
K
X
k=1
!
αk
ξk ∆k − ξek
.
1 + αk
(72)
We proceed to obtain the expression of Σ22·1 . Observe that
−1
h
i D1
T
..
T
Σ21 Σ−1
B1T · · · BK
.
11 Σ12 = ξ
= ξT
K
X
BkT Dk−1 Bk
k=1
!
−1
DK
B1
..
. ξ
BK
ξ
From (69), we further obtain
0
0
.
..
.. "
.
#
−1
ΓIk 1|Ik |
T
T −1
··· 0 =
Bk Dk Bk = 1|Ik | 0 · · ·
.
1 + αk
.
.
0
1T
Γ−1 1|Ik |
|I | k
k
1+αk
..
0
which is a K × K diagonal matrix. Thus,
2
Σ22·1 = Σ22 − Σ21 Σ−1
11 Σ12 = kξk + γ −
=
K
X
k=1
.
K
X
k=1
∈ RK×K ,
αk
ξ2
1 + αk k
1
ξ 2 + γ.
1 + αk k
(73)
e Recall that the inverse formula of
Finally, it remains to calculate the determinant of Σ and Σ.
a block matrix is
!
A B
det
= det(A) det(D − CA−1 B)
C D
for the invertible matrix A. We thus obtain
e = det Σ11 · Σ
e 22·1 ,
det Σ
det Σ = det Σ11 · Σ22·1 ,
e 22·1 = Σ
e 22 − Σ
e 21 Σ−1 Σ
e
with Σ
11 12 . This gives
det Σ
e 22·1
= log Σ22·1 − log Σ
log
e
det Σ
(74)
Following the same arguments from (73), one can derive
e 22·1 =
Σ
K
X
k=1
Therefore, collecting (72) and (74) yields
2KL PAe, PA
= Σ−1
22·1
e 2+γ+
kξk
K
X
k=1
1 e2
ξ + γ.
1 + αk k
αk
ξk ∆k − ξek
1 + αk
45
(75)
!
e 22·1 ,
− 1 + log Σ22·1 − log Σ
e 22·1 specified in (73) and (75), respectively. Given our specification of A = A(0)
with Σ22·1 and Σ
e = A(i) in (61) and (62), we further have
and A
ξ = 0 =⇒ Σ22·1 = γ,
and
η2s
+ γ,
2
2
1 + Imax /σmin
1 + Imax /σmin
2
= maxk∈[K] |Ik | and σmin
= mini∈[p] σi2 . Consequently, KL PAe, PA can be simplified as
(75)
with Imax
e 2 = η 2 s,
ξe = ηw(i) =⇒ kξk
e 22·1 ≥
Σ
2KL PAe, PA
=
≤
1
e 2+γ =
kξk
η2s
e 22·1
+ log γ − log Σ
γ
η2s
η2s
+γ
+ log γ − log
2
γ
1 + Imax /σmin
≤
η2s
η2s
− 2
2 )γ
γ
η s + (1 + Imax /σmin
=
2
Imax /σmin
+ (η 2 s)/γ
η2s
·
2
γ 1 + Imax /σmin
+ (η 2 s)/γ
≤
η2s
2 ,
σmin
(the concavity of log function)
2 . This concludes the proof of Lemma 14.
by using γ ≥ σmin
A.4
Proofs for the results from Section 4.3
We first prove the three statements of Theorem 6, then present the proofs of Remark 4. Without
loss of generality, we assume that the signed permutation P is identity.
Proof of Theorem 6. We first give the proof for part (a). Then, for ease of the presentation, we
prove part (c) first and then part (b).
b b) ⊆ supp(A b). In addition,
Proof of part (a). Recall that Lemma 12 immediately implies supp(A
J
J
b From the way we construct A
bb, we have supp(A
bb) ⊆
Theorem 3 yields Iba ⊆ Ia ∪ J1a , for any a ∈ K.
I
I
b ⊆ supp(A).
supp(AIb). Therefore, we have proved supp(A)
On the other hand, for any (j, a) ∈ supp(AJ2 ), we know |βaj | > 2µ. This and the fact that
kβbj − β j k∞ ≤ 2µ, immediately gives
|βbaj | ≥ |βaj | − kβbj − β j k∞ > 0,
bJ ).
which implies supp(AJ2 ) ⊆ supp(A
2
bia ) = sign(Aia ) for any (i, a) ∈
b
To show sign(ASb) = sign(ASb), since Lemma 10 guarantees sign(A
b we focus on any fixed (j, a) ∈ Sb and j ∈ J.
b First, we consider the case A
bja = βbaj > 0.
Sb and i ∈ I,
Removing super indices, if βba > 0, (58) gives β̄a > µ. Thus, βa ≥ β̄a − kβ − β̄k∞ > 0 by recalling
bja > 0, (j, a) ∈ Sb and j ∈ J,
b we have Aja > 0.
kβ − β̄k∞ ≤ µ. So far, we have shown that, for any A
b
b
Since the same argument holds for any Aja < 0, the proof of sign(ASb) = sign(ASb) is completed.
46
Proof of part (c). Recall that, for any i ∈ [p] and a ∈ [K],
b a ⇐⇒ A
bia =
i∈G
6 0.
i ∈ Ga ⇐⇒ Aia 6= 0,
We start our proof by rewriting the equivalent expression of TFPP and TFNP:
TFPP =
P
i∈[p],a∈[K] 1{Aia
P
bia 6= 0}
= 0, A
i∈[p],a∈[K] 1{Aia
= 0}
,
TFNP =
P
i∈[p],a∈[K] 1{Aia
P
bia = 0}
6= 0, A
i∈[p],a∈[K] 1{Aia
6= 0}
.
b ⊆ supp(A). Thus,
We first show TFPP = 0. From the result of part (a), we know supp(A)
X
bia 6= 0} = 0,
1{Aia = 0, A
i∈[p],a∈[K]
which implies TFPP = 0.
In order to prove the result of TFNP, observe
X
X
1{Aia 6= 0} = |I| +
si .
(76)
i∈J
i∈[p],a∈[K]
b we partition [p] = I ∪J1 ∪J2 ∪J3 = I ∪(L1 ∪L2 )∪J2 ∪J3
with si = kAi. k0 for each j ∈ J. For given I,
with L1 = Ib∩J1 and L2 = J1 \L1 . Let us consider the set I ∪L1 first. Theorem 3 implies I ∪L1 = Ib
bb, we have
and Iba \ Ia ⊆ J1a . From the way we construct A
I
X
i∈I∪L1 ,a∈[K]
bia = 0} =
1{Aia 6= 0, A
X
i∈L1 ,a∈[K]
bia = 0}.
1{Aia 6= 0, A
Since the definition of J1 implies that, for any j ∈ J1a and a ∈ [K], |Aja | ≥ 1−4δ/ν and |Ajb | ≤ 4δ/ν,
for any b 6= a, this implies
X
X
bia = 0} =
1{Aia 6= 0, A
1{Aib 6= 0} = ti ,
b6=a
a∈[K]
for any i ∈ J1a ∩ L1 and a ∈ [K]. Thus, we have
X
X
bia = 0} =
ti .
1{Aia 6= 0, A
(77)
i∈L1
i∈I∪L1 ,a∈[K]
Next we consider the set L2 . On the event E, for any i ∈ J1a ∩ L2 , we have
bia | ≥ |Aia | − kA
b − Ak∞ ≥ 1 −
|A
bia =
Thus, A
6 0, which implies
X
i∈L2 ,a∈[K]
4δ
− 2µ > 0.
ν
bia = 0} ≤
1{Aia 6= 0, A
47
X
i∈L2
ti .
(78)
bJ ) which yields
Then we consider the set J2 . Part (a) gives supp(AJ2 ) = supp(A
2
X
bia = 0} = 0.
1{Aia 6= 0, A
(79)
i∈J2 ,a∈[K]
bja 6= 0
Finally, we consider the set J3 . By examining the proof of Part (a), it is easy to verify that A
if |Aja | ≥ (2µ) ∨ (4δ/ν), for any j ∈ J3 and a ∈ [K]. Thus,
X
X
bia = 0} ≤
1{Aia 6= 0, A
ti .
(80)
i∈J3
i∈J3 ,a∈[K]
At last, combining (76) - (80) gives
TFNP =
P
i∈[p],a∈[K] 1{Aia
P
bia = 0}
6= 0, A
i∈[p],a∈[K] 1{Aia
6= 0}
≤
P
j∈J1 ∪J3 tj
|I| +
P
j∈J
sj
.
b a ) and GFNP(G
b a ) by the following:
Proof of part (b). Similarly, we can express GFPP(G
ba ) =
GFPP(G
P
i∈[p] 1{Aia
P
bia 6= 0}
= 0, A
i∈[p] 1{Aia
= 0}
ba ) =
GFNP(G
,
P
i∈[p] 1{Aia
P
bia = 0}
6= 0, A
i∈[p] 1{Aia
6= 0}
.
b GFPP(G
b a ) = 0 follows immediately by noting that
For any given a ∈ [K],
ba |
|(Ga )c |
|(Ga )c ∩ G
b a ),
= PK
GFPP(G
0 = TFPP ≥ PK
c|
c|
|(G
)
|(G
)
b
b
b=1
b=1
b a ) = 0 if (Ga )c = 0. To show the expression of GFNP(G
b a ), by the
with the convention GFPP(G
definition of I and Theorem 3, we obtain
X
X
bia = 0} = 0,
1{Aia 6= 0, A
1{Aia 6= 0} = |Ia |.
i∈I
i∈I
The latter immediately implies
X
i∈[p]
1{Aia 6= 0} = |Ia | +
X
sai
i∈J
In addition, following the same arguments in the proof of part (b), we have
X
X
X
bia = 0} = 0,
bia = 0} =
1{Aia 6= 0, A
1{Aia 6= 0, A
i∈J2
i∈J1 ∪J3
tai .
i∈J1 ∪J3 \J1a
Combining these two concludes the proof.
Proofs of Remark 4.
PK
We briefly verify the first claim. It suffices to verify
a=1 |Aja | ≤ 1 which is equivalent with
48
b Recall (16), since β j is feasible, the optimality of βbj immediately gives
kβbj k1 ≤ 1, for any j ∈ J.
kβbj k1 ≤ 1.
To verify the expression of TFNP in the second claim, we assume tj = t and sj = s, for j ∈ J, and
|J1 | + |J3 | = α(|I| + |J3 |). Note that |I| + |J1 | + |J2 | + |J3 | = p implies |J1 | + |J3 | = αp/(1 + α). We
therefore obtain
TFNP ≤
=
t(|J1 | + |J3 |)
t
=
s|J
|+|I|
2
s(|J1 | + |J3 |) + s|J2 | + |I|
s+
· 1+α
α
p
t
by (1 + α)(|I| + |J2 |) = p ,
|I|+s|J
|
s + α1 · |I|+|J22|
as desired.
We verify the third claim. On the event E, when J2 = J, Remark 2 yields Ib = I, Ib = I and
b − Ck∞ ≤ δ and
Jb = J. After careful examination of the proof of Lemma 11, we derive that kC
maxj∈J kθbj − θk∞ ≤ δ, on the event E. Therefore, choosing λ = δ and µ = 3kΩk∞,1 δ proves the
claim, following the proof of Theorems 4 and 6 step by step.
Finally, we verify the fourth claim on the hard-threshold estimator βej for any j ∈ J. For simplicity,
we remove the super indices. Recall that, βe is defined coordinate-wisely by β̄a 1{|β̄a | > µ} with
µ = 5kΩk∞,1 δ 0 .
First, we show kβe − βk∞ ≤ 2µ. For any a ∈ [K] such that |β̄a | ≤ µ, we have
|βea − βa | = |βa | ≤ kβ̄ − βk∞ + |β̄a | ≤ 2µ,
e − Ak∞ ≤ 2µ where
while the same bound is obtained above for the case |β̄a | > µ. This proves kA
e combines A
bb and βej for each j ∈ J.
b To prove the same rate in Theorem 4 for A,
e it suffices to
A
I
j
show that Lemma 12 still holds for βe . Recall that, on the event E, we have kβ̄ − βk∞ ≤ µ. For
any βa = 0, we thus have |β̄a | ≤ kβ̄ − βk∞ ≤ µ, which implies βea = 0. This concludes the proof of
e
Theorem 4 for A.
e and we write (i, a) ∈ Sb if |A
eia | =
To show part (a) of Theorem 6, let Sb denote the support of A
6 0.
b
Let (i, a) ∈ S be arbitrary fixed and consider the following two cases:
b from Theorem 3 and the way we construct A
bb, we have |A
eia | = 1. Thus, |Aia | ≥
- If i ∈ I,
I
eia | − kA
e − Ak∞ ≥ 1 − 2µ > 0.
|A
b then |A
eia | = |βei | = |β̄ i | > µ. Therefore, |Aia | = |β i | ≥ |β̄ i | − kβ̄ i − β i k∞ > 0.
- If i ∈ J,
a
a
a
a
e ⊆ supp(A). To show supp(AJ ) ⊆ supp(A),
e for any (i, a) ∈
Thus, we have proved that supp(A)
2
e
e
supp(AJ2 ), by the definition of J2 , |Aia | > 2µ. Thus, |Aia | ≥ |Aia | − kA − Ak∞ ≥ 0. Therefore,
e
(i, a) ∈ supp(A).
e b) = sign(A b), since Lemma 10 guarantees sign(A
bia ) = sign(Aia ) for any (i, a) ∈
To show sign(A
S
S
b we focus on each (i, a) ∈ Sb and i ∈ J.
b Assuming A
eia = βei > 0, we know βei = β̄ i > µ.
Sb and i ∈ I,
a
a
a
eia > 0 with
Since Aia = βai ≥ β̄ai − kβ̄ i − β i k∞ > 0, we have proved that Aia > 0 for any A
49
b Since the same argument holds for any A
eia < 0, we conclude the proof of
(i, a) ∈ Sb and i ∈ J.
e
sign(ASb) = sign(ASb).
The same conclusion in part (b) and (c) of Theorem 6 holds for GFPP, GFNP, TFPP and
e as it shares the same property in part (a).
TFNP based on the hard-threshold estimator A,
B
Cross-validation illustration
We consider a simple case, when C is diagonal and the signed permutation matrix P is I, to illustrate our cross-validation method.
Example 1. Let C = diag(τ, τ, τ ), I = {1, 2}, {3, 4}, {5, 6} and
1
0
0
−1
0
0
0
1
0
0
1
0
,
A=
0 −1
0
0 −1
0
0.4 0.6 0
−0.5 0 0.4
∗
τ
0
T
AI CAI =
0
0
0
τ
∗
0
0
0
0
0
0
∗
τ
0
0
0
0
τ
∗
0
0
0
0
0
0
∗
τ
0
0
0
,
0
τ
∗
where we use ∗ to reflect the fact that our algorithm ignores the diagonal elements. For the true I
bI = AI ,
and I, we have A
b T
b (1) − AI CA
Σ
I
II
For
=
we obtain
F-off
≤
≤
b (1) − ΣII
Σ
II
b (1) − ΣII
Σ
II
b (1) − Σij
max Σ
ij
i6=j
F-off
F-off
b T − ΣII
+ AI CA
I
F-off
p
b
+ |I|(|I| − 1) · kC − Ck∞ .
(2)
b
∨ max Σij − Σij ,
i6=j
1
b (1)
b T
CV (I) = q
ΣII − AI CAI
|I| |I| − 1
F-off
≤ 2.
Suppose that Ib = {1, 2}, {3, 5}, {4, 6} , so Ib = I, yet Ib =
6 I, we would have
∗ τb1 0 0 0 0
τb1 ∗ 0 0 0 0
0 0 ∗ 0 τb 0
2
bbC
bA
bT =
A
,
I
Ib
0 0 0 ∗ 0 τb3
0 0 τb2 0 ∗ 0
0 0 0 τb3 0 ∗
∗
∆τ1 0
∆τ1
∗
0
0
0
∗
bbC
bA
bT − Σ bb =
A
I
II
Ib
0
0
−τ
0
0
τb2
0
0
0
50
0
0
−τ
∗
0
τb3
0
0
τb2
0
∗
−τ
0
0
0
,
τb3
−τ
∗
baa from (11) for each a ∈ [K].
b Thus, the
Here ∆τa = τba − τa , using estimates τba defined in lieu of C
cross-validation criterion in (39) would satisfy
1
b b bT
q
AIbC AIb − ΣIbIb
b |I|
b −1
|I|
s
τ2
4τ 2 + 2b
τ22 + 2b
≥
3 − 2.
b |I|
b −1
|I|
b ≥
CV (I)
1
b (1)
−q
ΣIbIb − ΣIbIb
F-off
b |I|
b −1
|I|
F-off
From noting that |b
τa − τ | ≤ , for a = 2, 3, it gives
r
2
2
b ≥ 4τ − 4τ + 2 − 2 > 2 ≥ CV (I),
CV (I)
15
for τ ≥ 9. We conclude in this example, with Ib = I, incorrectly specifying I will induce a large
b 6= K and Ib =
loss. It is easily verified that this is also the case when Ib = I but K
6 I.
b For instance,
On the other hand, suppose we mistakenly included some non-pure variable in I.
b
suppose we found I = {1, 2}, {3, 4}, {5, 6, 7} . Then we would have
∗
τ
0
0
0 0 0.4τ
∗ τb1 0 0 0 0 0
∗
0
0
0 0 −0.4τ
τ
τb1 ∗ 0 0 0 0 0
0
0 0 ∗ τb2 0 0 0
0
∗
τ
0 0 0.6τ
bb0 C
bA
bT = 0 0 τb2 ∗ 0 0 0 .
, A
ΣIb0 Ib0 =
0
0
τ
∗
0
0
0.6τ
0
I
Ib
0
0 0 0 0 ∗ τb τb
0
0
0
∗ τ
0
3
3
0
0
0
τ ∗
0
0
0 0 0 0 τb3 ∗ τb3
0.4τ −0.4τ 0.6τ 0.6τ 0 0
∗
0 0 0 0 τb3 τb3 ∗
We thus have
bb0 C
bA
bT − Σ b0 b0
A
I I
I
Ib0
∗
∆τ1
0
=
0
0
0
−0.4τ
∆τ1
∗
0
0
0
0
0.4τ
0
0
∗
∆τ2
0
0
−0.6τ
0
0
∆τ2
∗
0
0
−0.6τ
0
0
−0.4τ
0
0
0.4τ
0
0
−0.6τ
0
0
−0.6τ
∗
∆τ3
τb3
∆τ3
∗
τb3
τb3
τb3
∗
and, by similar arguments, for τ ≥ 12, we find
r
4b
τ32 + 4 × 0.36τ 2 + 4 × 0.16τ 2
0
b
CV (I ) ≥
− 2 > 2.
42
Thus, the cross-validation loss in this example will be large even if only one non-pure variable is
mistakenly classified as pure variable. In rare cases, the cross-validation criterion might miss a very
small subset of I but this can be rectified in our later estimation of AJ .
51
C
C.1
Additional Simulation Results
Related work on the estimation of A
As we explained in Section 4.4, the existing procedures for estimating A in (1) are developed
for models satisfying identifiability conditions different than our (i)-(iii). Specifically, Bai and
Li [2012] propose to first optimize, via EM, a quasi-likelihood objective under the identifiability
conditions (a) C = IK and (b) AT Γ−1 A is diagonal. The major advantage of this setting is that the
computationally demanding EM algorithm only needs to determine A and Γ as C = IK is given.
b with the property
The EM algorithm, however, is only guaranteed to find stationary point B
bT B
b is diagonal. In the context of this problem, as the authors note, the EM algorithm
that B
requires a delicate initialization and is computationally demanding, even if only one of K, n and
p is moderately large. Next, the authors propose to link this estimator with an estimator of a
model no longer satisfying (a) and (b) as identifiability conditions, but satisfying instead (1) C is
an arbitrary positive definite matrix; (2) There exists a known set S of K pure variables, with only
one pure variable per latent factor allowed. No further sparsity conditions on A are imposed. To
bB
b T . This yields
estimate A under (2), they suggest to solve for A and C the equation ACAT = B
bS may not be invertible, and
e=B
bB
b −1 of A. However, when K is relatively large, B
the estimator A
S
−1
eS c is not sparse
b = IK , the submatrix A
eS = B
bS B
the estimator may not exist. Finally, although A
S
e
in general. One possibility is to threshold A, but it is unclear how to choose the correct threshold
level, for the following reason. Although the authors establish the asymptotic limit of the MLE of
A under (1) and (2), the estimator of A explained above is not guaranteed to be the MLE in this
model: if it exists, it is a transformation of a stationary point that estimates parameters under the
model specifications (a) and (b), different from (1) and (2). The immediate practical implication
b around A under (1) and (2) is not known, which makes the thresholding
is that the variation of A
b difficult to assess. For all these reasons, we cannot compare numerically our estimation
level of A
procedure with the procedure proposed in Bai and Li [2012], even in the (unrealistic) case when
the pure variable set is known.
C.2
LOVE for non-overlapping cluster estimation
In applications, one may not have prior information on whether the clusters may overlap or not.
Thus, one would prefer a clustering method that works well in both overlapping and non-overlapping
scenarios. In the previous section, we have demonstrated that LOVE outperforms the existing
clustering methods if data are generated from a model that yields variable clusters with overlaps.
In this section, we study the numerical performance of the proposed method under non-overlapping
data generating schemes.
To generate data with non-overlapping clusters, we set the number of variables in each cluster
to be 20. We generate the diagonal elements of C from the uniform distribution in [1, 2] and
use the same method as in Section 5.2 to generate the off-diagonal elements. The variance σj2
of the error Ej is generated from the uniform distribution in [3, 4]. In Table 3, we compare the
sensitivity and specificity of the proposed method with the CORD estimator [Bunea et al., 2016a]
under non-overlapping scenarios, where the sensitivity and specificity are defined in (81). The
52
CORD estimator can be viewed as a benchmark method for variable clustering without overlaps
and is shown to outperform K-means and hierarchical clustering, via an extensive numerical study
presented in Bunea et al. [2016a]. For this reason, we only focus on the comparison between LOVE
and CORD. From Table 3, we see that for small p (i.e., p = 100) the performance of LOVE
is only slightly worse than CORD. As p increases, the specificity of LOVE and that of CORD
remain close to 1, but LOVE yields in fact higher sensitivity than CORD when n = 300. This
confirms that the performance of the proposed method is comparable to the benchmark method
under non-overlapping scenarios. Of course, LOVE is much more flexible as it can detect possible
overlaps.
p
n = 300
LOVE
100
500
1000
n = 500
CORD
LOVE
CORD
SN
SP
SN
SP
SN
SP
SN
SP
0.87
(0.09)
0.86
(0.05)
0.84
(0.05)
0.90
(0.10)
0.98
(0.01)
0.97
(0.02)
0.92
(0.05)
0.82
(0.03)
0.78
(0.03)
0.98
(0.02)
0.98
(0.01)
0.97
(0.01)
0.93
(0.05)
0.87
(0.04)
0.87
(0.04)
0.97
(0.03)
0.99
(0.00)
1.00
(0.00)
0.98
(0.02)
0.94
(0.02)
0.90
(0.02)
1.00
(0.01)
1.00
(0.01)
1.00
(0.01)
Table 3: Sensitivity (SN) and specificity (SP) of the proposed method (LOVE) and CORD under
non-overlapping scenarios. Numbers in parentheses are the simulation standard errors.
C.3
Comparison with other overlapping clustering algorithms
We adopt the same data generating procedure except that we set K = 10 and the negative entries
of A are replaced by their absolute values, since existing overlapping clustering algorithms typically
return an estimator of A with positive entries. We compare the proposed method with the following
overlapping clustering algorithms: fuzzy K-means, and fuzzy K-medoids [Krishnapuram et al.,
2001], the latter being more robust to noise and outliers. We describe the methods briefly in what
follows. Both of them aim to estimate a degree of membership matrix M ∈ Rp×K by minimizing the
ej = (X1j , ..., Xnj ),
average within-cluster L2 or L1 distances [Bezdek, 2013]. Specifically, denote X
n
e = {X
e1 , ..., X
ep }. Let W = {w1 , ..., wK }, where wk ∈ R , be a subset of X
e with K elements.
and X
The fuzzy algorithms aim to find the set W such that J(W ) defined as
J(W ) =
p X
K
X
j=1 k=1
ej , wk ),
Mjk r(X
is minimized. Here, Mjk > 0 can be interpreted as the degree of membership matrix which is a
ej , wk ). Some commonly used expressions of Mjk are shown by Krishnapuram
known function of r(X
53
ej , wk ) is a measure of dissimilarity between X
ej and wk . For instance,
et al. [2001]. In addition, r(X
2
if r(x, θ) = kx−θk2 , this corresponds to the fuzzy K-means. Similarly, the fuzzy K-medoids is given
e is computationally infeasible,
by r(x, θ) = kx − θk1 . Since searching over all possible subsets of X
an approximate algorithm for minimizing J(W ) is proposed by Krishnapuram et al. [2001], we refer
to their original paper for further details.
Their degree of membership matrix M plays the same role as our allocation matrix A, but is
typically non-sparse. In order to construct overlapping clusters based on M one needs to specify a
cut-off value v and assign variable j to cluster k if Mjk > v. Moreover, the number of clusters K
is a required input of the algorithm. In the simulations presented in this section we set K = 10 for
these two methods, which have been implemented by the functions FKM and FKM.med in R.
We compare their performance with our proposed method LOVE. We emphasize that our
method does not require the specification of K and that the tuning parameters are chosen in
a data adaptive fashion, as explained in the previous sections. We follow the pairwise approach of
Wiwie et al. [2015] for this comparison. Recall that G = (G1 , ..., GK ) denotes the true overlapping
b 1 , ..., G
b ) to denote clusters computed from an
clusters. For notational simplicity, we use Gb = (G
K̄
b to be different from K. For
algorithm. Since LOVE estimates the number of clusters, we allow K
any pair 1 ≤ j < k ≤ p, define
n
o
b b for some 1 ≤ a ≤ K and 1 ≤ b ≤ K
b ,
T Pjk = 1 if j, k ∈ Ga and j, k ∈ G
n
o
b b for any 1 ≤ a ≤ K and 1 ≤ b ≤ K
b ,
T Njk = 1 if j, k ∈
/ Ga and j, k ∈
/G
n
o
b b for some 1 ≤ b ≤ K
b ,
F Pjk = 1 if j, k ∈
/ Ga for any 1 ≤ a ≤ K and j, k ∈ G
n
o
b b for any 1 ≤ b ≤ K
b .
F Njk = 1 if j, k ∈ Ga for some 1 ≤ a ≤ K and j, k ∈
/G
and we define
TP =
X
1≤j<k≤p
T Pjk , T N =
X
T Njk , F P =
1≤j<k≤p
X
1≤j<k≤p
F Pjk , F N =
X
F Njk .
1≤j<k≤p
We use sensitivity (SN) and specificity (SP) to evaluate the performance of different methods,
where
TN
TP
SP =
, and SN =
.
(81)
TN + FP
TP + FN
Recall that for the fuzzy methods, variable j belongs to cluster k if the estimated membership
matrix Mjk is beyond a cut-off v, i.e., Mjk > v. We search for the optimal cut-off v in a grid
{0.01, 0.1, ..., 0.3} such that SP + SN is maximized. The corresponding sensitivity and specificity
for LOVE, fuzzy K-means (F-Kmeans) and fuzzy K-medoids (F-Kmed) are shown in Figure 3. To
save space, we only present the results for p = 500 since the other scenarios illustrate the same
patterns. The following findings are observed. First, the F-Kmeans is superior to F-Kmed in most
scenarios in terms of both sensitivity and specificity. Second, LOVE clearly outperforms these
two existing methods and its specificity and sensitivity are very close to 1, which implies that our
method leads to very few false positives and false negatives. The conclusions hold with n from
300 to 1000. Moreover, we reiterate that the true value K = 10 is used as input in the competing
54
●
●
●
1.0
●
●
●
●
●
●
●
0.8
0.8
1.0
methods, whereas it is estimated from the data in LOVE. This illustrates the net advantage of the
proposed method over the existing overlapping clustering methods, for data generated from Model
(1).
●
●
●
●
●
●
●
●
●
0.4
500
700
0.6
0.0
0.0
LOVE
F−Kmeans
F−Kmed
300
0.4
●
0.2
Specificity
0.6
●
●
0.2
Sensitivity
●
●
900
300
sample size
500
700
900
sample size
Figure 3: Plot of specificity and sensitivity for LOVE, fuzzy K-means (F-Kmeans), and fuzzy
K-medoids (F-Kmed) when p = 500.
55
| 10math.ST
|
MACS: AN AGENT-BASED MEMETIC MULTIOBJECTIVE
arXiv:1206.1305v1 [cs.CE] 6 Jun 2012
OPTIMIZATION ALGORITHM APPLIED TO SPACE TRAJECTORY
DESIGN
Massimiliano Vasile – Federico Zuiani
Space Advanced Research Team, University of Strathclyde
James Weir Building 75 Montrose Street G1 1XJ, Glasgow, United Kingdom
[email protected], [email protected]
ABSTRACT
This paper presents an algorithm for multiobjective optimization that blends together a
number of heuristics. A population of agents combines heuristics that aim at exploring
the search space both globally and in a neighborhood of each agent. These heuristics
are complemented with a combination of a local and global archive. The novel agentbased algorithm is tested at first on a set of standard problems and then on three specific
problems in space trajectory design. Its performance is compared against a number of
state-of-the-art multiobjective optimisation algorithms that use the Pareto dominance
as selection criterion: NSGA-II, PAES, MOPSO, MTS. The results demonstrate that
the agent-based search can identify parts of the Pareto set that the other algorithms
were not able to capture. Furthermore, convergence is statistically better although the
variance of the results is in some cases higher.
1
KEYWORDS
Multiobjective optimisation; trajectory optimisation; memetic algorithms; multiagent
systems.
1. INTRODUCTION
The design of a space mission steps through different phases of increasing complexity.
In the first phase, a trade-off analysis of several options is required. The trade-off
analysis compares and contrasts design solutions according to different criteria and
aims at selecting one or two options that satisfy mission requirements. In mathematical
terms, the problem can be formulated as a multiobjective optimization problem.
As part of the trade-off analysis, multiple transfer trajectories to the destination need to
be designed. Each transfer should be optimal with respect to a number of criteria. The
solution of the associated multiobjective optimization problem, has been addressed, by
many authors, with evolutionary techniques. Coverstone et al. [1] proposed the use
of multiobjective genetic algorithms for the optimal design of low-thrust trajectories.
Dachwald et al. proposed the combination of a neurocontroller and of a multiobjective
evolutionary algorithm for the design of low-thrust trajectories [2]. In 2005 a study
by Lee et al. [3] proposed the use of a Lyapunov controller with a multiobjective evolutionary algorithm for the design of low-thrust spirals. More recently, Schütze et al.
proposed some innovative techniques to solve multiobjective optimization problems
for multi-gravity low-thrust trajectories. Two of the interesting aspects of the work
of Schütze et al. are the archiving of ǫ- and ∆-approximated solutions, to the known
best Pareto front [4], and the deterministic pre-pruning of the search space [5]. In
2009, Delnitz et al. [6] proposed the use of multiobjective subdivision techniques for
the design of low-thrust transfers to the halo orbits around the L2 libration point in the
Earth-Moon system. Minisci et al. presented an interesting comparison between an
EDA-based algorithm, called MOPED, and NSGA-II on some constrained and unconstrained multi-impulse orbital transfer problems [7].
In this paper, a hybrid population-based approach that blends a number of heuristics is
2
proposed. In particular, the search for Pareto optimal solutions is carried out globally
by a population of agents implementing classical social heuristics and more locally by
a subpopulation implementing a number of individualistic actions. The reconstruction
of the set of Pareto optimal solutions is handled through two archives: a local and a
global one.
The individualistic actions presented in this paper are devised to allow each agent to
independently converge to the Pareto optimal set. Thus creating its own partial representation of the Pareto front. Therefore, they can be regarded as memetic mechanisms
associated to a single individual. It will be shown that individualistic actions significantly improve the performance of the algorithm.
The algorithm proposed in this paper is an extension of the Multi-Agent Collaborative
Search (MACS), initially proposed in [8, 9], to the solution of multiobjective optimisation problems. Such an extension required the modification of the selection criterion,
for both global and local moves, to handle Pareto dominance and the inclusion of new
heuristics to allow the agents to move toward and along the Pareto front. As part of
these new heuristics, this papers introduces a dual archiving mechanism for the management of locally and globally Pareto optimal solutions and an attraction mechanisms
that improves the convergence of the population.
The new algorithm is here applied to a set of known standard test cases and to three
space mission design problems. The space mission design cases in this paper consider
spacecraft equipped with a chemical engine and performing a multi-impulse transfer.
Although these cases are different from some of the above-mentioned examples, that
consider a low-thrust propulsion system, nonetheless the size and complexity of the
search space is comparable. Furthermore, it provides a first test benchmark for multiimpulsive problems that have been extensively studied in the single objective case but
for which only few comparative studies exist in the multiobjective case [7].
The paper is organised as follows: section two contains the general formulation of the
problem, the third section starts with a general introduction to the multi-agent collaborative search algorithm and heuristics before going into some of the implementation
details. Section four contains a set of comparative tests that demonstrates the effec-
3
tiveness of the heuristics implemented in MACS. The section briefly introduces the
algorithms against which MACS is compared and the two test benchmarks that are
used in the numerical experiments. It then defines the performance metrics and ends
with the results of the comparison.
2. PROBLEM FORMULATION
A general problem in multiobjective optimization is to find the feasible set of solutions
that satisfies the following problem:
min f (x)
(1)
x∈D
where D is a hyperrectangle defined as D = xj | xj ∈ [blj buj ] ⊆ R, j = 1, ..., n
and f is the vector function:
f : D → Rm ,
f (x) = [f1 (x), f2 (x), ..., fm (x)]T
(2)
The optimality of a particular solution is defined through the concept of dominance:
with reference to problem (1), a vector y ∈ D is dominated by a vector x ∈ D if
fj (x) < fj (y) for all j = 1, ..., m. The relation x ≺ y states that x dominates y.
Starting from the concept of dominance, it is possible to associate, to each solution in
a set, the scalar dominance index:
Id (xj ) = |{i | i ∧ j ∈ Np ∧ xi ≺ xj }|
(3)
where the symbol |.| is used to denote the cardinality of a set and Np is the set of the
indices of all the solutions. All non-dominated and feasible solutions form the set:
X = {x ∈ D | Id (x) = 0}
(4)
Therefore, the solution of problem (1) translates into finding the elements of X. If X
is made of a collection of compact sets of finite measure in Rn , then once an element
4
of X is identified it makes sense to explore its neighborhood to look for other elements
of X. On the other hand, the set of non dominated solutions can be disconnected and
its elements can form islands in D. Hence, restarting the search process in unexplored
regions of D can increase the collection of elements of X.
The set X is the Pareto set and the corresponding image in criteria space is the Pareto
front. It is clear that in D there can be more than one Xl containing solutions that are
locally non-dominated, or locally Pareto optimal. The interest is, however, to find the
set Xg that contains globally non-dominated, or globally Pareto optimal, solutions.
3. MULTIAGENT COLLABORATIVE SEARCH
The key motivation behind the development of multi-agent collaborative search was to
combine local and global search in a coordinated way such that local convergence is
improved while retaining global exploration [9]. This combination of local and global
search is achieved by endowing a set of agents with a repertoire of actions producing
either the sampling of the whole search space or the exploration of a neighborhood
of each agent. More precisely, in the following, global exploration moves will be
called collaborative actions while local moves will be called individualistic actions.
Note that not all the individualistic actions, described in this paper, aim at exploring
a neighborhood of each agent, though. The algorithm presented in this paper is a
modification of MACS to tackle multiobjective optimization problems. In this section,
the key heuristics underneath MACS will be described together with their modification
to handle problem (1) and reconstruct X.
3.1. GENERAL ALGORITHM DESCRIPTION
A population P0 of npop virtual agents, one for each solution vector xi , with i =
1, ..., npop , is deployed in D. The population evolves through a number of generations.
At every generation k, the dominance index (3) of each agent xi,k in the population
Pk is evaluated. The agents with dominance index Id = 0 form a set Xk of nondominated solutions. Hence, problem (1) translates into finding a series of sets Xk
such that Xk → Xg for k → kmax with kmax possibly a finite number.
5
The position of each agent in D is updated through a number of heuristics. Some
are called collaborative actions because are derived from the interaction of at least two
agents and involve the entire population at one time. The general collaborative heuristic
can be expressed in the following form:
xk = xk + S(xk + uk )uk
(5)
where uk depends on the other agents in the population and S is a selection function
which yields 0 if the candidate point xk + uk is not selected or 1 if it is selected (see
Section 3.2). In this implementation a candidate point is selected if its dominance index
is better or equal than the one of xk . A first restart mechanism is then implemented to
avoid crowding. This restart mechanism is taken from [9] and prevents the agents from
overlapping or getting too close. It is governed by the crowding factor wc that defines
the minimum acceptable normalized distance between agents. Note that, this heuristic
increases the uniform sampling rate of D, when activated, thus favoring exploration.
On the other hand, by setting wc small the agents are more directed towards local
convergence.
After all the collaborative and restart actions have been implemented, the resulting
updated population Pk is ranked according to Id and split in two subpopulations: Pku
and Pkl . The agents in each subpopulation implement sets of, so called, individualistic
actions to collect samples of the surrounding space and to modify their current location.
In particular, the last npop − fe npop agents belong to Pku and implement heuristics that
can be expressed in a form similar to Eq. (5) but with uk that depends only on xk .
The remaining fe npop agents belong to Pkl and implement a mix of actions that aim
at either improving their location or exploring the neighborhood Nρ (xi,k ), with i =
1, ..., fe npop . Nρ (xi,k ) is a hyperectangle centered in xi,k . The intersection Nρ (xi,k )∩
D represents the local region around agent xi,k that one wants to explore. The size of
Nρ (xi,k ) is specified by the value ρ(xi,k ): the ith edge of Nρ has length 2ρ(xi,k ) max{buj −
xi,k [j], xi,k [j] − blj }.
The agents in Pkl generate a number of perturbed solutions ys for each xi,k , with s =
6
1, ..., smax . These solutions are collected in a local archive Al and a dominance index
is computed for all the elements in Al . If at least one element ys ∈ Al has Id = 0
then xi,k ← ys . If multiple elements of Al have Id = 0, then the one with the
largest variation, with respect to xi,k , in criteria space is taken. Figure 1(a) shows
three agents (circles) with a set of locally generated samples (stars) in their respective
neighborhoods (dashed square). The arrows indicate the direction of motion of each
agent. The figures shows also the local archive for the first agent Al1 and the target
global archive Xg .
The ρ value associated to an agent is updated at each iteration according to the rule
devised in [9]. Furthermore, a value s(xi,k ) is associated to each agent xi,k to specify
the number of samples allocated to the exploration of Nρ (xi,k ). This value is updated
at each iteration according to the rule devised in [9].
The adaptation of ρ is introduced to allow the agents to self-adjust the neighborhood
removing the need to set a priori the appropriate size of Nρ (xi,k ). The consequence
of this adaptation is an intensification of the local search by some agents while the
others are still exploring. In this respect, MACS works opposite to Variable Neighborhood Search heuristics, where the neighborhood is adapted to improve global exploration, and differently than Basin Hopping heuristics in which the neighborhood is
fixed. Similarly, the adaptation of s(xi,k ) avoids setting a priori an arbitrary number of
individualistic moves and has the effect of avoiding an excessive sampling of Nρ (xi,k )
when ρ is small. The value of s(xi,k ) is initialized to the maximum number of allowable individualistic moves smax . The value of smax is here set equal to the number
of dimensions n. This choice is motivated by the fact that a gradient-based method
would evaluate the function a minimum of n times to compute an approximation of
the gradient with finite differences. Note that the set of individualistic actions allows
the agents to independently move towards and within the set X, although no specific
mechanism is defined as in [10]. On the other hand the mechanisms proposed in [10]
could further improve the local search and will be the subject of a future investigation.
All the elements of Xk found during one generation are stored in a global archive Ag .
The elements in Ag are a selection of the elements collected in all the local archives.
7
Figure 1: Illustration of the a) local moves and archive and b) global moves and
archive.
Figure 1(b) illustrates three agents performing two social actions that yield two samples (black dots). The two samples together with the non-dominated solutions coming
from the local archive form the global archive. The archive Ag is used to implement an
attraction mechanism that improves the convergence of the worst agents (see Sections
3.4.1). During the global archiving process a second restart mechanism that reinitializes a portion of the population (bubble restart) is implemented. Even this second
restart mechanism is taken from [9] and avoids that, if ρ collapses to zero, the agent
keeps on sampling a null neighborhood.
Note that, within the MACS framework, other strategies can be assigned to the agents
to evaluate their moves in the case of multiple objective functions, for example a decomposition technique [11]. However, in this paper we develop the algorithm based
only on the use of the dominance index.
3.2. COLLABORATIVE ACTIONS
Collaborative actions define operations through which information is exchanged between pairs of agents. Consider a pair of agents x1 and x2 , with x1 ≺ x2 . One of the
two agents is selected at random in the worst half of the current population (from the
point of view of the property Id ), while the other is selected at random from the whole
population. Then, three different actions are performed. Two of them are defined by
adding to x1 a step uk defined as follows:
uk = αrt (x2 − x1 ),
(6)
and corresponding to: extrapolation on the side of x1 (α = −1, t = 1), with the further
constraint that the result must belong to the domain D (i.e., if the step uk leads out of
D, its size is reduced until we get back to D); interpolation (α = 1), where a random
point between x1 and x2 is sampled. In the latter case, the shape parameter t is defined
8
Algorithm 1 Main MACS algorithm
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
Initialize a population P0 of npop agents in D, k = 0, number of function evaluations neval = 0, maximum number of function evaluations Ne , crowding factor
wc
for all i = 1, ..., npop do
xi,k = xi,k + S(xi,k + ui,k )ui,k
end for
Rank solutions in Pk according to Id
Re-initialize crowded agents according to the single agent restart mechanism
for all i = fe npop , ..., npop do
Generate np mutated copies of xi,k .
Evaluate the dominance of each mutated copy yp against xi,k , with p =
1, ..., np
if ∃p|yp ≺ xi,k then
p̄ = arg maxp kyp − xi,k k
xi,k ← yp̄
end if
end for
for all i = 1, ..., fe npop do
Generate s < smax individual actions us such that ys = xi,k + us
if ∃s|ys ≺ xi,k then
s̄ = arg maxs kys − xi,k k
xi,k ← ys̄
end if
Store candidate elements ys in the local archive Al
Update ρ(xi,k ) and s(xi,k )
end for
S S
S
Form Pk = Pkl Pku and Âg = Ag Al Pk
Compute Id of all the elements in Âg
Ag = {x|x ∈ Âg ∧ Id (x) = 0 ∧ kx − xAg k > wc }
Re-initialize crowded agents in Pk according to the second restart mechanism
Compute attraction component to Ag for all xi,k ∈ Pk \ Xk
k =k+1
Termination Unless neval > Ne , GoTo Step 2
as follows:
t = 0.75
s(x1 ) − s(x2 )
+ 1.25
smax
(7)
The rationale behind this definition is that we are favoring moves which are closer to
the agent with a higher fitness value if the two agents have the same s value, while in
the case where the agent with highest fitness value has a s value much lower than that
of the other agent, we try to move away from it because a small s value indicates that
improvements close to the agent are difficult to detect.
9
The third operation is the recombination operator, a single-point crossover, where,
given the two agents: we randomly select a component j; split the two agents into
two parts, one from component 1 to component j and the other from component j + 1
to component n; and then we combine the two parts of each of the agents in order
to generate two new solutions. The three operations give rise to four new samples,
denoted by y1 , y2 , y3 , y4 . Then, Id is computed for the set x2 , y1 , y2 , y3 , y4 . The
element with Id = 0 becomes the new location of x2 in D.
3.3. INDIVIDUALISTIC ACTIONS
Once the collaborative actions have been implemented, each agent in Pku is mutated
a number of times: the lower the ranking the higher the number of mutations. The
mutation mechanisms is not different from the single objective case but the selection is
modified to use the dominance index rather than the objective values.
A mutation is simply a random vector uk such that xi,k + uk ∈ D. All the mutated
solution vectors are then compared to xi,k , i.e. Id is computed for the set made of
the mutated solutions and xi,k . If at least one element yp of the set has Id = 0 then
xi,k ← yp . If multiple elements of the set have Id = 0, then the one with the largest
variation, with respect to xi,k , in criteria space is taken.
Each agent in Pkl performs at most smax of the following individualistic actions: inertia, differential, random with line search. The overall procedure is summarized in
Algorithm 2 and each action is described in detail in the following subsections.
3.3.1. Inertia
If agent i has improved from generation k − 1 to generation k, then it follows the
direction of the improvement (possibly until it reaches the border of D), i.e., it performs
the following step:
ys = xi,k + λ̄∆I
where λ̄ = min{1, max{λ : ys ∈ D}} and ∆I = (xi,k − xi,k−1 ).
10
(8)
3.3.2. Differential
This step is inspired by Differential Evolution [12]. It is defined as follows: let xi1 ,k , xi2 ,k , xi3 ,k
be three randomly selected agents; then
ys = xi,k + e [xi1 ,k + F (xi3 ,k − xi2 ,k )]
(9)
with e a vector containing a random number of 0 and 1 (the product has to be intended
componentwise) with probability 0.8 and F = 0.8 in this implementation. For every
component ys [j] of ys that is outside the boundaries defining D then ys [j] = r(buj −
blj ) + blj , with r ∈ U (0, 1). Note that, although this action involves more than one
agent, its outcome is only compared to the other outcomes coming from the actions
performed by agent xi,k and therefore it is considered individualistic.
3.3.3. Random with Line Search
This move realizes a local exploration of the neighborhood N (xi,k ). It generates a first
random sample ys ∈ N (xi,k ). Then if ys is not an improvement, it generates a second
sample ys+1 by extrapolating on the side of the better one between ys and xi,k :
ys+1 = xi,k + λ̄ α2 rt (ys − xi,k ) + α1 (ys − xi,k )
(10)
with λ̄ = min{1, max{λ : ys+1 ∈ D}} and where α1 , α2 ∈ {−1, 0, 1}, r ∈
U (0, 1) and t is a shaping parameter which controls the magnitude of the displacement.
Here we use the parameter values α1 = 0, α2 = −1, t = 1, which corresponds to
extrapolation on the side of xi,k , and α1 = α2 = 1, t = 1, which corresponds to
extrapolation on the side of ys .
The outcome of the extrapolation is used to construct a second order one-dimensional
model of Id . The second order model is given by the quadratic function fl (σ) =
a1 σ 2 +a2 σ+a3 where σ is a coordinate along the ys+1 −ys direction. The coefficients
a1 , a2 and a3 are computed so that fl interpolates the values Id (ys ), Id (ys+1 ) and
Id (xi,k ). In particular, for σ = 0 fl = Id (ys ), for σ = 1 fl = Id (ys+1 ) and for
11
σ = (xi,k − ys )/kxi,k − ys k fl = Id (xi,k ). Then, a new sample ys+2 is taken at the
minimum of the second-order model along the ys+1 − ys direction.
Algorithm 2 Individual Actions in Pkl
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
s = 1, stop = 0
if xi,k ≺ xi,k−1 then
ys = xi,k + λ̄(xi,k − xi,k−1 )
with λ̄ = min{1, max{λ : ys ∈ D}}.
end if
if xi,k ≺ ys then
s=s+1
ys = e[xi,k − (xi1 ,k + (xi3 ,k − xi2 ,k ))]
∀j|ys (j) ∈
/ D, ys (j) = r(bu (j) − bl (j)) + bl (j),
with j = 1, ..., n and r ∈ U (0, 1)
else stop = 1
end if
if xi,k ≺ ys then
s=s+1
Generate ys ∈ Nρ (xi,k ).
Compute ys+1 = xi,k + λ̄rt (xi,k − ys )
with λ̄ = min{1, max{λ : ys ∈ D}}
and r ∈ U (0, 1).
Compute ys+2 = λ̄σmin (ys+1 − ys )/k(ys+1 − ys )k,
with σmin = arg minσ {a1 σ 2 + a2 σ + Id (ys )},
and λ̄ = min{1, max{λ : ys+2 ∈ D}}.
s=s+2
else stop = 1
end if
Termination Unless s > smax or stop = 1 , GoTo Step 4
The position of xi,k in D is then updated with the ys that has Id = 0 and the longest
vector difference in the criteria space with respect to xi,k . The displaced vectors ys
generated by the agents in Pkl are not discarded but contribute to a local archive Al ,
one for each agent, except for the one selected to update the location of xi,k . In order
to rank the ys , the following modified dominance index is used:
Iˆd (xi,k ) = j | fj (ys ) = fj (xi,k ) κ+
j | fj (ys ) > fj (xi,k )
(11)
where κ is equal to one if there is at least one component of f(xi,k ) = [f1 , f2 , ..., fm ]T
which is better than the corresponding component of f(ys ), and is equal to zero other-
12
wise.
Now, if for the sth outcome, the dominance index in Eq. (11) is not zero but is lower
than the number of components of the objective vector, then the agent xi,k is only
partially dominating the sth outcome. Among all the partially dominated outcomes
with the same dominance index we select the one that satisfies the condition:
min
s
f(xi,k )) − f(ys ) , e
where e is the unit vector of dimension m, e =
[1,1,1,...,1]T
√
m
(12)
. All the non-dominated and
selected partially dominated solutions form the local archive Al .
3.4. THE LOCAL AND GLOBAL ARCHIVES AL AND AG
Since the outcomes of one agent could dominate other agents or the outcomes of other
agents, at the end of each generation, every Al and the whole population Pk are added
to the current global archive Ag . The global Ag contains Xk , the current best estimate of Xg . The dominance index in Eq.(3) is then computed for all the elements
S
S
in Âg = Ag l Al Pk and only the non-dominated ones with crowding distance
kxi,k − xAg k > wc are preserved (where xAg is an element of Ag ).
3.4.1. Attraction
The archive Ag is used to direct the movements of those agents that are outside Xk .
All agents, for which Id 6= 0 at step k, are assigned the position of the elements in Ag
and their inertia component is recomputed as:
∆I = r(xAg − xi,k )
(13)
More precisely, the elements in the archive are ranked according to their reciprocal distance or crowding factor. Then, every agent for which Id 6= 0 picks the least crowded
element xAg not already picked by any other agent.
13
3.5. STOPPING RULE
The search is stopped when a prefixed number Ne of function evaluations is reached.
At termination of the algorithm the whole final population is inserted into the archive
Ag .
4. COMPARATIVE TESTS
The proposed optimization approach was implemented in a software code, in Matlab,
called MACS. In previous works [8,9], MACS was tested on single objective optimization problems related to space trajectory design, showing good performances. In this
work, MACS was tested at first on a number of standard problems, found in literature,
and then on three typical space trajectory optimization problems.
This paper extends the results presented in Vasile and Zuiani (2010) [13] by adding
a broader suite of algorithms for multiobjective optimization to the comparison and a
different formulation of the performance metrics.
4.1. TESTED ALGORITHMS
MACS was compared against a number of state-of-the-art algorithms for multiobjective
optimization. For this analysis it was decided to take the basic version of the algorithms
that is available online. Further developments of the basic algorithms have not been
considered in this comparison and will be included in future works. Note that, all the
algorithms selected for this comparative analysis use Pareto dominance as selection
criterion.
The tested algorithms are: NSGA-II [14], MOPSO [15], PAES [16] and MTS [17]. A
short description of each algorithm with their basic parameters follows.
4.1.1. NSGA-II
The Non-Dominated Sorting Genetic Algorithm (NSGA-II) is a genetic algorithm
which uses the concept of dominance class (or depth) to rank the population. A crowding factor is then used to rank the individuals within each dominance class. Optimiza-
14
tion starts from a randomly generated initial population. The individuals in the population are sorted according to their level of Pareto dominance with respect to other individuals. To be more precise, a fitness value equal to 1 is assigned to the non-dominated
individuals. Non-dominated individuals form the first layer (or class). Those individuals dominated only by members of the first layer form the second one and are assigned
a fitness value of 2, and so on. In general, for dominated individuals, the fitness is
given by the number of dominating layers plus 1. A crowding factor is then assigned
to each individual in a given class. The crowding factor is computed as the sum of
the Euclidean distances, in criteria space, with respect to other individuals in the same
class, divided by the interval spanned by the population along each dimension of the
objective space. Inside each class, the individuals with the higher value of the crowding
parameter obtain a better rank than those with a lower one.
At every generation, binary tournament selection, recombination, and mutation operators are used to create an offspring of the current population. The combination of
the two is then sorted according to dominance first and then to crowding. The nondominated individuals with lowest crowding factor are then used to update the population.
The parameters to be set are the size of the population, the number of generations, the
crossover and mutation probability, pc and pm , and distribution indexes for crossover
and mutation, ηc and ηm , respectively. Three different ratios between population size
and number of generations were considered: 0.08, 0.33 and 0.75. The values pc and
pm were set to 0.9 and 0.2 respectively and kept constant for all the tests. The values,
5, 10 and 20, were considered for ηc , while tests were run for values of ηm equal to 5,
25 and 50.
4.1.2. MOPSO
MOPSO is an extension of Particle Swarm Optimization (PSO) to multiobjective problems. Pareto dominance is introduced in the selection criteria for the candidate solutions to update the population. MOPSO features an external archive which stores all the
non-dominated solutions and at the same time is used to guide the search process of the
15
swarm. This is done by introducing the possibility to direct the movement of a particle
towards one of the less crowded solutions in the archive. The solution space is subdivided into hypercubes through an adaptive grid. The solutions in the external archive
are thus reorganized in these hypercubes. The algorithm keeps track of the crowding
level of each hypercube and promotes movements towards less crowded areas. In a
similar manner, it also gives priority to the insertion of new non-dominated individuals
in less crowded areas if the external archive has already reached its predefined maximum size. For MOPSO three different ratios between population size and number of
generations were tested: 0.08 0.33 0.75. It was also tested with three different numbers
of subdivisions of the solution space: 10, 30, and 50. The inertia component in the
motion of the particles was set to 0.4 and the weights of the social and individualistic
components were se to 1
4.1.3. PAES
PAES is a (1 + 1) Evolution Strategy with the addition of an external archive to store
the current best approximation of the Pareto front. It adopts a population of only a
single chromosome which, at every iteration, generates a mutated copy. The algorithm
then preserves the non-dominated one between the parent and the candidate solution. If
none of the two dominates the other, the algorithm then checks their dominance index
with respect to the solutions in the archive. If also this comparison is inconclusive,
then the algorithm selects the one which resides in the less crowded region of the
objective space. To keep track of the crowding level of the objective space, the latter is
subdivided in an n-dimensional grid. Every time a new solution is added to (or removed
from) the archive, the crowding level of the corresponding grid cell is updated. PAES
has two main parameters that need to be set, the number of subdivisions in the space
grid and the mutation probability. Values of 1,2 and 4 were used for the former and 0.6
0.8 and 0.9 were used for the latter.
16
4.1.4. MTS
MTS is an algorithm based on pattern search. The algorithm first generates a population
of uniformly distributed individuals. At every iteration, a local search is performed by
a subset of individuals. Three different search patterns are included in the local search:
the first is a search along the direction of each decision variable with a fixed step length;
the second is analogous but the search is limited to one fourth of all the possible search
directions; the third one also searches along each direction but selects only solutions
which are evenly spaced on each dimension within a predetermined upper bound and
lower bound. At each iteration and for each individual, the three search patterns are
tested with few function evaluations to select the one which generates the best candidate
solutions for the current individual. The selected one is then used to perform the local
search which will update the individual itself. The step length along each direction,
which defines the size of the search neighbourhood for each individual, is increased if
the local search generated a non-dominated child and is decreased otherwise. When
the neighbourhood size reaches a predetermined minimum value, it is reset to 40% of
the size of the global search space. The non-dominated candidate solutions are then
used to update the best approximation of the global Pareto front. MTS was tested with
a population size of 20, 40 and 80 individuals.
4.2. PERFORMANCE METRICS
Two metrics were defined to evaluate the performance of the tested multiobjective optimizers:
Mspr =
Mconv =
Mp
1 X
fj − g i
min 100
Mp i=1 j∈Np
gi
Np
gj − fi
1 X
min 100
Np i=1 j∈Mp
gj
(14)
(15)
where Mp is the number of elements, with objective vector g, in the true global Pareto
front and Np is the number of elements, with objective vector f , in the Pareto front that
a given algorithm is producing. Although similar, the two metrics are measuring two
17
different things: Mspr is the sum, over all the elements in the global Pareto front, of the
minimum distance of all the elements in the Pareto front Np from the the ith element in
the global Pareto front. Mconv , instead, is the sum, over all the elements in the Pareto
front Np , of the minimum distance of the elements in the global Pareto front from the
ith element in the Pareto front Np .
Therefore, if Np is only a partial representation of the global Pareto front but is a
very accurate partial representation, then metric Mspr would give a high value and
metric Mconv a low value. If both metrics are high then the Pareto front Np is partial
and poorly accurate. The index Mconv is similar to the mean Euclidean distance [15],
although in Mconv the Euclidean distance is normalized with respect to the values of the
objective functions, while Mspr is similar to the generational distance [18], although
even for Mspr the distances are normalized.
Given n repeated runs of a given algorithm, we can define two performance indexes:
pconv = P (Mconv < tolconv ) or the probability that the index Mconv achieves a value
less than the threshold tolconv and pspr = P (Mspr < tolspr ) or the probability that
the index Mspr achieves a value less than the threshold tolconv .
According to the theory developed in [7, 19], 200 runs are sufficient to have a 95%
confidence that the true values of pconv and pspr are within a ±5% interval containing
their estimated value.
Performance index (14) and (15) tend to uniformly weigh every part of the front. This
is not a problem for index (14) but if only a relatively small portion of the front is
missed the value of performance index (15) might be only marginally affected. For this
reason, we slightly modified the computation of the indexes by taking only the MP∗ and
NP∗ solutions with a normalised distance in criteria space that was higher than 10−3 .
The global fronts used in the three space tests were built by taking a selection of about
2000 equispaced, in criteria space, nondominated solutions coming from all the 200
runs of all the algorithms.
18
4.3. PRELIMINARY TEST CASES
For the preliminary tests, two sets of functions, taken from the literature [14, 15], were
used and the performance of MACS was compared to the results in [14] and [15].
Therefore, for this set of tests, MTS was not included in the comparison. The function
used in this section can be found in Table 1.
Table 1: Multiobjective test functions
The first three functions were taken from [15]. Test cases Deb and Scha are two
examples of disconnected Pareto fronts, Deb2 is an example of problem presenting
multiple local Pareto fronts, 60 in this two dimensional case.
The last three functions are instead taken from [14]. Test case ZDT 2 has a concave
Pareto front with a moderately high-dimensional search space. Test case ZDT 6 has
a concave Pareto front but because of the irregular nature of f1 there is a strong bias
in the distribution of the solutions. Test case, ZDT4, with dimension 10, is commonly
recognized as one of the most challenging problems since it has 219 different local
Pareto fronts of which only one corresponds to the global Pareto-optimal front.
As a preliminary proof of the effectiveness of MACS, the average Euclidean distance
of 500 uniformly spaced points on the true optimal Pareto front from the solutions
stored in Ag by MACS was computed and compared to known results in the literature.
MACS was run 20 times to have a sample comparable to the one used for the other
algorithms. The global archive was limited to 200 elements to be consistent with [14].
The value of the crowding factor wc , the threshold ρtol and the convergence ρmin were
kept constant to 1e-5 in all the cases to provide good local convergence.
To be consistent with [15], on Deb, Scha and, Deb2, MACS was run respectively for
4000, 1200 and 3200 function evaluations. Only two agents were used for these lower
dimensional cases, with fe = 1/2. On test cases ZDT 2, ZDT 4 and ZDT 6, MACS
was run for a maximum of 25000 function evaluations to be consistent with [14], with
three agents and fe = 2/3 for ZDT 2 and four agents and fe = 3/4 on ZDT 4 and
ZDT 6.
19
The results on Deb, Scha and, Deb2 can be found in Table 2, while the results on
ZDT 2, ZDT 4 and ZDT 6 can be found in 3.
On all the smaller dimensional cases MACS performs comparably to MOPSO and better than PAES. It also performs than NSGA-II on Deb and Deb2. On Scha MACS
performs apparently worse than NSGA-II, although after inspection one can observe
that all the elements of the global archive Ag belong to the Pareto front but not uniformly distributed, hence the higher value of the Euclidean distance. On the higher
dimensional cases, MACS performs comparably to NSGA-II on ZDT2 but better than
all the others on ZDT4 and ZDT6. Note in particular the improved performance on
ZDT4.
Table 2: Comparison of the average Euclidean distances between 500 uniformly
space points on the optimal Pareto front for various optimization algorithms:
smaller dimension test problems.
Table 3: Comparison of the average Euclidean distances between 500 uniformly
space points on the optimal Pareto front for various optimization algorithms:
larger dimension test problems.
On the same six functions a different test was run to evaluate the performance of different variants of MACS. For all variants, the number of agents, fe , wc , ρtol and ρmin was
set as before, but instead of the mean Euclidean distance, the success rates pconv and
pspr were measured for each variant. The number of function evaluations for ZDT 2,
ZDT 4, ZDT 6 and Deb2 is the same as before, while for Scha and Deb it was reduced respectively to 600 and 1000 function evaluations given the good performance
of MACS already for this number of function evaluations. Each run was repeated 200
times to have good confidence in the values of pconv and pspr .
Four variants were tested and compared to the full version of MACS. Variant MACS no
local does not implement the individualistic moves and the local archive, variant MACS
ρ = 1 has no adaptivity on the neighborhood, its size is kept fixed to 1,variant MACS
ρ = 0.1 has the size of the neighborhood fixed to 0.1, variant MACS no attraction has
the attraction mechanisms not active.
20
The result can be found in Table 4. The values of tolconv and tolspr are respectively
0.001 and 0.0035 for ZDT 2, 0.003 and 0.005 for ZDT 4, 0.001 and 0.025 for ZDT 6,
0.0012 and 0.035 for Deb, 0.0013 and 0.04 for Scha, 0.0015 and 0.0045 for Deb2.
This thresholds were selected to highlight the differences among the various variants.
The table shows that the adaptation mechanism is beneficial in some cases although,
in others, fixing the value of ρ might be a better choice. This depends on the problem
and a general rule is difficult to derive at present. Other adaptation mechanisms could
further improve the performance.
The use of individualistic actions coupled with a local archive is instead fundamental,
so is the use of the attraction mechanism. Note, however, how the attraction mechanism penalizes the spreading on biassed problems like ZDT 6, this is expected as it
accelerates convergence.
Table 4: Comparison of different variants of MACS.
4.4. APPLICATION TO SPACE TRAJECTORY DESIGN
In this section we present the application of MACS to three space trajectory problems: a two-impulse orbit transfer from a Low Earth Orbit (LEO) to a high-eccentricity
Molniya-like orbit, a three-impulse transfer from a LEO to Geostationary Earth Orbit
(GEO) and a multi-gravity assist transfer to Saturn equivalent to the transfer trajectory
of the Cassini mission. The first two cases are taken from the work of Minisci et al. [7].
In the two-impulse case, the spacecraft departs at time t0 from a circular orbit around
the Earth (the gravity constant is µE = 3.9860105 km3 s−2 ) with radius r0 = 6721
km and at time tf is injected into an elliptical orbit with eccentricity eT = 0.667
and semimajor axis aT = 26610 km. The transfer arc is computed as the solution
of a Lambert’s problem [20] and the objective functions are the transfer time T =
tf − t0 and the sum of the two norms of the velocity variations at the beginning and
at the end of the transfer arc ∆vtot . The objectives are functions of the solution vector
x = [t0 tf ]T ∈ D ⊂ R2 The search space D is defined by the following intervals
t0 ∈ [0 10.8], and tf ∈ [0.03 10.8].
21
In the three-impulse case, the spacecraft departs at time t0 from a circular orbit around
the Earth with radius r0 = 7000 km and after a transfer time T = t1 + t2 is injected
into a circular orbit with radius rf = 42000. An intermediate manoeuvre is performed
at time t0 + t1 and at position defined in polar coordinates by the radius r1 and the
angle θ1 . The objective functions are the total transfer time T and the sum of the
three impulses ∆vtot . The solution vector in this case is x = [t0 , t1 , r1 , θ1 , tf ]T ∈
D ⊂ R5 . The search space D is defined by the following intervals t0 ∈ [0 1.62],
t1 ∈ [0.03 21.54], r1 ∈ [7010 105410], θ1 ∈ [0.01 2π − 0.01], and t2 ∈ [0.03 21.54].
The Cassini case consists of 5 transfer arcs connecting a departure planet, the Earth,
to the destination planet, Saturn, through a sequence of swing-by’s with the planets:
Venus, Venus, Earth, Jupiter. Each transfer arc is computed as the solution of a Lambert’s problem [21] given the departure time from planet Pi and the arrival time at
planet Pi+1 . The solution of the Lambert’s problems yields the required incoming
and outgoing velocities at each swing-by planet vin and vrout . The swing-by is modeled through a linked-conic approximation with powered maneuvers [22], i.e., the mismatch between the required outgoing velocity vrout and the achievable outgoing velocity vaout is compensated through a ∆v maneuver at the pericenter of the gravity
assist hyperbola. The whole trajectory is completely defined by the departure time t0
and the transfer time for each leg Ti , with i = 1, ..., 5. The normalized radius of the
pericenter rp,i of each swing-by hyperbola is derived a posteriori once each powered
swing-by manoeuvre is computed. Thus, a constraint on each pericenter radius has to
be introduced during the search for an optimal solution. In order to take into account
this constraint, one of the objective functions is augmented with the weighted violation
of the constraints:
f (x) = ∆v0 +
4
X
∆vi + ∆vf +
4
X
wi (rp,i − rpmin,i )2
(16)
i=1
i=1
for a solution vector x = [t0 , T1 , T2 , T3 , T4 , T5 ]T . The objective functions are, the toP
tal transfer time T = 5i Ti and f (x). The minimum normalized pericenter radii are
rpmin,1 = 1.0496, rpmin,2 = 1.0496, rpmin,3 = 1.0627 and rpmin,4 = 9.3925. The
22
search space D is defined by the following intervals: t0 ∈ [−1000, 0]MJD2000, T1 ∈
[30, 400]d, T2 ∈ [100, 470]d, T3 ∈ [30, 400]d, T4 ∈ [400, 2000]d, T5 ∈ [1000, 6000]d.
The best known solution for the single objective minimization of f (x) is fbest =
4.9307 km/s, with xbest = [−789.753, 158.2993, 449.3859, 54.7060, 1024.5896, 4552.7054]T .
4.4.1. Test Results
For this second set of tests, each algorithm was run for the same number of function
evaluations. In particular, consistent with the tests performed in the work of Minisci et
al., [7] we used 2000 function evaluations for the two-impulse case and 30000 for the
three-impulse case. For the Cassini case, instead, the algorithms were run for 180000,
300000 and 600000 function evaluations.
Note that the version of all the algorithms used in this second set of tests is the one that
is freely available online, written in c/c++. We tried in all cases to stick to the available
instructions and recommendations by the author to avoid any bias in the comparison.
The thresholds values for the two impulse cases was taken from [7] and is tolconv =
0.1, tolspr = 2.5. For the three-impulse case instead we considered tolconv = 5.0,
tolspr = 5.0. For the Cassini case we used tolconv = 0.75, tolspr = 5, instead. These
values were selected after looking at the dispersion of the results over 200 runs. Lower
values would result in a zero value of the performance indexes of all the algorithms,
which is not very significant for a comparison.
MACS was tuned on the three-impulse case. In particular, the crowding factor wc , the
threshold ρtol and the convergence ρmin were kept constant to 1e-5, which is below
the required local convergence accuracy, while fe and npop were changed. A value of
1e-5 is expected to provide good local convergence and good density of the samples
belonging to the Pareto front. Table 5 reports the value of performance indexes pconv
and pspr over 200 runs of MACS with different settings. The index pspr and the index
pconv have different, almost opposite, trends. However, it was decided to select the
setting that provides the best convergence,i.e. npop = 15 and fe = 1/3. This setting
will be used for all the tests in this paper.
On top of the complete algorithm, two variants of MACS were tested: one without
23
individualistic moves and local archive, denoted as no local in the tables, and one with
no attraction towards the global archive Ag , denoted as no att in the tables. Only these
two variants are tested on these cases as they displayed the most significant impact in
the previous standard test cases and more importantly were designed specifically to
improve performances.
NSGA-II, PAES, MOPSO and MTS were tuned as well on the three-impulse case. In
particular, for NSGA-II the best result was obtained for 150 individuals and can be
found in Table 6. A similar result could be obtained for MOPSO, see Table 7. For
MTS only the population was changed while the number of individuals performing
local moves was kept constant to 5. The results of the tuning of MTS can be found in
Table 9. For the tuning of PAES the results can be found in Table 8.
All the parameters tuned in the three impulse case were kept constant except for the
population size of NSGA-II and MOPSO. The size of the population of NSGA-II and
MOPSO was set to 100 and 40 respectively on the two impulse case and was increased
with the number of function evaluations in the Cassini case. In particular for NSGAII the following ratios between population size and number of function evaluations
was used: 272/180000, 353/300000, 500/600000. For MOPSO the following ratios
between population size and number of function evaluations was used: 224/180000,
447/300000, 665/600000. This might not be the best way to set the population size for
these two algorithms but it is the one that provided better performance in these tests.
Note that the size of the global archive for MACS was constrained to be lower than
the size of the population of NSGA-II, in order to avoid any bias in the computation of
Mspr .
The performance of all the algorithms on the Cassini case can be found in Table 12 for
a variable number of function evaluations.
Figure 2: Three-impulse test case: a) Complete Pareto front, b) close-up of the
Pareto fronts
Table 5: Indexes pconv and pspr for different settings of MACS
24
Figure 3: Cassini test case: a) Complete Pareto front, b) close-up of the Pareto
fronts
For the three-impulse case, MACS was able to identify an extended Pareto front (see
Fig.2(a) and Fig. 2(b) where all the non-dominated solutions from all the 200 runs are
compared to the global front), compared to the results in [7]. The gap in the Pareto
front is probably due to a limited spreading of the solutions in that region. Note the
cusp due the transition between the condition in which 2-impulse solutions are optimal
and the condition in which 3-impulse solutions are optimal.
Table 11 summarizes the results of all the tested algorithms on the two-impulse case.
The average value of the performance metrics is reported with, in brackets, the associated variance over 200 runs. The two-impulse case is quite easy and all the algorithms
have no problems identifying the front. However, MACS displays a better convergence
than the other algorithms while the spreading of MOPSO and MTS is superior to the
one of MACS.
The three-impulse case is instead more problematic (see Table 10). NSGA-II is not
able to converge to the upper-left part of the front and therefore the convergence is 0
and the spreading is comparable to the one of MACS. All the other algorithms perform
poorly with a value of almost 0 for the performance indexes. This is mainly due to the
fact that no one can identify the vertical part of the front.
Note that the long tail identified by NSGA-II is actually dominated by two points belonging to the global front (see 2(b)).
Table 12 reports the statistics for the Cassini problem. On top of the performance of the
three variants tested on the other two problems, the table reports also the result for 10
agents and fe = 5. For all numbers of function evaluations MACS has better spreading
than NSGA-II because NSGA-II converges to a local Pareto front. Nonetheless NSGAII displays a more regular behavior and a better convergence for low number of function
evaluations although it never reaches the best front. The Pareto fronts are represented
in Fig. 3(a) and Fig. 3(b) (even in this case the figures represent all the non-dominated
solutions coming from all the 200 runs). Note that, the minimum f returned by MACS
25
is the best known solution for the single objective version of this problem (see [9]). All
the other algorithms perform quite poorly on this case.
Of all the variants of MACS tested on these problems the full complete one is performing the best. As expected, in all three cases, removing the individualistic moves
severely penalizes both convergence and spreading. It is interesting to note that removing the attraction towards the global front is also comparatively bad. On the two
impulse case it does not impact the spreading but reduces the convergence, while on
the Cassini case, it reduces mean and variance but the success rates are zero. The observable reason is that MACS converges slower but more uniformly to a local Pareto
front.
Finally, it should be noted that mean and variance seem not to capture the actual performance of the algorithms. In particular they do not capture the ability to identify the
whole Pareto front as the success rates instead do.
Table 6: NSGAII tuning on the 3-impulse case
Table 7: PAES tuning on the 3-impulse case
5. CONCLUSIONS
In this paper we presented a hybrid evolutionary algorithm for multiobjective optimization problems. The effectiveness of the hybrid algorithm, implemented in a code
called MACS, was demonstrated at first on a set of standard problems and then its performance was compared against NSGA-II, PAES, MOPSO and MTS on three space
trajectory design problems. The results are encouraging as, for the same computational
effort (measured in number of function evaluations, MACS was converging more accuTable 8: MOPSO tuning on the 3-impulse case.
Table 9: MTS tuning on the 3-impulse
26
rately than NSGA-II on the two-impulse case and managed to find a previously undiscovered part of the Pareto front of the three-impulse case. As a consequence, on the
three-impulse case, MACS, has better performance metrics than the other algorithms.
On the Cassini case NSGA-II appears to converge better to some parts of the front although MACS yielded solutions with better f and identifies once more a part of the
front that NSGA-II cannot attain. PAES and MTS do not perform well on the Cassini
case, while MOPSO converges well locally but, with the settings employed in this
study, yielded a very poor spreading.
From the experimental tests in this paper we can argue that the following mechanisms
seem to be particularly effective: the use of individual local actions with a local archive
as they allow the individuals to move towards and within the Pareto set; the use of an
attraction mechanism as it accelerates convergence.
Finally it should be noted that all the algorithms tested in this study use the Pareto
dominance as selection criterion. Different criteria, like the decomposition in scalar
subproblems, can be equality implemented in MACS, without disrupting its working
principles, and lead to different performance results.
ACKNOWLEDGEMENTS
The authors would like to thank Dr. Edmondo Minisci and Dr. Guilio Avanzini for the
two and three impulse cases and the helpful advice.
REFERENCES
[1] V. Coverstone-Caroll, J.W. Hartmann, and W.M. Mason. Optimal multi-objective
low-thrust spacecraft trajectories. Computer Methods in Applied Mechanics and
Engineering, 186:387–402, 2000.
[2] B. Dachwald. Optimization of interplanetary solar sailcraft trajectories using evolutionary neurocontrol. Journal of Guidance, and Dynamics, February 2004.
[3] S. Lee, P. von Allmen, W. Fink, A.E. Petropoulos, and R.J. Terrile. Multiobjective evolutionary algorithms for low-thrust orbit transfer optimization. In
27
Proceedings of the Genetic and Evolutionary Computation Conference (GECCO
2005), Washington DC, USA, June 25–29 2005.
[4] O. Schütze, M. M. Vasile, and C.A. Coello Coello. Approximate solutions in
space mission design. In Parallel Problem Solving from Nature (PPSN 2008),
Dortmund, Germany, September 13–17 2008.
[5] O. Schütze, M. Vasile, O. Junge, M. Dellnitz, and D. Izzo. Designing optimal
low-thrust gravity-assist trajectories using space pruning and a multi-objective
approach. Engineering Optimization, 41(2), February 2009.
[6] M. Dellnitz, S. Ober-Blbaum, M. Post, O. Schtze, and Bianca Thiere. A multiobjective approach to the design of low thrust space trajectories using optimal
control. Celestial Mechanics and Dynamical Astronomy, 105(1), February 2009.
[7] E. Minisci and G. Avanzini. Optimisation of orbit transfer manoeuvres as a
test benchmark for evolutionary algorithms. In Proceedings of the 2009 IEEE
Congress on Evolutionary Computation (CEC2009), Trondheim, Norway, May
2009.
[8] M. Vasile. A behavioral-based meta-heuristic for robust global trajectory optimization. In Proceedings of the 2007 IEEE Congress on Evolutionary Computation (CEC2007), Singapore, September 2007.
[9] M. Vasile and M. Locatelli. A hybrid multiagent approach for global trajectory
optimization. Journal of Global Optimization, 44(4):461–479, August. 2009.
[10] Oliver Schütze, Adriana Lara, Gustavo Sanchez, and Carlos A. Coello Coello.
HCS: A new local search strategy for memetic multi-objective evolutionary algorithms. IEEE Transactions on Evolutionary Computation (to appear), 2010.
[11] Q. Zhang and H. Li. Moea/d: A multiobjective evolutionary algorithm based on
decomposition. IEEE Transactions on Evolutionary Computation, 11(6), November.
28
[12] K.V. Price, R.M. Storn, and J.A. Lampinen. Differential Evolution. A Practical
Approach to Global Optimization. Natural Computing Series. Springer, 2005.
[13] M. Vasile and F. Zuiani. A hybrid multiobjective optimization algorithm applied
to space trajectory optimization. In Proceedings of the IEEE International Conference on Evolutionary Computation, Barcelona, Spain, July 2010.
[14] K. A. Deb, A. Pratap, and T. Meyarivan. Fast elitist multi-objective genetic algorithm: Nga-ii. Kangal report no. 200001, KanGAL, 2000.
[15] C. Coello and M. Lechuga. Mopso: A proposal for multiple objective particle
swarm optimization. Technical report evocinv-01-2001., CINVESTAV. Instituto
Politecnico Nacional. Col. San Pedro Zacatenco. Mexico., 2001.
[16] J.D. Knowles and D.W. Corne. The pareto archived evolution strategy : A new
baseline algorithm for pareto multiobjective optimisation. In Proceedings of the
IEEE International Conference on Evolutionary Computation, Washington DC,
US, 1999.
[17] Lin-Yu Tseng and Chun Chen. Multiple trajectory search for multiobjective optimization. In Proceedings of the IEEE International Conference on Evolutionary
Computation, Singapore, 25-28 September 2007.
[18] D.A. Van Veldhuizen and G.B. Lamont. Evolutionary computation and convergence to a pareto front. In Late Breaking papers at the Genetic Programming,
1998.
[19] M. Vasile, E. Minisci, and M. Locatelli. On testing global optimization algorithms
for space trajectory design. In AIAA/AAS Astrodynamic Specialists Conference,
Honolulu, Hawaii, USA, Aug 2008.
[20] G. Avanzini. A simple lambert algorithm. Journal of Guidance, Control, and
Dynamics, 31(6):1587–1594, Nov.–Dec. 2008.
[21] R. Battin. An Introduction to the Mathematics and Methods of Astrodynamics.
AIAA, 1999.
29
[22] D. R. Myatt, V.M. Becerra, S.J. Nasuto, and J.M. Bishop.
mization tools for mission analysis and design.
Final rept. esa ariadna itt
ao4532/18138/04/nl/mv,call03/4101, ESA/ESTEC, 2004.
30
Global opti-
Table 10: Summary of metrics Mconv and Mspr , and associated performance
indexes pconv and pspr , on the three impulse test cases.
Table 11: Metrics Mconv and Mspr , and associated performance indexes pconv
and pspr , on the two impulse test cases.
Table 12: Metrics Mconv and Mspr , and associated performance indexes pconv
and pspr , on the Cassini case.
31
NOMENCLATURE
Ag
global archive
Al
local archive
aT
semimajor axis
D
search space
eT
eccentricity
f
cost function
fe
fraction of the population doing local moves
Id
dominance index
Mconv
convergence metrics
Mspr
spreading metrics
Nρ
neighborhood of solution x
Ne
maximum number of allowed function evaluations
neval
number of function evaluations
npop
population size
Pi
i-th planet
Pk
population at generation k
pconv
percentage of success on convergence
pspr
percentage of success on spreading
r
random number
rp
pericentre radius
rpmin
minimum pericentre radius
S
selection function
s
resource index
T
transfer time
t0
departure time
ti
manoeuvre time
tf
final time
U
uniform distribution
32
u
variation of the solution x
X
Pareto optimal set
x
solution vector
y
mutate individual
wc
tolerance on the maximum allowable crowding
Greek symbols
∆v
variation of velocity
ρ
size of the neighborhood Nρ
µE
gravity constant
θi
true anomaly of manoeuvre i
33
LIST OF TABLES
Table 1: Multiobjective test functions
Scha
x ∈ [−5, 10]
Deb
x1 , x2 ∈ [0, 1]
α = 2; q = 4
Deb2
x1 ∈ [0, 1]
x2 ∈ [−30, 30]
ZDT2
xi ∈ [0, 1];
i = 1, . . . , n
n = 30
ZDT4
x1 ∈ [0, 1];
xi ∈ [−5, 5];
i = 2, . . . , n
n = 10
ZDT6
xi ∈ [0, 1];
i = 1, . . . , n
n = 10
2
f2 = (x
− 5)
−x
−2 + x
f1 =
4−x
−4 + x
f1 = x1
h
if
if
if
if
f2 = (1 + 10x2 ) 1 −
i
x1
sin(2πqx
)
1
1+10x2
x≤1
1<x<3
3<x≤4
x>4
x1
1+10x2
α
−
f1 = x1
f2 = g(x1 , x2 )h(x1 ,q
x2 ); g(x1 , x2 ) = 11 + x22 − 10 cos(2πx2 )
n
1 − fg1 iff1 ≤ g
h(x1 , x2 ) =
0 otherwise
9 Pn
g = 1 + n−1
i=2 xi
f1 2
h=1−( g )
f1 = x1 ; f2 = gh
g = 1 + 10(n
q − 1) +
f1
g
h=1−
f1 = x1 ; f2 = gh
Pn
2
i=2 [xi
− 10 cos(2πqxi )];
q Pn
i=2 xi
g = 1 + 9 4 n−1
h = 1 − ( fg1 )2
f1 = 1 − exp(−4x1 ) sin6 (6πx1 ); f2 = gh
Table 2: Comparison of the average Euclidean distances between 500 uniformly
space points on the optimal Pareto front for various optimization algorithms:
smaller dimension test problems.
Approach
MACS
NSGA-II
PAES
MOPSO
Deb2
Scha
Deb
1.542e-3
(5.19e-4)
0.094644
(0.117608)
0.259664
(0.573286)
0.0011611
(0.0007205)
3.257e-3
(5.61e-4)
0.001594
(0.000122)
0.070003
(0.158081)
0.00147396
(0.00020178)
7.379e-4
(6.36e-5)
0.002536
(0.000138)
0.002881
(0.00213)
0.002057
(0.000286)
34
Table 3: Comparison of the average Euclidean distances between 500 uniformly
space points on the optimal Pareto front for various optimization algorithms:
larger dimension test problems.
Approach
MACS
NSGA-II
PAES
ZDT2
ZDT4
ZDT6
9.0896e-4
(4.0862e-5)
0.000824
(<1e-5)
0.126276
(0.036877)
0.0061
(0.0133)
0.513053
(0.118460)
0.854816
(0.527238)
0.0026
(0.0053)
0.296564
(0.013135)
0.085469
(0.006644)
Table 4: Comparison of different versions of MACS.
Approach
Metric
ZDT2
ZDT4
ZDT6
Scha
Deb
Deb2
MACS
pconv
pspr
pconv
pspr
pconv
pspr
pconv
pspr
pconv
pspr
83.5%
22.5%
14%
1%
84%
22%
56%
5%
21%
0%
75%
28%
0%
0%
22%
7%
42%
15%
0.5%
0.5%
77%
58.5%
45%
34%
78%
63%
57%
61%
14.5%
78.5%
73%
38.5%
0.5%
0%
37%
0%
78%
88%
0.5%
0%
70.5%
83%
72.5%
4%
92%
54%
85%
94.5%
88.5%
0%
60%
67.5 %
11%
15%
21%
38%
42%
74%
0%
0%
MACS
no local
MACS
ρ=1
MACS
ρ = 0.1
MACS
no attraction
Table 5: Indexes pconv and pspr for different settings of MACS on the 3-impulse
case
pconv
npop = 5
npop = 10
npop = 15
fe = 1/3
fe = 1/2
fe = 2/3
45.5%
48.0%
45.0%
55.5%
51.0%
52.5%
61.0%
55.5%
43.0%
pspr
fe = 1/3
fe = 1/2
fe = 2/3
npop = 5
68.5%
65.0%
67.5%
npop = 10
62.0%
57.0%
51.0%
npop = 15
56.0%
46.0%
36.5%
35
Table 6: NSGAII tuning on the 3-impulse case
Mean Mconv
ηc / ηm
5
10
20
Mean Mspr
ηc / ηm
5
10
20
pconv
ηc / ηm
5
10
20
5
36.1
32.3
31.7
25
38.3
39.4
39.6
50
43.0
40.6
42.5
5
6.77
5.91
5.78
25
7.24
7.50
7.50
50
8.08
7.81
8.16
5
0.0%
0.0%
0.0%
25
0.0%
0.0%
0.0%
50
0.0%
0.0%
0.0%
Var Mconv
ηc / ηm
5
10
20
Var Mspr
ηc / ηm
5
10
20
pspr
ηc / ηm
5
10
20
5
201.0
182.0
175.0
25
202.0
172.0
183.0
50
185.0
182.0
169.0
5
9.97
9.74
9.75
25
9.47
8.68
8.53
50
7.25
8.34
8.04
5
44.8%
57.8%
61.0%
25
37.7%
33.1%
32.5%
50
23.4%
29.2%
24.7%
Table 7: MOPSO tuning on the 3-impulse case.
Mean Mconv
Particles/Subdivisions
50
100
150
Mean Mspr
Particles/Subdivisions
50
100
150
pconv
Particles/Subdivisions
50
100
150
10
59.1
50.0
47.8
30
50.2
43.3
41.4
50
41.5
41.3
39.4
10
16.1
14.7
14.8
30
13.5
12.2
11.9
50
12.3
11.6
11.4
10
0%
0%
0%
30
0%
0%
0%
50
0%
0%
0%
36
Var Mconv
Particles/Subdivisions
50
100
150
Var Mspr
Particles/Subdivisions
50
100
150
pspr
Particles/Subdivisions
50
100
150
10
1080.0
591.0
562.1
30
1010.2
721.1
550.2
50
713.3
778.1
608.2
10
41.3
32.8
30.0
30
37.3
25.8
22.2
50
24.0
24.5
22.5
10
0.5%
0.5%
0.5%
30
2.5%
4.5%
2%
50
3.5%
4%
4%
Table 8: PAES tuning on the 3-impulse case
Mean Mconv
Subdivisions/Mutation
1
2
4
Mean Mspr
Subdivisions/Mutation
1
2
4
pconv
Subdivisions/Mutation
1
2
4
0.6
53.7
52.8
53.0
0.8
70.6
70.2
70.2
0.9
70.2
70.0
70.1
0.6
14.2
13.6
13.8
0.8
27.7
27.6
27.7
0.9
36.7
36.6
36.6
0.6
0.0%
0.0%
0.0%
0.8
0.0%
0.0%
0.0%
0.9
0.0%
0.0%
0.0%
Var Mconv
Subdivisions/Mutation
1
2
4
Var Mspr
Subdivisions/Mutation
1
2
4
pspr
Subdivisions/Mutation
1
2
4
0.6
525.0
479.0
453.0
0.8
275.0
266.0
266.0
0.9
297.0
305.0
311.0
0.6
20.0
17.5
17.2
0.8
14.3
14.6
15.8
0.9
17.3
16.8
17.0
0.6
0.0%
0.0%
0.0%
0.8
0.0%
0.0%
0.0%
0.9
0.0%
0.0%
0.0%
Table 9: MTS tuning on the 3-impulse
3imp
Mconv
Mspr
Population
Mean
Var
Mean
Var
pconv
pspr
20
17.8
97.6
12.6
34.7
1.0%
0.5%
40
22.6
87.8
19.9
26.2
0.0%
0.0%
80
23.6
73.2
18.4
18.6
0.0%
0.0%
Table 10: Summary of metrics Mconv and Mspr , and associated performance
indexes pconv and pspr , on the three impulse test cases.
Metric
MACS
MACS
no local
MACS
no att
NSGA-II
PAES
MOPSO
MTS
Mconv
5.53
(15.1)
5.25
(3.73)
61.0%
56.0%
7.58
(26.3)
6.03
(3.95)
40.5%
36%
154.7
(235.0)
9.16
(2.07)
0.0%
0.0%
31.7
(175.0)
5.78
(9.75)
0.0%
61.0%
53.0
(453.0)
13.8
(17.2)
0.0%
0.0%
39.4
(608.1)
11.4
(22.5)
0%
4.0%
17.8
(97.6)
12.6
(34.7)
1.0%
0.5%
Mspr
pconv
pspr
37
Table 11: Metrics Mconv and Mspr , and associated performance indexes pconv
and pspr , on the two impulse test cases.
Metric
MACS
MACS
no local
MACS
no att
NSGA-II
PAES
MOPSO
MTS
Mconv
0.0077
(1e-3)
2.89
(0.49)
98.5%
29.5%
0.0039
(1.4e-2)
6.87
(7.36)
91.5%
0%
0.0534
(9.5e-3)
3.08
0.943
84%
24%
0.283
(4.35e-3)
2.47
(0.119)
0%
62.5%
0.198
(0.332)
332.0
(2.61e4)
75.5%
0.5%
0.378
(0.0636)
2.11
(2.65)
9%
94.5%
0.151
(0.083)
1.95
(1.41)
57.5%
85.5%
Mspr
pconv
pspr
38
Table 12: Metrics Mconv and Mspr , and associated performance indexes pconv
and pspr , on the Cassini case.
Approach
Metric
180k
300k
600k
MACS
5/15
Mconv
Mspr
pconv
pspr
6.50 (229.1)
12.7 (126.0)
6%
27.5%
4.48 (107.1)
11.1 (83.1)
14%
31%
3.91 (62.6)
8.64 (35.5)
21.5%
41%
MACS
5/10
Mconv
Mspr
pconv
pspr
6.74 (217.9)
12.1 (83.0)
8%
26.0%
5.56 (136.4)
10.3 (63.7)
10%
31.5%
3.14 (50.1)
8.11 (30.0)
25.5%
45.5%
MACS
no local
Mconv
Mspr
pconv
pspr
13.5 (436.2)
31.1 (274.3)
1.0%
1.0%
10.2 (350.1)
27.9 (278.9)
1.5%
2.5%
7.86 (190.2)
21.9 (226.7)
2.5%
6%
MACS
no att
Mconv
Mspr
pconv
pspr
2.62 (1.46)
27.8 (83.2)
0.0%
0.0%
2.2 (0.88)
22.9 (58.0)
0.0%
0.0%
1.82 (0.478)
18.2 (28.0)
1.0%
0.0%
NSGA-II
Mconv
Mspr
pconv
pspr
2.43 (18.0)
11.6 (71.4)
17.5%
15.5%
1.99 (16.8)
11.0 (47.5)
24.0%
12.5%
1.24 (1.62)
8.78 (28.2)
29.0%
25.0%
MOPSO
Mconv
Mspr
pconv
pspr
2.62 (7.33)
28.0 (308.3)
0.5%
0.0%
2.4 (2.57)
24.6 (260.4)
1.0%
0.0%
2.14 (0.94)
21.8 (231.3)
1.0%
0.5%
PAES
Mconv
Mspr
pconv
pspr
24.0 (54.5)
30.1 (47.3)
0.0%
0.0%
19.8 (32.9)
26.0 (33.9)
0.0%
0.0%
15.2 (16.6)
21.4 (19.5)
0.0%
0.0%
MTS
Mconv
Mspr
pconv
pspr
3.71 (1.53)
18.1 (18.2)
0.0%
0.0%
3.39 (1.67)
15.6 (13.4)
0.0%
0.0%
3.02 (1.69)
13.1 (8.46)
0.0%
0.0%
39
LIST OF FIGURES
40
(a)
(b)
Figure 1: Illustration of the a) local moves and archive and b) global moves and
archive.
41
Pareto Front Comparison
30
Global
MACS
NSGA−II
25
20
T [h]
Part of the front not identified by NSGAII
15
10
Dominated tail
5
0
0
50
100
150
∆v
tot
200
[km/s]
250
300
350
(a)
Pareto Front Comparison
8
Global
MACS
NSGA−II
7
6
T [h]
5
4
Dominated tail
3
2
1
0
10
20
30
40
∆ vtot [km/s]
50
60
(b)
Figure 2: Three-impulse test case: a) Complete Pareto front, b) close-up of the
Pareto fronts
42
Pareto Front Comparison
7000
Global
MACS
NSGA−II
6000
T [day]
5000
4000
3000
2000
1000
0
50
100
150
200
250
f [km/s]
(a)
Pareto Front Comparison
Global
MACS
NSGA−II
6000
5500
T [day]
5000
4500
4000
3500
3000
2500
2000
4
5
6
7
f [km/s]
8
9
10
(b)
Figure 3: Cassini test case: a) Complete Pareto front, b) close-up of the Pareto
fronts
43
Pareto Front Comparison
7000
Global
NSGA−II
MACS
6000
T [day]
5000
4000
3000
2000
1000
0
50
100
150
f [km/s]
200
250
Close−up of the Cassini Pareto fronts
7000
Global
NSGA−II
MACS
T [day]
6000
5000
4000
3000
2000
5
5.5
6
6.5
f [km/s]
7
7.5
8
| 5cs.CE
|
Statistical Query Lower Bounds for Robust Estimation
of High-Dimensional Gaussians and Gaussian Mixtures
arXiv:1611.03473v2 [cs.LG] 17 May 2017
Ilias Diakonikolas∗
University of Southern California
[email protected]
Daniel M. Kane†
University of California, San Diego
[email protected]
Alistair Stewart
University of Southern California
[email protected]
May 18, 2017
Abstract
We describe a general technique that yields the first Statistical Query lower bounds for a range of
fundamental high-dimensional learning problems involving Gaussian distributions. Our main results are
for the problems of (1) learning Gaussian mixture models (GMMs), and (2) robust (agnostic) learning
of a single unknown Gaussian distribution. For each of these problems, we show a super-polynomial
gap between the (information-theoretic) sample complexity and the computational complexity of any
Statistical Query algorithm for the problem. Statistical Query (SQ) algorithms are a class of algorithms
that are only allowed to query expectations of functions of the distribution rather than directly access
samples. This class of algorithms is quite broad: a wide range of known algorithmic techniques in
machine learning are known to be implementable using SQs. Moreover, for the unsupervised learning
problems studied in this paper, all known algorithms with non-trivial performance guarantees are SQ or
are easily implementable using SQs.
Our SQ lower bound for Problem (1) is qualitatively matched by known learning algorithms for
GMMs. At a conceptual level, this result implies that – as far as SQ algorithms are concerned – the
computational complexity of learning GMMs is inherently exponential in the dimension of the latent
space – even though there is no such information-theoretic barrier. Our lower bound for Problem (2)
implies that the accuracy of the robust learning algorithm in [DKK+ 16] is essentially best possible
among all polynomial-time SQ algorithms. On the positive side, we also give a new (SQ) learning
algorithm for Problem (2) achieving the information-theoretically optimal accuracy, up to a constant
factor, whose running time essentially matches our lower bound. Our algorithm relies on a filtering
technique generalizing [DKK+ 16] that removes outliers based on higher-order tensors.
Our SQ lower bounds are attained via a unified moment-matching technique that is useful in other
contexts and may be of broader interest. Our technique yields nearly-tight lower bounds for a number
of related unsupervised estimation problems. Specifically, for the problems of (3) robust covariance
estimation in spectral norm, and (4) robust sparse mean estimation, we establish a quadratic statistical–
computational tradeoff for SQ algorithms, matching known upper bounds. Finally, our technique can
be used to obtain tight sample complexity lower bounds for high-dimensional testing problems. Specifically, for the classical problem of robustly testing an unknown mean (known covariance) Gaussian, our
technique implies an information-theoretic sample lower bound that scales linearly in the dimension.
Our sample lower bound matches the sample complexity of the corresponding robust learning problem
and separates the sample complexity of robust testing from standard (non-robust) testing. This separation
is surprising because such a gap does not exist for the corresponding learning problem.
∗
†
Supported by NSF Award CCF-1652862 (CAREER) and a Sloan Research Fellowship.
Supported by NSF Award CCF-1553288 (CAREER) and a Sloan Research Fellowship.
1 Introduction
1.1 Background and Overview For the unsupervised estimation problems considered here, the input is
a probability distribution which is accessed via a sampling oracle, i.e., an oracle that provides i.i.d. samples
from the underlying distribution. Statistical Query (SQ) algorithms are a restricted class of algorithms that
are only allowed to query expectations of functions of the distribution rather than directly access samples.
This class of algorithms is quite broad: a wide range of known algorithmic techniques in machine learning
are known to be implementable using SQs. These include spectral techniques, moment and tensor methods,
local search (e.g., Expectation Maximization), and many others (see, e.g., [CKL+ 06, FGR+ 13] for a detailed
discussion). Moreover, for the unsupervised learning problems studied in this paper, all known algorithms
with non-trivial performance guarantees are SQ or are easily implementable using SQs.
A number of techniques have been developed in information theory and statistics to characterize the
sample complexity of inference tasks. These involve both techniques for proving sample complexity upper
bounds (e.g., VC dimension, metric/bracketing entropy) and information-theoretic lower bounds (e.g., Fano
and Le Cam methods). On the other hand, computational lower bounds have been much more scarce in
the unsupervised setting. Perhaps surprisingly, it is possible to prove unconditional lower bounds on the
computational complexity of any SQ algorithm that solves a given learning problem. Given the ubiquity and
generality of SQ algorithms, an SQ lower bound provides strong evidence of the problem’s computational
intractability.
In this paper, we describe a general technique that yields the first Statistical Query lower bounds for a
range of fundamental high-dimensional learning problems involving Gaussian distributions. Such problems
are ubiquitous in applications across the data sciences and have been intensely investigated by different
communities of researchers for several decades. Our main results are for the problems of (1) learning Gaussian mixture models (GMMs), and (2) robust (agnostic) learning of a single unknown Gaussian distribution.
In particular, we show a super-polynomial gap between the (information-theoretic) sample complexity and
the computational complexity of any Statistical Query algorithm for these problems. In more detail, our SQ
lower bound for Problem (1) is qualitatively matched by known learning algorithms for GMMs (all of which
can be implemented as SQ algorithms). For Problem (2), we give a new (SQ) algorithm in this paper whose
running time nearly matches our SQ lower bound.
Our SQ lower bounds are attained via a unified moment-matching technique that is useful in other
contexts and may be of broader interest. Our technique yields nearly-tight lower bounds for a number of
related unsupervised estimation problems. Specifically, for the problems of (3) robust covariance estimation
in spectral norm, and (4) robust sparse mean estimation, we establish a quadratic statistical–computational
tradeoff for SQ algorithms, matching known upper bounds.
Finally, we use our technique to obtain tight sample complexity lower bounds for high-dimensional
testing problems. Specifically, for the classical problem of robustly testing an unknown mean (known covariance) Gaussian, our technique implies an information-theoretic lower bound that scales linearly in the
dimension. This lower bound matches the sample complexity of the corresponding robust learning problem
and separates the sample complexity of robust testing from standard (non-robust) testing. This separation is
surprising because such a gap does not exist for the corresponding learning problem.
Before we discuss our contributions in detail, we provide the necessary background for the Statistical
Query model and the unsupervised estimation problems that we study.
Statistical Query Algorithms. A Statistical Query (SQ) algorithm relies on an oracle that given any
bounded function on a single domain element provides an estimate of the expectation of the function on
a random sample from the input distribution. This computational model was introduced by Kearns [Kea98]
in the context of supervised learning as a natural restriction of the PAC model [Val84]. Subsequently, the
SQ model has been extensively studied in a plethora of contexts (see, e.g., [Fel16b] and references therein).
1
A recent line of work [FGR+ 13, FPV15, FGV15, Fel16a] developed a framework of SQ algorithms
for search problems over distributions – encompassing the distribution estimation problems we study in
this work. It turns out that one can prove unconditional lower bounds on the computational complexity
of SQ algorithms via the notion of Statistical Query dimension. This complexity measure was introduced
in [BFJ+ 94] for PAC learning of Boolean functions and was recently generalized to the unsupervised setting [FGR+ 13, Fel16a]. A lower bound on the SQ dimension of a learning problem provides an unconditional lower bound on the computational complexity of any SQ algorithm for the problem.
Remark. We would like to emphasize here that the SQ lower bounds shown in this paper apply to the
running time of an SQ algorithm and not on its sample complexity (when we simulate the SQ algorithm by
drawing samples to answer its SQ queries). Specifically, for all learning problems considered in this paper,
there exist straightforward SQ algorithms (that can be simulated with sample access to the distribution) with
near-optimal sample complexity, albeit with exponential running time. Specifically, lower bounds on the SQ
dimension of the corresponding problems establish lower bounds on the running time of any SQ algorithm
for the problem – not on its sample complexity.
Learning Gaussian Mixture Models. A mixture model is a convex combination of distributions of known
type. The most commonly studied case is a Gaussian mixture model (GMM). An n-dimensional
k-GMM
P
is a distribution in Rn that is composed of k unknown Gaussian components, i.e., F = ki=1 wi N (µi , Σi ),
where the weights wi , mean vectors µi , and covariance matrices Σi are unknown. The problem of learning
a GMM from samples has received tremendous attention in statistics and, more recently, in TCS. A long
line of work initiated by Dasgupta [Das99, AK01, VW02, AM05, KSV08, BV08] provides computationally efficient algorithms for recovering the parameters of a GMM under separability assumptions. Subsequently, efficient parameter learning algorithms have been obtained [MV10, BS10, HP15] under minimal
information-theoretic separation assumptions. The related problems of density estimation and proper learning have also been extensively studied [FOS06, SOAJ14, DK14, MV10, HP15, LS15]. In density estimation
(resp. proper learning), the goal is to output some hypothesis (resp. GMM) that is close to the unknown
mixture in total variation distance.
The sample complexity of density estimation (and proper learning) for n-dimensional k-GMMs, up
to variation distance ǫ, is easily seen to be poly(n, k, 1/ǫ) – without any assumptions.(In Appendix A,
we describe a simple SQ algorithm for this learning problem with sample complexity poly(n, k, 1/ǫ), albeit
exponential running time). Given that there is no information-theoretic barrier for learnability in this setting,
the following question arises: Is there a poly(n, k, 1/ǫ) time algorithm for density estimation (or proper
learning) of n-dimensional k-GMMs? This question has been raised as an open problem in a number of
settings (see, e.g., [Moi14, Dia16] and references therein).
For parameter learning, the situation is somewhat subtle: In full generality, the sample complexity
is of the form poly(n) · (1/γ)Ω(k) , where the parameter γ > 0 quantifies the “separation” between the
components. Even in one-dimension, a sample complexity lower bound of (1/γ)Ω(k) is known [MV10,
HP15]1 . The corresponding “hard” instances [MV10, HP15] consist of GMMs whose components have
large overlap, so many samples are required to distinguish between them. Is this the only obstacle towards
a poly(n, k) time parameter learning algorithm? Specifically, suppose that we are given an instance of the
problem with the additional promise that the components are “nearly non-overlapping” – so that poly(n, k)
samples suffice for the parameter learning problem as well. (In Appendix B, we show that when the total
variation distance between any pair of components in the given mixture is close to 1, parameter learning
reduces to proper learning; hence, there is a poly(n, k)-sample parameter learning (SQ) algorithm that runs
in exponential time.) Is there a poly(n, k) time parameter learning algorithm for such instances?
1
To circumvent the information-theoretic bottleneck of parameter learning, a related line of work has studied parameter learning
in a smoothed setting [HK13, BCMV14, ABG+ 14, GHK15].
2
In summary, the sample complexity of both versions of the learning problem is poly(n)f (k). On the
other hand, the running time of all known algorithms for either version scales as ng(k) , where g(k) ≥ k.
This runtime is super-polynomial in the sample complexity of the problem for super-constant values of k
and is tight for these algorithms, even for GMMs with almost non-overlapping components. The preceding
discussion is summarized in the following:
Question 1.1. Is there a poly(n, k)-time density estimation algorithm for n-dimensional k-GMMs? Is there
a poly(n, k)-time parameter learning algorithm for nearly non-overlapping n-dimensional k-GMMs?
Robust Learning of a Gaussian. In the preceding paragraphs, we were working under the assumption
that the unknown distribution generating the samples is exactly a mixture of Gaussians. The more general
and realistic setting of robust (or agnostic) learning – when our assumption about the model is approximately
true – turns out to be significantly more challenging. Specifically, until recently, even the most basic setting
of robustly learning an unknown mean Gaussian with identity covariance matrix was poorly understood.
Without corruptions, this problem is straightforward: The empirical mean gives a sample-optimal efficient
estimator. Unfortunately, the empirical estimate is very brittle and fails in the presence of corruptions.
The standard definition of agnostically learning a Gaussian (see, e.g., Definition 2.1 in [DKK+ 16] and
references therein) is the following: Instead of drawing samples from a perfect Gaussian, we have access to a
distribution D that is promised to be close to an unknown Gaussian G – specifically ǫ-close in total variation
distance. This is the only assumption about the distribution D, which may otherwise be arbitrary: the ǫfraction of “errors” can be adversarially selected. The goal of an agnostic learning algorithm is to output
a hypothesis distribution H that is as close as possible to G (or, equivalently, D) in variation distance.
Note that the minimum variation distance, dTV (H, G), information-theoretically achievable under these
assumptions is Θ(ǫ), and we would like to obtain a polynomial-time algorithm with this guarantee.
Agnostically learning a single high-dimensional Gaussian is arguably the prototypical problem in robust
statistics [Hub64, HRRS86, HR09]. Early work in this field [Tuk75, DG92] studied the sample complexity
of robust estimation. Specifically, for the case of an unknown mean and known covariance Gaussian, the
Tukey median [Tuk75] achieves O(ǫ)-error with O(n/ǫ2 ) samples (see, e.g., [CGR15] for a simple proof).
Since Ω(n/ǫ2 ) samples are information-theoretically necessary – even without noise – the robustness requirement does not change the sample complexity of the problem.
The computational complexity of agnostically learning a Gaussian is less understood. Until recently,
√
all known polynomial time estimators could only guarantee error of Θ(ǫ n). Two recent works [DKK+ 16,
LRV16] made a first step in designing robust polynomial-time estimators for this problem. The results
of [DKK+ 16] apply in the standard agnostic model; [LRV16] works in a weaker model – known as Huber’s contamination model [Hub64] – where the noisy distribution D is of the form (1 − ǫ)G + ǫN ,
where N is an unknown “noise” distribution. For the problem of
√ robustly estimating +an unknown mean
log n), while [DKK 16] obtains error
Gaussian
N
(µ,
I),
[LRV16]
obtains
an
error
guarantee
of
O(ǫ
p
O(ǫ log(1/ǫ)), independent of the dimension2 .
A natural and important open problem, put forth by these works [DKK+ 16, LRV16], is the following:
Question 1.2. Is there a poly(n/ǫ)- time agnostic learning algorithm, with error O(ǫ), for an n-dimensional
Gaussian?
Statistical–Computational Tradeoffs. A statistical–computational tradeoff refers to the phenomenon that
there is an inherent gap between the information-theoretic sample complexity of a learning problem and its
computational sample complexity, i.e, the minimum sample complexity attainable by any polynomial time
2
The algorithm
p of [LRV16] can be extended to work in the standard agnostic model at the expense of an increased error
guarantee of O(ǫ log n log(1/ǫ)).
3
algorithm for the problem. The prototypical example is the estimation of a covariance matrix under sparsity
constraints (sparse PCA) [JL09, CMW13, CMW15], where a nearly-quadratic gap between informationtheoretic and computational sample complexity has been established (see [BR13b, WBS16b]) – assuming
the computational hardness of the planted clique problem.
For a number of high-dimensional learning problems (including the problem of robustly learning a
Gaussian under the total variation distance), it is known that the robustness requirement does not change
the information-theoretic sample complexity of the problem. On the other hand, it is an intriguing possibility that injecting noise into a high-dimensional learning problem may change its computational sample
complexity.
Question 1.3. Does robustness create inherent statistical–computational tradeoffs for natural high-dimensional
estimation problems?
In this work, we consider two natural instantiations of the above general question: (i) robust estimation
of the covariance matrix in spectral norm, and (ii) robust sparse mean estimation. We give basic background
for these problems in the following paragraphs.
For (i), suppose we have sample access to a (zero-mean) n-dimensional unknown-covariance Gaussian, and we want to estimate the covariance matrix with respect to the spectral norm. It is known (see, e.g.,
[Ver12]) that O(n/ǫ2 ) samples suffice so that the empirical covariance is within spectral error at most ǫ from
the true covariance; and this bound is information-theoretically optimal, to constant factors, for any estimator. For simplicity, let us assume that the desired accuracy is a small positive constant, e.g., ǫ = 1/10. Now
suppose that we observe samples from a corrupted Gaussian in Huber’s contamination model (the weaker
adversarial model) where the noise rate δ ≪ 1/10. First, it is not hard to see that the injection of noise
does not change the information-theoretic sample complexity of the problem: there exist (computationally
inefficient) robust estimators (see, e.g., [CGR15]) that use O(n) samples. (There is a straightforward SQ
algorithm for this problem as well that uses O(n) samples, but again runs in exponential time.) On the other
hand, if we are willing to use Õ(n2 ) samples, a polynomial-time robust estimator with constant spectral
error guarantee is known [DKK+ 16, DKK+ 17a]3 . The immediate question that follows is this:
Is there a computationally efficient robust covariance estimator in spectral error that uses a
strongly sub-quadratic sample size, i.e., O(n2−c ) for a constant 0 < c < 1?
For (ii), suppose we want to estimate the mean µ ∈ Rn of an identity covariance Gaussian up to
ℓ2 -distance ǫ, under the additional promise that µ is k-sparse, and suppose that k ≪ n1/2 . It is wellknown that the information-theoretic sample complexity of this problem is O(k log n/ǫ2 ), and the truncated empirical mean achieves the optimal bound. For simplicity, let us assume that ǫ = 1/10. Now suppose that we observe samples from a corrupted sparse mean Gaussian (in Huber’s contamination model),
where the noise rate δ ≪ 1/10. As in the setting of the previous paragraph, the injection of noise does
not change the information-theoretic sample complexity of the problem: there exist a (computationally
inefficient) robust SQ algorithm for this problem (see [Li17]) that use O(k log n) samples. Two recent
works [Li17, DBS17] gave polynomial time robust algorithms for robust sparse mean estimation with sample complexity Õ(k2 log n). In summary, in the absence of robustness, the information-theoretically optimal
sample bound is known to be achievable by a computationally efficient algorithm. In contrast, in the presence of robustness, there is a quadratic gap between the information-theoretic optimum and the sample
complexity of known polynomial-time algorithms. The immediate question is whether this gap is inherent:
We note that the robust covariance estimators of [DKK+ 16, DKK+ 17a] provide error guarantees under the Mahalanobis
distance, which is stronger than the spectral norm. Under the stronger metric, Ω(n2 ) samples are information-theoretically required
even without noise.
3
4
Is there a computationally efficient robust k-sparse mean estimator that uses a strongly subquadratic sample size , i.e., O(k2−c ) for a constant 0 < c < 1?
It is conjectured in [Li17] that a quadratic gap is in fact inherent for efficient algorithms.
High-Dimensional Hypothesis Testing. So far, we have discussed the problem of learning an unknown
distribution that is promised to belong (exactly or approximately) in a given family (Gaussians, mixtures
of Gaussians). A related inference problem is that of hypothesis testing [NP33, LR05]: Given samples
from a distribution in a given family, we want to distinguish between a null hypothesis and an alternative
hypothesis. Starting with [GR00, BFR+ 00], this broad question has been extensively investigated in TCS
with a focus on discrete probability distributions. A natural way to solve a distribution testing problem is to
learn the distribution in question to good accuracy and then check if the corresponding hypothesis is close to
one satisfying the null hypothesis. This testing-via-learning approach is typically suboptimal and the main
goal in this area has been to obtain testers with sub-learning sample complexity.
In this paper, we study natural hypothesis testing analogues of the high-dimensional learning problems
discussed in the previous paragraphs. Specifically, we study the sample complexity of (i) robustly testing an
unknown mean Gaussian, and (ii) testing a GMM.
To motivate (i), we consider arguably the most basic high-dimensional testing task: Given samples from
a Gaussian N (µ, I), where µ ∈ Rn is unknown, distinguish between the case that µ = 0 versus kµk2 ≥ ǫ.
(The latter condition is equivalent, up to constant factors, to dTV (N (µ, I), N (0, I)) ≥ ǫ.) The classical test
for this task is Hotelling’s T-squared statistic [Hot31], which is unfortunately not defined when the sample
size is smaller than the dimension [ZB96]. More recently, testers that succeed in the sub-linear regime have
been developed [SD08] (also see [ZB96, CQ10]). In Appendix C, we give a simple and natural tester for
√
this problem that uses O( n/ǫ2 ) samples, and show that this sample bound is information-theoretically
optimal, up to constant factors.
Now suppose that our Gaussianity assumption about the unknown distribution is only approximately
satisfied. Formally, we are given samples from a distribution D on Rn which is promised to be either (a)
a standard Gaussian N (0, I), or (b) a δ-noisy version of N (µ, I), where µ ∈ Rn satisfies kµk2 ≥ ǫ, and
the noise rate δ satisfies δ ≪ ǫ. The robust hypothesis testing problem is to distinguish, with high constant
probability, between these two cases. Note that condition (b) implies that dTV (D, N (0, I)) = Ω(ǫ), and
therefore the two cases are distinguishable. 4
Robust hypothesis testing is of fundamental importance and has been extensively studied in robust
statistics [HR09, HRRS86, Wil97]. Perhaps surprisingly, it is poorly understood in the most basic settings,
even information-theoretically. Specifically, the sample complexity of our aforementioned robust mean
testing problem has remained open. It is easy to see that the tester of Appendix C fails in the robust setting.
On the other hand, the testing-via-learning approach implies a sample upper bound of O(n/ǫ2 ) for our
robust testing problem – by using, e.g., the Tukey median. The following question arises:
Question 1.4. Is there an information-theoretic gap between robust testing and non-robust testing? What is
the sample complexity of robustly testing the mean of a high-dimensional Gaussian?
We conclude with our hypothesis testing problem regarding GMMs: Given samples from a distribution D
on Rn , we want to distinguish between the case that D = N (0, I), or D is a 2-mixture of identity covariance
Gaussians. This is a natural high-dimensional testing problem that we believe merits investigation in its own
right. The obvious open question here is whether there exists a tester for this problem with sub-learning
sample complexity.
4
Robust testing should not be confused with tolerant testing, where the completeness is relaxed. In our context, tolerant testing
corresponds to distinguishing
√ between dTV (D, N (0, I)) ≤ ǫ/2 versus dTV (D, N (0, I)) ≥ ǫ, where D = N (µ, I), and is easily
seen to be solvable with O( n/ǫ2 ) samples as well.
5
1.2 Our Results The main contribution of this paper is a general technique to prove lower bounds for
a range of high-dimensional estimation problems involving Gaussian distributions. We use analytic and
probabilistic ideas to construct explicit families of hard instances for the estimation problems described in
Section 1.1. Using our technique, we prove super-polynomial Statistical Query (SQ) lower bounds that
answer Questions 1.1 and 1.2 in the negative for the class of SQ algorithms. We also show that the observed
quadratic statistical–computational gap for robust sparse mean estimation and robust spectral covariance
estimation is inherent for SQ algorithms. As an additional important application of our technique, we obtain
information-theoretic lower bounds on the sample complexity of the corresponding testing problems. (We
note that our testing lower bounds apply to all algorithms.) Specifically, we answer Question 1.4 in the
affirmative, by showing that the robustness requirement makes the Gaussian testing problem informationtheoretically harder. In the body of this section, we state our results and elaborate on their implications and
the connections between them.
SQ Lower Bound for Learning GMMs. Our first main result is a lower bound of nΩ(k) on the complexity
of any SQ algorithm that learns an arbitrary n-dimensional k-GMM to constant accuracy (see Theorem 4.1
for the formal statement):
Theorem 1.1 (SQ Lower Bound for Learning GMMs). Any SQ algorithm that learns an arbitrary nΩ(1)
≥ nΩ(k) queries to an
dimensional k-GMM to constant accuracy, for all n ≥ poly(k), requires 2n
−O(k)
SQ oracle of precision n
.
Theorem 1.1 establishes a super-polynomial gap between the information-theoretic sample complexity of
learning GMMs and the complexity of any SQ learning algorithm for this problem. It is worth noting that
our hard instance is a family of high-dimensional
GMMs whose components are almost non-overlapping.
P
Specifically, for each GMM F = ki=1 wi N (µi , Σi ) in the family, the total variation distance between any
pair of Gaussian components can be made as large as 1 − 1/poly(n, k). More specifically, for our family of
hard instances, the sample complexity of both density and parameter learning is Θ(k · log n) (the standard
cover-based algorithm that achieves this sample upper bound is SQ). In contrast, any SQ learning algorithm
for this family of instances requires runtime at least nΩ(k) .
At a conceptual level, Theorem 1.1 implies that – as far as SQ algorithms are concerned – the computational complexity of learning high-dimensional GMMs is inherently exponential in the dimension of the
latent space – even though there is no such information-theoretic barrier in general. Our SQ lower bound
identifies a common barrier of the strongest known algorithmic approaches for this learning problem, and
provides a rigorous explanation why a long line of algorithmic research on this front either relied on strong
separation assumptions or resulted in runtimes of the form nΩ(k) .
SQ Lower Bound for Robustly Learning a Gaussian. Our second main result concerns the agnostic
learning of a single n-dimensional Gaussian. We prove two SQ lower bounds with qualitatively similar
guarantees for different versions of this problem. Our first lower bound is for the problem of agnostically
learning a Gaussian with unknown mean and identity covariance. Roughly speaking, we show that any SQ
1/4
algorithm that solves this learning problem to accuracy O(ǫ) requires complexity nΩ(log (1/ǫ)) . We show
(see Theorem 5.1 for a more detailed statement):
Theorem 1.2 (SQ Lower Bound for Robust Learning of Unknown Mean Gaussian). Let ǫ > 0, 0 < c ≤ 1/2,
and n ≥ poly(log(1/ǫ)). Any SQ algorithm that robustly learns an n-dimensional Gaussian N (µ, I),
c/2
Ω(1)
≥ nΩ(log(1/ǫ) ) queries to an SQ oracle
within total variation distance O(ǫ log(1/ǫ)1/2−c ), requires 2n
c/2
of precision n−Ω(log(1/ǫ) ) .
Some comments are in order. First, Theorem 1.2 shows a super-polynomial gap between the sample complexity of agnostically learning an unknown mean Gaussian and the complexity of SQ learning algorithms
6
for this problem. As mentioned in the introduction, O(n/ǫ2 ) samples information-theoretically suffice to
agnostically learn an unknown mean Gaussian to within error O(ǫ). Second, the robust learning algo+
rithm
SQ model, and achieves error
pof [DKK 16] runs in poly(n, 1/ǫ) time, can be implemented in thep
O(ǫ log(1/ǫ)). As a corollary of Theorem 1.2, we obtain that the O(ǫ log(1/ǫ)) error guarantee of
the [DKK+ 16] algorithm is best possible among all polynomial-time SQ algorithms.
Roughly speaking, Theorem 1.2 shows that any SQ algorithm that solves the (unknown mean Gaussian)
1/4
robust learning problem to accuracy O(ǫ) needs to have running time at least nΩ(log (1/ǫ)) , i.e., quasipolynomial in 1/ǫ. It is natural to ask whether this quasi-polynomial lower bound can be improved to, say,
exponential, e.g., nΩ(1/ǫ) . We show
√ that the lower bound of Theorem 1.2 is qualitatively tight. We design
an (SQ) algorithm that uses Oǫ (n log(1/ǫ) ) SQ queries of inverse quasi-polynomial precision. Moreover,
we can turn this SQ algorithm into an algorithm in the sampling oracle model with similar complexity.
Specifically, we show (see Theorem 8.7 and Corollary 8.8):
Theorem 1.3 (SQ Algorithm for Robust Learning of Unknown Mean Gaussian). Let D be a distribution
on
√
O( log(1/ǫ))
n
n
R such that dTV (D, N (µ, I))
)
√ ≤ ǫ for some µ ∈ R . There is an SQ algorithm that uses Oǫ (n
O( log(1/ǫ))
n
, and outputs µ
e ∈ R such that dTV (N (e
µ, I), N (µ, I)) ≤ O(ǫ). The
SQ’s to D of precision ǫ/n
SQ algorithm can be turned into an algorithm√(in the sample model) with the same error guarantee that has
sample complexity and running time Oǫ (nO( log(1/ǫ)) ).
Theorems 1.2 and 1.3 give a qualitatively tight characterization of the complexity of robustly learning
an unknown mean Gaussian in the standard agnostic model, where the noisy distribution D is such that
dTV (D, N (µ, I)) ≤ ǫ. Equivalently, D satisfies (1 − ǫ1 )D + ǫ1 N1 = (1 − ǫ2 )N (µ, I) + ǫ2 N2 , where
N1 , N2 are unknown distributions and ǫ1 + ǫ2 ≤ ǫ. A weaker error model, known as Huber’s contamination
model in the statistics literature [Hub64, HRRS86, HR09], prescribes that the noisy distribution D is of the
form D = (1−ǫ)N (µ, I)+ǫN , where N is an unknown distribution. Intuitively, the difference is that in the
former model the adversary is allowed to subtract good samples and add corrupted ones, while in the latter
the adversary is only allowed to add corrupted ones. We note that the lower bound of Theorem 1.2 does
not apply in Huber’s contamination model. This holds for a reason: Concurrent work [DKK+ 17b] gives
a poly(n/ǫ) time algorithm with O(ǫ) error for robustly learning N (µ, I) in Huber’s model. Hence, as a
corollary, we establish a computational separation between these two models of corruptions. We provide an
intuitive justification in Section 1.3.
Our second super-polynomial SQ lower bound is for the problem of robustly learning a zero-mean
unknown covariance Gaussian with respect to the spectral norm. Specifically, we show (see Theorem 5.12
for a detailed statement):
Theorem 1.4 (SQ Lower Bound for Robust Learning of Unknown Covariance Gaussian). Let ǫ > 0, 0 <
c ≤ 1, and n ≥ poly(log(1/ǫ)). Any SQ algorithm that, given access to an ǫ-corrupted n-dimensional
e with kΣ
e − Σk2 ≤ O(ǫ log(1/ǫ)1−c ), requires at least
Gaussian N (0, Σ), with I/2 Σ 2I, returns Σ
c/4
c/4
Ω(1)
≥ nΩ(log(1/ǫ) ) queries to an SQ oracle of precision n−Ω(log(1/ǫ) ) .
2n
Similarly, Theorem 1.4 shows a super-polynomial gap between the information-theoretic sample complexity and the complexity of any SQ algorithm for this problem. As mentioned in the introduction, O(n/ǫ2 )
samples information-theoretically suffice to agnostically learn the covariance to within spectral error O(ǫ).
Second, the robust learning algorithm of [DKK+ 16] runs in poly(n, 1/ǫ) time, can be implemented in the
SQ model, and achieves error O(ǫ log(1/ǫ)) in Mahalanobis distance (hence, also in spectral norm). Again,
the immediate corollary is that the O(ǫ log(1/ǫ)) error guarantee of the [DKK+ 16] algorithm is best possible among all polynomial-time SQ algorithms. The lower bound of Theorem 1.4 does not apply in Huber’s
contamination model. This holds for a reason: [DKK+ 17b] gives a poly(n) · 2poly log(1/ǫ) time algorithm
with O(ǫ) error in Huber’s model.
7
Statistical–Computational Tradeoffs for SQ algorithms. Our next SQ lower bounds establish nearly
quadratic statistical–computational tradeoffs for robust spectral covariance estimation and robust sparse
mean estimation. We note that both these lower bounds also hold in Huber’s contamination model. For
the former problem, we show (see Theorem 6.1 for the formal statement):
Theorem 1.5. Let 0 < c < 1/6, and n sufficiently large. Any SQ algorithm that, given access to an
ǫ-corrupted N (0, Σ), where ǫ ≤ c/ ln(n) for kΣk2 ≤ poly(n/ǫ), and returns Σ̃ with Σ̃/2 Σ 2Σ̃,
c/3
requires at least 2Ω(n ) queries to an SQ oracle of precision γ = O(n)−(1−5c/2) .
We note that, in order to simulate a single query of the above precision, we need to draw Ω(1/γ 2 ) =
Ω(n2−5c ) samples from our distribution. Roughly speaking, Theorem 1.5 shows that if an SQ algorithm
c/3
uses less than this many samples, then it needs to run in 2Ω(n ) time. This suggests a nearly-quadratic
statistical-computational tradeoff for this problem.
For robust sparse mean estimation we show (see Theorem 6.6 for the detailed statement):
Theorem 1.6. Fix any 0 < c < 1 and let n ≥ 8k2 . Any SQ algorithm that, given access to an ǫ-corrupted
N (µ, I), where ǫ = k−c/4 , and µ ∈ Rn is promised to be k-sparse with kµk2 = 1, and outputs a hypothesis
c
vector µ
b satisfying kb
µ − µk2 ≤ 1/2, requires at least nΩ(ck ) queries to an SQ oracle of precision γ =
O(k)3c/2−1 .
Similarly, to simulate a single query of the above precision, we need to draw Ω(1/γ 2 ) = Ω(k2−3c )
samples from our distribution. Hence, any SQ algorithm that uses this many samples requires runtime at
c
least nΩ(ck ) . This suggests a nearly-quadratic statistical-computational tradeoff for this problem.
Sample Complexity Lower Bounds for High-Dimensional Testing. We now turn to our informationtheoretic lower bounds on the sample complexity of the corresponding high-dimensional testing problems.
For the robust Gaussian mean testing problem in Huber’s contamination model, we show (see Theorem 7.5
for a more detailed statement):
Theorem 1.7 (Sample Complexity Lower Bound for Robust Testing of Unknown Mean Gaussian). Fix
ǫ > 0. Any algorithm with sample access to a distribution D on Rn which satisfies either (a) D = N (0, I)
or (b) D is a δ-noisy N (µ, I), and kµk2 ≥ ǫ, and distinguishes between the two cases with probability 2/3
requires (i) Ω(n) samples if δ = ǫ/100, (ii) Ω(n1−c ) samples if δ = ǫ/nc/4 , for any constant 0 < c < 1.
As stated in the Introduction, without the robustness requirement, for any constant ǫ > 0, the Gaussian
√
mean testing problem can be solved with Oǫ ( n) samples. Hence, the conceptual message of Theorem 1.7
is that robustness makes the Gaussian mean testing problem information-theoretically harder. In particular,
the sample complexity of robust testing is essentially the same as that of the corresponding learning problem.
Theorem 1.7 can be viewed as a surprising fact because it implies that the effect of robustness can be very
different for testing versus learning of the same distribution family. Indeed, recall that the sample complexity
of robustly learning an ǫ-corrupted unknown mean Gaussian, up to error O(ǫ), is O(n/ǫ2 ) – i.e., the same
as in the noiseless case.
As a final application of our techniques, we show a sample complexity lower bound for the problem of
testing whether a spherical GMM is close to a Gaussian (see Theorem 7.6 for the detailed statement):
Theorem 1.8 (Sample Complexity Lower Bound for Testing a GMM). Any algorithm with sample access to
a distribution D on Rn which satisfies either (a) D = N (0, I), or (b) D = (1/2)N (µ1 , I)+ (1/2)N (µ2 , I),
such that dTV (D, N (0, I)) ≥ ǫ, and distinguishes between the two cases with probability at least 2/3
requires Ω(n/ǫ2 ) samples.
Similarly, the sample lower bound of Theorem 1.8 is optimal, up to constant factors, and coincides with
the sample complexity of learning the underlying distribution.
8
1.3 Our Approach and Techniques In this section, we provide a detailed outline of our approach and
techniques. The structure of this section is as follows: We start by describing our Generic Lower Bound
Construction, followed by our main applications to the problems of Learning GMMs and Robustly Learning
an Unknown Gaussian. We continue with our applications to statistical–computational tradeoffs. We then
explain how our generic technique can be used to obtain our Sample Complexity Testing Lower Bounds,
which rely on essentially the same hard instances as our SQ lower bounds. We conclude with a sketch of
our new (SQ) Algorithm for Robustly Learning an Unknown Mean Gaussian to optimal accuracy.
Generic Lower Bound Construction. The main idea of our lower bound construction is quite simple: We
construct a family of distributions D that are standard Gaussians in all but one direction, but are somewhat
different in the remaining direction (Definition 3.1). Effectively, we are hiding the interesting information
about our distributions in this unknown choice of direction. By exploiting the simple fact that it is possible
to find exponentially many nearly-orthogonal directions (Lemma 3.7), we are able to show that any SQ
algorithm with insufficient precision needs many queries in order to learn an unknown distribution from D.
To prove our generic SQ lower bound, we need to bound from below the SQ-dimension of our hard
family of distributions D. Roughly speaking, the SQ-dimension of a distribution family (Definition 2.11)
corresponds to the number of nearly uncorrelated distributions (with respect to some fixed distribution) in
the family (see Definitions 2.9 and 2.10). It is known that a lower bound on the SQ-dimension implies a
corresponding lower bound on the number and precision of queries of any SQ algorithm (see Lemma 2.12).
More concretely, our hard families of distributions are constructed as follows: Given a distribution A on
the real-line, we define a family of high-dimensional distributions Pv (x), for v ∈ Sn a unit n-dimensional
vector. The distribution Pv gives a copy of A in the v-direction, while being an independent standard
Gaussian in the orthogonal directions (Definition 3.1). Our hard family will be the set D = {Pv | v ∈ Sn }.
For the sake of the intuition, we make two observations: (1) If A and N (0, 1) have substantially different
moments of degree at most m, for some m, then Pv and N (0, I) can be easily distinguished by comparing
their mth -order moment tensors. Since these tensors can be approximated in roughly nm queries (and
time), the aforementioned lower bound construction would necessarily fail unless the low-order moments
of A match the corresponding low-order moments of G. We show that, aside from a few mild technical
conditions (see Condition 3.2), this moment-matching condition is essentially sufficient for our purposes.
If the degree at most m tensors agree, we need to approximate tensors of degree m + 1. Intuitively, in
order to extract useful information from these higher degree tensors, one needs to approximate essentially
all of the nm+1 many such tensor entries. (2) A natural approach to distinguish between Pv and N (0, I)
would be via random projections. As a critical component of our proof, we show (see Lemma 3.5) that a
random projection of Pv will be exponentially close to N (0, 1) with high probability. Therefore, a random
projection-based algorithm would require exponentially many random directions until it found a good one.
We now proceed with a somewhat more technical description of our proof. To bound from below the
SQ-dimension of our hard family of distributions, we proceed
R as follows: The definition of the pairwise
correlation (Definition 2.9) implies we need to show that Pv Pv′ /G ≈ 1, where G ∼ N (0, I) is the
Gaussian measure, for any pair of unit vectors v, v ′ that are nearly orthogonal. To prove this fact, we make
essential use of the Gaussian (Ornstein–Uhlenbeck) noise operator and its properties (see, e.g., [O’D14]).
We explain this connection in the following paragraph.
By construction of the distributions Pv , Pv′ , it follows that in the directions perpendicular to both v and
v ′ , the relevant factors integrate to 1. Letting y = v · x and z = v ′ · x and letting y ′ , z ′ be the orthogonal
directions to y and z, we need to consider the integral
Z
A(y)A(z)G(y ′ )G(z ′ )/G(x) .
9
Fixing y and integrating over the orthogonal direction, we get
Z
Z
A(y)/G(y) A(z)G(z ′ )dy ′ .
Now, if v and v ′ are (exactly) orthogonal, z = y ′ and the inner integral equals G(y). When this is not the
case, the A(z) term is not quite vertical and the G(z ′ ) term not quite horizontal, so instead what we get is
only nearly Gaussian. In general, the inner integral is equal to
Uv·v′ A(y) ,
√
where Ut is the member of the Ornstein–Uhlenbeck semigroup, Ut f (z) = E[f (tz + 1 − t2 G)]. We show
that this quantity is close to a Gaussian, when v · v ′ is close to 0 (see Lemma 3.4).
The core idea of the analysis relies on the fact that Ut A is a smeared out version of A. As such, it
only retains the most prominent features of A, namely its low-order moments. In fact, we are able to show
that if A and G agree in their first m moments, then Ut A is Om (tm )-close to a Gaussian (see Lemma 3.5),
and thus the integral in question is Om ((|v · v ′ |)m )-close to 1. This intuition is borne out in a particularly
clean way by writing A/G in the basis of Hermite polynomials. The moment-matching condition implies
that the decomposition involves none of the Hermite polynomials of degrees 1 through m. However, the
Ornstein–Uhlenbeck operator, Ut , is diagonalized by the basis Hi G with eigenvalue ti . Thus, if A − G can
be written in this basis with no terms of degree less than m, applying Ut decreases the size of the function
by a multiple of approximately tm .
So far, we have provided a proof sketch of the following statement (Lemma 3.4): When two unit vectors
′
v, v are nearly orthogonal, then the distributions Pv , Pv′ are nearly uncorrelated. Since, for 0 < c <
c
1/2, we can pack 2Ω(n ) unit vectors v onto the sphere so that their pairwise inner products are at most
nc−1/2 (Lemma 3.7), we obtain an SQ-dimension lower bound of our hard family. In particular, to learn
c
the distribution Pv , for unknown v, any SQ algorithm requires either 2Ω(n ) queries or queries of accuracy
better than O(n)(m+1)(c−1/2) (Proposition 3.3). This completes the proof sketch of our generic construction.
In our main applications, we construct one-dimensional distributions A satisfying the necessary momentmatching conditions for m taken to be super-constant, thus obtaining super-polynomial SQ lower bounds.
For our quadratic statistical–computational tradeoffs, we match a constant number of moments. In the following paragraphs, we explain how we apply our framework to bound the SQ dimension for: (i) learning
k-GMMs to constant accuracy, (ii) robustly learning an ǫ-corrupted Gaussian to accuracy O(ǫ), and (iii)
robustly estimating a Gaussian covariance within constant spectral error and robustly estimating a sparse
Gaussian mean to constant ℓ2 -error. In all cases, we construct a distribution A on the real-line that satisfies the necessary moment-matching conditions such that the family D = {Pv | v ∈ Sn } belongs in the
appropriate class, e.g., is a k-GMM for (i), an ǫ-corrupted Gaussian for (ii), etc.
SQ Lower Bound for Learning k-GMMs. The properties of our one-dimensional distribution A are summarized in Proposition 4.2. Specifically, we construct a distribution A on the real line that is a k-mixture
of one-dimensional “skinny” Gaussians, Ai , that agrees with N (0, 1) on the first m = 2k − 1 moments
(condition (i)). For technical reasons, we require that the chi-squared divergence of A to N (0, 1) is bounded
from above by an appropriate quantity (condition (iv)). The Gaussian components, Ai , have the same variance and appropriately bounded means (condition (ii)). We can also guarantee that the components Ai are
almost non-overlapping (condition (iii)). This implies that the corresponding high-dimensional distributions
Pv , P′v will be at total variation distance close to 1 from each other when the directions v, v ′ are nearly
orthogonal, and moreover their means will be sufficiently separated.
To establish the existence of a distribution A with the above properties, we proceed in two steps: First,
we construct (Lemma 4.3) a discrete one-dimensional distribution B supported on k points, lying in an
10
√
O( k) length interval, that agrees with N (0, 1) on the first k moments. The existence of such a distribution
B essentially follows from standard tools on Gauss-Hermite quadrature. The distribution A is then obtained
(Corollary 4.4) by adding a zero-mean skinny Gaussian to an appropriately rescaled version of B. Additional
technical work (Lemmas 4.5 and 4.6) gives the other conditions.
Our family of hard high-dimensional instances will consist of GMMs that look like almost non-overlapping
“parallel pancakes” and is reminiscent of the family of instances considered in Brubaker and Vempala [BV08].
For the case of k = 2, consider a 2-GMM where both components have the same covariance that is far from
spherical, the vector between the means is parallel to the unit eigenvector with smallest eigenvalue, and
the distance between the means is a large multiple of the standard deviation in this direction (but a small
multiple of that in the orthogonal direction). This family of instances was considered in [BV08], who gave
an efficient spectral algorithm to learn them.
Our lower bound construction can be thought of as k “parallel pancakes” in which the means lie in a
one-dimensional subspace, corresponding to the smallest eigenvalue of the identical covariance matrices of
the components. All n − 1 orthogonal directions will have an eigenvalue of 1, which is much larger than
the smallest eigenvalue. In other words, for each unit vector v, the k-GMM Pv will consist of k “skinny”
Gaussians whose mean vectors all lie in the direction of v. Moreover, each pair of components
will have total
√
variation distance very close to 1 and their mean vectors are separated by Ω(1/ k). We emphasize once
more that our hard family of instances is learnable with O(k log n) samples – both for density estimation
and parameter estimation. On the other hand, any SQ learning algorithm for the family requires nΩ(k) time.
SQ Lower Bounds for Robustly Learning Unknown Gaussian. In the agnostic model, there are two
types of adversarial noise to handle: subtractive noise – corresponding to the good samples removed by
the adversary – and additive noise – corresponding to the bad points added by the adversary. The approach
of [DKK+ 16] does not do anything
p to address subtractive noise, but shows that this type of noise can incur
“small” error, e.g., at most O(ǫ log(1/ǫ)) for the case of unknown mean. For additive noise, [DKK+ 16]
uses an iterative spectral algorithm to filter out outliers.
For concreteness, let us consider the case of robustly learning N (µ, I). Intuitively, achieving error O(ǫ)
in the agnostic model is hard for the following reason: the two types of noise can collude so that the first few
moments of the p
corrupted distribution are indistinguishable from those of a Gaussian whose mean vector
has distance Ω(ǫ log(1/ǫ)) from the true mean.
To formalize this intuition, for our robust SQ learning lower bound, we construct a distribution A on
the real line that agrees with N (0, 1) on the first m = Ω(log1/4 (1/ǫ)) moments and is ǫ/100-close in
total variation distance to G′ = N (ǫ, 1) (see Proposition 5.2). We achieve this by taking A to be the
Gaussian N (ǫ, 1) outside its effective support, while in the effective support we add an appropriate degreem univariate polynomial p satisfying the appropriate moment conditions. By expressing this polynomial as
a linear combination of appropriately scaled Legendre polynomials, we can prove that its L1 and L∞ norms
within the effective support of G′ are much smaller than ǫ (see Lemma 5.6). This result is then used to
bound from above the distance of A from G′ , which gives our SQ lower bound.
We use a similar technique to prove our SQ lower bound for robust covariance estimation in spectral
norm. Specifically, we construct a distribution A that agrees with N (0, 1) on the first m = Ω(log(1/ǫ))
moments and is ǫ/100-close in total variation distance to G′ = N (0, (1 − δ)2 ), for some δ = O(ǫ) (see
Proposition 5.13). We similarly take A to be the Gaussian G′ outside its effective support, while in the
effective support we add an appropriate degree-m univariate polynomial p satisfying the appropriate moment
conditions. The analysis proceeds similarly as above.
Statistical–Computational Tradeoffs for SQ algorithms. For robust covariance estimation in spectral
norm, our one-dimensional distribution is selected to be A = (1 − ǫ)N (0, σ) + ǫN1 , where N1 is a mixture
of 2 unit-variance Gaussians with opposite means. By selecting σ appropriately, we can have A match
the first 3 moments of N (0, 1), see Theorem 6.1. For robust sparse mean estimation, it suffices to take
11
A = (1 − δ)N (ǫ, 1) + δN1 , where N1 is a unit-variance Gaussian selected so that E[A] = 0. An important
aspect of both these constructions is that the chi-squared distance χ2 (A, N (0, 1)) needs to be as small as
possible. Indeed, since we only match a small number of moments, our bound on χ2 (A, N (0, 1)) crucially
affects the accuracy of our SQ queries (Proposition 3.3).
Sample Complexity Testing Lower Bounds. Our sample complexity lower bounds follow from standard
information-theoretic
arguments, and rely on the same lower bound instances and correlation bounds (i.e.,
R
bounds on Pv Pv′ /G) established in our SQ lower bounds. In particular, we consider the problem of distinguishing between the distribution G ∼ N (0, I) and the distribution Pv for a randomly chosen unit vector
v ∈ Sn using N independent samples. Let G⊗N denote the distribution on N independent samples from G,
and P⊗N
the distribution obtained by picking a random v and then taking N independent samples from Pv .
v
If it is possible to reliably distinguish between these cases, itR must be
QNthe case that the chi-squared diver⊗N
⊗N
gence χ(Pv , G ) is substantially larger than 1. This is v,v′ ,xi i=1 Pv (xi )Pv′ (xi )/G(xi )dvdv ′ dxi .
Note that after fixing v and v ′ the above integral separates as a product, giving
Z
v,v′
Z
Pv (x)Pv′ (x)/G(x)dx
N
dvdv ′ .
(1)
Note that the inner integral was bounded from above by roughly (1 + (v · v ′ )m ). A careful analysis of the
distribution of the angle between two random unit vectors allows us to show that, unless N = Ω(n), the
chi-squared divergence is close to 1, and thus that this testing problem is impossible.
Algorithm for Robustly Learning Unknown Mean Gaussian. We give an SQ algorithm with O(ǫ)error for robustly learning an unknown mean Gaussian, showing that our corresponding SQ lower bound is
qualitatively tight. Our algorithm builds on the filter technique of [DKK+ 16], generalizing it to the more
involved setting of higher-order tensors.
As is suggested by our SQ lower bounds, the obstacle to learning the mean robustly, is that there are
ǫ-noisy Gaussians that are Ω(ǫ)-far in variation distance from a target Gaussian G, and yet match G in
all of their first O(log1/4 (1/ǫ)) moments. For our algorithm to circumvent this difficulty, it will need to
approximate all of the tth -order tensors for t ≤ k = Ω(log1/4 (1/ǫ)). Note that this already requires nk SQ
queries.
The first thing we will need to show is that k moments suffice, for an appropriate parameter k. Because of our lower bound construction, we know that k needs to be at least Ω(log1/4 (1/ǫ)). We show
that k = O(log1/2 (1/ǫ)) suffices. Specifically, we prove a one-dimensional moment-matching lemma
(Lemma 8.1) establishing the following: If an ǫ-noisy one-dimensional Gaussian approximately matches a
reference Gaussian G in all of its first k moments, where k = Θ(log1/2 (1/ǫ)) (i.e., quadratically larger than
our lower bound), then it must be O(ǫ)-close to G in variation distance. We note that it suffices to prove this
statement in the one-dimensional case, as we can just project onto the line between the means.
We now proceed to describe ourpalgorithm: Using the basic filter algorithm from [DKK+ 16], we start by
learning the true mean to error O(ǫ log(1/ǫ)). By translating, we can assume that the mean is this close to
0. We need to robustly approximate the low-order moments of our target Gaussian G′ . This is complicated
by the fact that even a small fraction of errors can have a huge impact on the moments of the distribution.
However, any large errors are easily detectable. In particular, if any tth moment tensor differs substantially
from that of the standard Gaussian, it will necessarily imply the presence of errors. In particular, it will
allow us to construct a polynomial p so that E[p(X)] − E[p(G′ )] (where X is a noisy version of G′ ) is much
larger than ǫkp(G′ )k2 . If this is the case, then many of our errors, x, must have p(x) very far from the mean.
By standard concentration inequalities, this will allow us to identify these points as almost certainly being
errors. This in turn lets us build a filter to clean-up our distribution X, making it closer to G′ .
12
Repeatedly applying filters as necessary, we can reduce to the case where the higher-order moments of
X are close to the higher-order moments of G. This will tell us that, in almost all directions, the first k
moments of X match the corresponding moments of G. By our moment-matching lemma, this will imply
that the mean of G′ is close to 0 in these directions. We will then only need to approximate the mean of
the projection of G′ onto the low-dimensional subspace V in which these moments fail to match. This
approximation can be done in a brute-force manner (in time exponential in dim(V ), which is still relatively
small), completing the description of the algorithm.
1.4 Related Work This work studies learning and testing high-dimensional structured distributions. Distribution learning and testing are two of the most fundamental inference tasks in statistics with a rich history
(see, e.g., [NP33, BBBB72, DG85, Sil86, Sco92, DL01, LR05]) that date back to Karl Pearson. The main
criteria to evaluate the performance of an estimator are its sample complexity and its computational complexity. Despite intensive investigation for several decades by different communities, the (sample and/or
computational) complexity of many learning and testing problems is still not well-understood, even for
some surprisingly simple high-dimensional settings. In the past few decades, a long line of work within
TCS [KMR+ 94, Das99, FM99, AK01, VW02, CGG02, MR05, BV08, KMV10, MV10, BS10, DDS12a,
DDS12b, CDSS13, DDO+ 13, CDSS14a, CDSS14b, ADLS17, DDS15, DDKT16, DKS16b, DKS16a] has
focused on designing efficient estimators in a variety of settings. We have already mentioned the most
relevant references for the specific questions we consider in Section 1.1.
With respect to computational lower bounds for unsupervised estimation problems, the most relevant
references are the works [FGR+ 13, FPV15, KV16] that show SQ lower bounds for the planted clique and
related planted-like problems. It should be noted that, beyond the fact that we also use the concept of
SQ dimension, our techniques are entirely different than theirs. Prior work by Feldman, O’Donnell, and
Servedio [FOS08] implicitly showed an SQ lower bound of nΩ(log k) for the problem of learning k-mixtures
of product distributions over {0, 1}n . This was obtained by a straightforward reduction from the problem of
learning k-leaf decision trees over n Boolean variables. Our lower bound construction for learning GMMs
is entirely different from [FOS08] that relied on the obvious combinatorial structure of the discrete setting.
A related line of work gives statistical-computational tradeoffs for sparse PCA [BR13a, BR13b, MW15,
WBS16a], based on various computational hardness assumptions. These results are of similar flavor as our
statistical–computational tradeoffs for SQ algorithms (Theorems 1.5 and 1.6). An important difference between these tradeoffs and the super-polynomial SQ lower bounds we prove in this paper (Theorems 1.1, 1.2,
and 1.4) is that the aforementioned sparse problems are known to be tractable if we increase the sample size
by a quadratic factor beyond the information-theoretic limit. In contrast, our main SQ lower bound results
establish a super-polynomial gap between the information-theoretic limit and the computational complexity
of any SQ algorithm.
Finally, we remark that in the supervised setting of PAC learning Boolean functions, a number of hardness results are known based on various complexity assumptions, see, e.g., [KKMS08, KS06, FGKP06,
KK14, DLS14, Dan16] for the problems of learning halfspaces and learning intersections thereof.
1.5 Discussion and Future Directions The main contribution of this paper is a technique that gives essentially tight SQ lower bounds for a number of fundamental high-dimensional learning problems, including
learning GMMs and robustly learning a single Gaussian. To the best of our knowledge, these are the first
such lower bounds for high-dimensional distribution learning problems in the continuous setting. As a
corollary, we provide a rigorous explanation of the observed (super-polynomial) gap between the sample
complexity of these problems and the runtime of the best known algorithms.
Our work naturally raises a number of interesting future directions. A natural open problem is to extend
our lower bound technique to broader families of high-dimensional distributions. More concretely, is there
a kω(1) poly(n) SQ lower bound for learning k-mixtures of n-dimensional spherical Gaussians? Note that
13
our nΩ(k) lower bound does not apply for the spherical case, as it crucially exploits the structure of the
covariance matrices. In fact, faster learning algorithms for the spherical case are known [SOAJ14], albeit
with exponential dependence on the number k of components. More broadly, can we extend our techniques
to other families of structured high-dimensional distributions (e.g., mixtures of other distribution families)?
1.6 Organization The structure of this paper is as follows: In Section 2, we introduce basic notation,
definitions, and a number of useful facts that will be required throughout the paper. Our SQ lower bounds
are established in Sections 3–6. Specifically, in Section 3, we give our generic high-dimensional SQ lower
bound construction, assuming the existence of a one-dimensional density satisfying the necessary moment
conditions. In Sections 4, 5, and 6, we construct the appropriate one-dimensional densities, thereby establishing our SQ lower bounds. Specifically, Sections 4 and 5 give our super-polynomial SQ lower bounds for
the problems of learning GMMs and robustly learning an unknown Gaussian. Section 6 gives our quadratic
statistical–computational tradeoffs (for SQ algorithms) for the problems of robust covariance estimation in
spectral norm and robust sparse mean estimation. Section 7 gives our (information-theoretic) sample complexity lower bounds for high-dimensional testing. Finally, in Section 8 we present our (SQ) algorithm for
robustly learning an unknown mean Gaussian with optimal accuracy, whose runtime qualitatively matches
our SQ lower bound from Section 5.
Acknowledgements. This project evolved over a number of years. We would like to thank Vitaly Feldman
for answering numerous questions about the Statistical Query model; Andy Drucker for useful discussions
on Question 1.1; Anup B. Rao for asking a question that motivated Theorem 6.1; Weihao Kong and Gregory
Valiant for useful discussions on Question 1.4; and Ankur Moitra and Eric Price for feedback on a previous
version of this paper.
2 Definitions and Preliminaries
2.1 Notation and Basic Definitions For n ∈ Z+ , we denote by [n] the set {1, 2, . . . , n}. We will denote
by Sn the Euclidean unit sphere in Rn . If v is a vector, we will let kvk2 denote its Euclidean norm. If M is
a matrix, we will let kM k2 denote its spectral norm, and kM kF denote its Frobenius norm.
Our basic object of study is the Gaussian (or Normal) distribution and finite mixtures of Gaussians:
Definition 2.1. The n-dimensional Gaussian distribution N (µ, Σ) with mean vector µ ∈ Rn and covariance
matrix Σ ∈ Rn×n is the distribution with probability density function
1
T −1
−n/2
1/2
f (x) = (2π)
det(Σ) exp − (x − µ) Σ (x − µ) .
2
Definition 2.2. An n-dimensional k-mixture
of Gaussians (k-GMM) is a distribution onP
Rn with probability
Pk
density function defined by F (x) = j=1 wj N (µj , Σj ), where wj ≥ 0, for all j, and kj=1 wj = 1.
Throughout the paper, we will make extensive use of the pdf of the standard one-dimensional Gaussian
N (0, 1), which we will denote by G(x).
Definition 2.3. The total variation distance between two distributions (with probability density functions)
R
def
P, Q : Rn → R+ is defined to be dTV (P, Q) = (1/2) · kP − Qk1 = (1/2) · x∈Rn |P(x) − Q(x)|dx. The
R
def R
χ2 -divergence of P, Q is χ2 (P, Q) = x∈Rn (P(x) − Q(x))2 /Q(x)dx = x∈Rn P2 (x)/Q(x)dx − 1.
2.2 Formal Problem Definitions We record here the formal definitions of the problems that we study.
Our first problem of interest is learning a mixture of k arbitrary high-dimensional Gaussians:
14
Definition 2.4 (Density Estimation/Proper Learning of GMMs). Let Gn,k be the family of n-dimensional
k-GMMs. The problem of density estimation for Gn,k is the following: Given ǫ > 0 and sample access to
an unknown P ∈ Gn,k , with probability 9/10, output a hypothesis distribution H such that dTV (H, P) ≤ ǫ.
The problem of proper learning for Gn,k is the same with the additional requirement that H ∈ Gn,k .
We also
Pkconsider the more challenging task of parameter estimation: Given samples from a distribution
P = i=1 Gi , where Gi is a weighted Gaussian, the goal is to learn a distribution Q that can be written as
P
Q = ki=1 Hi , where Hi is a weighted Gaussian and dTV (Hi , Gi ) is small for all i.
Our next question is the problem of robustly learning a Gaussian in the standard agnostic model:
Definition 2.5 (Robust Learning of Unknown Gaussian). The problem of robust (agnostic) learning of an
unknown Gaussian is the following: Given ǫ > 0, and sample access to an unknown distribution D with
dTV (D, N (µ, Σ)) ≤ ǫ, for some µ ∈ Rn and Σ ∈ Rn×n , output a hypothesis distribution H such that, with
probability at least 9/10, it holds that dTV (H, N (µ, Σ)) = f (ǫ), for some f : R+ → R+ .
Our SQ lower bounds apply to two special cases of this problem: when µ is unknown and Σ = I, and
when µ = 0 and Σ is unknown. For the latter case, our lower bound applies even for learning with respect
to the spectral norm (which is weaker than approximation in variation distance).
We now define the problem of robustly testing a Gaussian in Huber’s model. We remind the reader that
our tight sample complexity lower bound applies to this weaker model as well.
Definition 2.6 (Robust Testing of Unknown Mean Gaussian). The problem of robust testing of an unknown
mean Gaussian is the following: Given ǫ > 0, 0 < δ < ǫ/2 and sample access to an unknown distribution
D over Rn with the promise that one of the following two cases is satisfied: (i) D = N (0, I), or (ii)
D = (1 − δ)N (µ, I) + δN1 , where N1 is an unknown noise distribution and kµk2 ≥ ǫ, the goal is to
correctly distinguish between the two cases with confidence probability 2/3.
Finally, our problem of testing GMMs is the following:
Definition 2.7 (Testing Spherical GMMs). The problem of testing of a spherical GMM is the following:
Given ǫ > 0 and sample access to an unknown distribution D with the promise that one of the following two
cases is satisfied: (a) D = N (0, I), or (b) D is a 2-GMM w1 N (µ1 , I) + w2 N (µ2 , I) in Rn which satisfies
dTV (D, N (0, I)) ≥ ǫ, distinguish between the two cases with probability at least 2/3.
2.3 Basics on Statistical Query Algorithms over Distributions We begin by recording the necessary
definitions of Statistical algorithms for problems over distributions. All the definitions and facts in this
section are from [FGR+ 13]. We start by defining a general search problem over distributions.
Definition 2.8 (Search problems over distributions). Let D be a set of distributions over Rn , let F be a set
of solutions and Z : D → 2F be a map from a distribution D ∈ D to a subset of solutions Z(D) ⊆ F that
are defined to be valid solutions for D. The distributional search problem Z over D and F is to find a valid
solution f ∈ Z(D) given access to (an oracle or samples from) an unknown D ∈ D.
For general search problems over a distribution, we define SQ algorithms as algorithms that do not see
samples from the distribution but instead have access to an SQ oracle. We consider two types of SQ oracles
from the literature.
1. STAT(τ ): For a tolerance parameter τ > 0 and any bounded function f : Rn → [−1, 1], STAT(τ )
returns a value v ∈ [Ex∼D [f (x)] − τ, Ex∼D [f (x)] + τ ].
15
2. VSTAT(t): For a sample size parameter t > 0 and any bounded function f : Rn→ [0, 1], VSTAT(t)
q
Varx∼D [f (x)]
1
returns a value v ∈ [Ex∼D [f (x)] − τ, Ex∼D [f (x)] + τ ], where τ = max t ,
,
t
where Varx∼D [f (x)] = Ex∼D [f (x)] (1 − Ex∼D [f (x)]).
The first oracle was defined by Kearns [Kea98] and the second was introduced in [FGR+ 13]. These oracles
are known to be polynomially equivalent [FGR+ 13]. Also note that these oracles can return any value within
the given tolerance, and therefore can make adversarial choices.
The main technical tool that allows us to prove unconditional lower bounds on the complexity of SQ
algorithms is an appropriate notion of Statistical Query (SQ) dimension. Such a notion was defined in the
context of PAC learning of Boolean functions in [BFJ+ 94], and subsequently generalized to search problems
over distributions in [FGR+ 13]. We will require the simpler definition from Section 3 of that work that relies
on pairwise correlations:
Definition 2.9 (Pairwise Correlation). The pairwise correlation of two distributions with probability density
functions D1 , D2 : Rn → R+ with respect to a distribution with density D : Rn → R+ , where the support
def R
of D contains the supports of D1 and D2 , is defined as χD (D1 , D2 ) = Rn D1 (x)D2 (x)/D(x)dx − 1.
We remark that when D1 = D2 in the above definition, the pairwise correlation is identified with the
def R
between D1 and D, i.e., χ2 (D1 , D) = Rn D1 (x)2 /D(x)dx − 1.
We will also need the following definition:
χ2 -divergence
Definition 2.10. We say that a set of m distributions D = {D1 , . . . , Dm } over Rn is (γ, β)-correlated
relative to a distribution D over Rn if
(
γ if i 6= j
|χD (Di , Dj )| ≤
β if i = j.
We are now ready to define our notion of dimension:
Definition 2.11 (Statistical Query Dimension). For β, γ > 0, a search problem Z over a set of solutions
F, and a class of distributions D over Rn , let m be the maximum integer such that there exists a reference
distribution D over Rn and a finite set of distributions DD ⊆ D such that for any solution f ∈ F, Df =
DD \ Z −1 (f ) is (γ, β)-correlated relative to D and |Df | ≥ m. We define the statistical (query) dimension
with pairwise correlations (γ, β) of Z to be m and denote it by SD(Z, γ, β).
Our lower bounds proceed by bounding from below the statistical query dimension of the considered
distribution learning problems. The corresponding lower bounds on the complexity of SQ algorithms for
these problems are a corollary of the following result from [FGR+ 13]:
Lemma 2.12 (Corollary 3.12 in [FGR+ 13]). Let Z be a search problem over a set of solutions F and a
class of distributions D over Rn . For γ, β > 0, let s = SD(Z,
γ, β). For any γ ′ > 0, any SQ algorithm for
√
Z requires at least s · γ ′ /(β − γ) queries to the STAT( γ + γ ′ ) or VSTAT(1/(3(γ + γ ′ ))) oracles.
3 Statistical Query Lower Bounds: From One–Dimension to High–Dimensions
All our statistical query lower bounds are shown in two steps: We first construct a one-dimensional density
A satisfying certain technical conditions, and then use A to construct a high-dimensional distribution which
is Gaussian in all but one directions. The second step is the same for all the problems that we consider. To
formally define it, we require the following construction:
16
Definition 3.1 (High-Dimensional Hidden Direction Distribution). For a distribution A on the real line
with probability density function A(x) and a unit vector v ∈ Rn , consider the distribution over Rn with
probability density function
Pv (x) = A(v · x) exp −kx − (v · x)vk22 /2 /(2π)(n−1)/2 .
That is, Pv is the product distribution whose orthogonal projection onto the direction of v is A, and onto the
subspace perpendicular to v is the standard (n − 1)-dimensional normal distribution.
Suppose that we have constructed a one-dimensional distribution A satisfying the following condition:
Condition 3.2. Let m ∈ Z+ . The distribution A on R is such that (i) the first m moments of A agree with
the first m moments of N (0, 1), and (ii) χ2 (A, N (0, 1)) is finite.
Note that Condition 3.2-(ii) above implies that the distribution A has a probability density function
(pdf), which we will denote by A(x). We will henceforth blur the distinction between a distribution and its
pdf. The main result of this section is the following:
Proposition 3.3. Given a distribution A on R that satisfies Condition 3.2 for some m ∈ Z+ and a constant
0 < c < 1/2, consider the set of distributions {Pv }v∈Sn , for n ≥ mΩ(1/c) . For a given ǫ > 0, suppose that
dTV (Pv , Pv′ ) > 2ǫ whenever |v · v ′ | is at most 1/8. Then, any SQ algorithm which, given access to Pv (x)
c/2
for an unknown v ∈ Sn , outputs a hypothesis
Q with dTV (Q, Pv ) ≤ ǫ needs at least 2Ω(n ) ≥ nm+1
p
queries to STAT(O(n)−(m+1)(1/4−c/2) χ2 (A, N (0, 1))) or to VSTAT O(n)(m+1)(1/2−c) /χ2 (A, N (0, 1)) .
An intuitive interpretation of the proposition is as follows: If we do not want our SQ algorithm to use
a number of queries exponential in nΩ(1) , then we would need Ω(n)Ω(m+1) samples to simulate a single
statistical query.
The rest of this section is devoted to the proof of Proposition 3.3. In Section 3.1, we prove a correlation
bound which is the main technical ingredient for the proof. In Section 3.2, we show a simple packing for
unit vectors over the sphere and put the pieces together to complete the proof.
3.1
Main Correlation Lemma The main technical result of this section is the following:
Lemma 3.4 (Correlation Lemma). Let m ∈ Z+ . If the distribution A over R agrees with the first m moments
of N (0, 1), then for all v, v ′ ∈ Rn , we have that
|χN (0,I) (Pv , Pv′ )| ≤ |v · v ′ |m+1 χ2 (A, N (0, 1)) .
(2)
Note that we may assume that χ2 (A, N (0, 1)) is finite, otherwise the lemma statement is trivial. Hence,
we can henceforth assume that Condition 3.2 is satisfied. In particular the distributions A, Pv and Pv′ all
have probability density functions.
To prove Lemma 3.4 we proceed as follows: We start by bounding the χ2 -divergence between the onedimensional projection of Pv onto v ′ and N (0, 1). As well as being a critical component towards the proof
of Lemma 3.4, this fact can be used to show that random projections of Pv are close to N (0, 1) with high
probability. Specifically, we show:
Lemma 3.5. Let Q be the distribution of v ′ · X, for X ∼ Pv . Then, we have that
χ2 (Q, N (0, 1)) ≤ (v · v ′ )2(m+1) χ2 (A, N (0, 1)) .
17
Proof. Let θ be the angle between v and v ′ . Let x, y be orthogonal coordinates for the plane spanned by v
and v ′ , with the x-axis in the v ′ direction. Note that Pv is a product of a distribution on this plane and a
standard Gaussian perpendicular to it. On this plane, Pv is a product of A and N (0, 1). Thus, we have that
Z
Pv (x)dx
Q(x) =
′
Zx:v ·x=x
A(x cos θ + y sin θ)G(x sin θ − y cos θ)dy .
=
y∈R
Let Uθ be the linear operator that maps f : R → R to
Z
f (x cos θ + y sin θ)G(x sin θ − y cos θ)dy ,
y∈R
so that Q = Uθ (A). We will show that we can expand A as a linear combination of eigenfunctions of Uθ . √
Let Hei (x) denote the i-th (probabilists’) Hermite polynomial. We note
R that the functions Hei (x)G(x)/ i!,
for i ∈ N, are orthonormal with respect to the inner product hf, gi = x∈R f (x)g(x)/G(x)dx. Indeed, by
p
using the fact that the Hermite functions, which can be written as He(x) RG(x), are a complete orthonormal family for L2 (R), we get that any function f : R → R such that R f (x)g(x)/G(x)dx
< ∞ is
R∞
almost everywhere equal to a linear combination of these Hei (x)G(x)/i!. Since −∞ A(x)2 /G(x)dx =
1 + χ2 (A, N (0, 1)) is finite, we can write
A(x) =
∞
X
√
ai Hei (x)G(x)/ i! .
(3)
i=0
Using the orthogonality of Hei (x) we can extract these coefficients, since
h
√ i Z
ai Hei (x)2 G(x)/i!dx = ai .
EX∼A Hei (X)/ i! =
(4)
R
Since A agrees with the first m moments of the standard Gaussian, for 0 ≤ i ≤ m, we have that
√
√
EX∼A [Hei (X)/ i!] = EX∼N (0,1) [Hei (X)/ i!] = δi,0 .
This implies that a0 = 1 and a1 , . . . , am = 0. Thus, we have
A(x) = G(x) +
∞
X
√
ai Hei (x)G(x)/ i! .
(5)
i=m+1
The orthonormality of these functions with respect to the inner product hf, gi =
allows us to express the χ2 -divergence in terms of these coefficients:
Z ∞
(A(x) − G(x))2 /G(x)dx
χ2 (A, N (0, 1)) =
=
=
Z
−∞
∞
−∞
∞
X
∞
X
i=m+1
R
R f (x)g(x)/G(x)dx
also
!2
√
ai Hei (x)G(x)/ i! /G(x)dx
a2i .
(6)
i=m+1
18
Now we consider the effect of Uθ on this orthogonal family. From the definition of Uθ , we have
Z ∞
Uθ (Hei G)(x) =
Hei (x cos θ + y sin θ)G(x sin θ − y cos θ)dy .
−∞
We will use the well-known fact that Hei (x)G(x) is an eigenfunction of Uθ :
Fact 3.6. We have that: Uθ (Hei G)(x) = cosi (θ)Hei (x)G(x) .
For completeness, we include a proof in Appendix D.
We can now use these eigenfunctions and eigenvalues with Equation (5) to get an expression for Uθ A:
Uθ A(x) = G(x) +
∞
X
√
ai cosi θHei (x)G(x)/ i! ,
(7)
i=m+1
which can be used to express its χ2 -divergence:
∞
X
χ2 (Uθ A, N (0, 1)) =
a2i cos2i θ
i=m+1
≤ cos2(m+1) θ
= cos
2(m+1)
∞
X
a2i
i=m+1
2
θ · χ (A, N (0, 1)) ,
(8)
where the last line uses (6). Recalling that Q = Uθ A and cos θ = v · v ′ , this completes the proof.
Proof of Lemma 3.4. We first show that the correlation between the high-dimensional densities Pv and Pv′
needed for Lemma 3.4 can be reduced to a one-dimensional correlation. Just as in the proof of Lemma
3.5, let θ = arccos(v · v ′ ) and let x, y be coordinates for the plane spanned by v and v ′ with the x-axis in
the v ′ direction. Each of Pv and Pv′ is a product of a distribution on this plane and a standard Gaussian
perpendicular to it. On this plane, they are both products of A and N (0, 1) with different rotations applied.
Thus, we have that
Z
Pv (x)Pv′ (x)/G(x)dx
χN (0,I) (Pv , Pv′ ) + 1 =
n
ZR∞ Z ∞
=
A(x)G(y)A(x cos θ + y sin θ)G(x sin θ − y cos θ)/G(x)G(y)dxdy
−∞ −∞
Z ∞
Z ∞
A(x cos θ + y sin θ)G(x sin θ − y cos θ)dydx
A(x)/G(x) ·
=
−∞
−∞
= 1 + χN (0,1) (A, Uθ A) .
Now we can bound from above this correlation in terms of the χ2 -divergences of both distributions from
N (0, 1), one of which we can bound using Lemma 3.5:
Z
|A(x) − G(x)||Uθ A(x) − G(x)|/G(x)dx
|χN (0,1) (A, Uθ A)| ≤
sZ
sRZ
≤
=
p
R
(A(x) − G(x))2 /G(x)dx ·
χ2 (A, N (0, 1))
≤ | cos
m+1
2
·
χ2 (U
R
(Uθ A(x) − G(x))2 /G(x)dx
θ A, N (0, 1))
(θ)| · χ (A, N (0, 1)) ,
where the first line follows by triangle inequality, the second inequality is Cauchy-Schwarz, and the last line
uses Lemma 3.5. The proof of Lemma 3.4 is now complete.
19
3.2 Proof of Proposition 3.3 We note that our distribution learning problem can be expressed as a search
problem in the sense of [FGR+ 13]. Consider the following search problem Z: given access to (an oracle or
samples from) Pv , for an unknown unit vector v, find a distribution f such that dTV (Pv , f ) ≤ ǫ. Thus, for
us, the set of solutions F is the set of all distributions on Rn , and D ⊆ F is the set of Pv for all unit vectors
v. For a unit vector v, Z(Pv ) is the set of all distributions f such that dTV (Pv , f ) ≤ ǫ. For a distribution f
on Rd , Z −1 (f ) is the set of Pv such that dTV (f, Pv ) ≤ ǫ.
To prove a lower bound on the statistical dimension, we will take D = N (0, I), D = {Pv : v ∈ Sn }
and construct a suitable finite set DD .
c
Lemma 3.7. For any 0 < c < 1/2, there is a set S of at least 2Ω(n ) unit vectors in Rn such that for each
pair of distinct v, v ′ ∈ S, it holds |v · v ′ | ≤ O(nc−1/2 ).
c
The lemma can be shown by a probabilistic argument. Specifically, if we take |S| = 2Ω(n ) unit vectors
uniformly at random from the unit sphere Sn , then the desired event happens with positive probability. The
proof is given in Appendix D.
c
Proof of Proposition 3.3. For a given constant 0 < c < 1/2, fix a set S of 2Ω(n ) unit vectors in Rn
satisfying the statement of Lemma 3.7. Let DD = {Pv | v ∈ S}. Then, by Lemma 3.4, we have that, for
v, v ′ ∈ S with v 6= v ′ , it holds
χN (0,I) (Pv , Pv′ ) ≤ |v · v ′ |m+1 χ2 (A, N (0, 1)) = Ω(n)−(m+1)(1/2−c) χ2 (A, N (0, 1)) .
If v = v ′ , then χN (0,I) (Pv , Pv ) = χ2 (Pv , N (0, I)) = χ2 (A, N (0, 1)). We thus have that, for
def
def
γ = Ω(n)−(m+1)(1/2−c) χ2 (A, N (0, 1)) and β = χ2 (A, N (0, 1)) ,
DD is (γ, β) correlated with respect to D = N (0, I).
Since dTV (Pv , Pv′ ) > 2ǫ for distinct v and v ′ in S, for any distribution f over Rn , we have that
Z −1 (f ) = {Pv : v ∈ S and dTV (Pv , f ) ≤ ǫ} has |Z −1 (f )| ≤ 1 using the triangle inequality for dTV .
We thus have that DD \ Z −1 (f ) is (γ, β) correlated with respect to D = N (0, I), and |DD \ Z −1 (f )| ≥
c
|S| − 1 = 2Ω(n ) . That is,
c
SD (Z, γ, β) ≥ 2Ω(n ) .
def
An application of Lemma 2.12 for γ ′ = γ = Ω(n)−(m+1)(1/2−c) χ2 (A, N (0, 1)), we obtain that any SQ
c
algorithm requires at least 2Ω(n ) n−(m+1)(1/2−c) calls to the
p
STAT O(n)−(m+1)(1/4−c/2) χ2 (A, N (0, 1)) or VSTAT O(n)(m+1)(1/2−c) /χ2 (A, N (0, 1))
oracle to solve Z. From our assumption that n ≥ mΩ(1/c) , it follows that n ≥ Ω((m + 1) log n)2/c , and
c/2
c/2
therefore 2Ω(n ) ≥ nm+1 . Hence, the total number of required queries is at least 2Ω(n ) ≥ nm+1 . This
completes the proof.
4 SQ Lower Bound for Learning Gaussian Mixtures
The main result of this section is the following:
Theorem 4.1. Fix 0 < ǫ < 1. Any SQ algorithm that given SQ access to a k-mixture P of n-dimensional
Gaussians, N (µi , Σi ), i ∈ [k], for n ≥ Ω(k 8 log(1/ǫ)3 ), which are promised to satisfy dTV (Pi , Pj ) ≥ 1−ǫ,
for all i 6= j, and moreover are such that
1/2
1/2
max max kµi − µj k2 , max kΣi k2
k
≤ poly(k, log(1/ǫ)) min 1/kΣ−1
,
2
i
i,j
i
i
20
and outputs a distribution Q with dTV (P, Q) ≤ 1/2, needs at least 2Ω(n
or to VSTAT O(n)k/3 .
1/8 )
≥ n2k calls to STAT O(n)−k/6
Remark. We remark that the well-conditioned assumption in Theorem 4.1 (i.e., that the distances between
the means and the largest and smallest eigenvalues of any covariance matrix are bounded) guarantees that
an SQ algorithm with a bounded number of SQ queries is possible.
The proof of Theorem 4.1 follows by an application of the framework developed in Section 3 and the
following proposition:
Proposition 4.2. For any ǫ > 0, there exists a distribution A on R that is a mixture of k Gaussians Ai ,
i ∈ [k], and satisfies the following conditions:
(i) A agrees with N (0, 1) on the first 2k − 1 moments.
√
1
and mean of magnitude O( k).
(ii) Each Gaussian component Ai has variance Θ k2 log2 (k+1/ǫ)
(iii) It holds dTV (Ai , Aj ) ≥ 1 − ǫ, for all i 6= j.
(iv) We have χ2 (A, N (0, 1)) ≤ exp(O(k)) log(1/ǫ).
(v) In the high-dimensional construction of Definition 3.1, we have that dTV (Pv , Pv′ ) ≥ 1/2 whenever
|v · v ′ | ≤ 1/2.
Given Proposition 4.2, the proof of Theorem 4.1 follows easily.
Proof of Theorem 4.1. First note that the distribution A given by Proposition 4.2, satisfies Condition 3.2 for
m = 2k − 1. By Proposition 3.3, applied for c = 1/4, any algorithm that is given SQ access to Pv , for an
1/8
unknown unit vector v ∈ S, and outputs a distribution Q with dTV (P, Q) ≤ ǫ, needs at least 2Ω(n ) ≥ n2k
calls to
p
STAT O(n)−k/4 · exp(O(k)) log(1/ǫ) or VSTAT O(n)k/2 /(exp(O(k)) log(1/ǫ)) .
For n = Ω(k8 log(1/ǫ)3 ), we have nk/2 ≥ (exp(O(k)) log(1/ǫ))3 , and so we need precision O(n)−k/6 for
STAT or O(n)k/3 for VSTAT.
It remains to show that Pv is a mixture of k Gaussians that satisfies the necessary conditions. Note
that Pv , when expressed in an appropriate basis, is a product of the mixture of k univariate Gaussians
and the standard (n − 1)-dimensional normal distribution.
Recall that
the product of two Gaussians is
Pk
1
′
a Gaussian. If A =
i=1 wi N (µi , δ), where δ = Θ k 2 log(k+1/ǫ) (by Proposition 4.2 (ii)), then we
Pk
′
T . We can bound the variation distance between two
have that Pv =
i=1 wi N vµi , I − (1 − δ)vv
components by:
dTV N (vµ′i , I − (1 − δ)vv T ), N (vµ′j , I − (1 − δ)vv T ) = dTV N (µ′i , δ), N (µ′j , δ) ≥ 1 − ǫ ,
by Proposition 4.2 (iii). Also, we have that
n
o
1/2
max maxi,j kµi − µj k2 , maxi kΣi k2
√
max {maxi,j kµi − µj k2 , 1}
=
≤ O( k/δ) ≤ poly(k log(1/ǫ)) .
1/2
δ
mini 1/kΣ−1
i k2
This completes the proof.
21
4.1
Proof of Proposition 4.2 We start with the following lemma:
Lemma 4.3. There is a discrete distribution B on the real line, supported on k points,
that agrees with
√
N (0, 1) on the first 2k − 1 moments. All points x in the support of B have |x| = O( k).
Proof. This lemma essentially follows from standard techniques for Gaussian quadrature [AS72]. Given a
(possibly infinite) interval [a, b], a weighting function ω(x), and an integer k > 0, we can find xi and wi for
1 ≤ i ≤ k such that
Z b
k
X
wi p(xi ) ,
ω(x)p(x)dx =
a
i=1
for all polynomials p(x) of degree at most 2k − 1. The Gauss-Hermite quadrature is a standard implemen2
x ’s to be the
tation of this general scheme on the interval (−∞, ∞) with ω(x) = e−x . Here, we take the
√ i
2k−1 k! π
roots of the k-th (physicist’s) Hermite polynomial Hk (x). Then, we have that wi = k2 H (xi )2 .
k−1
2
√1 e−x /2 ,
2π
the pdf of N (0, 1). To do this, we need to rescale
√
def
the above wi and xi , and use the probabilist’s Hermite polynomials
Hek (x) = 2−k/2 Hk (x/ 2). We claim
√
k!
that we can take the x′i ’s to be the roots of Hek (x), i.e., x′i = 2xi and wi′ = k2 Hek−1
(x′ )2 . Indeed, we have
We would like to take ω(x) = G(x) :=
i
k
X
wi′ p(x′i ) =
i=1
k
X
k!
√
k2 Hek−1 (
i=1
k
k−1 √
X
2xi
)2
√
p( 2xi )
√
2 k! π
1
√
2xi )
·
·
p(
k2 Hk−1 (xi )2
π
i=1
Z ∞ √
1
2
p( 2y)e−y dy
=√
π −∞
Z ∞
√
1
2
=√
p(x)e−x /2 (1/ 2)dx
π
Z ∞ −∞
p(x)G(x)dx ,
=
=
−∞
for all polynomials p(x) of degree at most 2k − 1.
R∞
P
Note that all the weights are nonnegative by definition. Also note that ki=1 wi′ = −∞ 1 · G(x) = 1.
We take B to be the probability distribution with probability wi′ of being x′i , for each 1 ≤ i ≤ k. Then we
have
Z ∞
k
X
′ ′j
j
wi x i =
xj G(x)dx = EX∼N (0,1) [X j ] ,
EX∼B [X ] =
i=1
−∞
for all integers 1 ≤ j ≤ 2k − 1.
√
k), and so all roots of
It is known (see, e.g., [Sze89]) that all roots of Hk (x) have
absolute
value
O(
√
Hek . Hence, all points x in the support of B have |x| = O( k). This completes the proof.
On the other hand, if we want χ2 (A, N (0, 1)) to be finite, we need to have a mixture of Gaussians each
with positive variance δ > 0.
Corollary 4.4. For any 0 < δ < 1, there is a distribution A on R that is a mixture of k Gaussians each with
variance δ that agrees
√ with N (0, 1) on the first 2k − 1 moments. The means of all the Gaussian components
have magnitude O( k).
22
Proof. By rescaling the distribution B given by Lemma
4.3, we can find a discrete distribution B ′ supported
√
on k points with absolute value no bigger than O( k) that agrees with the first
√ 2k − 1 moments of N (0, 1 −
δ). The rescaled distribution B ′ assigns probability mass wi′ to the points 1 − δx′i , for 1 ≤ i ≤ k. Let
X ∼ B ′ , X ′ ∼ N (0, 1 − δ), and Y ∼ N (0, δ) that is independent of X, X ′ . We take A to be the distribution
of X + Y . Then, we have
j
j
X
X
i
i
E[X ′i ]E[Y j−i ] = E[(X ′ + Y )j ] ,
E[X i ]E[Y j−i ] =
E[(X + Y )j ] =
j
j
i=0
i=0
for all integers 1 ≤ j ≤ 2k − 1. By standard facts about Gaussians, √
X ′ +Y is distributed as N (0, 1). Finally,
note that the distribution of X + Y is a mixture of k Gaussians N ( 1 − δx′i , δ) with weights wi′ .
To appropriately set the parameter δ, we need to consider the high-dimensional construction (Definition 3.1):
√
Lemma 4.5. For v, v ′ ∈ Sn with |v · v ′ | ≤ 1/2, we have that dTV (Pv , Pv′ ) ≥ 1 − O k δ log(1/δ) csc θ .
Proof. We write Ai , for 1 ≤ i ≤ k, for the Gaussians N (µi , δ) that p
A is a mixture of. Fix ǫ > 0. By a
Chernoff bound, Ai is within the interval [µi − a, µi + a], where a = 2 δ log(1/ǫ) with probability at least
1 − ǫ.
We again consider the plane spanned by v and v ′ . Let x, y be the orthogonal coordinates with v in the
direction of the x-axis. Similarly, let x′ , y ′ be the orthogonal coordinates with v ′ in the direction of the
x′ -axis. Let θ be the angle between v and v ′ . We have that
Z ∞ Z ∞
Z
min{Pv (x), Pv′ (x)}dx =
min{A(x)G(y), A(x′ )G(y ′ )}dxdy
x=−∞ y=−∞
x
Z ∞ Z ∞
=
min{A(x)G(y), A(x′ )G(y ′ )} csc θdxdx′
′
x=−∞ x =−∞
Z ∞ Z ∞
min{Ai (x)G(y), Aj (x′ )G(y ′ )} csc θdxdx′
≤ k max
i,j
x=−∞ x′ =−∞
Z µi +a Z µj +a
min{Ai (x)G(y), Aj (x′ )G(y ′ )} csc θdxdx′
≤ kǫ + k max
i,j
≤ kǫ +
x=µi −a x′ =µj −a
2
k max a csc θ max
min{Ai (x)G(y), Aj (x′ )G(y ′ )}
i,j
x,x′ ∈R
√
≤ kǫ + ka2 csc θ/(2π δ)
√
= kǫ + k δ csc θ log(1/ǫ)/π .
√
√
R
Taking ǫ = δ, we obtain that x min{Pv (x), Pv′ (x)}dx ≤ O(k δ log(1/δ) csc θ). On the other hand,
Z
1
|Pv (x) − Pv′ (x)|dx
dTV (Pv , Pv′ ) =
2 x
Z
1
(max{Pv (x), Pv′ (x)} − min{Pv (x), Pv′ (x)}) dx
=
2 x
Z
1
(Pv (x) + Pv′ (x) − 2 min{Pv (x), Pv′ (x)}) dx
=
2 x
Z
= 1 − min{Pv (x), Pv′ (x)}dx
x
√
≥ 1 − O(k δ log(1/δ) csc θ) .
This completes the proof.
23
This gives an upper bound on δ. We don’t want δ to be too small, because of the following lemma:
√
Lemma 4.6. We have that χ2 (A, N (0, 1)) ≤ exp(O(k))/ δ.
Proof. Each component Ai , for 1 ≤ i ≤ k, satisfies the following:
Z
2
1 + χ (Ai , N (0, 1)) = Ai (x)2 /G(x)dx
x
Z
1
exp −(x − µi )2 /δ + x2 /2 dx
=√
2πδ Zx
1
=√
exp −x2 (1/δ − 1/2) + 2µi x/δ − µ2i /δ dx
2πδ Zx
1
=√
exp −(x − 2µi /(2 − δ))2 ((2 − δ)/2δ) + 2µ2i /(δ(2 − δ)) − µ2i /δ dx
2πδ Zx
1
=√
exp −(x − 2µi /(2 − δ))2 ((2 − δ)/2δ) + 2µ2i /(δ(2 − δ)) − (2 − δ)µ2i /δ(2 − δ) dx
2πδ x
√
Z p
2 exp(µ2i /(2 − δ))
p
1/ 2π(2δ/(2 − δ)) exp −(x − 2µi /(2 − δ))2 (1/δ − 1/2) dx
=
(2 − δ)δ
x
√
2
2 exp µi /(2 − δ)
p
=
(2 − δ)δ
√
≤ exp(O(k))/ 2δ .
Thus, for the mixture A we have that:
2
1 + χ (A, N (0, 1)) =
≤
≤
XX
i
j
i
j
i
j
XX
XX
wi wj /(1 − δ)
Z
Ai (x)Aj (x)/G(x)dx
x
q
wi wj /(1 − δ) (1 + χ2 (Ai , N (0, 1)) (1 + χ2 (Aj , N (0, 1))
√
wi wj /(1 − δ) · exp(O(k))/ 2δ
XX
√
wi wj
= exp(O(k))/ 2δ ·
i
√
= exp(O(k))/ 2δ · 1 .
j
This completes the proof.
The following simple lemma helps us enforce the condition that the Gaussian components are wellseparated:
1
Lemma 4.7. Given ǫ > 0, if δ ≤ O k log(1/ǫ)
, then dTV (Ai , Aj ) ≥ 1 − ǫ.
√
Proof. It is known (see, e.g., [Sze89]) that the difference between two roots of Hk (x) is√Ω(1/ k). Thus,
the same is true of Hek (x) and by our construction, we have that |µi − µj | ≥ Ω((1 − δ)/pk), for i 6= j. By
standard Chernoff bounds, with probability at least 1 − ǫ/2, Ai lies in the range (µi − p 2δ ln(2/ǫ), µi +
p
p2δ ln(2/ǫ)). Similarly, with probability at least 1 − ǫ/2, Aj lies in the range (µj − 2δ ln(2/ǫ),
õj +
2δ
ln(2/ǫ)).
If
these
intervals
are
disjoint,
we
have
d
(A
,
A
)
≥
1−ǫ.
This
holds
when
(1−δ)/
k) =
TV
i
j
p
Ω( δ log(1/ǫ)), which is true when δ ≤ O(1/k log(1/ǫ)).
24
We now have all the necessary tools to prove Proposition 4.2. We take δ = C/(k2 log2 (k + 1/ǫ)) for a
sufficiently small constant C > 0. Combined with Corollary 4.4, this gives condition (ii). For condition (i),
note that, by Corollary 4.4, A agrees with N (0, 1) on the first 2k − 1 moments. Since δ was selected to be
smaller than O(1/k log(1/ǫ)), Lemma 4.7 gives condition (iii). Lemma 4.6 gives condition (iv). Finally, by
our choice of δ and Lemma 4.5, we get condition (v). This completes the proof.
5 SQ Lower Bounds for Robust Learning of a Gaussian
In this section, we prove our super-polynomial SQ lower bounds for robustly learning a high-dimensional
Gaussian. In Section 5.1, we show our lower bound for robustly learning an unknown mean spherical
Gaussian. In Section 5.2, we give our lower bound for robustly learning a zero mean unknown covariance
Gaussian with respect to the spectral norm.
5.1 Robust Learning Lower Bound for Unknown Mean Gaussian In this subsection, we use the framework of Section 3 to prove the following theorem:
Theorem 5.1. Let 0 < ǫ < 1 and n ≥ log(1/ǫ)Ω(1) . Fix any M ∈ Z+ such that M = O(log1/2 (1/ǫ)),
where the universal constant in the O(·) is assumed to be sufficiently small. Any algorithm that, given
SQ access to a distribution P on Rn which satisfies dTV (P, N (µ, I)) ≤ ǫ for an unknown µ ∈ Rn with
kµk2 ≤ poly(n/ǫ), and returns a hypothesis distribution Q with dTV (Q, P) ≤ O(ǫ log(1/ǫ)1/2 /M 2 ),
1/12
requires at least 2Ω(n ) ≥ nM calls to STAT O(n)−M/6 or to VSTAT O(n)M/3 .
The theorem will follow from the following proposition:
Proposition 5.2. For any δ > 0 and m ∈ Z+ , there is a distribution A on R satisfying the following
conditions:
(i) A and N (0, 1) agree on the first m moments.
p
(ii) dTV (A, N (δ, 1)) ≤ O(δm2 / log(1/δ)).
(iii) χ2 (A, N (0, 1)) = O(δ).
Before we prove Proposition 5.2, we show how Theorem 5.1 easily follows from it using the machinery
developed in Section 3.
Proof of Theorem 5.1. We can assume without loss of generality that ǫ > 0 is smaller than a sufficiently
small universal constant. We apply Propositions 3.3 and 5.2, with the parameter c set to c = 1/6, m = M ,
and δ = Cǫ ln(1/ǫ)1/2 /M 2 , where C > 0 is a sufficiently large constant. By Proposition
5.2, we have that
p
(i) A and N (0, 1) agree on the first m moments, (ii) dTV (A, N (δ, 1)) ≤ O(δm2 / log(1/δ)) = O(ǫ) and
(iii) χ2 (A, N (0, 1)) = O(δ). Note that for any unit vector v ∈ Sn , it holds that
dTV (Pv , N (vδ, I)) = dTV (A, N (δ, 1)) ≤ O(Cǫ) .
Therefore, for any unit vectors v, v ′ with |v · v ′ | ≤ 1/8 we have that:
dTV (Pv , Pv′ ) ≥ dTV (N (δv, I), N (δv ′ , I)) − dTV (Pv , N (vδ, I)) − dTV (Pv′ , N (v ′ δ, I))
≥ Ω(δkv − v ′ k2 ) − O(Cǫ)
√
= Ω(δ 2 − 2v · v ′ ) − O(Cǫ)
= Ω(δ) − O(Cǫ) = Ω(δ) ,
25
(9)
where the first line is the triangle inequality, the second line uses (9) and that dTV (N (µ1 , I), N (µ2 , I)) =
Ω(kµ1 − µ2 k) when kµ1 − µ2 k is smaller than an absolute constant, the third line uses the assumption that
|v · v ′ | ≤ 1/8, and the last line follows from the definition of δ.
We want to apply Proposition 3.3 with its “local parameter” ǫ taken to be Ω(δ). The assumption on n
in the statement of Theorem 5.1, i.e., n ≥ log(1/ǫ)Ω(1) , implies the condition n ≥ mΩ(1) in the statement
of Proposition 3.3. By our choice of c = 1/6, we conclude
that any
SQ algorithm for our learning problem
√
1/12 )
Ω(n
M
−M/6
δ or to VSTAT O(n)M/3 /δ to produce
requires at least 2
≥ n queries to STAT O(n)
a hypothesis distribution Q with
dTV (Q, Pv ) ≤ O(δ) ≤ O(ǫ ln(1/ǫ)1/2 /M 2 ) ,
where we used the assumption that C is sufficiently large. This completes the proof.
Proof of Proposition 5.2 The rest of this section is devoted to the proof of Proposition 5.2. We start
by describing the outline of the proof. We then provide a number of intermediate useful lemmas that we
subsequently combine to complete the proof.
p
The proof plan proceeds as follows. For some C = Θ( log(1/δ)), we define the one-dimensional
distribution A to be:
• For x ∈
/ [−C, C], we define A(x) = G(x − δ).
• For x ∈ [−C, C], we define A(x) = G(x − δ) + p(x), where p(x) is the degree-m polynomial with
RC
−C p(x)dx = 0 and
Z
C
i
p(x)x dx =
−C
Z
∞
−∞
(G(x) − G(x − δ))xi dx ,
(10)
for 1 ≤ i ≤ m. (We note that p is unique after fixing m, C and δ.)
We need to show that we can
p find appropriate values for the parameters m, C, and δ such that the L1 -norm
of p(x) is at most O(δm2 / log(1/δ)) and that A(x) is non-negative. To achieve that, we will express p(x)
as a linear combination of (appropriately scaled) Legendre polynomials, a family of orthogonal polynomials
on [−C, C]. Rather than directly showing that the first m moments agree, we will instead want that the
expectations of the first m scaled Legendre polynomials agree. Bounds on the coefficients of the Legendre
polynomials in p(x) allow us to obtain bounds on the L1 and L∞ norms of p(x) on [−C, C]. Choosing m,
δ, and C appropriately will complete the proof of the proposition.
Properties of Legendre Polynomials
we will need:
We start by recording the properties of Legendre polynomials that
Fact 5.3. [Sze89] The Legendre polynomials, Pk (x), for k ∈ Z+ , satisfy the following properties:
(i) Pk (x) is a degree-k polynomial, P0 (x) = 1, and P1 (x) = x.
R1
(ii) −1 Pi (x)Pj (x)dx = (2/(2i + 1))δi,j for all i, j ≥ 0.
(iii) |Pk (x)| ≤ 1 for all |x| ≤ 1.
(iv) Pk (x) = (−1)k Pk (−x).
P⌊k/2⌋
(v) Pk (x) = (1/2k ) i=0 ki
2k−2i k−2i
x
.
k
26
As a simple corollary we obtain the following lemma:
Corollary 5.4. We have:
(i) |Pk (x)| ≤ (4|x|)k for all |x| ≥ 1.
√
R1
(ii) −1 |Pk (x)|dx ≤ O(1/ k).
We are now ready to proceed with the formal proof. The main technical result of this section is the following
lemma:
P
3/2 /C 2 ), for 0 ≤ k ≤ m.
Lemma 5.5. We can write p(x) = m
k=0 ak Pk (x/C), where |ak | = O(δk
Before we give the proof of Lemma 5.5, we deduce two corollaries that will be useful in the proof of
Proposition 5.2. First, we can obtain bounds on the L1 and L∞ norms of p(x) on [−C, C]. As an immediate
corollary of Lemma 5.5 and the aforementioned properties of Legendre polynomials, we deduce:
RC
Corollary 5.6. We have that: −C |p(x)|dx ≤ O(δm2 /C) and |p(x)| ≤ δm5/2 /C 2 , for all x ∈ [−C, C].
We now bound from above the desired χ2 -divergence:
Lemma 5.7. χ2 (A, N (0, 1)) = O δ2 + δm5/2 /C 2 · (C 2 δ2 + max|x|≤C |p(x)|/G(x)) .
Proof. We have the following:
Z ∞
A(x)2 /G(x)dx − 1
χ2 (A, N (0, 1)) =
=
−∞
∞
Z
−∞
Z
G(x − δ)2 /G(x)dx − 1 +
For the first term, we note that:
Z ∞
Z
2
G(x − δ) /G(x)dx =
−∞
∞
−∞
C
−C
2p(x)G(x − δ)/G(x)dx +
Z
C
p(x)2 /G(x)dx .
−C
G(x − 2δ) exp(δ2 )dx = exp(δ2 ) ≤ 1 + 2δ2 .
We bound the second term from above as follows:
Z C
Z
2p(x)G(x − δ)/G(x)dx =
−C
=
C
−C
Z C
−C
2p(x) exp(xδ − δ2 /2)dx
p(x) · (1 + xδ + O(C 2 δ2 ))dx
2 2
≤ C|a0 | + O(Cδ|a1 |) + C δ
Z
≤ 0 + O(δ2 /C) + O(δ3 m5/2 ) ,
C
−C
|p(x)|dx
where the last lines uses Lemma 5.5 and Corollary 5.6. Finally, for the third term we have:
Z
C
−C
p(x)2 /G(x)dx ≤ δm5/2 /C 2 max |p(x)|/G(x) ,
x∈[−C,C]
where the inequality follows from Corollary 5.6. This completes the proof of Lemma 5.7.
27
Proof of Lemma 5.5 We first note that we can express p(x) as a linear combination of scaled Legendre
polynomials whose coefficients are explicitly given by integrals:
RC
P
Claim 5.8. We can write p(x) = m
k=0 ak Pk (x/C), where ak = ((2k + 1)/2C) −C Pk (x/C)p(x)dx.
Proof. Since p(x) has degree at most m and the set of polynomials Pk (x/C),
Pm0 ≤ k ≤ m, contains a
polynomial of each degree from 0 to m, there exists ak ∈ R such that p(x) = k=0 ak Pk (x/C).
RC
It follows from Fact 5.3 (ii) and a change of variables that −C Pi (x/C)Pj (x/C)dx = (2C/(2i+1))δi,j ,
for all i, j ≥ 0. We can use this to extract the ak ’s. For 1 ≤ k ≤ m, we have
RC
Pk (x/C)p(x)dx =
RC
Pk (x/C)
ai Pi (x/C)dx =
i=0
−C
−C
m
P
m
P
ai
i=0
RC
Pk (x/C)Pi (x/C)dx = (2C/(2k+1))ak .
−C
Since the first m moments of p are fixed, via 10, we obtain:
Z
C
p(x)Pk (x/C)dx =
−C
Z
∞
−∞
(G(x) − G(x − δ))Pk (x/C)dx ,
(11)
for any 0 ≤ k ≤ m.
Since we will apply this with the parameter 1/δ exponential in m and C, we will be able to ignore O(δ2 )
terms. We use Taylor’s theorem to expand (G(x) − G(x − δ)) up to second order terms:
Fact 5.9. G(x) − G(x − δ) = xG(x)δ + (ξ(x)2 − 1)/2 · G(ξ(x))δ2 , for some x ≤ ξ(x) ≤ x + δ.
R∞
By (11) and Fact 5.9, to bound the magnitude of the ak ’s, it suffices to bound the terms −∞ Pk (x/C)xG(x)dx
R∞
and −∞ Pk (x)(ξ(x)2 − 1)/2 · G(ξ(x))dx. This is done in the following two lemmas.
√
R∞
Lemma 5.10. For k ≤ 4C, we have that −∞ Pk (x/C)xG(x)dx ≤ O( k/C).
Proof. When k is even, using Fact 5.3 (iv), we have that Pk (x/C)xG(x) = −(Pk (−x/C) · (−x)G(−x)),
and so the integral is zero. When k is odd, we can rewrite Fact 5.3 (v) in ascending order of terms, by using
the change of variables j = (k + 1)/2 − i, as
k
Pk (x) = (1/2 )
(k+1)/2
X
j=1
k
(k − 1)/2 + j
k + 2j − 1 2j−1
x
.
2j − 1
By standard results about the moments of Gaussians, for all j ≥ 1, we have that
Q
1)!! := ji=1 (2i + 1). Thus, we can write
Z
∞
−∞
Pk (x/C)xG(x)dx =
Z
∞
k
1/2
−∞
k
= (1/2 )
(k+1)/2
X
j=1
(k+1)/2
X
j=1
k
(k − 1)/2 + j
k
(k − 1)/2 + j
28
R∞
−∞ x
2j G(x)dx
= (2j +
k + 2j − 1 2j
x G(x)/C 2j−1 dx
2j − 1
k + 2j − 1
(2j + 1)!!/C 2j−1 .
2j − 1
Note that this quantity is non-negative. We can bound it from above as follows:
Z ∞
(k+1)/2
X
k
k + 2j − 1
k
Pk (x/C)xG(x)dx = (1/2 )
(2j + 1)!!/C 2j−1
(k
−
1)/2
+
j
2j
−
1
−∞
j=1
X
√
k + 2j − 1
(1/ k)
(2j + 1)!!/C 2j−1
2j − 1
X
√
(1/ k)(k + 2j − 1)2j−1 (2j + 1)!!/C 2j−1 (2j − 1)!
X
√
(1/ k)2(k + 2j − 1)2j−1 /C 2j−1
(k+1)/2
≤
j=1
(k+1)/2
≤
j=1
(k+1)/2
≤
j=1
X
√
√ (k+1)/2
(2k/C)2j−1 ≤ 8 k/C .
≤ (2/ k)
j=1
The proof of Lemma 5.10 is now complete.
R∞
Lemma 5.11. For any integer 1 ≤ k ≤ 4C, we have that −∞ Pk (x)(ξ(x)2 − 1)/2 · G(ξ(x))dx ≤ O(1) ,
for any function ξ(x) with x ≤ ξ(x) ≤ x + δ, for all x ∈ R.
Proof. We separate this integral into the interval [−C, C] and the tails. We can use Fact 5.3 (iii) to bound
the integral on [−C, C], as follows:
Z C
Z C
Pk (x)(ξ(x)2 − 1)/2 · G(ξ(x))dx ≤
(ξ(x)2 − 1)/2 · G(ξ(x))dx
−C
≤
Z
−C
C
−C
(|x| + δ + 1)2 /2 · G(min{0, |x| − δ})dx
≤ O(δ) +
Z
C+δ
−C−δ
(|x| + 2δ + 1)2 /2 · G(x)dx
≤ O δ + EX∼G [1] + EX∼G [|X|] + EX∼G [X 2 ]
= O(1) .
For the tails, we need Corollary 5.4(i). For the right tail, [C, ∞), we have
Z ∞
Z ∞
2
Pk (x)(ξ(x) − 1)/2 · G(ξ(x))dx ≤
(4|x|/C)k (ξ(x)2 − 1)/2 · G(ξ(x))dx
C
ZC∞
≤
(4|x|/C)k (x + δ)2 G(x − δ)dx
ZC∞
≤
(4/C)k |x + 2δ|k+2 G(x − δ)dx
C
Z ∞
≤
(4/C)k |x|k+2 · (1 + 2δ/C)k G(x − δ)dx
C−δ
Z ∞
≤2
(4/C)k |x|k+2 G(x − δ)dx
−∞
≤ O((4/C)k (k + 3)!!)
√
≤ O((4 k/C)k ) ≤ O(1) .
29
A similar bound holds for the left tail, which completes the proof.
Putting everything together, gives Lemma 5.5.
To prove Proposition 5.2, we need to set C appropriately and check the bounds on m needed for A(x)
to satisfy the necessary properties.
p
Proof of Proposition 5.2. Note that unless δ is sufficiently small and m2 ≤ O( log(1/δ)),
p taking A =
N (0, 1), instead of using our√construction, satisfies the proposition. We will take C = Θ( log(1/δ)), and
so we can assume that m ≤ C.
Recall that A(x) is defined to be G(x − δ) + p(x) on [−C, C] and G(x − δ) outside of [−C, C]. Firstly,
A(x) needs to be the pdf of a distribution. Since
Z ∞
Z C
(G(x) − G(x − δ))Pk (x/C)dx
Pk (x/C)p(x)dx =
−C
R∞
−∞
for k = 0, when Pk (x) = 1, we have that −∞ A(x)dx = 1. We also need that A(x) is non-negative, i.e.,
that A(x) = G(x − δ) + p(x) ≥ 0 for all x ∈ [−C, C]. Note that
G(x) + p(x) ≥ G(C + δ) − δm5/2 /C 2 ,
p
using Corollary 5.6. Since m2 ≤ C, we need G(C + δ) ≥ δC 3/4 . This holds when C = ln(1/δ) − δ,
p
p
3/4
since then we have G(C +δ) = δ/2π ≥ δ ln(1/δ) for sufficiently small δ. Note that this also implies
that A(x) ≤ 2G(x − δ) for all x, and |p(x)| ≤ G(x) for all −C ≤ x ≤ C.
The second of these and Lemma 5.7 imply (iii). For (i), by construction, we have that the first m
moments agree.
A satisfies (ii), since by Lemma 5.6 ,
Z
p
1 C
|p(x)|dx ≤ O(δm2 /C) = O(δm2 / log(1/δ)).
dTV (A, N (δ, 1)) =
2 −C
The proof of Proposition 5.2 is now compete.
5.2 Robust Learning Lower Bound for Unknown Covariance Gaussian In this subsection, we prove
an SQ lower bound for robustly learning the covariance matrix of a high-dimensional Gaussian with known
mean. We note that our lower bound applies even for spectral norm approximation. In particular, we show:
Theorem 5.12. Let ǫ > 0 and n ≥ Ω(log2 (1/ǫ)). Fix any M ∈ Z+ such that M = O(log1/4 (1/ǫ)),
where the universal constant in the O(·) is assumed to be sufficiently small. Any algorithm that, given SQ
access to a distribution P on Rn which has dTV (P, N (0, Σ)) ≤ ǫ for some positive-definite Σ ∈ Rn×n with
e with kΣ
e − Σk2 ≤ O(ǫ log(1/ǫ)/M 4 ), requires at least 2n2/15 ≥ nM
I/2 Σ 2I, and returnsa matrix Σ
calls to STAT O(n)−M/6 or to VSTAT O(n)M/3 .
The theorem will follow from the following proposition:
Proposition 5.13. For any 1/3 > δ > 0 and integer log(1/δ)1/4 ≫ m > 0, there is a distribution A on R
satisfying the following conditions:
(i) A and N (0, 1) agree on the first m moments.
(ii) dTV (A, N (0, (1 − δ)2 ) ≤ O(δm4 / log(1/δ)).
(iii) χ2 (A, N (0, 1)) = O(1 + m8 δ3/2 /log(1/δ)5/2 ).
30
As in the previous subsection, Theorem 5.12 follows easily from Proposition 3.3 and Proposition 5.13.
Proof of Theorem 5.12. We apply Propositions 3.3 and 5.13 with m = M and δ = 2ǫ ln(1/ǫ)/M 4 . By
Proposition 5.13, we have that (i) A and N (0, 1) agree on the first M moments, (ii) dTV (A, N (0, (1 −
δ)2 )) ≤ O(δm4 / log(1/δ)) = O(ǫ) and (iii) χ2 (A, N (0, 1)) = O(1).
Note that we cannot directly apply Proposition 3.3, since we are not aiming to learn within small total
variation distance. Instead, we are interested in a different search problem, that of finding an approximation
e to the covariance Σ with kΣ
e − Σk2 ≤ ǫ log(1/ǫ)/M 4 , where Σ is the covariance of a mean 0 Gaussian
Σ
within ǫ total variation distance. Note that for Pv , we have that Σ = I − (1 − (1 − δ)2 )vv T . We need to
e the set Z −1 (Σ)
e = {Pv :
argue that this search problem has at most one solution in S, i.e., that for any Σ,
2
T
4
−1
e
e
v ∈ S and kΣ−I −(1−(1−δ) )vv k2 ≤ ǫ log(1/ǫ)/M } has |Z (Σ)| ≤ 1, where S is as in Lemma 3.7.
Lemma 5.14. For S as in Lemma 3.7 with c = 1/6 and with n larger than a sufficiently large constant,
e − I − (1 − (1 − δ)2 )vv T k2 ≤ ǫ log(1/ǫ)/M 4 }| ≤ 1, for all Σ.
e
|{Pv : v ∈ S and kΣ
e and let v, v ′ be distinct
Proof. Suppose for a contradiction that this set has size at least 2 for some Σ
2
T
e − Σv k2 ≤
elements. Let Σv = I − (1 − (1 − δ) )vv and define Σv′ similarly. Then we have kΣ
4
4
e
ǫ log(1/ǫ)/M and kΣ − Σv′ k2 ≤ ǫ log(1/ǫ)/M . By the triangle inequality, we have kΣv − Σv′ k2 ≤
2ǫ log(1/ǫ)/M 4 . However, we also have that |v ·v ′ | ≤ O(n−1/3 ) ≤ 1/2. Now we get that v T Σv = (1− δ)2 ,
but v T Σv′ v = (1 − |v · v ′ |2 ) · 1 + |v · v ′ |2 · (1 − δ)2 ≥ 3/4 + (1/4)(1 − δ)2 . We thus obtain
kΣv − Σv′ k2 ≥ v T (Σv′ − Σv )v ≥ (3/4)(1 − (1 − δ)2 ) ≥ (3/2)δ − O(δ2 ) > δ ,
where the last inequality assumes that ǫ is at most an appropriately small universal constant. Since δ =
2ǫ ln(1/ǫ)/M 4 , this leads to a contradiction.
Now we use the the proof of Proposition 3.3 with c = 1/6 and ǫ taken to be Ω(δ). Our condition on n,
8 ). We conclude that it requires at least 2n2/15 ≥ nM calls to
n ≥ ln(1/ǫ)2 , implies
the condition n ≥ Ω(m
e such that kΣ
e − Σk2 ≤ ǫ log(1/ǫ)/M 4 .
STAT O(n)−M/6 or to VSTAT O(n)M/3 to produce a a Σ
Proof of Proposition 5.13. Similarly to the previous subsection, we choose to define the univariate distribution A to have probability density function given by
A(x) = G(x/(1 − δ))/(1 − δ) − p(x)1[−C,C] ,
p
where C is a sufficiently small multiple of log(1/δ) and p(x) is the unique degree-m polynomial that
causes APand G(x) (the pdf of N (0, 1)) to have the same
R ∞ first m moments. Once again, we may write
m
p(x) = k=0 ak Pk (x/C), where ak = ((2k + 1)/2C) −∞ (G(x) − G(x/(1 − δ))/(1 − δ))Pk (x/C)dx.
The bulk of our proof will now be in bounding the ak ’s.
The first thing to note is that since G(x) − G(x/(1 − δ)) is even, ak is 0 for k odd. For k even, we will
31
need to compute this expression using Fact 5.3 (v). In particular, we have that
Z
2k + 1 ∞
ak =
(G(x) − G(x/(1 − δ))/(1 − δ))Pk (x/C)dx
2C
−∞
Z
⌊k/2⌋
X k
2k − 2i
2k + 1 ∞
−k
(G(x) − G(x/(1 − δ))/(1 − δ))2
(x/C)k−2i dx
=
2C
i
k
−∞
i=0
2k + 1
=
2C
=
Z
∞
−∞
−k
(G(x) − G(x/(1 − δ))/(1 − δ))2
⌊k/2⌋
2k + 1 X
2k+1 C
j=0
=
k
k/2 + j
k + 2j
C −2j
k
Z
∞
−∞
⌊k/2⌋
X
k
k/2 + j
j=0
k + 2j
(x/C)2j dx
k
(G(x) − G(x/(1 − δ))/(1 − δ))x2j dx
⌊k/2⌋
k
k + 2j
2k + 1 X
C −2j (2j − 1)!!(1 − (1 − δ)2 j)
2k+1 C
k/2 + j
k
j=0
⌊k/2⌋
δ(2k + 1) X k + 2j
≤
C −2j (2j − 1)!!j
C
2j
j=0
∞
δ(2k + 1) X 2k 2j
≤
C
C
j=1
3
≤ δ10k C
−3
,
where in the last step we assume that k is less than a sufficiently small multiple of C.
It is now clear that A is a pseudo-distribution that matches its first m moments with N (0, 1). Firstly,
in order to check that A is a distribution, it
clear that A(x) > 0 for |x| > C. For |x| ≤ C we have
Pis
m
4 −3 . Since this is smaller than δ 1/2 <
that |A(x) − G(x/(1 − δ))/(1 − δ)| ≤
k=0 |ak | ≤ δ10m C
G(x/(1 − δ))/(1 − δ), we have that A(x) ≥ 0 everywhere.
Next, we need to bound from above dTV (A, N (0, 1 − δ)), i.e., the L1 -distance between A(x) and
G(x/(1 − δ))/(1 − δ). This in turn is at most
m Z
X
k=0
C
−C
|ak Pk (x/C)|dx ≤
m
X
δ20k3 C −2 = O(δm4 / log(1/δ)).
k=0
Finally, we need to bound from above χ2 (A, N (0, 1)). Note that
χ2 (A, N (0, 1)) ≤ O(χ2 (N (0, 1 − δ), N (0, 1)) + χ2 (A − N (0, 1 − δ), N (0, 1))) .
It is easy to see that χ2 (N (0, 1 − δ), N (0, 1)) = O(1 + δ). On the other hand, we have that
χ2 (A − N (0, 1 − δ), N (0, 1))) ≤ (m + 1)
m
X
a2k
k=0
= O(m7 δ2 C −6 )
= O(m8 δ3/2 C
This completes the proof.
32
Z
C
Pk (x/C)2 /G(x)dx
−C
m Z C
X
k=0
−5
).
−C
G(x)−2 dx
6 Statistical and Computational Tradeoffs
In this section, we prove our SQ lower bounds establishing statistical-computational tradeoffs for two natural
robust estimation problems. In Section 6.1, we give a sharp-tradeoff for the problem of robustly estimating
the covariance matrix in spectral norm. In Section 6.2, we show such a tradeoff for robust sparse mean
estimation.
6.1 Robust Estimation of Covariance Matrix in Spectral Norm In this subsection, we establish an SQ
lower bound for robust covariance estimation in spectral norm. Our SQ lower bound provides evidence for
the existence of a statistical-computational tradeoff for this problem. Roughly speaking, we show that, for
any constant c > 0, given samples from a corrupted n-dimensional Gaussian N (0, Σ), any computationally
efficient SQ algorithm that approximates Σ within a factor of 2 requires Ω(n2−c ) samples. Our lower bound
applies even to the weaker Huber contamination model.
We note that the information-theoretic optimum for this problem is known to be Θ(n) samples (and is
achievable by an exponential time SQ algorithm). Hence, our lower bound establishes a nearly-quadratic
gap in the sample complexity between efficient and inefficient SQ algorithms for this problem. Formally,
we show:
Theorem 6.1. Let 0 < c < 1/6, and n > 125. Any algorithm that, given SQ access to a distribution
P on Rn of the form P = (1 − ǫ)N (0, Σ) + ǫN1 , where ǫ ≤ c/ ln(n) and N1 is a noise distribution,
for some covariance Σ with kΣk2 ≤ poly(n/ǫ), and returns a matrix Σ̃ with Σ̃/2 Σ 2Σ̃, requires
c/3
at least 2Ω(n ) calls to STAT O(n)−(1−5c/2) or to VSTAT O(n)2−5c . Furthermore, the result holds
even when the noise distribution N1 is a mixture of 2 Gaussians.
Proof. Let ǫ = c/ ln(n). We consider the following mixture of 3 Gaussians:
p
p
A = (1 − ǫ)N (0, (1/5 − ǫ)/(1 − ǫ)) + (ǫ/2) · N ( 4/(5ǫ), 1) + ǫ/2 · N (− 4/(5ǫ), 1) .
Note that A is symmetric about 0 and so, for X ∼ A, we have EX∼A [X] = EX∼A [X 3 ] = 0. The variance
of A is VarX∼A [X] = EX∼A [X 2 ] = (1/5 − ǫ) + ǫ · 4/(5ǫ) + ǫ = 1. That is, A agrees with N (0, 1) on the
first 3 moments. We need a bound on χ2 (A, N (0, 1)). For this, we use the following three easy facts (see
Appendix D for the simple proofs):
Fact 6.2. For distributions B, C, D and w ∈ [0, 1], we have that χ2 (wB + (1 − w)C, D) = w2 χ2 (B, D)+
(1 − w)2 χ2 (C, D) + 2w(1 − w)χD (B, C).
Fact 6.3. For µ, µ′ ∈ R, we have that χN (0,1) (N (µ′ , 1), N (µ, 1)) = exp(µµ′ ) − 1.
p
Fact 6.4. We have that χ2 (N (0, σ 2 ), N (0, 1)) = 2/σ 4 − 1/σ 2 − 1.
Note that 1/6 ≤ (1/5 − ǫ)/(1 − ǫ) ≤ 1/5. Fact 6.4 now yields
p
√
χ2 (N (0, (1/5 − ǫ)/(1 − ǫ)), N (0, 1)) ≤ 2/(1/6)2 + 1/(1/6) − 1 = 78 − 1 ≤ 8 .
Using Fact 6.2, we can write:
p
χ2 (A, N (0, 1)) ≤ (1 − ǫ)2 χ2 (N (0, (1/5 − ǫ)/(1 − ǫ)), N (0, 1)) + (ǫ2 /2)χ2 (N ( 4/5ǫ, 1), N (0, 1))
p
p
p
+ (ǫ2 /2)χN (0,1) (N ( 4/5ǫ, 1), N (− 4/5ǫ, 1)) + (1 − ǫ)ǫχN (0,1) (N (0, (1/5 − ǫ)), N ( 4/5ǫ, 1))
≤ 8 + (ǫ2 /2)(exp(4/(5ǫ)) − 1) + 0
q
p
+ (1 − ǫ)ǫ χ2 (N ( 4/5ǫ, 1), N (0, 1))χ2 (N (0, (1/5 − ǫ)/(1 − ǫ)), N (0, 1))
√
≤ 8 + ǫ2 exp(4/(5ǫ)) + ǫ 8 exp(2/(5ǫ)) ≤ O(1 + exp(1/ǫ))
≤ O(nc ) .
33
Note that we cannot directly apply Proposition 3.3, since we are not aiming to learn within small variation
distance. Instead, we are interested in a different search problem, that of approximating the covariance Σv
of the (1 − ǫ) weight component of Pv to within a factor of 2. We need to argue that this search problem
has at most one solution in S, i.e., that for any Σ, the set Z −1 (Σ) = {Pv : v ∈ S and Σ Σv 2Σ} has
|Z −1 (Σ)| ≤ 1, where S is as in Lemma 3.7.
Lemma 6.5. For S as in Lemma 3.7, |{Pv : v ∈ S and Σ Σv 2Σ}| ≤ 1 for all Σ.
Proof. Suppose for a contradiction that |Z −1 (Σ)| ≥ 2 for some Σ. Then there are distinct v, v ′ ∈ S with
Σ Σv 2Σ and Σ Σv′ 2Σ. However, we have that |v · v ′ | ≤ O(nc−1/2 ) ≤ n−1/3 . Now
v T Σv v = (1/5 − ǫ)/(1 − ǫ) < 1/5, but
v T Σv′ v = (1 − |v · v ′ |2 ) · 1 + |v · v ′ |2 · (1/5 − ǫ)/(1 − ǫ) ≥ 1 − 5|v · v ′ |2 /6 ≥ 1 − n−1/3 > 4/5 .
Thus, we need v T Σv ≤ 2v T Σv v < 2/5, but v T Σv ≥ v T Σv′ v/2 > 2/5. This is a contradiction and so
|Z −1 (Σ)| < 1.
Now the proof of Proposition 3.3 applies and we obtain that any algorithm that outputspa Σ̃ satisfying the
c/2
desired conditions uses at least 2Ω(n ) ≥ nm+1 queries to STAT(O(n)−(m+1)(1/4−c/2) χ2 (A, N (0, 1)))
or to VSTAT(O(n)(m+1)(1/2−c) /χ2 (A, N (0, 1))), where m = 3. Now substituting χ2 (A, N (0, 1) ≤
c/2
O(nc ), we get that we need at least 2Ω(n ) queries to STAT(O(n)−(1−5c/2) ) or to VSTAT(O(n)(2−5c) ).
The proof is now complete.
6.2 Robust Sparse Mean Estimation In this subsection, we establish an SQ lower bound for robust
sparse mean estimation. Our SQ lower bound gives evidence for the existence of a statistical-computational
tradeoff for this problem. Roughly speaking, we show that, for any constant c > 0, given samples from
a corrupted n-dimensional Gaussian N (µ, I), where the mean vector µ is k-sparse, any computationally
efficient SQ algorithm that approximates the true mean requires Ω(k2−c ) samples. Our lower bound applies
even to the weaker Huber contamination model.
We note that the information-theoretic optimum for this problem is known to be Θ(k log n) (and is
achievable by an exponential time SQ algorithm). Hence, our lower bound establishes a nearly-quadratic
gap in the sample complexity between efficient and inefficient SQ algorithms. Formally, we show:
Theorem 6.6. Fix any constant 0 < c < 1. Let k, n ∈ Z+ be sufficiently large positive integers satisfying
n ≥ 8k2 . Any algorithm which, given SQ access to a distribution P on Rn such that P = (1 − δ)N (µ, I) +
δN1 , where N1 is an arbitrary distribution, δ = ǫ/kc/4 , and µ ∈ Rn is promised to be k-sparse with
c
kµk2 = ǫ, and outputs a hypothesis vector µ
b satisfying kb
µ − µk2 ≤ ǫ/2, requires at least Ω(nck /8 ) queries
to STAT(O(k)3c/2−1 ) or to VSTAT(O(k)2−3c ).
To prove our result, we will use the framework of Section 3: We will construct a suitable one-dimensional
distribution A and consider an appropriate collection of distributions Pv , but this time only for k-sparse unit
vectors v on Rn . We start by showing an analogue of Lemma 3.7 for k-sparse vectors, and then use it to
prove an analogue of Proposition 3.3. Our analogue of Lemma 3.7 is the following:
Lemma 6.7. Fix a constant 0 < c < 1. There exists a set S of k-sparse unit vectors on Rn of cardinality
c
|S| = ⌊nck /8 ⌋ such that for each pair of distinct vectors v, v ′ ∈ S we have that |v · v ′ | ≤ 2kc−1 .
Proof. Let √
D be the uniform distribution over the set of vectors v on Rn that have exactly k coordinates
equal to 1/ k and the rest n − k coordinates equal to zero. Consider the distribution D ′ of the inner product
v · v ′ , with v and v ′ independently drawn from D. Then, we have that v · v ′ = i/k, where i is the number of
34
non-zero coordinates that v ′ and v ′ have in common. Note that the distribution D ′ does not change if we fix
v ′ and therefore we have that
k
n−k
n
′
Pr[v · v = i/k] =
/
.
i
k−i
k
In other words the random variable k(v ·v ′ ) is distributed as the hypergeometric distribution with parameters
(n, k, k). By standard tail bounds on the hypergeometric distribution, for t > 0, we have
Pr k(v · v ′ ) ≥ k(k/n + t) ≤ exp (−k · KL (t + k/n||k/n)) ,
where KL(a||b) = a ln(a/b) + (1 − a) ln( 1−a
1−b ). We apply this concentration inequality in a regime where
a ≥ 9b, in which case KL(a||b) ≥ a ln(a/b) − 1 ≥ (a/2) ln(a/b).
def
Fix any constant 0 < c < 1. We apply the above concentration bound for t = kc−1 . Recalling the
assumption n ≥ k2 , we have that t + k/n ≥ 9k/n, and therefore
Pr k(v · v ′ ) ≥ k(k/n + t) ≤ exp (−k · KL (t + k/n||k/n))
≤ exp (−kt ln (tn/k))
−kt
tn
n −kc
=
=
k
k2−c
≤ n−ck
c /2
.
ckc /4
c
Now if we let S be a set of ⌊nck /8 ⌋ unit vectors drawn independently from D, there are ⌊n 2 ⌋ <
c
⌊nck /2 ⌋ distinct pairs of v, v ′ ∈ S, and by a union bound the probability there exist distinct v, v ′ ∈ S with
c
c
(v · v ′ ) ≥ 2kc−1 is less than ⌊nck /2 ⌋n−ck /2 < 1. Thus, there exists a set S such that all distinct pairs
v, v ′ ∈ S satisfy |v · v ′ | ≤ 2kc−1 . This completes the proof.
Before we proceed with the proof of Theorem 6.6, we make a useful observation: By following the
proof of Proposition 3.3 using Lemma 6.7 instead of Lemma 3.7, mutatis mutandis, we obtain:
Proposition 6.8. Given a distribution A over R that satisfies Condition 3.2 for some m ∈ Z+ , and any
constant 0 < c < 1, consider the set of distributions Pv for v ∈ Sn that are k-sparse, with n ≥ max{2(m +
1) ln n, 8k2 }. For a given ǫ > 0, suppose that dTV (Pv , Pv′ ) > 2ǫ whenever |v·v ′ | is smaller than 1/8. Then,
any SQ algorithm which, given access to Pv (x) for an unknown k-sparse v ∈ Sn , outputs a p
hypothesis Q
c /8
ck
2(m+1)
−(m+1)(1/2−c)
with dTV (Q, Pv ) ≤ ǫ needs at least ⌊n
⌋≥k
queries to STAT(O(k)
χ2 (A, N (0, 1)))
(m+1)(1−2c)
2
or to VSTAT(O(k)
/χ (A, N (0, 1))).
The above proposition can be used for m = 1 to establish a similar but quantitatively somewhat weaker
SQ lower bound. We can make a crucial improvement to this proposition for the specific A we use in the
proof below.
Proof of Theorem 6.6. We select the one-dimensional distribution A as follows:
A = (1 − δ)N (ǫ, 1) + δN (−(1 − δ)ǫ/δ, 1) ,
where δ = ǫk−c/4 . Note that A has mean 0, i.e., matches m = 1 moments of N (0, 1).
2
2
2 2
We could use Facts 6.2 and 6.3 to obtain
√ χ (A, N (0, 1) ≤ O(ǫ exp(ǫ /δ )). However, this would
require the parameter δ to be equal to ǫ/ c ln k to get the required bounds from Proposition 6.8. The
issue here is that χ2 (A, N (0, 1)) is much bigger than the variance of A, which means that the correlation
inequality |χN (0,1) (Pv , Pv′ )| ≤ (v · v ′ )2 χ2 (A, N (0, 1)) is far from tight for most v and v ′ . For our choice
of A, we prove the following lemma:
35
Lemma 6.9. If A = (1 − δ)N (ǫ, 1) + δN (−(1 − δ)ǫ/δ, 1), then for v, v ′ ∈ Sn , we have
1 + |χN (0,1) (Pv , Pv′ )| ≤ exp ǫ4 (v · v ′ )2 /δ4 .
√
P
Proof. Let θ be the angle between v and v ′ . As in (3), we willP
use the expansion A(x) = √∞
i=0 ai Hei (x)G(x)/ i!.
i
As in (7), we also have the expansion Uθ A(x) = G(x) + ∞
i=2 ai cos θHei (x)G(x)/ i!. Thus, we can
write
χN (0,I) (Pv , Pv′ ) = χN (0,1) (A, Uθ A)
Z ∞
=
(A(x) − G(x))(Uθ A − G(x))/G(x)dx
−∞
! ∞
!
Z ∞ X
∞
X
√
√
ai Hei (x)G(x)/ i!
=
ai cosi θHei (x)G(x)/ i! /G(x)dx
−∞
=
∞
X
i=1
i=2
a2i cosi θ .
i=2
Since A is a distribution with mean zero, we have a0 = 1, a1 = 0. We need to take advantage of the fact
than χ2 (A, N (0, 1)).
that for our selected probability density function A, the coefficient a22 is much smaller
√
We can find the ai explicitly using (4), which gives that ai = EX∼A [Hei (x)/ i!]. We have the following
well-known fact:
Fact 6.10. For µ ∈ R, we have EX∼N (µ,1) [Hei (x)] = µi .
Proof. Note that the i-th derivative of G(x) isP
(−1)i Hei (x)G(x). Using Taylor’s theorem, we can expand
∞
i
G(x − µ) around x to obtain G(x − µ) =
i=0 µ Hei (x)G(x)/i! . Taking the expectation of Hei (x)
extracts the i-th term, establishing the fact.
Thus, we have
√
i!ai = EX∼A [Hei (x)]
= (1 − δ)EX∼N (ǫ,1) [Hei (x)] + δEX∼N (−(1−δ)ǫ/δ,1) [Hei (x)]
= (1 − δ)ǫi + δ(−(1 − δ)ǫ/δ)i .
√
In addition to a0 = 1,a1 = 0, we can derive the bound |ai | ≤ (ǫ/δ)i / i!. Recalling the special case a1 = 0
and summing over i, we have
|χN (0,I) (Pv , Pv′ )| ≤
≤
∞
X
i=1
∞
X
i=2
a2i | cos θ|i
(ǫ/δ)2i | cos θ|i /i!
= exp ǫ2 |v · v ′ |/δ2 − 1 − ǫ2 |v · v ′ |/δ2 .
To complete the proof of the lemma, it is sufficient to show that exp(x)−x ≤ exp(x2 ) for all x ≥ 0. We note
that both expressions are 1 and have derivative 0 at x = 0. It suffices to show that d2 (exp(x) − x)/dx2 ≤
d2 (exp(x2 )/dx2 for x ≥ 0. Note that
2
d2 ex d2 (ex − x)
2
2
2
2
/
= (2ex + 4x2 ex )e−x ≥ 2ex −x = 2e(x−1/2) −1/4 ≥ 2e−1/4 > 1 .
2
2
dx
dx
This completes the proof.
36
We now have all the necessary ingredients to complete the proof of Theorem 6.6. For distinct k-sparse
unit vectors v, v ′ ∈ S, where S is given by Lemma 6.7, we have that
|χN (0,1) (Pv , Pv′ )| ≤ exp ǫ4 (v · v ′ )2 /δ4 − 1
≤ exp ǫ4 4k2c−2 /δ4 − 1
≤ exp 4k2c−2 · kc − 1
= e4 exp k3c−2 − 1
≤ e4 k3c−2 .
Following the proof of Proposition 3.3, we now have that it takes Ω(nck
or to VSTAT(O(k)2−3c ) to learn Pv .
c /8
) queries to STAT(O(k)3c/2−1 )
7 Sample Complexity Lower Bounds for High–Dimensional Testing
In this section, we use our framework to prove information-theoretic lower bounds on the sample complexity
of our two high-dimensional testing problems: (i) robustly testing the mean of a single unknown mean
identity covariance Gaussian in Huber’s contamination model, and (ii) (non-robustly) testing between a
single spherical Gaussian and a mixture of 2 spherical Gaussians.
Both these statements follow from the structural results established in the previous sections using the
following proposition:
Proposition 7.1. Let A be a distribution on R such that A has mean 0 and χ2 (A, N (0, 1)) is finite. Then,
there is no algorithm that, for any n, given N < n/(8χ2 (A, N (0, 1))) samples from a distribution D over
Rn which is either N (0, I) or Pv , for some unit vector v ∈ Rn , correctly distinguishes between the two
cases with probability at least 2/3.
Proof. At a high-level, the proof of the proposition uses the structure of the set of Pv ’s and standard
information-theoretic arguments.
Suppose that, after fixing the dimension n, the algorithm takes at most N samples. We can consider the
testing algorithm as a (possibly randomized) function from N -tuples of samples to its output. For a distribution D, let D ⊗N denote the distribution over independent N -tuples drawn from D. We write f (D ⊗N )
for the Bernoulli distribution that gives the output of the algorithm given a single sample of D ⊗N . Let QN
be the distribution obtained by choosing v uniformly at random over the unit sphere Sn , and then drawing
N samples from Pv . Then, f (QN ) should be “NO” with probability at least 2/3, since the probability that
⊗N ) is “YES” with probability at least
each f (P⊗N
v ) is “NO” is at least 2/3. On the other hand, f (N (0, I)
2/3. By the data processing inequality, it follows that
dTV (QN , N (0, I)⊗N ) ≥ dTV (f (QN ), f (N (0, I)⊗N )) ≥ 1/3 .
Suppose for the sake of contradiction that N < n/(8χ2 (A, N (0, 1))). Then, we claim that
dTV (QN , N (0, I)⊗N ) < 1/3 .
37
Indeed, we have that:
4dTV (QN , N (0, I)⊗N )2 + 1 ≤ χ2 (QN , N (0, I)⊗N ) + 1 =
Z
Z
N
Y
(1)
(N ) 2
G(x(i) )dx(N ) . . . dx(1)
QN (x , . . . x ) /
···
=
=
=
=
≤
Z
Z
Z
x(1) ∈Rn
x(1) ∈Rn
v∈Sn
Z
Z
Zv∈Sn Z
v∈Sn
···
Z
v′ ∈Sn
v′ ∈S
v′ ∈S
x(N) ∈Rn
x(N) ∈Rn
Z
i=1
Z
x(1) ∈Rn
v∈Sn
···
Z
Z
v′ ∈S
(1)
(N )
(1)
(N )
PN
)PN
)/
v (x , . . . , x
v′ (x , . . . , x
n
N
Y
G(x(i) )dv ′ dvdx(N ) . . . dx(1)
i=1
N
Y
x(N) ∈Rn i=1
Pv (x(i) )Pv′ (x(i) )/G(x(i) )dx(N ) . . . dx(1) dv ′ dv
(1 + χN (0,I) (Pv , Pv′ ))N dv ′ dv
n
n
1 + |v · v ′ |2 χ2 (A, N (0, 1))
N
dv ′ dv ,
where the last line follows from Lemma 3.4, since A satisfies Condition 3.2 for m = 1. We will need the
following facts about the Beta function B(x, y):
R π/2
Fact 7.2.
(i) For x > −1, y > −1 we have that: 0 sinx (θ) cosy (θ) = B((x + 1)/2, (y + 1)/2)/2.
(ii) For all x, y ∈ R, we have that B(x, y + 1) = B(x, y) · y/(x + y).
Consider choosing v and v ′ independently uniformly at random over Sn . To obtain our tight sample
complexity results, we will need a more precise analysis for the distribution of the angle θ between v and
v ′ . Specifically, we show the following:
Lemma 7.3. If we choose v and v ′ uniformly at random from Sn , the angle θ between them is distributed
with the probability density function sinn−2 (θ)/B((n − 1)/2, 1/2), where B(x, y) is the Beta function.
Proof. Since a rotation of the sphere moves both v and v ′ , θ is invariant under such rotations. Thus, we
get the same distribution by fixing v ′ = e1 , the unit vector in the x1 -direction, and choosing v uniformly at
random over the sphere. Now we have that cos θ = x1 . Let Sn (r) denote the surface area of the sphere of
radius r in (n + 1) dimensions and note that Sn (r) = r n Sn (1).
For any measurable function f ,we have that
Z π
Z
Z
f (cos θ) sinn−2 (θ)Sn−2 (1)dθ .
f (x1 )Sn−2 (x2 )ds =
f (x1 )dΣ =
0
S2
Sn
We thus have that the pdf of θ is proportional to sinn−2 (θ). Taking f (x1 ) ≡ 1, note that
Z π
Z
sinn−2 (θ)Sn−2 (1)dθ = B((n − 1)/2, 1/2) ,
1dΣ =
Sn−2 (1) =
Sn
0
using Fact 7.2 (i). We thus have that the pdf of θ is
sinn−2 (θ)/Sn−2 (1) = sinn−2 (θ)/B((n − 1)/2, 1/2) .
This completes the proof.
38
We now have that
4d2TV
⊗N
(QN , N (0, I)
)+1≤
Z
Z
v′ ∈S
1 + |v · v ′ |2 χ2 (A, N (0, 1))
N
dv ′ dv
n
n
Zv∈S
π
N
=
1 + | cos θ|2 χ2 (A, N (0, 1)) sinn−2 (θ)/B((n − 1)/2, 1/2)dθ
0
=2
Z
π/2
1 + cos2 (θ)χ2 (A, N (0, 1))
0
N
sinn−2 (θ)/B((n − 1)/2, 1/2)dθ
N Z π/2
X
N
χ2 (A, N (0, 1))i cos2i (θ) sinn−2 (θ)/B((n − 1)/2, 1/2)dθ
2
=
i
0
i=0
N
X
N 2
=
χ (A, N (0, 1))i B((n − 1)/2, i + 1/2)/B((n − 1)/2, 1/2) .
i
i=0
PN
We rewrite the above sum as i=0 bi , where
N 2
bi =
χ (A, N (0, 1))i B((n − 1)/2, i + 1/2)/B((n − 1)/2, 1/2) .
i
N N
/ i = (N − i)/(i + 1) and using
Note that b0 = 1. Now consider the ratio bi+1 /bi . Note that that i+1
Fact 7.2, we have that B((n − 1)/2, i + 3/2)/B((n − 1)/2, i + 1/2) = (i + 1/2)/(i + n/2). Therefore, it
follows that
bi+1 /bi = χ2 (A, N (0, 1)) · (N − i)/(i + n/2) · (i + 1/2)/(i + 1) .
For all i, we have
bi+1 /bi ≤ χ2 (A, N (0, 1)) · N/(n/2) = 2N χ2 (A, N (0, 1))/n .
When N ≤ n/(8χ2 (A, N (0, 1))), we have bi+1 /bi ≤ 1/4 for i ≥ 0, and therefore
N
X
i=0
Hence, we have
bi ≤ 1/(1 − 1/4) = 4/3 .
4d2TV (QN , N (0, I)⊗N ) + 1 ≤ 1/3 < 4/9 .
This implies that dTV (QN , N (0, I)⊗N ) < 1/3, which is the desired contradiction. This completes the
proof.
Remark 7.4. It is worth noting that matching m > 1 many moments does not seem to help in the setting
of the previous proposition, as long as χ2 (A, N (0, 1)) ≤ 1. This may seem to some extent unsurprising,
given that O(n) samples suffice for some of the learning problems we consider here. On the other hand,
we consider it somewhat surprising looking at the proof of Proposition 7.1. Specifically, for general m, we
would have that
N
X
N 2
2
⊗N
4dTV (QN , N (0, I) ) + 1 ≤
χ (A, N (0, 1))i B((n−1)/2, i(m+1)/2+1/2)/B((n−1)/2, 1/2) .
i
i=0
Note that the ratio of one term to the next approximately grows as N χ2 (A, N (0, 1))i(m−1)/2 /n(m+1)/2 . For
this to be less than 1/2, for all 0 ≤ i ≤ N , we need N (m+1)/2 χ2 (A, N (0, 1)) ≤ O(n(m+1)/2 ). Thus, we
need at least N = Ω(n/(χ2 (A, N (0, 1)))2/m ) samples . This suggests that we should be able to obtain a
tighter lower bound if χ2 (A, N (0, 1)) > 1 using this technique. We omit the details here, as we are mainly
interested in the regime χ2 (A, N (0, 1)) ≤ O(1) for our applications in this paper.
39
Using Proposition 7.1, we establish the two main results of this section:
Theorem 7.5 (Sample Complexity Lower Bound for Robustly Testing Unknown Mean Gaussian). There is
no algorithm that, for every ǫ > 0 and positive integer n, given fewer than Ω(n) samples from a distribution
P on Rn which is promised to satisfy either (a) P = N (0, I) or (b) P = (1 − δ)N (µ, I) + δN1 , where
δ = ǫ/100, kµk2 ≥ ǫ and the noise distribution N1 is a spherical Gaussian, can distinguish between the
two cases with probability at least 2/3.
If instead, for any constant 0 < c < 1, we are promised that P = (1 − δ)N (µ, I) + δN1 , where
δ = ǫ/nc/4 in case (b), then no algorithm that takes less than Ω(n1−c ) samples can distinguish between (a)
and (b) with probability at least 2/3.
Proof. Let δ be the noise rate. We will take δ = ǫ/100 or δ = ǫ/nc/4 . In both cases, we select our
one-dimensional distribution to be the following:
A = (1 − δ)N (ǫ, 1) + δN (−(1 − δ)ǫ/δ, 1) .
Now Lemma 6.9 yields that for v, v ′ ∈ Sn , we have
1 + |χN (0,1) (Pv , Pv′ )| ≤ exp ǫ4 (v · v ′ )2 /δ4 .
We will not apply Proposition 7.1 directly but follow its proof using the aforementioned stronger correlation bound. We have:
Z
Z
(1 + χN (0,I) (Pv , Pv′ ))N dv ′ dv
4dTV (QN , N (0, I)⊗N )2 + 1 ≤
′
Zv∈Sn Zv ∈Sn
exp N ǫ4 (v · v ′ )2 /δ4 dv ′ dv
≤
′
n v ∈Sn
Zv∈S
π
exp N ǫ4 (cos θ)2 /δ4 sinn−2 (θ)/B ((n − 1)/2, 1/2) dθ
=
0
=
∞
X
i=0
(N ǫ4 /δ4 )i B((n − 1)/2, i + 1/2)/i!B((n − 1)/2, 1/2) .
Now note that the ratio of the (i + 1)-th term to the i-th term of the corresponding series is
2N ǫ4
N ǫ4 /δ4 i + 1/2
·
≤
.
i + 1 i + n/2
nδ4
When N ≤ nδ4 /2ǫ4 , the 0-th term is 1 and the ratio of the (i + 1)-th to i-th term is less than 1/4. Therefore,
the above sum is less than 1/(1 − 1/4) = 4/3, which implies that
√
dTV (QN , N (0, I)⊗N ) ≤ 1/ 12 < 1/3 .
Following the proof of Proposition 7.1, we conclude that no algorithm satisfying the necessary conditions
exists. To complete the proof, note that for δ = ǫ/100, we need at least Ω(n) samples. And for δ = n−c/4 ǫ,
we need at least Ω(n1−c ) samples.
Theorem 7.6 (Sample Complexity Lower Bound for Testing GMMs). Fix 0 < ǫ < 1 and n ∈ Z+ . There is
no algorithm that, given less than Ω(n/ǫ2 ) samples from a distribution P on Rn that is promised to be either
(a) P = N (0, I), or (b) P is a mixture of two Gaussians each with weight 1/2 and identity covariance, such
that dTV (P, N (0, I)) ≥ ǫ, distinguishes between the two cases with probability at least 2/3.
40
Proof. We choose our one-dimensional distribution as
A = (1/2)N (−δ, 1) + (1/2)N (δ, 1) ,
where we set (with hindsight) δ = Θ(ǫ1/2 ).
Note that A has mean 0. By Claims 6.2 and 6.3, we have that
χ2 (A, N (0, 1)) = (1/2) exp(δ2 )/2 + exp(−δ2 )/2 − 1 = (1/2) cosh(δ2 ) − 1 = Θ(δ4 ) .
Similarly, it can be shown that dTV (A, N (0, 1)) = Θ(δ2 ), which is Ω(ǫ) by our choice of δ. Applying
Proposition 7.1 completes the proof.
8 SQ Algorithms for Robustly Learning and Testing a Gaussian
The structure of this section is as follows: In Section 8.1, we prove a moment–matching structural result that
forms the basis of our algorithms. In Section 8.2, we present our robust testing algorithm, and in Section 8.3
we give our robust learning algorithm.
8.1 One-Dimensional Moment Matching Lemma The main result of this section is the following structural result:
p
Lemma 8.1. Let G ∼ N (0, 1). For δ > ǫ > 0, define k = 2⌈ǫ ln(1/ǫ)/δ⌉. Let G′ be an ǫ-noisy onedimensional Gaussian with unit variance so that for all t ≤ k we have that the tth moments of G and G′
agree to within an additive (t − 1)!(δ/ǫ)t ǫ/t. Then, we have that dTV (G, G′ ) = O(δ).
e = N (µ, 1) be such that dTV (G′ , G)
e ≤ ǫ. We can assume without loss of generality that µ > 0.
Proof. Let G
p
Looking at just the mean andp
variance suffices to get dTV (G, G′ ) = O(δ log(1/δ)) using techniques
similar to the proof of the O(ǫ log(1/ǫ)) filter algorithm from [DKK+ 16]. This allows us to focus on the
case that µ ≥ 1. Formally, we have the following claim:
e is at least 1.
Claim 8.2. Lemma 8.1 holds when the mean µ of G
Proof. We start by noting that we can assume ǫ > 0 is smaller that a sufficiently small universal constant.
Assuming otherwise and recalling that δ > ǫ gives that dTV (G, G′ ) ≤ 1 = O(δ), in which case the lemma
statement is trivial.
Let µ′ be the mean of G′ . We can write
e + ǫ′ E − ǫ′ L ,
G′ = G
(12)
e ≥ ǫ′ L .
G
(13)
e ≤ ǫ. Moreover, it holds
for distributions E, L with disjoint supports where ǫ′ = dTV (G′ , G)
Since k ≥ 2 by definition, the lemma assumptions imply that |µ′ | ≤ δ and EX∼G′ [X 2 ] ≤ δ2 /ǫ. By (12)
we have that µ′ = µ + ǫ′ EX∼E [X] − ǫ′ EX∼L [X], and similarly EX∼G′ [(X − µ′ )2 ] = 1 + (µ − µ′ )2 +
ǫ′ EX∼E [(X − µ′ )2 ] − ǫ′ EX∼L [(X − µ′ )2 ]. Therefore, we get
ǫ′ EX∼E [(X − µ′ )2 ] ≤ δ2 /ǫ − 1 − (µ′ − µ)2 + ǫ′ EX∼L [(X − µ)2 ] .
41
Note that PrX∼L [|X − µ′ | ≥ T ] ≤ PrX∼Ge [|X − µ′ | ≥ T ]/ǫ′ . As in Corollary 8.8 of [DKK+ 16], we
have that
Z ∞
′ 2
PrX∼L [|X − µ′ | ≥ T ]2T dT
EX∼L [(X − µ ) ] =
0
Z ∞
2T min 1, PrX∼N (µ,1) [|X − µ′ | ≥ T ]/ǫ′ dT
≤
Z0 ∞
2T min{1, exp(−((T − |µ − µ′ |)2 /2)2T /ǫ′ }dT
≤
0
Z √2 ln(1/ǫ′ )+|µ−µ′ |
Z ∞
e
=
2T dT + √
exp(−T 2 /2)2(T + |µ − µ′ |)/dTV (G′ , G)dT
2 ln(1/ǫ′ )
0
′
′ 2
= O(ln(1/ǫ ) + 1 + |µ − µ | ) .
By combining the above, we get that
ǫ′ EX∼E [(X − µ′ )2 ] ≤ δ2 /ǫ − 1 − (1 − O(ǫ′ ))|µ − µ′ |2 + O(ǫ′ ln(1/ǫ′ )) ,
and thus EX∼E [(X − µ′ )2 ] ≤ δ2 /(ǫǫ′ ).
′ 2
′ 2
′ 2
However, the means of L and E, µL and µE have
p |µL − µ | ′ ≤′ EX∼L [(X −pµ ) ] and |µE′ − µ | ≤
′
2
′
′
′
EX∼E [(X − µ ) ],p
and therefore ǫ |µL − µ | ≤ O(ǫ ln 1/ǫ′ + ǫ |µ − µ|) ≤ O(ǫ ln 1/ǫ + ǫ|µ − µ|) and
′
′
ǫ |µE − µ| ≤ O(ǫ δ2 /ǫǫ′ ) = O(δ).
p
We thus have that µ′ − µ = ǫ′ µE − ǫ′ µL satisfies |µ′ − µ| ≤ O(δ) + O(ǫ ln 1/ǫ) + O(ǫ|µ − µ′ |). Since
ǫ is sufficiently small, we have |µ′ − µ| ≤ O(δ) + 1/2. Since |µ′ | ≤ δ, We thus have that µ ≤ O(δ) + 1/2.
Since we assumed that µ ≥ 1, we get that δ = Ω(µ − 1/2) = Ω(1/2) and so µ = O(δ). We therefore
conclude that dTV (G, G′ ) = O(µ) = O(δ).
e = Θ(µ). Suppose for the sake
We will henceforth assume that µ ≤ 1 and thus we have that dTV (G, G)
′
e ≥
of contradiction that dTV (G, G ) = Ω(δ), for a sufficiently large constant in the big-Ω. Then, dTV (G, G)
2
e
dTV (G, G) − ǫ ≫ δ. We may assume that µ ≥ C δ for a sufficiently large constant C > 0. Thus, we have
that ǫ ≤ µ/C 2 ≤ 1/C 2 .
The proof will proceed as follows: Let f (x) = sin(xCǫ/µ). We note that f (x) has a simple expectation
e and we can easily get a lower bound on their difference. We will also use the Taylor series
under G or G,
for f (x) and our moment bounds to derive an upper bound on this difference which contradicts this lower
bound.
For x ∈ R+ , we want an expression for EX∼N (x,1) [f (X)]. By standard facts on the Fourier transform,
we have that
EX∼N (x,1) [exp(−iωX)] = exp(−iωx)EX∼N (0,1) [exp(−iωX)] = exp(−ω 2 /2 − iωx) ,
for any ω ∈ R. Since sin(xCǫ/µ) = (exp(−ixCǫ/µ) − exp(ixCǫ/µ))/2i, we obtain that
EX∼N (x,1) [f (X)] = exp(−(Cǫ/µ)2 /2) sin(xCǫ/µ) .
Therefore, EX∼N (0,1) [f (X)] = 0 and EX∼N (µ,1) [f (X)] ≥ exp(1/C) sin(Cǫ) > (C/2)ǫ, and thus
EX∼G′ [f (X)] ≥ EX∼Ge [f (X)] − ǫ > (C/3)ǫ.
(14)
Let h be the degree-(k − 1) Taylor polynomial of f plus the term (Cxǫ/µ)k /k!. By the Lagrange form of
the remainder in Taylor’s theorem, we have that h(x) − (Cxǫ/µ)k /k! − f (x) ≤ f (k) (ξ)xk /k!, for some
42
ξ ∈ [0, x]. Since k is even, we have that the k-th derivative of f , |f (k) (ξ)| = (Cǫ/µ)k | sin(ξCǫ/µ)| ≤
(Cǫ/µ)k . Thus, we get
f (x) ≤ h(x) ≤ f (x) + 2(xCǫ/µ)k /k! .
Our goal will be to show that EX∼G′ [h(X)] is substantially larger than EX∼N (0,1) [h(X)], which will contradict the assumption about approximately matching moments. We start by considering EX∼N (µ,1) [h(X)]
versus EX∼N (0,1) [h(X)]. We can write
EX∼N (µ,1) [h(X)] − EX∼N (0,1) [h(X)]
h
i
h
i
= EX∼N (µ,1) [f (X)] − EX∼N (0,1) [f (X)] + O EX∼N (µ,1) (XCǫ/µ)k /k! − EX∼N (0,1) (XCǫ/µ)k /k!
h
i
h
i
,
≥ (C/3)ǫ + O EX∼N (µ,1) (XCǫ/µ)k /k! − EX∼N (0,1) (XCǫ/µ)k /k!
where the last inequality follows from (14). To bound this latter term, we make the following claim:
Claim 8.3. We have that
√
|G(x − µ) − G(x)| ≤ O(µ)G(x/ 2).
Proof. First, we note that G(x − µ)/G(x) = exp(−2xµ + µ2 /2).
Recalling our assumption that 0 ≤ µ < 1, for |x| ≤ O(1/µ),
√ we have that |G(x − µ) − G(x)| ≤
2
O(|x|µ + µ )G(x) ≤ O((|x| + 1)µ)G(x).
√ Then, since G(x)/G(x/ 2) = O(G(x)) ≤ O(1/(|x| + 1)), we
get |G(x − µ) − G(x)| ≤p
O(µ)G(x/ 2).
For |x| ≥ 5/µ,√since 2 ln(1/µ) + 2 ≤ 2 ln(1/µ) + 3 ≤ 2/µ + 3 ≤ x, we have that G(|x| − 2) ≤ µ.
Thus, G(x)/G(x/ 2) ≤ O(G(x)) ≤ O(µ) and
√
G(x − µ)/G(x/ 2) = O(exp(−x2 /4 + µx − µ2 /2))
√
= O(G((x − 2µ)/ 2) exp(µ2 /2))
≤ O(G(|x| − 2)) ≤ O(µ) ,
√
and hence |G(x − µ) − G(x)| ≤ O(µ)G(x/ 2).
Using the previous claim, we note that
Z
∞
−∞
√
(xCǫ/µ)k /k! |G(x − δ) − G(x)| dx = O(ǫ)EX∼N (0,1/2) [(Cǫ/µ)k /k!] = O(ǫ/(C/ 2)k ) = O(ǫ).
Therefore, we have that
EX∼Ge [h(X)] ≥ EX∼G [h(X)] + (C/3)ǫ.
Next, we wish to compare EX∼G′ [h(X)] to EX∼Ge [h(X)]. Using (12), we have the following for ǫ′ =
e
dTV (G′ , G):
EX∼G′ [h(X)] = EX∼Ge [h(X)] + ǫ′ (EX∼E [h(X)] − EX∼L [h(X)])
h
i
≥ EX∼Ge [h(X)] + ǫ′ EX∼E [f (X)] − EX∼L [f (X)] − EX∼L (Xǫ/δ)k /k!
h
i
= EX∼Ge [h(X)] + O(ǫ) − ǫ′ EX∼L (XCǫ/µ)k /k! .
e ≥ ǫ′ L, it follows that L satisfies the concentration inequality
From (13), i.e., G
Pr [|X − µ| > T ] ≤ 2 exp(−T 2 /2)/ǫ′ .
X∼L
43
We now proceed to bound the subtractive term from above:
h
i
ǫ′ EX∼L (XCǫ/µ)k /k!
≤ ǫ′
∞
X
J=0
Pr
X∼L
h
|X − µ| ≥ 2J
p
log(1/ǫ)
i
k
p
2(J + 1) log(1/ǫ) + µ (Cǫ/µ)k /k!
∞
k
p
X
p
2ǫJ (2(J + 2) log(1/ǫ)Cǫ/µ)k /k!
≤ ǫ 3 log(1/ǫ)Cǫ/µ /k! +
J=1
≤ǫ·O
p
log(1/ǫ)Cǫ/µ
k
/k! ·
1+
∞
X
ǫJ (J + 2)k
J=1
!
k √
p
log(1/ǫ)ǫ/Cδ /k! ≤ ǫ · O
log(1/ǫ)ǫ/Cδ /( 2πk(k/e)k )
≤ǫ·O
p
√
√
≤ ǫO( log(1/ǫ)ǫ/Cδk)k / k ≤ O(ǫ/ kC k ) ≤ O(ǫ) .
p
k
Therefore, we conclude that
EX∼G′ [h(X)] ≥ EX∼N (µ,1) [h(X)] + O(ǫ) ≥ EX∼N (0,1) [h(X)] + (C/4)ǫ .
On the other hand, P
recalling that h is the degree-(k − 1) Taylor expansion of f (x) = sin(xCǫ/µ), we
k−1
i
i
i
can write h(x) =
i=0 ai x , with ai = O((Cǫ/µ) /i!) = O((ǫ/Cδ) /i!). Therefore, the difference
EX∼G′ [h(X)] − EX∼N (0,1) [h(X)] is the sum over i of ai times the difference in the ith moments, which
by assumption is at most
ǫ
k−1
k−1
X
X
(1/i)2 = O(ǫ) .
(i − 2)!(δ/Cǫ)i O((ǫ/δ)i /i!) = O(ǫ)
i=1
i=0
This contradicts the fact that their difference is at least (C/4)ǫ, and concludes the proof.
8.2 Robust Testing Algorithm In this subsection, we give a robust testing algorithm, i.e., an algorithm
that distinguishes between an ǫ-noisy Gaussian and N (0, I). This algorithm will form the basis for our
robust learning algorithm of the following subsection.
Theorem 8.4 (Robust Testing Algorithm). Let G′ be an ǫ-noisy version of an n-dimensional Gaussian
with identity covariance. Let δ be at least a sufficiently large constant multiple of ǫ. There
p exists an SQ
algorithm that makes O(nk ) queries to STAT(ǫ · O(n log(n/ǫ)2 )−k ) where k = 2⌈O(ǫ log(1/ǫ)/δ)⌉,
and distinguishes between the cases that G′ is the standard normal distribution N (0, I), and the case that
G′ is at least δ-far from N (0, I). The algorithm has running time nO(k) .
By simulating the statistical queries with samples, we obtain:
Corollary 8.5. Given sample access to G′ , an ǫ-noisy version of an n-dimensional Gaussian with identity
covariance and ǫ, δ > 0 with δ be at least a sufficiently large constant multiple of ǫ, there is an algorithm that
with probability 9/10 distinguishes between the cases that G′ is the standard normal distribution N (0, I),
and the case that G′ is at leastp
δ-far from N (0, I) and requires at most (n log(1/ǫ))O(k) /ǫ2 samples and
running time where k = 2⌈O(ǫ log(1/ǫ)/δ)⌉.
e = N (µ, I) be such that dTV (G′ , G)
e ≤ ǫ. We need
Proof. In the case when G′ is δ-far from N (0, I), let G
′
′
to show that we can distinguish between the cases G = N (0, I) and G is an ǫ-noisy version of a Gaussian
44
e = N (µ, I) with dTV (G,
e G) ≥ δ − ǫ. We assume from now on that the completeness case is to show that
G
e
dTV (G, G) ≥ δ, since replacing δ with δ + ǫ does not affect the statement of the theorem.
The algorithm is quite simple. Let C be a sufficiently large universal constant such that the O(δ) total
variation distance bound in Lemma 8.1 is less than Cδ. We assume that δ > 4Cǫ.
Robust Testing Algorithm:
• For each coordinate axis 1 ≤ i ≤ n, use STAT to approximate PrX∼G′ [X ≤ ǫ] and PrX∼G′ [X ≥ ǫ]
to within ǫ/2. If any of these approximations are bigger than 1/2 + ǫ, then output “NO”.
p
• Let k = 2⌈2Cǫ ln(1/ǫ)/δ⌉. Let C ′ be a sufficiently large constant.
p
• Using access to STAT, find all the mixed moments of X ∼ G′ , conditioned on kXk2 ≤ C ′ k n log(n/ǫ),
of order at most k to within error n−k/2 ǫ/2.
• If the difference between any moment of order t ≤ k that we measured and that of N (0, I) is more
than ((t − 1)!(δ/2Cǫ)t /t − 1) · n−k/2 ǫ, then output “NO”.
• Otherwise, output “YES”.
The idea is to use Lemma 8.1 with the approximations the moments. However, we have the issue that
the STAT oracle
p can only be used to approximate the expectation of a bounded function. Using the condition
kXk2 ≤ C ′ k n log(n/ǫ) allows us to avoid this. But we first need to show that conditioning on it does not
affect the moments too much and does not move the distribution far in total variational distance.
Note that the first step of the algorithm will reject if the median of N (µ, I) projected onto any coordinate
axis is outside of the interval [−ǫ, ǫ]. If this occurs, then µ 6= 0. If this step does not reject, then µ projected
√
onto any coordinate axis is O(ǫ), and
so kµk2 ≤ O(ǫ n).
p
The condition kxk2 ≤ C ′ k n log(n/ǫ) ensures that the any degree less than k monomial in X
is at most (C ′ nk2 log(n/ǫ)))k/2 . Thus, we can approximate this expectation to precision n−k/2 ǫ using
√
STAT(ǫ · O(n log(n/ǫ)2 )−k ). However, since kµk2 ≤ O(ǫ n), by standard concentration bounds, the
probability that X ∼ N (µ, I) does
not satisfy this condition is at most ǫ · (C ′ nk2 log(n/ǫ)))−k . Let G′′ be
p
G′ conditioned on kXk2 ≤ C ′ k n log(n/ǫ). If G′ = N (0, I), then we need to show that the moments of
G′′ and G′ are within ln−k/2 (ǫ/2). To show this, we use the following lemma:
Lemma 8.6. For 0 ≤ δ ≤ exp(−k), the difference
between any mixed moment of degree at most k of
p
N (0, I) and N (0, I) conditioned on kxk2 ≤ O( nk log(1/δ)), is at most δ.
Proof. Let X be distributed as N (0, I). Let ǫ be δ/(k + ln 1/δ)(k−1) C for ap
sufficiently large constant
C. Thus,
ln(1/ǫ) = ln(1/δ) + ln C + (k − 1) ln(k + ln(1/δ)). Let T be 2n log 1/ǫ. Then, T =
p
OC ( nk log(1/δ)) and by standard concentration inequalities, we have that Pr[kXk2 ≥QT ] ≤ ǫ.
For a ∈ Nn with kak1 ≤ k, consider the monomial of degree
ma (x) = ni=1 xai i . First we
Qnat most k,
E(Xiai ) = 0, sinceQthe odd moments
consider its mean and variance. If any ai is odd, E(ma )(X) = i=1 Q
n
ai
ai /2 (a /2)! ≤
of Xi ∼ N (0, 1) are zero. If all ai are even, then E(ma )(X) = Q
i
i=1 E(Xi ) =
i2
Q
2a
n
a a ! ≤ 2k k!.
2] =
i
E(X
)
=
2
2k/2 (k/2)!. For the variance,
we
have
Var[m
(X)]
≤
E[m
(X)
i
a
a
i
i=1
i
i
√
Let pa (X) = ma (X)/2k/2 k!. Then we have that 0 ≤ E[pa (X)] ≤ 1 and Var[pa (X)] ≤ 1.
By the standard concentration inequality given in Lemma 8.16 below, we have that for all t > 0,
Pr[|pa (X)| ≥ t + 1] ≤ exp(2 − (t/R)2/k ) for some R > 0. Thus, we have Pr[|pa (X)| ≥ c + 1] ≤ ǫ, for
45
c = R(ln(1/ǫ) − 2)k/2 . Let I(x) be the indicator function of kxk2 ≥ T . Then we have that
|E[I(X)pa (X)]| ≤ E[I(X)|pa (X)|]
Z ∞
Pr[I(X)|pa (X)| ≥ t]dt
=
0
Z ∞
Z c+1
exp(2 − (t − 1/R)2/k )dt
ǫdt +
≤
c+1
0
Z ∞
exp(2 − (t/R)2/k )dt
= (c + 1)ǫ +
Zc ∞
= (c + 1)ǫ +
exp(2 − x)(dt/dx)dx
ln(1/ǫ)
Z ∞
exp(2 − x)xk/2−1 dx
= (c + 1)ǫ + (Rk/2)
(where x = (t/R)2/k )
ln(1/ǫ)
k/2−1
= (c + 1)ǫ + (Rk/2)ǫ ·
k/2
≤ R ln(1/ǫ)
where the integral
ma (x), we have
R∞
2
X
j=0
((k/2)!/(k/2 − j)!) ln(1/ǫ)k/2−j
ǫ + (k /8)ǫ(k + log(1/ǫ))k/2−1
≤ O(k2 ǫ(k + log(1/ǫ))k/2−1 ) ,
ln(1/ǫ) exp(2
− x)xk/2−1 dx is calculated explicitly below in Claim 8.18. In terms of
√
|E[I(X)ma (X)]| ≤ O(k2 ǫ(k + log(1/ǫ))k/2−1 2k/2 k!) ≤ O(ǫ(k + log(1/ǫ)(k−1) ) ≤ O(δ/C) ≤ δ/2 .
Then, for X ′ distributed as N (0, I) conditioned on kX ′ k2 ≤ T , we have
|E[ma (X)] − E[ma (X ′ )]| = |E[ma (X)] − E[ma (X)(1 − I(X))]/(1 − Pr[kXk2 > T ])|
= |(E[I(X)ma (X)] − E[ma (X)] Pr[kXk2 > T ]) /1 − Pr[kXk2 > T ])|
√
≤ δ/2 + 2k/2 kǫ /(1 − ǫ)
≤ 2δ/3(1 − ǫ) ≤ δ .
p
p
Applying Lemma 8.6 for δ = n−k/2 ǫ, noting that C ′ k n log(n/ǫ) = Ω(C ′ nk log(1/δ)) yields that
the moments of G′′ and G′ = N (0, I) are within ln−k/2 (ǫ/2). Thus, in this case, the approximations of the
moments of G′′ are within n−k/2 ǫ of the moments of G′ .
For the soundness case, we just note that since (t − 1)!(δ/2Cǫ)t ǫ/t − 1 ≥ (δ/2Cǫ)/2 − 1 ≥ 1, the
bounds on the moments we need to fail are bigger than the precision of the statistical queries we use to
approximate them, and therefore we never output “NO” when G′ = N (0, I).
e Then G′′ is a 2ǫ-noisy
Now suppose that G′ is an ǫ-noisy version of an identity covariance Gaussian G.
e
e
version of G. We will denote µ the mean vector of G and will assume that kµk2 ≥ δ. We need to show that
the algorithm outputs “NO”.
Consider the unit vector v = µ/kµk2 which has v · µ ≥ δ. Consider the one-dimensional distributions
ev of the form v · X, where either X ∼ G′′ or X ∼ G.
e Note that G
ev = N (kµk2 , 1) has mean larger
G′′v and G
′′
ev ) ≤ 2ǫ. We can now apply the contrapositive of Lemma 8.1 with δ/C in place
than δ and that dTV (Gv , G
46
of δ and 2ǫ in place of ǫ, which implies that there is a t ≤ k such that the t-th moment of G′v is more than
(t − 1)!(δ/2Cǫ)t ǫ/t far from that of N (0, 1). That is,
|EX∼G′ [(v · X)t ] − EX∼N (0,I) [(v · X)t ]| ≥ (t − 1)!(δ/2Cǫ)t ǫ/t .
Q ai
n
Now consider the polynomial (v·x)t . Note that the
coefficient
of
a
monomial
i xi , for a ∈ Z>0 , kak1 = t,
Q
n
ai
t
is given by the multinomial theorem as a1 ,...,a
i=1 vi . The L1 -norm of its coefficients is the same as the
n
Q
L1 -norm of entries of the rank-t tensor v ⊗t , which has i-th entry tj=1 vij , for i ∈ {0, . . . , n}t , since there
Q
t
entries in this symmetric tensor which are given by ni=1 viai for any a. The Frobenius norm of
are a1 ,...,a
n
Q
P
P
Q
v ⊗t , the L2 -norm of its entries, is i∈{0,...,n}t tj=1 vi2j = tj=1 ni=1 vi2 = 1. Since there are nt entries,
the L1 -norm of its entries must be at most nt/2 ≤ nk/2 . We can write EX∼G′′ [(v·X)t ]−EX∼N (0,I) [(v·X)t ]
as a linear combination of differences in moments with these coefficients, and we can thus bound this from
above by the product of the L1 -norm of
Qthe coefficients and the L∞ -norm of the differences in moments.
Hence, there must be some moment E[ i Xiai ] with a ∈ Zn>0 , kak ≤ k, such that
"
"
#
#
Y a
Y a
EX∼G′
Xi i − EX∼N (0,I)
Xi i ≥ (t − 1)!(δ/2Cǫ)t ǫ/t · n−k/2 .
i
i
This in turn means that the difference in the approximation of this moment of G′′ and that of N (0, I) is at
most ((t − 1)!(δ/2Cǫ)t /t − 1) · n−k/2 ǫ. Thus, the testing algorithm outputs “NO”.
8.3 Robust Learning Algorithm In this section, we build on the testing algorithm of the previous section
to design our robust learning algorithm. Formally, we prove:
Theorem 8.7. Let G′ be an ǫ-noisy version of an n-dimensional Gaussian with identity covariance matrix,
N (µ, I) with kµk2 ≤ poly(n, ǫ). There is an algorithm that, given statistical query access√to G′ , outO( log(1/ǫ))
+√
puts an approximation
µ
e to the mean µ such that√
kµ − µ
ek2 ≤ O(ǫ). The algorithm uses
√
√ n
O(
2log(1/ǫ)
log(1/ǫ))
calls to STAT(ǫ/(n ln(1/ǫ))O(
log(1/ǫ))
) and has running time nO(
log(1/ǫ))
O(
+2log(1/ǫ)
log(1/ǫ))
By simulating the statistical queries with samples, we obtain:
Corollary 8.8. Given sample access to G′ , an ǫ-noisy version of an n-dimensional Gaussian N (µ, I),√there
is an algorithm that
µ
e with kµ−e
µk2 ≤ O(ǫ) and requires (n log(1/ǫ))O( log(1/ǫ)) /ǫ2
√
√ with probability 9/10 outputs
samples and nO(
log(1/ǫ))
O(
/ǫ2 + 2log(1/ǫ)
log(1/ǫ))
time.
The work [DKK+ 16] gives algorithms which can compute an approximation µ′ with kµ − µ′ k2 ≤
p
O(ǫ log(1/ǫ)). These algorithms can be expressed as Statistical Query algorithms. However, due to the
model of adversary used for robustness in [DKK+ 16], the algorithms were expressed there in terms of operations on sets of samples that were drawn before the execution of the algorithm. The filtering algorithms
work by successively removing samples from this set and then computing expectations of the current set of
remaining samples. The samples that are removed are those that satisfy an explicit condition, we say that
they are rejected by a filter. We can implement these algorithms as SQ algorithms by replacing expectations
of the current set of remaining samples with the conditional expectation of the input distribution, conditioned
on all previous filters accepting. This is similar to the filtering algorithm for learning binary Bayesian networks given in [DKS16c]. Even there, we still used samples to compute the threshold for the filter. We note
that using arguments similar to those we use for the algorithm below, all theses algorithms can be expressed
as SQ algorithms. In particular, this is the case for Algorithm Filter-Gaussian-Unknown-Mean,
p
which we will use as a black box pre-processing step to approximate the mean within O(ǫ log(1/ǫ)).
Instead of dealing with moments, i.e., the expectations of monomials, directly, we will consider expectations of Hermite polynomials, which have a simpler form for normal distributions.
47
.
Definition Q
8.9. We define multi-dimensional normalized
Hermite polynomials as follows: for a ∈ Zn ,
Q
Hea (x) = ni=1 Heai (xi ). We define n(a) = ni=1 ai !.
Thus we have the following:
Fact 8.10. EX∼N (0,I) [Hea (X)Heb (X)] = δab n(a).
We define a linear combination of these Hermite polynomials with degree t associated with a tensor
of rank t. For i ∈ {0, . . . , n}t , we define the count vector c(i) ∈ Zn≥0 such that c(i)j is the number of
coordinates of i that are j.
Definition 8.11. For a rank-t tensor A over Rn , we define
X
√
hA (x) =
Hec(a) (x)/ t! .
i∈{0,...,n}t
We are now ready to describe our learning algorithm.
Robust Learning Algorithm:
p
1. Let k = 2⌈ ln(1/ǫ))⌉.
p
2. Compute an approximation µ′ with kµ′ − µk2 ≤ O(ǫ log(1/ǫ)) by iterating Algorithm
Filter-Gaussian-Unknown-Mean from [DKK+ 16]. We change the origin so that µ′ = 0.
p
3. Let N be the filter that accepts when kxk2 ≤ 2n log(1/ǫ).
√
4. For 1 ≤ t ≤ k, let Pet be the rank-t tensor with i1 , . . . , it entry given by t! times the result of asking
an SQ oracle for EX∼G′ [hc(i) (X)] conditioned on N accepting to within precision ǫ/nt/2 .
5. While kPet kF ≥ ǫΩ(log(1/ǫ))t/2 for some t,
′
• Let t′ be the least t such that kPet′ kF ≥ ǫΩ(log(1/ǫ))t /2 .
√
P
• Let A = Pet′ /kPet′ kF . Let hA (x) =
i1 ,...it′ Ai Hec(i) (x)/ t! For each positive integer T ,
approximate
Pr [|hA (X)| ≥ T + 1]
e
X∼G
until one is found that is at least
′
′
3 exp(2 − Ω(T )2/t )) + ǫ/Cn2t
for a sufficiently large constant C. Let F be the filter that accepts when |hA (x)| ≤ T + 1.
• Recalculate Pet , for all 1 ≤ t ≤ k, where all expectations are conditioned N and the filters F
from all previous iterations.
6. End while.
7. For 1 ≤ t ≤ k, compute the SVD of M (Pet ), Pet considered as an n × nt−1 matrix, and let Vt ⊆ Rn
be the subspace spanned by all right singular vectors of Pt with singular value more than ǫ.
8. Let V be the span of V1 , . . . , Vk .
9. Let S ⊂ V be a set of unit vectors of size | dim(V )|O(dim(V )) such that for any unit vector v ∈ V ,
there is a v ′ ∈ S with kv − v ′ k2 ≤ 1/2.
48
√
10. For each v ∈ S, compute the median mv of v T X, for X ∼ G′ , to within ǫ/ pdim V using bisection
and statistical queries to approximate the Pr[v T X ≤ m] for m = µ′ + O(ǫ log(1/ǫ)). (We don’t
need to condition on any filters here).
11. Find a feasible point µ
eV of the LP µ
eV ∈ V with |v T µ
eV − v T mv | ≤ O(ǫ) for all v ∈ S.
12. Return µ
eV .
Using similar techniques to those used to express this algorithm in terms of Statistical Queries, we can
e
run Algorithm Filter-Gaussian-Unknown-Mean using poly(n/ǫ) time and calls to STAT(O(ǫ/poly(n))).
Note that we can approximate conditional expectations easily as a ratio of expectations approximated
by two SQ queries. Since, as we will show, our filters only throw away at most an O(ǫ) fraction of points,
we will not need to increase the precision beyond a constant factor to
√ do this.
O( log(1/ǫ))
. To show that we can use the
The algorithm needs approximate
expectations to within ǫ/n
√
O( log(1/ǫ))
) to obtain this, we need to note that√the distributions we approxioracle STAT(ǫ/(n ln(1/ǫ))
mate the expectations of are supported p
in an interval of length (n ln(1/ǫ))O( log(1/ǫ)) . Thanks to the naive
pruning of Step 3, only x with kxk2 ≤ 2n log(1/ǫ) contribute to these expectations. This suffices
√ to show
that the maximum value of all polynomials we consider on any such x is at most (n ln(1/ǫ))O(
We need to show the following for the filter step of our algorithm:
log(1/ǫ))
.
Proposition 8.12. The loop in Step 5 takes O(n2k ) iterations and all filters together accept with probability
at least 1 − O(ǫ).
8.3.1 Proof of Proposition 8.12 We now proceed with the proof. By standard concentration bounds, N
accepts with probability 1 − O(ǫ). Let F ′ be the event that N and all filters F from previous iterations
accept. We assume inductively that PrG′ [F ′ ] ≤ O(ǫ), and need to show that the same holds if we include
the filter F produced in the current iteration.
p
e = N (µ, I) be a Gaussian with dTV (G′ , G)
e ≤ ǫ such that kµk2 ≤ O(ǫ log(1/ǫ)). We write
Let G
G′ |F ′ for the distribution obtained by conditioning on F ′ . Since PrG′ [F ′ ] ≤ Cǫ, we have that
e ≤ dTV (G′ , G)
e + dTV (G′ , G′ |F ′ ) ≤ (C + 1)ǫ .
dTV (G′ |F, G)
e + dTV (G′ |F, G)E
e − dTV (G′ |F, G)L,
e
Thus, for the first iteration, we have G′ |F ′ = G
for distributions E
and L with disjoint supports.
e + wE E − wL L, for distributions E and L with disjoint
For any iteration, we will write G′ |F ′ = wGe G
supports, where wE + wL = O(ǫ) and wGe = 1 + O(ǫ). In the first iteration, we will take wGe = 1 and
e
wE = wL = dTV (G′ |F ′ , G).
We will need properties of the polynomials hA (x) for the analysis. In particular, we show the following:
Lemma 8.13. Given a rank-t symmetric tensor A over Rn , let hA (x) be as in Definition 8.11. Then, we
have:
(i) EX∼N (0,I) [hA (X)2 ] = kAk2F .
(ii) If
PB is a rank-t tensor with Bi =
i Ai Bi .
√
t!EX∼P [Hec(i) (X)], for a distribution P, then EX∼P [hA (X)] =
(iii) We can recover A from hA (x) using
√
t!Ai1 ,...,it =
∂
∂xi1
· · · ∂x∂i hA (x).
t
(iv) If O is an orthogonal matrix, then hA (Ox) = hB (x) for a symmetric rank-t tensor B with kBkF =
kAkF .
49
√
(v) If B is a rank-t tensor with Bi = t!EX∼P [Hec(i) (X)], for a distribution P, and j > 0, v ∈ Rn ,
√
then EX∼P [Hej (v · X)]/ t! = B(v, . . . , v).
Proof. For (i), we first need
Pto get an expression for√the coefficients of each Hea (x), since they appear
multiple times in hA (x) = i∈{0,...,n}t Ai Hec(a) (x)/ t!. Let c−1 (a) be a function mapping a ∈ Zt≥0 with
−1
kak1 = t to i ∈ {1, . . . , n}t , with c(c−1 (a)) = a for all a. Since A is
symmetric, the choice of c does
t
not affect Ac−1 (a) for any a. Note that, for a given a, there are a1 ,...,an = t!/n(a) possible i with c(i) = a.
Thus, we have:
X
√
hA (x) =
Ai Hec(a) (x)/ t!
i∈{0,...,n}t
X
=
kak1 =t
=
√
t!/n(a) · Ac−1 (a) Hea (x)/ t!
X √
t!/n(a) · Ac−1 (a) Hea (x) .
kak1 =t
Now, by orthogonality of Hea (x) with distinct a, we have that, for X ∼ N (0, I), it holds:
X
E[hA (X)2 ] =
t!/n(a)2 · A2c−1 (a) E[Hea (X)2 ]
kak1 =t
=
X
kak1 =t
=
X
i
t!/n(a) · A2c−1 (a)
A2i = kAk2F .
For (ii), we now have that
√ X
EX∼P [hA (X)] = 1/ t! ·
Ai EX∼P [Hec(i) (x)]
=
X
i
Ai Bi .
i
Q
For (iii), note that a with kak1 = t, Hea (x) has only one monomial of degree t, which is xai i . Thus,
given i ∈ {1, . . . , n}t , there is only one a with kak1 = t and ∂x∂i · · · ∂x∂i Hea (x) 6= 0, which is a = c(i)
t
1
and has
∂
∂
∂
∂ Y c(i)j
Hec(i) (x) =
xj
···
···
∂xi1
∂xit
∂xi1
∂xit
j
Y
=
c(i)j ! = n(c(i)) .
j
Thus, we have
√
∂
∂
hA (x) = n(c(i)) · t!/n(a) · Ac−1 (a)
···
∂xi1
∂xit
√
= t!Ai .
For (iv), consider the linear transformation of rank-t tensors O⊗t , which for our purposes can be defined as
the unique function such that (O ⊗t A)(v1 , . . . , vt ) = A(OT v1 , . . . , OT vt ), for all rank-t tensors over Rn , A,
50
and vectors v1 . . . vt ∈ Rn . It is a fact that kO⊗t AkF = kAkF . Now,
P we∂gconsider the t-th order directional
derivatives of a function f , ∇v1 , · · · ∇vt f (x), where ∇v g(x) = i vi ∂xi (x). These can be expressed in
terms of the rank-t tensor F , with Fi = ∂x∂i · · · ∂x∂i f (x), as ∇v1 , . . . ∇vt f (x) = F (v1 , . . . vt ). We have
t
1
that, for all i,
∂
∂
hA (Ox)
···
∂xi1
∂xit
= ∇ei1 · · · ∇eit hA (Ox)
= ∇OT ei · · · ∇OT eit hA (x)
1
√
= t!A(OT ei1 , . . . , O T eit )
√
= t!(O⊗t A)i
∂
∂
h ⊗t (a) .
...
=
∂xi1
∂xit O A
Since hA (Ox) and hO⊗t A (x) are both multivariate polynomials of degree t, that these derivatives agree
means that the coefficients of all monomials of degree-t agree. We thus have hA (Ox) = hO⊗t A (x) +
p(x), where p is a polynomial of degree at most t − 1. Since hO⊗t A (x) is a linear combination of Hermite polynomials of degree t, which are orthogonal to all polynomials of degree smaller than t, we have
EX∼N (0,I) [hO⊗t A (X)p(X)] = 0, and so
1 = kAkF = EX∼N (0,I) [hA (Ox)2 ]
= EX∼N (0,I) [hO⊗t A (X)2 ] + EX∼N (0,I) [p(X)2 ] + 2EX∼N (0,I) [hO⊗t A (X)p(X)]
= kO ⊗t AkF + EX∼N (0,I) [p(X)2 ] + 0
= 1 + EX∼N (0,I) [p(X)2 ] .
Since EX∼N (0,I) [p(X)2 ] = 0, we must have p(x) ≡ 0, and thus
hA (Ox) = hO⊗t A (x) .
Taking B = O⊗t A gives (iv).
For (v), let O be an orthogonal matrix that gives a rotation mapping e1 to v, a = {t, 0, . . . , 0} and T1 be
the rank-t tensor with (1, . . . , 1) entry 1 and every other entry 0. Then, we can rewrite
√
√
Het (v · x) = Het ((Ox)1 ) = Hea ((Ox)1 ) = t!hT1 (Ox) = t!hO⊗t T1 (x) .
For any i, we have
(O⊗t T1 )i = (O ⊗t T1 )(ei1 , . . . , eit )
= T1 (OT ei1 , . . . , O T eit )
=
t
Y
(OT eij )1
j=1
=
t
Y
(Oe1 )ij
j=1
=
t
Y
vi j .
j=1
51
Thus we have O⊗t T1 = v ⊗t , the tensor with entries
Qt
j=1 vij ,
and so
√
EX∼P [Het (v · X)]/ t! = EX∼P [hv⊗t (X)]
=
X
Bi
i
t
Y
vij = B(v, . . . , v) .
(by (ii))
j=1
This completes the proof.
√
We write Pt′ or Gt′ for the rank-t′ tensor with entries t!E[hc(i) (X)], where X is distributed according
e respectively. We know that kPet′ kF ≥ ǫΩ(log(1/ǫ)t′ /2 ).
to G′ |F ′ or G
′
Lemma 8.14. When kPet′ kF ≥ ǫΩ(log(1/ǫ)t /2 ), we have |EX∼G′ [hA (X)]| ≥ ǫ · Ω(log(1/ǫ))t/2 .
Proof. The assumption on the SQ errors imply that the corresponding entries of Pt′ and Pet′ are within
′
ǫ/nt /2 . It follows that
√
kPt′ − Pet′ kF ≤ t′ !ǫ .
Using Lemma 8.13 (iii), we have
EX∼G′ [hA (X)] =
X
Ai (Pt′ )i
i∈[n]t′
= 1/kPet kF ·
X
i∈[n]
t′
(Pet′ )i (Pt′ )i
= 1/kPet kF · kPet′ k2F +
X
i∈[n]
t′
(Pet′ )i (Pt′ − Pet′ )i
≥ kPet′ kF − kPt′ − Pet′ kF ≥ kPet′ kF −
p
t′
′
≥ ǫΩ(log(1/ǫ)t /2 )) − ǫ log(1/ǫ)
√
t′ !ǫ
′
≥ ǫΩ(log(1/ǫ)t /2 ) .
Lemma 8.15. EX∼Ge [hA (X)] = O(ǫ
p
log(1/ǫ))t and EX∼Ge [hA (X)2 ] = O(1).
Proof. Since for all a ∈ Zk>0 , EX∼N (0,I) [Hea (X)] = 0, we have EX∼N (0,I) [hA (X] = 0. By Lemma 8.13
(i), EX∼N (0,I) [hA (X)2 ] = kAk2F = 1.
e = N (µ, I) instead of N (0, I). Consider a rotation given
We need to take these expectations under G
by an orthogonal matrix O that maps kµk2 e1 to µ. By Lemma 8.13 (i), there is a symmetric rank-t tensor
B with kBkF = 1 such that hA (OX) = hB (X). Now we have that
EX∼G′ [hA (X)] = EX∼N (kµk2 e1 ,I) [hA (OX)] = EX∼N (kµk2 e1 ,I) [hB (X)] ,
and similarly for hA (X)2 .
n−1
for the vector dropping the first coordinate of a, we have
For a ∈ Zn≥0 with a 6= 0, writing a−1 ∈ Z≥0
EX∼N (kµk2 e1 ,I) [Hea (X)] = EX∼N (0,I) [Hea−1 (X)]EX∼N (kµk2 ,1) [Hea1 (X)]
= δa−1 ,0 EX∼N (0,1) [Hea1 (X + kµk2 )] .
52
Note that there is only one index i such that c(i) is zero in all except the first coordinate, and so we have
√
EX∼Ge [hA (X)] = EX∼N (kµk2 e1 ,I) [hB (X)] = B1,...,1 EX∼N (0,1) [Het (X + kµk2 )/ t!] .
Since kBkF = 1, we have
√
EX∼Ge [hB (X)] ≤ |EX∼N (0,1) [Het (X + kµk2 )/ t!]|.
i
By standard results, we have that dHe
dx (x) = iHei−1 (x), and so by Taylor’s theorem we have Hei (x +
Pi
j
kµk2 ) = j=0 ji kµk2 Hei−j (x). Thus,
|EX∼N (0,1) [Het (X + kµk2 )]| =
=
This gives
t
X
t
i=0
kµkt2
i
kµki2 |EX∼N (0,1) [Het−i (X)]
.
p
√
√
EX∼Ge [hA (X)] ≤ kµkt2 / t! = O(ǫ log(1/ǫ))t / t! ,
as required.
Similarly, we have, for all a, b ∈ Zn>0 ,
EX∼N (kµk2 e1 ,I) [Hea (X)Heb (X)] = EX∼N (0,I) [Hea−1 (X)Heb−1 (X)]EX∼N (kµk2 ,1) [Hea1 (X)Heb1 (X)]
= δa−1 b−1 · EX∼N (0,1) [Hea1 (X + kµk2 )Heb1 (X + kµk2 )] .
When kak1 = kbk1 = t, if a−1 = b−1 , then a = b (since a1 = b1 = t − kak1 ). For 1 ≤ j ≤ t, we have:
!2
j
X
j
EX∼N (0,1) [Hej (X + kµk2 )2 ] = EX∼N (0,1)
kµki2 Hej−i (x)
i
i=0
=
=
=
≤
j 2
X
j
i=0
j
X
i=0
j
X
i=0
j
X
j
i
2
2
kµk2i
2 |EX∼N (0,1) [Hej−i (X) ]
kµk2i
2 i!
2
kµk2i
2 (j!/(j − i)!) /i!
2
kµk2i
2 j i/i!
i=0
j
X
≤2
i
2−2i
i=0
(since kµk2 ≤ 1/2k ≤ 1/2j)
≤3.
Putting these together, for a, b with kak1 = kbk1 = t, we have
EX∼N (kµk2 e1 ,I) [Hea (X)Heb (X)] ≤ 3δab .
The sum of squares of coefficients of all Hea (x) in hB (X) is EX∼N (0,1) [hB (X)2 ] = kBkF = 1, and so we
have that EX∼N (kµk2 e1 ,I) [hB (X)2 ]| ≤ 3. Finally, recall that EX∼G [h(A)2 ] = EX∼N (kµk2 e1 ,I) [hB (X)2 ]|,
and so this is O(1), as required.
53
Now consider the equation
EX∼G′ [hA (X)] = wGe EX∼Ge [hA (X)] + wE EX∼E [hA (X)] − wL EX∼L [hA (X)].
We know that the LHS is Ω(ǫ log(1/ǫ)t/2 ) and that the first term on the RHS is smaller. Therefore, one
e we will use standard concentration inequalities to show
of the last two terms is small. Since wL L ≤ G,
t/2
that wL EX∼L [hA (X)] is O(ǫ log(1/ǫ) ). If we cannot find a filter, then wE E ≤ G′ must satisfy similar
concentration inequalities, which would imply that wE EX∼E [hA (X)] is O(ǫ log(1/ǫ)t/2 ). Since some term
on the RHS must be bigger than this, we can find a filter.
Lemma 8.16. For X ∼ N (0, I), if p(x) is a degree-d polynomial with E[p(X)2 ] ≤ 1, we have that
Pr [|p(X)| ≥ T + E[p(X)]] ≤ exp(2 − Ω(T )2/d )) .
′
Lemma 8.17. We have that wL |EX∼L [hA (X)]| ≤ ǫ · O(log(1/ǫ))t /2 ).
Proof. We start with the following claim:
Claim 8.18. For any R > 0, d ∈ Z+ , ǫ > 0, and exp(−(a/R)2/d ) = ǫ, we have
Z ∞
exp(−(T /R)2/d )T dT ≤ (d2 /2)ǫ(d + ln(1/ǫ))d−1 .
a
Proof. Note that (a/R)2/d = ln(1/ǫ). First, we change variables to x = (T /R)2/d to obtain
Z ∞
Z ∞
dT
2/d
exp(−x)xd/2
exp(−(T /R) )T dT =
dx
dx
ln(1/ǫ)
ln(1/ǫ)
Z ∞
exp(−x)xd/2 · (Rd/2)xd/2−1 dx
=
ln(1/ǫ)
Z ∞
exp(−x)xd−1 dx .
= (Rd/2)
ln(1/ǫ)
We can now integrate by parts
Z
Z ∞
d−1
d−1
exp(−x)x dx = ǫ ln(1/ǫ)
+ (d − 1)
ln(1/ǫ)
∞
exp(−x)xd−2 dx .
ln(1/ǫ)
By a simple induction, we have
Z
∞
exp(−x)xd−1 dx = ǫ
ln(1/ǫ)
d−1
X
j=0
d!/(d − j)! ln(1/ǫ)(d−j−1) .
Now we have
Z
∞
2/d
exp(−(T /R)
)T dt = (d/2)ǫ
a
d−1
X
j=0
d!/(d − j)! ln(1/ǫ)(d−j−1)
2/d
≤ (d/2) exp(−(a/R)
2
)
d−1
X
j=0
d−1
≤ (d /2)ǫ(d + ln(1/ǫ))
54
.
dj ln(1/ǫ)(d−j−1)
Let µh = EX∼Ge [hA (X)]. We have the following sequence of inequalities:
Z ∞
T wL Pr [|hA (X)| > T ]dT
wL EX∼L| [hA (X)2 ] =
X∼L
0
Z ∞
T min{wL , Pr [|hA (X)| > T ]}dT
≤
e
X∼G
0
Z ∞
′
T min{wL , exp(2 − Ω(T − µh )2/t ))}dT
≤
Z0 ∞
′
T min{wL , exp(2 − ((T − µh )/R)2/t ))}dT
(for some R > 0)
≤
0
Z ∞
′
(T + µh ) min{wL , exp(2 − (T /R)2/t ))}dT
=
−µ
Z ∞
Z ch
′
exp(2 − (T /R)2/t )(T + µh )dT
(T + µh )wL dT +
=
c
−µh
′
≤ wL (c + µh )2 /2 + e2 (c + µh )/c ·
t′
′2
′
Z
(where c = (R ln 1/wL )t /2 )
∞
′
exp(−(T /R)2/t )T dT
c
′
≤ wL O(ln(1/wL )) + O(t wL (t + ln(1/wL ))t −1 )
′
≤ ǫ · O(ln(1/ǫ))t ,
p
where the last line follows from t′ ≤ k = O( log(1/ǫ)) ≤ O(log(1/ǫ)). Then, by an application of the
Cauchy-Schwarz inequality, we have that
q
t′ /2
2E
2
).
wL |EX∼L [hA (X)]| ≤ wL EX∼L [|hA (X)|] ≤ wL
X∼L [hA (X)] ≤ ǫ · O(log(1/ǫ))
This completes the proof of the lemma.
′
Lemma 8.19. If PrX∼G′ |F ′ [|hA (X)| ≥ T + 1] ≤ O(exp(2 − Ω(T )2/d )) + ǫ/(2n)2t ), for all integers T ,
′
then wE |EX∼E [hA (X)]| ≤ O(ǫ ln(1/ǫ)t /2 ).
′ includes the filter M , we have that the support of G′ |F ′ and the support of E includes only
Proof. Since F p
x with kxk2 ≤ 2n ln(1/ǫ):
p
p
Claim 8.20. When kxk2 ≤ 2n ln(1/ǫ), then |hA (x)| ≤ (2n ln(1/ǫ))t .
p
Proof. Note that t′ ≤ k ≤ 2n log(1/ǫ). Using the explicit formula for the coefficient Hei (x), we can
55
show that for |x| ≤
p
2n ln(1/ǫ), with k ≥ i, the Hei (x) is dominated by its leading coefficient:
|Hei (x)| =
≤
≤
≤
⌊i/2⌋
X
j=0
⌊i/2⌋
i!(−1)j xi−2j /j!(i − 2j)!2j
X
j!(3/2)j |x|i−2j
X
p
j!(3/2)j ( 2n ln(1/ǫ))i−2j
j=0
⌊i/2⌋
j=0
⌊i/2⌋
X
(2n log(1/ǫ))(i−j)/2
j=0
≤ 2 · (n log(1/ǫ))i/2 .
Therefore, for any a ∈ Zn≥0 with
P
i
ai = t, and kxk2 ≤
|ha (x)| =
≤
′
n
Y
i=1
n
Y
i=1
p
2n log(1/ǫ), we have that
p
|Heai (x)/ ai !|
p
ai
2 n log(1/ǫ)
p
≤ (2 n log(1/ǫ))t .
t
the L1 -norm of the entries is at most nt/2 . Thus, we have that
Since kAkF = 1 and
p A has n tentries, p
t/2
|hA (x)| ≤ n · (2 n log(1/ǫ)) = (2n log(1/ǫ))t .
We note that since
′
Pr
[|hA (X)| ≥ T + 1] ≤ O(exp(2 − Ω(T )2/d )) + ǫ/(2n)2t ) ,
Pr
[|hA (X)| ≥ T + 2] ≤ O(exp(2 − Ω(T )2/d )) + ǫ/(2n)2t ) ,
X∼G′ |F ′
for integers T , that
X∼G′ |F ′
′
for all T .
56
Similarly to the proof of Lemma 8.17, we obtain:
Z ∞
T wE Pr [|hA (X)| > T ]dT
wE EX∼E [hA (X)2 ] =
X∼E
0
√
Z
t
(2n
log(1/ǫ))
=
≤
≤
≤
≤
T wE Pr [|hA (X)| > T ]dT
Z
Z
Z
Z
0
(2n
√
X∼E
log(1/ǫ))t
T min{wE , Pr ′ [|hA (X)| > T ]}dT
0
(2n
√
(2n
√
0
−2
(2n
√
X∼G
log(1/ǫ))t
′
T min{wE , O(exp(2 − Ω(T − 2)2/t ) + ǫ/Cn2t )}dT
log(1/ǫ))t −1
′
(T + 2) min{wE , O(exp(2 − (T /R)2/t ) + ǫ/Cn2t )}dT
log(1/ǫ))t
2t′
O(T ǫ/(2n) )dT +
0
Z
c
(T + 2)wE dT
−2
Z ∞
O(exp(−(T /R)2/t ))T dT
(where c = (R ln 1/wE )t/2 )
+ (c + 1)/c ·
c
p
′
′
= (2n log(1/ǫ))2t · ǫ/(2n)2t + O(wE c2 /2) + O(t′2 wE (t′ + ln(1/wE ))t −1 )
≤ ǫ · (4 log(1/ǫ)t + wE · O(ln 1/wE )t
≤ ǫ · O(log(1/ǫ))t .
Then, by the Cauchy-Schwarz inequality, we conclude that
q
t/2
2E
2
.
wE |EX∼E [hA (X)]| ≤ wE EX∼E [|hA (X)|] ≤ wE
X∼E [hA (X)] ≤ ǫ · O(log(1/ǫ))
This completes the proof.
As an immediate consequence, we obtain:
Corollary 8.21. There is an integer 0 ≤ T ≤ O(n
Pr
X∼G′ |F ′
p
log(1/ǫ))t such that
′
[|hA (X)| ≥ T + 1] ≥ 3 exp(2 − Ω(T )2/d )) + 2ǫ/Cn2t .
We can now prove the following crucial lemma:
Lemma 8.22. The algorithm finds a T with
Pr
X∼G′ |F ′
′
[|hA (X)| ≥ T + 1] ≤ 3 exp(2 − Ω(T )2/d )) + ǫ/Cn2t .
Proof.
8.21, such a T exists, and therefore our algorithm will find one after enumerating
p By Corollary
t
O(n log(1/ǫ)) possibilities.
Let F be the event that the new filter accepts. In the next iterations, we will use G′ |F ′ ∩ F instead of
G′ |F ′ . We need to show that the parameters wGe , wE and wL improve in such a way that we only need a
bounded number of iterations:
e + w′ E ′ − w′ L′ , where L′ and E ′ have disjoint supports
Claim 8.23. We can write G′ |F ′ ∩ F = w′e G
E
L
G
′
′ , w ′ > 0 and w ′ + w ′ ≤ w + w − ǫ/Cn2t . The probability that the filter rejects is at most
wE
E
L
L
E
L
′ − w ′ ).
O(wE + wL − wE
L
57
Proof. This proof is very similar to that of Claim 26 from [DKS16c]. Let ¬F be the event that the filter
′
rejects, i.e., that |hA (X)| ≥ T + 1. We have that PrG′ |F ′ [¬F ] ≥ 3 exp(2 − Ω(T )2/d )) + ǫ/Cn2t . On the
other hand, by the concentration inequality, PrGe [¬F ] ≤ exp(2 − Ω(T )2/d )). Thus, we have that
′
Pr [¬F ] ≥ 3 Pr[¬F ] + ǫ/Cn2t .
G′ |F ′
e
G
e E and L yields for the event ¬F that
However, the defining relation between G′ |F ′ and G,
Pr [¬F ] ≥ wGe Pr[¬F ] + wE Pr[¬F ] − wE Pr[¬F ] .
G′ |F ′
E
e
G
E
Since
Pr [¬F ] ≤ wGe Pr′ [¬F ] + wE Pr[¬F ] ,
G′ |F ′
and wGe ≤ 1 + O(ǫ), we must have
E
G
Pr [¬F ] ≤ (2 + O(ǫ))wE Pr[¬F ] ,
G′ |F ′
E
and
Pr′ [¬F ] ≤ (1/3 + O(ǫ))wE Pr[¬F ].
E
G
Then, we get that
′
′
1 − Pr [¬F ] (G |F )(x) = 1 − Pr [¬F ] (G′ |F ′ ∩ ¬F )(x)
G′ |F ′
G′ |F ′
e
e
= wGe G(x) + wE 1 − Pr[¬F ] E(x) + wL 1 − Pr[¬F ] L(x) − wGe Pr′ [¬F ]G(x)
.
E
L
G
Thus, we have
wL (1 − PrL [¬F ]) − wGe PrG′ [¬F ]
1 − PrPe [¬F ]
≤ wL + (1 + O(ǫ)) Pr′ [¬F ] + O ǫ Pr[¬F ]
G
Pe
≤ wL + (1/3 + O(ǫ)) wE Pr[¬F ] + O ǫ Pr [¬F ] .
′
wL
=
E
Also we have
G′ |F ′
wE (1 − PrE [¬F ])
1 − PrG′ |F ′ [¬F ]
≤ wE 1 − Pr[¬F ] + O ǫ Pr [¬F ] .
′
wE
=
E
Thus,
wL + wE −
′
wL
−
′
wE
G′ |F ′
≥ (2/3 − O(ǫ)) wE Pr[¬F ] − O ǫ Pr[¬F ]
E
≥ (1/3 − O(ǫ)) Pr [¬F ] ≥ ǫ/Cn
Pe
2t′
G′ |F ′
Note that the penultimate inequality also gives that
′
′
Pr [¬F ] ≤ (3 + O(ǫ)) (wL + wE − wL
− wE
).
G′ |F ′
This completes the proof.
Proposition 8.12 now follows using induction on the iterations.
58
.
8.3.2
Completing the Proof of Correctness
Lemma 8.24. dim(V ) ≤ O(log(1/ǫ))k .
Proof. After leaving the filter loop, for all 1 ≤ t ≤ k, we have that kPet kF ≤ ǫO(log(1/ǫ))t/2 . M (Pet ) has
the same Frobenius norm, and thus the L2 -norm of its singular values, when considered as a matrix. Thus,
t/2 singular values bigger than ǫ. So, we have that dim(V ) = O(log(1/ǫ))t/2 ,
there are at most O(log(1/ǫ))
t
Pk
and so dim(V ) ≤ t=1 dim Vt ≤ O(log(1/ǫ))k .
Let µV be the projection of µ onto the subspace V . Now we can show using our moment matching
lemma that it suffices to approximate µV .
Lemma 8.25. We have that kµV − µk2 ≤ O(ǫ).
Proof. Let v = (µV − µ)/kµV − µk2 . Note that v is a unit vector perpendicular to V and we need to show
that v T µ ≤ O(ǫ). We will apply Lemma 8.1 to G′′ , the projection G′ conditioned on the event that all the
e G′ |F ′ ) ≤ O(ǫ), and so we have that
filters we produced accept F ′ , onto v. Note that G′ |F ′ has dTV (G,
′′
T
dTV (G , N (0, v µ)) ≤ O(ǫ). We can bound the expectation of the Hermite polynomials as follows, for
1 ≤ t ≤ k:
√
√
|EX∼G′′ [Het (X)/ t!]| = |EX∼G′ [Het (v · X)/ t!]| = |P t (v, . . . , v)|
= |(v ⊗t−1 )T M (P t )v| ≤ kv ⊗t−1 k2 kM (P t )vk2
≤ 1 · O(ǫ) .
√
√
On the other hand, we have EX∼N (0,1) [Het (X)/ t!] = 0, for 1 ≤ t ≤ k. For t = 0, Het (X)/ t! = 1,
which has expectation 1 under both G′′ and N (0, 1). We want to consider the difference in the expectations
of X t , for√1 ≤ t ≤ k. We can write xt as a linear combination of Hermite polynomials, xt =
Pt
we have that EX∼N (0,1) [(X t )2 ] =
Pi=02 ai Hei (x)/ i. Using the orthonormality of these polynomials,
t 2
t
i ai . On the other hand, by standard results, EX∼N (0,1) [(X ) ] = 2 t!. Thus, we have:
|EX∼G′′ [X t ] − EX∼N (0,1) [X t ]| =
t
X
i=0
√
√
ai EX∼G′′ [Hei (X)/ i!] − EX∼N (0,1) [Hei (X)/ i!]
≤ O(ǫ) ·
≤ O(ǫ) ·
t
X
ai
i=0
√
t·
√
2t t!
√
√ √
Note that for t ≥ 10, t2t t! ≤ (t − 1)!/t. Thus, there is a constant c > 0 such that this O(ǫ) · t · 2t t!
is smaller than (t − 1)!ct ǫ/t, for all 1 ≤ t ≤ k. Now we can apply Lemma 8.1 with
p δ = cǫ and obtain
T
that |v µ| ≤ O(δ) = O(ǫ). We need to set k to be a sufficiently high multiple of ǫ ln(1/ǫ) to make this
work.
It remains to analyze the rest of the algorithm and show that µ
eV it produces is close to µV .
Lemma 8.26. We can construct a set S ⊂ V of unit vectors of size dim(V )O(dim(V )) such that for any unit
vector v ∈ V , there is a v ′ ∈ S with kv − v ′ k2 ≤ 1/2, in time dim(V )O(dim(V )) .
59
Proof. Let ℓ = dim(V ). We will construct such a cover for Rℓ and translate that to V by using the orthonormal basis for V given by the right singular vectors of M (Pt√
) with singular values bigger than ǫ. We
can divide the cube [−1, 1]ℓ into ℓO(ℓ) cubes of side length 1/(2 ℓ). For each cube, we check if it has
a corner with L2 -norm ≥ 1 and a corner with L2 -norm ≤ 1. If it does not, it does not contain any unit
vectors so we can ignore it. If the cube does contain any unit vectors, then if its center is v, we add the
normalized vector
v/kvk2 to S. Since there is a unit vector v ′ in the cube, and all vectors in the cube have
√
kv ′ − vk2 ≤ ℓkv ′ − vk∞ ≤ 1/4, we have |kvk2 − 1| ≤ 1/4, and so kv − v/kvk2 k2 ≤ 1/4. Thus, for any
unit vector v ′ in this cube, we have kv ′ − v/kvk2 k2 ≤ 1/4 + 1/4 ≤ 1/2. Since every unit v ′ is in some cube
whose normalized center we added to S, we are done.
Firstly, we note that in order to approximate v T µ, it is sufficient to find an x with PrX∼G′ [v T X ≥ x] =
1/2 + O(ǫ):
Lemma 8.27. For all x ∈ R with | PrX∼G′ [v T X ≥ x] − 1/2| ≤ 3ǫ, we have that |v T µ − x| ≤ O(ǫ) .
e projected onto v has G(x − v T µ) ≥ 1/2 for all x with |x − v T µ| ≤ O(ǫ).
Proof. First note that the pdf of G
T
Supposing that |x − v µ| ≥ 8ǫ, we have that | PrX∼Ge [v T X ≥ x] − 1/2| ≥ 4ǫ, and so | PrX∼G′ [v T X ≥
e ≥ 3ǫ.
x] − 1/2| ≥ 4ǫ − dTV (G′ , G)
To show that we can find such a point by bisection, we need to show that there is an interval of such
points of reasonable length where we are looking for them:
Lemma 8.28. Given a unit vector v ∈ Rn , there is an interval [a, b] such that
• for all x ∈ [a, b], we have that | PrX∼G′ [v T X ≥ x] − 1/2| ≤ 2ǫ,
• b − a = Θ(ǫ),
• and |a|, |b| ≤ O(ǫ
p
log 1/ǫ).
Proof. We can take [a, b] to be the set of x with | PrX∼Ge [v T X ≥ x] − 1/2| ≤ ǫ. This is an interval since
e ≤ 2ǫ.
PrX∼Ge [v T X ≥ x] is monotone. All x in it have | PrX∼G′ [v T X ≥ x] − 1/2| ≤ ǫ + dTV (G′ , G)
T
T
T
Thus,
pby the previous lemma, |b − v µ|, |a
p− v µ| ≤ O(ǫ) and thus |b − a| ≤ O(ǫ), and since |v µ| ≤
O(ǫ log 1/ǫ), we have that |a|, |b| ≤ O(ǫ log 1/ǫ).
Thus, we obtain:
Lemma 8.29. Given a unit vector v ∈ Rn , we can find an mv with |v T µ−mv | ≤ O(ǫ) using O(log log 1/ǫ)
statistical queries of precision ǫ/2.
Proof. We use bisection to find a point where our SQ approximation pe to PrX∼G′ [v T X ≥ x] is within 5ǫ/2
of 1/2. If we find such a point, it has |v T µ − mv | ≤ O(ǫ), by Lemma 8.27. Lemma
p8.28 yields that there
is an interval [a, b] of length O(ǫ) containing such points in the interval |x| ≤ O(ǫ log(1/ǫ)). Indeed, if
our test point x has pe > 1/2 + 5ǫ/2, then x > b and if pe < 1/2 − 5ǫ/2, then x < a. Thus, [a, b] remains a
subinterval of the interval we are considering.
We now have that µv is a feasible point of the LP considered in Step 11. The following lemma completes
the proof:
Lemma 8.30. Any feasible point of the LP considered in Step 11, µ
eV has kµV − µ
eV k2 ≤ O(ǫ).
60
Proof. Consider the vector v = (µV − µ
eV )/kµV − µ
eV k2 . Note that v is in V , since µC , µ
eV are. Since
′
′
v is a unit vector in V , there is a v ∈ S with kv − v k2 ≤ 1/2. Since µ
eV is a solution to the LP,
v ′T (µV − µ
eV ) ≤ O(ǫ). Thus, we have that
kµV − µ
eV k2 = v T (µV − µ
eV )
= v ′T (µV − µ
eV ) + (v − v ′ )T (µV − µ
eV )
≤ O(ǫ) + kµV − µ
eV k2 /2 .
Therefore, kµV − µ
eV k2 ≤ O(ǫ), as required.
Proof of Theorem 8.7. Since the LP has a feasible point, we can find such a point µ
eV that has kµV − µ
eV k2 ≤
O(ǫ). By the previous lemma, we have that kµV −µk2 ≤ O(ǫ). Thus, the algorithm is correct. All statistical
queries are of the claimed precision. We need to get bounds on the running time and number of statistical
queries.
Step 2 that uses the algorithm from [DKK+ 16], takes poly(n/ǫ) time and statistical queries. Finding Pet ,
for each 1 ≤ t ≤ k, takes nt statistical queries, giving nO (k) time total. There are at most O(n2k ) iterations
of the loop. Each iteration takes poly(n/ǫ) time and statistical queries to find T , and nt statistical queries
to recompute Pet . It suffices to compute the SVD to within Frobenius norm 1/poly(n/ǫ), which takes time
O(k)
k
. Computing it takes
poly(n/ǫ). The set S has size dim(V )O(dim(V )) = log(1/ǫ)kO(log(1/ǫ)) = 2log(1/ǫ)
O(k)
O(k)
log(1/ǫ)
log(1/ǫ)
statistical queries and time. The LP has
. Approximating the medians takes 2
time 2
O(k)
O(k)
O(k)
log(1/ǫ)
bits, and so with a
constraints and log(1/ǫ)
variables. The size of the LP is 2log(1/ǫ)
2
O(k)
time.
polynomial time LP solver, we can get 2log(1/ǫ)
We thus have that the total time and statistical queries are both at most
√
√
log(1/ǫ)O( log(1/ǫ))
O(k)
log(1/ǫ)O(k)
O( log(1/ǫ))
.
+2
n
poly(1/ǫ) + 2
=n
References
[ABG+ 14]
J. Anderson, M. Belkin, N. Goyal, L. Rademacher, and J. R. Voss. The more, the merrier: the
blessing of dimensionality for learning large gaussian mixtures. In Proceedings of The 27th
Conference on Learning Theory, COLT 2014, pages 1135–1164, 2014.
[ADLS17]
J. Acharya, I. Diakonikolas, J. Li, and L. Schmidt. Sample-optimal density estimation in
nearly-linear time. In Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA 2017, pages 1278–1289, 2017.
[AK01]
S. Arora and R. Kannan. Learning mixtures of arbitrary Gaussians. In Proceedings of the 33rd
Symposium on Theory of Computing, pages 247–257, 2001.
[AM05]
D. Achlioptas and F. McSherry. On spectral learning of mixtures of distributions. In Proceedings of the Eighteenth Annual Conference on Learning Theory (COLT), pages 458–469,
2005.
[AS72]
M. Abramowitz and I. Stegun. Handbook of Mathematical Functions. Dover, 1972.
[BBBB72]
R.E. Barlow, D.J. Bartholomew, J.M. Bremner, and H.D. Brunk. Statistical Inference under
Order Restrictions. Wiley, New York, 1972.
61
[BCMV14] A. Bhaskara, M. Charikar, A. Moitra, and A. Vijayaraghavan. Smoothed analysis of tensor
decompositions. In Symposium on Theory of Computing, STOC 2014, pages 594–603, 2014.
[BFJ+ 94]
A. Blum, M. Furst, J. Jackson, M. Kearns, Y. Mansour, and S. Rudich. Weakly learning DNF
and characterizing statistical query learning using Fourier analysis. In Proceedings of the
Twenty-Sixth Annual Symposium on Theory of Computing, pages 253–262, 1994.
[BFR+ 00]
T. Batu, L. Fortnow, R. Rubinfeld, W. D. Smith, and P. White. Testing that distributions are
close. In IEEE Symposium on Foundations of Computer Science, pages 259–269, 2000.
[BR13a]
Q. Berthet and P. Rigollet. Complexity theoretic lower bounds for sparse principal component
detection. In COLT 2013 - The 26th Annual Conference on Learning Theory, pages 1046–
1066, 2013.
[BR13b]
Q. Berthet and P. Rigollet. Optimal detection of sparse principal components in high dimension. Ann. Statist., 41(4):1780–1815, 2013.
[BS10]
M. Belkin and K. Sinha. Polynomial learning of distribution families. In FOCS, pages 103–
112, 2010.
[BV08]
S. C. Brubaker and S. Vempala. Isotropic PCA and Affine-Invariant Clustering. In Proc. 49th
IEEE Symposium on Foundations of Computer Science, pages 551–560, 2008.
[CDSS13]
S. Chan, I. Diakonikolas, R. Servedio, and X. Sun. Learning mixtures of structured distributions over discrete domains. In SODA, pages 1380–1394, 2013.
[CDSS14a] S. Chan, I. Diakonikolas, R. Servedio, and X. Sun. Efficient density estimation via piecewise
polynomial approximation. In STOC, pages 604–613, 2014.
[CDSS14b] S. Chan, I. Diakonikolas, R. Servedio, and X. Sun. Near-optimal density estimation in nearlinear time using variable-width histograms. In NIPS, pages 1844–1852, 2014.
[CFJ13]
T. Cai, J. Fan, and T. Jiang. Distributions of angles in random packing on spheres. Journal of
Machine Learning Research, 14(1):1837–1864, 2013.
[CGG02]
M. Cryan, L. Goldberg, and P. Goldberg. Evolutionary trees can be learned in polynomial time
in the two state general Markov model. SIAM Journal on Computing, 31(2):375–397, 2002.
[CGR15]
M. Chen, C. Gao, and Z. Ren. Robust covariance matrix estimation via matrix depth. CoRR,
abs/1506.00691, 2015.
[CKL+ 06]
C.-T. Chu, S. K. Kim, Y. A. Lin, Y. Yu, G. Bradski, A. Y. Ng, and K. Olukotun. Map-reduce
for machine learning on multicore. In Proceedings of the 19th International Conference on
Neural Information Processing Systems, NIPS’06, pages 281–288, Cambridge, MA, USA,
2006. MIT Press.
[CMW13]
T. T. Cai, Z. Ma, and Y. Wu. Sparse pca: Optimal rates and adaptive estimation. Ann. Statist.,
41(6):3074–3110, 12 2013.
[CMW15]
T. Cai, Z. Ma, and Y. Wu. Optimal estimation and rank detection for sparse spiked covariance
matrices. Probability Theory and Related Fields, 161(3):781–815, 2015.
[CQ10]
S. X. Chen and Y. L. Qin. A two-sample test for high-dimensional data with applications to
gene-set testing. Ann. Statist., 38(2):808–835, 04 2010.
62
[Dan16]
A. Daniely. Complexity theoretic limitations on learning halfspaces. In Proceedings of the
48th Annual Symposium on Theory of Computing, STOC 2016, pages 105–117, 2016.
[Das99]
S. Dasgupta. Learning mixtures of Gaussians. In Proceedings of the 40th Annual Symposium
on Foundations of Computer Science, pages 634–644, 1999.
[DBS17]
S. Du, S. Balakrishnan, and A. Singh. Computationally efficient robust estimation of sparse
functionals. CoRR, abs/1702.07709, 2017.
[DDKT16]
C. Daskalakis, A. De, G. Kamath, and C. Tzamos. A size-free CLT for poisson multinomials
and its applications. In Proceedings of STOC’16, 2016.
[DDO+ 13]
C. Daskalakis, I. Diakonikolas, R. O’Donnell, R.A. Servedio, and L. Tan. Learning Sums of
Independent Integer Random Variables. In FOCS, pages 217–226, 2013.
[DDS12a]
C. Daskalakis, I. Diakonikolas, and R.A. Servedio. Learning k-modal distributions via testing.
In SODA, pages 1371–1385, 2012.
[DDS12b]
C. Daskalakis, I. Diakonikolas, and R.A. Servedio. Learning Poisson Binomial Distributions.
In STOC, pages 709–728, 2012.
[DDS15]
A. De, I. Diakonikolas, and R. Servedio. Learning from satisfying assignments. In Proceedings of the 26th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2015, pages
478–497, 2015.
[DG85]
L. Devroye and L. Györfi. Nonparametric Density Estimation: The L1 View. John Wiley &
Sons, 1985.
[DG92]
D. L. Donoho and M. Gasko. Breakdown properties of location estimates based on halfspace
depth and projected outlyingness. Ann. Statist., 20(4):1803–1827, 12 1992.
[Dia16]
I. Diakonikolas. Learning structured distributions. In P. Bühlmann, P. Drineas, M. Kane, and
M.J. van Der Laan, editors, Handbook of Big Data, Chapman & Hall/CRC Handbooks of
Modern Statistical Methods, chapter 15, pages 267–284. Taylor & Francis, 2016.
[DK14]
C. Daskalakis and G. Kamath. Faster and sample near-optimal algorithms for proper learning
mixtures of gaussians. In Proceedings of The 27th Conference on Learning Theory, COLT
2014, pages 1183–1213, 2014.
[DKK+ 16]
I. Diakonikolas, G. Kamath, D. M. Kane, J. Li, A. Moitra, and A. Stewart. Robust estimators in
high dimensions without the computational intractability. In Proceedings of FOCS’16, 2016.
[DKK+ 17a] I. Diakonikolas, G. Kamath, D. M. Kane, J. Li, A. Moitra, and A. Stewart. Being robust (in
high dimensions) can be practical. CoRR, abs/1703.00893, 2017.
[DKK+ 17b] I. Diakonikolas, G. Kamath, D. M. Kane, J. Li, A. Moitra, and A. Stewart. Robustly learning
a gaussian: Getting optimal error, efficiently. CoRR, abs/1704.03866, 2017.
[DKS16a]
I. Diakonikolas, D. M. Kane, and A. Stewart. The fourier transform of poisson multinomial
distributions and its algorithmic applications. In Proceedings of STOC’16, 2016.
[DKS16b]
I. Diakonikolas, D. M. Kane, and A. Stewart. Optimal learning via the fourier transform for
sums of independent integer random variables. In Proceedings of COLT 2016, pages 831–849,
2016. Full version available at https://arxiv.org/abs/1505.00662.
63
[DKS16c]
I. Diakonikolas, D. M. Kane, and A. Stewart. Robust learning of fixed-structure bayesian
networks. CoRR, abs/1606.07384, 2016.
[DL01]
L. Devroye and G. Lugosi. Combinatorial methods in density estimation. Springer Series in
Statistics, Springer, 2001.
[DLS14]
A. Daniely, N. Linial, and S. S.-Shwartz. From average case complexity to improper learning
complexity. In Symposium on Theory of Computing, STOC 2014, pages 441–448, 2014.
[Fel16a]
V. Feldman. A general characterization of the statistical query complexity.
abs/1608.02198, 2016.
[Fel16b]
V. Feldman. Statistical query learning. In Encyclopedia of Algorithms, pages 2090–2095.
2016.
[FGKP06]
V. Feldman, P. Gopalan, S. Khot, and A. Ponnuswami. New results for learning noisy parities
and halfspaces. In Proc. 47th IEEE Symposium on Foundations of Computer Science (FOCS),
pages 563–576, 2006.
[FGR+ 13]
V. Feldman, E. Grigorescu, L. Reyzin, S. Vempala, and Y. Xiao. Statistical algorithms and a
lower bound for detecting planted cliques. In Proceedings of STOC’13, pages 655–664, 2013.
[FGV15]
V. Feldman, C. Guzman, and S. Vempala. Statistical query algorithms for stochastic convex
optimization. CoRR, abs/1512.09170, 2015.
[FM99]
Y. Freund and Y. Mansour. Estimating a mixture of two product distributions. In Proceedings
of the 12th Annual COLT, pages 183–192, 1999.
[FOS06]
J. Feldman, R. O’Donnell, and R. Servedio. PAC learning mixtures of Gaussians with no
separation assumption. In Proc. 19th Annual Conference on Learning Theory (COLT), pages
20–34, 2006.
[FOS08]
J. Feldman, R. O’Donnell, and R. A. Servedio. Learning mixtures of product distributions
over discrete domains. SIAM J. Comput., 37(5):1536–1564, 2008.
[FPV15]
V. Feldman, W. Perkins, and S. Vempala. On the complexity of random satisfiability problems
with planted solutions. In Proceedings of the Forty-Seventh Annual ACM on Symposium on
Theory of Computing, STOC, 2015, pages 77–86, 2015.
[GHK15]
R. Ge, Q. Huang, and S. M. Kakade. Learning mixtures of gaussians in high dimensions. In
Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory of Computing, STOC
2015, pages 761–770, 2015.
[GR00]
O. Goldreich and D. Ron. On testing expansion in bounded-degree graphs. Technical Report
TR00-020, Electronic Colloquium in Computational Complexity, 2000.
[HK13]
D. Hsu and S. M. Kakade. Learning mixtures of spherical gaussians: moment methods and
spectral decompositions. In Innovations in Theoretical Computer Science, ITCS ’13, pages
11–20, 2013.
[Hot31]
H. Hotelling. The generalization of student’s ratio. Ann. Math. Statist., 2(3):360–378, 08
1931.
64
CoRR,
[HP15]
M. Hardt and E. Price. Tight bounds for learning a mixture of two gaussians. In Proceedings
of the Forty-Seventh Annual ACM on Symposium on Theory of Computing, STOC 2015, pages
753–760, 2015.
[HR09]
P.J. Huber and E. M. Ronchetti. Robust statistics. Wiley New York, 2009.
[HRRS86]
F. R. Hampel, E. M. Ronchetti, P. J. Rousseeuw, and W. A. Stahel. Robust statistics. The
approach based on influence functions. Wiley New York, 1986.
[Hub64]
P. J. Huber. Robust estimation of a location parameter. Ann. Math. Statist., 35(1):73–101, 03
1964.
[JL09]
I. M. Johnstone and A. Y. Lu. On consistency and sparsity for principal components analysis
in high dimensions. Journal of the American Statistical Association, 104(486):682–693, 2009.
[Kea98]
M. Kearns. Efficient noise-tolerant learning from statistical queries. Journal of the ACM,
45(6):983–1006, 1998.
[KK14]
A. R. Klivans and P. Kothari. Embedding hard learning problems into gaussian space. In
Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques,
APPROX/RANDOM 2014, pages 793–809, 2014.
[KKMS08] A. Kalai, A. Klivans, Y. Mansour, and R. Servedio. Agnostically learning halfspaces. SIAM
Journal on Computing, 37(6):1777–1805, 2008.
[KMR+ 94] M. Kearns, Y. Mansour, D. Ron, R. Rubinfeld, R. Schapire, and L. Sellie. On the learnability
of discrete distributions. In Proc. 26th STOC, pages 273–282, 1994.
[KMV10]
A. T. Kalai, A. Moitra, and G. Valiant. Efficiently learning mixtures of two Gaussians. In
STOC, pages 553–562, 2010.
[KS06]
A. Klivans and A. Sherstov. Cryptographic hardness for learning intersections of halfspaces.
In Proc. 47th IEEE Symposium on Foundations of Computer Science (FOCS), pages 553–562,
2006.
[KSV08]
R. Kannan, H. Salmasian, and S. Vempala. The spectral method for general mixture models.
SIAM J. Comput., 38(3):1141–1156, 2008.
[KV16]
R. Kannan and S. Vempala. Beyond spectral: Tight bounds for planted gaussians. CoRR,
abs/1608.03643, 2016.
[Li17]
J. Li. Robust sparse estimation tasks in high dimensions. CoRR, abs/1702.05860, 2017.
[LR05]
E. L. Lehmann and J. P. Romano. Testing statistical hypotheses. Springer Texts in Statistics.
Springer, 2005.
[LRV16]
K. A. Lai, A. B. Rao, and S. Vempala. Agnostic estimation of mean and covariance. In
Proceedings of FOCS’16, 2016.
[LS15]
J. Li and L. Schmidt. A nearly optimal and agnostic algorithm for properly learning a mixture
of k gaussians, for any constant k. CoRR, abs/1506.01367, 2015. To appear in COLT’17.
[Moi14]
A. Moitra. Algorithmic aspects of machine learning.
http://people.csail.mit.edu/moitra/docs/bookex.pdf.
65
2014.
Available online at
[MR05]
E. Mossel and S. Roch. Learning nonsingular phylogenies and Hidden Markov Models. In To
appear in Proceedings of the 37th Annual Symposium on Theory of Computing (STOC), 2005.
[MV10]
A. Moitra and G. Valiant. Settling the polynomial learnability of mixtures of Gaussians. In
FOCS, pages 93–102, 2010.
[MW15]
T. Ma and A. Wigderson. Sum-of-squares lower bounds for sparse pca. In Proceedings of the
28th International Conference on Neural Information Processing Systems, NIPS’15, pages
1612–1620, 2015.
[NP33]
J. Neyman and E. S. Pearson. On the problem of the most efficient tests of statistical hypotheses. Philosophical Transactions of the Royal Society of London. Series A, Containing Papers
of a Mathematical or Physical Character, 231(694-706):289–337, 1933.
[O’D14]
R. O’Donnell. Analysis of Boolean Functions. Cambridge University Press, New York, NY,
USA, 2014.
[Sco92]
D.W. Scott. Multivariate Density Estimation: Theory, Practice and Visualization. Wiley, New
York, 1992.
[SD08]
M. S. Srivastava and M. Du. A test for the mean vector with fewer observations than the
dimension. Journal of Multivariate Analysis, 99(3):386 – 402, 2008.
[Sil86]
B. W. Silverman. Density Estimation. Chapman and Hall, London, 1986.
[SOAJ14]
A. T. Suresh, A. Orlitsky, J. Acharya, and A. Jafarpour. Near-optimal-sample estimators for
spherical gaussian mixtures. In Advances in Neural Information Processing Systems (NIPS),
pages 1395–1403, 2014.
[Sze89]
G. Szegö. Orthogonal Polynomials, volume XXIII of American Mathematical Society Colloquium Publications. A.M.S, Providence, 1989.
[Tuk75]
J.W. Tukey. Mathematics and picturing of data. In Proceedings of ICM, volume 6, pages
523–531, 1975.
[Val84]
L. Valiant. A theory of the learnable. Communications of the ACM, 27(11):1134–1142, 1984.
[Ver12]
R. Vershynin. How close is the sample covariance matrix totheactual covariance matrix? Journal of Theoretical Probability, 25(3):655–686, 2012.
[VW02]
S. Vempala and G. Wang. A spectral algorithm for learning mixtures of distributions. In
Proceedings of the 43rd Annual Symposium on Foundations of Computer Science, pages 113–
122, 2002.
[WBS16a]
T. Wang, Q. Berthet, and R. J. Samworth. Statistical and computational trade-offs in estimation
of sparse principal components. Ann. Statist., 44(5):1896–1930, 10 2016.
[WBS16b]
T. Wang, Q. Berthet, and R.J. Samworth. Statistical and computational trade-offs in estimation
of sparse principal components. Ann. Statist., 44(5):1896–1930, 2016.
[Wil97]
R. R. Wilcox. Introduction to robust estimation and hypothesis testing. Statistical modeling
and decision science. Acad. Press, San Diego, Calif. [u.a.], 1997.
[ZB96]
H. Saranadasa Z. Bai. Effect of high dimension: by an example of a two sample problem.
Statist. Sinica,, 6:311–329, 1996.
66
Appendix
A
Sample Complexity Upper Bound for Learning GMMs
In this section, we show that learning a k-mixture of n-dimensional Gaussians to variation distance error ǫ
is easy information theoretically. In particular, we have:
Theorem A.1. Given ǫ > 0 and positive integers k and n, there exists an algorithm that, given a probability
distribution P which is a k-mixture of n-dimensional Gaussians, takes O(n2 k3 log2 (k)/ǫ5 ) samples from P
and with probability at least 2/3 returns a distribution Q with dTV (P, Q) < ǫ.
Note that the algorithm given in Theorem A.1 will not be computationally efficient.
The basic idea of Theorem A.1 will be to make many guesses as to the mixture, at least one of which
is close, and then run a tournament to find the true answer. We approximate the mixture by first guessing
approximations to the weights and then approximating each individual Gaussian. If we had polynomially
many samples from a single part of the mixture, it would be easy to learn:
Lemma A.2 (Folklore). Let G be an n-dimensional Gaussian, and let δ > 0. There exists a polynomial
time algorithm that, given O(n2 /δ2 ) independent samples from G, returns a probability distribution P so
that, with probability at least 2/3, dTV (G, P) < δ.
Note that we can easily improve the success probability in Lemma A.3 to 1− δ at the cost of multiplying
the sample complexity by log(1/δ). In particular, we have:
Corollary A.3. Let G be an n-dimensional Gaussian, and let ǫ > 0. There exists a polynomial time algorithm that given M independent samples from G returns a probability distribution P so that with probability
at least 1 − exp(−Ω(M ǫ2 /d2 )), we have dTV (G, P) < ǫ.
Unfortunately, we cannot simply run this algorithm for each component in our mixture, since we do not
know which samples come from which component. However, if we manage to correctly guess where each
sample comes from this will not be an issue.
Proposition A.4. Given ǫ > 0 and positive integers k and n, there exists an algorithm that given a probability distribution P, which is a k-mixture of n-dimensional Gaussians, and Θ(n2 k3 log(k)/ǫ3 ) independent
samples from P, returns a set of exp(O(n2 k3 log2 (k)/ǫ3 )) distributions Qi so that with probability at least
2/3 there exists an i so that dTV (Pi , Qi ) < ǫ.
Proof. If our algorithm is given N samples, it will return a qi for each function f : [N ] → [k]. Intuitively, f
encodes our guess as to which sample came from which component of the mixture. Note that there are only
exp(N log(k)) many such f ’s.
The algorithm is quite simple. Let s1 , s2 , . . . , sN be our samples, and let Si = {sj : f (j) = i}. Letting
A be the algorithm from Corollary A.3 with δ taken to be ǫ/(10k), we let
Qf =
k
X
|Si |
i=1
N
A(Si ).
We claim that at least one of these works
Pkwith probability 2/3.
In particular, let P be the mixture i=1 wi Gi . Consider the case where f correctly guesses which part
of the mixture each sample was taken from. In particular, f (i) = j if and only if si was taken from Gj . We
claim that, with probability at least 2/3, this choice of f leads to dTV (Qf , P) < ǫ. We will henceforth use
Si to denote the set of samples actually taken from the ith component of the mixture.
67
Firstly, note that by standard concentration bounds, we have that |SNi | − wi < ǫ/(10k), for all i, with
probability at least 9/10.
Secondly, note that after conditioning on which samples of P were taken form which part, the samples
themselves are independent samples from the appropriate Gi ’s, that with probability at least 9/10 we have
that dTV (Gi , A(Si )) < ǫ/(10k) for all i with |Si | ≫ n2 k2 log(k)/ǫ2 .
We claim that if both of the above conditions hold (which happens with probability at least 2/3) that
dTV (P, Qf ) < ǫ. Letting S be the set of indices i so that wi < ǫ/(5k), we note that for i 6∈ S that
|Si | ≫ n2 k2 log(k)/ǫ2 . We then have that
k
k
X
1 X
|Si |
dTV (P, Qf ) =
wi Gi −
A(Si )
2
N
i=1
i=1
1
k
X
|Si |
1
≤
wi Gi −
A(Si )
2
N
2
i=1
k
k
|Si |
|Si |
1X
1X
max wi ,
wi −
+
|Gi − A(Si )|1
≤
2
N
2
N
i=1
i=1
X
1X
(wi + ǫ/(10k))|Gi − A(Si )|1 +
dTV (Gi , A(Si ))
≤ ǫ/20 +
2
i∈S
X
1X
≤ ǫ/20 +
3ǫ/(10k) +
ǫ/(10k)
2
i∈S
i6∈S
i6∈S
≤ ǫ/20 + 3ǫ/20 + ǫ/20
<ǫ.
This completes the proof.
Theorem A.1 now follows immediately form a standard tournament argument (see, e.g., [DL01, DDS12b,
DDS15]).
B Sample Complexity Upper Bound for Parameter Estimation of Separated GMMs
Next we consider P
the more complicated task of parameter estimation. In particular, given samples from a
distribution P = ki=1 Gi , where each Gi is a weighted Gaussian, we would like to learn a distribution
P
Q that is not only close to P but that can be written as Q = ki=1 Hi with kHi − Gi k1 small for all i.
Now, in general, this task will require number of samples exponential in k, simply because there are pairs
of mixtures that are ǫΩ(k) -close in variation distance and yet ǫ-far in terms of their individual components.
However, we will show that if the components are separated, this cannot be the case and thus learning the
distribution in variation distance will be sufficient.
Before we begin, we need to clarify
R our notion of separation. Given two pseudo-distributions, p and q,
we define their overlap as V (p, q) := min(dp, dq). We should note that if p and q are honest distributions,
then dTV (p, q) = 1 − V (p, q). We have the following theorem:
P
P
Theorem B.1. There exists a constant C so that if we have two mixtures p = ki=1 Gi and q = ki=1 Hi ,
where p and q are normalized distributions with Hi , Gi weighted Gaussians, such that dTV (p, q) < (δ/k)C
for some sufficiently small δ > 0, and so that for any i 6= j, V (Gi , Gj ), V (Hi , Hj ) < (δ/k)C , then there
exists a permutation π : [k] → [k] so that kGi − Hπ(i) k1 < δ for all i.
68
We begin by producing a proxy for the overlap between distributions. In particular, for pseudo-distributions
p and q, we define
Z
p
h(p, q) = − log
dpdq .
Notice that if p and q are true distributions, this is related to the Hellinger distance by H(p, q) = 2(1 −
e−h(p,q) ). We also note the relationship to the overlap:
Lemma B.2. If p and q are pseudo-distributions with L1 norm at most 1, then
V (p, q) = exp(−Θ(h(p, q)) + O(1)) .
Proof. On the one hand, there is an easy upper bound
Z
Z p
dpdq = exp(−h(p, q)).
V (p, q) = min(dp, dq) ≤
The lower bound is by Cauchy-Schwarz
exp(−h(p, q)) =
Z p
This completes our proof.
dpdq ≤
Z
1/2 Z
1/2
p
≤ 2V (p, q).
min(dp, dq)
max(dp, dq)
Ideally we would like to show that h is nearly a metric for Gaussians. Namely that h(A, C) =
O(h(A, B) + h(B, C)). This would imply that Hi could not have large overlap with more than one Gj ,
since if V (Hi , Ga ) and V (Hi , Gb ) were both large, then h(Hi , Ga ), h(Hi , Gb ) would be small and therefore, h(Ga , Gb ) would be small. This would contradict our assumption that Ga and Gb have small overlap.
Unfortunately, this is not true. In one dimension, a very wide Gaussian may have non-trivial overlap with
two narrow Gaussians with widely separated means, neither of which overlaps the other substantially. We
will need to develop techniques to deal with this circumstance.
To do this, we introduce an intermediate notation. If Gi = wi N (µi , Σi ) are weighted Gaussians, we
define
hΣ (G1 , G2 ) := h(N (0, Σ1 ), N (0, Σ2 )).
This is useful because it does satisfy an approximate triangle inequality.
Proposition B.3. For F, G, H weighted Gaussians, we have that
hΣ (F, H) = O(hΣ (F, G) + hΣ (G, H)).
Before we prove this, we will first need to find an approximation to hΣ .
Lemma B.4. If G and H are weighted Gaussians with covariance matrices A and B respectively, then
hΣ (G, H) = Θ Σλ eigenvalue of B −1/2 AB −1/2 min(| log(λ)|, | log(λ)|2 ) .
Proof. By making an appropriate change of variables, we can assume that H has identity covariance and G
has covariance B −1/2 AB −1/2 . Thus, it suffices to consider the case where B = I. In this case, we may
69
diagonalize A to get A = diag(λi ). We then have that
hΣ (G, H) = h(N (0, A), N (0, I))
= − log (2π)−n/2
= − log
=
n
X
n
Y
−1/4
λi
n
Y
−1/4
λi
i=1
Z
exp −
−1/2
((1 + λ−1
i )/2)
i=1
!
n
X
!
x2i /2(1/(2λi ) + 1/2) dx
i=1
!
log(λi )/4 + log((1 + λ−1
i )/2)/2.
i=1
We claim that
log(λ)/4 + log((1 + λ−1 )/2)/2 = Θ(min(| log(λ)|, | log(λ)|2 )).
To see this note that when λ = 1 + ǫ for small values of ǫ, the left hand side above is
(ǫ/4 − ǫ2 /8 + O(ǫ3 )) + (−ǫ/4 + 3ǫ/16 + O(ǫ3 )) = ǫ2 /16 + O(ǫ3 ) = Θ(ǫ2 ).
On the other hand, when λ ≫ 1, this is asymptotic to | log(λ)|/4, and when λ ≪ 1, it is similarly asymptotic
to − log(λ)/4. Finally, since it is easily verified that log(λ)/4+log((1+λ−1 )/2)/2 is never 0 unless λ = 1,
this proves the claim, from which our lemma follows easily.
We will also need the following fact about eigenvalues of a product of matrices:
Lemma B.5. Let A and B be symmetric matrices with eigenvalues ν1 ≥ ν2 ≥ . . . ≥ νn > 0 and µ1 ≥
µ2 ≥ . . . ≥ µn > 0, respectively. Let λ1 ≥ λ2 ≥ . . . ≥ λ2n > 0 be the sorting of the νi and µi together.
Let M be a matrix with M T M = A. Then, the kth largest eigenvalue of M T BM is at most λ2k .
Proof. We need to show that there is an (n − k + 1)-dimensional subspace V so that for v ∈ V we have
that vA1/2 BA1/2 v ≤ λ2k |v|2 . Suppose that λ1 , . . . , λk−1 contains m of the νi and k − m − 1 of the µi . Let
V be the subspace of vectors v so that v is perpendicular to the top m eigenvectors of A and so that A1/2 v
is perpendicular to the top m − k − 1 eigenvalues of B. Then
vM T BM v ≤ λk |M v|2 = λk vAv ≤ λ2k |v|2 .
This completes the proof.
We are now ready to prove Proposition B.3.
Proof. Let F, G, H have covariance matrices A, B, C respectively. Let Σ1 = A−1/2 BA−1/2 , Σ2 =
B −1/2 CB −1/2 and Σ3 = A−1/2 CA−1/2 = (A−1/2 B 1/2 )Σ2 (B 1/2 A−1/2 ). Let the eigenvalues of Σi be
(i)
(i)
(i)
λ1 ≥ λ2 ≥ . . . ≥ λn > 0. Let f (x) = max(0, min(log(x), log 2 (x))). We have by Lemma B.4 that
!
n
X
(1)
(1)
f (λi ) + f (1/λi ) ,
hΣ (F, G) = Θ
i=1
hΣ (G, H) = Θ
n
X
(2)
f (1/λi )
(3)
f (λi )
(3)
f (1/λi )
i=1
hΣ (F, H) = Θ
n
X
!
(2)
f (λi ) +
i=1
70
+
!
,
.
(3)
(1)
On the other hand, Lemma B.5 says that λi is at most the square of the ith largest of the λj
Therefore,
!
n
n
n
X
X
X
(2)
(1)
(3)
f (λi ) .
f (λi ) +
f (λi ) = O
i=1
i=1
i=1
(2)
and λj .
Similarly, by considering the inverses of these matrices, we find that
n
X
(3)
f (1/λi ) = O
n
X
(1)
f (1/λi ) +
(2)
!
f (1/λi ) .
i=1
i=1
i=1
n
X
Together these complete the proof.
In addition to this, we need to know what else contributes to h(G, H). We define
hµ (G, H) = h(G, H) − hΣ (G, H).
We make the following claim:
Proposition B.6.
−1
hµ (w1 N (µ1 , Σ1 ), w2 N (µ2 , Σ2 )) = −1/2 log(w1 w2 )+inf ((x−µ1 )Σ−1
1 (x−µ1 )+(x−µ2 )Σ2 (x−µ2 ))/4.
x
Proof. We have that
h(w1 N (µ1 , Σ1 ), w2 N (µ2 , Σ2 ))
Z
−1
−1
−n/2
−1/4
(2π)
(det(Σ1 Σ2 ))
exp(−((x − µ1 )Σ1 (x − µ1 ) + (x − µ2 )Σ2 (x − µ2 ))/4)dx .
= − 1/2 log(w1 w2 ) − log
−1
Letting x0 achieve the minimum value of ((x − µ1 )Σ−1
1 (x − µ1 ) + (x − µ2 )Σ2 (x − µ2 ))/4, this is
−1
−1/2 log(w1 w2 ) + ((x0 − µ1 )Σ−1
1 (x0 − µ1 ) + (x0 − µ2 )Σ2 (x0 − µ2 ))/4
Z
−1
−1
−n/2
−1/4
+ log
(2π)
(det(Σ1 Σ2 ))
exp(−(x − x0 )(Σ1 + Σ2 )(x − x0 )/4)dx .
Noting that the term at the end is simply hΣ (w1 N (µ1 , Σ1 ), w2 N (µ2 , Σ2 )) completes the proof.
We need one further proposition from which Theorem B.1 will follow easily.
Proposition B.7. Under
the assumptions of Theorem B.1, for each i there exists at most one j so that
√
C
h(Gi , Hj ) > (δ/k) .
To prove this, we will need one further lemma:
√
Lemma B.8. If h(Gi , Hj ) > (δ/k) C , with ΣG and ΣH the covariance matrices of the corresponding
Gaussians, then for A a sufficiently large constant (independent of C) ΣG ≤ AΣH .
Proof. Suppose for sake of contradiction that this is not the case. By making a change of variables, we can
′
assume that ΣG = I. This
p means that ΣH has some eigenvector v with eigenvalue less than 1/A. Let H be
1/4
log(k/δ) in the direction closer to the mean of Gi . We have that
Hj translated by C
√
h(Hj , H ′ ) = hµ (Hj , H ′ ) = Θ(A C log(k/δ)).
71
Therefore, V (Hj , H ′ ) = (δ/k)Ω(A
√
C) .
On the other hand,
h(Gi , H ′ ) = hΣ (Gi , H ′ ) + hµ (Gi , H ′ )
√
≤ hΣ (Gi , H) + O( C log(k/δ))
√
≤ h(Gi , H) + O( C log(k/δ))
√
= O( C log(k/δ)).
√
This means that V (G, H ′ ) = (δ/k)O( C) .
√
P
P
This means that p = ℓ Gℓ has V (p,√H ′ ) = (δ/k)O( C) , and since p is close to q = ℓ Hℓ , there must
be some ℓ so that V (H ′ , Hℓ ) > (δ/k)O( C) . Note that ℓ here cannot be j. On the other hand, this implies
that
h(Hj , Hℓ ) = hΣ (Hj , Hℓ ) + hµ (Hj , Hℓ )
√
≤ hΣ (H ′ , Hℓ ) + O(hµ (H ′ , Hℓ ) + A C log(k/δ))
√
≤ O(h(H ′ , Hℓ ) + A C log(k/δ))
√
= O(A C log(k/δ)).
Therefore, V (Hℓ , Hj ) = (δ/k)O(A
the proof.
√
C) ,
which for C ≫ A2 contradicts our assumptions. This completes
We are now prepared to prove Proposition B.7.
√
Proof. Suppose for sake of contradiction that V (Gi , Hj ), V (Gi , Hℓ ) > (δ/k) C for some j 6= ℓ. Then, by
Lemma B.8, we have that all of the covariance matrices of Gi , Hj , Hℓ are comparable to each other (namely
each is no more than a constant√multiple of any other). We claim that this implies that hµ (Hj , Hℓ ) =
O(hµ (Hj , Gi ) + hµ (Hℓ , Gi ) + C log(k/δ)). This is because, letting ΣG be the covariance matrix of Gi ,
′
′
and letting
√ wG = |Gi |1 , wH = |Hj |1 , wH = |Hℓ |1 , we have the following: First, each of wG , wH , wH =
exp(O( C log(δ/k))), because each distribution has large overlap with some other. Next, we have that
√
hµ (Gi , Hj ) = O( C log(δ/k)) + inf Θ((x − µGi )ΣG (x − µGi ) + (x − µHj )ΣG (x − µHh )
x
√
= O( C log(δ/k)) + Θ((µGi − µHj )ΣG (µGi − µHj )).
Similarly,
√
hµ (Gi , Hℓ ) = O( C log(δ/k)) + Θ((µGi − µHℓ )ΣG (µGi − µHℓ )),
and
√
hµ (Hj , Hℓ ) = O( C log(δ/k)) + Θ((µHj − µHℓ )ΣG (µHj − µHℓ )).
This implies that
hµ (Hj , Hℓ ) = O(hµ (Hj , G) + hµ (G, Hℓ )).
Therefore, we have that
h(Hj , Hℓ ) = hΣ (Hj , Hℓ ) + hµ (Hj , Hℓ )
= O(hΣ (Hj , G) + hµ (Hj , G) + hΣ (G, Hℓ ) + hµ (G, Hℓ ))
= O(h(Hj , G) + h(G, Hℓ ))
= O(log(1/V (Hj , G)) + log(1/V (G, Hℓ )))
√
= O( C log(k/δ)).
72
√
However, this implies that V (Hj , Hℓ ) = (δ/k)O(
This completes our proof.
C) ,
a contradiction.
We are now ready to prove Theorem B.1
√
Proof. For each Gi that has overlap more than (δ/k) C with some Hj , let π(i) be that j. For other i, define
π(i) arbitrarily subject to π being √
a permutation.
Note that V (Gi , Hj ) < (δ/k) C for any j 6= π(i). Also note that
V (Gi , q) ≥ V (Gi , p) − |p − q|1 = |Gi |1 − 2(δ/k)C .
On the other hand,
V (Gi , q) ≤
X
V (Gi , Hj )
j
≤ V (Gi , Hπ(i) ) +
X
V (Gi , Hj )
j6=π(i)
≤ V (Gi , Hπ(i) ) + δ/3.
Therefore V (Gi , Hπ(i) ) ≥ |Gi |1 − δ/2. It is also at most |Gi |1 − δ/2. On the other hand |Gi − Hπ(i) |1 =
|Gi |1 + |Hπ(i) |1 − 2V (Gi , Hπ(i) ) ≤ δ. This completes the proof.
C
Testing the Mean of a High-Dimensional Gaussian
√
Theorem C.1. There exists an algorithm that given ǫ > 0 and k = O( n/ǫ2 ) samples from an ndimensional Gaussian G = N (µ, I) distinguishes between the cases
• µ=0
• kµk2 > ǫ
with probability at least 2/3.
Proof. The tester is fairly simple. Let Xi be the ith sample, and let
k
1 X
Xi .
Z := √
k i=1
The algorithm returns “YES” if kZk22 < ǫ2 k/2 + n and “NO”
√ otherwise.
To show correctness, note that Z is distributed as N (µ k, I). If µ = 0, then kZk22 has mean n and
variance O(n), and so it is less than n + ǫ2k/2 with probability at least 2/3, assuming that k is a sufficiently
√
large multiple of n/ǫ2 . On the other hand, if kµk2 > ǫ, we note that kZk22 has mean n + kkµk22 and
√
variance O(n) + O(kkµk22 ). Thus, if kkµk22 ≫ n, the algorithm rejects with probability 2/3. Again, this
√
happens if kµk2 > ǫ and k is a sufficiently large multiple of n/ǫ2 . This completes the proof.
We also note that this tester can be implemented in the SQ model simply by verifying that each
√
coordinate-wise median has absolute value less than ǫ/ n, which can be verified by showing that Pr(xi >
√
0) = 1/2 + O(ǫ/ n).
We also show that the tester above is sample-optimal, up to a constant factor:
√
Theorem C.2. There is no algorithm that given k = o( n/ǫ2 ) samples from an n-dimensional Gaussian
G = N (µ, I) distinguishes between the cases
73
• µ=0
• kµk2 > ǫ
with probability at least 2/3.
Proof. Suppose for sake of contradiction that such an algorithm does exist. Consider the following scenario:
√
Let µ be taken from the distribution N (0, (2ǫ/ n)I). Note that kµk2 > ǫ with probability at least 9/10.
Let Y1 , Y2 , . . . , Yk be independent samples taken from N (µ, I). And let Z1 , . . . , Zk be independent samples
from N (0, I). Assuming that our algorithm exists, it can distinguish between a sample from Y1 , . . . , Yk
and a sample from Z1 , . . . , Zk with probability better than 1/2. This means that these distributions must
have constant variational distance. However, note that the vector (Z1 , . . . , Zk ) is simply a standard nkdimensional Gaussian. The vector (Y1 , . . . , Yk ) on the other hand is an nk-dimensional Gaussian with mean
0 and with
2
1 + 2ǫ /n , if ab = cd
Cov(Yab , Ycd ) = 2ǫ2 /n ,
if b = d and a 6= c
0,
otherwise .
By standard results, G′ = N (0, Σ) has constant variation distance from N (0, I) if and only if kΣ−IkF ≫ 1.
Taking Σ to be the covariance matrix for the Y ’s, we have that
√
kΣ − Ik2F = nk(2ǫ/ n)2 = 4kǫ2 /n = o(1) .
This implies that the distribution on Y ’s is close, in total variation distance, to the distribution on Z’s, and
gives a contradiction.
D
Omitted Proofs
D.1 Proof of Fact 3.6 We will need the following claim:
Claim D.1. We have that:
Hei (x cos θ + y sin θ) =
i
X
i
j
j=1
cosj θ sini−j θHej (x)Hei−j (y) .
Proof. The Hei (x) are monic polynomials: the lead term is xi with coefficient 1. Thus, all the degree-i
terms of Hei (x cos θ + y sin θ) are given by
i
X
i
(x cos θ + y sin θ)i
cosj θ sini−j θxj y i−j .
j
j=1
It follows that the degree-i terms of the LHS and RHS of the lemma agree. Therefore, we have
i
X
i
Hei (x cos θ + y sin θ) = p(x, y) +
cosj θ sini−j θHej (x)Hei−j (y) ,
j
j=1
for some polynomial p(x, y) of degree at most i − 1. We need to show that p(x, y) is identically zero. To
show this we consider E[Hei (X cos θ + Y sin θ)2 ], for (X, Y ) ∼ N (0, I). Since the Gaussian is unaltered
by rotations, by a change of coordinates we have that:
Z ∞Z ∞
2
E[Hei (X cos θ + Y sin θ) ] =
Hei (x cos θ + y sin θ)2 G(x)G(y)dxdy
−∞
Z−∞
∞ Z ∞
=
Hei (x′ )2 G(x′ )G(y ′ )dx′ dy ′ = i! .
−∞
−∞
74
However, pairs of distinct Hej (x)Hei−j (y) are orthogonal to each other and they are all orthogonal to the
lower degree polynomial p(x, y). Thus, we have
E[Hei (X cos θ + Y sin θ)2 ] = E[p(X, Y )2 ] +
i 2
X
i
j=1
= E[p(X, Y )2 ] +
j
i 2
X
i
j=1
= E[p(X, Y )2 ] + i!
j
cos2j θ sin2(i−j) θi!(i − j)!
i
X
i
j=1
2
cos2j θ sin2(i−j) θE[Hej (X)2 Hei−j (Y )2 ]
j
cos2j θ sin2(i−j) θ
= E[p(X, Y ) ] + i!(cos2 θ + sin2 θ)i = E[p(X, Y )2 ] + i! .
We must therefore have that E[p(X, Y )2 ] = 0. Since the Gaussian has positive pdf everywhere, this implies
that p(X, Y ) is identically zero.
We now have:
Uθ (Hei G)(x) =
=
=
Z
∞
Z−∞
∞
Hei (x cos θ + y sin θ)G(x cos θ + y sin θ)G(x sin θ − y cos θ)dy
Hei (x cos θ + y sin θ)G(x)G(y)dy
−∞
i
X
j=1
i
Z ∞
i
j
i−j
Hej (x)Hei−j (y)G(x sin θ − y cos θ)dy
cos θ sin θ
j
−∞
= cos θHei (x)G(x) ,
since
R∞
−∞ Hei−j (y)G(y)dy
= δij . This completes the proof.
D.2 Proof of Lemma 3.7 We use the following lemma:
Lemma D.2 (Proposition 1 from [CFJ13]). Given any 0 < ǫ < π/2, let θ be the angle between two random
unit vectors uniformly distributed over Sn . Then we have that:
√
Pr[|θ − π/2| ≥ ǫ] ≤ O( n(cos ǫ)n−2 ) .
As a corollary, we have:
Corollary D.3. Let θ be the angle between two random unit vectors uniformly distributed over Sn . Then we
have that:
Pr | cos θ| ≥ Ω(n−α ) ≤ exp −Ω(n1−2α ) ,
for any 0 ≤ α ≤ 1/2.
Proof. We apply Lemma D.2 with ǫ = n−α . If n−α = O(1), the result is trivial, so we may assume that
ǫ ≤ 1/100. Then we have that cos ǫ ≤ 1 − ǫ2 /2 + ǫ2 /24 ≤ 1 − ǫ2 /3 ≤ exp(ǫ2 /4). Lemma D.2 now gives
that
√
Pr |θ − π/2| ≥ n−α ≤ O n exp(−n−2α /4)n−2 ≤ exp(−n1−2α /5) .
Note that if |θ − π/2|≤n−α , it follows that | cos θ| ≤ |θ − π/2| ≤ n−α .
75
Using Corollary D.3 for α = 1/2 − c, and a union bound over all pairs of distinct vectors in S, the
probability that there exist v 6= v ′ ∈ S such that |v · v ′ |≥Ω(nc−1/2 ) is less than
|S|2 2−Ω(n
2c )
<1.
Therefore, the set S will satisfy the statement of Lemma 3.7 with positive probability, as desired.
D.3 Proof of Fact 6.2 We have:
Z
2
1 + χ (wB + (1 − w)C, D) = (wB(x) + (1 − w)C(x))2 /D(x)dx
Z
Z
Z
2
2
2
2
=w
B(x) /D(x)dx + (1 − w)
C(x) /D(x)dx + 2w(1 − w) B(x)C(x)/D(x)dx
= w2 (1 + χ2 (B, D)) + (1 − w)2 (1 + χ2 (C, D)) + 2w(1 − w)(1 + χD (B, C))
= 1 + w2 χ2 (B, D) + (1 − w)2 χ2 (C, D) + 2w(1 − w)χD (B, C) .
This completes the proof.
D.4 Proof of Fact 6.3 By definition, we can write
Z ∞
′
G(x − µ′ )G(x − µ)/G(x)dx
1 + χN (0,1) (N (µ , 1), N (µ, 1)) =
−∞
Z ∞
√
= (1/ 2π) ·
exp −(x − µ′ )2 /2 − (x − µ)2 /2 + x2 /2 dx
Z−∞
∞
√
exp −x2 /2 + (µ′ + µ)x − µ2 /2 − µ′2 /2 dx
= (1/ 2π) ·
−∞
Z ∞
G(x − µ − µ′ ) exp −µ2 /2 + −µ′2 /2 + (µ + µ′ )2 /2 dx
=
−∞
= exp(µ′ µ) .
This completes the proof.
D.5 Proof of Fact 6.4 By definition, we have that
Z ∞
2
2
G(x/σ)2 /G(x)dx
1 + χ (N (0, σ ), N (0, 1)) = (1/σ)
−∞
Z ∞
1
= √
exp(x2 /2 − x2 /σ 2 )dx
2
σ 2π −∞
p
Z ∞
p
2/σ 2 − 1
1
p
G(x/
2/σ 2 − 1)dx
=
·
σ2
2/σ 2 − 1
−∞
p
2/σ 2 − 1 p
=
= 2/σ 4 − 1/σ 2 .
σ2
This completes the proof.
76
| 8cs.DS
|
PROC. OF THE 15th PYTHON IN SCIENCE CONF. (SCIPY 2017)
1
MatchPy: A Pattern Matching Library
Manuel Krebber‡∗ , Henrik Bartels‡ , Paolo Bientinesi‡
arXiv:1710.06915v1 [cs.PL] 16 Oct 2017
F
Abstract—Pattern matching is a powerful tool for symbolic computations, based
on the well-defined theory of term rewriting systems. Application domains include algebraic expressions, abstract syntax trees, and XML and JSON data.
Unfortunately, no lightweight implementation of pattern matching as general and
flexible as Mathematica exists for Python [Pö16], [Hao14], [Sch14], [Jen15].
Therefore, we created the open source module MatchPy which offers similar
pattern matching functionality in Python using a novel algorithm which finds
matches for large pattern sets more efficiently by exploiting similarities between
patterns.
Index Terms—pattern matching, symbolic computation, discrimination nets,
term rewriting systems
Introduction
Pattern matching is a powerful tool which is part of many
functional programming languages as well as computer algebra
systems such as Mathematica. It is useful for many applications
including symbolic computation, term simplification, term rewriting systems, automated theorem proving, and model checking. In
this paper, we present a Python-based pattern matching library and
its underlying algorithms.
The goal of pattern matching is to find a match substitution given a subject term and a pattern which is a term with
placeholders [BN98]. The substitution maps placeholders in the
pattern to replacement terms. A match is a substitution that can
be applied to the pattern yielding the original subject. As an
example consider the subject f (a) and the pattern f (x) where x is a
placeholder. Then the substitution σ = {x 7→ a} is a match because
σ ( f (x)) = f (a). This form of pattern matching without any special
properties of function symbols is called syntactic matching. For
syntactic patterns, the match is unique if it exists.
Among the existing systems, Mathematica [WR16] arguably
offers the most expressive pattern matching. Its pattern matching
offers similar expressiveness as Python’s regular expressions, but
for symbolic tree structures instead of strings. While pattern
matching can handle nested expressions up to arbitrary depth,
regular expressions cannot properly handle such nesting. Patterns
are used widely in Mathematica, e.g. in function definitions or for
manipulating terms. It is possible to define custom function symbols which can be associative and/or commutative. Mathematica
also offers sequence variables which can match a sequence of
terms instead of a single one. These are especially useful when
* Corresponding author: [email protected]
‡ RWTH Aachen University, AICES, HPAC Group
c 2017 Manuel Krebber et al. This is an open-access article
Copyright ○
distributed under the terms of the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium,
provided the original author and source are credited.
working with variadic function symbols. Mathics [Pö16] is an
open source computer algebra system written in Python that aims
to replicate the syntax and functionality of Mathematica.
To our knowledge, no existing work covers pattern matching
with function symbols which are either commutative or associative
but not both at the same time. However, there are functions
with those properties, e.g. matrix multiplication or arithmetic
mean. Most of the existing pattern matching libraries for Python
only support syntactic patterns. Associativity, commutativity and
sequence variables make multiple distinct matches possible for a
single pattern. In addition, pattern matching with either associativity or commutativity is NP-complete in both cases [BKN87].
While the pattern matching in SymPy [MSP+ 17] can work with
associative/commutative functions, it is limited to finding a single
match. Nonetheless, for some applications it is interesting to find
all possible matches for a pattern, e.g. because matches need to
be processed further recursively to solve an optimization problem.
Furthermore, SymPy does not support sequence variables and is
limited to a predefined set of mathematical operations.
In many applications, a fixed set of patterns is matched
repeatedly against different subjects. The simultaneous matching of multiple patterns is called many-to-one matching, as
opposed to one-to-one matching which denotes matching with
a single pattern. Many-to-one matching can gain a significant
speed increase compared to one-to-one matching by exploiting
similarities between patterns. This has already been the subject of research for both syntactic [Chr93], [Grä91], [NWE97]
and associative-commutative pattern matching [KL91], [BCR93],
[LM94], [BCR+ 95], [Eke95], [KM01], but not with the full
feature set described above. Discrimination nets [BW84] are the
state-of-the-art solution for many-to-one matching. Our goal is
to generalize this approach to support all the aforementioned
features.
In this paper, we present the open-source library for Python
MatchPy which provides pattern matching with sequence variables
and associative/commutative function symbols. In addition to
standard one-to-one matching, MatchPy also includes an efficient
many-to-one matching algorithm that uses generalized discrimination nets. First, we give an overview of what MatchPy can be
used for. Secondly, we explain some of the challenges arising
from the non-syntactic pattern matching features and how we
solve them. Then we give an overview of how many-to-one
matching is realized and optimized in MatchPy. Next, we present
our experiments where we observed significant speedups of the
many-to-one matching over one-to-one matching. Finally, we draw
some conclusions from the experiments and propose future work
on MatchPy.
2
PROC. OF THE 15th PYTHON IN SCIENCE CONF. (SCIPY 2017)
Usage Overview
Operation
MatchPy can be installed using pip and all necessary classes can
be imported from the toplevel module matchpy. Expressions in
MatchPy consist of constant symbols and operations. For patterns,
wildcards can also be used as placeholders. We use Mathematica’s
notation for wildcards, i.e. we append underscores to wildcard
names to distinguish them from symbols.
MatchPy can be used with native Python types such as list
and int. The following is an example of how the subject [0,
1] can be matched against the pattern [x_, 1]. The expected
match here is the replacement 0 for x_. We use next because we
only want to use the first (and in this case only) match of the
pattern:
>>> x_ = Wildcard.dot(’x’)
>>> next(match([0, 1], Pattern([x_, 1])))
{’x’: 0}
In addition to regular (dot) variables, MatchPy also supports
sequence wildcards. They can match a sequence of arguments
and we denote them with two or three trailing underscores for
plus and star wildcards, respectively. Star wildcards can match
an empty sequence, while plus wildcards require at least one
argument to match. The terminology is borrowed from regular
expressions where *, + and . are used for similar concepts.
>>> y___ = Wildcard.star(’y’)
>>> next(match([1, 2, 3], Pattern([x_, y___])))
{’x’: 1, ’y’: (2, 3)}
In the following, we omit the definition of new variables as they
can be done in the same way. In addition to native types, one
can also define custom operations by creating a subclass of the
Operation class:
class MyOp(Operation):
name = ’MyOp’
arity = Arity.variadic
associative = True
commutative = True
The name is a required attribute, while the others are optional and
influence the behavior of the operations. By default, operations
are variadic and neither commutative nor associative. Nested
associative operations have to be variadic and are automatically
flattened. Furthermore, regular variables behave similar to sequence variables as arguments of associative functions, because
the associativity allows arbitrary parenthesization of arguments:
>>> next(match(MyOp(0, 1, 2), Pattern(MyOp(x_, 2))))
{’x’: MyOp(0, 1)}
The argument of commutative operations are automatically sorted.
Note that patterns with commutative operations can have multiple
matches, because their arguments can be reordered arbitrarily.
>>> list(match(MyOp(1, 2), Pattern(MyOp(x_, z_))))
[{’x’: 2, ’z’: 1}, {’x’: 1, ’z’: 2}]
We can use the CustomConstraint class to create a constraint
that checks whether a is smaller than b:
a_lt_b = CustomConstraint(lambda a, b: a < b)
The lambda function gets called with the variable substitutions
based on their name. The order of arguments is not important and
it is possible to only use a subset of the variables in the pattern.
With this constraint we can define a replacement rule that basically
describes bubble sort:
>>> pattern = Pattern([h___, b_, a_, t___], a_lt_b)
>>> rule = ReplacementRule(pattern,
lambda a, b, h, t: [*h, a, b, *t])
Symbol
Arity
Multiplication
Addition
×
+
variadic
variadic
Transposition
Inversion
Inversion and Transposition
T
−1
unary
unary
unary
−T
Properties
associative
associative,
commutative
TABLE 1: Linear Algebra Operations
The replacement function gets called with all matched variables
as keyword arguments and needs to return the replacement. This
replacement rule can be used to sort a list when applied repeatedly
with replace_all:
>>> replace_all([1, 4, 3, 2], [rule])
[1, 2, 3, 4]
Sequence variables can also be used to match subsequences that
match a constraint. For example, we can use the this feature to
find all subsequences of integers that sum up to 5. In the following
example, we use anonymous wildcards which have no name and
are hence not part of the match substitution:
>>> x_sums_to_5 = CustomConstraint(
...
lambda x: sum(x) == 5)
>>> pattern = Pattern([___, x__, ___], x_sums_to_5)
>>> list(match([1, 2, 3, 1, 1, 2], pattern))
[{’x’: (2, 3)}, {’x’: (3, 1, 1)}]
More examples can be found in MatchPy’s documentation
[Kre17a].
Application Example: Finding matches for a BLAS kernel
BLAS is a collection of optimized routines that can compute specific linear algebra operations efficiently [LHKK79], [DCHH88],
[DCHD90]. As an example, assume we want to match all subexpressions of a linear algebra expression which can be computed by
the ?TRMM BLAS routine. These have the form α × op(A) × B
or α × B × op(A) where op(A) is either the identity function or
transposition, and A is a triangular matrix. For this example, we
leave out all variants where α 6= 1.
In order to model the linear algebra expressions, we use the
operations shown in Table 1. In addition, we have special symbol
subclasses for scalars, vectors and matrices. Matrices also have a
set of properties, e.g. they can be triangular, symmetric, square,
etc. For those patterns we also use a special kind of dot variable
which is restricted to only match a specific kind of symbol. Finally,
we construct the patterns using sequence variables to capture the
remaining operands of the multiplication:
A_ = Wildcard.symbol(’A’, Matrix)
B_ = Wildcard.symbol(’B’, Matrix)
A_is_triangular = CustomConstraint(
lambda A: ’triangular’ in A.properties)
trmm_patterns = [
Pattern(Times(h___,
A_is_triangular),
Pattern(Times(h___,
A_is_triangular),
Pattern(Times(h___,
A_is_triangular),
Pattern(Times(h___,
A_is_triangular),
]
A_, B_, t___),
Transpose(A_), B_, t___),
B_, A_, t___),
B_, Transpose(A_), t___),
MATCHPY: A PATTERN MATCHING LIBRARY
With these patterns, we can find all matches for the ?TRMM
routine within a product. In this example, M1, M2 and M3 are
matrices, but only M3 is triangular:
>>> expr = Times(Transpose(M3), M1, M3, M2)
>>> for i, pattern in enumerate(trmm_patterns):
...
for substitution in match(expr, pattern):
...
print(’{} with {}’.format(i, substitution))
0 with {A -> M3, B -> M2, t -> (), h -> ((M3)^T, M1)}
1 with {A -> M3, B -> M1, t -> (M3, M2), h -> ()}
2 with {A -> M3, B -> M1, t -> (M2), h -> ((M3)^T)}
3
Sequence Variables in Commutative Functions
The distribution of n subjects subterms onto m sequence variables
within a commutative function symbol can yield up to mn distinct
solutions. Enumerating all of the solutions is accomplished by
generating and solving several linear Diophantine equations. As an
example, lets assume we want to match f(a, b, b, b) with
f(x___, y__, y__) where f is commutative. This means
that the possible distributions are given by the non-negative integer
solutions of these equations:
As can be seen in the output, a total of three matches are found.
1
=
xa + 2ya
3
=
xb + 2yb
Design Considerations
There are plenty of implementations of syntactic matching and
the algorithms are well known. Implementing pattern matching
for MatchPy poses some challenges such as associativity and
commutativity.
Associativity/Sequence variables
Associativity enables arbitrary grouping of arguments for matching: For example, 1 + a + b matches 1 + x_ with {x 7→
a + b} because we can group the arguments as 1 + (a + b).
Basically, when regular variables are arguments of an associative
function, they behave like sequence variables. Both can result
in multiple distinct matches for a single pattern. In contrast,
for syntactic patterns there is always at most one match. This
means that the matching algorithm needs to be non-deterministic
to explore all potential matches for associative terms or terms
with sequence variables. We employ backtracking with the help
of Python generators to enable this. Associative matching is NPcomplete [BKN87].
Commutativity
Matching commutative terms is difficult because matches need
to be found independent of the argument order. Commutative
matching has been shown to be NP-complete, too [BKN87]. It
is possible to find all matches by matching all permutations of
the subjects arguments against all permutations of the pattern
arguments. However, with this naive approach, a total of n!m!
combinations have to be matched where n is the number of subject
arguments and m the number of pattern arguments. It is likely
that most of these combinations do not match or yield redundant
matches.
Instead, we interpret the arguments as a multiset, i.e. an orderless collection that allows repetition of elements. Also, we use
the following order for matching the subterms of a commutative
term:
1. Constant arguments
2. Matched variables, i.e. variables that already have a
value assigned in the current substitution
3. Non-variable arguments
4. Repeat step 2
5. Regular variables
6. Sequence variables
Each of those steps reduces the search space for successive
steps. This also means that if one step finds no match, the
remaining steps do not have to be performed. Note that steps 3, 5
and 6 can yield multiple matches and backtracking is employed to
check every combination. Since step 6 is the most involved, it is
described in more detail in the next section.
xa determines how many times a is included in the substitution for
x. Because y__ requires at least one term, we have the additional
constraint ya + yb ≥ 1. The only possible solution xa = xb = yb =
1 ∧ ya = 0 corresponds to the match substitution {x 7→ (a, b), y 7→
(b)}.
Extensive research has been done on solving linear Diophantine equations and linear Diophantine equation systems [Wei60],
[Bon67], [Lam88], [CF89], [AHL00]. In our case the equations
are actually independent expect for the additional constraints for
plus variables. Also, the non-negative solutions can be found more
easily. We use an adaptation of the algorithm used in SymPy which
recursively reduces any linear Diophantine equation to equations
of the form ax + by = d. Those can be solved efficiently with the
Extended Euclidian algorithm [MVO96]. Then the solutions for
those can be combined into a solution for the original equation.
All coefficients in those equations are likely very small since
they correspond to the multiplicity of sequence variables. Similarly, the number of variables in the equations is usually small as
they map to sequence variables. The constant is the multiplicity of
a subject term and hence also usually small. Overall, the number
of distinct equations that are solved is small and the solutions are
cached. This reduces the impact of the sequence variables on the
overall run time.
Optimizations
Since most applications for pattern matching repeatedly match a
fixed set of patterns against multiple subjects, we implemented
many-to-one matching for MatchPy. The goal of many-to-one
matching is to utilize similarities between patterns to match them
more efficiently. In this section, we give a brief overview of the
many-to-one matching algorithm used by MatchPy. Full details
can be found in the master thesis [Kre17b].
Many-to-one Matching
MatchPy includes two additional algorithms for matching:
ManyToOneMatcher and DiscriminationNet. Both enable matching multiple patterns against a single subject much
faster than matching each pattern individually using match. The
latter can only be used for syntactic patterns and implements a
state-of-the-art deterministic discrimination net. A discrimination
net is a data structure similar to a decision tree or a finite automaton [Chr93], [Grä91], [NWE97]. The ManyToOneMatcher
utilizes a generalized form of non-deterministic discrimination
nets that support sequence variables and associative function
symbols. Furthermore, as elaborated in the next section, it can
also match commutative terms.
In Figure 1, an example for a non-deterministic discrimination
net is shown. It contains three patterns that match Python lists:
4
PROC. OF THE 15th PYTHON IN SCIENCE CONF. (SCIPY 2017)
]
0
[y , 0]
y
Patterns
Subjects
Patterns
Subjects
1
x
a
1 1
x
a
1
2
x
a
2 2
x
a
2
[
1
]
x
[1, x
]
]
[1]
Fig. 2: Example for Order in Bipartite Graph.
Fig. 1: Example Discrimination Net.
One matches the list that consists of a single 1, the second one
matches a list with exactly two elements where the last element
is 0, and the third pattern matches any list where the first element
is 1. Note, that these patterns can also match nested lists, e.g. the
second pattern would also match [[2, 1], 0].
Matching starts at the root and proceeds along the transitions.
Simultaneously, the subject is traversed in preorder and each symbol is checked against the transitions. Only transitions matching
the current subterm can be used. Once a final state is reached,
its label gives a list of matching patterns. For non-deterministic
discrimination nets, all possibilities need to be explored via backtracking. The discrimination net allows to reduce the matching
costs, because common parts of different pattern only need to be
matched once. For non-matching transitions, their whole subtree is
pruned and all the patterns are excluded at once, further reducing
the match cost.
In Figure 1, for the subject [1, 0], there are two paths and
therefore two matching patterns: [y_, 0] matches with {y 7→ 1}
and [1, x___] matches with {x 7→ 0}. Both the y-transition and
the 1-transition can be used in the second state to match a 1.
Compared to existing discrimination net variants, we added
transitions for the end of a compound term to support variadic
functions. Furthermore, we added support for both associative
function symbols and sequence variables. Finally, our discrimination net supports transitions restricted to symbol classes (i.e.
Symbol subclasses) in addition to the ones that match just a specific symbol. We decided to use a non-deterministic discrimination
net instead of a deterministic one, since the number of states of
the later would grow exponentially with the number of patterns.
While the DiscriminationNet also has support for sequence
variables, in practice the net became to large to use with just a
dozen patterns.
Commutative Many-to-one Matching
Many-to-one matching for commutative terms is more involved.
We use a nested CommutativeMatcher which in turn uses
another ManyToOneMatcher to match the subterms. Our approach is similar to the one used by Bachmair and Kirchner in their
respective works [BCR+ 95], [KM01]. We match all the subterms
of the commutative function in the subject with a many-to-one
matcher constructed from the subpatterns of the commutative
function in the pattern (except for sequence variables, which
are handled separately). The resulting matches form a bipartite
graph, where one set of nodes consists of the subject subterms
and the other contains all the pattern subterms. Two nodes are
connected by an edge iff the pattern matches the subject. Such
an edge is also labeled with the match substitution(s). Finding
an overall match is then accomplished by finding a maximum
matching in this graph. However, for the matching to be valid, all
the substitutions on its edges must be compatible, i.e. they cannot
have contradicting replacements for the same variable. We use
the Hopcroft-Karp algorithm [HK73] to find an initial maximum
matching. However, since we are also interested in all matches
and the initial matching might have incompatible substitutions, we
use the algorithm described by Uno, Fukuda and Matsui [FM94],
[Uno97] to enumerate all maximum matchings.
To avoid yielding redundant matches, we extended the bipartite
graph by introducing a total order over its two node sets. This
enables determining whether the edges of a matching maintain
the order induced by the subjects or whether some of the edges
"cross". Formally, for all edge pairs (p, s), (p0 , s0 ) ∈ M we require
(s ≡ s0 ∧ p > p0 ) =⇒ s > s0 to hold where M is the matching,
s, s0 are subjects, and p, p0 are patterns. An example of this is
given in Figure 2. The order of the nodes is indicated by the
numbers next to them. The only two maximum matchings for
this particular match graph are displayed. In the left matching,
the edges with the same subject cross and hence this matching
is discarded. The other matching is used because it maintains the
order. This ensures that only unique matches are yielded. Once
a matching for the subpatterns is obtained, the remaining subject
arguments are distributed to sequence variables in the same way
as for one-to-one matching.
Experiments
To evaluate the performance of MatchPy, we conducted experiments on an Intel Core i5-2500K 3.3 GHz CPU with 8GB of
RAM. Our focus is on relative performance of one-to-one and
many-to-one matching rather than the absolute performance.
Linear Algebra
The operations for the linear algebra problem are shown in Table
1. The patterns all match BLAS kernels similar to the example
pattern which was previously described. The pattern set consists of
199 such patterns. Out of those, 61 have an addition as outermost
operation, 135 are patterns for products, and 3 are patterns for
single matrices. A lot of these patterns only differ in terms of
constraints, e.g. there are ten distinct patterns matching A × B
with different constraints on the two matrices. By removing the
sequence variables from the product patterns, these pattern can be
made syntactic when ignoring the multiplication’s associativity.
In the following, we refer to the set of patterns with sequence
variables as LinAlg and the set of syntactic product patterns as
Syntactic.
The subjects were randomly generated such that matrices
had random properties and each factor could randomly be transposed/inverted. The number of factors was chosen according to
a normal distribution with µ = 5. The total subject set consisted
of 70 random products and 30 random sums. Out of the pattern
set, random subsets were used to examine the influence of the
pattern set size on the matching time. Across multiple subsets
and repetitions per subject, the mean match and setup times
were measured. Matching was performed both with the match
MATCHPY: A PATTERN MATCHING LIBRARY
5
Setup Time
power
Match Time
60
Match Time [ms]
Setup Time [ms]
10
40
20
match
MTOM
trailer
5
arglist
atom
0
0
50
100
50
150
Number of Patterns
100
150
testlist gexp
Number of Patterns
Fig. 3: Timing Results for LinAlg.
Speedup
Fig. 6: AST of the isinstance pattern.
Break Even
8
Number of Subjects
20
Speedup
isinstance ( any , ( any . . . any ) )
15
10
5
0
50
100
6
4
2
0
150
reaching a speedup of up to 60. However, because it also has the
highest setup time, it only outperforms the many-to-one matcher
after about 100 subjects for larger pattern set sizes. In practice, the
discrimination net is likely the best choice for syntactic patterns, as
long as the discrimination net does not grow to large. In the worst
case, the size of the discrimination net can grow exponentially in
the number of patterns.
50
Number of Patterns
100
150
Number of Patterns
Fig. 4: Comparison for LinAlg.
Abstract Syntax Trees
function and the ManyToOneMatcher (MTOM). The results
are displayed in Figure 3.
As expected, both setup and match times grow with the pattern
set size. The growth of the many-to-one match time is much slower
than the one for one-to-one matching. This is also expected since
the simultaneous matching is more efficient. However, the growth
of setup time for the many-to-one matcher beckons the question
whether the speedup of the many-to-one matching is worth it.
Figure 4 depicts both the speedup and the break even point for
many-to-one matching for LinAlg. The first graph indicates that
the speedup of many-to-one matching increases with larger pattern
sets. But in order to profit from that speedup, the setup cost of
many-to-one matching must be amortized. Therefore, the second
graph shows the break even point for many-to-one matching in
terms of number of subjects. If for a given number of patterns and
subjects the corresponding point is above the line, then many-toone matching is overall faster. In this example, when matching
more than eight times, many-to-one matching is overall always
faster than one-to-one matching.
For the syntactic product patterns we compared the
match function, the ManyToOneMatcher (MTOM) and the
DiscriminationNet (DN). Again, randomly generated subjects were used. The resulting speedups and break even points are
displayed in Figure 5.
In this case, the discrimination net is the fastest overall
Speedup
Number of Subjects
Speedup
60
Break Even
MTOM
DN
40
20
0
100
MTOM
DN
50
0
20
40
60
80
100
Number of Patterns
120
20
40
60
80
100
Number of Patterns
Fig. 5: Comparison for Syntactic.
120
Python includes a tool to convert code from Python 2 to Python
3. It is part of the standard library package lib2to3 which has
a collection of "fixers" that each convert one of the incompatible
cases. To find matching parts of the code, those fixers use pattern
matching on the abstract syntax tree (AST). Such an AST can
be represented in the MatchPy data structures. We converted
some of the patterns used by lib2to3 both to demonstrate the
generality of MatchPy and to evaluate the performance of manyto-one matching. Because the fixers are applied one after another
and can modify the AST after each match, it would be difficult to
use many-to-one matching for lib2to3 in practice.
The following is an example of such a pattern:
power<
’isinstance’
trailer< ’(’ arglist< any ’,’ atom< ’(’
args=testlist_gexp< any+ >
’)’ > > ’)’ >
>
It matches an isinstance expression with a tuple as second
argument. Its tree structure is illustrated in Figure 6. The corresponding fixer cleans up duplications generated by previous fixers.
For example isinstance(x, (int, long)) would be converted
by another fixer into isinstance(x, (int, int)), which in
turn is then simplified to isinstance(x, int) by this fixer.
Out of the original 46 patterns, 36 could be converted to
MatchPy patterns. Some patterns could not be converted, because
they contain features that MatchPy does not support yet. The
features include negated subpatterns (e.g. not atom<’(’ [any]
’)’>) and subpatterns that allow an aritrary number of repetitions
(e.g. any (’,’ any)+).
Furthermore, some of the AST patterns contain alternative
or optional subpatterns, e.g. power<’input’ args=trailer<’(’
[any] ’)’>>. These features are also not directly supported by
MatchPy, but they can be replicated by using multiple patterns.
For those lib2to3 patterns, all combinations of the alternatives
were generated and added as invividual patterns. This resulted in
about 1200 patterns for the many-to-one matcher that completely
cover the original 36 patterns.
6
For the experiments, we used a file that combines the examples
from the unittests of lib2to3 with about 900 non-empty lines.
We compared the set of 36 patterns with the original matcher
and the 1200 patterns with the many-to-one matcher. A total of
about 560 matches are found. Overall, on average, our manyto-one matcher takes 0.7 seconds to find all matches, while the
matcher from lib2to3 takes 1.8 seconds. This yields a speedup
of approximately 2.5. However, the construction of the many-toone matcher takes 1.4 seconds on average. However, this setup
cost will be amortized by the faster matching for sufficiently large
ASTs. The setup time can also mostly be eliminated by saving the
many-to-one matcher to disk and loading it once required.
Compared to the one-to-one matching in MatchPy, the manyto-one matching achieves a speedup of about 60. This is due to
the fact that for any given subject less than 1% of patterns match.
By taking into account the setup time of the many-to-one matcher,
the break even point for it is at about 200 subjects.
Conclusions
We have presented MatchPy, a pattern matching library for Python
with support for sequence variables and associative/commutative
functions. This library includes algorithms and data structures
for both one-to-one and many-to-one matching. Because nonsyntactic pattern matching is NP-hard, in the worst case the
pattern matching times grows exponentially with the length of
the pattern. Nonetheless, our experiments on real world examples
indicate that many-to-one matching can give a significant speedup
over one-to-one matching. However, the employed discrimination
nets come with a one-time construction cost which needs to be
amortized to benefit from their speedup. In our experiments, the
break even point for many-to-one matching was always reached
well within the typical number of subjects for the respective
application. Therefore, many-to-one matching is likely to result
in a compelling speedup in practice.
For syntactic patterns, we also compared the syntactic discrimination net with the many-to-one matcher. As expected, discrimination nets are faster at matching, but also have a significantly
higher setup time. Furthermore, the number of states can grow
exponentially with the number of patterns, making them unsuitable
for some pattern sets. Overall, if applicable, discrimination nets
offer better performance than a many-to-one matcher.
Which pattern matching algorithm is the fastest for a given application depends on many factors. Hence, it is not possible to give
a general recommendation. Yet, the more subjects are matched
against the same pattern set, the more likely it is that many-to-one
outperforms one-to-one matching. In the experiments, a higher
number of patterns lead to an increase of the speedup of manyto-one matching. In terms of the size of the many-to-one matcher,
the growth of the net was sublinear in our experiments and still
feasible for large pattern sets. The efficiency of using many-toone matching also heavily depends on the actual pattern set, i.e.
the degree of similarity and overlap between the patterns.
PROC. OF THE 15th PYTHON IN SCIENCE CONF. (SCIPY 2017)
Additional pattern features
In the future, we plan to implement similar functionality to the
Repeated, Sequence, and Alternatives functions from
Mathematica. These provide another level of expressive power
which cannot be fully replicated with the current feature set
of MatchPy. Another useful feature are context variables as
described by Kutsia [Kut06]. They allow matching subterms at
arbitrary depths which is especially useful for structures like
XML. With context variables, MatchPy’s pattern matching would
be as powerful as XPath [RDS17] or CSS selectors [RJE17] for
such structures. Similarly, function variables which can match a
function symbol would also be useful for those applications.
Integration
Currently, in order to use MatchPy, existing data structures must
be adapted to provide their children via an iterator. Where that is
not possible, for example because the data structures are provided
by a third party library, translation functions need to be applied.
Also, some native data structures such as dicts are currently not
supported directly. Therefore, it would be useful, to have a better
way of using existing data structures with MatchPy.
In particular, easy integration with SymPy is an important goal,
since it is a popular tool for working with symbolic mathematics.
SymPy already implements a form of pattern matching which
is less powerful than MatchPy. It lacks support for sequence
variables, symbol wildcards and constraints. Each constant symbol
in SymPy can have properties that allow it to be commutative or
non-commutative. One benefit of this approach is easier modeling
of linear algebra multiplication, where matrices and vectors do
not commute, but scalars do. Better integration of MatchPy with
SymPy would provide the users of SymPy with more powerful
pattern matching tools. However, Matchpy would require selective
commutativity to be fully compatible with SymPy. Also, SymPy
supports older Python versions, while MatchPy requires Python
3.6.
Performance
If pattern matching is a major part of an application, its running
time can significantly impact the overall speed. Reimplementing parts of MatchPy as a C module would likely result in a
substantial speedup. Alternatively, adapting part of the code to
Cython could be another option to increase the speed [BBS09],
[WLØ09]. Furthermore, generating source code for a pattern set
similar to parser generators for formal grammars could improve
matching performance. While code generation for syntactic pattern matching has been the subject of various works [Aug85],
[FM01], [Mar08], [MRV03], its application with the extended
feature set of MatchPy is another potential area of future research.
Also, additonal research on the viability of pattern matching with
increasingly complex and large subjects or patterns is desirable.
Parallelizing many-to-one matching is also a possibility to increase
the overall speed which is worth exploring.
Functional pattern matching
Future Work
We plan on extending MatchPy with more powerful pattern
matching features to make it useful for an even wider range
of applications. The greatest challenge with additional features
is likely to implement them for many-to-one matching. In the
following, we discuss some possibilities for extending the library.
Since Python does not have pattern matching as a language
feature, MatchPy could be extended to provide a syntax similar
to other functional programming languages. However, without a
switch statement as part of the language, there is a limit to the
syntax of this pattern expression. The following is an example of
what such a syntax could look like:
MATCHPY: A PATTERN MATCHING LIBRARY
with match(f(a, b)):
if case(f(x_, y_)):
print("x={}, y={}".format(x, y)))
elif case(f(z_)):
....
There are already several libraries for Python which implement
such a functionality for syntactic patterns and native data structures (e.g. MacroPy [Hao14], patterns [Sch14] or PyPatt [Jen15]).
However, the usefulness of this feature needs further evaluation.
R EFERENCES
[AHL00]
Karen Aardal, Cor A. J. Hurkens, and Arjen K. Lenstra. Solving
a system of linear diophantine equations with lower and upper
bounds on the variables. Mathematics of Operations Research,
25(3):427–442, August 2000. doi:10.1287/moor.25.3.
427.12219.
[Aug85]
Lennart Augustsson. Compiling pattern matching. In Proc.
of a Conference on Functional Programming Languages and
Computer Architecture, pages 368–381, New York, NY, USA,
1985. Springer-Verlag New York, Inc.
[BBS09]
Stefan Behnel, Robert W. Bradshaw, and Dag Sverre Seljebotn.
Cython tutorial. In Gaël Varoquaux, Stéfan van der Walt, and
Jarrod Millman, editors, Proceedings of the 8th Python in Science
Conference, pages 4–14, Pasadena, CA USA, 2009.
[BCR93] Leo Bachmair, Ta Chen, and I. V. Ramakrishnan. Associativecommutative Discrimination Nets, pages 61–74. 1993. doi:
10.1007/3-540-56610-4_56.
[BCR+ 95] Leo Bachmair, Ta Chen, I. V. Ramakrishnan, Siva Anantharaman,
and Jacques Chabin. Experiments with associative-commutative
discrimination nets. In Proceedings of the 14th International Joint
Conference on Artificial Intelligence - Volume 1, IJCAI’95, pages
348–354. Morgan Kaufmann, 1995.
[BKN87] Dan Benanav, Deepak Kapur, and Paliath Narendran. Complexity of matching problems. Journal of Symbolic Computation, 3(1):203–216, February 1987. doi:10.1016/S07477171(87)80027-5.
[BN98]
Franz Baader and Tobias Nipkow. Term Rewriting and All That.
Cambridge University Press, New York, NY, USA, 1998.
[Bon67]
James Bond. Calculating the general solution of a linear diophantine equation. The American Mathematical Monthly, 74(8):955,
October 1967. doi:10.2307/2315274.
[BW84]
Alan Bundy and Lincoln Wallen. Discrimination Net, pages 31–
31. 1984. doi:10.1007/978-3-642-96868-6_59.
[CF89]
Michael Clausen and Albrecht Fortenbacher. Efficient solution
of linear diophantine equations. Journal of Symbolic Computation, 8(1-2):201–216, July 1989. doi:10.1016/s07477171(89)80025-2.
[Chr93]
Jim Christian. Flatterms, discrimination nets, and fast term
rewriting. Journal of Automated Reasoning, 10(1):95–113, 1993.
doi:10.1007/BF00881866.
[DCHD90] J. J. Dongarra, Jeremy Du Croz, Sven Hammarling, and I. S.
Duff. A set of level 3 basic linear algebra subprograms. ACM
Transactions on Mathematical Software, 16(1):1–17, March 1990.
doi:10.1145/77626.79170.
[DCHH88] Jack J. Dongarra, Jeremy Du Croz, Sven Hammarling, and
Richard J. Hanson. An extended set of FORTRAN basic linear algebra subprograms. ACM Transactions on Mathematical
Software, 14(1):1–17, March 1988. doi:10.1145/42288.
42291.
[Eke95]
S. M. Eker. Associative-commutative matching via bipartite graph
matching. The Computer Journal, 38(5):381–399, May 1995.
doi:10.1093/comjnl/38.5.381.
[FM94]
K. Fukuda and T. Matsui. Finding all the perfect matchings
in bipartite graphs. Applied Mathematics Letters, 7(1):15–18,
January 1994. doi:10.1016/0893-9659(94)90045-0.
[FM01]
Fabrice Le Fessant and Luc Maranget. Optimizing pattern
matching. In Proceedings of the Sixth ACM SIGPLAN International Conference on Functional Programming - ICFP ’01.
Association for Computing Machinery (ACM), 2001. doi:
10.1145/507635.507641.
[Grä91]
Albert Gräf. Left-to-right tree pattern matching. In Proceedings
of the 4th International Conference on Rewriting Techniques and
Applications, RTA-91, pages 323–334, New York, NY, USA,
1991. Springer-Verlag New York, Inc.
7
[Hao14]
Li Haoyi. Macropy. https://github.com/lihaoyi/macropy, 2014.
URL: https://github.com/lihaoyi/macropy.
[HK73]
John E. Hopcroft and Richard M. Karp. An n5/2 algorithm
for maximum matchings in bipartite graphs. SIAM Journal on
Computing, 2(4):225–231, December 1973. doi:10.1137/
0202019.
[Jen15]
Grant Jenks. Pypatt. https://pypi.python.org/pypi/pypatt, 2015.
URL: https://pypi.python.org/pypi/pypatt.
[KL91]
E. Kounalis and D. Lugiez. Compilation of pattern matching with
associative-commutative functions. In TAPSOFT ’91, pages 57–
73. Springer Nature, 1991. doi:10.1007/3-540-539824_4.
[KM01]
Hélène Kirchner and Pierre-Etienne Moreau. Promoting rewriting
to a programming language: A compiler for non-deterministic
rewrite programs in associative-commutative theories. International Journal of Foundations of Computer Science, 11(2):207–
251, March 2001. doi:10.1142/S0129054101000412.
[Kre17a] Manuel
Krebber.
Matchpy
documentation.
http://matchpy.readthedocs.io/,
06
2017.
URL:
http://matchpy.readthedocs.io/.
[Kre17b] Manuel Krebber. Non-linear associative-commutative many-toone pattern matching with sequence variables. Master’s thesis,
2017. arXiv:1705.00907.
[Kut06]
Temur Kutsia. Context sequence matching for XML. Electronic
Notes in Theoretical Computer Science, 157(2):47–65, May 2006.
doi:10.1016/j.entcs.2005.12.045.
[Lam88]
J. L. Lambert. Finding a partial solution to a linear system of
equations in positive integers. Computers & Mathematics with
Applications, 15(3):209–212, 1988. doi:10.1016/08981221(88)90172-1.
[LHKK79] C. L. Lawson, R. J. Hanson, D. R. Kincaid, and F. T. Krogh. Basic
linear algebra subprograms for fortran usage. ACM Transactions
on Mathematical Software, 5(3):308–323, September 1979. doi:
10.1145/355841.355847.
[LM94]
D. Lugiez and J. L. Moysset. Tree automata help one to solve
equational formulae in AC-theories. Journal of Symbolic Computation, 18(4):297–318, October 1994. doi:10.1006/jsco.
1994.1049.
[Mar08]
Luc Maranget. Compiling pattern matching to good decision trees.
In Proceedings of the 2008 ACM SIGPLAN Workshop on ML
- ML ’08. Association for Computing Machinery (ACM), 2008.
doi:10.1145/1411304.1411311.
[MRV03] Pierre-Etienne Moreau, Christophe Ringeissen, and Marian Vittek. A Pattern Matching Compiler for Multiple Target Languages,
pages 61–76. 2003. doi:10.1007/3-540-36579-6_5.
[MSP+ 17] Aaron Meurer, Christopher P. Smith, Mateusz Paprocki, Ondřej
Čertík, Sergey B. Kirpichev, Matthew Rocklin, AMiT Kumar,
Sergiu Ivanov, Jason K. Moore, Sartaj Singh, Thilina Rathnayake, Sean Vig, Brian E. Granger, Richard P. Muller, Francesco
Bonazzi, Harsh Gupta, Shivam Vats, Fredrik Johansson, Fabian
Pedregosa, Matthew J. Curry, Andy R. Terrel, Štěpán Roučka,
Ashutosh Saboo, Isuru Fernando, Sumith Kulal, Robert Cimrman, and Anthony Scopatz. Sympy: Symbolic computing in
python. Peerj Computer Science, 3:e103, January 2017. doi:
10.7717/peerj-cs.103.
[MVO96] Alfred J. Menezes, Scott A. Vanstone, and Paul C. Van Oorschot.
Handbook of Applied Cryptography. CRC Press, Inc., Boca
Raton, FL, USA, 1st edition, 1996.
[NWE97] Nadia Nedjah, Colin D. Walter, and Stephen E. Eldridge. Optimal
Left-to-right Pattern-matching Automata, pages 273–286. 1997.
doi:10.1007/BFb0027016.
[Pö16]
Jan Pöschko. Mathics. http://mathics.github.io/, October 2016.
URL: http://mathics.github.io/.
[RDS17] Jonathan Robie, Michael Dyck, and Josh Spiegel. XML path
language (XPath) 3.1. W3C recommendation, W3C, March 2017.
[RJE17]
Florian Rivoal, Tab Atkins Jr., and Elika Etemad. CSS snapshot
2017. W3C note, W3C, January 2017.
[Sch14]
Alexander
Schepanovski.
patterns.
https://github.com/Suor/patterns, 2014.
URL: https:
//github.com/Suor/patterns.
[Uno97]
Takeaki Uno. Algorithms for enumerating all perfect, maximum
and maximal matchings in bipartite graphs. In Algorithms and
Computation, pages 92–101. Springer Nature, 1997. doi:10.
1007/3-540-63890-3_11.
[Wei60]
Robert Weinstock. Greatest common divisor of several integers and an associated linear diophantine equation. The Amer-
8
[WLØ09]
[WR16]
PROC. OF THE 15th PYTHON IN SCIENCE CONF. (SCIPY 2017)
ican Mathematical Monthly, 67(7):664, August 1960. doi:
10.2307/2310105.
Ilmar M. Wilbers, Hans Petter Langtangen, and Åsmund Ødegård.
Using cython to speed up numerical python programs. In Proceedings of Mekit’09, pages 495–512, 2009.
Inc. Wolfram Research. Mathematica, 2016.
URL: https://www.wolfram.com.
| 6cs.PL
|
The One Hundred Layers Tiramisu:
Fully Convolutional DenseNets for Semantic Segmentation
arXiv:1611.09326v3 [cs.CV] 31 Oct 2017
Simon Jégou1 Michal Drozdzal2,3 David Vazquez1,4 Adriana Romero1 Yoshua Bengio1
1
Montreal Institute for Learning Algorithms 2 École Polytechnique de Montréal
3
Imagia Inc., Montréal, 4 Computer Vision Center, Barcelona
[email protected], [email protected], [email protected],
[email protected], [email protected]
Abstract
State-of-the-art approaches for semantic image segmentation are built on Convolutional Neural Networks (CNNs).
The typical segmentation architecture is composed of (a)
a downsampling path responsible for extracting coarse semantic features, followed by (b) an upsampling path trained
to recover the input image resolution at the output of the
model and, optionally, (c) a post-processing module (e.g.
Conditional Random Fields) to refine the model predictions.
Recently, a new CNN architecture, Densely Connected
Convolutional Networks (DenseNets), has shown excellent
results on image classification tasks. The idea of DenseNets
is based on the observation that if each layer is directly connected to every other layer in a feed-forward fashion then
the network will be more accurate and easier to train.
In this paper, we extend DenseNets to deal with the problem of semantic segmentation. We achieve state-of-the-art
results on urban scene benchmark datasets such as CamVid
and Gatech, without any further post-processing module
nor pretraining. Moreover, due to smart construction of
the model, our approach has much less parameters than
currently published best entries for these datasets. Code
to reproduce the experiments is publicly available here :
https://github.com/SimJeg/FC-DenseNet
Figure 1. Diagram of our architecture for semantic segmentation.
Our architecture is built from dense blocks. The diagram is composed of a downsampling path with 2 Transitions Down (TD) and
an upsampling path with 2 Transitions Up (TU). A circle represents concatenation and arrows represent connectivity patterns in
the network. Gray horizontal arrows represent skip connections,
the feature maps from the downsampling path are concatenated
with the corresponding feature maps in the upsampling path. Note
that the connectivity pattern in the upsampling and the downsampling paths are different. In the downsampling path, the input to
a dense block is concatenated with its output, leading to a linear
growth of the number of feature maps, whereas in the upsampling
path, it is not.
1. Introduction
Convolutional Neural Networks (CNNs) are driving major advances in many computer vision tasks, such as image classification [29], object detection [25, 24] and semantic image segmentation [20]. The last few years have witnessed outstanding improvements on CNN-based models.
Very deep architectures [29, 11, 31] have shown impressive results on standard benchmarks such as ImageNet [6]
or MSCOCO [19]. State-of-the-art CNNs heavily reduce
the input resolution through successive pooling layers and,
1
thus, are well suited for applications where a single prediction per input image is expected (e.g. image classification
task).
Fully Convolutional Networks (FCNs) [20, 27] were introduced in the literature as a natural extension of CNNs to
tackle per pixel prediction problems such as semantic image segmentation. FCNs add upsampling layers to standard
CNNs to recover the spatial resolution of the input at the
output layer. As a consequence, FCNs can process images
of arbitrary size. In order to compensate for the resolution loss induced by pooling layers, FCNs introduce skip
connections between their downsampling and upsampling
paths. Skip connections help the upsampling path recover
fine-grained information from the downsampling layers.
Among CNN architectures extended as FCNs for semantic segmentation purposes, Residual Networks (ResNets)
[11] make an interesting case. ResNets are designed to
ease the training of very deep networks (of hundreds of
layers) by introducing a residual block that sums two signals: a non-linear transformation of the input and its identity
mapping. The identity mapping is implemented by means
of a shortcut connection. ResNets have been extended to
work as FCNs [4, 8] yielding very good results in different segmentation benchmarks. ResNets incorporate additional paths to FCN (shortcut paths) and, thus, increase the
number of connections within a segmentation network. This
additional shortcut paths have been shown not only to improve the segmentation accuracy but also to help the network optimization process, resulting in faster convergence
of the training [8].
Recently, a new CNN architecture, called DenseNet, was
introduced in [13]. DenseNets are built from dense blocks
and pooling operations, where each dense block is an iterative concatenation of previous feature maps. This architecture can be seen as an extension of ResNets [11], which performs iterative summation of previous feature maps. However, this small modification has some interesting implications: (1) parameter efficiency, DenseNets are more efficient in the parameter usage; (2) implicit deep supervision,
DenseNets perform deep supervision thanks to short paths
to all feature maps in the architecture (similar to Deeply
Supervised Networks [18]); and (3) feature reuse, all layers can easily access their preceding layers making it easy
to reuse the information from previously computed feature
maps. The characteristics of DenseNets make them a very
good fit for semantic segmentation as they naturally induce
skip connections and multi-scale supervision.
In this paper, we extend DenseNets to work as FCNs by
adding an upsampling path to recover the full input resolution. Naively building an upsampling path would result
in a computationally intractable number of feature maps
with very high resolution prior to the softmax layer. This
is because one would multiply the high resolution feature
maps with a large number of input filters (from all the layers below), resulting in both very large amount of computation and number of parameters. In order to mitigate this
effect, we only upsample the feature maps created by the
preceding dense block. Doing so allows to have a number of
dense blocks at each resolution of the upsampling path independent of the number of pooling layers. Moreover, given
the network architecture, the upsampled dense block combines the information contained in the other dense blocks
of the same resolution. The higher resolution information
is passed by means of a standard skip connection between
the downsampling and the upsampling paths. The details of
the proposed architecture are shown in Figure 1. We evaluate our model on two challenging benchmarks for urban
scene understanding, Camvid [2] and Gatech [22], and confirm the potential of DenseNets for semantic segmentation
by improving the state-of-the-art.
Thus, the contributions of the paper can be summarized
as follows:
• We carefully extend the DenseNet architecture [13] to
fully convolutional networks for semantic segmentation, while mitigating the feature map explosion.
• We highlight that the proposed upsampling path, built
from dense blocks, performs better than upsampling
path with more standard operations, such as the ones
in [27].
• We show that such a network can outperform current
state-of-the-art results on standard benchmarks for urban scene understanding without neither using pretrained parameters nor any further post-processing.
2. Related Work
Recent advances in semantic segmentation have been devoted to improve architectural designs by (1) improving
the upsampling path and increasing the connectivity within
FCNs [27, 1, 21, 8]; (2) introducing modules to account for
broader context understanding [36, 5, 37]; and/or (3) endowing FCN architectures with the ability to provide structured outputs [16, 5, 38].
First, different alternatives have been proposed in the literature to address the resolution recovery in FCN’s upsampling path; from simple bilinear interpolation [10, 20, 1]
to more sophisticated operators such as unpooling [1, 21]
or transposed convolutions [20]. Skip connections from
the downsampling to the upsampling path have also been
adopted to allow for a finer information recovery [27]. More
recently, [8] presented a thorough analysis on the combination of identity mapping [11] and long skip connections [27]
for semantic segmentation.
Second, approaches that introduce larger context to semantic segmentation networks include [10, 36, 5, 37]. In
3. Fully Convolutional DenseNets
As mentioned in Section 1, FCNs are built from a downsampling path, an upsampling path and skip connections.
Skip connections help the upsampling path recover spatially detailed information from the downsampling path, by
reusing features maps. The goal of our model is to further
exploit the feature reuse by extending the more sophisticated DenseNet architecture, while avoiding the feature explosion at the upsampling path of the network.
In this section, we detail the proposed model for semantic segmentation. First, we review the recently proposed
DenseNet architecture. Second, we introduce the construction of the novel upsampling path and discuss its advantages
w.r.t. a naive DenseNet extension and more classical architectures. Finally, we wrap up with the details of the main
architecture used in Section 4.
3.1. Review of DenseNets
Figure 2. Diagram of a dense block of 4 layers. A first layer is applied to the input to create k feature maps, which are concatenated
to the input. A second layer is then applied to create another k
features maps, which are again concatenated to the previous feature maps. The operation is repeated 4 times. The output of the
block is the concatenation of the outputs of the 4 layers, and thus
contains 4 ∗ k feature maps
[10], an unsupervised global image descriptor is computed
added to the feature maps for each pixel. In [36], Recurrent Neural Networks (RNNs) are used to retrieve contextual information by sweeping the image horizontally and
vertically in both directions. In [5], dilated convolutions
are introduced as an alternative to late CNN pooling layers
to capture larger context without reducing the image resolution. Following the same spirit, [37] propose to provide
FCNs with a context module built as a stack of dilated convolutional layers to enlarge the field of view of the network.
Third, Conditional Random Fields (CRF) have long been
a popular choice to enforce structure consistency to segmentation outputs. More recently, fully connected CRFs
[16] have been used to include structural properties of the
output of FCNs [5]. Interestingly, in [38], RNN have been
introduced to approximate mean-field iterations of CRF optimization, allowing for an end-to-end training of both the
FCN and the RNN.
Finally, it is worth noting that current state-of-the-art
FCN architectures for semantic segmentation often rely on
pre-trained models (e.g. VGG [29] or ResNet101 [11]) to
improve their segmentation results [20, 1, 4].
Let x` be the output of the `th layer. In a standard CNN,
x` is computed by applying a non-linear transformation H`
to the output of the previous layer x`−1
x` = H` (x`−1 ),
(1)
where H is commonly defined as a convolution followed by
a rectifier non-linearity (ReLU) and often dropout [30].
In order to ease the training of very deep networks,
ResNets [11] introduce a residual block that sums the identity mapping of the input to the output of a layer. The resulting output x` becomes
x` = H` (x`−1 ) + x`−1 ,
(2)
allowing for the reuse of features and permitting the gradient to flow directly to earlier layers. In this case, H is
defined as the repetition (2 or 3 times) of a block composed
of Batch Normalization (BN) [14], followed by ReLU and
a convolution.
Pushing this idea further, DenseNets [13] design a more
sophisticated connectivity pattern that iteratively concatenates all feature outputs in a feedforward fashion. Thus, the
output of the `th layer is defined as
x` = H` ([x`−1 , x`−2 , ..., x0 ]),
(3)
where [ ... ] represents the concatenation operation. In this
case, H is defined as BN, followed by ReLU, a convolution
and dropout. Such connectivity pattern strongly encourages
the reuse of features and makes all layers in the architecture receive direct supervision signal. The output dimension
of each layer ` has k feature maps where k, hereafter referred as to growth rate parameter, is typically set to a small
value (e.g. k = 12). Thus, the number of feature maps in
DenseNets grows linearly with the depth (e.g. after ` layers,
the input [x`−1 , x`−2 , ..., x0 ] will have ` × k feature maps).
A transition down is introduced to reduce the spatial dimensionality of the feature maps. Such transformation is
composed of a 1×1 convolution (which conserves the number of feature maps) followed by a 2 × 2 pooling operation.
In the remainder of the article, we will call dense block
the concatenation of the new feature maps created at a given
resolution. Figure 2 shows an example of dense block construction. Starting from an input x0 (input image or output
of a transition down) with m feature maps, the first layer of
the block generates an output x1 of dimension k by applying
H1 (x0 ). These k feature maps are then stacked to the previous m feature maps by concatenation ([x1 , x0 ]) and used
as input to the second layer. The same operation is repeated
n times, leading to a new dense block with n × k feature
maps.
3.2. From DenseNets to Fully Convolutional
DenseNets
The DenseNet architecture described in Subsection 3.1
constitutes the downsampling path of our Fully Convolutional DenseNet (FC-DenseNet). Note that, in the downsampling path, the linear growth in the number of features
is compensated by the reduction in spatial resolution of each
feature map after the pooling operation. The last layer of the
downsampling path is referred to as bottleneck.
In order to recover the input spatial resolution, FCNs introduce an upsampling path composed of convolution, upsampling operations (transposed convolutions or unpooling
operations) and skip connections. In FC-DenseNets, we
substitute the convolution operation by a dense block and
an upsampling operation referred to as transition up. Transition up modules consist of a transposed convolution that
upsamples the previous feature maps. The upsampled feature maps are then concatenated to the ones coming from
the skip connection to form the input of a new dense block.
Since the upsampling path increases the feature maps spatial resolution, the linear growth in the number of features
would be too memory demanding, especially for the full
resolution features in the pre-softmax layer.
In order to overcome this limitation, the input of a dense
block is not concatenated with its output. Thus, the transposed convolution is applied only to the feature maps obtained by the last dense block and not to all feature maps
concatenated so far. The last dense block summarizes the
information contained in all the previous dense blocks at
the same resolution. Note that some information from earlier dense blocks is lost in the transition down due to the
pooling operation. Nevertheless, this information is available in the downsampling path of the network and can be
passed via skip connections. Hence, the dense blocks of the
upsampling path are computed using all the available feature maps at a given resolution. Figure 1 illustrates this idea
in detail.
Therefore, our upsampling path approach allows us to
build very deep FC-DenseNets without a feature map explosion. An alternative way of implementing the upsampling
path would be to perform consecutive transposed convolutions and complement them with skip connections from the
downsampling path in a U-Net [27] or FCN-like [20] fashion. This will be further discussed in Section 4
3.3. Semantic Segmentation Architecture
In this subsection, we detail the main architecture, FCDenseNet103, used in Section 4.
First, in Table 1, we define the dense block layer, transition down and transition up of the architecture. Dense
block layers are composed of BN, followed by ReLU, a
3 × 3 same convolution (no resolution loss) and dropout
with probability p = 0.2. The growth rate of the layer is set
to k = 16. Transition down is composed of BN, followed
by ReLU, a 1 × 1 convolution, dropout with p = 0.2 and a
non-overlapping max pooling of size 2 × 2. Transition up
is composed of a 3 × 3 transposed convolution with stride 2
to compensate for the pooling operation.
Second, in Table 2, we summarize all Dense103 layers.
This architecture is built from 103 convolutional layers : a
first one on the input, 38 in the downsampling path, 15 in
the bottleneck and 38 in the upsampling path. We use 5
Transition Down (TD), each one containing one extra convolution, and 5 Transition Up (TU), each one containing a
transposed convolution. The final layer in the network is a
1 × 1 convolution followed by a softmax non-linearity to
provide the per class distribution at each pixel.
It is worth noting that, as discussed in Subsection 3.2, the
proposed upsampling path properly mitigates the DenseNet
feature map explosion, leading to reasonable pre-softmax
feature map number of 256.
Finally, the model is trained by minimizing the pixelwise cross-entropy loss.
4. Experiments
We evaluate our method on two urban scene understanding datasets: CamVid [2], and Gatech [22]. We trained our
models from scratch without using any extra-data nor postprocessing module. We report the results using the Intersection over Union (IoU) metric and the global accuracy
(pixel-wise accuracy on the dataset). For a given class c,
predictions (oi ) and targets (yi ), the IoU is defined by
P
(oi == c ∧ yi == c)
,
IoU (c) = Pi
i (oi == c ∨ yi == c)
(4)
where ∧ is a logical and operation, while ∨ is a logical or
operation. We compute IoU by summing over all the pixels
i of the dataset.
Layer
Batch Normalization
ReLU
3 × 3 Convolution
Dropout p = 0.2
Transition Down (TD)
Batch Normalization
ReLU
1 × 1 Convolution
Dropout p = 0.2
2 × 2 Max Pooling
Transition Up (TU)
3 × 3 Transposed Convolution
stride = 2
Table 1. Building blocks of fully convolutional DenseNets. From left to right: layer used in the model, Transition Down (TD) and Transition
Up (TU). See text for details.
4.1. Architecture and training details
We initialize our models using HeUniform [12] and train
them with RMSprop [33], with an initial learning rate of
1e − 3 and an exponential decay of 0.995 after each epoch.
All models are trained on data augmented with random
crops and vertical flips. For all experiments, we finetune
our models with full size images and learning rate of 1e − 4.
We use validation set to earlystop the training and the finetuning. We monitor mean IoU or mean accuracy and use
patience of 100 (50 during finetuning).
We regularized our models with a weight decay of 1e−4
and a dropout rate of 0.2. For batch normalization, we use
current batch statistics at training, validation and test time.
Architecture
Input, m = 3
3 × 3 Convolution, m = 48
DB (4 layers) + TD, m = 112
DB (5 layers) + TD, m = 192
DB (7 layers) + TD, m = 304
DB (10 layers) + TD, m = 464
DB (12 layers) + TD, m = 656
DB (15 layers), m = 896
TU + DB (12 layers), m = 1088
TU + DB (10 layers), m = 816
TU + DB (7 layers), m = 578
TU + DB (5 layers), m = 384
TU + DB (4 layers), m = 256
1 × 1 Convolution, m = c
Softmax
Table 2. Architecture details of FC-DenseNet103 model used in
our experiments. This model is built from 103 convolutional layers. In the Table we use following notations: DB stands for Dense
Block, TD stands for Transition Down, TU stands for Transition
Up, BN stands for Batch Normalization and m corresponds to the
total number of feature maps at the end of a block. c stands for the
number of classes.
4.2. CamVid dataset
CamVid1 [2] is a dataset of fully segmented videos for
urban scene understanding. We used the split and image
resolution from [1], which consists of 367 frames for training, 101 frames for validation and 233 frames for test. Each
frame has a size 360 × 480 and its pixels are labeled with
11 semantic classes. Our models were trained with crops
of 224 × 224 and batch size 3. At the end, the model is
finetuned with full size images.
In Table 3, we report our results for three networks with
respectively (1) 56 layers (FC-DenseNet56), with 4 layers
per dense block and a growth rate of 12; (2) 67 layers (FCDenseNet67) with 5 layers per dense block and a growth
rate of 16; and (3) 103 layers (FC-DenseNet103) with a
growth rate k = 16 (see Table 2 for details). We also trained
an architecture using standard convolutions in the upsampling path instead of dense blocks (Classic Upsampling). In
the latter architecture, we used 3 convolutions per resolution
level with respectively 512, 256, 128, 128 and 64 filters, as
in [27]. Results show clear superiority of the proposed upsampling path w.r.t. the classic one, consistently improving the IoU significantly for all classes. Particularly, we
observe that unrepresented classes benefit notably from the
FC-DenseNet architecture, namely sign, pedestrian, fence,
cyclist experience a crucial boost in performance (ranging
from 15% to 25%).
As expected, when comparing FC-DenseNet56 or FCDenseNet67 to FC-DenseNet103, we see that the model
benefits from having more depth as well as more parameters.
When compared to other methods, we show that FCDenseNet architectures achieve state-of-the-art, improving
upon models with 10 times more parameters. It is worth
mentioning that our small model FC-DenseNet56 already
outperforms popular architectures with at least 100 times
more parameters.
It is worth noting that images in CamVid correspond to
video frames and, thus, the dataset contains temporal information. Some state-of-the-art methods such as [17] incorporate long range spatio-temporal regularization to the out1 http://mi.eng.cam.ac.uk/research/projects/VideoRec/CamVid/
put of a FCN to boost their performance. Our model is able
to outperform such state-of-the-art model, without requiring any temporal smoothing. However, any post-processing
temporal regularization is complementary to our approach
and could bring additional improvements.
Unlike most of the current state-of-the-art methods, FCDenseNets have not been pretrained on large datasets such
as ImageNet [6] and could most likely benefit from such
pretraining. More recently, it has been shown that deep networks can also boost their performance when pretrained on
data other than natural images, such as video games [26, 28]
or clipart [3], and this an interesting direction to explore.
Figure 3 shows some qualitative segmentation results on
the CamVid dataset. Qualitative results are well aligned
with the quantitative ones, showing sharp segmentations
that account for a lot of details. For example, trees, column
poles, sidewalk and pedestrians appear very well sketched.
Among common errors, we find that thin details found in
trees can be confused with column poles (see fifth row),
buses and trucks can be confused with buildings (fourth
row), and shop signs can be confused with road signs (second row).
4.3. Gatech dataset
Gatech2 [23] is a geometric scene understanding dataset,
which consists of 63 videos for training/validation and 38
for testing. Each video has between 50 and 300 frames
(with an average of 190). A pixel-wise segmentation map is
provided for each frame. There are 8 classes in the dataset:
sky, ground, buildings, porous (mainly trees), humans, cars,
vertical mix and main mix. The dataset was originally built
to learn 3D geometric structure of outdoor video scenes and
the standard metric for this dataset is mean global accuracy.
We used the FC-DenseNet103 model pretrained on
CamVid, removed the softmax layer, and finetuned it for
10 epochs with crops of 224 × 224 and batch size 5. Given
the high redundancy in Gatech frames, we used only one
out of 10 frames to train the model and tested it on all full
resolution test set frames.
In Table 4, we report the obtained results. We compare
the results to the recently proposed method for video segmentation of [34], which reports results of their architecture
with 2D and 3D convolutions. Frame-based 2D convolutions do not have temporal information. As it can be seen
in Table 4, our method gives an impressive improvement of
23.7% in global accuracy with respect to previously published state-of-the-art with 2D convolutions. Moreover, our
model (trained with only 2D convolutions) also achieves a
significant improvement over state-of-the-art models based
on spatio-temporal 3D convolutions (3.4% improvement).
2 http://www.cc.gatech.edu/cpl/projects/videogeometriccontext/
5. Discussion
Our fully convolutional DenseNet implicitly inherits
the advantages of DenseNets, namely: (1) parameter efficiency, as our network has substantially less parameters than other segmentation architectures published for the
Camvid dataset; (2) implicit deep supervision, we tried including additional levels of supervision to different layers of
our network without noticeable change in performance; and
(3) feature reuse, as all layers can easily access their preceding layers not only due to the iterative concatenation of
feature maps in a dense block but also thanks to skip connections that enforce connectivity between downsampling
and upsampling path.
Recent evidence suggest that ResNets behave like ensemble of relatively shallow networks [35]: ”Residual networks avoid the vanishing gradient problem by introducing
short paths which can carry gradient throughout the extent
of very deep networks”. It would be interesting to revisit
this finding in the context of fully convolutional DenseNets.
Due to iterative feature map concatenation in the dense
block, the gradients are forced to be passed through networks of different depth (with different numbers of nonlinearities). Thus, thanks to the smart connectivity patterns, FC-DenseNets might represent an ensemble of variable depth networks. This particular ensemble behavior
would be very interesting for semantic segmentation models, where the ensemble of different paths throughout the
model would capture the multi-scale appearance of objects
in urban scene.
6. Conclusion
In this paper, we have extended DenseNets and made
them fully convolutional to tackle the problem semantic image segmentation. The main idea behind DenseNets is captured in dense blocks that perform iterative concatenation
of feature maps. We designed an upsampling path mitigating the linear growth of feature maps that would appear in
a naive extension of DenseNets.
The resulting network is very deep (from 56 to 103 layers) and has very few parameters, about 10 fold reduction
w.r.t. state-of-the-art models. Moreover, it improves stateof-the-art performance on challenging urban scene understanding datasets (CamVid and Gatech), without neither additional post-processing, pretraining, nor including temporal information.
Aknowledgements
The authors would like to thank the developers of
Theano [32] and Lasagne [7]. Special thanks to Frédéric
Bastien for his work assessing the compilation issues.
Thanks to Francesco Visin for his well designed dataloader [9], as well as Harm de Vries for his support
76.1
82.2
84.0
85.6
Classic Upsampling
FC-DenseNet56 (k=12)
FC-DenseNet67 (k=16)
FC-DenseNet103 (k=16)
7
7
7
7
20
1.5
3.5
9.4
73.5
77.6
80.2
83.0
72.2
72.0
75.4
77.3
92.4
92.4
93.0
93.0
66.2
73.2
78.2
77.3
Global accuracy
88.7
89.0
89.0
91.3
Mean IoU
71.0
74.6
76.2
77.2
31.0
50.1
55.5
57.2
62.5
86.9
85.9
88.7
88.0
−
79.0
88.3
25.1
31.1
52.1
50.5
55.2
58.9
65.8
66.9
86.8
88.9
90.8
91.5
Cyclist
77.8
81.5
82.6
84.0
46.4
63.1
48.9
58.8
57.0
61.6
65.3
66.1
Sidewalk
58.5
Pole
87.0
Fence
Car
52.0
Pedestrian
Sky
68.7
Road
Tree
29.5
29.5
252
32.3
134.5
37.3
140.8
140.8
Sign
!
!
!
!
!
!
!
!
Building
# parameters (M)
Pretrained
SegNet [1]
Bayesian SegNet [15]
DeconvNet [21]
Visin et al. [36]
FCN8 [20]
DeepLab-LFOV [5]
Dilation8 [37]
Dilation8 + FSO [17]
Model
13.4
25.3
17.9
16.0
60.5
24.8
32.7
42.3
46.9
49.9
86.2
n/a
n/a
n/a
91.2
92.2
92.2
92.5
41.7
48.4
56.3
59.1
24.4
27.2
35.8
37.6
19.9
14.3
23.4
16.9
72.7
75.4
75.3
76.0
26.9
31.8
40.9
43.9
90.0
92.8
94.7
94.5
37.7
37.9
58.4
59.6
22.7
26.2
30.7
37.1
30.8
32.6
38.4
37.8
69.6
79.9
81.9
82.2
Table 3. Results on CamVid dataset. Note that we trained our own pretrained FCN8 model
Acc.
Model
2D models (no time)
2D-V2V-from scratch [34]
55.7
FC-DenseNet103
79.4
3D models (incorporate time)
3D-V2V-from scratch [34]
66.7
3D-V2V-pretrained [34]
76.0
[6]
[7]
Table 4. Results on Gatech dataset
[8]
in network parallelization, and Tristan Sylvain. We acknowledge the support of the following agencies for research funding and computing support: Imagia Inc., Spanish projects TRA2014-57088-C2-1-R & 2014-SGR-1506,
TECNIOspring-FP7-ACCI grant.
[9]
References
[1] V. Badrinarayanan, A. Kendall, and R. Cipolla. Segnet: A
deep convolutional encoder-decoder architecture for image
segmentation. CoRR, abs/1511.00561, 2015.
[2] G. J. Brostow, J. Shotton, J. Fauqueur, and R. Cipolla.
Segmentation and recognition using structure from motion
point clouds. In European Conference on Computer Vision
(ECCV), 2008.
[3] L. Castrejon, Y. Aytar, C. Vondrick, H. Pirsiavash, and
A. Torralba. Learning aligned cross-modal representations
from weakly aligned data. CoRR, abs/1607.07295, 2016.
[4] L. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L.
Yuille. Deeplab: Semantic image segmentation with deep
convolutional nets, atrous convolution, and fully connected
crfs. CoRR, abs/1606.00915, 2016.
[5] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and
A. L. Yuille. Semantic image segmentation with deep con-
[10]
[11]
[12]
[13]
[14]
[15]
[16]
volutional nets and fully connected crfs. In International
Conference of Learning Representations (ICLR), 2015.
J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei.
ImageNet: A Large-Scale Hierarchical Image Database. In
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009.
S. Dieleman, J. Schlter, C. Raffel, E. Olson, and et al.
Lasagne: First release., Aug. 2015.
M. Drozdzal, E. Vorontsov, G. Chartrand, S. Kadoury, and
C. Pal. The importance of skip connections in biomedical
image segmentation. CoRR, abs/1608.04117, 2016.
A. R. F. Visin. Dataset loaders: a python library to load and
preprocess datasets. https://github.com/fvisin/
dataset_loaders, 2017.
C. Gatta, A. Romero, and J. van de Weijer. Unrolling loopy
top-down semantic feedback in convolutional deep networks.
In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) workshop, 2014.
K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. CoRR, abs/1512.03385, 2015.
K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into
rectifiers: Surpassing human-level performance on imagenet
classification. CoRR, abs/1502.01852, 2015.
G. Huang, Z. Liu, K. Q. Weinberger, and L. van der
Maaten. Densely connected convolutional networks. CoRR,
abs/1608.06993, 2016.
S. Ioffe and C. Szegedy. Batch normalization: Accelerating
deep network training by reducing internal covariate shift.
CoRR, abs/1502.03167, 2015.
A. Kendall, V. Badrinarayanan, and R. Cipolla. Bayesian
segnet: Model uncertainty in deep convolutional encoderdecoder architectures for scene understanding. CoRR,
abs/1511.02680, 2015.
P. Krähenbühl and V. Koltun. Efficient inference in fully
connected crfs with gaussian edge potentials. In Advances in
Neural Information Processing Systems (NIPS). 2011.
Figure 3. Qualitative results on the CamVid test set. Pixels labeled in yellow are void class. Each row represents (from left to right):
original image, original annotation (ground truth) and prediction of our model.
[17] A. Kundu, V. Vineet, and V. Koltun. Feature space optimization for semantic video segmentation. In IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 2016.
[18] C. Lee, S. Xie, P. W. Gallagher, Z. Zhang, and Z. Tu. Deeplysupervised nets. In International Conference on Artificial
Intelligence and Statistics (AISTATS), 2015.
[19] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollr, and C. L. Zitnick. Microsoft coco: Common
objects in context. In European Conference on Computer Vision (ECCV), 2014.
[20] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional
networks for semantic segmentation. In IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 2015.
[21] H. Noh, S. Hong, and B. Han. Learning deconvolution network for semantic segmentation. arXiv preprint
arXiv:1505.04366, 2015.
[22] S. H. Raza, M. Grundmann, and I. Essa. Geometric context from video. IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2013.
[23] S. H. Raza, M. Grundmann, and I. Essa. Geometric context
from video. In IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2013.
[24] J. Redmon, S. K. Divvala, R. B. Girshick, and A. Farhadi.
You only look once: Unified, real-time object detection.
CoRR, abs/1506.02640, 2015.
[25] S. Ren, K. He, R. B. Girshick, and J. Sun. Faster R-CNN:
towards real-time object detection with region proposal networks. CoRR, abs/1506.01497, 2015.
[26] S. R. Richter, V. Vineet, S. Roth, and V. Koltun. Playing
for data: Ground truth from computer games. In European
Conference on Computer Vision (ECCV), 2016.
[27] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and
Computer-Assisted Intervention (MICAI), 2015.
[28] G. Ros, L. Sellart, J. Materzynska, D. Vazquez, and A. M.
Lopez. The synthia dataset: A large collection of synthetic
images for semantic segmentation of urban scenes. In IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR), 2016.
[29] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR,
abs/1409.1556, 2014.
[30] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning
Research, 15:1929–1958, 2014.
[31] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. E. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. CoRR, abs/1409.4842,
2014.
[32] Theano Development Team. Theano: A Python framework
for fast computation of mathematical expressions. arXiv eprints, abs/1605.02688, May 2016.
[33] T. Tieleman and G. Hinton. rmsprop adaptive learning. In
COURSERA: Neural Networks for Machine Learning, 2012.
[34] D. Tran, L. D. Bourdev, R. Fergus, L. Torresani, and
M. Paluri. Deep end2end voxel2voxel prediction. CoRR,
abs/1511.06681, 2015.
[35] A. Veit, M. J. Wilber, and S. J. Belongie. Residual networks
are exponential ensembles of relatively shallow networks.
CoRR, abs/1605.06431, 2016.
[36] F. Visin, M. Ciccone, A. Romero, K. Kastner, K. Cho,
Y. Bengio, M. Matteucci, and A. Courville. Reseg: A recurrent neural network-based model for semantic segmentation. In IEEE Conference on Computer Vision and Pattern
Recognition (CVPR) workshop, 2016.
[37] F. Yu and V. Koltun. Multi-scale context aggregation by dilated convolutions. In International Conference of Learning
Representations (ICLR), 2016.
[38] S. Zheng, S. Jayasumana, B. Romera-Paredes, V. Vineet,
Z. Su, D. Du, C. Huang, and P. Torr. Conditional random
fields as recurrent neural networks. In International Conference on Computer Vision (ICCV), 2015.
| 7cs.IT
|
On the Fine-grained Complexity of One-Dimensional
Dynamic Programming
arXiv:1703.00941v1 [cs.CC] 2 Mar 2017
Marvin Künnemann
Ramamohan Paturi
Stefan Schneider
University of California, San Diego
March 6, 2017
Abstract
In this paper, we investigate the complexity of one-dimensional dynamic programming, or
more specifically, of the Least-Weight Subsequence (LWS) problem: Given a sequence of n data
items together with weights for every pair of the items, the task is to determine a subsequence S
minimizing the total weight of the pairs adjacent in S. A large number of natural problems
can be formulated as LWS problems, yielding obvious Opn2 q-time solutions.
In many interesting instances, the Opn2 q-many weights can be succinctly represented. Yet
except for near-linear time algorithms for some specific special cases, little is known about
when an LWS instantiation admits a subquadratic-time algorithm and when it does not. In
particular, no lower bounds for LWS instantiations have been known before. In an attempt
to remedy this situation, we provide a general approach to study the fine-grained complexity
of succinct instantiations of the LWS problem. In particular, given an LWS instantiation
we identify a highly parallel core problem that is subquadratically equivalent. This provides
either an explanation for the apparent hardness of the problem or an avenue to find improved
algorithms as the case may be.
More specifically, we prove subquadratic equivalences between the following pairs (an LWS
instantiation and the corresponding core problem) of problems: a low-rank version of LWS
and minimum inner product, finding the longest chain of nested boxes and vector domination,
and a coin change problem which is closely related to the knapsack problem and pmin, `qconvolution. Using these equivalences and known SETH-hardness results for some of the
core problems, we deduce tight conditional lower bounds for the corresponding LWS instantiations. We also establish the pmin, `q-convolution-hardness of the knapsack problem. Furthermore, we revisit some of the LWS instantiations which are known to be solvable in near-linear
time and explain their easiness in terms of the easiness of the corresponding core problems.
1
Introduction
Dynamic programming (DP) is one of the most fundamental paradigms for designing algorithms
and a standard topic in textbooks on algorithms. Scientists from various disciplines have developed
DP formulations for basic problems encountered in their applications. However, it is not clear
This research is supported by the Simons Foundation. This research is support ed by NSF grant CCF-1213151
from the Division of Computing and Communication Foundations. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the
National Science Foundation.
1
whether the existing (often simple and straightforward) DP formulations are in fact optimal or
nearly optimal. Our lack of understanding of the optimality of the DP formulations is particularly
unsatisfactory since many of these problems are computational primitives.
Interestingly, there have been recent developments regarding the optimality of standard DP
formulations for some specific problems, namely, conditional lower bounds assuming the Strong
Exponential Time Hypothesis (SETH) [27]. The longest common subsequence (LCS) problem is
one such problem for which almost tight conditional lower bounds have been obtained recently. The
LCS problem is defined as follows: Given two strings x and y of length at most n, compute the
length of the longest string z that is a subsequence of both x and y. The standard DP formulation
for the LCS problem involves computing a two-dimensional table requiring Opn2 q steps. This
algorithm is only slower than the fastest known algorithm due to Masek and Paterson [34] by a
polylogarithmic factor. However, there has been no progress in finding more efficient algorithms
for this problem since the 1980s, which prompted attempts as early as in 1976 [6] to understand
the barriers for efficient algorithms and to prove lower bounds. Unfortunately, there have not
been any nontrivial unconditional lower bounds for this or any other problem in general models
of computation. This state of affairs prompted researchers to consider conditional lower bounds
based on conjectures such as 3-Sum conjecture [19] and more recently based on ETH [28] and
SETH [27]. Researchers have found ETH and SETH to be useful to explain the exact complexity
of several NP-complete problems (see the survey paper [33]). Surprisingly, Ryan Williams [39] has
found a simple reduction from the CNF-SAT problem to the orthogonal vectors problem which
under SETH leads to a matching quadratic lower bound for the orthogonal vectors problem. This
in turn led to a number of conditional lower bound results for problems in P (including LCS and
related problems) under SETH [7, 1, 11, 2, 23]. Also see [37] for a recent survey.
The DP formulation of the LCS problem is perhaps the conceptually simplest example of a twodimensional DP formulation. In the standard formulation, each entry of an n ˆ n table is computed
in constant time. The LCS problem belongs to the class of alignment problems which, for example,
are used to model similarity between gene or protein sequences. Conditional lower bounds have
recently been extended to a number of alignment problems [9, 7, 1, 11, 3].
In contrast, there are many problems for which natural quadratic-time DP formulations compute
a one-dimensional table of length n by spending Opnq-time per entry. In this work, we investigate
the optimality of such DP formulations and obtain new (conditional) lower bounds which match
the complexity of the standard DP formulations.
1-dimensional DP: The Least-Weight Subsequence (LWS) Problem. In this paper, we
investigate the optimality of the standard DP formulation of the LWS problem. A classic example
of an LWS problem is airplane refueling [25]: Given airport locations on a line, and a preferred
distance per hop k (in miles), we define the penalty for flying k 1 miles as pk ´ k 1 q2 . The goal is
then to find a sequence of airports terminating at the last airport that minimizes the sum of the
penalties. We now define the LWS problem formally.
Problem 1.1 (LWS). We are given a sequence of n ` 1 data items x0 , . . . , xn , weights wi,j P
t´W, . . . , W u Y t8u for every pair i ă j of indices where the weights may also be functions of the
values of data items xi , and an arbitrary function g : Z Ñ Z. The LWS problem is to determine
T rns which is defined by the following DP formulation.
T r0s “ 0,
T rjs “ min gpT risq ` wi,j
0ďiăj
2
for j “ 1, . . . , n.
(1)
To formulate airplane refueling as an LWS problem, we let xi be the location of the i’th airport,
g be the identity function, and wi,j “ pxj ´ xi ´ kq2 .
In the definition of the LWS problem, we did not specify the encoding of the problem (in
particular, the type of data items and the representation of the weights wi,j ) so we can capture a
larger variety of problems: it not only encompasses classical problems such as the pretty printing
problem due to Knuth and Plass [31], the airplane refueling problem [25] and the longest increasing
subsequence (LIS) [18], but also the unbounded subset sum problem [36, 10], a more general coin
change problem that is effectively equivalent to the unbounded knapsack problem, 1-dimensional
k-means clustering problem [24], finding longest R-chains (for an arbitrary binary relation R), and
many others (for a more complete list of problems definitions, see Section 2).
Under mild assumptions on the encoding of the data items and weights, any instantiation of the
LWS problems can be solved in time Opn2 q using (1) for determining the values T rjs, j “ 1, . . . , n in
time Opnq each. However, the best known algorithms for the LWS problems differ quite significantly
in their time complexity. Some problems including the pretty printing, airline refueling and LIS
turn out to be solvable in near-linear time, while no subquadratic algorithms are known for the
unbounded knapsack problem or for finding the longest R-chain.
The main goal of the paper is to investigate the optimality of the LWS DP formulation for
various problems by proving conditional lower bounds.
Succinct LWS instantiations. In the extremely long presentation of an LWS problem, the
weights wi,j are given explicitly. This is however not a very interesting case from a computational
point of view, as the standard DP formulation takes linear time (in the size of the input) to
compute T rns. In the example of the airplane refueling problem the size of the input is only Opnq
assuming that the values of the data items are bounded by some polynomial in n. For such succinct
representations, we ask if the quadratic-time algorithm based on the standard LWS DP formulation
is optimal. Our approach is to study several natural succinct versions of the LWS problem (by
specifying the type of data items and the weight function1 ) and determine their complexity. We
refer to Section 2 for examples of succinct instantiations of the LWS problem.
Our Contributions and Results. The main contributions of our paper include a general framework for reducing succinct LWS instantiations to what we call the core problems and proving
subquadratic equivalences between them. The subquadratic equivalences are interesting for two
reasons. First, they allows us to conclude conditional lower bounds for certain LWS instantiations,
where previously no lower bounds are known. Second, subquadratic (or more general fine-grained)
equivalences are more useful since they let us translate hardness as well as easiness results.
Our results include tight (up to subpolynomial factors) conditional lower bounds for several
LWS instantiations with succinct representations. These instantiations include the coin change
problem, low rank versions of the LWS problem, and the longest subchain problems. Our results
are somewhat more general. We propose a factorization of the LWS problem into a core problem
and a fine-grained reduction from the LWS problem to the core problem. The idea is that core
problems (which are often well-know problems) capture the hardness of the LWS problem and act
as a potential barrier for more efficient algorithms. While we do not formally define the notion of
a core problem, we identify several core problems which share several interesting properties. For
example, they do not admit natural DP formulations and are easy to parallelize. In contrast, the
quadratic-time DP formulation of LWS problems requires the entries T ris to be computed in order,
suggesting that the general problem might be inherently sequential.
1 In
all our applications, the function g is the trivial identity function.
3
The reductions between LWS problems and core problems involve a natural intermediate problem, which we call the Static-LWS problem. We first reduce the LWS problem to the Static-LWS
problem in a general way and then reduce the Static-LWS problem to a core problem. The first
reduction is divide-and-conquer in nature and is inherently sequential. The latter reduction is specific to the instantiation of the LWS problem. The Static-LWS problem is easy to parallelize and
does not have a natural DP formulation. However, the problem is not necessarily a natural problem.
The Static-LWS problem can be thought of as a generic core problem, but it is output-intensive.
In the other direction, we show that many of the core problems can be reduced to the corresponding LWS instantiations thus establishing an equivalency between LWS instantiations and their core
problems. This equivalence enables us to translate both the hardness and easiness results (i.e., the
subquadratic-time algorithms) for the core problems to the corresponding LWS instantiations.
The first natural succinct representation of the LWS problem we consider is the low rank
LWS problem, where the weight matrix W “ pwi,j q is of low rank and thus representable as
W “ L ¨ R where L and RT are pn ˆ nop1q q-matrices. For this low rank LWS problem, we identify
the minimum inner product problem (MinInnProd) as a suitable core problem. It is only natural
and not particularly surprising that MinInnProd can be reduced to the low-rank LWS problem
which shows the SETH-hardness of the low-rank LWS problem. The other direction is more
surprising: Inspired by an elegant trick of Vassilevska Williams and Williams [40], we are able to
show a subquadratic-time reduction from the (highly sequential) low-rank LWS problem to the
(highly parallel) MinInnProd problem. Thus, the very compact problem MinInnProd problem
captures exactly the complexity of the low-rank LWS problem (under subquadratic reductions).
We also show that the coin change problem is subquadratically equivalent to the pmin, `qconvolution problem. In the coin change problem, the weight matrix W is succinctly given as a
Toeplitz matrix. At this point, the conditional hardness of the pmin, `q-convolution problem is
unknown. The quadratic-time hardness of the pmin, `q-convolution problem would be very interesting, since it is known that the pmin, `q-convolution problem is reducible to the 3-sum problem
and the APSP problem, However, recent results give surprising subquadratic-time algorithms for
special cases of pmin, `q-convolution [14]. If these subquadratic-time algorithms extend to the
general pmin, `q-convolution problem, our equivalence result also provides a subquadratic-time
algorithm for the coin change problem and the closely related unbounded knapsack problem. As a
corollary, our reductions also give a quadratic-time pmin, `q-convolution-based lower bound for
the bounded case of knapsack.
We next consider the problem of finding longest chains: here, we search for the longest subsequence (chain) in the input sequence such that all adjacent pairs in the subsequence are contained in some binary relation R. We show that for any binary relation R satisfying certain
conditions the chaining problem is subquadratically equivalent to a corresponding (highly parallel) selection problem. As corollaries, we get equivalences between finding the longest chain of
nested boxes (NestedBoxes) and VectorDomination as well as between finding the longest subset chain (SubsetChain) and the orthogonal vectors (OV) problem. Interestingly, these results
have algorithmic implications: known algorithms for low-dimensional vector domination and lowdimensional orthogonal vectors translate to faster algorithms for low-dimensional NestedBoxes
and SubsetChain for small universe size.
Table 1 lists the LWS succinct instantiations (as discussed above) and their corresponding core
problems. All LWS instantiations and core problems considered in this paper are formally defined
in Section 2.
Finally, we revisit classic problems including the longest increasing subsequence problem, the
4
Name
Coin Change
LowRankLWS
R-chains
NestedBoxes
SubsetChain
Weights
Equivalent Core
Reference
Toeplitz matrix:
pmin, `q-convolution Theorem 4.8
wi,j “ wj´i
Remark: Subquadratically equivalent to UnboundedKnapsack
Low rank representation:
MinInnProd
Theorem 3.9
wi,j “ xσi , µj y
matrix induced by R:
SelectionpRq
Theorem 5.3
wi,j “ wj if Rpxi , xj q and 8 o/w
Theorem 5.7
Remark: Result below are corollaries.
wi,j “ ´1 if Bj contains Bi
VectorDomination
wi,j “ ´1 if Si Ď Sj
OrthogonalVectors
Table 1: Summary of our results
Name
Longest Increasing
Subsequence
Unbounded Subset
Sum
Concave 1-dim. DP
Weights
matrix induced by Ră :
wi,j “ ´1 if xi ă xj
Toeplitz t0, 8u matrix:
wi,j “ wj´i P t0, 8u
concave matrix:
wi,j ` wi1 ,j 1 ď wi1 ,j ` wi,j 1
for i ď i1 ď j ď j 1
Õpnq-time reducible to
Sorting
Convolution
SMAWK problem
Reference
[18],
Observation 2
[10],
Observation 3
[25, 21, 38],
Observation 4
Table 2: Near-linear time algorithms following from the proposed framework.
unbounded subset sum problem and the concave LWS problem and analyze the Static-LWS
instantiations to immediately infer that the corresponding core problem can be solved in near-linear
time. Table 2 gives an overview of some of the problems we look at in this context.
Related Work. LWS has been introduced by Hirschberg and Lamore [25]. If the weight function
satisfies the quadrangle inequality 2 formalized by Yao [41], one obtains the concave LWS problem, for
which they give an Opn log nq-time algorithm. Subsequently, improved algorithms solving concave
LWS in time Opnq were given [38, 21]. This yields a fairly large class of weight functions (including,
e.g., the pretty printing and airplane refueling problems) for which linear-time solutions exist. To
generalize this class of problems, further works address convex weight functions3 [20, 35, 30] as well
as certain combinations of convex and concave weight functions [16] and provide near-linear time
algorithms. For a more comprehensive overview over these algorithms and further applications of
the LWS problem, we refer the reader to Eppstein’s PhD thesis [17].
Apart from these notions of concavity and convexity, results on the succinct LWS problems
are typically more scattered and problem-specific (see, e.g., [18, 31, 10, 24]; furthermore, a closely
related recurrence to (1) pops up when solving bitonic TSP [15]). An exception to this rule is a
study of the parallel complexity of LWS [22].
Organization. Section 3 contains the result on low-rank LWS. This is also where we formally
2 See
3A
Section 2 for definitions.
weight function is convex if it satisfies the inverse of the quadrangle inequality.
5
introduce Static-LWS. Section 4 proves the subquadratic equivalence of the coin change problem
and pmin, `q-convolution, while Section 5 discusses chaining problems and their corresponding
selection (core) problem. Our results on near-linear time algorithms are given in Section 6.
2
Preliminaries
In this section, we state our notational conventions and list the main problems considered in this
work.
Problem A subquadratically reduces to problem B, denoted A ď2 B, if for any ε ą 0 there is
a δ ą 0 such that an algorithm for B with time Opn2´ε q implies an algorithm for A with time
Opn2´δ q. We call the two problems subquadratically equivalent, denoted A ”2 B, if there are
subquadratic reductions both ways.
We let rns :“ t1, . . . , nu. When stating running time, we use the notation Õp¨q to hide polylogarithmic factors. For a problem P , we write T P for its time complexity. We generally assume
the word-RAM model of computation with word size w “ Θplog nq. For most problems defined in
this paper, we consider inputs to be integers in the range t´W, . . . , W u where W fits in a constant
number of words4 . For vectors, we use d for the dimension and generally assume d “ nop1q .
Core Problems and Hypotheses. One of the most popular problems in the field of quadratictime conditional hardness is the following problem.
Problem 2.1 (Orthogonal Vectors (OV)). Given a1 , . . . , an , b1 , . . . , bn P t0, 1ud, determine if there
is a pair i, j satisfying xai , bj y “ 0.
Recall that for OV (and the related problems below) we assume d “ nop1q . Thus the naive
algorithm solves OV in time Opn2 ¨ dq “ Opn2`op1q q.
One of the reasons for the popularity of OV is its surprising connection to the Strong Exponential
Time Hypothesis (SETH) [27]: It states that for every ε ą 0 there is a k, such that the k-SAT
problem requires time Ωp2p1´εqn q. By an elegant reduction due to Williams [39], OV is quadratictime SETH-hard, i.e., there is no algorithm with running time time Opn2´ε q for any ε ą 0 unless
SETH is false.
We consider the following generalizations of OV.
Problem 2.2 (MinInnProd). Given a1 , . . . , an , b1 , . . . , bn P t´W, . . . , W ud and a natural number
r P Z, determine if there is a pair i, j satisfying xai , bj y ď r.
Problem 2.3 (AllInnProd). Given a1 , . . . , an P t´W, . . . , W ud and b1 , . . . , bn P t´W, . . . , W ud ,
determine for all j P rns, the value miniPrns xai , bj y.
Problem 2.4 (VectorDomination). Given a1 , . . . , an , b1 , . . . , bn P t´W, . . . , W ud determine if
there is a pair i, j such that ai ď bj component-wise.
Problem 2.5 (SetContainment). Given sets a1 , . . . , an , b1 , . . . , bn Ď rds given as vectors in
t0, 1ud determine if there is a pair i, j such that ai Ď bj .
Note that SetContainment is a special case of VectorDomination and computationally
equivalent to OV, as xa, by “ 0 if and only if a Ď b (in this slight misuse of notation we think of
the Boolean vectors a, b as sets and let b̄ denote the complement of b).
4 For
op1q
the purposes of our reductions, even values up to W “ 2n
6
would be fine.
Since subquadratic solutions to any of these problems trivially give a subquadratic solution to
OV, these problems are also quadratic-time SETH-hard. However, the converse does not necessarily hold. In particular, the strongest currently known upper bounds differ: while for OV and
SetContainment for small dimension d “ c ¨ logpnq, an n2´1{Oplog cq -time algorithm is known [4],
2
for VectorDomination the best known algorithm runs only in time n2´1{Opc log cq [26, 13].
Another fundamental quadratic-time problem is pmin, `q-convolution, defined below.
Problem 2.6 (pmin, `q-convolution). Given vectors a “ pa0 , . . . , an´1 q, b “ pb0 , . . . , bn´1 q P
t´W, . . . , W un , determine its pmin, `q-convolution a ˚ b defined by
pa ˚ bqk “
min
0ďi,jăn:i`j“k
ai ` b j
for all 0 ď k ď 2n ´ 2.
As opposed to the classical convolution, which we denote as a f b, solvable in time Opn log nq
using FFT, no strongly subquadratic algorithm for pmin, `q-convolution is known. Compared to
OV, we have less support for believing that no Opn2´ε q-time algorithm for pmin, `q-convolution
exists. In particular, interesting special cases can be solved in subquadratic-time [14] and there
are subquadratic-time co-nondeterministic and nondeterministic algorithms [8, 12]. At the same
time, breaking this long-standing quadratic-time barrier is a prerequisite for progress on refuting
the 3SUM and APSP conjectures. This makes it an interesting target particularly for proving
subquadratic equivalences, since both positive and negative resolutions of this open question appear
to be reasonable possibilities.
Succinct LWS Versions and Applications. In the definition of LWS (Problem 1.1) we did not
fix the encoding of the problem (in particular, the choice of data items, as well the representation
of the weights wi,j and the function g). Assuming that g can be determined in Õp1q and that
W “ polypnq, this problem can naturally be solved in time Õpn2 q, by evaluating the central
recurrence (1) for each j “ 1, . . . , n – this takes Õpnq time for each j, since we take the minimum
over at most n expressions that can be evaluated in time Õp1q by accessing the previously computed
entries T r0s, . . . , T rj ´ 1s as well as computing g. In all our applications, g will be the identity
function, hence it will suffice to define the type of data items and the corresponding weight matrix.
Throughout this paper, whenever we fix a representation of the weight matrix W “ pwi,j qi,j , we
denote the corresponding problem LWSpWq.
In the remainder of this section, we list problems considered in this paper that can be expressed
as an LWS instantiations. At this point, we typically give the most natural formulations of these
problems – the corresponding definitions as LWS instantiations are given in the corresponding
sections.
We start off with a natural succinct “low-rank” version of LWS.
Problem 2.7 (LowRankLWS). LowRankLWS is the LWS problem where the weight matrix
W is of rank d ! n. The input is given succinctly as two matrices A and B, which are pn ˆ dq- and
pd ˆ nq-matrices respectively, and W “ A ¨ B.
Alternatively, LowRankLWS may be interpreted in the following way: There are places
0, 1, . . . , n, each of which is equipped with an in- and an out-vector. The cost of going from place
i to j is then defined as the inner product of the out-vector of i with the in-vector of j, and the
task is to compute the minimum-cost monotonically increasing path to reach place n starting from
0. In Section 3, we prove subquadratic equivalence to MinInnProd.
We consider the following coin change problem and variations of Knapsack.
7
Problem 2.8 (CC). We are given a weight sequence w “ pw1 , . . . , wn q with wi P t´W, . . . , W u Y
t8u, i.e., the
ř the weight of the multiset of denominations I
ř coin with value i has weight wi . Find
such that iPI i “ n and the sum of the weights iPI wi is minimized.
Problem 2.9 (UnboundedKnapsack). We are given a sequence of profits p “ pp1 , . . . , pn q with
pi P t0, 1, . . . , W u, ř
i.e., the item of size i has profit
ř pi . Find the total profit of the multiset of
indices I such that iPI i ď n and the total profit iPI pi is maximized.
Note that if we replace multiset by set in the above definition, we obtain the bounded version
of the problem, which we denote by Knapsack.
We remark that our perspective on CC and UnboundedKnapsack (as well as UnboundedSubsetSum below) using LWS is slightly different than many classical accounts of Knapsack: We
define the problem size as the budget size instead of the number of items, thus our focus is on
pseudo-polynomial time algorithms for the typical formulations of these problems.
Note that we state the coin change problem as allowing positive or negative weights, but UnboundedKnapsack only allows for positive profits. Furthermore, CC is a minimization problem,
while UnboundedKnapsack is a maximization problem. For CC, the maximization problem is
trivially equivalent as we can negate all weights. Furthermore, we can freely translate the range of
the weights in the coin change problem by defining wi1 “ i ¨ M ` wi for all i and sufficiently large or
small M . The most significant difference between CC and UnboundedKnapsack is that for CC
the indices have to sum to exactly n, while for UnboundedKnapsack n is only an upper bound.
We will encounter an important generalization of the two problems above, defined as follows.
Problem 2.10 (oiCC). The output-intensive version of CC is to determine, given an input to
CC, the weight of the optimal multiset such that the denominations sum up to j for all 1 ď j ď n.
It is easy to see that oiCC is at least as hard as both CC and UnboundedKnapsack. We
will relate the above Knapsack variants to pmin, `q-convolution in Section 4.
In Section 6, we will revisit near-linear time algorithms for the following special case of the coin
change problem.
Problem 2.11 (UnboundedSubsetSum). Given a subset S Ď rns, determine whether there is a
multiset of elements of S that sums up to exactly n.
We also discuss problems where the goal is to find the longest chain among data items, where
the notion of a chain is defined by some binary relation R. We first give the definition of the general
problem which is parameterized by R.
Problem 2.12 (ChainLWS). Fix a set X of objects and a relation R Ď X ˆ X. The Weighted
Chain Least-Weight Subsequence Problem for R, denoted ChainLWSpRq, is the following problem:
Given data items x0 , . . . , xn P X, weights w1 , . . . , wn´1 P t´W, . . . , W u, find the weight of the
increasing sequence i0 “ 0 ă i1 ă i2 ă . . . ă ř
ik “ n such that for all j with 1 ď j ď k the pair
k´1
pxij´1 , xij q is in the relation R and the weight j“1 wij is minimized.
The following problems are specializations of this problem for different relations.
Problem 2.13 (NestedBoxes). Given n boxes in d dimensions, given as non-negative, d-dimensional
vectors pb1 , . . . , bn q, find the longest chain such that each box fits into the next (without rotation).
We say box that box a fits into box b if for all dimensions 1 ď i ď d, ai ď bi .
8
Problem 2.14 (SubsetChain). Given n sets from a universe U of size d, given as Boolean,
d-dimensional vectors pb1 , . . . , bn q, find the longest chain such that each set is a subset of the next.
Note that SubsetChain is a special case of NestedBoxes.
Problem 2.15 (LIS). Given a sequence of n integers x1 , . . . , xn , compute the length of the longest
subsequence that is strictly increasing.
Finally, we will briefly discuss the following class of LWS problems that turn out to be solvable
in near-linear time.
Problem 2.16 (ConcLWS). Given an LWS instance in which the weights satisfy the quadrangle
inequality
for i ď i1 ď j ď j 1 ,
wi,j ` wi1 ,j 1 ď wi1 ,j ` wi,j 1
solve it. The weights are not explicitly given, but each wi,j can be queried in constant time.
3
LowRankLWS
Let us first analyze the following canonical succinct representation of a low-rank weight matrix
W “ pwi,j qi,j : If W is of rank d ! n, we can write it more succinctly as W “ A ¨ B, where A and B
are pn ˆ dq- and pd ˆ nq matrices, respectively. We can express the resulting natural LWS problem
equivalently as follows.
Problem 3.1 (LowRankLWS). We define the following LWS instantiation LowRankLWS “
LWSpWLowRankq.
Data items: out-vectors µ0 , . . . , µn´1 P t´W, . . . , W ud , in-vectors σ1 , . . . , σn P t´W, . . . , W ud
Weights: wpi, jq “ xµi , σj y for 0 ď i ă j ď n
In this section, we show that this problem is equivalent, under subquadratic reductions, to the
following non-sequential problem.
Problem 3.2 (MinInnProd). Given a1 , . . . , an , b1 , . . . , bn P t´W, . . . , W ud and a natural number
r P Z, determine if there is a pair i, j satisfying xai , bj y ď r.
We first give a simple reduction from MinInnProd that along the way proves quadratic-time
SETH-hardness of LowRankLWS.
Lemma 3.3. It holds that T MinInnProdpn, d, W q ď T LowRankLWSp2n ` 1, d ` 2, dW q ` Opndq.
Proof. Given a1 , . . . , an , b1 , . . . , bn P t´W, . . . , W ud , let O “ p0, . . . , 0q P Zd be the all-zeroes vector
and define the following in- and out-vectors
µ0 “ pdW, 0, Oq,
µi “ p0, dW, ai q,
µn`j “ p0, 0, Oq,
σ2n`1 “ pdW, dW, Oq,
σi “ p0, 0, Oq,
σn`j “ pdW, 0, bj q,
for i “ 1, . . . , n,
for j “ 1, . . . , n.
To prove correctness, we show that in the constructed LowRankLWS instance, we have T r2n`1s “
mini,j xai , bj y, from which the results follows immediately. Inductively, we have T ris “ 0 for i “
1, . . . , n, since xµi1 , σi y “ 0 for all 0 ď i1 ă i ď n. Similarly, for j “ 1, . . . , n one can inductively show
9
that T rn ` js “ min1ďiďn,j 1 ďj xai , bj 1 y, using that xµ0 , σn`j y “ pdW q2 ě maxi,j xai , bj y, xµi , σn`j y “
xai , bj y and xµn`j 1 , σn`j y “ 0 for all 1 ď i, j ď n and j 1 ď j. Finally, using (1) xµ0 , σ2n`1 y “
pdW q2 ě maxi,j xai , bj y and T r0s “ 0, (2) xµi , σ2n`1 y “ pdW q2 ě maxi,j xai , bj y and T ris “ 0 for
i “ 1, . . . , n and (3) xµn`j , σ2n`1 y “ 0 and T rn ` js “ min1ďiďn,1ďj 1 ďj xai , bj 1 y for all j “ 1, . . . , n,
we can finally determine T r2n ` 1s “ mini,j xai , bj y.
To prove the other direction, we will give a quite general approach to compute the sequential
LWS problem by reducing to a natural static subproblem of LWS:
Problem 3.4 (Static-LWSpWq). Fix an instance of LWSpWq. Given intervals I :“ ta`1, . . . , a`
N u and J :“ ta`N `1, . . . , a`2N u, together with the correctly computed values T ra`1s, . . . , T ra`
N ` 1s, the Static Least-Weight Subsequence Problem (Static-LWS) asks to determine
T 1 rjs :“ min T ris ` wi,j
for all j P J.
iPI
Lemma 3.5 (LWSpWq ď2 Static-LWSpWq). For any choice of W, if Static-LWSpWq can be
solved in time OpN 2´ε q for some ε ą 0, then LWSpWq can be solved in time Õpn2´ε q.
Proof. In what follows, we fix LWS as LWSpWq and Static-LWS as Static-LWSpWq.
We define the subproblem Spti, . . . , ju, pti , . . . , tj qq that given an interval spanned by 1 ď i ď
j ď n and values tk “ min0ďk1 ăi T rk 1 s ` wk1 ,k for each point k P ti, . . . , ju, computes all values
T rks for k P ti, . . . , ju. Note that a call to Sprns, pw0,1 , . . . , w0,n qq solves the LWS problem, since
T r0s “ 0 and thus the values of tk , k P rns are correctly initialized.
We solve S using Algorithm 1. We briefly argue correctness, using the invariant that tk “
Algorithm 1 Reducing LWS to Static-LWS
1: function S(ti, . . . , ju, pti , . . . , tj q)
2:
if i “ j then
3:
return T ris Ð ti
4:
m Ð r j´i
2 s
5:
pT ris, . . . , T ri ` m ´ 1sq Ð Spti, . . . , i ` m ´ 1u, pti , . . . , ti`m´1 qq
6:
solve Static-LWS on the subinstance given by I :“ ti, . . . , i`m´1u and J :“ ti`m, . . . , i`
2m ´ 1u:
7:
Ź obtains values T 1 rks “ miniďk1 ăi`m T rk 1 s ` wk1 ,k for k “ i ` m, . . . , i ` 2m ´ 1.
1
8:
tk Ð minttk , T 1 rksu for all k “ i ` m, . . . , i ` 2m ´ 1.
9:
pT ri ` ms, . . . , T ri ` 2m ´ 1sq Ð Spti ` m, . . . , i ` 2m ´ 1u, pt1i`m , . . . , t1i`2m´1 qq
10:
if j “ i ` 2m then
11:
T rjs :“ minttj , miniďkăj T rks ` wk,j u.
12:
return pT ris, . . . , T rjsq
min0ďk1 ăi T rk 1 s ` wk1 ,k in every call to S. If S is called with i “ j, then the invariant yields
ti “ min0ďk1 ăi T rk 1 s ` wk1 ,i “ T ris, thus T ris is computed correctly. For the call in Line 5, the
invariant is fulfilled by assumption, hence the values pT ris, . . . , T ri ` m ´ 1sq are correctly computed.
For the call in Line 9, we note that for k “ i ` m, . . . , i ` 2m ´ 1, we have
t1k “ minttk , T 1 rksu “ mint min
T rk 1 s ` wk1 ,k ,
1
0ďk ăi
min
iďk1 ăi`m
10
T rk 1 s ` wk1 ,k u “
min
0ďk1 ăi`m
T rk 1 s ` wk1 ,k .
Hence the invariant remains satisfied. Thus, the values pT ri ` ms, . . . , T ri ` 2m ´ 1sq are correctly
computed. Finally, if j “ i ` 2m, we compute the remaining value T rjs correctly, since tj “
min0ďkăi T rks ` wk,j by assumption.
To analyze the running time T S pnq of S on an interval of length n :“ j ´ i ` 1, note that
each call results in two recursive calls of interval lengths at most n{2. In each call, we need an
additional overhead that is linear in n and T Static-LWS pn{2q. Solving the corresponding recursion
T S pnq ď 2T S pn{2q ` T Static-LWS pn{2q ` Opnq, we obtain that an OpN 2´ε q-time algorithm StaticLWS, with 0 ă ε ă 1 yields T LWSpnq ď T S pnq “ Opn2´ε q. Similarly, an OpN logc N q-time
algorithm for Static-LWS would result in an Opn logc`1 nq-time algorithm for LWS.
For the special case of LowRankLWS, it is straightforward to see that the static version boils
down to the following natural reformulation.
Problem 3.6 (AllInnProd). Given a1 , . . . , an P t´W, . . . , W ud and b1 , . . . , bn P t´W, . . . , W ud ,
determine for all j P rns, the value miniPrns xai , bj y. (Again, we typically assume that d “ nop1q and
op1q
W “ 2n .)
Lemma 3.7 (Static-LWSpWLowRank q ď2 AllInnProd). We have
T Static-LWSpWLowRank q pn, d, W q ď T AllInnProdpn, d ` 1, nW q ` Opndq.
Proof. Consider Static-LWSpWLowRankq. Let I “ ta ` 1, . . . , a ` N u, J “ ta ` N ` 1, . . . , a `
2N u and values T ra ` 1s, . . . , T ra ` N s be given. To determine T 1 rjs “ miniPI T ris ` wi,j for all
j P J, it is sufficient to solve AllInnProd on the vectors aa`1 , . . . , aa`N , ba`N `1 , . . . , ba`2N P
tnW, . . . , nW ud`1 defined by
ai :“ pµi , T risq
bj “ pσj , 1q,
for all i P I, j P J,
since then xai , bj y “ T ris ` xµi , σj y “ T ris ` wi,j . The claim immediately follows (note that
|T ris| ď nW ).
Finally, inspired by an elegant trick of [40], we reduce AllInnProd to MinInnProd.
Lemma 3.8 (AllInnProd ď2 MinInnProd). We have
?
T AllInnProdpn, d, W q ď Opn ¨ T MinInnProdp n, d ` 3, ndW 2 q ¨ log2 nW q.
Proof. We first observe that we can tune MinInnProd to also return a witness pi, jq with xai , bj y ď
r, if it exists. To do so, we replace each ai by the pd` 2q-dimensional vector a1i “ pai ¨n, pi ´ 1qn, ´1q
and similarly, each bj by the pd ` 2q-dimensional vector b1j “ pbj ¨ n, ´1, j ´ 1q. Clearly, we have
xa1i , b1j y “ xai , bj yn2 ´ pi ´ 1qn ´ pj ´ 1q. Thus xa1i , b1j y ď rn2 if and only if xai , bj y ď r since i, j P rns.
Using a binary search over r, we can find mini,j xa1i , b1j y, from whose precise value we can determine
also a witness, if it exists. Thus the running time witpn, d, W q for finding such a witness is bounded
by Oplog nW q ¨ T MinInnProdpn, d ` 2, nW q.
To solve AllInnProd, i.e., to compute pj :“ miniPrns xai , bj y for all j P rns, we employ a
parallel binary search. Consider in particular the following problem P: Given arbitrary r1 , . . . , rn ,
determine for all j P rns whether there exists i P rns
? such that xai , bj y ď rj . We will show below
that this problem can be solved in time Opn ¨ witp n, d ` 1, dW 2 qq. The claim then follows, since
starting from feasible intervals R1 “ ¨ ¨ ¨ “ Rn “ t´dW 2 , . . . , dW 2 u satisfying pj P Rj , we can
11
halve the sizes of each interval simultaneously by a single call to P. Thus, after OplogpdW qq
calls, the true
in the
T AllInnProdpn, d, wq “
? time guarantee
? values pj 2can be determined, resulting
MinInnProd
2
p n, d ` 3, ndW q log2 pnW qq, as desired.
Opn ¨ witp n, d ` 1, dW q ¨ logpdW qq “ Opn ¨ T
We complete the proof of the claim by showing how to solve P. Without loss of generality, we
2
can assume that rj ď dW?
for every j, since no larger inner product
may exist. We group the
?
vectors a1 , . . . , an in g :“ r ns groups A1 , . . . , Ag of size at most n each, and do the same for the
vectors b1 , . . . , bn to obtain B1 , . . . , Bg . Now, we iterate over all pairs of groups Ak , Bℓ , k, ℓ P rgs:
For each such choice of pairs, we do the following process. For each vector ai P Ak , we define
the pd ` 1q-dimensional vector ãi :“ pai , ´1q and for every vector bj P Bℓ , we define b̃j :“ pbj , rj q.
In the obtained instance tãi uaPAk , tb̃j ubPBℓ , we try to find some i, j such that xãi , b̃j y ď 0, which
is equivalent to xai , bj y ď rj . If we succeed in finding such a witness, we delete bj and b˜j (but
remember its witness) and repeat finding witnesses (an deleting the witnessed bj ) until we cannot
find any. The process then ends and we turn to the next pair of groups.
It is easy to see that for all j P rns, we have xai , bj y ď rj for some i P rns if and only if the
above process finds a witness for bj at some point. To argue about the running time, we charge the
running time of every call to witness finding to either (1) the pair Ak , Bℓ , if the call is the first call
in the process for Ak , Bℓ , or (2) to bj , if the call resulted from finding a witness for bj in the previous
call. Note that every pair Ak , Bℓ is charged by exactly one call and every bj is charged by at most
one call (since in after a witness for bj is found, we delete bj and no further
bj can be
? witness for
2
2
found). Thus
in
total,
we
obtain
a
running
time
of
at
most
pg
`
nq
¨
witp
n,
d
`
1,
dW
q
`
Opnq “
?
Opn ¨ witp n, d ` 1, dW 2 qq.
Theorem 3.9. We have LowRankLWS ”2 MinInnProd.
Proof. In Lemmas 3.3, 3.5, 3.7, and 3.8, we have proven
MinInnProd ď2 LowRankLWS “ LWSpWLowRank q
ď2 Static-LWSpWLowRank q ď2 AllInnProd ď2 MinInnProd,
proving the claim.
4
Coin Change and Knapsack Problems
In this section, we focus on the following problem related to Knapsack: Assume we are given
coins of denominations d1 , . . . , dm with corresponding weights w1 , . . . , wm and a target value n,
determine a way to represent n using these coins (where each coin can be used arbitrarily often)
minimizing the total sum of weights of the coins used. Since without loss of generality di ď n for
all i, we can assume that m ď n and think of n as our problem size. In particular, we describe the
input by weights w1 , . . . , wn where wi denotes the weight of the coin of denomination i (if no coin
with denomination i exists, we set wi “ 8). It is straightforward to see that this problem is an
LWS instance LWSpWcc q, where the weight matrix Wcc is a Toeplitz matrix.
Problem 4.1 (CC). We define the following LWS instantiation CC “ LWSpWcc q.
Data items: weight sequence w “ pw1 , . . . , wn q with wi P t´W, . . . , W u Y t8u
Weights: wi,j “ wj´i for 0 ď i ă j ď n
12
Translated into a Knapsack-type formulation (i.e., denominations are weights, weights are profits,
and the objective becomes to maximize the profit), the problem differs from UnboundedKnapsack
only in that it searches for the most profitable multiset of items of weight exactly n, instead of at
most n.
Problem 4.2 (UnboundedKnapsack). We are given a sequence of profits p “ pp1 , . . . , pn q with
pi P t0, 1, . . . , W u, ř
i.e., the item of size i has profit
ř pi . Find the total profit of the multiset of
indices I such that iPI i ď n and the total profit iPI pi is maximized.
The purpose of this section is to show that both CC and UnboundedKnapsack are subquadratically equivalent to the pmin, `q-convolution problem. Along the way, we also prove
quadratic-time pmin, `q-convolution-hardness of Knapsack. Recall the definition of pmin, `qconvolution.
Problem 4.3 (pmin, `q-convolution). Given vectors a “ pa0 , . . . , an´1 q, b “ pb0 , . . . , bn´1 q P
t´W, . . . , W un , determine its pmin, `q-convolution a ˚ b defined by
pa ˚ bqk “
min
0ďi,jăn:i`j“k
ai ` b j
for all 0 ď k ď 2n ´ 2.
As opposed to the classical convolution, which we denote as a f b, solvable in time Opn log nq
using FFT, no strongly subquadratic algorithm for pmin, `q-convolution is known. Compared to
the popular orthogonal vectors problem, we have less support for believing that no Opn2´ε q-time
algorithm for pmin, `q-convolution exists. In particular, interesting special cases can be solved in
subquadratic time [14] and there are subquadratic-time co-nondeterministic and nondeterministic
algorithms [8, 12]. At the same time, breaking this long-standing quadratic-time barrier is a prerequisite for progress on refuting the 3SUM and APSP conjectures. This makes it an interesting target
particularly for proving subquadratic equivalences, since both positive and negative resolutions of
this open question appear to be reasonable possibilities.
To obtain our result, we address two issues: (1) We show an equivalence between the problem
of determining only the value T rns, i.e., the best way to give change only for the target value n,
and to determine all values T r1s, . . . , T rns, which we call the output-intensive version. (2) We show
that the output-intensive version is subquadratic equivalent to pmin, `q-convolution.
Problem 4.4 (oiCC). The output-intensive version of CC is to determine, given an input to CC,
all values T r1s, . . . , T rns.
We first consider issue (2) and provide a pmin, `q-convolution-based lower bound for oiCC.
Lemma 4.5 (pmin, `qconv ď2 oiCC). We have T pmin,`qconv pn, W q ď T oiCCp6n, 4p2W ` 1qq `
Opnq.
Proof. We first do a translation of the input. Note that for any scalars α, β, we have pa`αq˚pb`βq “
pa ˚ bq ` α ` β. Let M :“ 2W ` 1. Without loss of generality, we may assume that
2M ď ai ď 3M
0 ď bj ď M
for all i “ 0, . . . , n ´ 1,
for all j “ 0, . . . , n ´ 1.
We now define a CC instance with a problem size n1 “ 6n and W 1 “ 4M by defining
w “ p4M qn ˝ pan´1 , . . . , a0 q ˝ p4M qn ˝ pbn´1 , . . . , b0 q ˝ p4M q2n .
13
We now claim that T r4n ` is “ pa ˚ bq2n´i for i “ 1, . . . , 2n, which immediately yields the lemma.
To do so, we will prove the following sequence of identities.
T ris “ 4M
T rn ` is “ an´i
T r2n ` is “ 4M
T r3n ` is “ bn´i
for i P rns,
for i P rns,
(2)
(3)
for i P r2ns,
(6)
for i P rns,
for i P rns,
T r4n ` is “ pa ˚ bq2n´i
(4)
(5)
In the last line, we define, for our convenience, pa ˚ bq2n´1 “ 4M (note that before, we defined only
the entries pa ˚ bqk with k ď 2n ´ 2).
For later convenience, observe that 0 ď wi ď 4M for all i P rn1 s. It is easy to see that this
implies 0 ď T ris ď 4M for i P rn1 s.
The identities in (2) are obvious.
To prove the identities in (3) inductively over i, recall that T rn ` is “ minj“1,...,n`i tT rn `
i ´ js ` wj u. Observe that T rn ` i ´ js ` wj ă 4M can only occur if j ě n ` 1 (since otherwise
wj “ 4M ), which implies n ` i ´ j ď n and T rn ` i ´ js “ 4M except for the case j “ n ` i. In
this case, we have T rn ` i ´ js ` wj “ T r0s ` wn`i “ an´i ď 4M .
To prove the identities in (4), observe that for 1 ď j ď 3n, we have wj ě 2M by assumption
mini ai ě 2M . Similarly, we have already argued that T ri1 s ě 2M for 1 ď i1 ď 2n. Thus, we can
inductively show that T r2n ` is “ mintT r0s ` w2n`i , minj“1,...,2n`i´1 T r2n ` i ´ js ` wj u “ 4M
using w2n`i “ 4M and that every sum in the inner minimum expression is at least 4M .
To prove the identities in (5), note that for T r3n ` i ´ js ` wj ă 4M to hold, we must have
either n ` 1 ď j ď 2n or 3n ` 1 ď j ď 3n ` i, since otherwise wj “ 4M . We observe that for
n ` 1 ď j ď 2n, we have wj ě mini ai ě 2M and T r3n ` i ´ js ě mini ai ě 2M . Thus, we may
assume that 3n ` 1 ď j ď 3n ` i. Note that in this case, we have T r3n ` i ´ js “ 4M except for
the case j “ 3n ` i, where we have T r3n ` i ´ js ` wj “ T r0s ` w3n`i “ bn´i ă 4M .
Finally, for the identities in (6), we might have T r4n ` is ` wj ă 4M only if n ` 1 ď j ď 2n or
3n ` 1 ď j ď 4n. First consider the case that i “ 1. We have
T r4n ` 1s “ mintw4n`1 ,
T r4n ` 1 ´ js `wj ,
min
n`1ďjď2n looooooomooooooon
“4M
min
T r4n ` 1 ´ js `wj u “ 4M.
3n`1ďjď4n looooooomooooooon
“4M
Inductively over 1 ă i ď 2n, we will prove T r4n ` is “ pa ˚ bq2n´i . By definition,
T r4n ` is “ mintw4n`i ,
min
n`1ďjď2n
T r4n ` i ´ js ` wj ,
min
3n`1ďjď4n
T r4n ` i ´ js ` wj u
T rn ` i ´ j 1 s ` bn´j 1 u
“ mintw4n`i , min
T r3n ` i ´ j 1 s ` an´j 1 , min
1
1
1ďj ďn
1ďj ďn
(7)
Note that
min
1ďj 1 ďn
T
rn ` i ´ j 1 s
loooooomoooooon
“4M for
j 1 ěi
or
`bn´j 1 “
min
maxt1,i´nuďj 1 ďminti´1,nu
an´pi´j 1 q ` bn´j 1 “ pa ˚ bq2n´i
j 1 ăi´n
where the last equation follows from noting that the choice of j 1 lets n ´ j 1 and n ´ pi ´ j 1 q range
over all admissible pairs of values in t0, . . . , n ´ 1u summing up to 2n ´ i. Similarly, we inductively
prove that
min T r3n ` i ´ j 1 s ` an´j 1 “
1ďj 1 ďn
min
maxt1,i´nuďj 1 ďminti´1,nu
14
an´pi´j 1 q ` bn´j 1 “ pa ˚ bq2n´i ,
since an´j 1 ě 2M and T r3n ` i ´ j 1 s ě 2M whenever j 1 ě i or j 1 ă i ´ n (where the last
regime uses T r4n ` i1 s “ pa ˚ bq2n´i1 ě 2M inductively for i1 ă i). Finally, since pa ˚ bq2n´i ď
pmaxi ai q ` pmaxj bj q ď 4M , we can simplify (7) to T r4n ` is “ pa ˚ bq2n´i .
Using the notion of Static-LWS, the other direction is straight-forward.
Lemma 4.6. We have oiCC ď2 Static-LWSpWcc q ď2 pmin, `qconv.
Proof. In Lemma 3.5, we have in fact reduced the output-intensive version of LWSpWq to our
static problem Static-LWSpWq, thus specialized to the coin change problem, we only need to
show that Static-LWSpWcc q subquadratically reduces to pmin, `q-convolution. Consider an
input instance to Static-LWS given by I “ ta ` 1, . . . , a ` N u, J “ ta ` N ` 1, . . . , a ` 2N u and
values T ris, i P I. Defining M :“ 2W ` 1 and the vectors
N times
hkkkkkkikkkkkkj
u :“ pnM, T ra ` 1s, . . . , T ra ` N s, nM, . . . , nM q,
v :“ pnM, w1 , . . . , w2N q,
we have pu ˚ vqN `k “ mini“1,...,N T ra ` is ` wN `k´i “ T 1 ra ` N ` ks for all k “ 1, . . . , N , thus
a pmin, `q-convolution of two p2n ` 1q-dimensional vectors solves Static-LWSpWcc q, yielding
the claim.
The last two lemmas resolve issue (2). We proceed to issue (1) and show that the outputintensive version is subquadratically equivalent to both CC and UnboundedKnapsack that only
ask to determine a single output number. We introduce the following notation for our convenience:
Recall that
ř weight wi denotes the weight of a coin of denomination i. For a multiset S Ď rns, we let
dpSq :“ iPS i denote its total denomination, i.e., sum of the denomination of theřcoins in S (where
multiples uses of the same coin is allowed, since S is a multiset). We let wpSq :“ iPS w
ři denote the
weight of the multiset. Analogously, when considering a Knapsack instance, ppSq “ i pi denotes
the total profit of the item (multi)set S.
It is trivial to see that UnboundedKnapsack ď2 oiCC. Furthermore, we can give the following
simple reduction from CC to UnboundedKnapsack.
Oberservation 1 (CC ď2 UnboundedKnapsack ď2 oiCC). We have
T CC pn, W q ď T UnboundedKnapsackpn, nW q ` Opnq and T UnboundedKnapsackpn, W q ď T oiCCpn, W q `
Opnq.
Proof. Given a CC instance, for every weight wi ă 8, we create an item of size i and profit
pi :“ i ¨ M ´ wi in our resulting UnboundedKnapsack instance for a sufficiently large constant
M ě nW . This way, all profits are positive and every multiset S whose sizes sum up to B has a profit
of ppSq “ B ¨ M ´ wpSq. Since M ě nW ě maxS,dpSqďn |wpSq|, this ensures that the maximumprofit multiset of total size/denomination at most n has a total size/denomination of exactly n.
Thus, the optimal multiset S ˚ has profit pps˚ q “ n ¨ M ´ minS:dpSq“n wpSq “ n ¨ M ´ T rns, from
which we can derive T rns, as desired.
Given an UnboundedKnapsack instance, we define for every item of size i and profit pi
the corresponding weight wi “ ´pi in a corresponding CC instance. It remains to compute all
T r1s, . . . , T rns in this instance and determining their minimum, concluding the reduction.
15
The remaining part is similar in spirit to Lemma 3.8: Somewhat surprisingly, the same general
approach works despite the much more sequential nature of the Knapsack/CoinChange problem –
this sequentiality can be taken care of by a more careful treatment of appropriate subproblems that
involves solving them in a particular order and feeding them with information gained during the
process.
In what follows, to clarify which instance is currently considered, we let T I denote the T -table of
the (oi)CC LWS problem (see Problem 1.1) corresponding to instance I. Dropping the superscript
always refers to T I .
?
Lemma 4.7 (oiCC ď2 CC). We have that T oiCCpn, W q ď OplogpnW q ¨ n ¨ T CC p24 n, 3n2 W qq.
?
Proof. Let I be an oiCC instance. To define our subproblems, we set N :“ r ns and define
N ranges W1 :“ t1, . . . , N u, . . . , WN :“ tpN ´ 1qN ` 1, . . . , N 2 u. To determine all T ris “
minS:dpSq“i wpSq, we will compute T ris for all i P Wj successively over all j “ 1, . . . , N . The case
of j “ 1 and j “ 2 can be computed by the naive algorithm in time OpN 2 q “ Opnq. Consider
now any fixed j ě 3 and assume that all values T ris for i P Wj 1 with j 1 ă j have already been
computed. We employ a parallel binary search. For every i P Wj , we set up a feasible range Ri
initialized to t´nW, . . . , nW u. We will maintain the invariant that T ris P Ri and will halve the
size of all feasible ranges Ri , i P Wj simultaneously using a small number of calls to the following
problem PpM, W̄ q: Given an instance J for CC specified by the weights w̃1 , . . . , w̃M , as well as
values r̃1 , . . . , r̃M P t´W̄ , . . . , W̄ u Y t´8, 8u, determine whether there exists an i P rM s with
T J ris ď r̃i , and if so, also return a witness i. We will later prove that this problem can be solved in
time T P pM, W̄ q “ OpT CC p2M, 3M 2 W̄ qq. Clearly, after OplogpnW qq rounds of this parallel binary
search, the feasible ranges consists of single values, thus determining the values of all T ris for i P Wj .
Since we will show that halving all feasible ranges for range Wj takes OpN q calls to Pp12N, nW q,
and we need to determine at most N ranges W3 , . . . , WN , the total time for this process amounts
to OplogpnW qN 2 ¨ T P p12N, nW qq “ OplogpnW qN 2 ¨ T CC p24N, 3n2 W qq.
We now describe how to use P to halve the size of all feasible ranges Ri , i P Wj : we set
ri to the median of Ri and aim to determine, for all i P Wj , whether T ris ď ri , i.e., whether
some multiset S with dpSq “ i and wpSq ď ri exists. We achieve this by the following process:
For every k “ 1, . . . , j, we consider only two ranges, namely Wk “ tpk ´ 1qN ` 1, . . . , kN u and
Wj´k Y Wj´k`1 “ tpj ´ k ´ 1qN ` 1, . . . , pj ´ k ` 1qN u. Let us first consider the case k ě 2. Here,
we can define the 2N -dimensional vectors a, b with
#
wpk´1qN `ℓ for ℓ P rN s,
aℓ “
8
for ℓ ą N,
bℓ “ T rpj ´ k ´ 1qN ` ℓs
for ℓ P r2N s.
(Note that all T ris, i P Wj´k Y Wj´k`1 for k ě 2 have already been computed by assumption.) We
are interested in all those values of the pmin, `q-convolution a˚b of these vectors that correspond
to summing up some wpk´1qN `ℓ with some T rpj ´ k ´ 1qN ` ℓ1 s such that pj ´ 2qN ` ℓ ` ℓ1 P Wj .
More specifically, we aim to determine whether there is some ℓ with pa ˚ bqN `ℓ ď rpj´1qN `ℓ . To do
so, we use the reduction from pmin, `q-convolution to oiCC given in Lemma 4.5 to create an
oiCC instance J . From this instance of problem size 12N we can read off the values of a ˚ b as a
certain interval in the corresponding T J -table. Thus, we can test whether pa ˚ bqN `ℓ ď rpj´1qN `ℓ
for some ℓ using Pp12N, nW q: for every ℓ, we let i be the unique index in the T J -table representing
16
the entry pa ˚ bqN `ℓ and set r̃i :“ rpj´1qN `ℓ . For all other i1 , we set r̃i1 “ ´8, thus enforcing that
those indices will never be reported.
For the special case k “ 1, we proceed slightly differently: Here, we define the 2N -dimensional
vectors a, b with
aℓ “ T rℓs
for ℓ P r2N s
#
T rpj ´ 2qN ` ℓs for ℓ P rN s
bℓ “
8
for ℓ ą N.
(Note that all necessary T ris, i P W1 Y W2 and T ris, i P Wj´1 have already been computed by
assumption.) Analogously to above, we use Pp12N, nW q to test whether pa ˚ bqN `ℓ ď rpj´1qN `ℓ
using the reduction from pmin, `q-convolution to oiCC given in Lemma 4.5.
Once an i P Wj has been reported to satisfy T ris ď ri for some witnessing subproblem given
by the ranges Wk and Wj´k Y Wj´k`1 for some k, we set ri :“ ´8 and repeat on the same
subproblem k (analogously to the approach of Lemma 3.8). Note that for every j, we have j ď N
subproblems and at most N many indices i P Wj that can be reported. Thus, we use at most
OpN q many calls to the subproblem P.
To briefly argue correctness, note that by construction, we only determine some i with T ris ď ri
if we have found a witness. For the converse, let k be the largest index such that the optimal multiset
for i includes a coin in Wk . Then the subproblem given by the ranges Wk and Wj´k Y Wj´k`1
will give a witness. This is obvious for k ě 2. For k “ 1, note that no weight in Wk1 with k 1 ą 1 is
used in an optimal multiset for T ris P Wj . In particular, the optimal multiset S can be represented
as S “ S 1 Y S 2 , where S 1 is a multiset of total denomination i1 P Wj´1 and S 2 is a multiset of
total denomination i ´ i1 P W1 Y W2 . Thus, in the instance constructed from a, b, we will find the
witness T ris ď T ri1 s ` T ri ´ i1 s ď ri .
We finally describe how to solve PpM, W̄ q in time T CCp2M, 3M 2 W̄ q. First consider the problem
without finding a witnessing i. Let w̃1 , . . . , w̃M , r̃1 , . . . , r̃M be an instance J of PpM, W̄ q. We define
a CC instance K of problem size 2M by giving the weights
wi1 :“ w̃i
1
w2M´i
for all i P rM s,
:“ ´3M W̄ ´ r̃i
for all i P rM s.
We claim that T K r2M s ď ´3M W̄ iff the input instance to P is a yes instance: First observe that
T K r1s “ T J r1s, . . . , T K rM s “ T J rM s since the first M weights agree for both J and K. Consider
the case that there is some i P rM s with T J ris ď r̃i . Then we have T Kr2M s ď T K ris ` w2M´i “
pT J ris ´ r̃i q ´ 3M W̄ ď ´3M W̄ , as desired. Conversely, assume that all T J ris ą r̃i . We distinguish
the cases whether the optimal subsequence S uses only weights among w̃1 , . . . , w̃M or not. In the
first case, since |w̃i | ď W for i P rM s, we have that wpSq ě 2M ¨ miniPrns |w̃i | ě ´2M W̄ ą ´3M W̄ .
Otherwise, S uses exactly one weight among w̃M`1 , . . . , w̃2M . Let this weight be w̃2M´i . Then
wpSq “ T K ris ` w̃2M´i “ pT J ris ´ r̃i q ´ 3M W̄ ą ´3M W̄ since T J ris ą r̃i , yielding the claim.
Very similar to Lemma 3.8, we can now tune the above reduction to also produce a witness i
such that T J ris ď r̃i . For this, we scale all weights wi1 , i P r2M s by a factor of M and subtract
a value of i ´ 1 for every wi1 , i P rM s. It is easy to see that a yes instance K attains some value
T K r2M s “ ´κ ¨ M ´ i for some integers κ ě 3 and 0 ď i ă n, where i ` 1 is a witness for
T J ri ` 1s ď r̃i`1 , thus computing T K r2M s lets us derive a witness as well. Thus, problem P can
be solved by a single call to T CCp2M, 3M 2 W̄ q.
17
The results above prove the following theorem.
Theorem 4.8. We have pmin, `qconv ”2 CC ”2 UnboundedKnapsack. Furthermore, the
bounded version of Knapsack admits no strongly subquadratic-time algorithm unless pmin, `qconvolution can be solved in strongly subquadratic time.
Proof. Lemmas 4.5 and 4.6 prove pmin, `qconv ”2 oiCC, while Observation 1 and Lemma 4.7
establish oiCC ”2 CC ”2 UnboundedKnapsack, yielding the first claim.
The second claim follows from inspecting the proofs of Lemma 4.5, Lemma 4.7 and the first
claim of Observation 1 and observing that we only reduce to CC/Knapsack instances in which
the optimal multiset (for each total size) is always a set, i.e., uses each element at most once.
5
Chain LWS
In this section we consider a special case of of Least-Weight Subsequence problems called the Chain
Least-Weight Subsequence. This captures problems in which edge weights are given implicitly by
a relation R that determines which pairs of data items we are allowed to chain – the aim is to find
the longest chain.
An example of a Chain Least-Weight Subsequence problem is the NestedBoxes problem.
Given n boxes in d dimensions, given as non-negative, d-dimensional vectors b1 , . . . , bn , find the
longest chain such that each box fits into the next (without rotation). We say box that box a fits
into box b if for all dimensions 1 ď i ď d, ai ď bi .
NestedBoxes is not immediately a least-weight subsequence problem, as for least weight subsequence problems we are given a sequence of data items, and require any sequence to start at the
first item and end at the last. We can easily convert NestedBoxes into a LWS problem by sorting
the vectors by the sum of the entries and introducing two special boxes, one very small box K such
that K fits into any box bi and one very large box J such that any bi fits into J.
We define the chain least-weight subsequence problem with respect to any relation R and consider a weighted version where data items are given weights. To make the definition consistent
with the definition of LWS the output is the weight of the sequence that minimizes the sum of the
weights.
Problem 5.1 (ChainLWS). Fix a set of objects X and a relation R Ď X ˆ X. We define the
following LWS instantiation ChainLWSpRq “ LWSpWChainLWSpRq q.
Data items: sequence
of objects x0 , . . . , xn P X with weights w1 , . . . , wn P t´W, . . . , W u.
#
wj if pxi , xj q P R,
for 0 ď i ă j ď n.
Weights: wi,j “
8 otherwise,
The input to the (weighted) chain least-weight subsequence problem is a sequence of data items,
and not a set. Finding the longest chain in a set of data items is NP-complete in general. For
example, consider the box overlap problem: The input is a set of boxes in two dimensions, given by
the top left corner and the bottom right corner, and the relation consists of all pairs such that the
two boxes overlap. This problem is a generalization of the Hamiltonian path problem on induced
subgraphs of the two-dimensional grid, which is an NP-complete problem [29].
We relate ChainLWSpRq to the class of selection problems with respect to the same relation
R.
18
Problem 5.2 (Selection Problem). Given data items a1 , . . . , an , b1 , . . . , bn and a relation Rpai , bj q,
determine if there is a pair i, j satisfying Rpai , bj q. We denote this selection problem with respect
to a relation R by SelectionpRq.
The class of selection problems includes several well studied problems including MinInnProd,
OV [39, 4] and VectorDomination [26].
We will use the selection problems in the search variant, where we find a pair satisfying the R
if such a pair exists. To reduce the the search variant to the decision variants in a fine-grained way,
we can use a simple, binary search type reduction from the decision problem to the search problem:
We give a subquadratic reduction from ChainLWSpRq to SelectionpRq that is independent
of R.
Theorem 5.3. For all relations R such that R can be computed in time subpolynomial in the
number of data items n, ChainLWSpRq ď2 SelectionpRq.
The proof is again based on Static-LWS and a variation on a trick of [40].
As an intermediate step, we define Static-ChainLWS as the equivalent of Static-LWS in the
special case for chains.
Problem 5.4 (Static-ChainLWS). Fix an instance of ChainLWSpRq. Given intervals I :“
ta ` 1, . . . , a ` N u and J :“ ta ` N ` 1, . . . , a ` 2N u for some a and N , together with the correctly
computed values T ra ` 1s, . . . , T ra ` N s, the Static Chain Least-Weight Subsequence Problem
(Static-ChainLWS) asks to determine
T 1 rjs :“
min
iPI:Rpi,jq
T ris ` wj
for all j P J.
Similar to the definition of ChainLWS, Static-ChainLWS is the special case of Static-LWS
where the the weights wi,j are restricted to be either wj or 8, depending on R. As a result, Lemma
3.5 applies directly.
Corollary 5.5 (ChainLWSpRq ď2 Static-LWSpRq). For any R, if Static-ChainLWSpRq can
be solved in time Opn2´ε q for some ε ą 0, then ChainLWSpRq can be solved in time Õpn2´ε q.
We now reduce Static-ChainLWSpRq to SelectionpRq with a variation on the trick by [40].
Lemma 5.6 (Static-ChainLWSpRq ď2 SelectionpRq). For all relations R such that R can
be computed in time subpolynomial in the number of data items n, Static-ChainLWSpRq ď2
SelectionpRq.
Proof. As a first step, we sort the data items ai , i P I “ ta ` 1, . . . , a ` N u by T ris in increasing
order and we will assume for the remainder of the proof that for all?a ` 1 ď i ă a ` N we have
T ris ď T ri ` 1s. We then split the set aa`1 , . . . , aa`N into g :“ r N s groups A1 , . . . , Ag with
Ai “ tapi´1qrN {gs , . . . , airN {gs´1 u. We split the set ba`N `1 , . . . , ba`2N into B1 , . . . , Bg in a similar
fashion. We then iterate over all pairs Ak , Bl with k, l P rgs in lexicographic order, and for each pair
we do the following. Call the oracle for SelectionpRq on the input Ak , Bl to find a pair ai , bj such
that the relation R is satisfied on the pair. If there is no such pair, move to the next pair Ak˚ , Bl˚
of sets of data items. If there is such a pair, find the first element ai˚ P Ak such that Rpai˚ , bj q
using a simple linear scan. As we first sorted A and iterate over sets Ak , Bl in lexicographic order,
we have T 1 rjs “ T ri˚ s ` wj . We then remove bj from Bl and repeat.
19
For the runtime analysis, we observe, that the oracle can find a pair of elements at most OpN q
times, as each time we find a pair we remove an element from the input. In the case where we do
find a pair of elements we do a linear scan that takes OpN {gq time. Furthermore, each pair of sets
Ak , Bl can fail to find a?pair at most once. Hence, if T Selection is the time to solve the selection
problem and using g “ N we get a time of
?
?
?
?
T pN q “ N T Selection p N q ` N pT Selection p N q ` N q “ N T Selection p N q
(8)
which is subquadratic if T Selection pN q is subquadratic.
Theorem 5.7. Let D be the set of possible data items. For any relation R such that
• There is a data item K such that pK, dq P R for all d P D.
• There is a data item J such that pd, Jq P R for all d P D.
• For any set of data items d1 , . . . , dn there is a sequence i1 , . . . , in such that for any j ă k,
pdij , dik q R R. This ordering can be computed in time Opn2´δ q for δ ą 0. We call this ordering
the natural ordering.
Then SelectionpRq ď2 ChainLWSpRq.
Proof. We construct an unweighted ChainLWS problem with all weights set to ´1, so that the
problem is to find the longest chain. Let a1 , . . . an and b1 , . . . , bn be the data items of SelectionpRq
and sort both sets according to the natural ordering. We claim that for the sequence of data items
K, a1 , . . . an , b1 , . . . , bn , J the weight of the least weight subsequence is ´3 exactly if there is a pair
pai , bj q P R. Because of the property of the natural ordering, any valid subsequence starting at K
and ending at J contains at most one element ai and at most one element bj . If there is a pair
pai , bj q P R, then the sequence K, ai , bj , J will have value ´3. If there is no such pair, any valid
sequence contains at most one element other than K and J and its value is therefore at least ´2.
The proof is in the appendix.
In the rest of the section we give some interesting instantiations of the subquadratic equivalence
of Selection and ChainLWS.
Corollary 5.8 (NestedBoxes ”2 VectorDomination). The weighted NestedBoxes problem
2
on d “ c log n dimensions can be solved in time n2´p1{Opc log cqq . For d “ ωplognq, the (unweighted)
NestedBoxes problem cannot be solved in time Opn2´ε q for any ε ą 0 assuming SETH.
Proof. Let R be the relation that contains all pairs of non-negative, d-dimensional vectors a, b such
that ai ď bi for all i. Now SelectionpRq is VectorDomination, and ChainLWSpRq is the
NestedBoxes problem.
Using the reduction from Theorem 5.3 and the algorithms for vector domination of the stated
runtime [26, 13] we immediately get an algorithm for NestedBoxes.
We apply Theorem 5.7 with J “ W d where W is the largest coordinate in all input vectors,
K “ 0d and use the sum of the coordinates of the boxes as the natural ordering. SETH-hardness
of NestedBoxes then follows from the SETH-hardness of vector domination [39].
20
If we restrict NestedBoxes and VectorDomination to Boolean vectors, then we get SubsetChain and SetContainment respectively. In this case the upper bound improves to n2´1{Oplog cq
[4].
We would like to point out that the definition of ChainLWS requires the input to be a sequence
of data items, and not a set. Consider the following definition:
Problem 5.9 (ChainSet). Let a set of data items data items tx0 , . . . , xn u, weights w1 , . . . , wn´1 P
t´W, . . . , W u and a relation Rpxi , xj q be given. The chain set problem for R, denoted ChainSetpRq
asks to find the weight sequence i0 , i1 , i2 , . . . , ik such that for all j with 1 ď j ď k the pair pxij´1 , xij q
řk´1
is in the relation R and the weight j“1 wij is minimized.
While ChainLWS can always be solved in quadratic time, ChainSet is NP-complete. For
example, consider the box overlap problem: The input is a set of boxes in two dimensions, given by
the top left corner and the bottom right corner, and the relation consists of all pairs such that the
two boxes overlap. This problem is a generalization of the Hamiltonian path problem on induced
subgraphs of the two-dimensional grid, which is an NP-complete problem [29]. This is a formal
barrier to a more general reduction than Theorem 5.7, as we need some mechanism to impose an
ordering on the data items.
6
Near-linear time algorithms
In this section, we classify problems to be solvable in near-linear time using the lens of our framework.
Note that in these instances, near-linear time solutions have already been known, however, our
focus on the static variants of LWS provides a simple, general approach to find fast algorithms
by identifying a simple “core” problem. Since in this paper, we generally ignore subpolynomial
factors in the running time, we concentrate here on the reduction from some LWS variant to its
corresponding core problem and disregard reductions in the other direction.
6.1
Longest Increasing Subsequence
The longest increasing subsequence problem LIS has been first investigated by Fredman [18], who
gave an Opn log nq-time algorithm and gave a corresponding lower bound based on Sorting. The
following LWS instantiation is equivalent to LIS.
Problem 6.1 (LIS). We define the following LWS instantiation LIS “ LWSpWLIS q.
Data items: integers
# x1 , . . . , xn P t1, . . . , W u
´1 if xi ă xj
Weights: wi,j “
8 ow.
It is straightforward to verify that ´T rns yields the value of the longest increasing subsequence of
x1 , . . . , xn . Using the static variant of LWS introduced in Section 3, we observe that LIS effectively
boils down to Sorting.
Oberservation 2. LIS can be solved in time Õpnq.
Proof. By Lemma 3.5, we can reduce LIS to the static variant Static-LWSpWLISq. It is straightforward to see that the latter can be reformulated as follows: Given pa1 , T r1sq, . . . , paN , T rN sq and
21
b1 , . . . , bN , determine for every j “ 1, . . . , N , the value T 1 rjs “ ´1 ` min1ďiďN,ai ăbj T ris. To do
so, it suffices to sort the first list as pai1 , T ri1 sq, . . . , paiN , T riN sq with ai1 ď ¨ ¨ ¨ ď aiN and the
second as bj1 , . . . , bjN with bj1 ď ¨ ¨ ¨ ď bjN . Finally, a single pass over both lists will do: For each
k “ 1, . . . , N , we search for the largest ℓ such that aiℓ ă bjk , then the T 1 -value corresponding to bjℓ
is ´1 ` min1ďℓ1 ďℓ T riℓ1 s. By this approach, it is easy to see that after sorting, these values can be
computed in time OpN q. For the exact running time, note that solving Static-LWSpWLIS q takes
time OpN log N q due to sorting, yielding a Opn log2 nq-time algorithm for LIS by Lemma 3.5.
6.2
Unbounded Subset Sum
UnboundedSubsetSum is a variant of the classical SubsetSum, in which repetitions of elements
are allowed. While improved pseudo-polynomial-time algorithms for SubsetSum could only recently be found [32, 10], there is a simple algorithm solving UnboundedSubsetSum in time
Opn log nq [10]. It can be cast into an LWS formulation as follows.
Problem 6.2 (UnboundedSubsetSum). We define the following LWS instantiation LIS “
LWSpWUSS q.
Data items: S Ď#rns
0 if j ´ i P S
Weights: wi,j “
8 ow.
Note that in this formulation, T rns “ 0 iff there is a multiset of numbers from S that sums up
to n. It is a straightforward observation that the static variant of UnboundedSubsetSum can be
solved by classical convolution, i.e., p¨, `q-convolution.
Oberservation 3. UnboundedSubsetSum can be solved in time Õpnq.
Proof. Noting that all weights wi,j are either 0 or 8, it is easy to see that the static variant
Static-LWSpWUSS q can be reformulated as follows: Given a subset X Ď I “ ta ` 1, . . . , a ` N u,
determine, for all j P J “ ta`N `1, . . . , a`2N u, whether there exists some i P X such that j ´i P S.
To do so, we do the following: We represent X as an N -bit vector x “ px1 , . . . , xN q P t0, 1uN with
xi “ 1 iff. a ` i P X. Furthermore, we represent the “relevant part” of S by defining a 2N -bit
vector s “ ps1 , . . . , s2N q P t0, 1u2N with si “ 1 iff. i P S. Then the p¨, `q-convolution r “ x f s of
x and s allows us to determine T 1 ra ` N ` js for j “ 1, . . . , N : this values is 0 iff rN `j ą 0 and
8 otherwise. Correctness follows from the observation that rN `j ą 0 is equivalent to the existence
of some i P rN s and k P r2N s with i ` k “ N ` j and xi “ sk “ 1. This in turn is equivalent to
a ` i P X and pa ` N ` jq ´ pa ` iq “ N ` j ´ i “ k P S, as desired.
Thus Static-LWSpWUSS q can be solved by a single convolution computation, which can be
performed in time OpN log N q. Thus by Lemma 3.5, this gives rise to a Opn log2 nq-time algorithm
for UnboundedSubsetSum.
6.3
Concave LWS
The concave LWS problem is a special case of LWS in which the weights satisfy the quadrangle
inequality. Since a complete description of the input instance consists of Ωpn2 q weights, we use
the standard assumption that each wi,j can be queried in constant time. This allows for sublinear
solutions in the input description, in particular there exist Opnq-time algorithms [38, 21].
22
Problem 6.3 (ConcLWS). We define the following LWS instantiation LIS “ LWSpWconc q.
Weights: wi,j given by oracle access, satisfying wi,j ` wi1 ,j 1 ď wi1 ,j ` wi,j 1 for i ď i1 ď j ď j 1 .
We revisit ConcLWS and its known connection to the problem of computing column (or row)
minima in a totally monotone5 pn ˆ nq-matrix, which we call the SMAWK problem because of
its remarkable Opnq-time solution called the SMAWK algorithm [5].
Oberservation 4. ConcLWS can be solved in time Õpnq.
Proof. The static variant of ConcLWS can be formulated as follows: Given intervals I “ ta `
1, . . . , a ` N u and J “ ta ` N ` 1, . . . , a ` 2N u, we define a matrix M :“ pmi,j qiPI,jPJ q with
mi,j “ T ris ` wi,j . It is easy to see that M is a totally monotone matrix since w satisfies the
quadrangle inequality. Note that the minimum of column j P J in M is miniPI T ris ` wi,j “ T 1 rjs
by definition. Thus, using the SMAWK algorithm we can determine all T 1 rjs in simultaneously in
time OpN q.
Thus by Lemma 3.5, we obtain an Opn log nq-time algorithm for ConcLWS.
Acknowledgments. We would like to thank Karl Bringmann and Russell Impagliazzo for helpful
discussions and comments.
References
[1] Amir Abboud, Arturs Backurs, and Virginia Vassilevska Williams. Quadratic-time hardness
of LCS and other sequence similarity measures. In Proc. 56th Annual IEEE Symposium on
Foundations of Computer Science (FOCS’15), pages 59–78, 2015.
[2] Amir Abboud, Thomas Dueholm Hansen, Virginia Vassilevska Williams, and Ryan Williams.
Simulating branching programs with edit distance and friends or: A polylog shaved is a lower
bound made. In Proc. 48th Annual ACM Symposium on Symposium on Theory of Computing
(STOC’16), 2016. To appear.
[3] Amir Abboud, Virginia Vassilevska Williams, and Oren Weimann. Consequences of faster
alignment of sequences. In Proc. 41st International Colloquium on Automata, Languages, and
Programming (ICALP’14), pages 39–51, 2014.
[4] Amir Abboud, Ryan Williams, and Huacheng Yu. More applications of the polynomial method
to algorithm design. In Proc. 26th Annual ACM-SIAM Symposium on Discrete Algorithms
(SODA’15), pages 218–230, 2015.
[5] Alok Aggarwal, Maria M. Klawe, Shlomo Moran, Peter W. Shor, and Robert E. Wilber. Geometric applications of a matrix-searching algorithm. Algorithmica, 2:195–208, 1987.
[6] Alfred V. Aho, Daniel S. Hirschberg, and Jeffrey D. Ullman. Bounds on the complexity of the
longest common subsequence problem. Journal of the ACM, 23(1):1–12, 1976.
5 A matrix M “ pm q
1
1
i,j i,j is totally monotone if for all i ă i and j ă j , we have that mi,j ą mi1 ,j implies that
mi,j 1 ą mi1 ,j 1 . For a more comprehensive treatment, we refer to [5, 21].
23
[7] Arturs Backurs and Piotr Indyk. Edit distance cannot be computed in strongly subquadratic
time (unless SETH is false). In Proc. 47th Annual ACM Symposium on Theory of Computing
(STOC’15), pages 51–58, 2015.
[8] David Bremner, Timothy M. Chan, Erik D. Demaine, Jeff Erickson, Ferran Hurtado, John
Iacono, Stefan Langerman, Mihai Patrascu, and Perouz Taslakian. Necklaces, convolutions,
and X+Y. Algorithmica, 69(2):294–314, 2014.
[9] Karl Bringmann. Why walking the dog takes time: Frechet distance has no strongly subquadratic algorithms unless SETH fails. In Proc. 55th Annual IEEE Symposium on Foundations of Computer Science (FOCS’14), pages 661–670, 2014.
[10] Karl Bringmann. A near-linear pseudopolynomial time algorithm for subset sum. In Proc. 28th
Annual ACM-SIAM Symposium on Discrete Algorithms (SODA’17), pages 1073–1084, 2017.
[11] Karl Bringmann and Marvin Künnemann. Quadratic conditional lower bounds for string
problems and dynamic time warping. In Proc. 56th Annual IEEE Symposium on Foundations
of Computer Science (FOCS’15), pages 79–97, 2015.
[12] Marco L. Carmosino, Jiawei Gao, Russell Impagliazzo, Ivan Mihajlin, Ramamohan Paturi, and
Stefan Schneider. Nondeterministic extensions of the strong exponential time hypothesis and
consequences for non-reducibility. In Proc. 7th ACM Conference on Innovations in Theoretical
Computer Science (ITCS’16), pages 261–270, 2016.
[13] Timothy M Chan. Speeding up the four russians algorithm by about one more logarithmic
factor. In Proc. 26th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA’15),
pages 212–217. Society for Industrial and Applied Mathematics, 2015.
[14] Timothy M. Chan and Moshe Lewenstein. Clustered integer 3sum via additive combinatorics.
In Proc. 47th Annual ACM Symposium on Theory of Computing, (STOC’15), pages 31–40,
2015.
[15] Mark de Berg, Kevin Buchin, Bart M. P. Jansen, and Gerhard J. Woeginger. Fine-grained
complexity analysis of two classic TSP variants. In Proc. 43rd International Colloquium on
Automata, Languages, and Programming (ICALP’16), pages 5:1–5:14, 2016.
[16] David Eppstein. Sequence comparison with mixed convex and concave costs. J. Algorithms,
11(1):85–101, 1990.
[17] David A. Eppstein. Efficient algorithms for sequence analysis with concave and convex gap
costs. PhD thesis, Columbia University, 1989.
[18] Michael L. Fredman. On computing the length of longest increasing subsequences. Discrete
Mathematics, 11(1):29 – 35, 1975.
[19] Anka Gajentaan and Mark H Overmars. On a class of o (n2) problems in computational
geometry. Computational geometry, 5(3):165–185, 1995.
[20] Zvi Galil and Raffaele Giancarlo. Speeding up dynamic programming with applications to
molecular biology. Theoretical Computer Science, 64(1):107–118, 1989.
24
[21] Zvi Galil and Kunsoo Park. A linear-time algorithm for concave one-dimensional dynamic
programming. Inf. Process. Lett., 33(6):309–311, 1990.
[22] Zvi Galil and Kunsoo Park. Parallel algorithms for dynamic programming recurrences with
more than O(1) dependency. J. Parallel Distrib. Comput., 21(2):213–222, 1994.
[23] Jiawei Gao, Russell Impagliazzo, Antonina Kolokolova, and Ryan Williams. Completeness for
first-order properties on sparse structures with algorithmic applications. In Proc. 28th Annual
ACM-SIAM Symposium on Discrete Algorithms (SODA’17), pages 2162–2181, 2017.
[24] A. Grønlund, K. Green Larsen, A. Mathiasen, J. Sindahl Nielsen, S. Schneider, and M. Song.
Fast Exact k-Means, k-Medians and Bregman Divergence Clustering in 1D. ArXiv e-prints,
January 2017. arXiv:1701.07204.
[25] Daniel S. Hirschberg and Lawrence L. Larmore. The least weight subsequence problem. SIAM
Journal on Computing, 16(4):628–638, 1987.
[26] Russell Impagliazzo, Shachar Lovett, Ramamohan Paturi, and Stefan Schneider. 0-1 integer
linear programming with a linear number of constraints. arXiv preprint arXiv:1401.5512, 2014.
[27] Russell Impagliazzo and Ramamohan Paturi. On the complexity of k-sat. Journal of Computer
and System Sciences, 62(2):367 – 375, 2001.
[28] Russell Impagliazzo, Ramamohan Paturi, and Francis Zane. Which problems have strongly
exponential complexity? Journal of Computer and System Sciences, 63(4):512–530, 2001.
[29] Alon Itai, Christos H Papadimitriou, and Jayme Luiz Szwarcfiter. Hamilton paths in grid
graphs. SIAM Journal on Computing, 11(4):676–686, 1982.
[30] Maria M. Klawe and Daniel J. Kleitman. An almost linear time algorithm for generalized
matrix searching. SIAM J. Discrete Math., 3(1):81–97, 1990.
[31] Donald E. Knuth and Michael F. Plass. Breaking paragraphs into lines. Softw., Pract. Exper.,
11(11):1119–1184, 1981.
[32] Konstantinos Koiliaris and Chao Xu. A faster pseudopolynomial time algorithm for subset
sum. In Proc. 28th Annual ACM-SIAM Symposium on Discrete Algorithms, (SODA’17), pages
1062–1072, 2017.
[33] Daniel Lokshtanov, Dániel Marx, and Saket Saurabh. Lower bounds based on the exponential
time hypothesis. Bulletin of the EATCS, 105:41–72, 2011.
[34] William J. Masek and Mike Paterson. A faster algorithm computing string edit distances.
Journal of Computer and System Sciences, 20(1):18–31, 1980.
[35] Webb Miller and Eugene W. Myers. Sequence comparison with concave weighting functions.
Bulletin of Mathematical Biology, 50(2):97–120, 1988.
[36] David Pisinger. Dynamic programming on the word RAM. Algorithmica, 35(2):128–145, 2003.
[37] Virginia Vassilevska Williams. Hardness of Easy Problems: Basing Hardness on Popular Conjectures such as the Strong Exponential Time Hypothesis (Invited Talk). In Proc. 10th International Symposium on Parameterized and Exact Computation (IPEC’15), pages 17–29, 2015.
25
[38] Robert E. Wilber. The concave least-weight subsequence problem revisited. J. Algorithms,
9(3):418–425, 1988.
[39] Ryan Williams. A new algorithm for optimal 2-constraint satisfaction and its implications.
Theoretical Computer Science, 348(2):357–365, 2005.
[40] Virginia Vassilevska Williams and Ryan Williams. Subcubic equivalences between path, matrix
and triangle problems. In Proc. 51st Annual IEEE Symposium on Foundations of Computer
Science (FOCS’10), pages 645–654, 2010.
[41] F. Frances Yao. Efficient dynamic programming using quadrangle inequalities. In Proc. 12th
Annual ACM Symposium on Theory of Computing (STOC’80), pages 429–435, 1980.
26
| 8cs.DS
|
arXiv:1504.06363v1 [cs.DS] 23 Apr 2015
On the Runtime of Randomized Local Search and
Simple Evolutionary Algorithms for Dynamic
Makespan Scheduling
Frank Neumann
Optimisation and Logistics
School of Computer Science
The University of Adelaide
Adelaide, Australia
Carsten Witt
DTU Compute
Technical University of Denmark
2800 Kgs. Lyngby, Denmark
February 20, 2018
Abstract
Evolutionary algorithms have been frequently used for dynamic optimization
problems. With this paper, we contribute to the theoretical understanding of this
research area. We present the first computational complexity analysis of evolutionary algorithms for a dynamic variant of a classical combinatorial optimization
problem, namely makespan scheduling. We study the model of a strong adversary
which is allowed to change one job at regular intervals. Furthermore, we investigate the setting of random changes. Our results show that randomized local search
and a simple evolutionary algorithm are very effective in dynamically tracking
changes made to the problem instance.
Optimization problems in real-world applications often change due to a changing
environment. Evolutionary algorithms, ant colony optimization and other bio-inspired
search heuristics have been frequently applied to dynamically changing problems.
An important approach to gain a theoretical understanding of evolutionary algorithms and other types of bio-inspired computation methods is the computational complexity analysis of these algorithms. During the last 20 years, a large body of results
and methods has been built up. This includes the development of methods for the analysis of bio-inspired computing [4, 6, 7, 25] and results for some of the best-known
combinatorial optimization problems such as the traveling salesperson problem [21],
set cover [5, 26], and makespan scheduling [22, 23] as well as different multi-objective
problems [8, 17, 19]. These studies often consider the algorithms called Randomized Local Search (RLS) and (1+1) EA, which we also investigate in this paper. Although these algorithms seem to be relatively simple, it should be noted that upper
bounds on the expected optimization time of these algorithms can often be translated
to population-based evolutionary algorithms with more complicated variation operators, e. g., crossover by increasing the upper bounds by only a linear factor with respect
1
to population and problem size [7]. We refer the reader to [1, 11, 16] for comprehensive
presentations of this research area.
In recent years, the computational complexity analysis of these algorithms on dynamically changing problems has gained increasing interest [9, 12, 14, 15, 18, 20].
We study one of the classical combinatorial optimization problems, namely makespan
scheduling on two machines. We consider RLS and (1+1) EA and analyze how they
are able to keep track of changes that occur to the processing times of the given jobs. In
our investigations, we examine two models of dynamic changes where in each iteration
at most the processing time of one job can be changed. In the adversary model, an adversary is able to change the processing time pi ∈ [L,U] of an arbitrary job i, possibly
repeated at regular intervals. First, we show that even for very frequent and arbitrary
changes, the algorithms are able to obtain solutions of discrepancy at most U frequently
during the run of the algorithm. Afterwards, we show that RLS and (1+1) EA can maintain solutions of discrepancy at most U if the period of changes is not too small. In the
random model, processing times are from the set {1, . . . , n} and an adversary is able to
pick the job i to be changed. The processing time pi of the chosen job is undergoing a
random change and is either increased or decreased by 1. For the random model, we
show that the (1+1) EA obtains solutions of discrepancy O(log n) in time O(n4 log n)
regardless of the initial solution and that the expected ratio between discrepancy and
makespan is at most 6/n at least once in a phase of O(n3/2 ) iterations.
The outline of the paper is as follows. We introduce the dynamic makespan problem
and the algorithms under investigation in Section 1. Our analyses for the adversary
model is presented in Section 2 and the random model is investigated in Section 3.
Finally, we finish with some conclusions.
1 Preliminaries
We investigate the performance of randomized local search and a simple evolutionary
algorithm for a dynamic version of the classical makespan problem. Given n jobs and
their processing times pi > 0, 1 ≤ i ≤ n, the goal is to assign each job to one of two
machines M1 and M2 such that the makespan is minimized. A candidate solution is
given by a vector x ∈ {0, 1}n, where job i is assigned to machine M1 if xi = 0 and
assigned to machine M2 if xi = 1, 1 ≤ i ≤ n.
The makespan of a candidate solution x is given by
)
(
f (x) = max
n
n
i=1
i=1
∑ pi (1 − xi), ∑ pi xi
x∗
and the goal is to find a solution which minimizes f . We denote by |M j | the load
of machine j = 1, 2. We consider the dynamic version of the problem where exactly
one job changes. We will also allow such changes to be repeated at regular intervals.
We assume pi ∈ [L,U], 1 ≤ i ≤ n, where L is a lower bound on the processing time of
any job and U is an upper bound. We denote by R = U/L the ratio between upper and
lower bound.
Randomized local search (RLS) (see Algorithm 1) starts with a candidate solution x
and produces in each iteration a new solution y by flipping one randomly chosen bit of
2
Algorithm 1: RLS.
choose x ∈ {0, 1}n;
while stopping criteria not fullfilled do
y ← x;
flip one bit of y chosen uniformly at random;
if f (y) ≤ f (x) then x ← y
Algorithm 2: (1+1) EA.
choose x ∈ {0, 1}n;
while stopping criteria not fullfilled do
y ← x;
flip each bit of y independently with prob. 1/n;
if f (y) ≤ f (x) then x ← y
x. (1+1) EA (see Algorithm 2) works with a more flexible mutation operator which flips
each bit with probability 1/n. The two introduced algorithms are standard benchmark
algorithms in the area of runtime analysis of evolutionary computation [1, 11, 16].
While evolutionary algorithms usually work with a larger population and potentially
also a crossover operator, usually positive statements on (1+1) EA transfer to elitist
population-based evolutionary algorithms by losing only a polynomial factor dependent on the problem and population size [7]. This holds for all results obtained in this
paper as well as long as there is in each iteration an inverse polynomial probability of
selecting each individual of the parent population, selection does not accept worsenings of the worst fitness value from the population, and only the variation operator of
(1+1) EA is applied.
We study the runtime behaviour of RLS and (1+1) EA on the introduced dynamic
makespan scheduling problem and their ability to obtain solutions of good discrepancy.
For our theoretical investigations, we do not consider any stopping criteria and measure
runtime by the number of iterations of the while-loop to achieve a solution of desired
quality. The expected number of iterations is referred to as the expected time to reach
the desired goal. In our investigations, we denote by
!
!
n
d(x) =
∑ pi (1 − xi)
i=1
n
−
∑ p i xi
,
i=1
the discrepancy of the solution x. We will study the expected time, for different scenarios, until RLS and (1+1) EA have produced solutions of small discrepancy.
We state an important property on the number of jobs on the fuller machine (i. e.,
the heavier loaded machine, which determines the makespan), which can easily be
derived by taking into account the upper (U) and lower (L) bound on the processing
times.
3
• Every solution has at least ⌈(P/2)/U⌉ ≥ ⌈(n/2)(L/U)⌉ = ⌈(n/2) · R−1⌉ jobs on
the fuller machine.
2 Adversary Model
In this section, we consider the case of a strong adversary. In one change, the adversary
is allowed to pick one job i to be changed and is able to choose an arbitrary new
processing time pi ∈ [L,U].
2.1 Obtaining a discrepancy of at most U
We start our analysis by presenting upper bounds for RLS and (1+1) EA to obtain a
discrepancy of at most U from any starting solution.
2.1.1 RLS
We first consider RLS and show that the algorithm obtains a solution of discrepancy
at most U in expected time O(n min{log n, log R}). This bound holds independently
of the initial solution and the number of changes made by the adversary. The only
requirement is that the adversary makes at most one change at a time. The proof uses
the fact that for RLS the number of jobs on the fuller machine does not increase until
the fuller machine switches.
Theorem 1 The expected time until RLS has obtained a solution x with d(x) ≤ U is
O(n min{log n, log R}) independently of the initial solution and the number of changes
made by the adversary.
Proof. We assume that we are starting with an arbitrary solution assigning the jobs to
the two machines. Let x be the current solution and consider in each point in time the
fuller machine. The number of jobs on the fuller machine does not increase as this
would lead to a larger discrepancy.
We claim that if the fuller machine switched (either by moving a single job or by
a single change of the adversary) then a solution of discrepancy at most U has been
obtained in the step before and after the switch. Note that moving one job to another
machine changes the load on each machine by at most U and that the adversary can
change the load on each machine by at most U − L. So, the step switching the fuller
machine (accepted or rejected) reduces the load on the fuller machine from P/2 + α ,
where P = ∑ni=1 pi , to P/2 − β where α + β ≤ U. This implies min{α , β } ≤ U/2 and
therefore a discrepancy of at most U directly before and/or after the fuller machine
has switched. Note, that such a step is only accepted by RLS iff β ≤ α and that a
discrepancy of at most U has been obtained if the step is accepted. On the other hand,
the case α < β which is rejected by RLS implies a discrepancy of at most U before the
switch.
The fuller machine has at least ⌈(n/2) · R−1 ⌉ jobs. Let k be the number of jobs
on the fuller machine. Then the probability to reduce the number of jobs on the fuller
4
machine is nk and the expected waiting time for such a step is n/k. Summing up, the
expected time to switch the fuller machine is at most
n
n
k
k=max{⌈(n/2)·R−1 ⌉,1}
∑
We have two cases. If R ≥ n/2, the sum is at most nHn = O(n log n), where Hn is
the n-th Harmnoic number. If R < n/2, the sum is at most n ln n + 1 − n ln(n/(2R)) =
O(n log R). Altogether, after at most O(n min{log n, log R}) steps a solution of discrepancy at most U has been obtained.
2.1.2 (1+1) EA
In Theorem 1, we exploited that accepted steps of RLS cannot increase the number
of jobs on the fuller machines. In contrast, the (1+1) EA may move few big jobs
from the fuller to the emptier machine and many small jobs the other way round if
the accumulated effect of the step decreases the discrepancy. Such multiple-bit flips,
which may increase the number of jobs on the fuller machine, arise in a similar way
in the analysis of the (1+1) EA on linear functions, where they complicate the analysis
considerably [24]. However, it is also known that the number of incorrectly set bits in
the (1+1) EA (corresponding to the number of jobs on the fuller machine) has a drift
towards 0. We are going to show that this drift leads in time O(n3/2 ) to the situation
that the fuller machine switches, which was analyzed in Theorem 1. We cannot show
the bound O(n log n) using the advanced drift techniques from [24] since the dynamics
of the job sizes do not allow us to use the potential function from the literature.
Theorem 2 The expected time until the (1+1) EA has obtained a solution x with d(x) ≤
U is O(n3/2 ) independently of the initial solution and the number of changes made by
the adversary.
Proof. We start with a given search point x0 , where the time index w. l. o. g. is 0.
W. l. o. g., M1 is the fuller machine w. r. t. x0 . We write ℓt to denote the load of M1
after t steps. Now, let T denote the first point in time where ℓt ≤ P/2 + U/2. At this
time, M1 might still be the fuller machine, which implies a discrepancy at most U. Only
if ℓT < P/2 −U/2, the discrepancy is greater than U. Note that ℓT −1 > P/2 +U/2 and
each job size is at most U. Each step resulting in ℓT < P/2−U/2 must flip at least 2 bits
and can be converted into a step resulting in ℓT ∈ [P/2 −U/2, P/2 +U/2] by conditioning on that a certain subset of bits do not flip. Note that the step defined by the stopping
time T may be required to flip already more than one bit to reach ℓT ≤ P/2 + U/2 or
even no bits may flip at all if the adversary is responsible for reaching ℓT ≤ P/2 +U/2;
in the latter case, already discrepancy at most U has been obtained. Note also that
flipping bits in addition to the ones required to reach ℓT ≤ P/2 + U/2 may result in a
rejected step. If we condition on the step flipping as few additional bits as possible, we
are guaranteed to enter the interval [P/2 − U/2, P/2 + U/2] for the load of the fuller
machine, resulting in an accepted step. The probability of not flipping a certain subset
of bits is at least (1 − 1/n)n ≥ e−2 . Hence, if the step leading to time T flips more than
5
the required bits, we repeat the following analysis and increase the expected time by a
factor of at most e2 .
We denote by N1 (xt ) the number of jobs on M1 with respect to xt , the current search
point after t steps. Based on this, we define the potential function
(
N1 (xt ) if t < T
d(xt ) :=
0
otherwise.
Hence, the potential function reflects the number of jobs on the fuller machine before
time T and is set to 0 afterwards. As we have argued, the discrepancy at time T is at
most U with probabiblity at least e−2 .
The aim now is to bound E[T ], which is achieved by bounding the expression
E[d(xt ) − d(xt+1 ) | xt ;t < T ] from below and performing drift analysis. In what follows, we use the notation Xt := d(xt ).
Since it is necessary to move at least one job from the fuller machine to change
the d-value, which happens with probability at least 1/(en) for each of these jobs, and
each job on the emptier machine switches machine with probability at most 1/n, we
get the bound on the drift
Xt
n − Xt
X2
E[Xt − Xt+1 | Xt ;t < T ] ≥
(1)
1−
= t2 ,
en
n
en
which is at least 1/(en2). Hence, despite the fact that the number of jobs on the fuller
may increase, its decreases in expectation. Since the maximal d-value is n, we get
E[T ] = O(n3 ) by additive drift analysis [6]. However, the pessimistic process analyzed
here has already been more closely investigated in the literature. It has been (apart
from irrelevant differences in details) modeled by a process called PO-EA by [10],
which was recently revisited by [2]. Using this analysis, the bound can be improved to
O(n3/2 ).
In the following, we present a self-contained proof of the O(n3/2 ) bound using a
novel potential function that is easier to handle than the one proposed in the literature.
Intuitively, our potential function exploits that the process mostly moves due to √
the
variance of the one-step change (instead of the very small drift) in the regime√Xt ≤ n
whereas it is governed by the actual drift E[Xt − Xt+1 | Xt ] when Xt is above n.
For x ≥ 0, let the potential function be
(
√
√
x(ln( n) + 2 − ln(x)) if x ≤ n,
g(x) :=
√
otherwise.
3 n − nx
We note that g(x) is monotone increasing and continuous on [0, n]. Moreover, the
derivative satisfies
( √
√
ln( n) + 1 − ln(x) if x ≤ n,
dg
′
= n
g (x) :=
dx
otherwise.
x2
6
and is non-increasing and continuous as well. Hence, g(x) is a concave function. The
second derivative equals
(
√
−1/x if x ≤ n,
d2 g
′′
g (x) := 2 =
dx
otherwise.
− 2n
x3
and satisfies g′′ (x) ≤ −1/x for x ≤ n.
By the mean-value theorem, we get for all x ≤ n and for y ≥ 0 that
g(x) − g(x − y) ≥ yg′ (x) ≥ g(x + y) − g(x).
(2)
Moreover, by developing Taylor expansions of g(x − y) and g(x + y) up to terms of
fourth order, it is easy to see that
(g(x) − g(x − y)) − (g(x + y) − g(x)) ≥ −
d2 g 1
≥ .
dx2
x
(3)
We are now going to analyze the drift of the process defined by Yt := g(Xt ). To this
end, it is useful to decompose the drift into a positive and negative part. Define
∆−
X := (Xt − Xt+1 ) · 1{Xt+1 ≤ Xt }
and
∆+
X := (Xt+1 − Xt ) · 1{Xt+1 ≥ Xt }
+
−
and
Y and∆Y with respect to the Y -process. Note that E[Xt − Xt+1 | Xt ] =
accordingly
∆+
−
E ∆X | Xt − E ∆X | Xt and accordingly for the drift of the Y -process.
Combining this decomposition with (2), we obtain
E[Yt − Yt+1 | Xt ]
+
′
≥ g′ (Xt ) · E ∆−
X | Xt − g (Xt )E ∆X | Xt
= g′ (Xt )E[Xt − Xt+1 | Xt ].
If Xt >
√
n, plugging in the expression for g′ (Xt ) and the bound (1) yields
E[Yt − Yt+1 | Xt ] ≥
1
n Xt2
= ,
·
Xt2 en2 en
which does √
not depend on Xt .
If Xt ≤ n, we combine the decomposition with (3) and get for some value a(Xt )
that
E[Yt − Yt+1 | Xt ]
+
1
E ∆−
≥ a(Xt ) +
X | Xt − a(Xt )E ∆X | Xt
Xt
−
E ∆X | Xt
≥
+ a(Xt )E[Xt − Xt+1 | Xt ]
X
−t
E ∆X | Xt
≥
Xt
7
since
E[Xt − Xt+1 | Xt ] ≥ 0 according to (1). Hence, we are left with a bound on
E ∆−
X | Xt . Here we again argue that the number of jobs decreases by 1 if one of
the
X
the fuller machine moves and no other jobs moves. Consequently,
from
t jobs
Xt
≥
E ∆−
|
X
and
t
X
en
1
E[Yt − Yt+1 | Xt ] ≥
en
√
1
if Xt ≤ n. Together with the bound derived above, we have E[Yt − Yt+1 | Xt ] ≥ en
for
every Xt ≤ n. Now, since Y0 ≤ 3n1/2, the additive drift theorem yields E[T ] ≤ 3en3/2 =
O(n3/2 ) as suggested.
2.2 Recovering a discrepancy of at most U
We now consider the situation where the algorithms have obtained a solution of discrepancy at most U and the processing time of one arbitrary job is changed afterwards.
We show an upper bound of O(min{R, n}) on the time needed to obtain a discrepancy
of U after this change.
Theorem 3 Let x be the current solution that has a discrepancy of at most U before
changing the processing time of a job on the fuller machine. Then, the expected time of
RLS and (1+1) EA to obtain a discrepancy of at most U is O(min{R, n}).
Proof. We use multiplicative drift analysis [3] to show the O(n) bound and consider
drift according to the discrepancy d(x). Let P = ∑ni=1 pi and Xt be the random variable
for d(x) of the search point x at time t ≥ 0. With respect to the parameters from
the multiplicative drift theorem, we have s0 ≤ U + (U − L), smin = U and therefore
s0 /smin ≤ 2. W. l. o. g., let 1, . . . , f be the jobs on the fuller machine and p1 , . . . p f be
their processing times. Furthermore let y(i) be the search point obtained by flipping the
bit i for i = 1, . . . , f . As long as the current solution x has discrepancy greater than U,
each of these single bit flips is accepted. We get
1 n−1 f
1
· ∑ (d(x) − d(y(i))
E[Xt − Xt+1 | Xt ] ≥ · 1 −
n
n
i=1
!
f
1
2
≥
∑ 2 · pi ≥ en (P/2 + d(x)/2))
en i=1
=
1
1
(P + d(x)) ≥
d(x).
en
en
We set δ = 1/(en) and get en ln(s0 /smin ) ≤ en ln 2 = O(n) as an upper bound.
It remains to show the O(R) bound. From the previous calculation, we already have
E[Xt − Xt+1 | Xt ] ≥
1
(P + d(x)) ≥ P/(en).
en
Using additive drift analysis, the expected time to reach a discrepancy of at most U
when starting with a solution x with d(x) ≤ U + (U − L) is
(U − L)/(P/(en)) ≤ en(U − L)/(nL) = e(R − 1) = O(R).
8
Altogether the upper bound is O(min{R, n}), which completes the proof.
The previous theorem implies that both algorithms are effectively tracking solutions of discrepancy O(U) if the time where no changes to the processing times are
happening is at least c · min{R, n}, where c is an appropriate constant. In particular, changes happening every c′ n iterations where c′ is an appropriate constant can be
tracked effectively regardless of the ratio R = U/L. Furthermore, a small ratio R, e. g.
a constant, implies that very frequent changes (every c′′ R iterations, c′′ an appropriate
constant) can be tracked by RLS and (1+1) EA. These statements can be obtained by
combining drift analysis with an averaging argument over a number of phases. Due to
space restrictions, this analysis is not spelt out here.
3 Random Model
We now consider a model with less adversarial power. Dynamic changes are still possible, but each change is limited in effect. More precisely, we consider a random model
as common in the average-case analysis of algorithms [23]. For simplicity, we consider the model where all jobs sizes are in {1, . . . , n}; generalizations to other sets are
possible. At each point of time, at most one job size can be changed by the adversary.
The adversary can only choose the job to change, but neither amount or direction of
change. If a job i is chosen to change, then its processing time changes from its current
value pi to one of the two values pi + 1 and pi − 1, each with probability 1/2. Two exceptions are made if pi = n, which results in job size n − 1, and if pi = 1, which results
in job size 2 afterwards. In other words, the size of each job performs a fair random
walk on {1, . . . , n}, with reflecting barriers. With respect to the initial job sizes, we
consider both arbitrary (worst-case) initializations and the case that the sizes are drawn
uniformly at random and independently from {1, . . . , n}. Then each initial job size is
(n + 1)/2 in expectation.
It is useful to denote the random processing time of job i at time t by the random
variable Xi (t). It is well known [13] that the random walk described by the process
Xi (t), t ≥ 0, has a stationary probability distribution given by
(
1/(2n − 2) if j = 1 or j = n
lim Pr(Xi (t) = j) =
t→∞
1/(n − 1) otherwise
Hence, the probability values in the stationary distribution differ from the initial uniform distribution by a factor of at most 2. It is also well known in the theory of random
walks that the so-called mixing time (informally, the time to get sufficiently close to
the stationary distribution) of the considered random walk is O(n2 ) steps. Hence, for
any i, j ∈ {1, . . ., n} and for any t ≥ cn2 , where t denotes the number of changes to
job pi and c is a sufficiently large constant, we have
c2
c1
≤ Pr(Xi (t) = j) ≤
n
n
for two constants c1 , c2 > 0. Hereinafter, we asssume this bracketing of Xi (t) to hold,
i. e., the mixing time has elapsed for every job.
9
The aim is to analyze the discrepancies obtainable in our model. We summarize in
the following lemma a useful property of the distribution of the processing times Xi (t),
and drop the time index for convenience. Roughly speaking, it shows that there are no
big gaps in the set of values that is taken by at least one job.
Lemma 4 Let φ (i) := |{X j | X j = i ∧ j ∈ {1, . . . , n}}|, where i ∈ {1, . . . , n}, be the
frequency of jobs size i. Let G := max{ℓ | ∃i : φ (i) = φ (i + 1) = · · · = φ (i + ℓ) = 0}
the maximum gap size, i. e. maximum size of intervals with zero frequency everywhere.
Then, for some constant c > 0,
Pr(G ≥ ℓ) ≤ n2−cℓ .
Proof. Recall that we assume to be close to the stationary distribution, more precisely
for each i, j ∈ {1, . . . , n}, we have Pr(X j = i) ≥ c1 /n. By considering disjoint events,
Pr(X j ∈ {i, . . . , i + ℓ}) ≥
c1 ℓ
n
for each ℓ ≤ n.
Then for each ℓ ≥ 1, we get from the independence of the job sizes that
c1 ℓ n
,
Pr(∀ j ∈ {1, . . ., n} : X j ∈
/ {i, . . . , i + ℓ}) ≤ 1 −
n
which is at most cℓ3 for some constant c3 < 1. Hence, by a union bound the probability
that there is an i such that for all j ∈ {1, . . . , n} : X j ∈
/ {i, . . . , i + ℓ} is at most ncℓ3 , which
−cℓ
equals n2 for some constant c > 0.
Hereinafter, with high probability means probability at least 1 − O(n−c ) for any
constant c > 0. We prove the following theorem stating that with high probability
discrepancy O(log n) can be reached in polynomial time. Its proof is inspired by the
average-case analysis from [23]. Note also that the theorem is restricted to the (1+1) EA
since its proof analyzes improvements made by swapping two jobs between the fuller
and emptier machine.
Theorem 5 Regardless of the initial solution, the following claim holds: with high
probability the time for the (1+1) EA after a one-time change to obtain a discrepancy
of at most O(log n) is O(n4 log n).
Proof. According to Lemma 4, there is for any constant c > 0 a sufficiently large constant c′ > 0 such that there is not gap of size G := c′ log n or larger with probabiilty at
least 1 − n · n−c−1 = 1 − n−c. In the following, we assume this maximum gap size to
hold.
If the current discrepancy is larger than G, then there must be either at least one
pair of jobs j, j′ with j on the fuller machine and j′ on the emptier machine such that
X j′ < X j and X j − X j′ ≤ G, or a job j on the fuller machine of size of at most G. To
see this, imagine that despite the gap size of at most G, there is no such pair as in the
first case. Then all jobs of size at least G must be on the fuller machine, resulting in the
second case.
10
Now, in the first case it is sufficient to swap jobs j and j′ to decrease the discrepancy by at least 1. In the second case, it is enough to move job j from the fuller to the
emptier machine to decrease the discrepancy by at least 1. In any case, the probability
of decreasing the discrepancy is at least (1 − 1/n)n−1 n12 = Ω(n−2 ). Since the maximum discrepancy is O(n2 ), the expected number of decreases is also at most O(n2 ).
Multiplying this with the waiting time for an improvement, we have an expected time
of O(n4 ). By a simple application of Markov’s inequality and repeating phases of
length cn4 for some constant c, it is easy to see that the time is O(n4 log n) with high
probability.
The previous theorem covers a worst-case initialization with all jobs on one machine, where the discrepancy can be up to n2 . Under random initialization, this is
unlikely to happen, as the following theorem shows.
√
Theorem 6 The expected discrepancy of the random initial solution is Θ(n n). Under
a random initial solution, the time for the (1+1) EA after a one-time change to obtain
a discrepancy of O(log n) is O(n3.5 log2 n) with high probability.
√
√
Proof. We prove that the initial discrepancy is Θ(n n) in expectation and O(n n log n)
with high probability. From the last property, the statement on the time to obtain a discrepancy of O(log n) follows with the same
√ ideas as in the proof of Theorem 5 if the
initial discrepancy is estimated with O(n n log n) instead of O(n2 ).
We are left with the proofs on the initial discrepancy. Let K denote the number of jobs thatpare initially put
√ on the first machine. Then E[K] = n/2 but also
E[|K − n/2|] ≤ Var(K) = Θ( n), where we used Jensen’s inequality and the fact
that K ∼ Bin(n, 1/2).√Moreover, by the properties of the binomial distribution we have
that√Pr(K ≥ n/2 + c n) = Ω(1) for some constant√c > 0. Altogether, E[|K − n/2|] =
Θ( n). In other words, there are in expectation Θ( n) more jobs on one machine than
on the other.
Each job size is initially uniformly distributed on {1, . . . , n} and√has expectation
(n +√1)/2. By linearity of expectation, the discrepancy is at least Θ( n)(n + 1)/2 =
Θ(n n). This consideration just subtracts the total load of the machine having the
minority of the jobs from the total load of the machine having the majority (hereinafter
called machine “majority”). Should this difference be negative, the discrepancy is still
positive. However, by approximating the sum of the job
√ sizes on machine “majority”
n), one can also see that the
by a normal distribution
with
standard
deviation
Θ(n
√
discrepancy is O(n n) even if machine “majority” is allowed to have
√ a total load less
than the other machine. Altogether the expected
discrepancy
is
Θ(n
n).
√
The statement that the discrepancy is O(n n log n) with high probability follows
by using Chernoff bounds on the difference
√ in the number of jobs between the two
machines (stating that there are at most O( n log n) more jobs on one machine than the
other), approximating the tails of the sum of the
√ job sizes on the machines by a normal
distribution and arguing that a deviation of c n log n from the mean has probability
2
e−Ω(c ) .
Finally, we turn to the case that job sizes change frequently. In the extreme case, at
every point of time one job size is allowed to increase or decrease by 1. Then it seems
11
hard to obtain a discrepancy of O(log n) as shown in Theorem 5. However, we can
apply the results from Section 2, noting that the maximum job size is n at any time. In
relation to the makespan, the discrepancy, which will also be at most n, is negligible.
Theorem 7 In the model with random changes, the following holds: the expected
time until the (1+1) EA (RLS) has obtained a solution with discrepancy at most n is
O(n3/2 ) (respectively O(n log n)) independently of the initial solution and the number
of changes. The expected ratio between discrepancy and makespan is at most 6/n then.
Proof. Since R = U = n in the notation of Theorem 1 and Theorem 2, we immediately
obtain the first statement of our theorem. To compute the expected ratio, note that at any
time the sum of all job sizes has an expected value of n(n + 1)/2 and is at least n2 /3 + n
with probability 1 − 2−Ω(n) using the approximation by Normal distribution. In this
case, the makespan must be at least n2 /6 + n/2, and the ratio is at most n/(n2 /6 +
n/2) ≤ 6/n − 3/n. If the sum of the job sizes is less than n2 /3 + n, then the ratio is at
most n/n since all job sizes are at least one. Altogether, the expected ratio is bounded
from above by 6/n − 3/n + 2−Ω(n) ≤ 6/n.
4 Conclusions
We have shown that randomized local search and evolutionary algorithms are provably successful in tracking solutions of good discrepancy for the dynamic makespan
scheduling. Investigating the adversary model, we have shown that the algorithms obtain solutions of discrepancy at most U every O(n log n) (for RLS) and every O(n3/2 )
(for (1+1) EA) iterations even if changes are arbitrary and frequent. Furthermore, such
a discrepancy is maintained if the period of changes is not too small. For the random model, we have shown that discrepancies of O(log n) are obtained and that a ratio
of at most 6/n between discrepancy and makespan is obtained frequently during the
optimization process.
Acknowledgements
Frank Neumann has been supported by the Australian Research Council (ARC) through
grants DP130104395 and DP140103400. Carsten Witt has been supported by the Danish Council for Independent Research (DFF) through grant 4002-00542. We thank
Mojgan Pourhassan and three anonymous reviewers for providing valuable feedback
that helped to improve the paper.
Bibliography
[1] Anne Auger and Benjamin Doerr. Theory of Randomized Search Heuristics: Foundations and Recent Developments. World Scientific Publishing Co., Inc., 2011.
[2] Sylvain Colin, Benjamin Doerr, and Gaspard Férey. Monotonic functions in EC:
anything but monotone! In Proc. of GECCO ’14, pages 753–760. ACM Press, 2014.
12
[3] Benjamin Doerr, Daniel Johannsen, and Carola Winzen. Multiplicative drift analysis. Algorithmica, 64(4):673–697, 2012.
[4] Stefan Droste, Thomas Jansen, and Ingo Wegener. On the analysis of the (1+1)
evolutionary algorithm. Theoretical Computer Science, 276:51–81, 2002.
[5] Tobias Friedrich, Jun He, Nils Hebbinghaus, Frank Neumann, and Carsten Witt.
Approximating covering problems by randomized search heuristics using multiobjective models. Evolutionary Computation, 18(4):617–633, 2010.
[6] Jun He and Xin Yao. Drift analysis and average time complexity of evolutionary
algorithms. Artificial Intelligence, 127(1):57–85, 2001.
[7] Jun He and Xin Yao. Towards an analytic framework for analysing the computation
time of evolutionary algorithms. Artificial Intelligence, 145(1-2):59–97, 2003.
[8] Christian Horoba. Exploring the runtime of an evolutionary algorithm for the
multi-objective shortest path problem. Evolutionary Computation, 18(3):357–381,
2010.
[9] Thomas Jansen and Christine Zarges. Evolutionary algorithms and artificial
immune systems on a bi-stable dynamic optimisation problem. In Proc. of
GECCO ’14, pages 975–982. ACM Press, 2014.
[10] Thomas Jansen. On the brittleness of evolutionary algorithms. In Proc. of
FOGA ’07, pages 54–69. ACM Press, 2007.
[11] Thomas Jansen. Analyzing Evolutionary Algorithms – The Computer Science
Perspective. Natural Computing Series. Springer, 2013.
[12] Timo Kötzing and Hendrik Molter. ACO beats EA on a dynamic pseudo-boolean
function. In Proc. of PPSN’ 12, pages 113–122. Springer, 2012.
[13] David A. Levin, Yuval Peres, and Elizabeth L. Wilmer. Markov Chains and Mixing Times. American Mathematical Society, 2009.
[14] Andrei Lissovoi and Carsten Witt. MMAS versus population-based EA on a
family of dynamic fitness functions. Algorithmica, 2015. In press, final version:
http://dx.doi.org/10.1007/s00453-015-9975-z.
[15] Andrei Lissovoi and Carsten Witt. Runtime analysis of ant colony optimization on
dynamic shortest path problems. Theoretical Computer Science, 561:73–85, 2015.
[16] Frank Neumann and Carsten Witt. Bioinspired Computation in Combinatorial
Optimization: Algorithms and Their Computational Complexity. Springer, 2010.
[17] Frank Neumann. Expected runtimes of a simple evolutionary algorithm for the
multi-objective minimum spanning tree problem. European Journal of Operational
Research, 181(3):1620–1629, 2007.
13
[18] Pietro S. Oliveto and Christine Zarges. Analysis of diversity mechanisms for
optimisation in dynamic environments with low frequencies of change. In Proc. of
GECCO ’13, pages 837–844. ACM Press, 2013.
[19] Chao Qian, Yang Yu, and Zhi-Hua Zhou. An analysis on recombination in multiobjective evolutionary optimization. Artificial Intelligence, 204:99–119, 2013.
[20] Philipp Rohlfshagen, Per Kristian Lehre, and Xin Yao. Dynamic evolutionary
optimisation: An analysis of frequency and magnitude of change. In Proc. of
GECCO ’09, pages 1713–1720. ACM Press, 2009.
[21] Andrew M. Sutton and Frank Neumann. A parameterized runtime analysis of
evolutionary algorithms for the Euclidean traveling salesperson problem. In Proc.
of AAAI ’12, 2012.
[22] Andrew M. Sutton and Frank Neumann. A parameterized runtime analysis of
simple evolutionary algorithms for makespan scheduling. In Proc. of PPSN ’12,
pages 52–61. Springer, 2012.
[23] Carsten Witt. Worst-case and average-case approximations by simple randomized
search heuristics. In Proc. of STACS ’05, pages 44–56. Springer, 2005.
[24] Carsten Witt. Tight bounds on the optimization time of a randomized search
heuristic on linear functions. Combinatorics, Probability & Computing, 22(2):294–
318, 2013.
[25] Yang Yu and Zhi-Hua Zhou. A new approach to estimating the expected first
hitting time of evolutionary algorithms. In Proc. of AAAI ’06. AIII Press, 2006.
[26] Yang Yu, Xin Yao, and Zhi-Hua Zhou. On the approximation ability of evolutionary optimization with application to minimum set cover: Extended abstract. In
Proc. of IJCAI ’13. IJCAI/AAAI, 2013.
14
| 9cs.NE
|
Neural Responding Machine for Short-Text Conversation
Lifeng Shang Zhengdong Lu Hang Li
Noah’s Ark Lab
Huawei Technologies Co. Ltd.
Sha Tin, Hong Kong
{Shang.Lifeng,Lu.Zhengdong,HangLi.HL}@huawei.com
arXiv:1503.02364v2 [cs.CL] 27 Apr 2015
Abstract
We propose Neural Responding Machine (NRM), a neural network-based response generator for
Short-Text Conversation. NRM takes the general encoder-decoder framework: it formalizes the
generation of response as a decoding process based on the latent representation of the input text,
while both encoding and decoding are realized with recurrent neural networks (RNN). The NRM
is trained with a large amount of one-round conversation data collected from a microblogging
service. Empirical study shows that NRM can generate grammatically correct and content-wise
appropriate responses to over 75% of the input text, outperforming state-of-the-arts in the same
setting, including retrieval-based and SMT-based models.
1 Introduction
Natural language conversation is one of the most challenging artificial intelligence problems,
which involves language understanding, reasoning, and the utilization of common sense knowledge. Previous works in this direction mainly focus on either rule-based or learning-based methods (Williams and Young, 2007; Schatzmann et al., 2006; Misu et al., 2012; Litman et al., 2000). These
types of methods often rely on manual effort in designing rules or automatic training of model with a particular learning algorithm and a small amount of data, which makes it difficult to develop an extensible
open domain conversation system.
Recently due to the explosive growth of microblogging services such as Twitter1 and Weibo2 , the
amount of conversation data available on the web has tremendously increased. This makes a data-driven
approach to attack the conversation problem (Ji et al., 2014; Ritter et al., 2011) possible. Instead of multiple rounds of conversation, the task at hand, referred to as Short-Text Conversation (STC), only considers one round of conversation, in which each round is formed by two short texts, with the former being
an input (referred to as post) from a user and the latter a response given by the computer. The research
on STC may shed light on understanding the complicated mechanism of natural language conversation.
Previous methods for STC fall into two categories, 1) the retrieval-based method (Ji et al., 2014),
and 2) the statistical machine translation (SMT) based method (Ritter et al., 2011). The basic idea of
retrieval-based method is to pick a suitable response by ranking the candidate responses with a linear or
non-linear combination of various matching features (e.g. number of shared words). The main drawbacks
of the retrieval-based method are the following
• the responses are pre-existed and hard to be customized for the particular text or requirement from
the task, e.g., style or attitude.
• the use of matching features alone is usually not sufficient for distinguishing positive responses from
negative ones, even after time consuming feature engineering. (e.g., a penalty due to mismatched
named entities is difficult to be incorporated into the model)
The SMT-based method, on the other hand, is generative. Basically it treats the response generation as
a translation problem, in which the model is trained on a parallel corpus of post-response pairs. Despite
1
2
https://twitter.com/.
http://www.weibo.com/.
its generative nature, the method is intrinsically unsuitable for response generation, because the responses
are not semantically equivalent to the posts as in translation. Actually one post can receive responses
with completely different content, as manifested through the example in the following figure:
Post
UserA
UserB
UserC
1.1
Having my fish sandwich right now
For god’s sake, it is 11 in the morning
Enhhhh... sounds yummy
which restaurant exactly?
Overview
In this paper, we take a probabilistic model to address the response generation problem, and propose
employing a neural encoder-decoder for this task, named Neural Responding Machine (NRM). The
neural encoder-decoder model, as illustrated in Figure 1, first summarizes the post as a vector representation, then feeds this representation to decoder to generate responses. We further generalize this
scheme to allow the post representation dynamically change during the generation process, following the
idea in (Bahdanau et al., 2014) originally proposed for neural-network-based machine translation with
automatic alignment.
For god's sake, it is 11 in the morning
Enhhhh... sounds yummy
which restaurant exactly?
Decoder
vector
Encoder
Having my fish sandwich right now
Figure 1: The diagram of encoder-decoder framework for automatic response generation.
NRM essentially estimates the likelihood of a response given a post. Clearly the estimated probability
should be complex enough to represent all the suitable responses. Similar framework has been used
for machine translation with a remarkable success (Kalchbrenner and Blunsom, 2013; Auli et al., 2013;
Sutskever et al., 2014; Bahdanau et al., 2014). Note that in machine translation, the task is to estimate
the probability of a target language sentence conditioned on the source language sentence with the same
meaning, which is much easier than the task of STC which we are considering here. In this paper, we
demonstrate that NRM, when equipped with a reasonable amount of data, can yield a satisfying estimator
of responses (hence response generator) for STC, despite the difficulty of the task.
Our main contributions are two-folds: 1) we propose to use an encoder-decoder-based neural network to generate a response in STC; 2) we have empirically verified that the proposed method, when
trained with a reasonable amount of data, can yield performance better than traditional retrieval-based
and translation-based methods.
1.2
RoadMap
In the remainder of this paper, we start with introducing the dataset for STC in Section 2. Then we
elaborate on the model of NRM in Section 3, followed by the details on training in Section 4. After that,
we report the experimental results in Section 5. In Section 6 we conclude the paper.
2 The Dataset for STC
Our models are trained on a corpus of roughly 4.4 million pairs of conversations from Weibo 3 .
3
The dataset and its English translation (by machine translation system) will be released soon.
2.1
Conversations on Sina Weibo
Weibo is a popular Twitter-like microblogging service in China, on which a user can post short messages
(referred to as post in the reminder of this paper) visible to the public or a group of users following
her/him. Other users make comment on a published post, which will be referred to as response. Just like
Twitter, Weibo also has the length limit of 140 Chinese characters on both posts and responses, making
the post-response pair an ideal surrogate for short-text conversation.
Training
Test Data
Labeled Dataset
(retrieval-based)
Fine Tuning
(SMT-based)
#posts
#responses
#pairs
#test posts
#posts
#responses
#labeled pairs
#posts
#responses
#pairs
219,905
4,308,211
4,435,959
110
225
6,017
6,017
2,925
3,000
3,000
Table 1: Some statistics of the dataset. Labeled Dataset and Fine Tuning are used by retrieval-based
method for learning to rank and SMT-based method for fine tuning, respectively.
2.2
Dataset Description
To construct this million scale dataset, we first crawl hundreds of millions of post-response pairs, and
then clean the raw data in a similar way as suggested in (Wang et al., 2013), including 1) removing trivial
responses like “wow”, 2) filtering out potential advertisements, and 3) removing the responses after first
30 ones for topic consistency. Table 1 shows some statistics of the dataset used in this work. It can
be seen that each post have 20 different responses on average. In addition to the semantic gap between
post and its responses, this is another key difference to a general parallel data set used for traditional
translation.
3 Neural Responding Machines for STC
The basic idea of NRM is to build a hidden representation of a post, and then generate the response
based on it, as shown in Figure 2. In the particular illustration, the encoder converts the input sequence
x = (x1 , · · · , xT ) into a set of high-dimensional hidden representations h = (h1 , · · · , hT ), which, along
with the attention signal at time t (denoted as αt ), are fed to the context-generator to build the context
input to decoder at time t (denoted as ct ). Then ct is linearly transformed by a matrix L (as part of the
decoder) into a stimulus of generating RNN to produce the t-th word of response (denoted as yt ).
In neural translation system, L converts the representation in source language to that of target language. In NRM, L plays a more difficult role: it needs to transform the representation of post (or some
part of it) to the rich representation of many plausible responses. It is a bit surprising that this can be
achieved to a reasonable level with a linear transformation in the “space of representation”, as validated
in Section 5.3, where we show that one post can actually invoke many different responses from NRM.
The role of attention signal is to determine which part of the hidden representation h should be emphasized during the generation process. It should be noted that αt could be fixed over time or changes
dynamically during the generation of response sequence y. In the dynamic settings, αt can be function
of historically generated subsequence (y1 , · · · , yt−1 ), input sequence x or their latent representations,
more details will be shown later in Section 3.2.
We use Recurrent Neural Network (RNN) for both encoder and decoder, for its natural ability to
summarize and generate word sequence of arbitrary lengths (Mikolov et al., 2010; Sutskever et al., 2014;
Cho et al., 2014).
Decoder
Attention Signal
Context Generator
Encoder
Figure 2: The general framework and dataflow of the encoder-decoder-based NRM.
͙͘
͙͘
͙͘
Figure 3: The graphical model of RNN decoder. The dashed lines denote the variables related to the
function g(·), and the solid lines denote the variables related to the function f (·).
3.1
The Computation in Decoder
Figure 3 gives the graphical model of the decoder, which is essentially a standard RNN language model
except conditioned on the context input c. The generation probability of the t-th word is calculated by
p(yt |yt−1 , · · · , y1 , x) = g(yt−1 , st , ct ),
(1)
where yt is a one-hot word representation, g(·) is a softmax activation function, and st is the hidden state
of decoder at time t calculated by
st = f (yt−1 , st−1 , ct ),
(2)
and f (·) is a non-linear activation function and the transformation L is often assigned as parameters of f (·). Here f (·) can be a logistic function, the sophisticated long short-term memory (LSTM) unit (Hochreiter and Schmidhuber, 1997), or the recently proposed gated recurrent
unit (GRU) (Chung et al., 2014; Cho et al., 2014). Compared to “ungated” logistic function, LSTM and
GRU are specially designed for its long term memory: it can store information over extended time steps
without too much decay. We use GRU in this work, since it performs comparably to LSTM on squence
modeling (Chung et al., 2014), but has less parameters and easier to train.
3.2
The Computation in Encoder
We consider three types of encoding schemes, namely 1) the global scheme, 2) the local scheme, and the
hybrid scheme which combines 1) and 2).
Global Scheme: Figure 4 shows the graphical model of the RNN-encoder and related context generator
for a global encoding scheme. The hidden state at time t is calculated by ht = f (xt , ht−1 ) (i.e. still
GRU unit), and with a trivial context generation operation, we essentially use the final hidden state hT
as the global representation of the sentence. The same strategy has been taken in (Cho et al., 2014) and
(Sutskever et al., 2014) for building the intermediate representation for machine translation. This scheme
however has its drawbacks: a vectorial summarization of the entire post is often hard to obtain and may
lose important details for response generation, especially when the dimension of the hidden state is not
big enough4 . In the reminder of this paper, a NRM with this global encoding scheme is referred to as
NRM-glo.
Context Generator
͙͘
Figure 4: The graphical model of the encoder in NRM-glo, where the last hidden state is used as the
context vector ct = hT .
Local Scheme: Recently, Bahdanau et al. (2014) and Graves (2013) introduced an attention mechanism that allowsPthe decoder to dynamically select and linearly combine different parts of the input
sequence ct = Tj=1 αtj hj , where weighting factors αtj determine which part should be selected to
generate the new word yt , which in turn is a function of hidden states αtj = q(hj , st−1 ), as pictorially
shown in Figure 5. Basically, the attention mechanism αtj models the alignment between the inputs
around position j and the output at position t, so it can be viewed as a local matching model. This local
scheme is devised in (Bahdanau et al., 2014) for automatic alignment between the source sentence and
the partial target sentence in machine translation. This scheme enjoys the advantage of adaptively focusing on some important words of the input text according to the generated words of response. A NRM
with this local encoding scheme is referred to as NRM-loc.
Attention Signal
Context Generator
͙͘
Figure 5: The graphical model P
of the encoder in NRM-loc, where the weighted sum of hidden sates is
used as the context vector ct = Tj=1 αtj hj .
4
Sutskever et al. (2014) has to use 4, 000 dimension for satisfying performance on machine translation, while
(Cho et al., 2014) with a smaller dimension perform poorly on translating an entire sentence.
3.3
Extensions: Local and Global Model
In the task of STC, NRM-glo has the summarization of the entire post, while NRM-loc can adaptively
select the important words in post for various suitable responses. Since post-response pairs in STC are
not strictly parallel and a word in different context can have different meanings, we conjecture that the
global representation in NRM-glo may provide useful context for extracting the local context, therefore
complementary to the scheme in NRM-loc. It is therefore a natural extension to combine the two models
by concatenating their encoded hidden states to form an extended hidden representation for each time
stamp, as illustrated in Figure 6. We can see the summarization hgT is incorporated into ct and αtj to
provide a global context for local matching. With this hybrid method, we hope both the local and global
information can be introduced into the generation of response. The model with this context generation
mechanism is denoted as NRM-hyb.
It should be noticed that the context generator in NRM-hyb will evoke different encoding mechanisms
in the global encoder and the local encoder, although they will be combined later in forming a unified
representation. More specifically, the last hidden state of NRM-glo plays a role different from that of
the last state of NRM-loc, since it has the responsibility to encode the entire input sentence. This role
of NRM-glo, however, tends to be not adequately emphasized in training the hybrid encoder when the
parameters of the two encoding RNNs are learned jointly from scratch. For this we use the following
trick: we first initialize NRM-hyb with the parameters of NRM-loc and NRM-glo trained separately,
then fine tune the parameters in encoder along with training the parameters of decoder.
͙͘
Context Generator
global encoder
͙͘
Attention Signal
local encoder
Figure
graphical model for the encoder in NRM-hyb, while context generator function is ct =
PT 6: The
l ; hg ], here [hl ; hg ] denotes the concatenation of vectors hl and hg
α
[h
j
j
j=1 tj j
T
T
T
To learn the parameters of the model, we maximize the likelihood of observing the original response
conditioned on the post in the training set. For a new post, NRMs generate their responses by using a
left-to-right beam search with beam size = 10.
4 Experiments
We evaluate three different settings of NRM described in Section 3, namely NRM-glo, NRM-loc, and
NRM-hyb, and compare them to retrieval-based and SMT-based methods.
4.1
Implementation Details
We use Stanford Chinese word segmenter 5 to split the posts and responses into sequences of words.
Although both posts and responses are written in the same language, the distributions on words for the
two are different: the number of unique words in post text is 125,237, and that of response text is 679,958.
We therefore construct two separate vocabularies for posts and responses by using 40,000 most frequent
words on each side, covering 97.8% usage of words for post and 96.2% for response respectively. All the
5
http://nlp.stanford.edu/software/segmenter.shtml
remaining words are replaced by a special token “UNK”. The dimensions of the hidden states of encoder
and decoder are both 1,000, and the dimensions of the word-embedding for post and response are both
620. Model parameters are initialized by randomly sampling from a uniform distribution between -0.1
and 0.1. All our models were trained on a NVIDIA Tesla K40 GPU using stochastic gradient descent
algorithm with mini-batch. The training stage of each model took about two weeks.
4.2
Competitor Models
Retrieval-based: with retrieval-based models, for any given post p∗ , the response r ∗ is retrieved from
a big post-response pairs (p, r) repository. Such models rely on three key components: a big repository,
sets of feature functions Φi (p∗ , (p, r)), and a machine learning model to combine these features. In this
work, the whole 4.4 million Weibo pairs are used as the repository, 14 features, ranging from simple
cosine similarity to some deep matching models (Ji et al., 2014) are used to determine the suitability of
a post to a given post p∗ through the following linear model
score(p∗ , (p, r)) =
X
ωi Φi (p∗ , (p, r)).
(3)
i
Following the ranking strategy in (Ji et al., 2014), we pick 225 posts and about 30 retrieved responses for
each of them given by a baseline retriever6 from the 4.4M repository, and manually label them to obtain
labeled 6,017 post-response pairs. We use ranking SVM model (Joachims, 2006) for the parameters ωi
based on the labeled dataset. In comparison to NRM, only the top one response is considered in the
evaluation process.
SMT-based: In SMT-based models, the post-response pairs are directly used as parallel data for training a translation model. We use the most widely used open-source phrase-based translation modelMoses (Koehn et al., 2007). Another parallel data consisting of 3000 post-response pairs is used to tune
the system. In (Ritter et al., 2011), the authors used a modified SMT model to obtain the “Response”
of Twitter “Stimulus”. The main modification is in replacing the standard GIZA++ word alignment
model (Och and Ney, 2003) with a new phrase-pair selection method, in which all the possible phrasepairs in the training data are considered and their associated probabilities are estimated by the Fisher’s
Exact Test, which yields performance slightly better than default setting8 . Compared to retrieval-based
methods, the generated responses by SMT-based methods often have fluency or even grammatical problems. In this work, we choose the Moses with default settings as our SMT model.
5 Results and Analysis
Automatic evaluation of response generation is still an open problem. The widely accepted evaluation
methods in translation (e.g. BLUE score (Papineni et al., 2002)) do not apply, since the range of the
suitable responses is so large that it is practically impossible to give reference with adequate coverage. It
is also not reasonable to evaluate with Perplexity, a generally used measurement in statistical language
modeling, because the naturalness of response and the relatedness to post can not be well evaluated.
We therefore resort to human judgement, similar to that taken in (Ritter et al., 2011) but with important
difference.
5.1
Evaluation Methods
We adopt human annotation to compare the performance of different models. Five labelers with at least
three-year experience of Sina Weibo are invited to do human evaluation. Responses obtained from the
five evaluated models are pooled and randomly permuted for each labeler. The labelers are instructed
to imagine that they were the authors of the original posts and judge whether a response (generated or
retrieved) is appropriate and natural to a input post. Three levels are assigned to a response with scores
from 0 to 2:
6
8
we use the default similarity function of Lucene 7
Reported results showed that the new model outperformed the baseline SMT model 57.7% of the time.
Post
ᝅཝ⾷॰䠂㘷ᱥޡѠӰ...འᕖҼ
There are always 8 players at the Italian restricted area. Unbelievable!
Related Criterion
Labels
Response1
ᡇᱥᝅཝ䱕Ⲻ⨹䘭θㅿᖻ∊䎑ᔶခȾ
I am a big fan of the Italy team, waiting for the football game to start
Logic Consistency
Unsuitable
Response2
ᝅཝⲺ伕⢟འ㗄Ҽ
Italian food is absolutely delicious.
Semantic Relevance
Unsuitable
Response3
འᕖҼ!
Unbelievable!
Generality
Neutral
Response4
૾૾૾ԃ❬ᱥ0φ0 Ⱦ䘎⋗ⵁࡦ䘑⨹Ⱦ
Ha!Ha!Ha! it is still 0:0, no goal so far.
Scenario Dependence
Neutral
Response5
䘏↙ᱥᝅཝᕅ䱨ᆾ䏩⨹Ⱦ
This is exactly the Italian defending style football game.
²²
Suitable
Figure 7: An example post and its five candidate responses with human annotation. The content of the
post implies that the football match is already started, while the author of Response1 is still waiting for
the match to start. Response2 talks about the food of Italy. Response3 is a widely used response, but it
is suitable to this post. Response4 states that the current score is still 0:0, it is a suitable response only in
this specific scenario.
Models
NRM-glo
NRM-loc
NRM-hyb
Rtr.-based
SMT-based
Mean Score
0.969
1.065
1.156
0.984
0.313
Suitable (+2)
34.0%
36.0%
39.3%
29.8%
5.6 %
Neutral (+1)
28.9%
34.5%
37.1%
38.7%
20.0%
Unsuitable (0)
37.1%
29.5%
23.6%
31.5%
74.4%
Agreement
0.397
0.247
0.309
0.346
0.448
Table 2: The results of evaluated methods. Mean score is the average value of annotated scores over all
annotations. (Rtr.-based means the retrieval-based method)
ID
Post
∅ᒪݹᰜ䜳ᶛ儎✝Ⱦ
1
High fever attacks me
HYHU\1HZ<HDU¶V'D\
৾㘷Ҽжθ䉘ᡇԢ
㓺θᡇᖾ儎ު
2 I gain one more year.
Grateful to my group, so
happy
NRM-glo
NRM-loc
NRM-hyb
Rtr.-based
ཐՇᚥθ໔ᕰ࣑⯡ރέ∅ᒪݹᰜ䜳ৱॱ䲘ⵁ ⾓ᰟᰛᓭགྷθ䓡։ ڛ儎✝«
⯻
ᓭέ
Go to see a doctor
Rest more to boost
Get well soon and stay +LJKIHYHU«
your immune system. HYHU\1HZ<HDU¶V'D\ healthy!
⭕ᰛᘡ҆έ≮䘒᭥ᤷ
֖έ
Happy birthday! Will
stand by you forever!
⭕ᰛᘡ҆έ≮䘒ᒪ䖱
┸Ӥέ
Happy birthday! Wish
you always stay young
and pretty!
㘷Ҽ㘷Ҽθᴾу侬 ᡇਠᱥ㘷Ҽж
Ӱέ
Getting old now. Time I just gain only one
more year.
has no mercy.
ਹ㍖ㅢжཟθඐᤷօθ ᗻ亱ᶛ᭥ᤷέࣖ⋯έ ඐᤷօθඐᤷсৱȾ ∅ཟ䜳ᱥ㍖伕ѱѿ㘻 ඐᤷօέ
ፊᵫȾ
3 First day of being a
Hold on, keep it up.
I come to offer my
I have been being a
Hold on!
support. Keep it up!
vegetarian. Hold on,
vegetarian everyday
CuiDuoLa
䘏〃ཟ≊ⵕᓊ䈛ᑜѠᐻ 䘏ᱥଠݵϋⵕ㗄έ 䱩ݿ᱄აθᗹ㡈⭻Ⱦ䘏〃ཟ≊ⵕуᱥⴌⲺȾᮽኧ ࠰ৱ䑅䶈ሱ
ᢴ⚫ ૾૾
䬻࠰ৱ䑅䶈䎅㣧
Where is it? It is so
Such a nice sunny day! It is indeed a very nice WenShanOHW¶VJRRXW
We should go out with
4
to get some
I am in a great mood. weather.
some cute guys to enjoy beautiful!
inspiration. Ha! Ha!
a great outing in such a
nice weather.
Figure 8: Some responses generated by different models (originally in Chinese with their literal English
translation), where the words in boldfaces are entity names.
• Suitable (+2): the response is evidently an appropriate and natural response to the post;
• Neutral (+1): the response can be a suitable response in a specific scenario;
• Unsuitable (0): it is hard or impossible to find a scenario where response is suitable.
To make the annotation task operable, the suitability of generated responses is judged from the following five criteria:
(a) Grammar and Fluency: Responses should be natural language and free of any fluency or grammatical errors;
(b) Logic Consistency: Responses should be logically consistent with the test post;
(c) Semantic Relevance: Responses should be semantically relevant to the test post;
(d) Scenario Dependence: Responses can depend on a specific scenario but should not contradict the
first three criteria;
(e) Generality: Responses can be general but should not contradict the first three criteria;
If any of the first three criteria (a), (b), and (c) is contradicted, the generated response should be labeled
as “Unsuitable”. The responses that are general or suitable to post in a specific scenario should be
labeled as “Neutral”. Figure 7 shows an example of the labeling results of a post and its responses. The
first two responses are labeled as “Unsuitable” because of the logic consistency and semantic relevance
errors. Response4 depends on the scenario (i.e., the current score is 0:0), and is therefore annotated as
“Neutral”.
Model A
Model B
NRM-loc
NRM-hyb
NRM-hyb
Rtr.-based
Rtr.-based
Rtr.-based
SMT
SMT
NRM-glo
NRM-glo
NRM-loc
NRM-glo
NRM-loc
NRM-hyb
NRM-hyb
Rtr.-based
Average
rankings
(1.463, 1.537)
(1.434, 1.566)
(1.465, 1.535)
(1.512, 1.488)
(1.533, 1.467)
(1.552, 1.448)
(1.785, 1.215)
(1.738, 1.262)
p value
2.01%
0.01%
3.09%
48.1%
6.20%
0.32%
0.00 %
0.00 %
Table 3: p-values and average rankings of Friedman test for pairwise model comparison. (Rtr.-based
means the retrieval-based method)
5.2
Results
Our test set consists of 110 posts that do not appear in the training set, with length between 6 to 22
Chinese words and 12.5 words on average. The experimental results based on human annotation are
summarized in Table 2, consisting of the ratio of three categories and the agreement among the five
labelers for each model. The agreement is evaluated by Fleiss’ kappa (Fleiss, 1971), as a statistical
measure of inter-rater consistency. Except the SMT-based model, the value of agreement is in a range
from 0.2 to 0.4 for all the other models, which should be interpreted as “Fair agreement”. The SMT-based
model has a relatively higher kappa value 0.448, which is larger than 0.4 and considered as “Moderate
agreement”, since the responses generated by the SMT often have the fluency and grammatical errors,
making it easy to reach an agreement on such unsuitable cases.
From Table 2, we can see the SMT method performs significantly worse than the retrieval-based and
NRM models and 74.4% of the generated responses were labeled as unsuitable mainly due to fluency
and relevance errors. This observation confirms with our intuition that the STC dataset, with one post
potentially corresponding to many responses, can not be simply taken as parallel corpus in a SMT model.
Surprisingly, more than 60% of responses generated by all the three NRM are labeled as “Suitable”
or “Neutral”, which means that most generated responses are fluent and semantically relevant to post.
Among all the NRM variants
• NRM-loc outperforms NRM-glo, suggesting that a dynamically generated context might be more effective than a “static” fixed-length vector for the entire post, which is consistent with the observation
made in (Bahdanau et al., 2014) for machine translation;
• NRM-hyp outperforms NRM-loc and NRM-glo, suggesting that a global representation of post is
complementary to dynamically generated local context.
The retrieval-based model has the similar mean score as NRM-glo, and its ratio on neutral cases
outperforms all the other methods. This is because 1) the responses retrieved by retrieval-based method
are actually wrote by human, so they do not suffer from grammatical and fluency problems, and 2)
the combination of various feature functions potentially makes sure the picked responses semantically
relevant to test posts. However the picked responses are not customized for new test posts, so the ratio
of suitable cases is lower than the three neural generation models.
To test statistical significance, we use the Friedman test (Howell, 2010), which is a non-parametric
test on the differences of several related samples, based on ranking. Table 3 shows the average rankings
over all annotations and the corresponding p-values for comparisons between different pairs of methods.
The comparison between retrieval-based and NRM-glo is not significant and their difference in ranking
is tiny. This indicates that the retrieval-based method is comparable to the NRM-glo method. The NRMhyb outperforms all the other methods, and the difference is statistically significant (p < 0.05). The
difference between NRM-loc and retrieval-based method is marginal (p = 0.062). SMT is significantly
worse than retrieval-based and NRM-hyb methods.
5.3
Case Study
Figure 8 shows some example responses generated by our NRMs (only the one with biggest likelihood
is given) and the comparable retrieval-based model. It is intriguing to notice that three NRM variants
give suitable but quite distinct responses, with different perspectives and choices of words. This, as we
conjecture, is caused by both the architecture variations among models as well as the variations from
random effects like the initialization of parameters. Another interesting observation is on the forth example, where the retrieval-based method returns a response with the a mismatched entity name “WenShan”,
which is actually a quite common problem for retrieval-based model, where the inconsistency details
(e.g., dates, named entities), which often render the response unsuitable, cannot be adequately considered in the matching function employed in retrieving the responses. In contrast, we observe that NRMs
tend to make general response and barely generate those details.
Post
R1
ਹ㍖ㅢжཟθඐᤷօθፊᵫȾ
First day of being a vegetarian. Hold on, CuiDuoLa
ඐᤷօθඐᤷ㍖伕έ
Hold on, Keep up your vegetarian diet.
ᜩ㾷ڛᓭ੍ϋᜩ㾷䮵ስ੍ϋ䛙ቧީ⌞ᡇⲺᗤঐȾ
R2 Would you like to be healthy? Would you like to live
long? If so, follow me on Weibo.
R3
䘏ቧᱥՖ䈪ѣⲺਹ㍖੍ϋ
This is the so-called vegetarian diet ?
ਹ侧Ҽᢃᴿ࣑≊ࠅ㛛
R4 Eat up so that you will have enough energy to go for
a diet!
䓡։ᱥ䶟ળⲺᵢ䫧θу㾷ᙛέ
R5 Health is the essence for winning your future battles.
No rush!
Figure 9: Multiple responses generated by the NRM-hyb.
We also use the NRM-hyb as an example to investigate the ability of NRM to generate multiple
responses. Figure 9 lists 5 responses to the same post, which are gotten with beam search with beam size
= 500, among which we keep only the best one (biggest likelihood) for each first word. It can be seen
that the responses are fluent, relevant to the post, and still vastly different from each other, validating our
initial conjecture that NRM, when fueled with large and rich training corpus, could work as a generator
that can cover a lot of modes in its density estimation.
6 Conclusions and Future Work
In this paper, we explored using encoder-decoder-based neural network system, with coined name Neural
Responding Machine, to generate responses to a post. Empirical studies confirm that the newly proposed
NRMs, especially the hybrid encoding scheme, can outperform state-of-the-art retrieval-based and SMTbased methods. Our preliminary study also shows that NRM can generate multiple responses with great
variety to a given post. In future work, we would consider adding the intention (or sentiment) of users as
an external signal of decoder to generate responses with specific goals.
References
[Auli et al.2013] Michael Auli, Michel Galley, Chris Quirk, and Geoffrey Zweig. 2013. Joint language and translation modeling with recurrent neural networks. In EMNLP, pages 1044–1054.
[Bahdanau et al.2014] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation
by jointly learning to align and translate. arXiv preprint arXiv:1409.0473.
[Cho et al.2014] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and
Yoshua Bengio. 2014. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078.
[Chung et al.2014] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical
evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555.
[Fleiss1971] Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological
bulletin, 76(5):378.
[Graves2013] Alex Graves.
arXiv:1308.0850.
2013.
Generating sequences with recurrent neural networks.
preprint
[Hochreiter and Schmidhuber1997] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory.
Neural computation, 9(8):1735–1780.
[Howell2010] David C. Howell. 2010. Fundamental Statistics for the Behavioral Sciences. PSY 200 (300) Quantitative Methods in Psychology Series. Wadsworth Cengage Learning.
[Ji et al.2014] Zongcheng Ji, Zhengdong Lu, and Hang Li. 2014. An information retrieval approach to short text
conversation. arXiv preprint arXiv:1408.6988.
[Joachims2006] Thorsten Joachims. 2006. Training linear svms in linear time. In SIGKDD, pages 217–226. ACM.
[Kalchbrenner and Blunsom2013] Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent continuous translation
models. In EMNLP, pages 1700–1709.
[Koehn et al.2007] Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola
Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, et al. 2007. Moses: Open source toolkit
for statistical machine translation. In Proceedings of the 45th annual meeting of the ACL on interactive poster
and demonstration sessions, pages 177–180. ACL.
[Litman et al.2000] Diane Litman, Satinder Singh, Michael Kearns, and Marilyn Walker. 2000. Njfun: a reinforcement learning spoken dialogue system. In Proceedings of the 2000 ANLP/NAACL Workshop on Conversational
systems, pages 17–20. ACL.
[Mikolov et al.2010] Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernockỳ, and Sanjeev Khudanpur. 2010.
Recurrent neural network based language model. In INTERSPEECH 2010, pages 1045–1048.
[Misu et al.2012] Teruhisa Misu, Kallirroi Georgila, Anton Leuski, and David Traum. 2012. Reinforcement learning of question-answering dialogue policies for virtual museum guides. In SIGDIAL, pages 84–93. ACL.
[Och and Ney2003] Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical
alignment models. Computational linguistics, 29(1):19–51.
[Papineni et al.2002] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for
automatic evaluation of machine translation. In Proceedings of the 40th annual meeting on association for
computational linguistics, pages 311–318. Association for Computational Linguistics.
[Ritter et al.2011] Alan Ritter, Colin Cherry, and William B Dolan. 2011. Data-driven response generation in
social media. In EMNLP, pages 583–593. Association for Computational Linguistics.
[Schatzmann et al.2006] Jost Schatzmann, Karl Weilhammer, Matt Stuttle, and Steve Young. 2006. A survey
of statistical user simulation techniques for reinforcement-learning of dialogue management strategies. The
knowledge engineering review, 21(02):97–126.
[Sutskever et al.2014] Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. 2014. Sequence to sequence learning with
neural networks. In NIPS, pages 3104–3112.
[Wang et al.2013] Hao Wang, Zhengdong Lu, Hang Li, and Enhong Chen. 2013. A dataset for research on shorttext conversations. In EMNLP, pages 935–945.
[Williams and Young2007] Jason D Williams and Steve Young. 2007. Partially observable markov decision processes for spoken dialog systems. Computer Speech & Language, 21(2):393–422.
| 9cs.NE
|
Synchronization Strings: Efficient and Fast Deterministic
Constructions over Small Alphabets∗
Kuan Cheng†
Bernhard Haeupler
‡
Xin Li§
Amirbehshad Shahrasbi
¶
Ke Wuk
arXiv:1803.03530v1 [cs.IT] 8 Mar 2018
March 12, 2018
Abstract
Synchronization strings are recently introduced by Haeupler and Shahrasbi [15] in the study
of codes for correcting insertion and deletion errors (insdel codes). A synchronization string
is an encoding of the indices of the symbols in a string, and together with an appropriate
decoding algorithm it can transform insertion and deletion errors into standard symbol erasures
and corruptions. This reduces the problem of constructing insdel codes to the problem of
constructing standard error correcting codes, which is much better understood. Besides this,
synchronization strings are also useful in other applications such as synchronization sequences
and interactive coding schemes. For all such applications, synchronization strings are desired to
be over alphabets that are as small as possible, since a larger alphabet size corresponds to more
redundant information added.
Haeupler and Shahrasbi [15] showed that for any parameter ε > 0, synchronization strings of
arbitrary length exist over an alphabet whose size depends only on ε. Specifically, [15] obtained
an alphabet size of O(ε−4 ), which left an open question on where the minimal size of such
alphabets lies between Ω(ε−1 ) and O(ε−4
). In this work, we partially bridge this gap by
providing
an improved lower bound of Ω ε−3/2 , and an improved upper bound of O ε−2 . We also
provide fast explicit constructions of synchronization strings over small alphabets.
Further, along the lines of previous work on similar combinatorial objects, we study the
extremal question of the smallest possible alphabet size over which synchronization strings can
exist for some constant ε < 1. We show that one can construct ε-synchronization strings over
alphabets of size four while no such string exists over binary alphabets. This reduces the extremal
question to whether synchronization strings exist over ternary alphabets.
∗
Supported in part by NSF grants CCF-1527110, CCF-1618280, CCF-1617713 and NSF CAREER award CCF1750808.
†
[email protected]. Department of Computer Science, Johns Hopkins University.
‡
[email protected]. Department of Computer Science, Carnegie Mellon University
§
[email protected]. Department of Computer Science, Johns Hopkins University.
¶
[email protected]. Department of Computer Science, Carnegie Mellon University
k
[email protected]. Department of Computer Science, Johns Hopkins University.
1
1
Introduction
This paper focuses on the study of a combinatorial object called synchronization string. Intuitively, a
synchronization string is a (finite or infinite) string that avoids similarities between pairs of intervals
in the string. Such nice properties and synchronization strings themselves can actually be motivated
from at least two different aspects: coding theory and pattern avoidance. We now discuss the
motivations and previous work in each aspect below.
1.1
Motivation and Previous Work in Coding Theory
The general and most important goal of the coding theory is to obtain a reliable transmission
of information in the presence of noise or adversarial error. Starting from the pioneering works
of Shannon, Hamming, and many others, coding theory has evolved into an extensively studied
field, with applications found in various areas in computer science. Regarding the general goal of
correcting errors, we now have a very sophisticated and almost complete understanding of how
to deal with symbol erasures and corruptions. On the other hand, the knowledge of codes for
synchronization errors such as insertions and deletions, has lagged far behind despite also being
studied intensively since the 1960s. In practice, this is one of the main reasons why communication
systems require a lot of effort and resources to maintain synchronization strictly.
One major difficulty in designing codes for insertion and deletion errors is that in the received
codeword, the positions of the symbols may have changed. This is in contrast to standard symbol
erasures and corruptions, where the positions of the symbols always stay the same. Thus, many of
the known techniques in designing codes for standard symbol erasures and corruptions cannot be
directly utilized to protect against insertion and deletion errors.
In this context, a recent work of Haeupler and Shahrasbi [15] introduced synchronization strings,
which enable a black-box transformation of Hamming-type error correcting codes to codes that
protect against insertions and deletions. Informally, a synchronization string of length n is an
encoding of the indices of the n positions into a string over some alphabet Σ, such that, despite
some insertion and deletion errors, one can still recover the correct indices of many symbols. With
the correct indices of these symbols, a standard error correcting code can then be used to recover the
original message. This then gives a code for insertion and deletion errors, which is the combination
of a standard error correcting code and a synchronization string.
The simplest example of a synchronization string is just to record the index of each symbol,
i.e, the string 1, 2, · · · , n. It can be easily checked that even if (1 − ε) fraction of these indices are
deleted, one can still correctly recover the positions of the remaining εn symbols. However, this
synchronization string uses an alphabet whose size grows with the length of the string. The main
contribution of [15] is to show that under a slight relaxation, there exist synchronization strings of
arbitrary length n over an alphabet with fixed size. Further, [15] provided efficient and streaming
methods to correctly recover the indices of many symbols from a synchronization string after being
altered by insertion and deletion errors. Formally, [15] defines ε-synchronization strings as follows.
A string S is an ε-synchronization string if the edit distance of any two consecutive substrings S[i, j)
and S[j, k) is at least (1 − ε)(k − i).
Using the construction and decoding methods for ε-synchronization strings, [15] gives a code
that for any δ ∈ (0, 1) and ε > 0, can correct δ fraction of insertion and deletion errors with rate
1 − δ − ε. Besides this, synchronization strings have found a variety of applications, such as in
synchronization sequences [22], interactive coding schemes [8, 9, 11, 10, 13, 19, 18, 16], coding
against synchronization errors [16, 17], and edit distance tree codes [3, 18].
For all such applications, synchronization strings are desired to be over alphabets that are as small
2
as possible, since a larger alphabet size corresponds to more redundant information added. Thus a
natural question here is how small the alphabet size can be. In [15], Haeupler and Shahrasbi showed
that ε-synchronization strings with arbitrary length exist over an alphabet of size O(ε−4 ), they also
gave a randomized polynomial time algorithm to construct such strings. In a very recent work [16],
they further gave various efficient deterministicconstructions for finite/infinite ε-synchronization
strings, which have alphabets of size poly ε−1 for some unspecified large polynomials. On the
other hand, the definition of synchronization strings implies that any ε−1 consecutive symbols in an
ε-synchronization string have to be distinct—providing an Ω ε−1 lower-bound for the alphabet
size.
1.2
Motivation and Previous Work in Pattern Avoidance
Apart from applications in coding theory and other communication problems involving insertions
and deletions, synchronization strings are also interesting combinatorial objects from a mathematical
perspective. As a matter of fact, plenty of very similar combinatorial objects have been studied
prior to this work.
A classical work of Axel Thue [26] introduces and studies square-free strings, i.e., strings that do
not contain two identical consecutive substrings. Thue shows that such strings exist over alphabets
of size three and provides a fast construction of such strings using morphisms. The seminal work of
Thue inspired further works on the same problem [27, 21, 7, 24, 25, 28] and problems with a similar
pattern avoidance theme.
Krieger et. al. [20] study strings that satisfy relaxed variants of square-freeness, i.e., strings that
avoid approximate squares. Their study provides several results on strings that avoid consecutive
substrings of equal length with small additive or multiplicative Hamming distance in terms of their
length. In each of these regimes, [20] gives constructions of approximate square free strings over
alphabets with small constant size for different parameters.
Finally, Camungol and Rampersad [5] study approximate squares with respect to edit distance,
which is equivalent to the ε-synchronization string notion except that the edit distance property is
only required to hold for pairs of consecutive substrings of equal length. [5] employs a technique
based on entropy compression to prove that such strings exist over alphabets that are constant in
terms of string length but exponentially large in terms of ε−1 . We note that the previous result of
Haeupler and Shahrasbi [15] already improves this dependence to O(ε−4 ).
Again, a main question addressed in most of the above-mentioned previous work on similar
mathematical objects is how small the alphabet size can be.
1.3
Our Results
In this paper we study the question of how small the alphabet size of an ε-synchronization string
can be. We address this question both for a specified ε and
for unspecified ε. In the first case we
−4 provided in [15] and the lower bound of
try to bridge
the
gap
between
the
upper
bound
of
O
ε
Ω ε−1 . In the second case we study the question of how small the alphabet size can be to ensure
the existence of an ε-synchronization string for some constant ε < 1. In both cases we also give
efficient constructions that improve previous results.
1.3.1
New Bounds on Minimal Alphabet Size for a given ε
Our first theorem gives improved upper bound and lower bound for the alphabet size of an
ε-synchronization string for a given ε.
3
Theorem 1.1. For any 0 < ε < 1, there exists an alphabet Σ of size O ε−2 such that an infinite
ε-synchronization string exists over Σ. In addition, ∀n ∈ N, a randomized algorithm can construct
an ε-synchronization string of length n in expected time O(n5 log n). Further, the alphabet
size of
any ε-synchronization string that is long enough in terms of ε has to be at least Ω ε−3/2 .
Next, we provide efficient and even linear-time constructions of ε-synchronization strings over
drastically smaller alphabets than the efficient constructions in [16].
Theorem 1.2. For every n ∈ N and any constant ε ∈ (0, 1), there is a deterministic construction
of a (long-distance) ε-synchronization string of length n over an alphabet of size O(ε−2 ) that runs
in poly(n) time. Further, there is a highly-explicit linear time construction of such strings over an
alphabet of size O(ε−3 ).
Moreover, in Section 4.3, we present a method to construct infinite synchronization strings using
constructions for finite ones that only increases the alphabet size by a constant factor—as opposed
to the construction in [16] that increases the alphabet size quadratically.
Theorem 1.3. For any constant 0 < ε < 1, there exists an explicit construction of an infinite
ε-synchronization string S over an alphabet of size O(ε−2 ). Further, there exists a highly-explicit
construction of an infinite ε-synchronization string S over an alphabet of size O(ε−3 ) such that for
any i ∈ N, the first i symbols can be computed in O(i) time and S[i, i + log i] can be computed in
O(log i) time.
1.3.2
Minimal Alphabet Size for Unspecified ε: Three or Four?
One interesting question that has been commonly addressed by previous work on similar combinatorial objects is the size of the smallest alphabet over which one can find such objects. Along
the lines of [26, 28, 20, 5], we study the existence of synchronization strings over alphabets with
minimal constant size.
It is easy to observe that no such string can exist over a binary alphabet since any binary
string of length four either contains two consecutive identical symbols or two consecutive identical
substrings of length two. On the other hand, one can extract constructions over constant-sized
alphabets from the existence proofs in [15, 16], but the unspecified constants there would be quite
large. In Section 6.2, for some ε < 1, we provide a construction of arbitrarily long ε-synchronization
strings over an alphabet of size four. This narrows down the question to whether such strings exist
over alphabets of size three.
To construct such strings, we introduce the notion of weak synchronization string, which requires
substrings to satisfy a similar property as that of an ε-synchronization string, except that the lower
bound on edit distance is rounded down to the nearest integer. We show that weak synchronization
strings exist over binary alphabets and use one such string to modify a ternary square-free string
([26]) into a synchronization string over an alphabet of size four.
Finally, in Appendix A, we provide experimental evidence for the existence of synchronization
strings over ternary alphabets by finding lower-bounds for ε for which ε-synchronization strings
over alphabets of size 3, 4, 5, and 6 might exist. Similar experiments have been provided for related
combinatorial objects in the previous work [20, 5].
1.3.3
Constructing Synchronization Strings Using Uniform Morphisms
Morphisms have been widely used in previous work as a tool to construct similar combinatorial
objects. A uniform morphism of rank r over an alphabet Σ is a function φ : Σ → Σr that maps
4
any symbol of an alphabet Σ to a string of length r over the same alphabet. Using this technique,
some similar combinatorial objects in previous work have been constructed by taking a symbol
from the alphabet and then repeatedly using an appropriate morphism to replace each symbol
with a string [28, 20]. Here we investigate whether such tools can also be utilized to construct
synchronization strings. In Section 6.1, we show that no such morphism can construct arbitrarily
long ε-synchronization strings for any ε < 1.
2
Some Notations and Definitions
Usually we use Σ (probably with some subscripts) to denote the alphabet and Σ∗ to denote all
strings over alphabet Σ.
Definition 2.1 (Subsequence). The subsequence of a string S is any sequence of symbols obtained
from S by deleting some symbols. It doesn’t have to be continuous.
Definition 2.2 (Edit distance). For every n ∈ N, the edit distance ED(S, S 0 ) between two strings
S, S 0 ∈ Σn is the minimum number of insertions and deletions required to transform S into S 0 .
Definition 2.3 (Longest Common Subsequence). For any strings S, S 0 over Σ, the longest common
subsequence of S and S 0 is the longest pair of subsequence that are equal as strings. We denote by
LCS(S, S 0 ) the length of the longest common subsequence of S and S 0 .
Note that ED(S, S 0 ) = |S| + |S 0 | − 2LCS(S, S 0 ) where |S| denotes the length of S.
Definition 2.4 (ε-synchronization string). A string S is an ε-synchronization string if ∀1 ≤ i <
j < k ≤ |S| + 1, ED(S[i, j), S[j, k)) > (1 − ε)(k − i).
Definition 2.5 (square-free string). A string S is a square free string if ∀1 ≤ i < i + 2l ≤ |S| + 1,
(S[i, i + l) and S[i + l, i + 2l)) are different as words.
We also introduce the following generalization of a synchronization string, which will be useful
in our deterministic constructions of synchronization strings.
Definition 2.6 (ε-synchronization circle). A string S is an ε-synchronization circle if ∀1 ≤ i ≤ |S|,
Si , Si+1 , . . . , S| S|, S1 , S2 , . . . , Si−1 is an ε-synchronization string.
3
ε-synchronization Strings and Circles with Alphabet Size O(ε−2 )
In this section we show that by using a non-uniform sample space together with the Lovász Local
Lemma, we can have a randomized polynomial time construction of an ε-synchronization string with
alphabet size O(ε−2 ). We then use this to give a simple construction of an ε-synchronization circle
with alphabet size O(ε−2 ) as well. Although the constructions here are randomized, the parameter ε
can be anything in (0, 1) (even sub-constant), while our deterministic constructions in later sections
usually require ε to be a constant in (0, 1).
We first recall the General Lovász Local Lemma.
Lemma 3.1. (General Lovász Local Lemma) Let A1 , ..., An be a set of bad events. G(V, E) is a
dependency graph for this set of events if V = {1, . . . , n} and each event Ai is mutually independent
of all the events {Aj : (i, j) ∈
/ E}.
5
If there exists x1 , ..., xn ∈ [0, 1) such that for all i we have
Y
(1 − xj )
Pr(Ai ) ≤ xi
(i,j)∈E
Then the probability that none of these events happens is bounded by
Pr[
n
^
i=1
Āi ] ≥
n
Y
(1 − xi ) > 0
i=1
Using this lemma, we have the following theorem showing the existence of ε-synchronization
strings over an alphabet of size O(ε−2 ).
Theorem 3.2. ∀ε ∈ (0, 1) and ∀n ∈ N, there exists an ε-synchronization string S of length n over
alphabet Σ of size Θ(ε−2 ).
Proof. Suppose |Σ| = c1 ε−2 where c1 is a constant. Let t = c2 ε−2 and 0 < c2 < c1 . The sampling
algorithm is as follows:
1. Randomly pick t different symbols from Σ and let them be the first t symbols of S. If t ≥ n,
we just pick n different symbols.
2. For t + 1 ≤ i ≤ n, we pick the ith symbol S[i] uniformly randomly from Σ \ {S[i − 1], . . . , S[i −
t + 1]}
Now we prove that there’s a positive probability that S contains no bad interval S[i, k] which violates
the requirement that ED(S[i, j], S[j + 1, k]) > (1 − ε)(k − i) for any i < j < k. This requirement is
equivalent to LCS(S[i, j], S[j + 1, k]) < 2ε (k − i).
Notice that for k − i ≤ t, the symbols in S[i, k] are completely distinct. Hence we only need to
consider the case where k − i > t. First, let’s upper bound the probability that an interval is bad:
εl
l
Pr[interval I of length l is bad] ≤
(|Σ| − t)− 2
εl
εl
el εl
(|Σ| − t)− 2
εl p
ε |Σ| − t −εl
≤(
)
e
= C −εl
≤
The first inequality holds because if the interval is bad, then it has to contain a repeating sequence
a1 a2 . . . ap a1 a2 . . . ap where p is at least εl2 . Such sequence can be specified via choosing εl positions
in the interval and the probability that a given sequence is valid for the string in this construction
εl
is at most (|Σ| − t)− 2 . The second inequality comes from Stirling’s inequality.
The inequality above indicates that the probability that an interval of length l is bad can be
upper bounded by C −εl , where C is a constant and can be arbitrarily large by modifying c1 and c2 .
Now we use general Lovász local lemma to show that S contains no bad interval with positive
probability. First we’ll show the following lemma.
Claim. The badness of interval I = S[i, j] is mutually independent of the badness of all intervals
that do not intersect with I.
6
Proof. Suppose the intervals before I that do not
0 . We denote the indicator variables
are I10 , . . . , Im
0
is,
(
(
0 if I is not bad
0 if Ik
b=
, bk =
1 if I is bad
1 if Ik
intersect with I are I1 , . . . , Im , and those after I
of each interval being bad as b, bk and b0k0 . That
is not bad
,
is bad
b0k0
(
0 if Ik0 0 is not bad
=
1 if Ik0 0 is bad
First we prove that there exists p ∈ (0, 1) such that ∀x1 , x2 , . . . , xm ∈ {0, 1},
Pr[b = 1|bk = xk , k = 1, . . . , m] = p
According to our construction, we can see that for any fixed prefix S[1, i − 1], the probability
that I is bad is a fixed real number p0 . That is,
∀ valid S̃ ∈ Σi−1 , Pr[b = 1|S[1, i − 1] = S̃] = p0
This comes from the fact that, the sampling of the symbols in S[i, k] only depends on the previous
h = min{i − 1, t − 1} different symbols, and up to a relabeling these h symbols are the same h
symbols (e.g., we can relabel them as {1, · · · , h} and the rest of the symbols as {h + 1, · · · , |Σ|}). On
the other hand the probability that b = 1 remains unchanged under any relabeling of the symbols,
since if two sampled symbols are the same, they will stay the same; while if they are different, they
will still be different. Thus we have:
Pr[b = 1|bk = xk , i = 1, . . . , m]
Pr[b = 1, bk = xk , i = 1, . . . , m]
=
Pr[bk = xk , k = 1, . . . , m]
P
Pr[b = 1, S[1, i − 1] = S̃]
= S̃P
S̃ Pr[S[1, i − 1] = S̃]
X Pr[b = 1, S[1, i − 1] = S̃] Pr[S[1, i − 1] = S̃]
)
=
(
P
0
Pr[S[1, i − 1] = S̃]
S̃ 0 Pr[S[1, i − 1] = S̃ ]
S̃
X
Pr[S[1, i − 1] = S̃]
)
=
(Pr[b = 1|S[1, i − 1] = S̃] P
0
S̃ 0 Pr[S[1, i − 1] = S̃ ]
S̃
=p0
X
Pr[S[1, i − 1] = S̃]
0
S̃ 0 Pr[S[1, i − 1] = S̃ ]
P
S̃
0
=p
In the equations, S̃ indicates all valid string that prefix S[1, i − 1] can be such that bk = xk , k =
1, . . . , m. Hence, b is independent of {bk , k = 1, . . . , m}. Similarly, we can prove that the joint
distribution of {b0k0 , k 0 = 1, . . . , m0 } is independent of that of {b, bk , k = 1, . . . , m}. Hence b is
independent of {bk , b0k0 , k = 1, . . . , m, k 0 = 1, . . . , m0 }, which means, the badness of interval I is
mutually independent of the badness of all intervals that do not intersect with I.
Obviously, an interval of length l intersects at most l + l0 intervals of length l0 . To use Lovász
local lemma, we need to find a sequence of real numbers xi,k ∈ [0.1) for intervals S[i, k] for which
Y
Pr[S[i, k]is bad] ≤ xi,k
(1 − xi0 ,k0 )
S[i,k]∩S[i0 ,k0 ]6=∅
7
The rest of the proof is the same as that of Theorem 5.7 in [15].
We propose xi,k = D−ε(k−i) for some constant D ≥ 1. Hence we only need to find a constant D
such that for all S[i, k],
n
Y
C −ε(k−i) ≤ D−ε(k−i) [1 − D−εl ]l+(k−i)
l=t
That is, for all
l0
∈ {1, ..., n},
C
−l0
≤D
−l0
n
Y
l+l0
[1 − D−εl ] ε
l=t
which means that
D
C≥Q
n
−εl ]
l=t [1 − D
l/l0 +1
ε
Notice that the righthand side is maximized when n = ∞, l0 = 1. Hence it’s sufficient to show that
D
C≥Q
∞
l=t [1
Let L = maxD>1 Q
∞
D
−εl ]
l=t [1 − D
l+1
ε
− D−εl ]
. We only need to guarantee that C > L.
We claim that L = Θ(1). Since that t = c2 ε−2 = ω(
D
Q∞
l=t [1
l+1
ε
− D−εl ]
l+1
ε
log
ε
1
ε
),
D
l+1 −εl
]
l=t [1 − ε D
< Q∞
<
=
=
1−
D
P∞ l+1
1−
1
ε
1−
1 2tD−εt
ε (1−D−ε )2
l=t
ε
D−εl
D
−εl
l=t (l + 1)D
P∞
D
D
=
1
1−
2
D− ε
ε3 (1−D−ε )2
(1)
(2)
(3)
(4)
(5)
Inequality (1) comes from the fact that (1 − x)α > 1 − αx, (2) comes from he fact that
Q∞
P∞
P∞
xt (1+t−tx)
2txt
l
< (1−x)
2 , x < 1.
i=1 (1 − xi ) ≥ 1 −
i=1 xi and (3) is a result from
l=t (l + 1)x =
(1−x)2
−1
D ε
We can see that for D = 7, maxε { ε23 (1−D
−ε )2 } < 0.9. Therefore (5) is bounded by a constant,
which means L = Θ(1) and the proof is complete.
Using a modification of an argument in [15], we can also obtain a randomized construction.
Lemma 3.3. There exists a randomized algorithm which for any ε ∈ (0, 1) and any n ∈ N, constructs
an ε-synchronization string of length n over alphabet of size O(ε−2 ) in expected time O(n5 log n).
Proof. The algorithm is similar to that of Lemma 5.8 in [15], using algorithmic Lovász Local lemma
[23] and the extension in [14]. It starts with a string sampled according to the sampling algorithm
in the proof of Theorem 3.2, over alphabet Σ of size Cε−2 for some large enough constant C. Then
8
the algorithm checks all O(n2 ) intervals for a violation of the requirements for ε-synchronization
string. If a bad interval is found, this interval is re-sampled by randomly choosing every symbol
s.t. each one of them is different from the previous t − 1 symbols, where t = c0 ε−2 with c0 being a
constant smaller than C.
One subtle point of our algorithm is the following. Note that in order to apply the algorithmic
framework of [23] and [14], one needs the probability space to be sampled from n independent
random variables P = {P1 , · · · , Pn } so that each event in the collection A = {A1 , · · · , Am } is
determined by some subset of P. Then, when some bad event Ai happens, one only resamples the
random variables that decide Ai . Upon first look, it may appear that in our application of the Lovász
Local lemma, the sampling of the i’th symbol depends on the the previous h = min{i − 1, t − 1}
symbols, which again depend on previous symbols, and so on. Thus the sampling of the i’th symbol
depends on the sampling of all previous symbols. However, we can implement our sampling process
as follows: for the i’th symbol we first independently generate a random variable Pi which is uniform
over {1, 2, · · · , |Σ| − h}, then we use the random variables {P1 , · · · , Pn } to decide the symbols, in
the following way. Initially we fix some arbitrary order of the symbols in Σ, then for i = 1, · · · , n,
to get the i’th symbol, we first reorder the symbols Σ so that the previous h chosen symbols are
labeled as the first h symbols in Σ, and the rest of the symbols are ordered in the current order as
the last |Σ| − h symbols. We then choose the i’th symbol as the (h + Pi )’th symbol in this new
order. In this way, the random variables {P1 , · · · , Pn } are indeed independent, and the i’th symbol
is indeed chosen uniformly from the |Σ| − h symbols excluding the previous h symbols. Furthermore,
the event of any interval S[i, k] being bad only depends on the random variables (Pi , · · · , Pk ) since
no matter what the previous h symbols are, they are relabeled as {1, · · · , h} and the rest of the
symbols are labeled as {h + 1, · · · , |Σ|}. From here, the same sequence of (Pi , · · · , Pk ) will result in
the same behavior of S[i, k] in terms of which symbols are the same. We can thus apply the same
algorithm as in [15].
Note that the time to get the i’th symbol from the random variables {P1 , · · · , Pn } is O(n log 1ε )
since we need O(n) operations each on a symbol of size Cε−2 . Thus resampling each interval takes
O(n2 log 1ε ) time since we need to resample at most n symbols. For every interval, the edit distance
can be computed using the Wagner-Fischer dynamic programming within O(n2 log 1ε ) time. [14]
shows that the expected number of re-sampling is O(n). The algorithm will repeat until no bad
interval can be found. Hence the overall expected running time is O(n5 log 1ε ).
√
Note that without loss of generality we can assume that ε > 1/ n because for smaller errors we
can always use the indices directly, which have alphabet size n. So the overall expected running
time is O(n5 log n).
We can now construct an ε-synchronization circle using Theorem 3.2.
Theorem 3.4. For every ε ∈ (0, 1) and every n ∈ N, there exists an ε-synchronization circle S of
length n over alphabet Σ of size O(ε−2 ).
Proof. First, by Theorem 3.2, we can have two ε-synchronization strings: S1 with length d n2 e over
Σ1 and S2 with length b n2 c over Σ2 . Let Σ1 ∩ Σ2 = ∅ and |Σ1 | = |Σ2 | = O(ε−2 ). Let S be the
concatenation of S1 and S2 . Then S is over alphabet Σ = Σ1 ∪ Σ2 whose size is O(ε−2 ). Now we
prove that S is an ε-synchronization circle.
∀1 ≤ m ≤ n, consider string S 0 = sm , sm+1 , . . . , sn , s1 , s2 , . . . , sm−1 . Notice that for two strings
T and T 0 over alphabet Σ, LCS(T, T 0 ) ≤ 2ε (|T |+|T 0 |) is equivalent to ED(T, T 0 ) ≥ (1−ε)(|T |+|T 0 |).
For any i < j < k, we call an interval S 0 [i, k] good if LCS(S 0 [i, j], S 0 [j + 1, k]) ≤ 2ε (k − i). It suffices
to show that ∀1 ≤ i, k ≤ n, the interval S 0 [i, k] is good.
Without loss of generality let’s assume m ∈ [d n2 e, n].
9
Intervals which are substrings of S1 or S2 are good intervals, since S1 and S2 are ε-synchronization
strings.
We are left with intervals crossing the ends of S1 or S2 .
If S 0 [i, k] contains sn , s1 but doesn’t contain sd n2 e : If j < n − m + 1, then there’s no common
subsequence between s0 [i, j] and S 0 [n − m + 2, k]. Thus
ε
ε
LCS(S 0 [i, j], S 0 [j + 1, k]) ≤ LCS(S 0 [i, j], S 0 [j + 1, n − m + 1]) ≤ (n − m + 1 − i) < (k − i)
2
2
If j ≥ n − m + 1, then there’s no common subsequence between S 0 [j + 1, k] and S 0 [i, n − m + 1].
Thus
ε
ε
LCS(S 0 [i, j], S 0 [j + 1, k]) ≤ LCS(S 0 [n − m + 2, j], S 0 [j + 1, k]) ≤ (k − (n − m + 2)) < (k − i)
2
2
Thus intervals of this kind are good.
Figure 1: Example where S 0 [i, k] contains sn , s1 but doesn’t contain sd n2 e
If S 0 [i, k] contains sb n2 c , sd n2 e but doesn’t contain sn : If j ≤ n − m + b n2 c + 1, then there’s
no common subsequence between S 0 [i, j] and S 0 [n − m + d n2 e + 1, k], thus
ε
n
LCS(S 0 [i, j], S 0 [j + 1, k]) ≤ LCS(S 0 [i, j], S 0 [j + 1, n − m + b c + 1]) < (k − i)
2
2
If j ≥ n−m+b n2 c+1, then there’s no common subsequence between S 0 [j+1, k] and S 0 [i, n−m+b n2 c+1].
Thus
n
ε
LCS(S 0 [i, j], S 0 [j + 1, k]) ≤ LCS(S 0 [n − m + d e + 1, j], S 0 [j + 1, k]) < (k − i)
2
2
10
Thus intervals of this kind are good.
Figure 2: Example where S 0 [i, k] contains sb n2 c , sd n2 e
If S 0 [i, k] contains sd n2 e and sn : If n − m + 2 ≤ j ≤ n − m + b n2 c + 1, then the common
subsequence is either that of S 0 [i, n − m + 1] and S 0 [n − m + d n2 e + 1, k] or that of S 0 [n − m + 2, j]
and S 0 [j + 1, n − m + b n2 c + 1]. This is because Σ1 ∩ Σ2 = ∅. Thus
LCS(S 0 [i, j], S 0 [j + 1, k])
n
≤ max{LCS(S 0 [i, n − m + 1], S 0 [n − m + d e + 1, k]),
2
n
0
0
LCS(S [n − m + 2, j], S [j + 1, n − m + b c + 1])}
2
ε
< (k − i)
2
11
Figure 3: Example where S 0 [i, k] contains sd n2 e and sn
If j ≤ n − m + 1, then there’s no common subsequence between S 0 [i, j] and S 0 [n − m + 2, n −
m + b n2 c + 1]. Thus
LCS(S 0 [i, j], S 0 [j + 1, k])
n
≤LCS(S 0 [i, j], S 0 [j + 1, n − m + 1]) + LCS(S 0 [i, j], S 0 [n − m + d e + 1, k])
2
ε
n
ε
< (n − m + 1 − i) + (n − d e)
2
2
2
ε
ε
≤ (n − m + 1 − i) + (k − (n − m + 2))
2
2
ε
= (k − 1 − i)
2
ε
< (k − i)
2
If j ≥ S 0 [n − m + d n2 e + 1], the proof is similar to the case where j ≤ n − m + 1.
This shows that S 0 is an -synchronization string. Thus by the definition of synchronization
circle, the construction gives an -synchronization circle.
4
Deterministic Constructions of Long-Distance Synchronization
Strings
In this section, we give deterministic constructions of synchronization strings. In fact, we consider
a generalized version of synchronization strings, i.e., f (l)-distance ε-synchronization strings first
defined by Haeupler and Shahrasbi [16]. Throughout this section, ε is considered to be a constant
in (0, 1).
12
Definition 4.1 (f (l)-distance ε-synchronization string). A string S ∈ Σn is an f (l)-distance εsynchronization string if for every 1 ≤ i < j ≤ i0 < j 0 ≤ n + 1, ED (S[i, j), S[i0 , j 0 )) > (1 − ε)(l) for
i0 − j ≤ f (l) where l = j + j 0 − i − i0 .
As a special case, 0-distance synchronization strings are standard synchronization strings.
Similar to [16], we focus on f (l) = n · 1l>c log n where 1l>c log n is the indicator function for
l > c log n. This function considers the edit distance of all pairs of large intervals and adjacent small
intervals.
Definition 4.2 (c-long-distance ε-synchronization strings). We call n·1l>c log n -distance ε-synchronization
strings c-long-distance ε-synchronization strings.
4.1
Polynomial Time Constructions of Long-Distance Synchronization Strings
Here, by combining the deterministic Lovász local lemma of Chandrasekaran et al. [6] and the
non-uniform sample space used in Theorem 3.2, we give a deterministic polynomial-time construction
of c-long ε-synchronization strings over an alphabet of size O(ε−2 ).
First we recall the following property of c−long synchronization strings.
Lemma 4.3 (Corollary 4.4 of [16]). If S is a string which satisfies the c-long-distance ε-synchronization
property for any two non-adjacent intervals of total length 2c log n or less, then it satisfies the property
for all pairs of non-adjacent intervals.
We now have the following theorem.
Theorem 4.4. For any n ∈ N and any constant ε ∈ (0, 1), there is a deterministic construction
of a O(1/ε)-long-distance ε-synchronization string of length n, over an alphabet of size O(ε−2 ), in
time poly(n).
Proof. To prove this, we will use the Lovśz Local Lemma and its deterministic algorithm in [6].
Suppose the alphabet is Σ with |Σ| = q = c1 ε−2 where c1 is a constant. Let t = c2 ε−2 and
0 < c2 < c1 . We denote |Σ| − t as q. The sampling algorithm of string S (1-index based)is as follows:
• Initialize an arbitrary order for Σ.
• For ith symbol:
– Denote h = min{t−1, i−1}. Generate a random variable Pi uniformly over {1, 2, . . . , |Σ|−
h}.
– Reorder Σ such that the previous h chosen symbols are labeled as the first h symbols in
Σ, and the rest are ordered in the current order as the last |Σ| − h symbols.
– Choose the (Pi + h)’th symbol in this new order as S[i].
Define the bad event Ai1 ,l1 ,i2 ,l2 as intervals S[i1 , i1 + l1 ) and S[i2 , i2 + l2 ) violating the c = O(1/ε)long-distance synchronization string property for i1 + l1 ≤ i2 . In other words, Ai1 ,l1 ,i2 ,l2 occurs if
and only if ED(S[i1 , i1 + l1 ), S[i2 , i2 + l2 )) ≤ (1 − ε)(l1 + l2 ), which is equivalent to LCS(S[i1 , i1 +
l1 ), S[i2 , i2 + l2 )) ≥ 2ε (l1 + l2 ).
Note that according to the definition of c-long distance ε-synchronization string and Lemma 4.3,
we only need to consider Ai1 ,l1 ,i2 ,l2 where l1 + l2 < c log n and c log n ≤ l1 + l2 ≤ 2c log n. Thus we
can upper bound the probability of Ai1 ,l1 ,i2 ,l2 ,
ε(l1 +l2 )
l1
l2
Pr [Ai1 ,l1 ,i2 ,l2 ] ≤
(|Σ| − t)− 2
ε(l1 + l2 )/2 ε(l1 + l2 )/2
13
≤
=
≤
ε(l1 +l2 )/2
ε(l1 +l2 )/2
ε(l1 +l2 )
l2 e
l1 e
(|Σ| − t)− 2
ε(l1 + l2 )/2
ε(l1 + l2 )/2
!ε(l1 +l2 )
√
2e l1 l2
p
ε(l1 + l2 ) |Σ| − t
!εl
!εl
el
e
p
p
=
= Ĉ εl ,
εl |Σ| − t
ε |Σ| − t
where l = l1 + l2 and Ĉ is a constant which depends on c1 and c2 .
However, to apply the deterministic Lovśz Local Lemma (LLL), we need to have two additional
requirements. The first requirement is that each bad event depends on up to logarithmically many
variables, and the second is that the inequalities in the Lovśz Local Lemma hold with a constant
exponential slack [16].
The first requirement may not be true under the current definition of badness. Consider for
example the random variables Pi1 , . . . , Pi1 +l1 −1 , Pi2 , Pi2 +l2 −1 for a pair of split intervals S[i1 , i1 +
l1 ), S[i2 , i2 + l2 ) where the total length l1 + l2 is at least 2c log n. The event Ai1 ,l1 ,i2 ,l2 may depend
on too many random variables (i.e., Pi1 , . . . , Pi2 +l2 −1 ).
To overcome this, we redefine the badness of the split interval S[i1 , i1 + l1 ) and S[i2 , i2 + l2 )
as follows: let Bi1 ,l1 ,i2 ,l2 be the event that there exists Pi1 +l1 , . . . , Pi2 −1 (i.e., the random variables
chosen between the two intervals) such that the two intervals generated by Pi1 . . . , Pi1 +l1 −1 and
Pi2 , . . . , Pi2 +l2 −1 (together with Pi1 +l1 , . . . , Pi2 −1 ) makes LCS(S[i1 , i1 + l1 ), S[i2 , i2 + l2 )) ≥ 2ε (l1 + l2 )
according to the sampling algorithm. Note that if Bi1 ,l1 ,i2 ,l2 does not happen, then certainly Ai1 ,l1 ,i2 ,l2
does not happen.
Notice that with this new definition of badness, Bi1 ,l1 ,i2 ,l2 is independent of {Pi1 +l1 , . . . , Pi2 −1 }
and only depends on {Pi1 . . . , Pi1 +l1 −1 , Pi2 , . . . , Pi2 +l2 }. In particular, this implies that Bi1 ,l1 ,i2 ,l2
is independent of the badness of all other intervals which have no intersection with (S[i1 , i1 + l1 ),
S[i2 , i2 + l2 )).
We now bound Pr[Bi1 ,l1 ,i2 ,l2 ]. When considering the two intervals S[i1 , i1 + l1 ), S[i2 , i2 + l2 ) and
their edit distance under our sampling algorithm, without loss of generality we can assume that the
order of the alphabet at the point of sampling S[i1 ] is (1, 2, . . . , q) just by renaming the symbols.
Now, if we fix the order of the alphabet at the point of sampling S[i2 ] in our sampling algorithm,
then S[i2 , i2 + l2 ) only depends on {Pi2 , . . . , Pi2 +l2 } and thus LCS(S[i1 , i1 + l1 ), S[i2 , i2 + l2 )) only
depends on {Pi1 . . . , Pi1 +l1 −1 , Pi2 , . . . , Pi2 +l2 }.
Conditioned on any fixed order of the alphabet at the point of sampling S[i2 ], we have that
LCS(S[i1 , i1 + l1 ), S[i2 , i2 + l2 )) ≥ 2ε (l1 + l2 ) happens with probability at most Ĉ εl by the same
computation as we upper bound Pr[Ai1 ,l1 ,i2 ,l2 ]. Note that there are at most q! different orders of the
alphabet. Thus by a union bound we have
Pr[Bi1 ,l1 ,i2 ,l2 ] ≤ Ĉ εl × q! = C εl ,
for some constant C.
In order to meet the second requirement of the deterministic algorithm of LLL, we also need to
find real numbers xi1 ,i1 +l1 ,i2 ,i2 +l2 ∈ [0, 1] such that for any Bi1 ,l1 ,i2 ,l2 ,
1.01
Y
Pr[Bi1 ,l1 ,i2 ,l2 ] ≤ xi1 ,l1 ,i2 ,l2
(1 − xi01 ,l10 ,i02 ,l20 )
[S[i1 ,i1 +l1 )∪S[i2 ,i2 +l2 )]∩[S[i01 ,i01 +l10 )∪S[i02 ,i02 +l20 )]6=∅
14
.
We propose xi1 ,l1 ,i2 ,l2 = D−ε(l1 +l2 ) for some D > 1 to be determined later. D has to be chosen such
that for any i1 , l1 , i2 , l2 and l = l1 + l2 :
e
p
ε |Σ|
1.01
!εl
D−εl
≤
Y
1−D
−ε(l10 +l20 )
(6)
[S[i1 ,i1 +l1 )∪S[i2 ,i2 +l2 )]∩[S[i01 ,i01 +l10 )∪S[i02 ,i02 +l20 )]6=∅
Notice that
Y
D−εl
0
0
1 − D−ε(l1 +l2 )
(7)
[S[i1 ,i1 +l1 )∪S[i2 ,i2 +l2 )]∩[S[i01 ,i01 +l10 )∪S[i02 ,i02 +l20 )]6=∅
≥ D
−εl
l0
Y
2cY
log n
1 − D−εl
0
[(l1 +l10 )+(l1 +l20 )+(l2 +l10 )+(l2 +l20 )]n
l0 =c log n l10 =1
×
= D
cY
log n
1 − D−εl
l00 =t
2cY
log n
−εl
l+l00
00
l0
Y
(8)
1 − D−εl
0
4(l+l0 )n
cY
log n
×
= D
−εl
1−D
−εl0
4l0 (l+l0 )n
×
"c log n
Y
(9)
1−D
−εl00
#l
×
cY
log n
1 − D−εl
00
l00
(10)
l00 =t
2cX
log n
≥ D−εl 1 −
l+l00
l00 =t
l0 =c log n
00
l00 =t
l0 =c log n l10 =1
2cY
log n
1 − D−εl
4l0 (l + l0 )n D
−εl0
l0 =c log n
"
× 1−
cX
log n
#l
D
−εl00
×
1−
l00 =t
≥ D−εl 1 −
cX
log n
!
00
l D
−εl00
(11)
l00 =t
2cX
log n
0
(4 · 2c log n(2c log n + 2c log n)n) D−εl
(12)
l0 =c log n
"
× 1−
∞
X
#l
D
−εl00
×
1−
l00 =t
∞
X
!
l00 D
−εl00
(13)
l00 =t
2cX
log n
"
−1
−c2 ε
× 1− D
= D 1 −
32c n log n D
1 − D−ε
l0 =c log n
!
D−c2 /ε (D−ε + c2 /ε2 − c2 D−ε /ε2 )
× 1−
(1 − D−ε )2
"
#l
−c2 ε−1
D
≥ D−εl 1 − 32c3 n log3 nD−εc log n 1 −
1 − D−ε
!
D−c2 /ε (D−ε + c2 /ε2 − c2 D−ε /ε2 )
× 1−
(1 − D−ε )2
−εl
2
2
#l
−εl0
15
(14)
(15)
Equation 8 holds because there are two kinds of pairs of intervals. The first kind contains all
pairs of intervals whose total length is between c log n and 2c log n intersecting with S[i1 , i1 + l1 ) or
S[i2 , i2 + l2 ). The number of such pairs is at most (l1 + l10 ) + (l1 + l20 ) + (l2 + l10 ) + (l2 + l20 ). The
second kind contains all adjacent intervals of total length less than c log n. Notice that according
to our sampling algorithm, every t consecutive symbols are distinct, thus any adjacent intervals
whose total length is less than t cannot be bad. Hence the second term contains intervals such that
t ≤ l00 = l100 + l200 ≤ c log n.
The rest of the proof is the same as that of Theorem 4.5 in [16].
Equation 11 comes from the fact that for 0 < x, y < 1:
(1 − x)(1 − y) > 1 − x − y
For D = 2 and c = 2/ε,
2−c2 /ε
=0
ε→0 1 − 2−ε
lim
Thus, for sufficiently small ε,
2−c2 /ε
1−2−ε
< 21 . Moreover,
0
32c2 n log2 nD−εl =
−c2 /ε
28 log3 n
= o(1)
ε3 n
−ε
2
−ε
2
(D +c2 /ε −c2 D /ε )
Finally, for sufficiently small ε, 1 − D
> 2−ε . Therefore, for sufficiently
(1−D−ε )2
small ε and sufficiently large n, 15 is satisfied under the condition:
Y
0
0
D−εl
1 − D−ε(l1 +l2 )
[S[i1 ,i1 +l1 )∪S[i2 ,i2 +l2 )]∩[S[i01 ,i01 +l10 )∪S[i02 ,i02 +l20 )]6=∅
1
1
4−εl
≥ 2−εl (1 − )(2−ε )l (1 − ) ≥
2
2
4
So for LLL to work, the following should be guaranteed:
e
p
ε |Σ| − t
!
εl
1.01
2
4−εl
42.02(1+ε)e
≤
⇐
≤ |Σ| − t
4
ε2
Hence the second requirement holds for |Σ| − t =
44.04 e2
ε2
= O(ε−2 ).
Corollary 4.4.1. For any n ∈ N and any constant ε ∈ (0, 1), there is a deterministic construction
of an ε-synchronization string of length n, over an alphabet of size O(ε−2 ), in time poly(n).
By a similar concatenation construction used in the proof of Theorem 3.4, we also have a
deterministic construction for synchronization circles.
Corollary 4.4.2. For any n ∈ N and any constant ε ∈ (0, 1), there is a deterministic construction
of an ε-synchronization circle of length n, over an alphabet of size O(ε−2 ), in time poly(n).
4.2
Deterministic linear time constructions of c-long distance ε-synchronization
string
Here we give a much more efficient construction of a c-long distance ε-synchronization string, using
synchronization circles and standard error correcting codes. We show that the following algorithm
gives a construction of c-long distance synchronization strings.
16
Algorithm 1 Explicit Linear Time Construction of c-long distance ε-synchronization string
Input:
• An ECC Cˆ ⊂ Σm
, with distance δm and block length m = c log n.
Ĉ
• An ε0 -synchronization circle SC = (sc1 , . . . , scm ) of length m over alphabet ΣSC .
Operations:
• Construct a code C ⊂ Σm such that
ˆ
C = {((ĉ1 , sc1 ), . . . , (ĉm , scm ))|(ĉ1 , . . . , ĉm ) ∈ C}
where Σ = ΣĈ × ΣSC .
• Let S be concatenation of all codewords C1 , . . . , CN from C.
Output: S.
To prove the correctness, we first recall the following theorem from [15].
Theorem 4.5 (Theorem 4.2 of [15]). Given an ε0 -synchronization string S with length n, and
0
an efficient ECC C with block length n, that corrects up to nδ 1+ε
1−ε0 half-errors, one can obtain an
0
insertion/deletion code C that can be decoded from up to nδ deletions, where C 0 = {(c01 , . . . , c0n )|∀i ∈
[n], c0i = (ci , S[i]), (c1 , . . . , cn ) ∈ C}.
We have the following property of longest common subsequence.
Lemma 4.6. Suppose T1 is the concatenation of `1 strings, T1 = S1 ◦ · · · ◦ S`1 and T2 is the
concatenation of `2 strings, T2 = S10 ◦ · · · ◦ S`0 2 . If there exists an integer t such that for all i, j, we
have LCS(Si , Sj0 ) ≤ t, then we have LCS(T1 , T2 ) ≤ (`1 + `2 )t.
Proof. We rename the strings in T2 by S`1 +1 , · · · , S`1 +`2 . Suppose the longest common subsequence
between T1 and T2 is T̃ , which can be viewed as a matching between T1 and T2 .
we can divide T̃ sequentially into disjoint intervals, where each interval corresponds to a common
subsequence between a different pair of strings (Si , Sj ), where Si is from T1 and Sj is from T2 . In
addition, if we look at the intervals from left to right, then for any two consecutive intervals and
their corresponding pairs (Si , Sj ) and (Si0 , Sj 0 ), we must have i0 ≥ i and j 0 ≥ j since the matchings
which correspond to two intervals cannot cross each other. Furthermore either i0 > i or j 0 > j as
the pair (Si , Sj ) is different from (Si0 , Sj 0 ).
Thus, starting from the first interval, we can label each interval with either i or j such that every
interval receives a different label, as follows. We label the first interval using either i or j. Then,
assuming we have already labeled some intervals and now look at the next interval. Without loss of
generality assume that the previous interval is labeled using i, now if the current i0 > i then we can
label the current interval using i0 ; otherwise we must have j 0 > j so we can label the current interval
using j 0 . Thus the total number of the labels is at most l1 + l2 , which means the total number of
the intervals is also at most l1 + l2 . Note that each interval has length at most t, therefore we can
upper bound LCS(T1 , T2 ) by (l1 + l2 )t.
1−ε0
Lemma 4.7. The output S in Algorithm 1 is an ε1 -synchronization circle, where ε1 ≤ 10(1 − 1+ε
δ).
0
17
Proof. Suppose Cˆ can correct up to δm half-errors. Then according to lemma 4.5, C can correct up
0
to 1−ε
1+ε0 δm deletions.
0
Let α = 1 − 1−ε
1+ε0 δ. Notice that C has the following properties:
1. LCS(C) = maxc1 ,c2 ∈C LCS(c1 , c2 ) ≤ αm
2. Each codeword in C is an ε-synchronization circle over Σ.
Consider any shift of the start point of S, we only need to prove that ∀1 ≤ i < j < k ≤
n, LCS(S[i, j], S[j + 1, k]) < ε21 (k − i).
Suppose S1 = S[i, j] and S2 = S[j + 1, k]. Let ε1 = 10α.
Case 1: k − i > m. Let |S1 | = s1 and |S2 | = s2 , thus s1 + s2 > m. If we look at each Sh for
h = 1, 2, then Sh can be divided into some consecutive codewords, plus at most two incomplete
codewords at both ends. In this sense each Sh is the concatenation of `h strings with `h < smh + 2.
An example of the worst case appears in Figure 4.
Figure 4: Example of the worst case, where j splits a codeword, and there are two incomplete
codewords at both ends.
Now consider the longest common subsequence between any pair of these strings where one is
from S1 and the other is from S2 , we claim that the length of any such longest common subsequence
is at most αm. Indeed, if the pair of strings are from two different codewords, then by the property
of the code C we know the length is at most αm. On the other hand, if the pair of strings are from
a single codeword (this happens when j splits a codeword, or when S[i] and S[k] are in the same
codeword), then they must be two disjoint intervals within a codeword. In this case, by the property
that any codeword is also a synchronization circle, the length of the longest common subsequence of
this pair is at most ε20 m.
2ε0
ε0
1−ε0
0
Note that α = 1 − 1−ε
1+ε0 δ ≥ 1 − 1+ε0 = 1+ε0 ≥ ε0 (since δ, ε0 ∈ (0, 1)). Thus 2 m < αm.
Therefore, by Lemma 4.6, we have
LCS(S1 , S2 )
s1
s2
<( + 2 +
+ 2)αm
m
m
=α(s1 + s2 + 4m)
(16)
<5α(s1 + s2 )
ε1
(k − i)
2
Case 2: If k − i ≤ m, then according to the property of synchronization circle SC, we know
that the longest common subsequence of S1 and S2 is less than ε20 (k − i) ≤ α(k − i) ≤ ε21 (k − i).
As a result, the longest common subsequence of S[i, j] and S[j + 1, k] is less than ε21 (k − i),
which means that S is an ε1 -synchronization circle.
=5α(k − i) =
Similarly, we also have the following lemma.
18
Lemma 4.8. The output S of algorithm 1 is a c-long distance ε-synchronization string of length
0
n = N m where N is the number of codewords in C, ε = 12(1 − 1−ε
1+ε0 δ).
Proof. By Lemma 4.7, S is an ε1 -synchronization string, thus the length of longest common
subsequence for adjacent intervals S1 , S2 with total length l < c log n is less than ε21 l. We only need
to consider pair of intervals S1 , S2 whose total length l ∈ [c log n, 2c log n].
Notice that the total length of S1 and S2 is at most 2c log n, which means that S1 and S2 each
intersects with at most 3 codewords from C. Using Lemma 4.6, we have that LCS(S1 , S2 ) ≤ 6αl.
0
Thus picking ε = max{12α, ε1 } = 12α = 12(1 − 1−ε
1+ε0 δ), S from algorithm 1 is a c-long distance
ε-synchronization circle.
We need the following code constructed by Guruswarmi and Indyk [12].
Lemma 4.9 (Theorem 3 of [12]). For every 0 < r < 1, and all sufficiently small ε > 0, there exists
−4 −1
a family of codes of rate r and relative distance (1 − r − ε) over an alphabet of size 2O(ε r log(1/ε))
such that codes from the family can be encoded in linear time and can also be uniquely decoded in
linear time from 2(1 − r − ε) fraction of half errors.
Lemma 4.10 (ECC by Brute Force Search). For any n ∈ N, any ε ∈ [0, 1], one can construct
n
εn
a ECC in time O(2εn ( 2e
ε ) n log(1/ε)) and space O(2 n log(1/ε)), with block length n, number of
εn
codewords 2 , distance d = (1 − ε)n, alphabet size 2e/ε.
We can now use Algorithm 1 to give a linear time construction of c − long − distance εsynchronization strings.
Theorem 4.11. For every n ∈ N and any constant 0 < ε < 1, there is a deterministic construction
of a c = O(ε−2 ) − long − distance ε-synchronization string S ∈ Σn where |Σ| = O(ε−3 ), in time
n
O(n). Moreover, S[i, i + log n] can be computed in O( log
) time.
ε2
Proof. Suppose we have an error correcting code Cˆ with distance rate 1 − ε0 =
ε
1+ 36
ε
1− 36
ε
(1 − 12
), message
rate rc = O(ε02 ), over an alphabet of size |Σc | = O(ε0−1 ), with block length m = O(ε0−2 log n).
ε
Let c = O(ε0−2 ) = O(ε−2 ). We apply Algorithm 1, using Ĉ and an 36
-synchronization circle
−2
SC of length m over an alphabet of size O(ε ). Here SC is constructed by Corollary 4.4.2 in
1− ε
time poly(m) = poly(log n). By Lemma 4.8, we have a c-long-distance 12(1 − 1+ 36ε (1 − ε0 )) = ε36
synchronization string of length m · |Σc |rc m ≥ n.
It remains to show that we can have such a Cˆ with linear time encoding. We use the code in
Lemma 4.9 as the outer code and the one in Lemma 4.10 as inner code. Let Cout be an instantiation
of the code in Lemma 4.9 with rate ro = εo = 13 ε0 , relative distance do = (1 − 2εo ) and alphabet size
−5
4
εo log n
2O(εo log(1/εo )) , and block length no = log(1/ε
, which is encodable and decodable in linear time.
o)
Further, according to Lemma 4.10 one can find a code Cin with rate ri = O(εi ) where εi = 13 ε0 ,
−6
relative distance 1 − εi , over an alphabet of size 2e
εi , and block length ni = O(εi log(1/εi )). Note
that since the block length and alphabet size are both constant because ε is a constant. So the
encoding can be done in constant time.
Concatenating Cout and Cin gives the desire code Cˆ with rate O(ε02 ), distance 1 − O(ε0 ) and
alphabet of size O(ε0−1 ) and block length O(ε0−2 log n). Moreover, the encoding of Cˆ can be done in
linear time, because the encoding of Cout is in linear time and the encoding of Cin is in constant
time.
n
Note that since every codeword of Cˆ can be computed in time O( log
), S[i, i + log n] can be
ε2
log n
computed in O( ε2 ) time.
19
Corollary 4.11.1. For every n ∈ N and any constant 0 < ε < 1, there is a deterministic
construction of an ε-synchronization string S ∈ Σn where |Σ| = O(ε−3 ), in time O(n). Moreover,
n
S[i, i + log n] can be computed in O( log
) time.
ε2
4.3
Explicit Constructions of Infinite Synchronization Strings
In this section we focus on the construction of infinite synchronization strings.
The definition of infinite synchronization strings is similar to that of standard synchronization
strings except that the length of the string is infinite.
Definition 4.12 (Infinite ε-synchronization string). A string S is an infinite ε-synchronization
string if it has infinite length and ∀1 ≤ i < j < k, ED(S[i, j), S[j, k)) > (1 − ε)(k − i).
To measure the efficiency of the construction of an infinite string, we consider the time complexity
for computing the first n elements of that string. An infinite synchronization string is said to have
an explicit construction if there is an algorithm that computes any position S[i] in time poly(i).
Moreover, it is said to have a highly-explicit construction if there is an algorithm that computes any
position S[i] in time O(log i).
We have the following algorithm.
Algorithm 2 Construction of infinite ε-synchronization string
Input:
• A constant ε ∈ (0, 1).
Operations:
• Let q ∈ N be the size of an alphabet large enough to construct an 2ε -synchronization string.
Let Σ1 and Σ2 be two alphabets of size q such that Σ1 ∩ Σ2 = ∅.
• Let k = 4ε . For i = 1, 2, . . ., construct an 2ε -synchronization string Ski of length k i , where
Ski is over Σ1 if i is odd and over Σ2 otherwise.
• Let S be the sequential concatenation of Sk , Sk2 , Sk3 , . . . , Skt , . . .
Output: S.
Lemma 4.13. If there is a construction of 2ε -synchronization strings with alphabet size q, then
Algorithm 2 constructs an infinite ε-synchronization string owith alphabet size 2q.
Proof. Algorithm 2 can be shown as in the figure below.
Figure 5: Sk and Sk3 are over alphabet Σ1 and Sk2 is over Σ2 .
Now we show that S is an infinite ε-synchronization string.
Claim. Let x < y < z be positive integers and let t be such that k t ≤ |S[x, z)| < k t+1 . Then
ED(S[x, y), S[y, z)) ≥ (1 − 2ε )(z − x)(1 − k2 ).
20
P
i+1
Proof. Let li be the index of S where Ski+1 starts. Then li = ij=1 k j = k k−1−k . Notice that
lt−1 < 2k t−1 and |S[x, z)| ≥ k t , one can throw away all elements of S[x, z) whose indices are less
t−1
than lt−1 without losing more than 2kkt = k2 fraction of the elements of S[x, z). We use S[x0 , z) to
denote the substring after throwing away the symbols before lt−1 . Thus x0 ≥ lt−1 .
Since x0 ≥ lt−1 , S[x0 , z) either entirely falls into a synchronization string Skl or crosses two
synchronization strings Skl and Skl+1 over two entirely different alphabets Σ1 and Σ2 . Thus the
edit distance of S[x0 , y) and S[y, z) is at least (1 − 2ε )(z − x).
Since k = 4ε , we have that
ε 2
2
ε
= 1−
(z − x) ≥ (1 − ε)(z − x).
(z − x) 1 −
ED(S[x, y), S[y, z)) ≥ 1 −
2
k
2
This shows that S is an ε-synchronization string.
If we instantiate Algorithm 2 using Corollary 4.4.1, then we have the following theorem.
Theorem 4.14. For any constant 0 < ε < 1, there exists an explicit construction of an infinite
ε-synchronization string S over an alphabet of size O(ε−2 ).
Proof. We combine Algorithm 2 and Corollary 4.4.1. In the algorithm, we can construct every
substring Ski in polynomial time with alphabet size q = O(ε−2 ), by Corollary 4.4.1. So the first n
symbols of S can be computed in polynomial time.
By Lemma 4.13, S is an infinite ε-synchronization string over an alphabet of size 2q = O(ε−2 ).
If we instantiate Algorithm 2 using Corollary 4.11.1, then we have the following theorem.
Theorem 4.15. For any constant 0 < ε < 1, there exists a highly-explicit construction of an infinite
ε-synchronization string S over an alphabet of size O(ε−3 ). Moreover, for any i ∈ N, the first i
symbols can be computed in O(i) time and S[i, i + log i] can be computed in O(log i) time.
Proof. Combine Algorithm 2 and Corollary 4.11.1. In the algorithm, we can construct every
substring Ski in linear time with alphabet size q = O(ε−3 ), by Corollary 4.11.1. So the first i
symbols can be computed in O(i) time. Also any substring S[i, i + log i] can be computed in time
O(log i).
By Lemma 4.13, S is an infinite ε-synchronization string over an alphabet of size 2q = O(ε−3 ).
5
Ω ε−3/2 Lower-Bound on Alphabet Size
The twin word problem was introduced by Axenovich, Person, and Puzynina [1] and further studied
by Bukh and Zhou [4]. Any set of two identical disjoint subsequences in a given string is called a
twin word. [1, 4] provided a variety of results on the relations between the length of a string, the
size of the alphabet over which it is defined, and the size of the longest twin word it contains. We
will make use of the following result from [4] that is built upon Lemma 5.9 from [2] to provide a
new lower-bound on the alphabet size of synchronization strings.
Theorem 5.1 (Theorem 3 from [4]). There exists a constant c so that every word of length n over
a q-letter alphabet contains two disjoint equal subsequences of length cnq −2/3 .
Further, Theorem 6.4 of [15] states that any ε-synchronization string of length n has to satisfy
ε-self-matching property which essentially means that it cannot contain two (not necessarily disjoint)
subsequences of length εn or more. These two requirements lead to the following inequality for an
ε-synchronization string of length n over an alphabet of size q.
cnq −2/3 ≤ εn ⇒ c0 ε−3/2 ≤ q
21
6
Synchronization Strings over Small Alphabets
In this section, we focus on synchronization strings over small constant-sized alphabets. We study the
question of what is the smallest possible alphabet size over which arbitrarily long ε-synchronization
strings can exist for some ε < 1, and how such synchronization strings can be constructed.
Throughout this section, we will make use of square-free strings introduced by Thue [26], which
is a weaker notion than synchronization strings that requires all consecutive equal-length substrings
to be non-identical. Note that no synchronization strings or square-free strings of length four
or more exist over a binary alphabet since a binary string of length four either contains two
consecutive similar symbols or two identical consecutive substrings of length two. However, for
ternary alphabets, arbitrarily long square-free strings exist and can be constructed efficiently using
uniform morphism [28]. In Section 6.1, we will briefly review this construction and show that no
uniform morphism can be used to construct arbitrary long synchronization strings. In Section 6.2,
we make use of ternary square-free strings to show that arbitrarily long ε-synchronization strings
exist over alphabets of size four for some ε < 1. Finally, in Section A, we provide experimental
lower-bounds on ε’ for which ε-synchronization strings exist over alphabets of size 3, 4, 5, and 6.
6.1
Morphisms cannot Generate Synchronization Strings
Previous works show that one can construct infinitely long square-free or approximate-square-free
strings using uniform morphisms. A uniform morphism of rank r over an alphabet Σ is a function
φ : Σ → Σr that maps any symbol out of an alphabet Σ to a string of length r over the same
alphabet. Applying the function φ over some string S ∈ Σ∗ is defined as replacing each symbol of S
with φ(S).
[20, 27, 21, 7, 28] show that there are uniform morphisms that generate the combinatorial
objects they study respectively. More specifically, one can start from any letter of the alphabet and
repeatedly apply the morphism on it to construct those objects. For instance, using the uniform
morphisms of rank 11 suggested in [28], all such strings will be square-free. In this section, we
investigate the possibility of finding similar constructions for synchronization strings. We will show
that no such morphism can possibly generate an infinite ε-synchronization strings for any fixed
0 < ε < 1.
The key to this claim is that a matching between two substrings is preserved under an application
of the uniform morphism φ. Hence, we can always increase the size of a matching between two
substrings by applying the morphism sufficiently many times, and then adding new matches to the
matching from previous steps.
Theorem 6.1. Let φ be a uniform morphism of rank r over alphabet Σ. Then φ does not generate
an infinite ε-synchronization string, for any 0 < ε < 1.
Proof. To prove this, we show that for any 0 < ε < 1, applying morphism φ sufficiently many
times over any symbol of alphabet Σ produces a strings that has two neighboring intervals which
contradict ε-synchronization property. First, we claim that, without loss of generality, it suffices to
prove this for morphisms φ for which φ(σ) contains all elements of Σ for any σ ∈ Σ. To see this,
consider the graph G with |Σ| vertices where each vertex corresponds to a letter of the alphabet
and there is a (σ1 , σ2 ) edge if φ(σ1 ) contains σ2 . It is straightforward to verify that after applying
morphism φ over a letter sufficiently many times, the resulting string can be split into a number of
substrings so that the symbols in any of them belong to a subset of Σ that corresponds to some
strongly connected component in G. As ε-synchronization string property is a hereditary property
over substrings, this gives that one can, without loss of generality, prove the above-mentioned claim
22
for morphisms φ for which the corresponding graph G is strongly connected. Further, let d be the
greatest common divisor of the size of all cycles in G. One can verify that, for some sufficiently
large k, φkd will be a morphism that, depending on the letter σ to perform recursive applications
of the morphism on, will always generate strings over some alphabet Σσ and φkd (σ 0 ) contains all
symbols of Σσ for all σ 0 ∈ Σσ . As proving the claim for φkd implies it for φ as well, the assumption
mentioned above does not harm the generality.
We now proceed to prove that for any morphism φ of rank r as described above, any positive
integer n ∈ N, and any positive constant 0 < δ < 1, there exists m ∈ N so that
n
1
m
m
− δ · rm
LCS(φ (a), φ (b)) ≥ 1 − 1 −
|Σ|2 r
for any a, b ∈ Σ where φm represents m consecutive applications of morphism φ and LCS(., .)
denotes the longest common substring.
Having such claim proved, one can take δ = (1 − ε)/2 and n large enough so that m applications
of φ over any pair of symbols entail strings with a longest common substring that is of a fraction
larger than 1 − (1 − ε) = ε in terms of the length of those strings. Then, for any string S ∈ Σ∗ ,
one can take two arbitrary consecutive symbols of φ(S) like S[i] and S[i + 1]. Applying morphism
φ, m more times on φ(S) makes the corresponding intervals of φm+1 (S) have an edit distance
that is smaller than 1 − ε fraction of their combined lengths. This shows that φm+1 (S) is not an
ε-synchronization string and finishes the proof.
Finally, we prove the claim by induction on n. For the base case of n = 1, given the assumption
of all members of Σ appearing in φ(σ) for all σ ∈ Σ, φ(a) and φ(b) have a non-empty common
subsequence. This gives that
1
1
LCS(φ(a), φ(b)) ≥ 1 = 1 − 1 −
·r > 1− 1−
− δ · r.
r
|Σ|2 r
Therefore, choosing m = 1 finishes the induction base.
We now prove the induction step. Note that by induction hypothesis, for some given n, one can
find m1 such that
n
1
δ
m1
m1
LCS(φ (a), φ (b)) ≥ 1 − 1 −
−
· rm1 .
|Σ|2 r
2
Now, let m2 = logr 2δ . Consider φm2 (a) and φm2 (b). Note that among all possible pairs of symbols
m
from Σ2 , one appears at least r|Σ|22 times in respective positions of φm2 (a) and φm2 (b). Let (a0 , b0 )
be such pair. As φ(a0 ) and φ(b0 ) contain all symbols of Σ, one can take one specific occurrence of
a fixed arbitrary symbol σ ∈ Σ in all appearances of the pair φ(a0 ) and φ(b0 ) to find a common
m
m +1
subsequence of size r|Σ|22 = r|Σ|22 r or more between φm2 +1 (a) and φm2 +1 (b) (See Figure 6).
𝜙 𝑚2 (𝑎)
𝜙 𝑚2 (⋅)
𝜙 𝑚2 +1 (𝑎)
𝜙(⋅)
𝑎
𝑎
𝑏
𝑐
𝑏
𝑎
𝜙(𝑎)
𝑏
𝑐
𝑎
𝑏
𝑎
𝑏
𝜙(𝑐)
𝜙(𝑏)
𝜎
𝜙(𝑐)
𝜙(𝑎)
𝜙(𝑏)
𝜎
𝜙
𝑚2
(𝑏)
𝜙(𝑏)
𝜎
𝜙(𝑎)
𝜙(𝑎)
𝜙(𝑏)
𝜎
𝜙
𝑚2 +1
(𝑏)
Figure 6: Induction step in Theorem 6.1; Most common pair (a0 , b0 ) = (b, a).
23
Note that one can apply the morphism φ further times over φm2 +1 (a) and φm2 +1 (b) and such
common subsequence will still be preserved; However, one might be able to increase the size of it
by adding new elements to the common subsequence from equal length pairs of intervals between
current common subsequence elements (denoted by blue dashed line in Figure 6). The total length
of such intervals is
1
1
r
δ
1−
− m2 +1 = 1 −
−
2
2
|Σ| r r
|Σ| r 2
or more. In fact, using the induction hypothesis, by applying the morphism m1 more times, one can
get the following for m = m1 + m2 + 1.
"
n
1
δ
1
m
m
+ 1− 1−
−
LCS(φ (a), φ (b)) ≥
|Σ|2 r
|Σ|2 r
2
#
1
δ
· 1−
rm
−
|Σ|2 r 2
"
#
n+1
1
≥ 1− 1−
− δ rm
|Σ|2 r
This completes the induction step and finishes the proof.
6.2
Synchronization Strings over Alphabets of Size Four
In this section, we show that synchronization strings of arbitrary length exist over alphabets of size
four. In order to do so, we first introduce the notion of weak ε-synchronization strings. This weaker
notion is very similar to the synchronization string property except the edit distance requirement is
rounded down.
Definition 6.2 (weak ε-synchronization strings). String S of length n is a weak ε-synchronization
string if for every 1 ≤ i < j < k ≤ n,
ED(S[i, j), S[j, k)) ≥ b(1 − ε)(k − i)c.
We start by showing that binary weak ε-synchronization strings exist for some ε < 1.
6.2.1
Binary Weak ε-Synchronization Strings
Here we prove that an infinite binary weak ε-synchronization string exists. The main idea is to take
a synchronization string over some large alphabet and convert it to a binary weak synchronization
string by mapping each symbol of that large alphabet to a binary string and separating each binary
encoded block with a block of the form 0k 1k .
Theorem 6.3. There exists a constant ε < 1 and an infinite binary weak ε-synchronization string.
Proof. Take some arbitrary ε0 ∈ (0, 1). According to [15], there exists an infinite ε0 -synchronization
string S over a sufficiently large alphabet Σ. Let k = dlog |Σ|e. Translate each symbol of S into k
binary bits, and separate the translated k-blocks with 0k 1k . We claim that this new string T is a
weak ε-synchronization binary string for some ε < 1.
First, call a translated k-length symbol followed by 0k 1k a full block. Call any other (possibly
empty) substring a half block. Then any substring of T is a half-block followed by multiple full
blocks and ends with a half block.
24
Let A and B be two consecutive substrings in T . Without loss of generality, assume |A| ≤ |B|
(because edit distance is symmetric). Let M be a longest common subsequence between A and B.
Partition blocks of B into the following 4 types of blocks:
1. Full blocks that match completely to another full block in A.
2. Full blocks that match completely but not to just 1 full block in A.
3. Full blocks where not all bits within are matched.
4. Half blocks.
The key claim is that the 3k elements in B which are matched to a type-2 block in A are not
contiguous and, therefore, there is at least one unmatched symbol in B surrounded by them. To see
this, assume by contradiction that all letters of some type-2 block in A are matched contiguously.
The following simple analysis over 3 cases contradicts this assumption:
• Match starts at middle of some translated k-length symbol, say position p ∈ [2, k]. Then the
first 1 of 1k in A will be matched to the (k − p + 2)-th 0 of 0k in B, contradiction.
• Match starts at 0-portion of 0k 1k block, say at the p-th 0. Then the p-th 1 of 1k in A will be
matched to the first 0 of 0k in B, contradiction.
• Match starts at 1-portion of 0k 1k block, say at the p-th 1. Then the p-th 0 of 0k in A will be
matched to the first 1 of 1k in B, contradiction.
Let the number of type-i blocks in B be ti . For every type-2 block, there is an unmatched letter
in A between its first and last matches. Hence, |A| ≥ |M | + t2 . For every type-3 block, there is an
unmatched letter in B within. Hence, |B| ≥ |M | + t3 . Therefore, |A| + |B| ≥ 2|M | + t2 + t3 .
Since |A| ≤ |B|, the total number of full blocks in both A and B is at most 2(t1 + t2 + t3 ) + 1.
(the additional +1 comes from the possibility that the two half-blocks in B allows for one extra full
block in A) Note t1 is a matching between the full blocks in A and the full blocks in B. So due to
the ε0 -synchronization property of S, we obtain the following.
t1 ≤
ε0
ε0
ε0
(2(t1 + t2 + t3 ) + 1) =⇒ t1 ≤
(t
+
t
)
+
2
3
2
1 − ε0
2(1 − ε0 )
|A|+|B|
Furthermore, t1 + t2 + t3 + 2 > |B|
3k ≥
6k . This, along with the above inequality, implies the
following.
|A| + |B|
1
4 − 3ε0
(t
+
t
)
+
>
.
2
3
0
0
1−ε
2(1 − ε )
6k
The edit distance between A and B is
ED(A, B) = |A| + |B| − 2|M | ≥ t2 + t3
1 − ε0
4 − 3ε0
1 − ε0
>
(|A| + |B|) −
>
(|A| + |B|) − 2.
6k
2
6k
Set ε = 1 −
1−ε0
18k .
If |A| + |B| ≥
1
1−ε ,
then
1 − ε0
1 − ε0
(|A| + |B|) − 2 ≥
− 2(1 − ε) (|A| + |B|)
6k
6k
= (1 − ε)(|A| + |B|) ≥ b(1 − ε)(|A| + |B|)c.
As weak ε-synchronization property trivially holds for |A| + |B| <
weak ε-synchronization string.
25
1
1−ε ,
this will prove that T is a
6.2.2
ε-Synchronization Strings over Alphabets of Size Four
A corollary of Theorem 6.3 is the existence of infinite synchronization strings over alphabets of size
four. Here we make use of the fact that infinite ternary square-free strings exist, which was proven
in previous work [26]. We then modify such a string to fulfill the synchronization string property,
using the existence of an infinite binary weak synchronization string.
Theorem 6.4. There exists some ε ∈ (0, 1) and an infinite ε-synchronization string over an alphabet
of size four.
Proof. Take an infinite ternary square-free string T over alphabet {1, 2, 3} [26] and some ε ∈ 11
12 , 1 .
Let S be an infinite weak binary ε0 = (12ε − 11)-synchronization string. Consider the string W that
is similar to T except that the i-th occurrence of symbol 1 in T is replaced with symbol 4 if S[i] = 1.
Note W is still square-free. We claim W is an ε-synchronization string as well.
Let A = W [i, j), B = W [j, k) be two consecutive substrings of W . If k − i < 1/(1 − ε), then
ED(A, B) ≥ 1 > (1 − ε)(k − i) by square-freeness.
Otherwise, k − i ≥ 1/(1 − ε) ≥ 12. Consider all occurrences of 1 and 4 in A and B, which form
consecutive subsequences As and Bs of S respectively. Note that |As | + |Bs | ≥ (k − i − 3)/4, because,
by square-freeness, there cannot be a length-4 substring consisting only of 2’s and 3’s in W .
By weak synchronization property,
ED(As , Bs ) ≥ b(1 − ε0 )(|As | + |Bs |)c
≥ b3(1 − ε)(k − i − 3)c > 3(1 − ε)(k − i) − 9(1 − ε) − 1
≥ (1 − ε)(k − i),
and hence, ED(A, B) ≥ ED(As , Bs ) ≥ (1 − ε)(k − i). Therefore, W is an ε-synchronization
string.
Organization and Acknowledgements
All detailed proofs appear in the appendix. Specifically, proofs in Section 3 appear in Appendix A,
proofs in Section 4 appear in Appendix B, and proofs in Section 6 appear in Appendix C.
The authors thank Raymond Kang for valuable discussions in the early stages of this work,
contributions to Theorem 6.1, and experiments in Appendix A. We also thank Noga Alon for
referring us to the previous work on the twin word problem.
References
[1] Maria Axenovich, Yury Person, and Svetlana Puzynina. A regularity lemma and twins in words.
Journal of Combinatorial Theory, Series A, 120(4):733–743, 2013.
[2] Paul Beame and Dang-Trinh Huynh-Ngoc. On the value of multiple read/write streams for
approximating frequency moments. In Foundations of Computer Science, 2008. FOCS’08.
IEEE 49th Annual IEEE Symposium on, pages 499–508. IEEE, 2008.
[3] Mark Braverman, Ran Gelles, Jieming Mao, and Rafail Ostrovsky. Coding for interactive
communication correcting insertions and deletions. IEEE Transactions on Information Theory,
63(10):6256–6270, 2017.
26
[4] Boris Bukh and Lidong Zhou. Twins in words and long common subsequences in permutations.
Israel Journal of Mathematics, 213(1):183–209, 2016.
[5] Serina Camungol, Narad Rampersad, et al. Avoiding approximate repetitions with respect to
the longest common subsequence distance. inv lve, page 657, 2016.
[6] Karthekeyan Chandrasekaran, Navin Goyal, and Bernhard Haeupler. Deterministic algorithms
for the lovász local lemma. SIAM Journal on Computing, 42(6):2132–2155, 2013.
[7] Max Crochemore. Sharp characterizations of squarefree morphisms. Theoretical Computer
Science, 18(2):221–226, 1982.
[8] Ran Gelles. Coding for interactive communication: A survey, 2015.
[9] Ran Gelles and Bernhard Haeupler. Capacity of interactive communication over erasure channels
and channels with feedback. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium
on Discrete Algorithms, pages 1296–1311. Society for Industrial and Applied Mathematics,
2015.
[10] Mohsen Ghaffari and Bernhard Haeupler. Optimal error rates for interactive coding ii: Efficiency
and list decoding. In Foundations of Computer Science (FOCS), 2014 IEEE 55th Annual
Symposium on, pages 394–403. IEEE, 2014.
[11] Mohsen Ghaffari, Bernhard Haeupler, and Madhu Sudan. Optimal error rates for interactive
coding i: Adaptivity and other settings. In Proceedings of the forty-sixth annual ACM symposium
on Theory of computing, pages 794–803. ACM, 2014.
[12] V. Guruswami and P. Indyk. Linear-time encodable/decodable codes with near-optimal rate.
IEEE Transactions on Information Theory, 51(10):3393–3400, Oct 2005. doi:10.1109/TIT.
2005.855587.
[13] Bernhard Haeupler. Interactive channel capacity revisited. In Foundations of Computer Science
(FOCS), 2014 IEEE 55th Annual Symposium on, pages 226–235. IEEE, 2014.
[14] Bernhard Haeupler, Barna Saha, and Aravind Srinivasan. New constructive aspects of the
lovász local lemma. Journal of the ACM (JACM), 58(6):28, 2011.
[15] Bernhard Haeupler and Amirbehshad Shahrasbi. Synchronization strings: codes for insertions
and deletions approaching the singleton bound. In Proceedings of the 49th Annual ACM
SIGACT Symposium on Theory of Computing, pages 33–46. ACM, 2017.
[16] Bernhard Haeupler and Amirbehshad Shahrasbi. Synchronization strings: Explicit constructions,
local decoding, and applications. arXiv preprint arXiv:1710.09795, 2017.
[17] Bernhard Haeupler, Amirbehshad Shahrasbi, and Madhu Sudan. Synchronization strings: List
decoding for insertions and deletions, 2018.
[18] Bernhard Haeupler, Amirbehshad Shahrasbi, and Ellen Vitercik. Synchronization strings:
Channel simulations and interactive coding for insertions and deletions. arXiv preprint
arXiv:1707.04233, 2017.
[19] Gillat Kol and Ran Raz. Interactive channel capacity. In Proceedings of the forty-fifth annual
ACM symposium on Theory of computing, pages 715–724. ACM, 2013.
27
[20] Dalia Krieger, Pascal Ochem, Narad Rampersad, and Jeffrey Shallit. Avoiding approximate
squares. In International Conference on Developments in Language Theory, pages 278–289.
Springer, 2007.
[21] John Leech. 2726. a problem on strings of beads. The Mathematical Gazette, 41(338):277–278,
1957.
[22] Hugues Mercier, Vijay K Bhargava, and Vahid Tarokh. A survey of error-correcting codes
for channels with symbol synchronization errors. IEEE Communications Surveys & Tutorials,
12(1), 2010.
[23] Robin A Moser and Gábor Tardos. A constructive proof of the general lovász local lemma.
Journal of the ACM (JACM), 57(2):11, 2010.
[24] Robert Shelton. Aperiodic words on three symbols. Journal für die Reine und Angewandte
Mathematik, 321:195–209, 1981.
[25] Robert O Shelton and Raj P Soni. Aperiodic words on three symbols. iii. Journal für die Reine
und Angewandte Mathematik, 330:44–52, 1982.
[26] A Thue. Über unendliche zeichenreihen (1906). Selected Mathematical Papers of Axel Thue.
Universitetsforlaget, 1977.
[27] Axel Thue. Über die gegenseitige Lage gleicher Teile gewisser Zeichenreihen, von Axel Thue...
J. Dybwad, 1912.
[28] Boris Zolotov. Another solution to the thue problem of non-repeating words. arXiv preprint
arXiv:1505.00019, 2015.
28
Appendices
A
Lower-bounds for ε in Infinite ε-Synchronization Strings
It is known from Section 6.2 that infinite synchronization strings exist over alphabet sizes |Σ| ≥ 4.
A natural question to ask is the optimal value of ε for each such |Σ|. Formally, we seek to discover
Bk = inf{ε : there exists an infinite ε-synchronization string with |Σ| = k}
for small values of k. To that end, a program was written to find an upper bound for Bk for k ≤ 6.
The program first fixes an ε, then exhaustively enumerates all possible ε-synchronization strings
over an alphabet size of k by increasing length. If the program terminates, then this ε is a proven
lower bound for Bk . Among every pair of consecutive substrings in each checked string that failed
the ε-synchronization property, we find the one that has the lowest edit distance relative to their
total length and such fraction would be a lower-bound for Bk as well. Such experimentally obtained
lower-bounds for alphabets of size 3, 4, 5, and 6 are listed in Table 1.
k
3
4
5
6
Bk ≥ ·
12/13
10/13
2/3
18/29
Table 1: Computationally proven lower-bounds of Bk
29
| 8cs.DS
|
PROC. OF THE 7th EUR. CONF. ON PYTHON IN SCIENCE (EUROSCIPY 2014)
53
Predictive Modelling of Toxicity Resulting from
Radiotherapy Treatments of Head and Neck Cancer
arXiv:1412.6399v1 [physics.med-ph] 19 Dec 2014
Jamie A Dean∗† , Liam C Welsh† , Kevin J Harrington† , Christopher M Nutting† , Sarah L Gulliford†
F
Abstract—In radiotherapy for head and neck cancer, the radiation dose delivered to the pharyngeal mucosa (mucosal lining of the throat) is thought to
be a major contributing factor to dysphagia (swallowing dysfunction), the most
commonly reported severe toxicity. There is a variation in the severity of dysphagia experienced by patients. Understanding the role of the dose distribution
in dysphagia would allow improvements in the radiotherapy technique to be
explored. The 3D dose distributions delivered to the pharyngeal mucosa of 249
patients treated as part of clinical trials were reconstructed. Pydicom was used
to extract DICOM (digital imaging and communications in medicine) data (the
standard file formats for medical imaging and radiotherapy data). NumPy and
SciPy were used to manipulate the data to generate 3D maps of the dose
distribution delivered to the pharyngeal mucosa and calculate metrics describing
the dose distribution. Multivariate predictive modelling of severe dysphagia,
including descriptions of the dose distribution and relevant clinical factors, was
performed using Pandas and SciKit-Learn. Matplotlib and Mayavi were used
for 2D and 3D data visualisation. A support vector classification model, with
feature selection using randomised logistic regression, to predict radiationinduced severe dysphagia, was trained. When this model was independently
validated, the area under the receiver operating characteristic curve was 0.54.
The model has poor predictive power and work is ongoing to improve the model
through alternative feature engineering and statistical modelling approaches.
Index Terms—radiotherapy, radiation oncology, head and neck cancer, dysphagia, pharyngeal mucosa, toxicity, predictive modelling, machine learning,
statistical learning
1
Fig. 1: Example of a head and neck treatment plan. The colour wash
shows the calculated 3D dose distribution and the pharyngeal mucosa
contours, delineated by a radiation oncologist, are shown in green.
I NTRODUCTION
Head and neck cancer is the fifth most common cancer worldwide, with an annual incidence of approximately 500,000
cases globally [Jemal]. Radiotherapy is the primary nonsurgical treatment of head and neck cancer and is commonly
given in combination with chemotherapy and/or surgery. The
aims of the treatment are to achieve loco-regional disease
control whilst preserving organ function. Modern radiotherapy
techniques allow the radiation dose delivered to the patient
to be modulated in order to create highly conformal dose
distributions, which minimise the doses delivered to normal
tissues in close proximity to the tumour. A typical example of
a head and neck radiotherapy treatment plan dose distribution
* Corresponding author: [email protected]
† The Institute of Cancer Research and The Royal Marsden NHS Foundation
Trust, London, UK
c 2014 Jamie A Dean et al. This is an open-access artiCopyright ○
cle distributed under the terms of the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction
in any medium, provided the original author and source are credited.
http://creativecommons.org/licenses/by/3.0/
is shown in Figure 1. However, there are still high rates of
toxicity, which reduce patients’ quality of life and limit the
amount of dose that can be delivered to the tumour and hence
the probability of controlling the disease. The most commonly reported severe radiation-induced toxicity, during and
following treatment, is dysphagia (swallowing dysfunction).
A variation in the severity of radiation-induced dysphagia
between patients is observed. The radiation dose delivered
to the pharyngeal mucosa (mucosal lining of the throat) is
thought to be a major contributing factor to this side effect
[Eisbruch]. Understanding the role of the radiotherapy dose
distribution in the onset of dysphagia would allow further
improvements to the radiotherapy technique, aiming to reduce
the number of patients who experience severe dysphagia.
Python allows seamless integration of radiotherapy data
handling, manipulation, processing and statistical modelling
using machine learning through the Pydicom, NumPy, SciPy,
Pandas and SciKit-Learn modules. In this manuscript we
outline our application of Python to manipulate radiotherapy
54
PROC. OF THE 7th EUR. CONF. ON PYTHON IN SCIENCE (EUROSCIPY 2014)
treatment plan data to map the planned dose distribution onto
the pharyngeal mucosa, compute metrics describing the dose
distributions delivered to the mucosa, and train and assess
multivariate predictive models of radiation-induced severe
dysphagia.
2
2.1
M ATERIALS AND M ETHODS
Patient Data
Data were available for 249 head and neck cancer patients
treated in clinical trials at The Royal Marsden NHS Foundation Trust [Nutting], [Guerrero-Urbano], [Powell]. All data
were anonymised. Toxicity outcome data were collected for
patients as part of the trial protocols. The severity of dysphagia was graded, according to the Common Terminology
Criteria for Adverse Events [CTCAE], on a scale from 0
to 5, with higher grades representing higher severity. These
toxicity grades were dichotomised into ‘mild’ and ‘severe’
dysphagia, with severe defined as grade 3 or higher. Grade
3 dysphagia corresponds to the patient requiring the insertion
of a percutaneous endoscopic gastrostomy (feeding) tube. It
requires a surgical procedure and represents a pronounced
detriment to a patient’s quality of life and is thus of high
clinical importance. This reasoning warranted the decision to
dichotomise the outcomes, as well as the choice of threshold.
In addition to its clinical importance, the cut point of grade 3
dysphagia is relatively objective compared with other toxicity
outcome measures and so minimises noise in the outcome data.
Clinical factors that are potentially relevant to dysphagia were
also collected as part of the trial protocols. These included
patient age, gender, comorbidities and concomitant treatments.
2.2
Preparation of Radiotherapy Treatment Plan Data
The pharyngeal mucosa was defined by head and neck radiation oncologists, who contoured the structure on axial
slices of computed tomography (CT) scans using either the
Pinnacle (Philips Radiation Oncology Systems, Fitchburg,
WI), CadPlan (Varian Medical Systems, Palo Alto, CA) or
Eclipse (Varian Medical Systems, Palo Alto, CA) proprietary
radiotherapy treatment planning software and the structure definition described in [Bhide]. The treatments were planned and
the doses calculated by physicists and dosimetrists, following
the trial protocols, using the same software.
All scientific computing tasks were performed using Python
v.2.7.6 [Python]. A script was used to anonymise the identifying data in the radiotherapy treatment plan files. When
treatment planning is performed, the radiation oncologist
names the organs that they delineate. This commonly leads to
inconsistencies in the names given to the same organ between
patients. A script was written to change the name of the
pharyngeal mucosa structure from the one assigned by the
radiation oncologist (the user selects the name to change, e.g.
‘Pharyngeal Mucosa’, ‘Pharynx’ or ‘pharyngeal mucosa’, from
a list of all of the structures) to ‘PM’. This allowed for batch
processing of the tasks relating to the radiotherapy treatment
plans. The user can pass a list of the patient identification
numbers and the name of the organ of interest, in this case
‘PM’, and the dose maps and metrics outlined below are
calculated for all patients, without the need for interventions
from the user.
2.3
Generating 3D Dose Maps
Radiotherapy treatment plans are produced using proprietary
software. There is a standardised format for radiotherapy
treatment plan data, DICOM RT, in which the treatment plan
data is exported from the planning software. The information
in a treatment plan is saved in four different types of file:
• CT - These files contain the computed tomography images used to model the patient and calculate the dose
distribution in the treatment planning software. There is
one file per 2D slice of the 3D image volume acquired.
• RTDOSE - This file contains the calculated dose distribution, described by points on a 3D grid.
• RTSTRUCT - This file contains the coordinates of the
contours of the structures delineated by the radiation
oncologist.
• RTPLAN - This file contains the parameters of the
radiation beams.
To study the dose-toxicity response relationship for the
pharyngeal mucosa, it was necessary to reconstruct the planned
dose to this structure from the DICOM RT files. This was
achieved using Pydicom v.0.9.3-1 [Pydicom], NumPy v.1.8.11 [NumPy], SciPy v.0.14.0-3 [SciPy] and Matplotlib v.1.3.1-9
[Matplotlib]. The resolutions of the dose grids are courser than
the image grid upon which the structure contour points are defined. Moreover, the dose grid does not cover the entire image
grid, only the volume that is irradiated. Therefore, the dose
grid was tri-linearly interpolated to match the resolution of the
image grid (tri-linear interpolation does not lead to noticeable
distortions of the dose distributions with the acquired dose grid
resolution of 2.5 mm x 2.5 mm x 2.5 mm) and the translational offset between the dose grid and image grid calculated
using the image coordinates located in the metadata of the
RTSTRUCT files. The matplotlib.path.Path.contains_points()
function and the coordinates of the pharyngeal mucosa structure contours were used to determine which voxels were
included in the structure. Since the pharyngeal mucosa is
a ‘wall’-type structure as opposed to a ‘solid’ volume the
numpy.logical_xor() function was used to exclude those voxels
located within one of the contours, but not part of the mucosa
structure. This informed a binary mask consisting of ‘1’s for
the voxels in the structure and ‘0’s for the voxels not included
in the structure. Multiplying the 3D dose grid with this binary
mask produced a 3D map of just the dose delivered to the
pharyngeal mucosa.
2.4
Extracting Dose Metrics
In most radiotherapy dose-outcome studies, the 3D dose
distribution is reduced to a cumulative dose-volume histogram
(DVH), describing the volume of an organ that receives each
dose level. DVHs were extracted from the 3D dose maps using
the NumPy histogram function:
import numpy as np
PREDICTIVE MODELLING OF TOXICITY RESULTING FROM RADIOTHERAPY TREATMENTS OF HEAD AND NECK CANCER
def dose_volume_histogram(organDose):
"""Computes the cumulative dose-volume
histogram"""
hist = np.histogram(organDose,
range = (0, 80), bins = 80,
density = True)[0]
dvh = 1.0 - np.cumsum(hist)
return dvh
The mean and maximum doses were also calculated using
NumPy.
2.5
Statistical Modelling
The volumes of pharyngeal mucosa receiving cumulative doses
in the range 10 Gy (V10) to 80 Gy (V80) in 5 Gy intervals
were extracted from the DVHs for inclusion as model covariates. The mean and maximum doses were also used as model
inputs. The clinical factors included as covariates were: age,
gender, smoking status (smoker at the time of treatment vs.
non-smoker at the time of treatment), alcohol status, induction
chemotherapy (binary) and concurrent chemotherapy (binary).
Data preparation for statistical modelling was carried out using
Pandas v.0.14.1-2 [Pandas].
In order for the model to be able to inform clinical practice it
must have high interpretability and low dimensionality. These
requirements informed the choice of modelling strategies employed. Feature selection is often challenging when working
with high-dimensionality data. An additional complication to
the feature selection problem associated with radiotherapy
dose-outcome studies is high collinearity between covariates.
This is due to the nature of the physical processes that govern
how the radiation dose is deposited in the patient. Adjacent
dose levels in the DVH are highly correlated making it
challenging to distinguish the dose levels that are important for
predicting toxicity outcomes from those that merely correlate
with the dose levels that cause toxicity. To have the best chance
of overcoming this problem a combination of large variation
in the DVHs between patients and a suitable feature selection
strategy must be employed.
Statistical modelling was performed using SciKit-Learn
v.0.15.1-1 [SciKit-Learn]. A subset of 25 patients (10 % of
the entire cohort) were separated out for use as an independent
test set of model performance:
import sklearn
from sklearn import cross_validation
# Split data into training set and independent test
#
set
xTrain, xTest, yTrain, yTest =
cross_validation.train_test_split(xData, yData,
test_size = 0.1, random_state = 42)
The model training pipeline consisted of data centring about
0 and scaling to unit variance followed by a feature selection
step and, finally, a model fitting step:
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model
import RandomizedLogisticRegression
from sklearn import svm
# z-scale data
scaler = StandardScaler()
# Feature selection method
55
featureSelection = RandomizedLogisticRegression(
fit_intercept = False)
# Model fitting method
modelFitting = svm.SVC(probability = True)
# Create pipeline
estimators = [(’scaler’, scaler),
(’featureSelection’, featureSelection),
(’modelFitting’, modelFitting)]
classifier = Pipeline(estimators)
Since the toxicity outcomes were known, supervised learning
techniques could be utilised. There is no obvious solution to
the problem of collinearity between the variables. Discriminant
analysis techniques, such as linear discriminant analysis and
quadratic discriminant analysis, cannot robustly handle correlated variables, making them inappropriate. One potential strategy is to remove the collinearity by performing PCA. However,
this results in a model with low interpretability and so would
not allow the causal features to be determined. An alternative
approach is to initially remove features using interpretable
dimensionality reduction techniques, for example, univariate
feature selection and recursive feature elimination and then
remove remaining correlated variables. These approaches,
however, may be unstable and suffer from reduced generalisability as different features may be selected with different
datasets. Randomised logistic regression (RLR) [Meinshausen]
was chosen for feature selection in order to maximise the
stability of the selected features. Support vector classification
(SVC) with linear and radial basis function kernels [Cortes]
was employed for model fitting as this technique is capable
of producing high performance, complex (incorporating nonlinearity and interactions), yet interpretable, models. A crossvalidated grid search, with stratified 5-fold cross-validation,
was used over the whole pipeline to tune the hyper-parameters
of the models used for feature selection and model fitting on
the reduced set of features:
from sklearn.grid_search import GridSearchCV
featureSelectionParams = {
’featureSelection__C’: [
0.01, 0.1, 1.0, 10.0, 100.0],
’featureSelection__scaling’: [0.25, 0.5, 0.75],
’featureSelection__selection_threshold’: [
0.3, 0.4, 0.5]}
modelFittingParams = {
’modelFitting__C’: [0.01, 0.1, 1.0, 10.0, 100.0],
’modelFitting__kernel’: [’linear’, ’rbf’],
’modelFitting__gamma’: [0.01, 0.1, 1.0]}
parameters = dict(featureSelectionParams.items()
+ modelFittingParams.items())
skf = cross_validation.StratifiedKFold(yTrain,
n_folds = 5)
gridSearch = GridSearchCV(clf, param_grid =
parameters, cv = 5, scoring = ’roc_auc’)
gridSearch.fit(xTrain, yTrain)
bestParameters, bestScore = gridSearch.best_params_,
gridSearch.best_score_
print ’Best Parameters:’, bestParameters
print ’Best Score:’, bestScore
To obtain an unbiased estimate of the model prediction score
on new data a nested 5-fold cross-validation was carried out,
with area under the receiver operating characteristic (ROC)
56
PROC. OF THE 7th EUR. CONF. ON PYTHON IN SCIENCE (EUROSCIPY 2014)
curve used as the scoring function:
# Nested 5-fold stratified cross-validation for model
#
evaluation
scores = cross_validation.cross_val_score(gridSearch,
xTrain, yTrain, cv = skf, scoring = ’roc_auc’)
print ’Cross-validation: ROC AUC =
%0.2f (+/- %0.2f)’ % (scores.mean(),
scores.std()*2)
All cross-validations were stratified to reduce variance in the
estimate of model performance. Data from the 25 patients in
the independent test dataset, not used in any part of the model
training process, were used to perform an independent test of
model performance:
# Test classifier on independent test data
yPred = gridSearch.predict(xTest)
print classification_report(yTest, yPred)
print ’ROC AUC: ’, roc_auc_score(yTest, yPred)
3
R ESULTS
Figure 2 shows an example of a 3D map of the dose delivered
to the pharyngeal mucosa. Mayavi v.4.3.1-2 [Mayavi] was
used for 3D data visualisation. This pharyngeal mucosa dose
distribution is typical of head and neck radiotherapy treatments. The region receiving the highest doses is included in the
treatment target volume and there is a dose fall-off superiorly
and inferiorly of this region.
Figure 3 shows the dose-volume histograms of the 249
patients included in the analysis. Matplotlib was used for 2D
data visualisation. It can be observed that there is variation in
the pharyngeal mucosa DVHs across the cohort. This variation
is largely due to the geometry of the treatment target volume:
both its size and location within the pharynx (throat).
The correlation matrix for the input variables is shown in
Figure 4. The correlation matrix highlights the high correlation
coefficients between adjacent dose levels in the DVH suggesting high collinearity.
Principal component analysis (PCA) was used for data
visualisation. Figure 5 displays the variance explanation and
Figure 6 shows the data projected into the first two principal
components space. Visualising the data using PCA shows
that the different outcomes are highly overlapping in the
first two principal components space (the first two principal
components explain a relatively large amount of the variance
in the data).
The model hyper-parameters chosen are shown in Table 1.
During model training the area under the ROC curve was
determined to be 0.54 +/- 0.23 and when the model was
validated on the independent dataset it was 0.54. Figure 7
shows the ROC curve for the independent validation.
4
Fig. 2: 3D map of pharyngeal mucosa dose distribution for one
patient.
D ISCUSSION
Python was deemed to be a well-suited platform for achieving
our aims of training predictive models of severe radiationinduced dysphagia. This is due to its strong functionality in
handling DICOM data, manipulation, processing and visualisation of 3D data, and machine learning.
The overlap of the two outcome classes in the first two principal components space upon PCA suggests that generalised
linear models, such as logistic regression, are unlikely to make
Fig. 3: DVHs for the 249 patients analysed. Blue curves represent
patients who did not experience severe dysphagia and red curves
represent patients who did.
PREDICTIVE MODELLING OF TOXICITY RESULTING FROM RADIOTHERAPY TREATMENTS OF HEAD AND NECK CANCER
Hyper-parameter
RLR regularisation constant (C)
RLR threshold
RLR scaling
SVC regularisation constant (C)
SVC kernel
SVC
kernel
hyperparameter (gamma)
57
Value
1.0
0.4
0.75
100.0
Radial
function
1.0
basis
TABLE 1: Model hyper-parameters.
Fig. 4: Correlation matrix of model covariates.
Fig. 7: ROC curve for the independent validation.
Fig. 5: Variance explanation against principal component index from
PCA analysis.
Fig. 6: Data projection into the first two principal components space.
Red points represent patients who experienced severe dysphagia, blue
points represent patients who did not experience severe dysphagia.
good classifiers in this instance. Non-linear kernel modelling
is likely to result in models with greater predictive power for
this dataset. During model hyper-parameter tuning, the SVC
kernel selected was the radial basis function, supporting this
observation.
The area under the ROC curve is low for both the internal
and external validation indicating that the model is unable to
correctly classify patients better than chance. The uncertainty
on the area under the ROC curve from the nested crossvalidation is large suggesting that the model is unstable. Our
present model is thus unsuitable to support clinical decision
making and inform on the causal features of radiation-induced
severe dysphagia. However, we are currently exploring improvements to our methodology with promising preliminary
results.
There are many potential reasons why our model suffers
from low predictive power. These include insufficient characterisation of chemotherapy treatments, smoking status or
alcohol status, the influence of other organs, not considered
in our study, involved in the onset of severe dysphagia, and
the impact of other factors not considered, such as surgery or
genetic susceptibility to radiation-induced mucosal damage.
However, we suspect that the low predictive power may,
at least in part, be due to limitations associated with the
dosimetric input data. In reducing the 3D dose distribution to
a DVH much of the information is discarded. We believe that
58
PROC. OF THE 7th EUR. CONF. ON PYTHON IN SCIENCE (EUROSCIPY 2014)
some of this discarded information is likely to be important
for predicting toxicity. We are developing several novel dose
metrics that more fully characterise the dose distribution,
which we aim to use as model inputs in the future. We
anticipate that these may improve the predictive power of the
models.
Furthermore, there are discrepancies between the planned
doses calculated using the treatment planning software (and
extracted for use in our study) and the doses that are actually
delivered to the patients. These are due to movement of the
patients’ internal anatomy whilst the radiation is delivered and
weight loss over the course of treatment. In the future we
plan to explore the magnitude of these effects and attempt to
develop strategies to minimise their impact.
The machine learning pipeline developed has been designed
to enable simple addition of different feature selection and
model fitting algorithms enabling alternative statistical techniques to be utilised in the future. It also transferable to other
toxicities and organs at risk from radiotherapy treatments, for
example, lung pneumonitis resulting from radiotherapy for the
treatment of lung cancer, as well as studying tumour control.
5
C ONCLUSIONS
In this study we have shown that Python can be successfully
applied to studies of radiotherapy dose-toxicity relationships.
The Pydicom, NumPy, SciPy, Pandas and SciKit-Learn modules allow for both manipulation and processing of the treatment planning data, and statistical modelling using machine
learning, making Python well suited to this type of study.
Whilst initial attempts to generate a predictive model of severe dysphagia were unsuccessful, preliminary investigation of
using novel dose metrics to characterise the dose distribution
appear promising. Ongoing work involves using NumPy and
SciPy to calculate novel dose metrics expected to influence
toxicity and the application of alternative statistical methods
within the machine learning pipeline.
R EFERENCES
[Jemal]
A. Jemal et al. Global cancer statistics, CA: A Cancer
Journal for Clinicians, 61:69-90, 2011.
[Eisbruch]
A. Eisbruch. Dysphagia and aspiration following
chemo-irradiation of head and neck cancer: major
obstacles to intensification of therapy, Annals of
Oncology, 15:363-364, 2004.
[Nutting]
C.M. Nutting et al. Parotid-sparing intensity modulated
versus conventional radiotherapy in head and neck cancer
(PARSPORT): a phase 3 multicentre randomised controlled trial, Lancet Oncology, 12:127-136, 2011.
[Guerrero-Urbano] T. Guerrero Urbano et al. A phase I study of doseescalated chemoradiation with accelerated intensity modulated radiotherapy in locally advanced
head and neck cancer, Radiotherapy and Oncology,
85:36-41, 2007.
[Powell]
C. Powell et al. Fatigue during chemoradiotherapy
for nasopharyngeal cancer and its relationship to
radiation dose distribution in the brain, Radiotherapy and Oncology, 110:416-421, 2014.
[CTCAE]
The National Cancer Institute. Common terminology criteria for adverse events v3.0, 2006.
[Bhide]
S.A. Bhide et al. Characteristics of response of oral
and pharyngeal mucosa in patients receiving chemoIMRT for head and neck cancer using hypofractionated
accelerated radiotherapy, Radiotherapy and Oncology,
97:86-91, 2010.
[Python]
[NumPy]
[SciPy]
[Pydicom]
[Pandas]
[SciKit-Learn]
[Matplotlib]
[Mayavi]
[Meinshausen]
[Cortes]
Python Software Foundation. Python Language Reference, version 2.7, Available at http://www.python.org.
T Oliphant et al. Numerical Python (NumPy), http://www.
numpy.org [Online; accessed 2014-09-29].
E. Jones et al. SciPy: Open source scientific tools
for Python, http://www.scipy.org/ [Online; accessed
2014-08-20], 2001.
D. Mason.
Pydicom,
https://code.google.com/p/
pydicom/ [Online; accessed 2014-08-20].
W. McKinney. Data structures for statistical computing in Python, Proceedings of the 9th Python in
Science Conference, 51-56, 2010.
F. Pedregosa et al. Scikit-learn: Machine learning in
Python, Journal of Machine Learning Research,
12:2825-2830, 2011.
J.D. Hunter. Matplotlib: A 2D graphics environment,
Computing in Science and Engineering, 9:90-95, 2007.
P. Ramachandran and G. Varoquaux. Mayavi: 3D
visualization of scientific data, IEEE Computing in
Science & Engineering, 13:40-51, 2011.
N. Meinshausen and P. Buhlmann. Stability selection,
Journal of the Royal Statistical Society: Series B,
72:417-473, 2010.
C. Cortes and V. Vapnik. Support-vector networks,
Machine learning, 20:273-297, 1995.
| 5cs.CE
|
Dynamic compensation and homeostasis:
a feedback control perspective
arXiv:1801.04959v1 [cs.SY] 15 Jan 2018
Michel Fliessa,c , Cédric Joinb,c,d
a LIX
(CNRS, UMR 7161), École polytechnique, 91128 Palaiseau, France.
[email protected]
b CRAN
(CNRS, UMR 7039), Université de Lorraine, BP 239,
54506 Vandœuvre-lès-Nancy, France
[email protected]
c AL.I.E.N.
(ALgèbre pour Identification & Estimation Numériques),
7 rue Maurice Barrès, 54330 Vézelise, France.
{michel.fliess, cedric.join}@alien-sas.com
d Projet
Non-A, INRIA Lille – Nord-Europe, France
Abstract
“Dynamic compensation” is a robustness property where a perturbed biological
circuit maintains a suitable output [Karin O., Swisa A., Glaser B., Dor Y., Alon
U. (2016). Mol. Syst. Biol., 12: 886]. In spite of several attempts, no fully
convincing analysis seems now to be on hand. This communication suggests an
explanation via “model-free control” and the corresponding “intelligent” controllers [Fliess M., Join C. (2013). Int. J. Contr., 86, 2228-2252], which are
already successfully applied in many concrete situations. As a byproduct this
setting provides also a slightly different presentation of homeostasis, or “exact
adaptation,” where the working conditions are assumed to be “mild.” Several
convincing, but academic, computer simulations are provided and discussed.
Keywords: Systems biology, homeostasis, exact adaptation, dynamic
compensation, integral feedback control, PID, model-free control, intelligent
proportional controller.
Research Report. École polytechnique, Palaiseau, France, January 2018
1. Introduction
In systems biology, i.e., an approach of growing importance to theoretical biology (see, e.g., Alon (2006); Klipp et al. (2016); Kremling (2012)), basic control notions, like feedback loops, are becoming more and more popular
(see, e.g., Åström et al. (2008); Cowan et al. (2014); Cosentino et al. (2011);
Del Vecchio et al. (2015)). This communication intends to show that a peculiar
feedback loop permits to clarify the concept of dynamic compensation (DC ) of
biological circuits, which was recently introduced by Karin et al. (2016). DC is
a robustness property. It implies, roughly speaking, that biological systems are
able of maintaining a suitable output despite environmental fluctuations. As
noticed by Karin et al. (2016) such a property arises naturally in physiological
systems. The DC of blood glucose, for instance, with respect to variation in
insulin sensitivity and insulin secretion is obtained by controlling the functional
mass of pancreatic beta cells.
The already existing and more restricted homeostasis, or exact adaptation,
deals only with constant reference trajectories, i.e., setpoints. It is achievable via
an integral feedback (see, e.g., Alon et al. (1999); Briat et al. (2016); Miao et al.
(2011); Stelling et al. (2004); Yi et al. (2000))
Remark 1.1. PIDs (see, e.g., Åström et al. (2008); O’Dwyer (2009)) read:
Z
u = KP e + KI e + KD ė
(1)
where
• u, y, y ⋆ are respectively the control and output variables, and the reference
trajectory.
• e = y − y ⋆ is the tracking error,
• KP , KI , KD ∈ R are the tuning gains.
To the best of our knowledge, they are, strangely enough, more or less missing
2
in the literature on theoretical biology,1 although they lead to the most widely
used control strategies in industry.
From KP = KD = 0 in Equation (1), the following integral feedback is
deduced:
u = KI
Z
e
(2)
Compare Equation (2) with the references above on homeostasis, and Somvanshi et al.
(2015). See Abouaı̈ssa et al. (2017b), and the references therein, for an application to ramp metering on freeways in order to avoid traffic congestion.
Conditions for DC have already been investigated by several authors: Karin et al.
(2017a,b); Sontag (2017); Villaverde et al. (2017). Parameter identification,
which plays a key rôle in most of those studies, leads, according to the own
words of Karin et al. (2017b), to some kind of “discrepancy,” which is perhaps not yet fully cleared up. We suggest therefore another roadmap, i.e.,
intelligent feedback controllers as defined by Fliess et al. (2013). Many concrete applications have already been developed all over the world. Some have
been patented. The bibliography contains for obvious reasons only recent works
in biotechnology: Bara et al. (2016); Join et al. (2017a); Lafont et al. (2015);
MohammadRidha et al. (2018); Tebbani et al. (2016).2
An unexpected byproduct is derived from Remark 1.1 and the comparison in
Abouaı̈ssa et al. (2017b) between Equation (2) and our intelligent proportional
controller (Fliess et al. (2013)). Exact adaptation means now a “satisfactory”
behavior thanks to the feedback loop defined by Equation (2) when the working
conditions are “mild.” The result by Karin et al. (2016) via a mechanism for DC
based on known hormonal circuit reactions, which states that exact adaptation
does not guarantee dynamical compensation, remains therefore valid in this new
context.
1 This
is of course less the case in synthetic biology, i.e., a blending between biology and
engineering (see, e.g., Del Vecchio et al. (2016) and the references therein).
2 A rather comprehensive bibliography of concrete applications is provided
Abouaı̈ssa et al. (2017a).
3
by
This exploratory research report is organized as follows. Intelligent controllers are summarized in Section 2, where the connection between classic PIs
and intelligent proportional controllers is also presented. Section 3, which is
hevily influenced by Abouaı̈ssa et al. (2017b), defines dynamic compensation
and exact adaptation. Section 4 displays various convincing, but academic,
computer experiments. Some concluding remarks may be found in Section 5.
2. Model-free control and intelligent controllers
See Fliess et al. (2013) for full details.
2.1. Generalities
2.1.1. The ultra-local model
The poorly known global description of the plant, which is assumed for
simplicity’s sake to be SISO (single-input single output),3 is replaced by the
ultra-local model :
y (ν) = F + αu
(3)
where:
• the control and output variables are respectively u and y;
• the derivation order ν is often equal to 1, sometimes to 2; in practice ν ≥ 3
has never been encountered;
• the constant α ∈ R is chosen by the practitioner such that αu and y (ν)
are of the same magnitude; therefore α does not need to be precisely
estimated.
The following comments might be useful:
• Equation (3) is only valid during a short time lapse and must be continuously updated;
3 For
a multivariable extension, see, e.g., Lafont et al. (2015); Menhour et al. (2017).
4
• F is estimated via the knowledge of the control and output variables u
and y;
• F subsumes not only the system structure, which most of the time will be
nonlinear, but also any external disturbance.
2.1.2. Intelligent controllers
Set ν = 2. Close the loop with the following intelligent proportional-integralderivative controller, or iPID,
u=−
F − ẏ ∗ + KP e + KI
α
R
e + KD ė
(4)
where:
• e = y − y ⋆ is the tracking error,
• KP , KI , KD ∈ R are the tuning gains.
When KI = 0, we obtain the intelligent proportional-derivative controller, or
iPD,
u=−
F − ẏ ∗ + KP e + KD ė
α
(5)
When ν = 1 and KI = KD = 0, we obtain the intelligent proportional controller,
or iP, which is the most important one,
u=−
F − ẏ ∗ + KP e
α
(6)
Combining Equations (3) and (6) yields:
ė + KP e = 0
(7)
where F does not appear anymore. The tuning of KP is therefore straightforward.
Remark 2.1. See Join et al. (2017b) for a comment on those various controllers.
5
2.1.3. Estimation of F
Assume that F in Equation (3) is “well” approximated by a piecewise constant function Fest .4 The estimation techniques below are borrowed from Fliess et al.
(2003, 2008) and Sira-Ramı́rez et al. (2014). Let us summarize two types of
computations:
1. Rewrite Equation (3) in the operational domain (see, e.g., Erdélyi (1962)):
sY =
Φ
+ αU + y(0)
s
where Φ is a constant. We get rid of the initial condition y(0) by multiplying both sides on the left by
Y +s
d
ds :
dU
Φ
dY
=− 2 +α
ds
s
ds
Noise attenuation is achieved by multiplying both sides on the left by s−2 ,
since integration with respect to time is a lowpass filter. It yields in the
time domain the realtime estimate, thanks to the equivalence between
d
ds
and the multiplication by −t,
Fest (t) = −
6
τ3
Z
t
[(τ − 2σ)y(σ) + ασ(τ − σ)u(σ)] dσ
(8)
t−τ
where τ > 0 might be quite small. This integral may of course be replaced
in practice by a classic digital filter.
2. Close the loop with the iP (6). It yields:
Z t
1
⋆
Fest (t) =
(ẏ − αu − KP e) dσ
τ t−τ
Remark 2.2. From a hardware standpoint, a real-time implementation of our
intelligent controllers is also cheap and easy (Join et al. (2013)).
2.2. PI and iP
Consider the classic continuous-time PI controller
Z
u(t) = kp e(t) + ki e(τ )dτ
4 This
is a weak assumption (see, e.g., Bourbaki (1976)).
6
(9)
R
A crude sampling of the integral e(τ )dτ through a Riemann sum I(t) leads to
Z
e(τ )dτ ≃ I(t) = I(t − h) + he(t)
where h is the sampling interval. The corresponding discrete form of Equation
(9) reads:
u(t) = kp e(t) + ki I(t) = kp e(t) + ki I(t − h) + ki he(t)
Combining the above equation with
u(t − h) = kp e(t − h) + ki I(t − h)
yields
u(t) = u(t − h) + kp (e(t) − e(t − h)) + ki he(t)
(10)
Remark 2.3. A trivial sampling of the “velocity form” of Equation (9)
u̇(t) = kp ė(t) + ki e(t)
(11)
yields
u(t) − u(t − h)
= kp
h
e(t) − e(t − h)
h
+ ki e(t)
which is equivalent to Equation (10).
Replace in Equation (6) F by ẏ(t) − αu(t − h) and therefore by
y(t) − y(t − h)
− αu(t − h)
h
It yields
u(t) = u(t − h) −
e(t) − e(t − h) KP
−
e(t)
hα
α
(12)
FACT.- Equations (10) and (12) become identical if we set
kp = −
1
,
αh
ki = −
KP
αh
(13)
Remark 2.4. This path breaking result was first stated by d’Andréa-Novel et al.
(2010):
7
• It is straightforward to extend it to the same type of equivalence between
PIDs and iPDs.
• It explains apparently for the first time the ubiquity of PIs and PIDs in
the industrial world.
3. Exact adaptation and dynamic compensation
Equation (11) shows that integral and proportional-integral controllers are
close when
1. ė remains small,
2. the reference trajectory y ∗ starts at the initial condition y(0) or, at least,
at a point in a neighborhood,
3. the measurement noise corruption is low.
The following conditions might be helpful:
• the reference trajectory y ∗ is “slowly” varying, and starts at the initial
condition y(0) or, at least, at a point in its neighborhood,5
• the disturbances and the corrupting noises are rather mild.
Then the performances of the integral controller should be decent: this is exact adaptation, or homeostasis. When the above conditions are not satisfied,
dynamic compensation means that one at least of the feedback loops in Section 2.1 is negative, i.e., fluctuations around the reference trajectory due to
perturbations and input changes are attenuated.6
5 Setpoints
6 The
are therefore recovered.
wording “negative feedback” is today common in biology, but, to some extent, ne-
glected in engineering, where it was quite popular long time ago (see, e.g., Küpfmüller et al.
(2017)). Historical details are given by Zeron (2008) and Del Vecchio et al. (2015).
8
4. Two computer experiments
The two academic examples below provide easily implementable numerical
examples. They are characterized by the following features:
• KI = 0.5 (resp. KI = 1) for the integral feedback in the linear (resp.
nonlinear) case.
• α = KP = 1 for the the iP (6) in both cases.
• The sampling period is 10ms.
• In order to be more realistic, the output is corrupted additively by a zeromean white Gaussian noise of standard deviation 0.01.
4.1. Linear case
Consider the input-output system defined by the transfer function
2(s + 1)
s2 + s + 1
Several reference trajectories are examined:
(i) Setpoint and 50% efficiency loss of the actuator: see Figures 1 see 2.
(ii) Slow connection between two setpoints: see Figures 3 and 4.
(iii) Fast connection: see Figures 5 and 6.
(iv) Complex reference trajectory: see Figures 7 and 8.
In the first scenario, the control efficiency loss is attenuated much faster by the
iP than by the integral feedback. The behaviors of the integral feedback and
the iP with respect to a slow connection are both good and cannot be really
distinguished. The situation change drastically with a fast connection and a
complex reference trajectory: the iP becomes vastly superior to the integral
feedback. Exact adaptation works well only in the second scenario, whereas
dynamic compensation yields always excellent results.
9
1.2
1.3
Output
Reference trajectory
Reference
1.2
1
1.1
0.8
1
0.9
0.6
0.8
0.4
0.7
0.6
0.2
0.5
0
0.4
0
5
10
15
20
25
30
35
40
45
50
0
5
10
15
20
Time(s)
25
30
35
40
45
50
Time(s)
(a) Control
(b) Output
Figure 1: Integral feedback, constant reference trajectory, control efficiency loss
1.2
1.3
Output
Reference trajectory
Reference
1.2
1
1.1
0.8
1
0.6
0.9
0.8
0.4
0.7
0.2
0.6
0
0.5
-0.2
0.4
0
5
10
15
20
25
30
35
40
45
50
0
Time(s)
5
10
15
20
25
30
35
Time(s)
(a) Control
(b) Output
Figure 2: iP, constant reference trajectory, control efficiency loss
10
40
45
50
1.2
0.5
1
0.4
Output
Reference trajectory
Reference
0.8
0.3
0.6
0.2
0.4
0.1
0.2
0
0
-0.1
-0.2
0
5
10
15
20
25
30
35
40
45
50
0
5
10
15
20
Time(s)
25
30
35
40
45
50
Time(s)
(a) Control
(b) Output
Figure 3: Integral feedback, slow connection
1.2
0.8
1
0.6
Output
Reference trajectory
Reference
0.8
0.4
0.6
0.2
0.4
0
0.2
-0.2
0
-0.4
-0.2
0
5
10
15
20
25
30
35
40
45
50
0
5
10
Time(s)
15
20
25
30
Time(s)
(a) Control
(b) Output
Figure 4: iP, slow connection
11
35
40
45
50
0.6
1.2
0.5
1
0.4
0.8
0.3
0.6
0.2
0.4
0.1
0.2
0
0
-0.1
-0.2
0
5
10
15
20
25
30
35
40
45
50
Output
Reference trajectory
Reference
0
5
10
15
20
Time(s)
25
30
35
40
45
50
Time(s)
(a) Control
(b) Output
Figure 5: Integral feedback, fast connection
0.8
1.2
0.6
1
0.4
0.8
0.2
0.6
0
0.4
-0.2
0.2
-0.4
0
-0.6
Output
Reference trajectory
Reference
-0.2
0
5
10
15
20
25
30
35
40
45
50
0
5
10
Time(s)
15
20
25
30
Time(s)
(a) Control
(b) Output
Figure 6: iP, fast connection
12
35
40
45
50
0.8
1.6
0.7
1.4
0.6
1.2
0.5
1
0.4
0.8
0.3
0.6
0.2
0.4
0.1
0.2
0
Output
Reference trajectory
Reference
0
-0.1
-0.2
0
5
10
15
20
25
30
35
40
45
50
0
5
10
15
20
Time(s)
25
30
35
40
45
50
Time(s)
(a) Control
(b) Output
Figure 7: Integral connection, complex reference trajectory
0.8
1.4
1.2
0.6
1
0.4
0.8
0.2
Output
Reference trajectory
Reference
0.6
0.4
0
0.2
-0.2
0
-0.4
-0.2
0
5
10
15
20
25
30
35
40
45
50
0
5
10
15
20
Time(s)
25
30
Time(s)
(a) Control
(b) Output
Figure 8: iP, complex reference trajectory
13
35
40
45
50
4.2. Nonlinear case
Consider
ẏ + y = u3 + Ppert
where Ppert is a perturbation. Introduce three scenarios:
(i) Setpoint without any perturbation, i.e., Ppert = 0: see Figures 9 and 10.
(ii) Setpoint with a sine wave perturbation which starts at t = 25s, i.e.,
Ppert (t) = 0.2 sin( 2π
5 (t − 25)) if t ≥ 25s, and Ppert (t) = 0 if t ≤ 25s:
see Figures 11 and 12.
(iii) Non-constant reference trajectory without any perturbation, i.e., Ppert =
0: see Figures 13 and 14.
A clear-cut superiority of the iP with respect to the integral feedback is indisputable. The behavior of dynamic compensation (resp. exact adaptation) is
always (resp. never) satisfactory.
Remark 4.1. Do not believe that integral feedbacks are never adequate if nonlinearities occur. See
• an example related to ramp metering in Abouaı̈ssa et al. (2017b),
• theoretical investigations in Sontag (2010).
5. Conclusion
In order to be fully convincing, this preliminary annoucement on homeostasis
extensions needs of course to exhibit true biological examples. In our context
noise corruption is also a hot topic (see, e.g., Briat et al. (2016); Sun et al.
(2010)). The estimation and identification techniques sketched in Section 2.1
might lead to a better understanding (see also Fliess (2006, 2008)).
14
1.4
1.4
Output
Reference trajectory
Reference
1.2
1.2
1
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0
0.2
0
5
10
15
20
25
30
35
40
45
50
0
5
10
15
20
Time(s)
25
30
35
40
45
50
Time(s)
(a) Control
(b) Output
Figure 9: Integral feedback, constant reference trajectory, without any perturbation
1.4
1.4
Output
Reference trajectory
Reference
1.2
1.2
1
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0
0.2
0
5
10
15
20
25
30
35
40
45
50
0
Time(s)
5
10
15
20
25
30
35
40
Time(s)
(a) Control
(b) Output
Figure 10: iP, constant reference trajectory, without any perturbation
15
45
50
1.4
1.4
Output
Reference trajectory
Reference
1.2
1.2
1
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0
0.2
0
5
10
15
20
25
30
35
40
45
50
0
5
10
15
20
Time(s)
25
30
35
40
45
50
Time(s)
(a) Control
(b) Output
Figure 11: Integral feedback, constant reference trajectory, with perturbation
1.2
1.4
1
1.2
0.8
1
0.6
0.8
0.4
0.6
0.2
0.4
Output
Reference trajectory
Reference
0
0.2
0
5
10
15
20
25
30
35
40
45
50
0
Time(s)
5
10
15
20
25
30
Time(s)
(a) Control
(b) Output
Figure 12: iP, constant reference trajectory, with perturbation
16
35
40
45
50
1.2
1.2
1
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0
0
-0.2
-0.2
0
5
10
15
20
25
30
35
40
45
50
Output
Reference trajectory
Reference
0
5
10
15
20
Time(s)
25
30
35
40
45
50
Time(s)
(a) Control
(b) Output
Figure 13: Integral feedback, non-constant reference trajectory, without any perturbation
1.2
1.2
1
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0
0
-0.2
Output
Reference trajectory
Reference
-0.2
0
5
10
15
20
25
30
35
40
45
50
0
Time(s)
5
10
15
20
25
30
35
40
Time(s)
(a) Control
(b) Output
Figure 14: iP, non-constant reference trajectory, without any perturbation
17
45
50
References
Abouaı̈ssa H., Alhaj Hasan O., Join C., Fliess M., Defer D. (2017a). Energy
saving for building heating via a simple and efficient model-free control design:
First steps with computer simulations. 21st Int. Conf. Syst. Theor. Contr.
Comput., Sinaia. https://hal.archives-ouvertes.fr/hal-01568899/en/
Abouaı̈ssa H., Fliess M., Join C. (2017b). On ramp metering: Towards a better
understanding of ALINEA via model-free control. Int. J. Contr., 90, 10181026.
Alon U. (2006). An Introduction to Systems Biology: Design Principles of Biological Circuits. Chapman and Hall.
Alon U., Surette M.G., Barkai N., Leibler S. (1999). Robustness in bacterial
chemotaxis. Nature, 397, 168-171.
d’Andréa-Novel B., Fliess M., Join C., Mounier H., Steux B. (2010). A mathematical explanation via “intelligent” PID controllers of the strange ubiquity
of PIDs. 18th Medit. Conf. Contr. Automat., Marrakech.
https://hal.archives-ouvertes.fr/inria-00480293/en/
Åström K.J., Murray R.M. (2008). Feedback Systems: An Introduction for Scientists and Engineers. Princeton University Press.
Bara O., Fliess M., Join C., Day J., Djouadi S.M. (2016). Model-free immune
therapy: A control approach to acute inflammation. Europ. Contr. Conf.,
Aalborg. https://hal.archives-ouvertes.fr/hal-01341060/en/
Bourbaki N. (1976). Fonctions d’une variable réelle. Hermann. English translation (2004): Functions of a Real Variable. Springer.
Briat C., Gupta A., Khammash M. (2016). Antithetic integral feedback ensures
robust perfect adaptation in noisy biomolecular networks. Cell Syst., 2, 15-26.
Cowan N.J., Ankarali M.M., Dyhr J.P., Madhav M.S., Roth E., Sefati S., Sponberg S., Stamper S.A., Fortune E.S., Daniel T.L. (2014). Feedback control as
18
a framework for understanding tradeoffs in biology. Integr. Compar. Biol., 54,
223-237.
Cosentino C., Bates D. (2011). An Introduction to Feedback Control in Systems
Biology. CRC Press.
Del
Vecchio
meets
D.,
synthetic
Dy
A.J.,
biology.
J.
Qian
Roy.
Y.
Soc.
(2016).
Interface,
Control
13:
theory
20160380.
http://dx.doi.org/10.1098/rsif.1916.0380
Del Vecchio D., Murray R.M. (2015). Biomolecular Feedback Systems. Princeton
University Press.
Erdélyi A. (1962). Operational Calculus and Generalized Functions. Holt Rinehart Winston.
Fliess M. (2006). Analyse non standard du bruit. C. R. Math., 342, 797-802.
Fliess M. (2008). Critique du rapport signal à bruit en communications
numériques. Revue Afric. Recher. Informat. Math. Appli., 9, 419-429.
https://hal.archives-ouvertes.fr/inria-00311719/en/
Fliess M., Join C. (2013). Model-free control. Int. J. Contr., 86, 2228-2252.
Fliess M., Sira-Ramı́rez H. (2003). An algebraic framework for linear identification. ESAIM Contr. Optimiz. Calc. Variat., 9, 151-168.
Fliess M., Sira-Ramı́rez H. (2008). Closed-loop parametric identification for
continuous-time linear systems via new algebraic techniques. H. Garnier & L.
Wang (Eds): Identification of Continuous-time Models from Sampled Data,
Springer, pp. 362-391.
Join C., Bernier J., Mottelet S., Fliess M., Rechdaoui-Guérin S., Azimi S.,
Rocher V. (2017a). A simple and efficient feedback control strategy for
wastewater denitrification. 20th World Congr. Int. Feder. Automat. Contr.,
Toulouse. https://hal.archives-ouvertes.fr/hal-01488199/en/
19
Join C., Chaxel F., Fliess M. (2013). “Intelligent” controllers on cheap and small
programmable devices. 2nd Int. Conf. Contr. Fault-Tolerant Syst., Nice.
https://hal.archives-ouvertes.fr/hal-00845795/en/
Join C., Delaleau E., Fliess M., Moog C.H. (2017b). Un résultat intrigant en
commande sans modèle. ISTE OpenScience Automat., 1, 9 p.
https://hal.archives-ouvertes.fr/hal-01628322/en/
Karin O., Alon U. (2017a). Biphasic response as a mechanism against mutant
takeover in tissue homeostasis circuits. Molec. Syst. Biol., 13, 933.
Karin O., Alon U., Sontag E. (2017b). A note on dynamical compensation and its relation to parameter identifiability. bioRxiv doi:
https://doi.org/10.1101/123489
Karin O., Swisa A., Glaser B., Dor Y., Alon U. (2016). Dynamical
compensation in physiological circuits. Mol. Syst. Biol., 12:
886. DOI
10.15252/msb.19167216
Klipp E., Liebermeister W., Wierling C., Kowald A. (2016). Systems Biology
(2nd ed.). Wiley-VCH.
Kremling A. (2012). Kompendium Systembiologie – Mathematische Modellierung und Modellanalyse. Vieweg + Teubner. English translation (2014):
Systems Biology. CRC Press.
Küpfmüller K., Mathis W., Reibiger, W. (2017). Theoretische Elektrotechnik –
Eine Einführung (20. Auflage).7 Springer.
Lafont F., Balmat J.-F., Pessel N., Fliess M. (2015). A model-free control strategy for an experimental greenhouse with an application to fault accommodation. Comput. Electron. Agricult., 110, 139–149.
7 The
first edition was published in 1932. Karl Küpfmüller was its sole author. He died in
1977.
20
Menhour L., d’Andréa-Novel B., Fliess M., Gruyer D., Mounier H (2017). An
efficient model-free setting for longitudinal and lateral vehicle control. Validation through the interconnected pro-SiVIC/RTMaps prototyping platform.
IEEE Trans. Intel. Transport. Syst., 2017. DOI: 10.1109/TITS.1917.2699283
Miao H., Xia X., Perelson A.S., Wu H. (2011). On identifiability of nonlinear
ODE models and applications in viral dynamics. SIAM Rev., 53, 3-39.
MohammadRidha T., Aı̈t-Ahmed M., Chailloux L., Krempf M., Guilhem I.,
Poirier J.-Y., Moog C.H. (2018). Model free iPID control for glycemia regulation of type-1 diabetes. IEEE Trans. Biomed. Engin., 65, 199-206.
O’Dwyer A. (2009). Handbook of PI and PID Controller Tuning Rules (3rd ed.).
Imperial College Press.
Sira-Ramı́rez H., Garcı́a-Rodrı́guez C., Cortès-Romero J., Luviano-Juárez A.
(2013). Algebraic Identification and Estimation Methods in Feedback Control
Systems. Wiley.
Somvanshi P.R., Patel A.K, Bhartiya S., Venkatesh K.V. (2015). Implementation of integral feedback control in biological systems. Wiley Interdisc. Rev.:
Syst. Biol. Med., 7, 301-316.
Sontag E.D. (2010). Remarks on feedforward circuits, adaptation, and pulse
memory. IET Syst. Biol., 4, 39-51.
Sontag
ity,
E.D.
(2017).
Dynamic
compensation,
and equivariances. PLoS Comput. Biol.,
parameter
13:
identifiabil-
e1005447. doi:
10.1371/journal.pcbi.1005447
Sun L., Becskei A. (2010). The cost of feedback control. Nature, 467, 163-164.
Stelling J., Sauer U., Szallasi Z., Doyle F.J., Doyle J. (2004). Robustness of
cellular functions. Cell, 118, 675-685.
Tebbani S., Titica M., Join C., Fliess M., Dumur D. (2016). Model-based versus model-free control designs for improving microalgae growth in a closed
21
photobioreactor: Some preliminary comparisons. 24th Medit. Conf. Contr.
Automat., Athens. https://hal.archives-ouvertes.fr/hal-01312251/en/
Villaverde A.F., Banga J.R. (2017). Dynamical compensation and structural
identifiability of biological models: Analysis, implications, and reconciliation.
PLoS Comput. Biol., 13, e1005878.
https://doi.org/10.1371/journal.pcbi.1005878
Yi T.-M., Huang Y., Simon M.I., Doyle J. (2000). Robust perfect adaptation
in bacterial chemotaxis through integral feedback control. Proc. Nat. Acad.
Sci., 97, 4649-4653.
Zeron E.S. (2008). Positive and negative feedback in engineering and biology.
Math. Model. Nat. Phenom., 3, 67-84.
22
| 3cs.SY
|
1
Framework of Channel Estimation for Hybrid
Analog-and-Digital Processing Enabled Massive
MIMO Communications
arXiv:1712.08846v1 [cs.IT] 23 Dec 2017
Leyuan Pan, Student Member, IEEE, Le Liang, Student Member, IEEE,
Wei Xu, Senior Member, IEEE, and Xiaodai Dong, Senior Member, IEEE
Abstract—We investigate a general channel estimation problem
in the massive multiple-input multiple-output (MIMO) system
which employs the hybrid analog/digital precoding structure
with limited radio-frequency (RF) chains. By properly designing
RF combiners and performing multiple trainings, the proposed
channel estimation can approach the performance of fullydigital estimations depending on the degree of channel spatial
correlation and the number of RF chains. Dealing with the hybrid
channel estimation, the optimal combiner is theoretically derived
by relaxing the constant-magnitude constraint in a specific singletraining scenario, which is then extended to the design of combiners for multiple trainings by Sequential and Alternating methods.
Further, we develop a technique to generate the phase-only RF
combiners based on the corresponding unconstrained ones to
satisfy the constant-magnitude constraints. The performance of
the proposed hybrid channel estimation scheme is examined
by simulations under both nonparametric and spatial channel
models. The simulation results demonstrate that the estimated
CSI can approach the performance of fully-digital estimations in
terms of both mean square error and spectral efficiency. Moreover, a practical spatial channel covariance estimation method
is proposed and its effectiveness in hybrid channel estimation is
verified by simulations.
Index Terms—Massive MIMO, channel estimation, hybrid
precoding, millimeter wave
I. I NTRODUCTION
M
ASSIVE multiple-input multiple-output (MIMO) is potentially one of the most promising and key technologies to meet the stringent performance requirements for
the next-generation, i.e., 5G, wireless communications [1],
[2]. It has attracted considerable research interests from
both academia and industry since the seminal work [3] was
published in 2010 [4]–[7]. Most notably, with an excessive
amount of antennas mounted at the transmitter and/or receiver,
signal processing, including both transmission precoding and
receiving combining, can be greatly simplified while achieving
highly optimal performance [8]–[11]. Simple linear precoding
schemes, such as zero-forcing (ZF), are virtually optimal and
comparable to the capacity-achieving nonlinear dirty paper
L. Pan and X. Dong are with the Department of Electrical and Computer
Engineering, University of Victoria, Victoria, BC V8P 5C2, Canada (Email:
[email protected], [email protected]).
L. Liang is with the School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA (Email:
[email protected]).
W. Xu is with the National Mobile Communications Research Laboratory, Southeast University, Nanjing, Jiangsu 210096, China (Email:
[email protected]).
coding (DPC). Thanks to the employment of millimeter wave
(mmWave) bands, a large number of antennas can be packed
into a small area due to the short wavelength [12].
However, the conventional signal processing is typically
performed at the baseband, which means that the signal
received from each antenna port needs to be properly filtered, down-converted, and then sampled, where the hardware
module performing such tasks is normally referred to as a
radio-frequency (RF) chain. Analogous procedure exists for
signal transmission. Given a large number of antennas in a
massive MIMO system, it would be formidable to feed each
antenna with a dedicated RF chain due to high cost and power
consumption. To circumvent the challenging requirement of
massive RF chains, an analog/digital hybrid structure has
been proposed for massive MIMO systems operated on both
centimeter wave (cmWave) [13] and mmWave [14] bands1. On
the transmitter side of a hybrid system, low-dimensional baseband signals (after digital processing) are converted to the RF
domain, feeding a phase-shifting network to properly adjust
the phases of transmission signals which are then transmitted
to wireless channels by antennas [12], [13]. The design of the
RF and baseband precoding/combining matrices in a hybrid
structure is a challenging problem and has been extensively
studied in recent years, e.g., [14], [16]–[19]. However, these
papers all assume the availability of channel state information
(CSI). Other beamforming solutions without the explicit need
of channel knowledge depending on iterative beam trainings
and multi-stage codebooks have been developed in [20]–[23].
These solutions, however, naturally have the common disadvantage that the beamforming converges towards a single beam
which is not capable to achieve multiplexing gains or support
multiple streams in multi-user systems. Hence, the explicit
channel estimation is one of the most important elements in
reaping all the advantages and gains of massive MIMO, and
unsurprisingly has been under extensive investigation. It is a
challenging task, yet is even more difficult to fulfill under the
limited RF-chain hybrid structure constraint.
The main task of the channel estimation in a hybrid precoding system is to recover the M -dimensional channel vector
from L observations at baseband, where L (< M ) is the
number of limited RF chains. The conventional fully-digital
MIMO channel estimation methods in the literature cannot
1 According to the notation in [15], the cmWave bands generally denote
1 − 10 GHz while mmWave denotes 10 − 100 GHz.
2
be applied in a hybrid structure to obtain full CSI [24]–
[28]. So far, only a few papers have studied channel estimation with limited RF chains in hybrid structured mmWave
communications. An adaptive compressed sensing solution for
the hybrid channel estimation was proposed in [29] from the
perspective of angular sparsity of mmWave channels. Scanning
in the angular domain is performed at both the transmitter
and receiver sides. As such, the complexity and resource
consumption of the proposed channel estimation scheme is
dominated by the sparsity of channel scattering. For instance,
it will consume much more resource with higher complexity
to achieve a desired angular resolution in channel estimation
when the scattering paths are rich, compared to the case of
very sparse channels. A similar method was developed in a
frequency-selective channel scenario [30]. In [31], channel estimation with phase-shifters or switches in a hybrid structured
system was discussed. Still, it employs the high-complexity
compressed sensing method with angular domain scanning to
estimate the angle of arrival/angle of departure (AoA/AoD) of
each scattering path. Few prior literature has considered the
general problem of channel estimation in a hybrid structured
massive MIMO system without significant dependence on the
channel sparsity. Since the degree of freedom (DoF) of the
received signals at baseband is limited by the number of RF
chains, it becomes prohibitively difficult to obtain satisfactory
higher-dimensional vector channel estimates, especially for
rich scattering environments, in a hybrid precoding system
using conventional channel estimation approaches.
In this paper, we consider the uplink channel estimation
of a multi-user massive MIMO system with the hybrid RFbaseband processing structure where the base station (BS) is
equipped with a large number of antennas but driven by a
far smaller number of RF chains, and each mobile station
(MS) is equipped with a single antenna, and propose an
efficient channel estimation scheme by exploiting the spatial
correlation of massive MIMO channels where both single
and multiple trainings are investigated. One may suspect the
necessity of investigation on multiple trainings. Theoretically,
one pilot symbol is optimal, and in fact enough, for a singleantenna user to assist uplink estimation of uncorrelated MIMO
channels under some mild assumptions [32]. In a massive
MIMO system with limited RF chains, however, only much
smaller L-dimensional (compared to the number of antennas,
M ) signal is captured at baseband per training. It is, therefore,
evident that a single training symbol becomes insufficient for
the BS to conduct full-dimensional channel estimation of M
(independent) channel coefficients. To approach the performance of fully-digital channel estimation, multiple trainings
are required to achieve the DoF of fully-digital baseband
signal measurements. That is, T L observations can be utilized
to estimate the M -dimensional MIMO channels, where T is
the number of trainings. Hence, it needs to properly design
RF combiners for different training phases to capture the
channel energy and then recover the CSI as accurately as
possible. Furthermore, how many trainings are required to
achieve the performance of fully-digital channel estimation?
Empirically, T = M/L training symbols can be utilized to
achieve the DoF of a fully-digital training in the estimation
of uncorrelated channels. However, such conclusions may not
hold for correlated channels. As revealed in both [33] and
[34], the optimal number of training symbols can be reduced
due to the fact that the dimension of statistically dominant
subspaces is less than the number of antennas. In practice,
massive MIMO and mmWave channels are inevitably spatially
correlated due to the limited number of propagation paths
and angular spreads [24], [29], [35]. Thus, less than M/L
trainings could be sufficient to achieve the fully-digital channel
estimation performance in such spatially correlated channels.
Finally, we summarize the main contributions of this paper as
follows:
•
•
•
RF Combiner Design for Single Training: We investigate the channel estimation in a hybrid precoding
system and formulate an optimization problem on channel
estimation following the minimum mean square error
(MMSE) criteria. To solve the problem, the constantmagnitude constraint of the RF combiner is temporarily
relaxed. By employing the properties of Block Generalized Rayleigh Quotient, the theoretical optimizer of
the formulated optimization problem is solved and the
corresponding optimal RF combiner is designed in the
single training scenario. Analyses of the designed combiner show that the mean square error (MSE) of the
channel estimation decreases when (1) the channel is
more spatially correlated, (2) more RF chains for channel
estimation are deployed, (3) larger transmission power
for training pilots is utilized. Moreover, the closed-form
expression of the MSE is derived and verified to be
quite accurate compared to the simulation results, thus
providing useful guidelines in practical system designs.
RF Combiner Design for Multiple Trainings: We formulate the optimization problem to design RF combiners
for multiple trainings and solve it by Sequential and
Alternating methods with the assistance of the singletraining result. The Sequential method can achieve stepwisely minimum MSE with low complexity while the
Alternating one solves the joint optimization problem
iteratively with high complexity, however, achieving the
local optimum. The performance and complexity of the
proposed RF combiners are examined in simulations
under both nonparametric and parametric channel models
[36].
Spatial Channel Covariance Estimation in HybridStructured System: To perform the proposed RF combiner design and channel estimation, the spatial channel
covariance matrix needs be known by the BS to perform
the design of RF combiners for channel estimation. In
this paper, a covariance estimation method is proposed in
the hybrid-structured massive MIMO system. From simulation results, it achieves comparable spectral efficiency
with the estimated covariance matrix compared to that
with the perfect one, which proves the effectiveness of
the proposed method.
Note that the considered system structure and channel estimation scheme is also applicable to mmWave channels [15],
[17], [37].
3
Apart from the contributions, we also present the main
characteristics differentiating this paper from the existing
works as summarized below.
• Compared to the typical fully-digital massive MIMO
systems, the signal dimension in the considered system
is reduced after the processing of the phase-shifting
network. That is, it converts the received M -dimensional
RF signal to L-dimensional baseband signal where M
denotes the number of antenna elements and L represents
the number of RF chains typically satisfying L < M
in hybrid precoding/combining systems. As such, the
channel estimation performed in baseband cannot utilize
the complete training information as the fully-digital
massive MIMO does.
• Compared to existing work on channel estimation in a
hybrid-structure mmWave system, no explicit usage of
channel sparsity is required by our proposed methods to
perform the estimation, which is adaptive to both nonsparse cmWave and sparse mmWave channel scenarios.
In addition, the complexity of our proposed scheme is
determined by the number of RF chains and training
phases which are fixed in an online system, rather than the
scattering circumstance which is time-variant and difficult
to be guaranteed.
Organization: The rest of this paper is organized as follows.
Section II illustrates the system model of the massive MIMO
system with the hybrid precoding structure and the channel
models adopted in this paper. In Section III, the channel
estimation schemes in the hybrid structure are proposed and
analyzed. Section ?? proposes the method to generate the
channel covariance matrices which can be employed in practical system implementations. The numerical and simulation
results are presented in Section V. Finally, Section VI concludes our work in this paper.
Notation: The boldface capital and lowercase letters are
used to denote matrices and vectors, respectively, while the
plain lowercase letters are scalars. Superscripts (·)∗ , (·)T and
(·)H stand for the conjugate, transpose and conjugate-transpose
of a vector or matrix, respectively. IM represents the identity
matrix of size M . Operators k · k, k · kF and tr(·) denote
the Euclidean norm of a vector, the Frobenius norm and the
trace of a matrix, respectively. blkdiag{A1 , · · · , AL } stands
for the block matrix with diagonal elements as A1 , · · · , AL
in sequence. For statistical vectors and matrices, E{·} and
Var{·} are utilized to represent the expectation and variance,
respectively. Moreover, x ∼ CN (0, Σ) represents the complex
Gaussian random vector x with zero mean and covariance matrix Σ. Finally, a ≻ b means that a majorizes (or dominates)
b and vice versa.
II. S YSTEM M ODEL
We investigate a single-cell hybrid massive MIMO communication structure as shown in Fig. 1 where K single-antenna
mobile stations (MSs) are served by a base station (BS)
equipped with M antennas driven by L RF chains which are
interfaced by an analog RF phase-shifting network converting
the high-dimensional received signal to low-dimensional one
RF Chain
Baseband
Processing
L
RF Chain
Base Station
Analog
RF
Combiner
F
gK
MS K
g1
MS 1
M
Mobile Station
Fig. 1: Block diagram of massive MIMO with a hybrid
structure.
through phase-only linear combing. Here, L < M means
limited RF chains while L = M implies full chains. The
combined RF signal after the phase-shifting network is then
down-converted to baseband and sampled by analog-to-digital
converter(s) for further digital processing [12].
Due to its reciprocity in TDD operation, the channel between the BS and each MS can be obtained by the uplink
training and used for downlink precoding within a coherence
interval [38]. Different from the downlink training employed
in [39], the uplink training requires the pilot sequences transmitted by all MSs to be orthogonal to each other to avoid
pilot contamination. Hence, the duration of each pilot sequence
should be at least K symbols to guarantee the orthogonality
conditions [38]. Among all, the identity matrix IK is one of
the simplest orthogonal pilots, which implies that the MSs
transmit pilot sequentially and all others keep mute when
a MS is transmitting pilot. Moreover, the performance of
channel estimation with identity matrix as pilot is the same
as that with others under block-fading channels. On the other
hand, it is more practical to design the channel-estimation-used
analog combiner based on the corresponding spatial channel
covariance matrix of a single MS per training, which can be
understood from the perspective of hybrid beamforming that
the beamformer is toward one of the uncorrelated MSs when
it is transmitting pilot signal and leads to better performance
in sparse mmWave channels where the number of scattering
paths is limited. Consequently, the channel estimation of each
MS can be performed one by one with similar manipulations.
For simplicity, we select a single MS to illustrate the hybrid
channel estimation framework in the following elaborations.
During the uplink training period, a single-antenna MS
transmits T pilots to the BS. Considering the tth (t =
1, 2, · · · , T ) pilot training, the received baseband signal at BS
after pilot compensation can be expressed as
√
√
yt = Ft ( ρgϕt + ñt ) ϕ∗t = ρFt g + Ft nt ,
(1)
where g ∈ CM×1 represents the uplink channel from the
MS to BS which is assumed to remain static throughout a
coherence interval, ϕt stands for the tth training symbol with
ϕt ϕ∗t = 1, ρ is the pilot power per transmission, ñt ∈ CM×1
denotes the additive white Gaussian noise (AWGN) with
zero mean and unit variance, nt , ñt ϕ∗t which persists
as CN (0, IM ) AWGN, and Ft ∈ CL×M is the RF phaseshifting matrix constructed by unit-magnitude elements. For
T pilot transmissions, the received signal can be stacked as a
4
concatenated vector yc = [y1T , y2T , · · · , yTT ]T ∈ CT L×1 given
by
√
(2)
yc = ρFc g + Fd nc ,
where the pilot power is identical for each training peT T
riod, the compact matrices Fc = [FT
1 , · · · , FT ] , Fd =
T
T T
blkdiag{F1 , · · · , FT }, and nc = [n1 , · · · , nT ] .
In this paper, we consider the spatially correlated MIMO
channel model, which is a typical case in most MIMO scenarios due to the limited number of incident paths and angular
spreads on BS [24]. The spatial channel covariance matrix can
be expressed by R = E{ggH }. Without loss of generality,
we normalize tr (R) = M . In Section III, R is assumed to
be known by the BS while the estimation of the covariance
matrix in practice will be discussed in Section IV.
through numerical approaches without sustainable resource
consumption due to the non-convex constraint.
To facilitate further analysis, we temporarily drop the
constant-magnitude constraint in the following design. This is
a good way to begin with because it is capable to derive some
closed-form expressions and provide guidelines for the design
of phase-only combiners. Note that the design of phase-only
combiners will be considered in Section III-C.
Without the phase-only constraint, it can be shown that the
optimal RF unconstrained combiner has a structure characterized in the following lemma.
Lemma 1. For any optimal solution to the unconstrained RF
combiners given their singular value decompositions (SVD) as
F̃opt
= Ut [Σt 0][Vt,L Vt,R ]H ,
t
(7)
it is safe to reconstruct the optimal combiners according to
III. C HANNEL E STIMATION WITH H YBRID S TRUCTURE
For channel estimation at the BS, the task is to obtain
the channel estimate in the MMSE sense with the designed
phase-shifting matrix. By employing the MMSE aimed linear
estimator, the channel estimation problem can be formulated
as follows:
minimize E kg − ĝk2
Fc ,W
subject to: ĝ = Wyc
√
yc = ρFc g + Fd nc
(3)
Fi ∈ F , i = 1, · · · , T
where W represents the baseband processing matrix, and F
denotes the set of all feasible phase-only RF combiners. In
order to minimize MSE, the optimal solution to W is the
well-known Weinner filter [40] given by
−1
√
H
,
(4)
W = ρRFH
c ρFc RFc + RFd
where Fc is assumed to be known as a prior and RFd , Fd FH
d
for notation simplification. Using (4), the objective function in
(3) equals
o
n
2
MSE =E kg − ĝk
−1
H
Fc R
=tr R − ρRFH
c ρFc RFc + RFd
−1
−1
,
(5)
=tr R−1 + ρFH
R
F
c
c
Fd
where the last equality holds due to the Woodbury matrix
identity [41] and it is assumed that R is in full rank. If R
is rank deficient, remedies similar to the one discussed in [33]
can be applied which does not affect the MSE. Combining
(3)–(5), the primal channel estimation problem is readily
equivalent to the following problem as
−1
−1
minimize tr R−1 + ρFH
c RFd Fc
Fc
(6)
subject to: Ft ∈ F , t = 1, · · · , T.
The optimization problem (6) aims to find the optimal and
feasible RF combiners Ft (t = 1, · · · , T ) for the T -length
training sequence. However, it is difficult to directly obtain
the optimal solution in either closed-form expressions or
H
= Vt,L
,
Fopt
t
which yields the same MSE performance as
t = 1, · · · , T
(8)
F̃opt
t
does, where
Proof: See Appendix A
Lemma 1 indicates that the optimal unconstrained RF
combiner Fopt
can always be restricted as row-unitary and
t
is thus independent of both Ut and Σt . Hence, we can set
both of them to be identity matrices and equivalently simplify
the MSE in (5) as
!−1
T
X
,
MSE = tr Λ−1 + ρ
Ṽt ṼtH
(9)
t=1
where Ṽt , UH Vt,L , Λ = diag{λ1 , · · · , λM } and U are the
eigenvalues and corresponding eigenvectors of R, i.e., R =
UΛUH , respectively. Without loss of generality, it is assumed
that the eigenvalues are arranged
the decreasing order, i.e.,
Pin
M
λ1 ≥ · · · ≥ λM , and hence
t=1 λt = tr(R) = M . In
addition, there exists ṼtH Ṽt = IL due to the unitary property
of both U and Vt,L . To design the optimal RF combiners, we
first investigate the single-pilot strategy, i.e., T = 1, and then
extend to the multiple pilot trainings where T > 1.
A. Optimal Combiner Design of Single Training
When T = 1, the MSE expressed in (9) reduces to
−1
Ṽ1H Λ
MSE = tr Λ − ΛṼ1 Ṽ1H ΛṼ1 + ρ−1 IM
(10a)
−1
H 2
−1
H
Ṽ1 Λ Ṽ1 ,
= tr(Λ) − tr
Ṽ1 Λ + ρ IM Ṽ1
(10b)
where (10a) follows from the Woodbury matrix identity [41]
and (10b) is derived on the fact that tr(AB) = tr(BA). From
(10), the design of the unconstrained optimal RF combiner in
(6) can be reformulated as the following optimization problem:
−1 H 2
H
−1
maximize tr
Ṽ1 Λ + ρ IM Ṽ1
Ṽ1 Λ Ṽ1
Ṽ1
subject to: Ṽ1H Ṽ1 = IL .
(11)
5
To solve (11), we employ the block generalized Rayleigh
quotient which is presented as Lemma 2 in Appendix B.
Applying Lemma 2, the objective function of (11) is the
block generalized Rayleigh quotient of Ṽ1 with respect to the
pencil (Λ2 , Λ + ρ−1 IM ). The property of block generalized
Rayleigh quotient described by Lemma 2 indicates that the
maximizer of (11) is the matrix spanned by the L generalized
eigenvectors corresponding to the L largest eigenvalues of the
pencil (Λ2 , Λ + ρ−1 IM ) if the matrix happens to satisfy the
constraint in (11). As Λ2 and Λ+ρ−1 IM are both diagonal, the
generalized eigenvalues and corresponding
ieigenvectors can
h
be directly obtained as Λ̃ = λ̃1 , · · · , λ̃M and Ũ = IM ,
respectively, where λ̃i = λ2i /(λi + ρ−1 ) (i = 1, · · · , M )
and λ̃1 ≥ · · · ≥ λ̃M according to the decreasing order of
λi . Therefore, the maximizer of (11) is constructed by the
eigenvectors corresponding to the largest L eigenvalues of the
pencil (Λ2 , Λ + ρ−1 IM ), denoted as
Ṽ1opt = Ũ[1:L] = IM[1:L] ,
(12)
where (Ṽ1opt )H Ṽ1opt = IL satisfies the constraint in (11). Now,
we can arrive at the following Theorem 1 which presents
the optimal design of the unconstrained single-training RF
combiner.
Theorem 1. The optimal unconstrained RF combiner for the
single pilot strategy can be designed as follows:
Fopt
= (U[1:L] )H ,
1
(13)
where U is the matrix spanned by the eigenvectors of R.
Proof: By substituting (12) into (8) with the transition
defined by Ṽt , UH Vt,L , the optimal unconstrained singletraining RF combiner can be obtained.
Theorem 1 implies that the RF combiner should receive the
training pilots along the largest L dominant eigen-directions
of R to minimize the MSE of channel estimations. More
explicitly, we substitute (12) into (10) and obtain the optimal
MSE as follows:
−1 !
IL
−1
MSE =tr
Λ +ρ
0M−L
=
L
X
l=1
M
X
λl
+
λl
1 + ρλl
=M −
l=L+1
L
X
l=1
λ2l
,
λl + 1/ρ
(14)
where (14) P
is obtained by applying the power constraint
M
defined by
i=1 λi = M . Prior to stating more insights
obtained from (14), we define a useful concept of one channel
being more spatially correlated than another in the following
way.
Definition 1. Let channel vectors g1 ∈ CM and g2 ∈ CM
have covariance matrices R1 ∈ RM×M and R2 ∈ RM×M ,
respectively. We say g1 is more spatially correlated than g2
if and only if λ1 ≻ λ2 , where λ1 and λ2 are composed of
the eigenvalues sorted in a descending order of R1 and R2 ,
respectively.
By Definition 1 and observing (14), we can conclude the
following corollaries on the design insights of the optimal
unconstrained RF combiner revealed in Theorem 1.
Corollary 1. Given the number of RF chains and pilot
power fixed, the MSE of channel estimation with the optimal
unconstrained single-training RF combiner decreases if the
channel is more spatially correlated.
Proof: See Appendix C.
Corollary 2. Given the channel correlation and pilot power
fixed, the MSE of channel estimation with the optimal unconstrained single-training RF combiner decreases with more RF
chains deployed.
Corollary 3. Given the channel correlation and number of
RF chains fixed, the MSE of channel estimation with the
optimal unconstrained single-training RF combiner decreases
with increasing pilot power.
Corollary 2 and 3 can be directly proved by checking the
fact that the MSE in (14) monotonously decreases with the
increase of ρ and/or L.
B. Combiner Design of Multiple Trainings
In a typical fully-digital MIMO system, the M -dimensional
signal can be observed at baseband to estimate the M dimensional MIMO channels. However, the hybrid structure
MIMO system can only capture L-dimensional signal from the
phase-shifting network fed by M antennas for each training.
Note that L < M in the limited RF chain scenario. Hence,
the observed low-dimensional signal is not sufficient to recover
the high-dimensional channel information for a single training.
To approach the performance of fully-digital channel estimation, multiple trainings can be employed to achieve the DoF
of fully-digital baseband signal measurements, namely, T L
observations can be utilized to estimate the M -dimensional
MIMO channels. Typically, it is assumed that T L ≤ M .
Therefore, the RF combiner for each training phase needs to
be properly designed to capture the channel energy for channel
recovery as accurate as possible.
In this subsection, we investigate the design of unconstrained combiners for multiple trainings, i.e., F1 , · · · , FT .
Intuitively, the multiple trainings can be performed by simply
repeating the single-pilot training for multiple times. Inspired
by Theorem 1 where the most dominated L eigen-directions of
R are selected to construct the combiner, it is heuristic to select the most dominated T L eigen-directions to establish the T
combiners for multiple trainings, namely, the Block Selection
method, where the RF combiner for the tth (t = 1, 2, · · · , T )
training can be composed of the tth most dominated Leigenvector block of R, as expressed by
Fopt
= (U[(t−1)L+1:tL] )H .
t
(15)
Evidently, the Block Selection method has low complexity, yet
leading to non-optimal performance.
In order to improve the performance of channel estimation
with multiple trainings, we formulate the optimization problem
6
to design combiners for multiple trainings by recalling (9)
shown as follows:
maximize MSE
Ṽt
subject to: ṼtH Ṽt = IL ,
(16)
t = 1, · · · , T
where the MSE of the objective function is expressed by
(9). The designed Ṽt can be employed to construct the RF
combiners according to Theorem 1. Hence, it is necessary to
find the optimizer of (16) to design RF combiners for multiple
trainings. Dealing with the optimization problem, we propose
two methods to solve it in the following part of this section.
1) Sequential Optimization: It has been investigated in
Subsection III-A that the closed-form expression of the optimal combiner can be obtained in the single-training scenario.
For multiple trainings, however, it is difficult to obtain the
global optimal solution directly. Nevertheless, we propose
a sequential approach, namely, the Sequential Optimization
(short as Sequential) method, to minimize MSE step-wisely
when T > 1. To illustrate the Sequential Optimization method,
recall (9) and reformulate it in the following form:
−1
−1
H
,
(17)
MSE = tr
ΓT + ρṼT ṼT
where
Γ−1
t
=
(
H
Γ−1
t−1 + ρṼt−1 Ṽt−1 , t > 1
−1
Λ ,
t = 1.
(18)
Following the iterative definition of Γ−1
denoted by (18), the
t
MSE can be minimized by the Sequential Optimization method
as follows.
Firstly, by setting T = 1 the problem reduces to the singlepilot training case which can be solved as presented in Section
III-A.
Subsequently, with T = 2, the MSE denoted by (17) can
be reformulated as
−1
−1
H
.
(19)
MSE = tr
Γ2 + ρṼ2 Ṽ2
By applying the same manipulations as in Section III-A, the
optimal Ṽ2 to minimize the MSE expressed by (19) can
be obtained. Similar to (12), the optimal Ṽ2 , say Ṽ2opt , is
constructed by the most dominated L eigen-directions of the
pencil (Γ22 , Γ2 +ρ−1 IM ). In other words, Ṽ2opt is combined by
L generalized eigenvectors corresponding to the L largest generalized eigenvalues of (Γ22 , Γ2 + ρ−1 IM ). For ease of exposition, we denote Γ2 = diag{γ1,2 , · · · , γM,2 }. The generalized
eigenvalues and eigenvectors of pencil (Γ22 , Γ2 + ρ−1 IM ) can
T
be obtained as γ̃2 = [γ̃1,2 , · · · , γ̃M,2 ] and IM , respectively,
2
−1
where γ̃j,2 = γj,2 /(γj,2 + ρ ) for j = 1, · · · , M . Therefore,
Ṽ2opt = IM[j1 ,··· ,jL ] ,
(20)
where the indices [j1 , · · · , jL ] are the first L numbers of
[j1 , · · · , jM ] which follow from the descending order of γ̃i,2
shown as γ̃j1 ,2 ≥ · · · ≥ γ̃jM ,2 .
In the similar manner as T = 2, we can obtain each ṼTopt
for any T > 2 successively according to the generalized
eigenvalues and eigenvectors of pencil (Γ2T , ΓT + ρ−1 IM ).
Finally, the combiners are derived by applying the definition
of Ṽi and Lemma 1. For instance, the optimal F2,opt can be
expressed by
(21)
Fopt
= UH
2
[j1 ,··· ,jL ] .
By reviewing (18) and the Sequential Optimization method,
a natural question arises here: Whether Γt is invertible or
not? To check this question, we start with considering a twostage training with Γ1 and Γ2 . Note that the known covariance
matrix R is assumed to be symmetric and full-rank2, thus it
is positive definite with all eigenvalues positive. Hence, Γ1
is composed of positive diagonal elements which indicates
that it is invertible. By setting T = 2 in (18), we have
−1
H
Γ−1
2 = Γ1 + ρṼ1 Ṽ1 . In addition to positive diagonal Γ1 ,
the optimal Ṽ1 solved in Section III-A is shown to be IM[1:L] .
H
Hence, it is guaranteed that Γ−1
1 + ρṼ1 Ṽ1 is always diagonal
with positive entries for ρ > 0. So is Γ2 invertible. Using the
similar manipulation, one can successively verify that ΓT −1
is composed of positive diagonal entries and hence ΓT is
diagonal and invertible for any T > 2.
Remark 1. In the aforementioned illustrations, each Ṽtopt
(t = 1, · · · , T ) during the Sequential Optimization has the
same structure in which they are constructed by column-wise
permutations of IM which denotes the set of generalized
eigenvectors of pencil (Γ2t , Γt + ρ−1 IM ), and the selected
columns correspond to the L largest generalized eigenvalues
of this pencil. Accordingly, each combiner for the multiple
trainings, namely Fopt
t , is the Hermitian of the matrix constructed by the consistent column-wise permutations of U
as the construction of Ṽtopt . Note that the eigenvectors of
the above pencil always compose an identity matrix due to
diagonal Γt .
Remark 2. Inevitably, the optimum Ṽtopt (t = 1, · · · , T )
obtained by the Sequential Optimization method is not a global
minimizer of the MSE in (9). Nevertheless, it minimizes the
MSE reformulated by (17) for each step of iterations, namely
the step-wise minimizer. The performance of such sub-optimal
combiners will be evaluated by numerical results in Section
V.
2) Alternating Optimization: For further performance enhancement, we solve (16) via joint optimization over Ṽt ’s. It
in many cases can achieve near optimal performance via the
alternating optimization to solve non-convex problems. More
specifically, we consider fixing all the other Ṽt (t 6= j) while
optimizing only a single Ṽj . And then, iterations are taken
to update each Ṽj (j = 1, · · · , T ) alternatively until convergence. This solution is named as Alternating Optimization
(short as Alternating) throughout this paper.
For an explicit exposure of alternating optimization, we
reformulate the MSE expressed by (9) to separate Ṽj from
the sum as given by
−1
H
Ṽ
, j = 1, · · · , T
+
ρ
Ṽ
MSE(Ṽj ) = tr
Q−1
j j
j
(22)
2 If R is rank deficient, the method mentioned below (5) can be employed
for adjustment.
7
PT
where Q−1
= Λ−1 + ρ t=1,t6=j Ṽt ṼtH . With given values
j
of Ṽt (t 6= j), Ṽj can be solved in the same manner as
Subsection III-A where Lemma 2 is applied. In this paper, the
relative MSE increment of each iteration round, namely,
(n)
(n−1)
(n)
(n−1)
ǫj = MSE Ṽj
− MSE Ṽj
,
MSE Ṽj
will be taken as the error measurement. The iterations continue
(n)
until ǫj falls below a prescribed tolerance ǫ and the last
(n)
iterate Ṽj is taken as the solution.
Finally, the unconstrained near-optimal RF combiners are
computed from the solved Ṽtopt (t = 1, · · · , T ) by applying
Lemma 1. The alternating optimization method is summarized
step-by-step in Algorithm 1.
Algorithm 1 Alternating Optimization to Design Optimal
Unconstrained Combiners for Multiple Trainings
Input: M , T and R
Output: Optimal unconstrained RF combiners for multiple
trainings, Fopt
(t = 1, · · · , T )
t
(0)
1: Initialize n = 0 and Ṽt for t = 1, · · · , T
2: loop
(n)
(n)
3:
Use Ṽt (t 6= j) to obtain Ṽj which minimizes
MSE expressed by (22) for j = 1, · · · , T sequentially
(n)
and update Ṽj with the new solved values
(n)
4:
5:
6:
7:
8:
9:
10:
11:
if
MSE Ṽj
(n−1)
−MSE Ṽj
(n−1)
MSE Ṽj
< ǫ then
Stop loop
else
Update n by 1 (i.e., n = n + 1)
end if
end loop
Compute Fopt
with Ṽtopt by applying Lemma 1
t
opt
Output Ft
Since Algorithm 1 is implemented via iterations, the convergence issue is needed to be addressed here. During each
(n)
iteration, the solution Ṽj always minimizes (22). Hence, the
objective MSE is monotonically decreasing as the increasing
of iterations until it achieves the minimal MSE. This proves
the convergence of Algorithm 1.
C. Design of Phase-only RF Combiners
In the previous subsections, we first omit the phase-only
constraints when designing the RF combiners. However, the
RF combiners are implemented using phase shifters in practical hybrid systems, which can only perform phase rotations
on the received RF signals. Hence, the magnitude of each
element of a RF combiner matrix must keep constant. In
this subsection, we propose a method to design the phaseonly RF combiners based on the unconstrained optimal ones.
Moreover, the performance loss incurred by the constantmagnitude constraint is characterized by numerical evaluation
shown in Section V-A3.
Inspired by the fact that the phase shifters only perform
phase adjustments on the received signal, we take out the phase
of each unconstrained combiner and construct the constantmagnitude RF combiners according to
(i,j)
Ft
(i,j)
= ejφt
(i,j)
denotes the phase
where φt
(i,j)
(i,j)
(i,j)
i.e., Ft,opt = at ejφt .
, t = 1, · · · , T,
(23)
of the (i,j)th element of Ft,opt ,
It is interesting to find that such heuristic phase-only RF
combiners achieve desirable performance via simulation verifications in Section V. Besides its effectiveness, the very low
complexity is another advantage of this design. Furthermore,
the performance of the phase-only combiners designed above
will be examined and analyzed in Subsections V-A3 and V-A4
in terms of both MSE of channel estimation and spectral
efficiency in hybrid precoded multiuser massive MIMO communications.
IV. E STIMATION OF S PATIAL C HANNEL C OVARIANCE
The prior knowledge of spatial channel correlations, which
can be characterized by the channel covariance matrix, is one
of the most important prerequisites to improve the performance
of not only channel estimations but also precoding designs
[42]. While it has been assumed to be known by the BS in
our proposed channel estimation framework, the covariance
estimation should not be overlooked in practice. Due to the
limited RF chains, it is infeasible to estimate the covariance
matrix in hybrid precoding systems by employing conventional
approaches [43]–[46]. Dealing with this problem, some of the
prior works have made their efforts, but still have limitations
when applied to practical hybrid-structured massive MIMO
or mmWave systems. [47] has proposed a parallel subspace
estimation and tracking by recursive least squares from partial
observations, namely the PETRELS method, which can be
employed in the hybrid massive MIMO system to perform
covariance estimation, whereas tremendous observations are
needed to achieve desirable performance due to no consideration on the symmetry property of a spatial channel covariance matrix. A less complicated method has been proposed
in [48], however, by employing a parallel switch network
which increases the hardware cost. In [49], a compressive
sensing based spatial channel covariance estimation has been
proposed while a strong condition is assumed that AoAs do not
change during the long-term covariance estimations which is
impractical. In this section, we propose a practical covariance
matrix estimation method which can be employed in a hybridstructured massive MIMO system with limited RF chains. The
estimates of spatial channel covariance matrix can be utilized
in our proposed channel estimation framework.
Let T = M/L and consider the ith (i = 1, · · · , Nc )
coherence interval, (2) can be rewritten as
√
(24)
yc [i] = ρFc [i]g[i] + Fd [i]nc [i],
where Fc [i] ∈ CM×M denotes the compacted phase
shifting matrix in the ith coherence interval, Fd [i] ,
T
T
ML×1
[FT
is the AWGN and Nc
1 [i], · · · , FT [i]] , nc [i] ∈ C
stands for the number of coherence intervals during which the
spatial channel
covariance matrix can be viewed constant, i.e.,
R = E g[i]g[i]H . For the purpose of simplifying trainings,
8
By substituting (26) into (25), it derives the estimate of the
spatial channel covariance matrix.
Note that it may need many trainings to approximate the
covariance matrix of the received signal with its sample covariance, which increases the overhead of channel estimation.
One may concern how much resource is consumed to estimate channel covariance in practical systems. Theoretically,
the covariance matrix exploits the second-order statistics of
the Wide-Sense Stationary Uncorrelated Scattering (WSSUS)
channel nature which varies slowly over time and frequency,
and a one-time estimate can be used for long-term communications [50]. In fact, the experimental results in [51] have
revealed that the spatial covariance matrix can be viewed
constant over 22.6 s, 9.7 s and 4.8 s in the urban, rural
and indoor environment, respectively. Considering a system
with 10 MSps, for example, the estimated spatial channel
covariance can be used for 1 × 109 times within 10 s, which is
far more than the consumption of channel usage on covariance
estimation. Therefore, the overhead of covariance estimation
is negligible in practice. Apart from overhead, the estimation
accuracy of covariance matrix and its impact on channel
estimation performance is another issue need to be considered
in practical systems. According to (25) and (26), the estimation
accuracy of R is determined by R̂yc , which can be increased
by extending the averaging time. In Section V, simulation
results will demonstrate the number of channel usages needed
to achieve the required estimation accuracy.
V. N UMERICAL
AND
S IMULATION R ESULTS
In this section, we evaluate the performance of the proposed
channel estimation scheme by employing both unconstrained
and phase-only RF combiners to perform uplink channel
estimation in the massive MIMO system with the hybrid
precoding structure. The evaluations are performed with nonparametric channel models followed by parametric ones.
In simulations, we consider both single-user and multi-user
scenarios. In the single-user scenario, a single-antenna MS is
served by a BS deployed in the hybrid analog/digital system
structure which is shown by Fig. 1 with K = 1. The multiuser scenario adopts K = 8 single-antenna MSs served by
the BS with the same structure as that employed in the singleuser case. Without explicit indications, the BS is equipped with
M = 64 antennas followed by L = 8 RF chains. As the power
of noise is normalized to 1, the transmission power ρ can be
used to denote the received signal-to-noise ratio (SNR). Each
simulation result is obtained by averaging over 10, 000 channel
realizations. Moreover, we adopt the low-complexity hybrid
precoding scheme proposed by Liang et al. in [17] to preprocess the transmission signal in downlink communications.
A. Performance with Nonparametric Channel Model
In this subsection, the nonparametric channel model denoted
by g = R1/2 h is considered, where h ∈ CM×1 stands for the
small-scale fading of the channel constructed by independently
and identically distributed (i.i.d.) Gaussian random variables,
i.e., h ∼ CN (0, IM ), and R is modeled by exponential
correlation model as [R]m,n = a|m−n| with [R]m,n denoting
the (m, n)th element of R and 0 ≤ a < 1 being a real number
that controls the correlation introduced to the channel model
[24], [25]. Here, a larger a corresponds to a more spatially
correlated channel and g degenerates to an i.i.d. Rayleigh
fading channel when a = 0.
0
Normalized MSE (dB)
it is durable to employ the same phase shifting matrix for each
coherence interval, i.e., Fc [i] = Fc . Further, we can select an
invertible Fc from the feasible constant-magnitude matrix set
F for training purpose, e.g., the DFT matrix. Hence, the spatial
covariance matrix of the channel can be derived from (24) as
−1 H
1
,
(25)
Fc
Ryc − Fd FH
R = F−1
d
c
ρ
where Ryc = E yc [i]yc [i]H denoting the covariance matrix
of the received signal.
From (25), the estimation of spatial channel covariance
matrix can be converted to that of the received signal covariance, where both phase shifting matrix Fc and pilot
transmission power ρ are assumed to be known by the BS. In
practice, the statistical covariance of the received signal can be
approximated with the sample covariance matrix constructed
as follows:
Nc
1 X
R̂yc =
yc [i]yc [i]H .
(26)
Nc i=1
L=[4,8,16,32,48]
-5
-10
-15
-20
-25
-20
Unconstrained combiner simulation
Unconstrained combiner closed-form
DFT combiner
Fully-digital estimation
-10
0
10
20
SNR (dB)
Fig. 2: Performance comparison of the single-training hybrid
channel estimation with different RF chains. (M = 64, a =
0.8)
1) Single Training Performance Evaluation: Fig. 2 shows
the normalized MSE performance of the hybrid channel estimations with the designed single-training unconstrained combiner at different numbers of RF chains. As a benchmark, the
performance of the fully-digital estimation is also presented.
Note that the fully-digital estimation is performed with single
training throughout this section. From the comparison, it is
evident to find that the MSE decreases with the increase of
the pilot training power, which verifies Corollary 3. Moreover,
with more RF chains equipped at the BS, the channel estimation performance is also improved, which confirms Corollary
2. Note that in Fig. 2, we also plot the estimation performance
9
with RF combiners composed of the columns randomly selected from a discrete Fourier transform (DFT) matrix. Its poor
performance as compared to the designed combiners proposed
in this paper justifies the necessity to properly design RF
combiners and the effectiveness of our design. Furthermore, it
is observed that the derived closed-form expressions are quite
accurate in characterizing the MSE performance of the channel
estimations throughout the whole SNR range, thus providing
valuable guidelines in practical system designs. Finally, the
figure also shows that the performance gap between the
fully-digital estimation and the limited-chain hybrid channel
estimations is tolerably small at low SNR regions. There is
only about 0.5 dB MSE gap between the hybrid 16-chain and
the conventional fully-digital estimations at SNR=0 dB, where
three-quarter RF chains are saved with the hybrid structure.
However, one may argue: how about the performance gap at
high SNR regions? Due to limited chains, the single training
is not sufficient to achieve the performance of fully-digital
channel estimation. To improve the performance of the hybrid
channel estimations at full SNR regions, multiple trainings are
introduced in this paper and the performance is examined in
the next subsection.
0
-5
0
T=[1,2,4,6,8]
-10
Normalized MSE (dB)
Normalized MSE (dB)
T=[1,2,4,6,8]
to find that even T = 2 trainings can outperform the fullydigital estimation. This phenomenon reveals the fact that, when
the dimension of the dominated subspaces of the correlated
channels is far less than M , the limited-chain estimation can
suppress noise better than the fully-digital one due to multiple
trainings, which results in the distinct performance gap in
the low SNR regions. On top of that, more performance
differences can also be compared from the observation of Fig.
3. In the figure, the normalized MSE of the Block Selection
method denoted by the triangle curve is always larger than
or equal to that of the other two methods, i.e., the Sequential
and Alternating solutions. The performance gap is particularly
obvious at the low SNR regions. Such performance loss of
Block Selection is caused by the fact that the noise is not
considered when designing the combiners for the second
and later pilot trainings. However, little performance gap
can be observed between Block Selection and the others at
high SNR regions due to the less significance of noise to
channel estimation. By comparing the performances between
the Sequential and Alternating solutions, it is evidenced that
they achieve almost the same MSE no matter at high or low
SNR regions which implies that the Sequential method can
achieve the local optimum in solving (16). On the other hand,
the lower designing complexity of the Sequential designates
its superiority than the Alternating method.
-15
-20
-25
-20
Block selection
Sequential solution
Alternating solution
Fully-digital estimation
-10
0
10
20
-5
-10
-15
-20
SNR (dB)
Fig. 3: Performance comparison of the multiple-training hybrid
channel estimation with different training times. (M = 64,
L = 8, a = 0.8)
2) Multiple Trainings Performance Evaluation: In this
paper, the design of unconstrained combiners for multiple
trainings is formulated as (16), which are solved by both
Sequential and Alternating approaches alongside of the intuitive Block Selection method. The performance is examined
in Fig. 3 with T = 1, 2, 4, 6 and 8. The figure shows that the
normalized MSE of the hybrid channel estimation decreases
as the increasing of training times due to the increased DoF
of baseband observations. Additionally, it is interesting to
find that the MSE performance with L = 8 and T = 8
almost achieves that of the fully-digital estimation at high
SNRs. This phenomenon verifies that the performance of fullydigital channel estimation can be achieved by the limited-chain
hybrid channel estimation with sufficient DoF of trainings,
i.e., T × L = M . In the low SNR regions, it is evident
-25
-20
Phase-only combiners
Unconstrained combiners
Fully-digital estimation
-10
0
10
20
SNR (dB)
Fig. 4: Performance comparison of channel estimation between
unconstrained and phase-only combiners with the Sequential
method under different pilot trainings. (M = 64, L = 8, a =
0.8)
3) Performance Evaluation of Phase-only RF Combiners:
In the previous two subsections, the performance of the hybrid
channel estimation with unconstrained RF combiners are evaluated. However, only phase shifters can be employed in the
current practical applications with hybrid precoding structure,
which means that only phase-only combiners are applicable
in practice. In this subsection, we examine the performance
in hybrid channel estimation with the phase-only combiners
and compare it to that of the corresponding unconstrained
ones, where the combiners for multiple trainings are designed
by employing the Sequential method. In Fig. 4, the dashed
10
Normalized MSE (dB)
a=[0.1,0.3,0.5,0.7,0.9]
-5
-10
-15
Phase-only combiners
Unconstrained combiners
-20
-10
0
10
20
SNR (dB)
Fig. 5: Channel estimation performance of both unconstrained
and phase-only combiners with the Sequential method under
different channel correlations. (M = 64, L = 8, T = 6)
lines marked by triangles denote the MSE performance of the
hybrid channel estimation with phase-only combiners which
are designed according to (23), where the performance of the
unconstrained combiners are denoted by the circled solid lines.
From comparisons, it is obvious that the performance loss
caused by the phase-only combiners is tolerable at high SNR
regions and is negligible throughout the low SNR regions.
For instance, there is only 1.5 dB MSE loss at SNR=20
dB while 0.08 dB loss at SNR=−20 dB when T = 8. The
figure also shows that the performance loss is smaller with
fewer trainings, i.e., smaller T . Therefore, the hybrid channel
estimation with phase-only combiners achieve the desirable
performance measured by the normalized MSE. Furthermore,
we also present the performance of hybrid channel estimation
at different channel correlation shown in Fig. 5. It is evident
that the normalized MSE of the hybrid channel estimation
decreases, thus the performance increases, as the increasing
of a, namely, with more spatially correlated channels, which
verifies Corollary 1. Moreover, it is obvious that the performance between the phase-only and unconstrained combiners
is tight throughout the SNR regions, e.g., there exist only 0.7
dB of MSE gap at SNR=20 dB when a = 0.9.
4) Spectral Efficiency Evaluation with Hybrid Channel
Estimation: Apart from the normalized MSE performance
presented in previous subsections, the spectral efficiency of
the hybrid channel estimation is examined in this part. In
the simulation, the spatial channel covariance matrix of each
user is estimated by the method proposed in Section IV with
300 training intervals. The following channel estimation is
performed with the estimated covariance matrix. Within each
coherence interval, K MSs transmit orthogonal training pilots
to the BS for channel estimations during uplink communications, while the BS broadcasts data to all MSs during the
downlink communications by employing the hybrid precoding
with the estimated channels. Note that the uplink and downlink
channels are assumed reciprocal in this simulation and the
spectral efficiency is calculated over the downlink data transmissions. Fig. 6 shows comparisons of the spectral efficiency
with the precoding scheme designed from the perfect CSI and
estimated CSI with perfect and estimated covariance matrices,
respectively, where the phase shifters are considered both ideal
with no quantizations and non-ideal with 2-bit, 3-bit and 4bit quantizations, respectively. From the figure, we observe
that the spectral efficiency curves denoting the estimated and
ideal covariance matrices are very tight which means that
the estimated covariance matrix with 300 training intervals is
accurate enough to be comparable with the ideal one when
employed in hybrid channel estimations. By recalling the
example in Section IV, it is shown that the overhead of
covariance estimation is only 2.4 × 10−6 per user, which is
negligible in practice. The figure further shows that the spectral
efficiency with the estimated CSI approaches that with the
perfect one, especially in low SNR region which is always
the scenario in realistic massive MIMO systems. Moreover,
we can see that the hybrid precoding scheme with estimated
CSI achieves highly desirable performance as compared to
the gene-assisted system where perfect CSI is available to
the transmitter, particularly in the low SNR regions, where
the negligible performance gap is introduced by the constantmagnitude constraint. Finally, the curves denoting different
phase quantizations show that the 2-bit quantization is not
sufficient to achieve the no-quantization performance while
the 3-bit, especially the 4-bit, quantized phase shifters produce
comparable spectral efficiency to the no-quantized ideal phase
shifters, with only slight performance gap, which means that
our proposed hybrid channel estimation is insensitive to the
quantization of phase shifters.
70
Spectral efficiency (bits/s/Hz)
0
60
Perfect CSI
Estimated CSI, perfect covariance
Estimated CSI, estimated covariance
50
40
No quant.
4-bit quant.
3-bit quant.
2-bit quant.
30
20
10
0
-20
-10
0
10
20
SNR (dB)
Fig. 6: Spectral efficiency performance with the hybrid precoding scheme using both estimated and perfect CSI where the
non-perfect CSI is estimated with both perfect and estimated
spatial channel covariance matrix. (M = 64, L = 8, K = 8,
T = 8, Nc = 300, ρ = 10 dB and a = 0.8)
B. Performance with Spatial Channel Model
In Subsection A, the nonparametric channel models are
employed in simulations to evaluate the performance of the
11
proposed channel estimation design in the hybrid precoding
system. In this subsection, the proposed channel estimation
framework is evaluated in the practical spatial channel model
(SCM) under suburban macro, urban macro and urban micro
environments. The SCM channels are generated from the
open source software provided by [52] with default settings
(6 scattering paths with 20 subpaths each, which is rich
scattering) and random seed 1. The spatial channel covariance
matrix is estimated by the proposed method in Section IV. Fig.
7 shows the spectral efficiency performances of the perfect
and estimated CSI obtained by the proposed method and by
the compressed sensing based channel estimation in [29],
respectively. From the figure, it is obvious that the spectral
efficiency with the proposed channel estimation method almost
overlaps that with the perfect CSI in all three environments
with only visible minor performance gap in the high SNR
region, i.e., 15 to 20 dB, which explicitly show the high
accuracy of the proposed channel estimation frame. Note that
only 6 trainings are employed in the simulation which verifies
that less than M/L trainings can be sufficient to achieve
the performance of the perfect CSI in spatially correlated
channels. Moreover, the comparisons with the compressed
sensing based hybrid channel estimations proposed in [29]
evidently show that our proposed method is more effective
in the rich scattering channels which verifies that no sparsity
is required by our proposed method to estimate CSI.
Spectral efficiency (bits/s/Hz)
80
70
60
50
40
A PPENDIX
A. Proof of Lemma 1
By applying the eigenvalue decomposition of a positivedefinite R, i.e., R = UΛUH , the objective function in (6)
can be derived as
−1
−1
MSE =tr R−1 + ρFH
c RFd Fc
−1
−1
=tr Λ−1 + ρUH FH
R
F
U
c
c
Fd
! !−1
p
X
H −1
,
FH
Fi U
=tr Λ−1 + ρUH
i (Fi Fi )
i=1
Perfect CSI: suburban macro
Perfect CSI: urban macro
Perfect CSI: urban micro
Estimated CSI: suburban macro
Estimated CSI: urban macro
Estimated CSI: urban micro
Ref.: suburban macro
Ref.: urban macro
Ref.: urban micro
(27)
where the last equality uses
−1
H −1
.
, · · · , (FT FH
R−1
T)
Fd = blkdiag (F1 F1 )
Substituting (7) into the above equation yields
!−1
T
X
H
,
UH Vi,L Vi,L
U
MSE = tr Λ−1 + ρ
30
(28)
(29)
i=1
which indicates that the MSE is independent of both Ui
and Σi . Hence, we can safely set both of them to identity
H
matrices and obtain Fopt
= Vi,L
from (7) without loss of the
i
optimality, which completes the proof.
20
10
0
-20
theoretical analyses have shown that the channel estimation
performance can improve when the channel is more spatially
correlated, the BS deploys more RF chains or the MSs increase
training power, which has been verified by simulation results.
In addition, the RF combiners have been designed in the
case of multiple trainings to increase the DoF of the received
signal measurements at BS. The simulation results reveal that
multiple trainings can approach the performance of fullydigital estimation with single training and even outperform
it at low SNR regions when the channel is highly correlated.
Finally, this paper has proposed a practical spatial channel
covariance matrix estimation method in the hybrid-structured
massive MIMO system. The simulation results have shown the
effectiveness of the proposed method under various scenarios
in terms of the spectral efficiency with the hybrid precoding.
-10
0
SNR (dB)
10
20
Fig. 7: Spectral efficiency performance with the hybrid precoding scheme in practical SCM channels using both perfect and
estimated CSI. The Ref. result follows the channel estimation
algorithm proposed in [29]. (M = 64, L = 8, T = 6,
Nc = 1000, ρ = 20 dB and 3-bit phase quantization)
VI. C ONCLUSIONS
This paper has investigated and proposed a channel estimation framework for massive MIMO systems employing a hybrid transceiver structure with a limited number of RF chains
and a phase shifting network. The optimal RF combiners have
been designed in the case of single training by exploiting
the channel covariance matrix and the complete CSI is estimated in the hybrid structured massive MIMO system. The
B. Block Generalized Rayleigh Quotient
Lemma 2. (Block Generalized Rayleigh Quotient) [53] Given
a full column rank matrix V, the block generalized Rayleigh
quotient with respect to the pencil (A, B) is defined as
−1 T
GRQ(V) = tr VT BV
V AV ,
(30)
where V ∈ CN ×T . Suppose the generalized eigenvalues
and the corresponding eigenvectors of the pencil (A, B) are
denoted by λ1 ≥ · · · ≥ λN and (v1 , · · · , vN ), respectively,
where the eigenvalues are arranged in decreasing order,
without loss of generality. For a matrix containing a subset
of distinct generalized eigenvectors, the generalized Rayleigh
quotient evaluates to the sum of the associated eigenvalues:
GRQ([vi1 , · · · , viT ]) =
T
X
j=1
λij
(31)
12
and bounds can be placed on the Rayleigh quotient of an
N × T matrix as
λ1 + λ2 + · · · + λT ≥ GRQ(V) ≥ λN −T +1 + · · · + λN , (32)
where an equality in this bound indicates that
V is composed of the extreme eigenvectors as
colsp(V)
=
colsp([v1 , · · · , vT ]) or colsp(V)
=
colsp([vN −T +1 , · · · , vN ]). Here colsp(·) represents the
column space spanned by the specified matrix.
C. Proof of Corollary 1
We start with the definition of function φ : RL → R as
φ(x) =
L
X
ψ(xl ),
(33)
l=1
x2
l
where ψ(xl ) , xl +1/ρ
. Note that the MSE shown in (14)
can be denoted by MSE = M − φ(λ). It is straightforward
to show that ψ(xl ) is convex and hence φ(x) is Schur-convex
according to [54]. Suppose channel g1 ∈ CM is more spatially
correlated than g2 ∈ CM which can be mathematically
expressed as λ1 ≻ λ2 , where λ1 = [λ1,1 , · · · , λM,1 ]T and
λ2 = [λ1,2 , · · · , λM,2 ]T denote the eigenvalues sorted in
descending order of the covariance matrices of channel g1
and g2 , respectively. Applying the majorization theory on the
Schur-convex function φ(x), there exists
φ(λ1 ) ≥ φ(λ2 ).
(34)
Hence, MSE(g1 ) ≤ MSE(g2 ), which proves Corollary 1.
R EFERENCES
[1] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K.
Soong, and J. C. Zhang, “What will 5G be?” IEEE J. Sel. Areas
Commun., vol. 32, no. 6, pp. 1065–1082, Jun. 2014.
[2] W. Chin, Z. Fan, and R. Haines, “Emerging technologies and research
challenges for 5G wireless networks,” IEEE Wireless Commun., vol. 21,
no. 2, pp. 106–112, Apr. 2014.
[3] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9,
no. 11, pp. 3590–3600, Nov. 2010.
[4] J. Hoydis, S. Ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
of cellular networks: How many antennas do we need?” IEEE J. Sel.
Areas Commun., vol. 31, no. 2, pp. 160–171, Feb. 2013.
[5] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive
MIMO for next generation wireless systems,” IEEE Commun. Mag.,
vol. 30, no. 1, pp. 186–195, Feb. 2014.
[6] S. Jin, X. Wang, Z. Li, and K.-K. Wong, “Zero-forcing beamforming in
massive MIMO systems with time-shifted pilots,” in Proc. IEEE ICC,
Jun. 2014, pp. 4801–4806.
[7] Y. Dai and X. Dong, “Power allocation for multi-pair massive MIMO
two-way AF relaying with linear processing,” IEEE Trans. Wireless
Commun., vol. 15, no. 9, pp. 5932–5946, Sep. 2016.
[8] L. Pan, Y. Dai, W. Xu, and X. Dong, “A novel block-shifted pilot design
for multipair massive MIMO relaying,” in Proc. EUSIPCO, Aug. 2016,
pp. 848–852.
[9] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta,
O. Edfors, and F. Tufvesson, “Scaling up MIMO: Opportunities and
challenges with very large arrays,” IEEE Signal Process. Mag., vol. 30,
no. 1, pp. 40–60, Jan. 2013.
[10] E. Björnson, L. Sanguinetti, J. Hoydis, and M. Debbah, “Optimal design
of energy-efficient multi-user MIMO systems: Is massive MIMO the
answer?” IEEE Trans. Wireless Commun., vol. 14, no. 6, pp. 3059 –
3075, Jun. 2015.
[11] L. Pan, Y. Dai, W. Xu, and X. Dong, “Multipair massive MIMO relaying
with pilot-data transmission overlay,” IEEE Trans. Wireless Commun.,
vol. PP, no. 99, pp. 1–1, Mar. 2017.
[12] R. W. Heath Jr., N. González-Prelcic, S. Rangan, W. Roh, and A. M.
Sayeed, “An overview of signal processing techniques for millimeter
wave MIMO systems,” IEEE J. Sel. Topics Signal Process, vol. 10,
no. 3, pp. 436–453, Apr. 2016.
[13] X. Zhang, A. F. Molisch, and S.-Y. Kung, “Variable-phase-shift-based
RF-baseband codesign for MIMO antenna selection,” IEEE Trans. Signal
Process., 2005.
[14] O. E. Ayach, R. W. Heath, Jr., S. Abu-Surra, S. Rajagopal, and Z. Pi,
“The capacity optimality of beam steering in large millimeter wave
MIMO systems,” in Proc. IEEE SPAWC, Jun. 2012, pp. 100–104.
[15] A. F. Molisch, V. V. Ratnam, S. Han, Z. Li, S. L. H. Nguyen,
L. Li, and K. Haneda, “Hybrid beamforming for massive MIMO
A survey,” arXiv preprint arXiv:1609.05078, Apr. 2016. [Online].
Available: https://arxiv.org/pdf/1609.05078.pdf
[16] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, Jr.,
“Spatially sparse precoding in millimeter wave MIMO systems,” IEEE
Trans. Wireless Commun., vol. 13, no. 3, pp. 1499–1513, Mar. 2014.
[17] L. Liang, W. Xu, and X. Dong, “Low-complexity hybrid precoding
in massive multiuser MIMO systems,” IEEE Wireless Commun. Lett.,
vol. 3, no. 6, pp. 653–656, Dec. 2014.
[18] A. Alkhateeb, G. Leus, and R. W. Heath Jr., “Limited feedback hybrid
precoding for multi-user millimeter wave systems,” IEEE Trans. Wireless
Commun., vol. 14, no. 11, pp. 6481–6494, Nov. 2015.
[19] A. Alkhateeb and R. W. Heath Jr., “Frequency selective hybrid precoding
for limited feedback millimeter wave systems,” IEEE Trans. Commun.,
vol. 64, no. 5, pp. 1801–1818, May 2016.
[20] J. Wang, Z. Lan, C.-W. Pyo, T. Baykas, C.-S. Sum, M. A. Rahman,
J. Gao, R. Funada, F. Kojima, H. Harada, and S. Kato, “Beam codebook
based beamforming protocol for multi-Gbps millimeter-wave WPAN
systems,” IEEE J. Sel. Areas Commun., vol. 27, no. 8, pp. 1390–1399,
Oct. 2009.
[21] L. Chen, Y. Yang, X. Chen, and W. Wang, “Multi-stage beamforming
codebook for 60GHz WPAN,” in Proc. ICST CHINACOM, Aug. 2011,
pp. 1–5.
[22] S. Hur, T. Kim, D. J. Love, J. V. Krogmeier, T. A. Thomas, and
A. Ghosh, “Millimeter wave beamforming for wireless backhaul and
access in small cell networks,” IEEE Trans. Commun., vol. 61, no. 10,
pp. 4391–4403, Oct. 2013.
[23] Y. M. Tsang, A. S. Y. Poon, and S. Addepalli, “Coding the beams:
improving beamforming training in mmWave communication system,”
in Proc. IEEE GLOBECOMM 2011, Kathmandu, Nepal, Dec. 2011, pp.
1–6.
[24] J. Choi, D. J. Love, and P. Bidigare, “Downlink training techniques for
FDD massive MIMO systems: open-loop and closed-loop training with
memory,” IEEE J. Sel. Topics Signal Process, vol. 8, no. 5, pp. 802–814,
Oct. 2014.
[25] M. Biguesh and A. B. Gershman, “Training-based MIMO channel
estimation: A study of estimator tradeoffs and optimal training signals,”
IEEE Trans. Signal Process., vol. 54, no. 3, pp. 884 – 893, Mar. 2006.
[26] C.-K. Wen, S. Jin, K.-K. Wong, J.-C. Chen, and P. Ting, “Channel estimation for massive MIMO using Gaussian-mixture Bayesian learning,”
IEEE Trans. Wireless Commun., vol. 14, no. 3, pp. 1356–1368, Mar.
2015.
[27] D. Fan, Z. Zhong, G. Wang, and F. Gao, “Channel estimation for wireless
cellular systems with massive linear receiving antennas,” in Proc. 10th
Int. Conf. Commun. and Netw. in China (ChinaCom), Aug. 2015, pp.
95–99.
[28] H. Xie, F. Gao, S. Zhang, and S. Jin, “UL/DL channel estimation for
TDD/FDD massive MIMO systems using DFT and angle reciprocity,”
in Proc. VTC Spring, May 2016, pp. 1–5.
[29] A. Alkhateeb, O. E. Ayach, G. Leus, and R. W. Heath, Jr., “Channel
estimation and hybrid precoding for millimeter wave cellular systems,”
IEEE J. Sel. Topics Signal Process, vol. 8, no. 5, pp. 831–846, Oct.
2014.
[30] Z. Gao, C. Hu, L. Dai, and Z. Wang, “Channel estimation for millimeterwave massive MIMO with hybrid precoding over frequency-selective
fading channels,” IEEE Commun. Lett., vol. 20, no. 6, pp. 1259–1262,
Jun. 2016.
[31] R. Méndez-Rial, C. Rusut, A. Alkhateeb, N. González-Prelcic, and R. W.
Heath Jr., “Channel estimation and hybrid combining for mmWave:
Phase shifters or switches?” in Proc. ITA, Feb. 2015, pp. 90–97.
[32] B. Hassibi and B. M. Hochwald, “How much training is needed in
multiple-antenna wireless links?” IEEE Trans. Inf. Theory, vol. 49, no. 4,
pp. 951–963, Apr. 2003.
[33] J. H. Kotecha and A. M. Sayeed, “Transmit signal design for optimal
estimation of correlated MIMO channels,” IEEE Trans. Signal Process.,
vol. 52, no. 2, pp. 546–557, Feb. 2004.
13
[34] E. Björnson and B. Ottersten, “A framework for training-based estimation in arbitrarily correlated Rician MIMO channels with Rician
disturbance,” IEEE Trans. Signal Process., vol. 58, no. 3, Mar. 2010.
[35] A. Alkhateebt, O. E. Ayach, G. Leus, and R. W. Heath, Jr., “Singlesided adaptive estimation of multi-path millimeter wave channels,” in
Proc. IEEE SPAWC, Jun. 2014, pp. 125–129.
[36] A. Forenza, D. J. Love, and R. W. Heath Jr., “Simplified spatial
correlation models for clustered MIMO channels with different array
configuration,” IEEE Trans. Veh. Technol., vol. 56, no. 4, pp. 1924–
1934, Jul. 2007.
[37] T. E. Bogale, L. B. Le, A. Haghighat, and L. Vandendorpe, “On the
Number of RF Chains and Phase Shifters, and Scheduling Design With
Hybrid AnalogDigital Beamforming,” IEEE Trans. Wireless Commun.,
vol. 15, no. 5, pp. 3311–3326, May 2016.
[38] T. L. Marzetta, “How much training is required for multiuser MIMO?”
in Proc. ACSSC, Oct. 2006, pp. 359–363.
[39] A. Alkhateeb, G. Leus, and R. W. Heath Jr., “Compressed sensing
based multi-user millimeter wave systems: How many measurements
are needed?” in Proc. IEEE ICASSP, Apr. 2015, pp. 1–5.
[40] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory, 1st ed. Prentice Hall PTR, Apr. 1993, vol. 1.
[41] N. J. Higham, Accuracy and Stability of Numerical Algorithms, 2nd ed.
SIAM, 2002.
[42] A. Liu and V. K. N. Lau, “Two-stage subspace constrained precoding
in massive MIMO cellular systems,” IEEE Trans. Wirelss Commun.,
vol. 14, no. 6, pp. 3271–3279, Jun. 2015.
[43] B. D. Carlson, “Covariance matrix estimation errors and diagonal
loading in adaptive arrays,” IEEE Trans. Aerosp. Electron. Syst, vol. 24,
no. 4, pp. 397–401, Jul. 1988.
[44] Y.-C. Liang and F. P. S. Chin, “Downlink channel covariance matrix
(DCCM) estimation and its applications in wireless DS-CDMA systems,” IEEE J. Sel. Areas Commun., vol. 19, no. 2, pp. 222–232, Feb.
2001.
[45] W. L. Melvin and G. A. Showman, “An approach to knowledge-aided
covariance estimation,” IEEE Trans. Aerosp. Electron. Syst, vol. 42,
no. 3, pp. 1021–1042, Jul. 2006.
[46] R. Méndez-Rial, N. González-Prelcic, and R. W. Heath Jr., “Augmented
covariance estimation with a cyclic approach in DOA,” in Proc. IEEE
ICASSP, Apr. 2015, pp. 2784–2788.
[47] Y. Chi, Y. C. Eldar, and R. Calderbank, “PETRELS: Parallel subspace
estimation and tracking by recursive least square from partial observations,” IEEE Trans. Signal Process., 2013.
[48] R. Méndez-Rial, N. González-Prelcic, and R. W. Heath Jr., “Adaptive
hybrid precoding and combining in mmWave multiuser MIMO systems
based on compressed covariance estimation,” in Proc. IEEE CAMSAP,
Dec. 2015, pp. 213–216.
[49] S. Park and R. W. Heath Jr., “Spatial channel covariance estimation for
mmWave hybrid MIMO architecture,” in Proc. IEEE Asilomar 2016,
2016, pp. 1–5.
[50] S. Haghighatshoar and G. Caire, “Massive MIMO channel subspace estimation from low-dimensional projections,” IEEE Trans. Signal Process.,
vol. 65, no. 2, pp. 303–318, Jan. 2017.
[51] I. Viering, H. Hofstetter, and W. Utschick, “Spatial long-term variations
in urban, rural and indoor environments,” in Proc. European Cooperation in the Field of Scienific and Technical Research (COST), Sep. 2002,
pp. 1–9.
[52] J. Salo, G. D. Galdo, J. Salmi, P. Kyosti, M. Milojevic, D. Laselva,
and C. Schneider, “MATLAB implementation of the 3GPP Spatial
Channel Model (3GPP TR 25.996),” Jan. 2005. [Online]. Available:
http://www.tkk.fi/Units/Radio/scm/
[53] C. G. Baker, “Riemannian manifold trust-region methods with applications to eigenproblems,” Ph.D. dissertation, Florida State University,
May 2008.
[54] A. W. Marshal, I. Olkin, and B. C. Arnold, Inequalities: Theory of
Majorization and Its Applications, 2nd ed. Springer, 2010.
| 7cs.IT
|
1
Low-complexity implementation of convex
optimization-based phase retrieval
Sercan Ö. Arık, and Joseph M. Kahn, Fellow, IEEE
Abstract— Phase retrieval has important applications in
optical imaging, communications and sensing. Lifting the
dimensionality of the problem allows phase retrieval to be
approximated as a convex optimization problem in a higherdimensional space. Convex optimization-based phase retrieval
has been shown to yield high accuracy, yet its low-complexity
implementation has not been explored. In this paper, we study
three fundamental approaches for its low-complexity
implementation: the projected gradient method, the Nesterov
accelerated gradient method, and the alternating direction
method of multipliers (ADMM). We derive the corresponding
estimation algorithms and evaluate their complexities. We
compare their performance in the application area of directdetection mode-division multiplexing. We demonstrate that they
yield small estimation penalties (less than 0.2 dB for transmitter
processing and less than 0.6 dB for receiver equalization) while
yielding low computational cost, as their implementation
complexities all scale quadratically in the number of unknown
parameters. Among the three methods, ADMM achieves
convergence after the fewest iterations and the fewest
computational operations.
Index Terms—Phase retrieval, convex optimization, alternative
direction method of multipliers, optical communications, modedivision multiplexing.
P
I. INTRODUCTION
HASE retrieval corresponds to the extraction of phase
information about signals solely from their measured
intensities, possibly with the aid of various transformations.
Phase retrieval problems are commonly encountered in
applications where phase measurements are impossible or
impractical. One such field is optics, due to the fundamental
difficulty of measuring phase at carrier frequencies of
hundreds of THz [1]. The applications of phase retrieval in
optics include imaging (crystallographic imaging [2-4],
speckle imaging [5], coherent diffractive imaging [6,7],
astronomical imaging [8,9]), sensing (wavefront sensing [11]),
and
communications
(direct-detection
mode-division
multiplexing, MDM [10]). There are also important
applications of phase retrieval outside of optics, such as multiinput multi-output (MIMO) wireless communications [12],
This manuscript was submitted in September 2017. This work was
supported in part by Huawei Technologies Co., Ltd.
S. Ö. Arık is with Baidu Silicon Valley Artificial Intelligence Lab,
Sunnyvale, CA 94089 (e-mail: [email protected]).
J. M. Kahn is with the E. L. Ginzton Laboratory, Department of Electrical
Engineering, Stanford University, Stanford, CA 94305 USA (e-mail:
[email protected]).
speech recognition [13] and compressed sensing [14-16].
The traditional setting of phase retrieval is in the recovery
of signals from measured intensities in the space/spatial
frequency or time/temporal frequency domains. This setting is
encountered in many applications, because many systems
(e.g., optical systems such as graded-index media, lenses, or
simply free-space propagation) transform signals from the
spatial domain into the spatial frequency domain, and many
detection systems measure only intensities. The approaches
typically used for phase retrieval in this traditional problem
setting are of the form of sequential gradient descent. For
example, the Gerchberg-Saxton algorithm is based on
iteratively transforming a candidate solution back and forth
between the two domains (e.g., space and spatial frequency),
imposing constraints in one domain before transforming to the
other [17,18]. Despite the good performance of such heuristic
approaches, there is no algorithm for phase retrieval that
guarantees achievement of a globally optimal solution at a low
(non-deterministic polynomial-time hard) computational
complexity. A fundamental challenge is the non-convexity of
the intensity constraint after an orthogonal transformation
[19].
Furthermore, in many applications, the problem setting for
phase retrieval is more general than that addressed by
traditional algorithms. An important generalization is
encountered in recovery of unknown signals from the
measured intensities of their inner products with known
signals. (When the known signals are complex exponentials,
we obtain the special case described in the previous
paragraph.) The inner product operation occurs in linear
systems when measurements are performed in the appropriate
domain. For example, for a linear MIMO optical channel at a
given frequency, the known signals may correspond to
(baseband envelopes of) known training sequence vectors, and
the signals to be recovered may correspond to rows of the
channel transfer matrix [10]. In such settings, generalization of
low-complexity heuristic approaches based on sequential
gradient descent is challenging, as there might not be an
orthogonal transformation between the known signals (e.g.,
Fourier
transformation)
with
a
low-complexity
implementation (e.g., fast Fourier transform algorithm). The
concept of sequential algorithms may be generalized in the
form of alternating minimization, based on alternation
between the missing phase information and the candidate
solutions [20], but such methods do not reliably achieve good
performance [2,10].
Addressing phase retrieval problems in a general setting,
2
Candes et al. recently proposed an approach of lifting the
optimization problem to a high-dimensional space, where a
convex optimization problem can efficiently approximate the
original phase retrieval problem [21]. Convex optimizationbased phase retrieval has been demonstrated to achieve high
performance in imaging [21-23] and in communications [10].
Implementation complexity is critical in many applications,
yet the low-complexity implementation of convex
optimization-based phase retrieval has not yet been explored.
For example, in short-reach optical communications [10],
transceiver power consumption is an important concern, and
can determine the practical feasibility of any proposed
technology. Specialized low-complexity algorithms are
required for such applications. Methods employed in generalpurpose optimization software, such as interior-point methods,
are typically not suitable for such applications.
In this paper, we study three fundamental methods for lowcomplexity implementation of convex optimization-based
phase retrieval: the projected gradient method in Section III,
the Nesterov accelerated gradient method in Section IV, and
the alternating direction method of multipliers (ADMM) in
Section V. For each method, we describe adaptive algorithms
and present the associated computational complexities. In
comparing the performance of these methods, we focus on an
application area in optical communications: direct-detection
MDM, whose architecture is described in detail in Section VI.
We demonstrate that all these methods yield excellent
performance at low implementation complexity in either
transmitter- or receiver-based MIMO processing.
We focus on the problem of extracting unknown signal
vectors from the magnitudes of their inner products with
known signal vectors. We assume that D-dimensional known
complex vectors x (n ) are used for estimation of Ddimensional unknown complex vectors h l . With additive
(n ) , the available measurements are
h l H x (n )
2
+ n l (n ) = d l (n ),
N
min
rank(Cl ) + λ
∑ x( )
n H
Cl x (n ) − d l
(n )
2
,
n =1
(1)
where 1 ≤ n ≤ N and 1 ≤ l ≤ L . We might have L = D , e.g.,
in direct-detection MDM for a D × D MIMO channel, or we
might have L >> D , e.g., in diffractive imaging when
multiple illuminations are used.
In the absence of noise, the formulation (1) can be
expressed equivalently as L different rank-minimization
problems. Defining D × D complex Hermitian matrix
parameters C l , we have
min rank(Cl )
. (2)
(n )
s.t. Cl ≥ 0 and x(n )H Cl x(n ) = dl , for 1 ≤ n ≤ N
Optimal solutions to (2) are obtained for the matrix parameters
ˆ = h H h . To consider the effect of noise, the equality
C
l
l
l
constraints in (2) can be expressed as an additive penalty term
in a maximum-likelihood formulation of signal retrieval with
additive noise . In this paper, we focus on spatially white zero-
(3)
s.t. Cl ≥ 0
where λ is a tradeoff constant and ≥ denotes that the matrix
is positive semidefinite. Generalization of (3) to other noise
types is possible by considering the maximum-likelihood
formulation for the corresponding noise statistics, which might
or might not be convex in x (n ) .
As rank-minimization problems are non-deterministic
polynomial-time hard, it is desired to apply a convex
relaxation that can approximate the optimization problem (3)
efficiently, while yielding a global optimal solution. An
efficient convex relaxation is based on substituting the trace
function [21] in place of the rank function, which yields
N
min
g (Cl ) = tr (Cl ) + λ
∑ x( )
n H
Cl x ( n ) − d l
(n )
n=1
2
.
(4)
s.t. Cl ≥ 0
After computing the optimal matrix parameter Ĉ l , its first
ˆ = ∑D χ
ˆ ˆ ˆ H
ˆ 1,1 u
ˆ 1,1 , where C
χ
l
k =1 i ,k u i ,k u i ,k
principal component
, yields the estimate for h l . The overall dimensionality of the
problem is squared in exchange for the benefit of convex
approximation.
We will next focus on three techniques that aim to
approximately solve the increased-dimensionality problem
efficiently with low computational complexity. In estimating
() ()
II. CONVEX OPTIMIZATION-BASED PHASE RETRIEVAL
noise n l
mean real-valued Gaussian noise, for which the rankminimization problem can be expressed as
*
computational complexity, we assume ℜ , ℑ
and ( )
operations have negligible cost. We quantify the
computational complexity of other operations, which are
generally performed on complex numbers, in terms of the
number of floating point operations (flops) required. We
assume optimal ordering of algebraic operations to minimize
complexity. For example, multiplication of a matrix, vector
and scalar can be done with lower complexity by first
multiplying the matrix and the vector and then the scalar, or
by first multiplying the vector and the scalar and then the
matrix, as opposed to first multiplying the scalar and the
matrix and then the vector. We also assume that any computed
result can be reused at multiple steps without being
recomputed, by utilizing a working memory that can store
intermediate results. The structure of the objective function in
(4) forms the basis for the reduced-complexity algorithms.
III. PROJECTED GRADIENT METHOD
Projected gradient methods represent an efficient first-order
approach for constrained convex optimization problems. The
principle is based on iteratively searching for optimal values
of the variables along a direction determined by gradient
values, while imposing the constraints at each step.
Starting with an initial guess Cl 0 , the projected gradient
method applies the following update equation recursively at
iteration step k + 1 :
(5)
Cl k +1 ← P Cl k − t k ∇g Cl k .
[]
[
]
( [ ] [ ] ( [ ]))
3
( ) denotes the gradient operator corresponding to
objective function and P( ) denotes the projection
In (5), ∇g
the
operator that is required to satisfy the constraints.
We will next derive the various operators and analyze their
computational complexities. The step size can be updated by a
search algorithm. A common approach is a back-tracking line
search algorithm:
while ( g C l k < g Cl k − t k ∇g Cl k ):
( [ ])
( [ ] [ ] ( [ ]))
do t [k ] = t [k ] ⋅ τ
(6)
where 0 < τ < 1 . The computational complexity of backtracking line search is dominated by the complexity of the
(
)
objective function computation, which is O D 2 LN flops per
iteration step.
Computation of ∇g Cl , the gradient with respect to C l ,
requires computing the following derivatives for 1 ≤ i ≤ D and
i +1 ≤ k ≤ D :
2
∂g (C)
⎛N
⎞
= 1 + 2λ ⎜ ∑ x i(n ) ℜ x (n )H C l x (n ) ⎟
∂ℜ(C lii )
⎝ n =1
⎠
,
(7)
N
2
⎛
(n ) (n ) ⎞
− 2λ ⎜ ∑ ℜ d l
xi
⎟
⎝ n =1
⎠
N
∂g (C)
⎛
⎞
= 2λℜ⎜ ∑ x i(n ) x (kn )* x (n )H C l x (n ) ⎟
∂ℜ(C lik )
⎝ n =1
⎠
,
(8)
⎛ N (n ) (n ) (n )* ⎞
− 2λℜ⎜ ∑ d l x i x k ⎟
⎝ n =1
⎠
∂g (C)
⎛ N (n ) (n )* (n )H
⎞
= 2λℑ⎜ ∑ xi x k x Cl x(n ) ⎟
∂ℑ(Clik )
⎝ n =1
⎠
.
(9)
N
⎛
(n ) (n ) (n )* ⎞
− 2λℑ⎜ ∑ dl xi x k ⎟
⎝ n =1
⎠
2
(
n)
Computation of the x
and x (n )* x (n ) terms require
( )
(
)
(
)
( )
(
)
i
i
k
( )
2
O ND flops in total. For all expressions in (7)-(9), the
x (n )H Cl x (n ) terms are computed once with a complexity of
2
( )
O D2 L flops. Given
(n )* (n )
x i(n ) , x i x k , and x (n )H Cl x (n )
terms,
of
computation
N
(n ) (n )* (n )H C x (n )
∑ xi x k x
l
n =1
N
(n ) 2 ℜ x (n )H C x (n ) require
l
∑ xi
n =1
(
)
(
)
or
2 N − 1 flops for each i, k
(
)
and l, yielding a total complexity of O D2 LN flops. To
construct all the derivatives, performing all the additions and
subtractions in (7)-(9) requires O D2 L flops. The total
( )
(
)
complexity of computing all the derivatives is O D2 LN flops.
Imposing the constraint is based on the projection operator
onto the positive-semidefinite cone:
⎛D
⎞ D
P⎜⎜ ∑ λiuiui H ⎟⎟ = ∑ max(0, λi )uiui H .
(10)
⎝ i =1
⎠ i =1
Computation of the projection operator requires eigenvalue
decomposition. Hardware-efficient eigenvalue decomposition
techniques depend on the size of the matrix. For D = 3,
techniques based on the analytical expressions of the
eigenvalues and eigenvectors are fastest [1]. For D > 25,
divide-and-conquer is the fastest algorithm [2]. For 3 < D <
25, (the range of interest for practical applications, e.g., in
MDM systems), tridiagonal QR iteration is the fastest
( )
algorithm, with a complexity of O D3 flops [24]. After
replacing any negative eigenvalues by 0, the reconstruction of
( )
the projection requires O D3
flops. Overall, applying
( )
projection L times requires in total O D 3 L flops per
iteration,
For the projected gradient method, the overall
computational cost is dominated by the computation of
gradient
and
projection
operations,
and
is
flops per iteration. The total
max O D 2 LN , O D 3 L
computational complexity is also proportional to the number
of iterations. Hence, minimization of the total number of
iterations is also crucial.
((
) ( ))
IV. NESTEROV ACCELERATED GRADIENT METHOD
Optimal first-order methods are derived from classical
projected gradient methods, with the goal of reducing the
number of iterations required to achieve convergence.
One modification of the projected gradient method,
pioneered by Nesterov [25], is the accelerated gradient method
[26], which has update equations:
Cl [k + 1] ← P(K l [k ] − t[k + 1]∇g (K l [k ])),
γ[k + 1] ←
1
⎛1
⎜ + 1+ 1
⎜2
4 γ[k ]2
⎝
⎞
⎟
⎟
⎠
,
(11)
(12)
γ[k + 1](1 − γ[k ])
(C l [k + 1]− C l [k ]) . (13)
γ[k ]
For determining the step size in (11), as above, the backtracking line search algorithm (6) can be used. The
fundamental distinction of the Nesterov accelerated gradient
method is the adjustment of the step size using the γ k
K l [k + 1] ← C l [k + 1] +
[]
[]
parameter. The evolution of γ k as a function of the iteration
number is shown in Fig. 1.
The Nesterov accelerated gradient method has a slightly
higher complexity than the projected gradient method. The
computational cost for the update step (12) is O 1 flops and
( )
()
2
for (13) is O D L flops. The major computational cost is step
((
) ( ))
(11), which requires max O D 2 LN , O D 3 L flops per
iteration, similar to the projected gradient algorithm. Again,
the total computational complexity is proportional to the
number of iterations. The modification for acceleration
reduces the number of steps required for convergence, despite
a slight increase in the computational cost per iteration step.
V. ALTERNATING DIRECTION METHOD OF MULTIPLIERS
A general approach to address constrained convex
optimization problems with complicated objective functions is
4
1
N
1
( N ) = λ x (n ) x (n )* x (n ) x (n )* .
Tikab
∑ i k b a
(22)
0.8
0.6
0.6
0.4
0.4
0.2
0.2
γ[k]
0.8
0
0
5
10
15
20
(γ[k + 1](1 − γ[k]))/γ[k]
n =1
0
25
Iteration number
Fig. 1
γ[k ]
vs. iteration number for the Nesterov accelerated projected
The minimization problem (17) reduces to a set of D 2 linear
equations. Consider the matrix representation of the set of
linear equations (20)
⎡ C l11 [k + 1] ⎤
⎢ C [k + 1] ⎥
l12
⎢
⎥
~
⎥ = b l(N ) [k ],
(23)
X (N ) + ρI ⎢
⎢
⎥
⎢C lD(D −1) [k + 1]⎥
⎢ C lDD[k +1] ⎥
⎣
⎦
(
)
( )
where b l N [k ] is a vector of size D 2 × 1 such that
(N ) − (E [k ] / 2 + ρD [k ]) − δ / 2 ,
b l(n(()i −1)D + k ) [k ] = U lik
lik
lik
i ,k
N
T
~ (N )
2
2
X = λ ∑ X(n )X(n ) is a matrix of size D × D , and
n =1
gradient method.
to decompose them into smaller problems, each of which is
then easier to handle. The ADMM technique combines the
benefits of the dual decomposition and augmented Lagrangian
methods for constrained optimization,
We start by observing that the optimization problem can be
equivalently expressed in equality constraint form:
min
g (Cl ) + I ≥0 (Dl )
s.t. Cl = Dl
,
(14)
( )
2
)
(n )* (n )
X((in−1
)D + k = xi x k is a vector of size D × 1 . For the case of
(~
)
invertible X(N ) + ρI , the solution of (23) is given by
⎡ Cl11[k + 1] ⎤
⎢ C [k + 1] ⎥
⎢ l12
⎥ ~
−1
⎢
⎥ = X(N ) + ρI bl(N )[k ] = F (N )bl(N )[k ]. (24)
⎢
⎥
⎢ClD(D −1)[k + 1]⎥
⎢ ClDD[k +1] ⎥
⎣
⎦
(
)
where I ≥0 D l is the indicator function such that
⎧ 0, D l ≥ 0
.
(15)
I ≥0 (D l ) = ⎨
⎩∞, otherwise
The corresponding augmented Lagrangian is in the form
L(C l , D l , E l ) = g (C l ) + I ≥0 (D l )
.
(16)
2
+ El , Dl − Cl + ρ Dl − Cl F
ADMM is based on replacing minimization of (16) by
alternating minimization, where each step optimizes the
matrix parameters C l , D l and E l separately:
Inverting an ordinary matrix of size D 2 × D 2 requires O D 6
flops using naïve algorithms such as Gauss-Jordan
C l [k + 1] ← argmin L(C l , D l [k ], E l [k ]) ,
(17)
D l [k + 1] ← arg min L(C l [k + 1], D l , E l [k ]) ,
(18)
N −1
T
T
⎛
⎞
= ⎜ λX (N ) X (N ) + λ ∑ X (n ) X (n ) + ρI ⎟
n =1
⎝
⎠
E l [k + 1] ← E l [k ]+ ρ (Dl [k + 1]− Cl [k + 1]).
(19)
( )
(
~
F (N ) = X (N ) + ρI
(
Cl
Dl
We now analyze the steps (17)-(19).
For step (17), we find the analytical solution by setting the
first partial derivatives with respect to Clik to zero (since the
second derivatives are always positive), which yield
D
(
)
(N ) =
ρClik [k + 1] + ∑ Clab [k + 1]Tikab
a ,b =1
( N ) − ⎛ Elik [k ] + ρD [k ]⎞ − δ i , k
Ulik
⎜
⎟
lik
2
⎝ 2
⎠
where δ i ,k is a Kronecker delta function, and
N
( N ) = λ d (n ) x (n ) x (n )* ,
U lik
∑ l
i
k
n =1
,
(20)
)
elimination, and requires O D4.746 flops with optimized
Coppersmith-Winograd-like construction methods [27].
Iterative computation of the matrix inverse term F ( N ) is also
possible using the Sherman–Morrison–Woodbury lemma,
exploiting its structure:
= F (N −1) −
−1
)
−1
.
(25)
F (N −1) X (N ) X ( N )T F (N −1)
1 / λ + X (N )T F (N −1) X (N )
~
Note that we have the initial conditions X (0) = 0 and
F (0) = I / ρ . Computation of X (N ) requires O D 2 flops.
4
Given X (N ) and F (N −1) , the update of F ( N ) requires
( )
( )
OD
flops. As the update is repeated for all n , the total complexity
of computing F ( N ) becomes O D 4 N flops. Computation of
(
)
(
)
( N ) terms requires an additional
all U lik
O D2 NL flops. Given
( )
( N ), computation of all b (N ) [k ] vectors requires
U lik
O D2 L
l
flops.
(21)
Overall,
((
2
computation
) (
4
))
of
Cl [k + 1]
max O D LN , O D N flops at each iteration.
requires
5
For step (18), an analytical solution can be obtained by
equating the gradient to zero. We have the gradient
2 ⎞
⎛ 1
∇ D[k ]⎜
El [k ] + (Dl [k ] − Cl [k + 1]) ⎟ =
⎜ 2ρ
⎟
F⎠
⎝
.
(26)
⎛ 1
⎞
2⎜⎜
El [k ] + (Dl [k ] − Cl [k + 1])⎟⎟
2
ρ
⎝
⎠
1
If Cl [k + 1] −
El [k ] ≥ 0 , then the optimal solution is
2ρ
1
D l [k + 1] = C l [k + 1] −
E l [k ] . For the general case, the
2ρ
⎛
⎞
1
optimal D l k + 1 is the projection P⎜⎜ C l [k + 1] −
E l [k ]⎟⎟ .
2
ρ
⎝
⎠
The overall update equation becomes
⎛
⎞
1
(27)
D l [k + 1] ← P⎜⎜ C l [k + 1] −
E l [k ]⎟⎟ .
2ρ
⎝
⎠
[
modulus) of each modal field amplitude is available at the
receiver. This has important consequences for MIMO signal
(a)
(b)
]
Computation of the
( )
1
C l [k + 1] −
E l [k ] terms requires
2ρ
(c)
O D2 L flops in total. The projection operator is applied L
times and dominates the computational complexity, requiring
O D 3 L flops per iteration.
( )
( )
) ( ))
Lastly, step (19) requires O D2 L flops at each step. The
overall
computational
complexity
per
step
is
2
4
3
flops
per
iteration,
and
is
max O D LN , O D N , O D L
again proportional to the total number of iterations.
The computational complexities of the three methods are
summarized in Table I.
((
) (
TABLE I
COMPUTATIONAL COMPLEXITIES OF PHASE RETRIEVAL METHODS
Computational complexity per iteration
( (
( (
) (
) (
))
))
Projected gradient
max O D 2 LN , O D 3 L
Nesterov
max O D 2 LN , O D 3 L
ADMM
max O D 2 LN , O D 4 N , O D 3 L
((
) (
) (
))
VI. APPLICATION EXAMPLE: DIRECT-DETECTION MODEDIVISION MULTIPLEXING
In this section, we study application of the phase retrieval
techniques described above in direct-detection MDM [10].
MDM aims to increase the data throughput per fiber by
employing a fiber supporting D propagating modes. At the
transmitter, D independent data signals are multiplexed onto
these modes, ideally providing a D-fold increase in
throughput. During transmission, the data signals are
intermixed by random coupling between modes, which
corresponds to a D × D channel transfer matrix that has offdiagonal entries. We assume that group velocity dispersion
and modal dispersion are negligible, such that memoryless
signal processing and channel estimation are sufficient.
Using direct detection simplifies receiver implementation,
making the technique potentially suitable for short-reach
communications. However, only the intensity (squared
Fig. 2 Direct detection MDM systems using: (a) digital MIMO precoding at
the transmitter, (b) optical MIMO precoding at the transmitter, and (c)
optical MIMO equalization at the receiver. Mux: Multiplexer, mod.:
modulator, det.: detector, I: in-phase, Q: quadrature.
processing, channel transfer matrix estimation and information
decoding. First, compensating for modal crosstalk requires
complex-valued D × D MIMO signal processing. As shown
in Fig. 2, direct-detection MDM can be implemented using a
digital MIMO precoder or optical MIMO precoder at the
transmitter, or using an optical MIMO equalizer at the receiver
(additional details are given in [10]). Second, adjustment of
the MIMO precoder or equalizer requires estimation of the
D × D channel transfer matrix using phase retrieval. Third,
information must be demodulated from the optical carrier
using noncoherent or differentially coherent detection.
Here, we are concerned with the second consequence,
namely, how to perform channel estimation using phase
retrieval. Estimation is facilitated using a known training
sequence, which is composed of N D-dimensional signals. For
a given D , the required size of the training signals can be
found empirically. If the number of training vectors N is not
large enough, high estimation error is obtained due to
insufficient measurement diversity. On the other hand,
increasing N beyond a certain threshold does not improve
estimation accuracy significantly, and it is desired to set the
value of N around this threshold. Note that for optical
precoding or equalization methods, the training vectors can be
transmitted at a low symbol rate, and high-speed digital signal
processing can be avoided. For all three methods (digital or
optical MIMO precoding or optical equalization), channel
estimation corresponds to the phase retrieval problems
described in Section 2.
6
For a MIMO precoder, the channel transfer matrix needs to
be estimated up to a constant diagonal unitary matrix [10]. We
have L = D , since for each transmitted sequence, D different
the receiver is thermal noise-limited, so the received modal
intensities are corrupted by additive, spatially white real
Gaussian noise. Realizations of the channel transfer matrix are
Group 1
(a)
2×2
MIMO
MMF
Spatial
Demux
Group 2
SNR = 12 dB
-1
a2
10
SNR = 14 dB
b1
4×4
MIMO
Crosstalk-free
Interior point-based
Projected gradient
Nesterov accelerated gradient
ADMM
SNR = 8 dB
a1
b2
-2
10
b3
(b)
Polarization
Beam
Spatial Splitter
Mode 1
4 × 4 MIMO Processing using Mach-Zehnder Array
Average BER
b4
-3
10
-4
10
-5
10
Spatial
Mode 2
-6
10
-7
10
Polarization
Converter
Fig. 3 (a) Demultiplexing using receiver MIMO processing in mode group
subspaces, (b) implementation of optical MIMO processing using MachZehnder array.
inner products are generated by the D rows of the channel
transfer matrix, i.e., h l in (1) corresponds to the l th row of the
channel transfer matrix.
For a MIMO equalizer, the channel transfer matrix needs
to be estimated up to a complex exponential, hence an
additional optimization procedure is used to estimate the
unknown phases of each row of the channel transfer matrix,
yielding L = D + 1 [10].
Considering L = D or L = D + 1 with Table I, for a MIMO
precoder or a MIMO equalizer, we estimate the computational
complexity per iteration to be max O D 3 N , O D 4 for the
projected gradient and Nesterov accelerated gradient methods,
and O D 4 N for the ADMM. In other words, they are of the
order of the dimensionality of the high-dimensional space,
which is the square of the number of unknowns.
In short-reach systems, the end-to-end link may be
optimized to have negligible modal crosstalk between mode
groups [28]. In this case, MIMO processing can be
implemented within each mode group separately. Fig. 3
illustrates processing in mode group subspaces in a system
using receiver MIMO equalization implemented by an array of
Mach-Zehnder interferometers [29]. Processing in mode group
subspaces can reduce the total hardware complexity of MIMO
processing, as the required number of Mach-Zehnder
interferometers scales with the square of the number of modes
processed. In addition, mode group subspace processing can
significantly reduce the computational complexity of channel
estimation, as the estimation algorithms can be implemented
separately for each mode group (e.g., of sizes D = 2 and
D = 4 for the system shown in Fig. 3).
((
(
) ( ))
)
VII. NUMERICAL RESULTS
In the simulations presented here, we consider a directdetection MDM system employing on-off keying. We assume
0
10
20
30
40
Number of iterations
50
60
70
Fig. 4 Average BER vs. number of iterations for a direct-detection MDM
system using on-off keying, transmitter MIMO precoding, D = 6 modes and
N = 300 training signals. Horizontal lines show the performance achieved in
the two reference cases: for the ideal crosstalk-free channel and when an
interior-point technique [28] is used to implement convex estimation-based
phase retrieval. ADMM: alternating direction method of multipliers.
chosen from an ensemble of fully random complex-valued
D × D unitary matrices to model worst-case modal crosstalk.
We assume D = 6 modes and, accordingly, we choose
N = 300 training signals to obtain low estimation error. We
assume that these N training signals are generated randomly,
with each vector element an independent, identically
distributed circularly symmetric Gaussian random variable.
The convex optimization formulation tradeoff constant is
chosen as λ = 10, as in [10], as this choice yielded the lowest
estimation error in the original convex optimization problem
formulation (4). For projected gradient and Nesterov
accelerated gradient methods, the backtracking line search
parameter is chosen as τ = 0.3, which is at an optimal point
when the tradeoff between the number of backtracking line
search iterations and the direction of search accuracy is
considered (when the value of τ is close to 1, convergence of
the backtracking line search would be slow and when the
value of τ value is too small, the estimation error would be
high when moving in the direction of the step size).
The system electrical signal-to-noise ratio (SNR) is defined
as
⎡
SNR = ⎢
⎣
where
D
∑i =1 y i
(n )
2
(D ⋅ σ )⎤⎥
2
,
(28)
⎦
denotes an average over training signals. Since (28) is
proportional to the square of the optical power per mode, a 1-dB
change of the optical power per mode corresponds to a 2-dB
change of 10 log10 (SNR ). A range of SNR values is
considered that corresponds to practical short-reach optical
links. Link performance is quantified by the average bit-error
ratio (BER). Note that short-reach communication systems
typically use forward error-correction codes with low
complexity, necessitating BER targets around 10-3 to 10-4.
7
We compare the performance of the three low-complexity
convex optimization-based phase retrieval techniques to the
performance achieved in two reference cases. The first
Average BER
10-2
Crosstalk-free
Interior point-based
Projected gradient
Nesterov
ADMM
10-3
10-4
10-5
7
8
9
10
11
12
13
14
SNR (dB)
Fig. 5 Average BER vs. SNR for a direct-detection MDM system using onoff keying, transmitter MIMO precoding, D = 6 modes and N = 300 training
signals. ADMM: alternating direction method of multipliers.
Average BER
10-2
Crosstalk-free
Interior point-based
Projected gradient
Nesterov
ADMM
10-3
10-4
10-5
7
8
9
10
11
SNR (dB)
12
13
14
Fig. 6 Average BER vs. SNR for a direct-detection MDM system using onoff keying, receiver MIMO equalization, D = 6 modes and N = 300 training
signals. ADMM: alternating direction method of multipliers.
reference case is the ideal crosstalk-free channel, where
additive noise is the only impairment. The second reference
case includes modal crosstalk and uses CVX [30], a wellknown interior-point-based package, for convex estimationbased phase retrieval to determine the MIMO precoder or
equalizer.
Fig. 4 shows the average BER vs. number of iterations
using transmitter MIMO precoding. The convergence speed is
rather insensitive to the SNR value. The projected gradient
method converges after ~50 iterations, the Nesterov
accelerated gradient method converges after ~10 iterations,
and the ADMM converges after just a few iterations. In the
remainder of the simulations, we fix the total number of
iterations to 70 for the projected gradient method, 30 for the
Nesterov accelerated gradient method and 5 for the ADMM,
to ensure convergence beyond the respective knees of the
curves shown in Fig. 4.
Table II provides estimates of the number of operations per
second required for channel estimation based on the
assumptions described in Sections III-V. We note that the
mapping of an algorithm to hardware can be optimized given
tradeoffs between the costs of computation, memory, and
communication on a chip. Hence, the numerical values in
Table II should be considered as rough order-of-magnitude
estimates for straightforward hardware mapping on a singlecore processor without parallelization of the computation.
The operation counts in Table II are computed assuming
channel estimation at intervals of 1, 10 and 100 ms. Based on
studies of channel dynamics [31], we estimate that short-reach
multimode links with D~6 modes will change on time scales
of ms to tens of ms [31]. Even if a direct-detection MDM
system must update channel estimates at 1 ms intervals, the
operation count for ADMM is only ~5 Gflop/s. This is about
two orders of magnitude smaller than the operation count for a
long-haul coherent MDM receiver [28].
TABLE II
OPERATIONS PER SECOND FOR CONVEX OPTIMIZATION-BASED PHASE
RETRIEVAL IN A DIRECT-DETECTION MDM SYSTEM USING TRANSMITTER
MIMO PRECODING, ASSUMING L = D = 6, N = 300.
Update interval
Projected gradient
Nesterov
ADMM
Operations per second required for channel
estimation for update intervals indicated
1 ms
10 ms
100 ms
~50 Glops/s
~5 Gflops/s
~500 Mflops/s
~20 Gflops/s
~2 Gflops/s
~200 Mflops/s
~5 Gflops/s
~0.5 Gflops/s
~50 Mflops/s
Fig. 5 shows the average BER vs. SNR using transmitter
MIMO precoding. All three techniques exhibit performance
very close to CVX, incurring SNR penalties (with respect to
crosstalk-free case) less than 0.3 dB for target average BERs
of 10-3 to 10-4.
Fig. 6 shows the average BER vs. SNR using receiver
MIMO
equalization.
Receiver
MIMO
equalization
necessitates estimation procedures to determine the transmitter
MIMO precoder, as well as additional estimation to
disambiguate common phase factors [10]. Consequently, there
are more opportunities for estimation errors than in transmitter
precoding. The projected gradient and Nesterov accelerated
gradient methods yield similar performance, incurring SNR
penalties (with respect to crosstalk-free case) less than 0.6 dB
for target average BERs of 10-3 to 10-4. For ADMM, the SNR
penalty is slightly higher, but still less than 0.8 dB for the
same target average BER range.
VIII. CONCLUSIONS
We have studied three low-complexity approaches for
implementing convex optimization-based phase retrieval: the
projected gradient method, the Nesterov accelerated gradient
method and ADMM. The projected gradient method is a firstorder technique, which aims to impose the constraint at each
iteration step. The Nesterov accelerated gradient method
achieves an acceleration of the projected gradient method by
using an adaptive step size parameter. The ADMM is based on
redefining the convex optimization objective for uncoupled
solutions of the separable objective functions using alternating
minimization. We have demonstrated application of the three
techniques to direct-detection MDM. All three methods
accurately approximate the solution to the original phase
retrieval problem, incurring small estimation penalties. The
computational complexities of all three methods scale with the
square of the number of unknowns per iteration. Among the
8
three, ADMM yields convergence after the smallest number of
iterations.
ACKNOWLEDGEMENTS
Discussions with Prof. Emmanuel J. Candes are gratefully
acknowledged.
REFERENCES
[1] A. Walther, “The question of phase retrieval in optics,” J. Mod. Opt. vol.
10, no. 1, pp. 41-49, 1963.
[2] E. J. Candes, X. Li, and M. Soltanolkotabi, “Phase retrieval via Wirtinger
flow: Theory and algorithms,” IEEE Trans. Inf. Theory, vol. 61, no. 4, pp.
1985-2007, Feb. 2015.
[3] F. Fogel, I. Waldspurger, A. d’Aspremont, “Phase retrieval for imaging
problems,” Math. Program. Comp. vol. 8, no. 3, pp. 311-355, Mar. 2016.
[4] R. P. Millane, “Phase retrieval in crystallography and optics,” J. Opt. Soc.
Am. A vol. 7, no. 3, pp. 394-411, 1990.
[5] R. Horisaki, R. Egami, J. Tanida, “Single-shot phase imaging with
randomized light (SPIRaL),” Opt. Expr. vol. 24, no. 4, pp. 3765-3773, Feb.
2016.
[6] M. H. Seaberg, A. d'Aspremont, and J. J. Turner, “Coherent diffractive
imaging using randomly coded masks,” App. Phys. Lett. vol. 107, no. 23,
231103, Nov. 2015.
[7] J. A. Rodriguez, R. Xu, C. C. Chen, Y. Zou, and J. Miao, “Oversampling
smoothness: an effective algorithm for phase retrieval of noisy diffraction
intensities,” J. App. Crystallogr. vol. 46, pp. 312-318, 2013.
[8] J. C. Dainty, and J. R. Fienup, “Phase retrieval and image reconstruction
for astronomy,” in Image Recovery: Theory and Applications, H. Stark,
(Academic Press Inc., 1987), pp. 231-275.
[9] F. Roddier, Adaptive Optics in Astronomy (Cambridge University Press,
1999).
[10] S. O. Arik and J. M. Kahn, “Direct-detection mode-division multiplexing
in modal basis using phase retrieval,” Opt. Lett. vol. 41, no. 18, pp. 42654268, Sep. 2016.
[11] R. A. Gonsalves and R. Chidlaw, “Wavefront sensing by phase retrieval,”
Proc. SPIE 0207, Applications of Digital Image Processing III no. 32, Dec.
1979.
[12] G. Wunder, H. Boche, T. Strohmer, and P. Jung, “Sparse signal
processing concepts for efficient 5G system design,” IEEE Access vol. 3, pp.
195-208 Feb. 2015.
[13] R. Balan, P. Casazza, and D. Edidin, “On signal reconstruction without
phase,” App. and Comp. Harmonic Anal. vol. 20, no. 3, pp. 345-356, May
2006.
[14] M. B. McCoy and J. A. Tropp, “The achievable performance of convex
demixing,” [Online]. Available: https://arxiv.org/abs/1309.7478.
[15] C. Hegde and R. B. Baraniuk, “Signal recovery on incoherent
manifolds,” IEEE Trans. on Info. Theory vol. 58, no. 12, pp/ 7204-7214, Jul.
2012.
[16] P. Schniter and S. Rangan, “Compressive phase retrieval via generalized
approximate message passing,” IEEE Trans. Signal Process. vol. 63, no. 4,
pp. 1043-1055, Feb. 2015.
[17] R. W. Gerchberg and W. O. Saxton, “A practical algorithm for the
determination of the phase from image and diffraction plane pictures,” Optik
vol. 35, pp. 237-246, 1972.
[18] J. R. Fienup, “Phase retrieval algorithms: a comparison,” App. Opt. vol.
21, no. 15, pp. 2758-2769, Aug. 1982.
[19] V. Elser, “Phase retrieval by iterated projections,” J. Opt. Soc. Am. A vol.
20, no. 1, pp. 40-55, 2003.
[20] P. Netrapalli, P. Jain, and S. Sanghavi, “Phase retrieval using alternating
minimization,” IEEE Trans. Signal Process. vol. 63, no. 18, pp. 4814-4826,
Jun. 2015.
[21] E. J. Candes, T. Strohmer, and V. Voroninski, “Phaselift: Exact and
stable signal recovery from magnitude measurements via convex
programming,” Commun. Pure and App. Math. vol. 66, no. 8, pp. 1241-1274,
Aug. 2013.
[22] E. J. Candes, Y. C. Eldar, T. Strohmer, and V. Voroninski, “Phase
retrieval via matrix completion,” SIAM J. Imaging Sci. vol. 57, no. 2, pp. 225251, 2015.
[23] L. H. Yeh, J. Dong, J. Zhong, L. Tian, M. Chen, G. Tang, M.
Soltanolkotabi, and L. Waller, “Experimental robustness of Fourier
ptychography phase retrieval algorithms,” Opt. Expr. vol. 23, no. 26, pp.
33214-33240, Dec. 2015.
[24] J. W. Demmel, Applied Numerical Linear Algebra (Siam, 1997).
[25] Y. Nesterov, Introductory Lectures on Convex Optimization, (Kluwer
Academic Publishers, 2004).
[26] A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding
algorithm for linear inverse problems,” SIAM J. Imaging Sci. vol. 2, no. 1, pp.
183–202, 2009.
[27] W. Williams, “Multiplying matrices faster than Coppersmith-Winograd,”
in Proceedings of ACM Symposium on Theory of Computing 2012, pp. 887898.
[28] S. O. Arık, K. P. Ho, J. M. Kahn, “Group delay management and multiinput multi-output signal processing in mode-division multiplexing systems,"
J. Lightw. Techn. vol. 34, no. 11, pp. 2867-2880, Jun. 2016.
[29] D. A. B. Miller, “Self-aligning universal beam coupler,” Opt. Expr. vol
21, no. 5, pp. 6360-6370, Mar. 2013.
[30] M. Grant and S. Boyd, "CVX: Matlab Software for Disciplined Convex
Programming," [Online]. Available: http://cvxr.com/cvx.
[31] K. Choutagunta and J. M. Kahn, “Dynamic Channel Modeling for ModeDivision Multiplexing", J. Lightw. Techn. vol. 35, no. 12, pp. 2451-2463, Jun.
2017.
| 7cs.IT
|
Compositional Construction of Finite State Abstractions
for Stochastic Control Systems
arXiv:1709.09546v1 [cs.SY] 27 Sep 2017
Kaushik Mallik, Sadegh Esmaeil Zadeh Soudjani, Anne-Kathrin Schmuck and Rupak Majumdar
Abstract— Controller synthesis techniques for continuous systems with respect to temporal logic specifications typically use a
finite-state symbolic abstraction of the system. Constructing this
abstraction for the entire system is computationally expensive,
and does not exploit natural decompositions of many systems
into interacting components. We have recently introduced a
new relation, called (approximate) disturbance bisimulation for
compositional symbolic abstraction to help scale controller
synthesis for temporal logic to larger systems.
In this paper, we extend the results to stochastic control
systems modeled by stochastic differential equations. Given
any stochastic control system satisfying a stochastic version of
the incremental input-to-state stability property and a positive
error bound, we show how to construct a finite-state transition
system (if there exists one) which is disturbance bisimilar to the
given stochastic control system. Given a network of stochastic
control systems, we give conditions on the simultaneous existence of disturbance bisimilar abstractions to every component
allowing for compositional abstraction of the network system.
I. I NTRODUCTION
In abstraction-based controller synthesis, a finite-state
symbolic model of a continuous system is used to synthesize
a symbolic controller for a logical specification, and the
controller is then refined to a controller for the original
system. This technique has recently gained a lot of attention due to two main advantages. First, it allows for fully
automated controller synthesis for systems with continuous
dynamics while handling complex specifications (given e.g.
as ω-regular languages) in addition to stability. Second,
it naturally accounts for the complex interplay between
discrete and continuous components within a control loop.
The soundness of the abstraction-based synthesis technique
relies on notions of behavioral closeness of the original
system and its abstraction, which is formalized using system
equivalence relations (see e.g. [11], [8] and the references
therein). Recently, abstraction-based controller synthesis has
been extended to stochastic control systems [12], [2]. In the
stochastic setting, behavioral closeness of the original system
and its abstraction is formalized using the n-th moment of
the trajectories.
Despite its nice theoretical properties and applicability
to many different system classes, abstraction based controller synthesis does not scale very well because both
the abstraction step and the controller synthesis step are
exponential in the dimension of the continuous state space.
This issue motivated us to propose disturbance bisimulation
[5], an equivalence relation that exploits the intrinsic compositionality of systems. Given a network of metric systems
All authors are with MPI-SWS, Kaiserslautern, Germany {kmallik,
sadegh, akschmuck, rupak}@mpi-sws.org
and their equivalently interconnected disturbance bisimilar
abstractions, the main result of [5] shows that the overall network system is also disturbance bisimilar to the network of
the abstractions. This result has an interesting consequence:
given a construction of a disturbance bisimilar abstraction
for a given system class, we can compositionally abstract
a network of systems from this class whenever all local
abstractions are guaranteed to exist simultaneously. In [5] we
exploited this fact for large networks of incremental input-tostate stable deterministic control systems and demonstrated
the effectiveness of our approach in a case study.
In this paper, we extend these results to the class of
stochastic control systems which satisfy a stochastic version
of the incremental input-to-state stability condition. Our main
contribution in this paper is to show how a stochastic control
system, which may be connected to other components within
a network, allows for the algorithmic computation of (1) a
metric system capturing its time-sampled dynamics and (2) a
metric system capturing its abstract symbolic dynamics, such
that the two constructed systems are disturbance bisimilar.
This construction allows us to use the results from [5] to provide a compositional abstraction-based controller synthesis
technique for a given network of continuous and stochastic
dynamical systems.
Our results relate to recent results in [13] on compositional
abstraction for stochastic systems. The main difference between our work and [13] is the type of abstraction: while
we work with finite state symbolic abstractions, their abstractions are infinite state. There have been some efforts for
improving scalability of abstraction techniques for stochastic
systems in a different setting, where abstract models are
Markov chains and the goal is to match distributions on states
up to a fixed horizon. In [9], the state space discretization
is done adaptively and, in [10], the abstract state space of a
monolithic system is represented compositionally.
The paper is organized as follows. After introducing preliminaries on stochastic control systems in Sec. II, we define
metric systems in Sec. III, and present how the two particular
metric systems discussed above can be obtained from a
stochastic control system. Given these two metric systems,
we give sufficient conditions for them to be disturbance
bisimilar in Sec. IV, after recalling the notion of disturbance
bisimulation from [5]. In Sec. V, we invoke results from [5]
to extend our result from a single stochastic control system
to a network of such systems. All proofs can be found in the
appendix.
II. S TOCHASTSIC C ONTROL S YSTEMS
Most part of this section is adapted from [12] to systems
with stochastic disturbance inputs.
A. Notation
We use the symbols N, R, R>0 , R≥0 and Z to denote the
set of natural, real, positive real, non-negative real numbers,
and integers, respectively. The symbols In , 0n , and 0n×m
denote the identity matrix, the zero vector, and the zero
matrix in Rn×n , Rn , and Rn×m , respectively. Given a vector
x ∈ Rn , we denote by xi the i-th element of x and by k x k
the infinity norm of x.
A continuous function γ : R≥0 → R≥0 is said to belong
to class K if it is strictly increasing and γ(0) = 0; γ is said
to belong to class K∞ if γ ∈ K and limr→∞ γ(r) = ∞.
A continuous function β : R≥0 × R≥0 → R≥0 is said to
belong to class KL if, for each fixed s, the function
β(·, s) : R≥0 → R≥0 belongs to class K∞ and, for each
fixed r, the map β(r, ·) : R≥0 → R≥0 is decreasing and
lims→∞ β(r, s) = 0. Let f : R≥0 → Rk be a measurable function. We define the (essential) supremum k f k
of f as k f k := ess(sup){k f (t) k | t ≥ 0}. A function
f is bounded if k f k < ∞. Given a square matrix M ,
we denote by T r(M ) the trace of M , and by λmin (M )
and λmax (M ) the minimum and maximum eigenvalue of
M respectively. Given a matrix
M = {mij } ∈ Rn×m , we
Pm
denote by k M k := max1≤i≤n j=1 |mij | the infinity norm
p
of M , and by kM kF := Tr (M M T ) the Frobenius norm
of M . We denote by Diag(a1 , . . . , an ) the diagonal matrix
with diagonal entries a1 , . . . , an ∈ R. If a1 , . . . , an are
matrices, then Diag(a1 , . . . , an ) is a block diagonal matrix
of appropriate dimension.
B. Stochastic Control System
We fix the probability space for the whole paper as
(Ω, F , P), where Ω is a sample space, F is a sigma algebra
over Ω representing the set of events, and P is a probability
measure. Let (Ω, F , P) admits a filtration F = (Fs )s≥0
which is complete and right continuous [4, p. 89]. Let
(Bs )s≥0 be a r-dimensional F-Brownian motion.
Definition 2.1: A stochastic control system is a tuple Σ =
(X, U, U, W, W, f, σ), where X = Rn is the state space,
U ⊆ Rm is the input set that is assumed to be compact, U is
a subset of set of all measurable, locally essentially bounded
functions of time from R≥0 to U , W ⊆ Rp is the disturbance
input space that is assumed to be compact, W is a set of
stochastic processes with elements ν : Ω × R≥0 → W , f :
X × U × W → X is a continuous function of its arguments
representing the drift of Σ, σ : X → Rn×r is a function
representing the diffusion of Σ.
A stochastic process ξ : Ω × R≥0 → X is called a solution
process of Σ if there exists µ ∈ U and ν ∈ W satisfying the
following stochastic differential equation:
process of Σ from the initial condition ξaµν (0) = a P-a.s.,
and under effect of input signal µ ∈ U and disturbance signal
ν ∈ W. Note that given any time instant t, ξaµν (t) represents
a random variable from Ω to X measurable in Ft .
We make the following two assumptions on stochastic control systems to ensure a unique global continuous solution.
Assumption 1 (Lipschitz condition): There exist constants
Lf , Lσ ∈ R≥0 such that the following inequalities hold
kf (x, u, w)−f (x′ , u′ , w′ )k ≤ Lf (kx−x′ k+ku−u′k+kw−
w′ k), and kσ(x) − σ(x′ )k ≤ Lσ kx − x′ k, for all x, x′ ∈ X,
u, u′ ∈ U and w, w′ ∈ W .
Assumption 2 (Linear growth): There exists a positive
constant K such that for all x ∈ X, u, ∈ U and w ∈ W ,
max(kf (x, u, w)k2 , kσ(x)k2 ) ≤ K 1 + kxk2 . (II.2)
Assump. 1 on Lipschitz continuity gives uniqueness and
Assump. 2 on linear growth gives global existence ([7,
Thm. 5.2.1]). The latter will also be used in Sec. V-B
(cf. Prop. 5.3) to provide an upper bound on the second
moment of the solution process.
In this paper the disturbances in the set W are allowed to
be stochastic. This is necessary because, as will be described
later, in our setting the disturbances play the role of trajectories of other stochastic control systems after interconnection.
For the results of this paper to hold, we require the process
(ξ, ν) : Ω × R≥0 → X × W to be an Itô process, i.e., (ξ, ν)
has to be the solution of a possibly time-inhomogeneous Itô
diffusion.
C. δ-ISS-Mq
We now generalize the notion of incremental input-tostate stability in the q-the moment (δ-ISS-Mq ) for stochastic
control systems from [12] by considering disturbances. In
the absence of noise, these notions correspond to δ-ISS for
deterministic systems [1].
Definition 2.2: A stochastic control system Σ =
(X, U, U, W, W, f, σ) is stochastically incrementally inputto-state stable in the q-th moment (δ-ISS-Mq ), if there exists
a KL function β and K∞ functions ρu and ρd such that
for any t ∈ R≥0 , any µ, µ′ ∈ U, any ν, ν ′ ∈ W, and any
Rn -valued random variables a and a′ that are measurable in
F0 , the following condition is satisfied:
(II.1)
q
q
E [kξaµν (t) − ξa′ µ′ ν ′ (t)k ] ≤ β E ka − a′ k , t
q
+ ρu (k µ − µ′ k) + ρd E k ν − ν ′ k . (II.3)
The δ-ISS-Mq property can be characterized in terms of
the existence of stochastic incremental Lyapnuov functions.
Definition 2.3: [12, Def. 3.2] Define the diagonal set ∆ as
∆ = {(x, x) | x ∈ Rn }. Consider a stochastic control system
Σ = (X, U, U, W, W, f, σ) and a continuous function V :
X × X → R≥0 which is smooth on {Rn × Rn }\∆. Function
V is called a δ-ISS-Mq Lyapunov function for Σ if there exist
K∞ functions α, α, σu , σd , and a constant κ ∈ R>0 such
that
P-almost surely (P-a.s.). For succinctness of representation,
we use the notation ξaµν to denote a stochastic solution
(i) α is a convex function, and α and σd are concave
functions;
d ξ(t) = f (ξ(t), µ(t), ν(t)) d t + σ(ξ(t)) d Bt ,
(ii) for any x, x′ ∈ X,
α(kx − x′ kq ) ≤ V (x, x′ ) ≤ α(kx − x′ kq );
(iii) for any x, x′ ∈ X, x 6= x′ , any u, u′ ∈ U , and any
w,w′ ∈ W ,
′
′
f (x, u, w)
Lu,u ,w,w V (x, x′ ) := [∂x V ∂x′ V ]
f (x′ , u′ , w′ )
∂x,x V ∂x,x′ V
1
σ(x) T
T
′
σ (x) σ (x )
+ Tr
σ(x′ )
∂x′ ,x V ∂x′ ,x′ V
2
′
′
≤ −κV (x, x ) + σu (k u − u k) + σd (k w − w′ kq ),
′
′
where Lu,u ,w,w is the infinitesimal generator ([7, Section
7.3]) associated to the stochastic control system (II.1), which
depends on two separate controls u, u′ ∈ U and two separate
disturbances w, w′ ∈ W . In this case we say that the
stochastic control system Σ admits a δ-ISS-Mq Lyapunov
function, witnessed by α, α, σu , σd , and κ ∈ R>0 .
Note that condition (i) is not required in the context of
deterministic control systems. Condition (ii) implies that the
growth rates of the functions α and α are linear, as a concave
function is supposed to dominate a convex one. These
conditions are not restrictive provided we are interested in
the dynamics of Σ on a compact subset D ⊂ Rn , which is
often the case in practice. It can be readily verified that the
δ-ISS-Mq Lyapunov function in Definition 2.3 is a stochastic
bisimulation function between Σ and itself, as defined in [3],
Def. 5.
The following theorem describes δ-ISS-Mq in terms of
the existence of δ-ISS-Mq Lyapunov functions. It generalizes
the corresponding theorem [12, Thm. 3.3] in the presence of
disturbances.
Theorem 2.4: A stochastic control system Σ is δ-ISS-Mq
if it admits a δ-ISS-Mq Lyapunov function.
In order to keep the notation simple, we present the
results only for second moment in the rest of paper with
the understanding that they can be generalized for other
moments.
The following lemma (compare [12, Lemma 3.4]) provides
a sufficient condition on a particular function V to be a δISS-Mq Lyapunov function.
Lemma 2.5: Let Σ = (X, U, U, W, W, f, σ) be a stochastic control system. Let P ∈ Rn×n be a symmetric positive
definite matrix. Consider the function V : Rn × Rn → R≥0
defined as:
V (x, x′ ) :=
1
(x − x′ )T P (x − x′ )
2
(II.4)
and satisfying
(x − x′ )T P (f (x, u, w) − f (x′ , u, w))+
(II.5)
2
1 √
P (σ(x) − σ(x′ ))
≤ −2e
κV (x, x′ ),
+
2
F
or, if f is differentiable, satisfying
(x − x′ )T P ∂x f (z, u, w)(x − x′ )+
1 √
+
P (σ(x) − σ(x′ ))
2
(II.6)
2
F
≤ −2e
κV (x, x′ ),
for all x, x′ , z ∈ X, for all u ∈ U , for all w ∈ W , and for
some constant κ
e ∈ R>0 . Then V is a δ-ISS-M2 Lyapunov
function for Σ.
D. Noisy and Noise Free Trajectories
In this section we provide an upper bound on the distance
between a stochastic state trajectory and its associated noisefree trajectory at any instant of time. This bound is a
generalization of the bound in [12, Lemma 3.10] to the case
when there is disturbance in the system. The bound will be
instrumental in proving closeness between the trajectories of
a stochastic control system and its abstraction in Sec. III.
Lemma 2.6: Consider a stochastic control system Σ =
(X, U, U, W, W, f, σ). Suppose there exists a δ-ISS-M2 Lyapunov function V of Σ s.t. its Hessian matrix in R2n×2n
satisfies 0 ≤ ∂x,x V (x, x′ ) ≤ P , for some positive semidefinite matrix P ∈ R2n×2n and for any x, x′ ∈ Rn . Define
ξ xµν as the solution of the ordinary differential equation
(ODE)
ξ˙ xµν (t) = f ξ xµν (t), µ(t), ν(t)
(II.7)
starting from the initial condition x. Then for any x in a
compact set D ⊂ Rn , any µ ∈ U and any ν ∈ W, we have
i
h
2
≤ h(σ, t),
E ξxµν (t) − ξ xµν (t)
where
1 √
h(σ, t) := α−1 k P k2 · n · min{n, r} · e−κt · L2σ ·
2
Z t
β supx∈D k x k2 , s + ρu (supu∈U k u k)
·
0
+ρd supw∈W k w k2 ds .
(II.8)
The non-negative valued function h tends to zero as t → 0,
t → ∞, or as Lσ → 0, where Lσ is the Lipschitz constant
introduced in Assump. 1.
Remark 2.7: Eqn. (II.8) gives a representation of the function h in terms of ρu and ρd . This representation of h can
be translated into a form using σu and σd instead, as shown
in [12, Lemma 3.10, Cor. 3.11].
III. F ROM S TOCHASTIC C ONTROL S YSTEMS
S YSTEMS
TO
M ETRIC
We now introduce (deterministic) metric systems and
interpret stochastic control systems and their abstractions as
metric systems. As in [12], [5], we consider metric systems
that are time sampled w.r.t. a globally fixed time sampling
parameter τ ∈ R>0 .
Definition 3.1: Given the probability space Ω and a time
sampling parameter τ ∈ R>0 , a stochastic metric system1
S = (X, U, Uτ , W, Wτ , δτ ) consists of a (possibly infinite)
set of states X, given by a set of random variables, and
equipped with a metric d : X×X → R≥0 , a set of piece-wise
1 Often, metric systems are defined with an additional output space and
an output map from states to the output space. We omit the output space
for notational simplicity; for us, the state and the output space coincide, and
the output map is the identity function.
constant inputs Uτ of duration τ taking values in U ⊆ Rm ,
i.e.,
Uτ = {µ : [0, τ ] → U | ∀t1 , t2 ∈ [0, τ ] . µ(t1 ) = µ(t2 )},
(III.1)
a set of disturbances Wτ taking values in W ⊆ Rp , i.e.,
Wτ ⊆ {ν : Ω × [0, τ ] → W },
(III.2)
and a transition function δτ : X × Uτ × Wτ → 2X . We write
µ,ν
x −−→ x′ if x′ ∈ δτ (x, µ, ν), and we denote the unique value
τ
of µ ∈ U over [0, τ ] by uµ ∈ U .
A deterministic metric system is a special type of a
stochastic metric system where the states are deterministic points (i.e. random variables with Dirac delta distributions), and disturbances are deterministic signals of the form
[0, τ ] → W .
If the metric system S is undisturbed, we define W =
{0}. In this case we occasionally represent S by the tuple
S = (X, U, Uτ , δτ ) and use δτ : X × Uτ → 2X with
the understanding that x′ ∈ δτ (x, µ, ν) holds for the zero
trajectory ν : R≥0 → {0} whenever x′ ∈ δτ (x, µ). By
slightly abusing notation we write x′ = δτ (x, µ, ν) as a short
form when the set δτ (x, µ, ν) = {x′ } is a singleton. If X, Uτ
and Wτ are finite (resp. countable), S is called finite (resp.
countable). We also assign to a transition x′ = δτ (x, µ, ν)
any continuous time evolution ξ : [0, τ ] → X s.t. ξ(0) = x
and ξ(τ ) = x′ .
In the following we introduce two approaches to capture
an abstracted version of the dynamics of a stochastic control
systems Σ by a metric system conforming to Def. 3.1. The
first approach results in a sampled time abstraction which
we denote by Pτ (Σ).
Definition 3.2: Given a stochastic control system Σ =
(X, U, U, W, W, f, σ), a time-sampling parameter τ ∈ R>0 ,
and a probability space (Ω, F , P), the discrete-time stochastic metric system induced by Σ is defined by Pτ (Σ) =
(Xr , U, Uτ , W, Wτ , δτ ) s.t. Xr is the set of all X-valued
random variables, Uτ and Wτ are defined over U and W ,
respectively, as in Eqn. (III.1)-(III.2) and δτ (x, µ, ν) = x′
if x and x′ are measurable in Ft and Ft+τ , respectively,
for some t ∈ R≥0 , and there exists a solution process ξ :
Ω × R≥0 → Rn of Σ satisfying ξ(0) = x and ξxµν (τ ) = x′
P-a.s. Since we allow any state to be initial, all states in X
need to be measurable on F0 . We equip Xr with the metric
1/2
d(x, x′ ) := E[k x − x′ k2 ]
.
Remark 3.3: Recall that the disturbances in the set W
are stochastic. Hence the above metric system must be
constructed by looking at the sampled version of (ξ, ν),
which is the solution process of the Itô diffusion associated
with (x, w).
The second approach additionally imposes a
quantization of the state, input and disturbance spaces
and results in a metric system denoted by Pτ ηω (Σ).
Before defining this system formally we introduce
notation for quantization. For any A ⊆ Rn and any
vector η with elements ηi > 0, we define [A]η :=
{(a1 , . . . , an ) ∈ A | ai = 2kηi , k ∈ Z, i = 1, . . . , n}.
For x ∈ Rn and vector λ with elements λi > 0, let
Bλ (x) = {x′ ∈ Rn | k xi − x′i k ≤ λi } denote the closed
rectangle centered at x. Note that for any λ ≥ η (elementwise), the collection of sets Bλ (q) with q ∈ [Rn ]η is a cover
of Rn , that is, Rn ⊆ ∪{Bλ (q) | q ∈ [Rn ]η }. We will use
this insight to discretize the state and the input space of Σ
using discretization parameters η and ω, respectively.
Also we need to define a vector-valued metric for comparing two disturbance vectors. Let A1 , . . . , Ak be a finite set
of metric spaces, where each AiQ
, i = 1, . . . , k has a metric
k
di : Ai × Ai → R≥0 . Let A = i=1 Ai . We construct the
k
metric e : A × A → R≥0 as an extension of the metrics di
on Ai : for any a = (a1 , . . . , ak ) and a′ = (a′1 , . . . , a′k ), we
define
e(a, a′ ) := (d1 (a1 , a′1 ), . . . , dk (ak , a′k )).
(III.3)
For the disturbance space W we allow the discretization
of W to be predefined. We make the following general
assumptions on the discretizaion of W which will be useful
when we deal with networks of stochastic control systems
in Sec. V.
Assumption 3: Let Σ = (X, U, U, W, W, f, σ) be a
stochastic control system. Then we assume that there exists
f ⊆ W , that exists a vector ε̃ ∈ Rp , and a
a countable set W
≥0
vector-valued metric e : W × W → Rp≥0 , s.t. for all w ∈ W
f for which
there exists a w̃ ∈ W
e(w, w̃) ≤ ε̃ and k w − w̃ k ≤ k e(w, w̃) k.
(III.4)
Using this assumption we formally define the abstract metric
system Pτ ηω (Σ) induced by Σ as follows.
Definition 3.4: Let Σ = (X, U, U, W, W, f, σ) be a
stochastic control system for which Assump. 3 holds. Given
three constants τ ∈ R>0 , η ∈ R>0 , and ω ∈ R>0 ,
the discrete-time discrete-space deterministic metric system
induced by Σ is defined by
f , Wτ ηω , δτ ηω )
Pτ ηω (Σ) = (Xτ ηω , [U ]ω , Uτ ηω , W
(III.5)
s.t. Xτ ηω = [X]η , Uτ ηω is defined over [U ]ω , as in (III.1),
f | ∀t, k ∈ [0, τ ] . ν(t) = ν(k)},
Wτ ηω := {ν : [0, τ ] → W
and
δτ ηω (x, µ, ν) = {x′ ∈ Xτ ηω | k ξ xµν (τ ) − x′ k ≤ η},
where ξ xµν (·) are the noise free trajectories of Σ defined via Eqn. (II.7). We equip Xτ ηω with the metric
d(x, x′ ) := k x − x′ k naturally inherited from X. We denote
f.
the unique value of ν ∈ Wτ ηω over [0, τ ] by wν ∈ W
Remark 3.5: Let us emphasize that even though Pτ (Σ) is
a stochastic metric system and Pτ ηω (Σ) is a deterministic
metric system, since we are interested in studying the closeness of their trajectories in the next section, it is important
that Xr and Xτ ηω are part of the same state space. We
interpret Xτ ηω to be contained in Xr , since a set of points
can be associated with a set of random variables with Dirac
delta distributions.
IV. D ISTURBANCE B ISIMULATION
This section contains the main contribution of the paper; after recalling the notion of disturbance bisimulation
from [5] we present sufficient conditions under which the
two metric systems Pτ ηω (Σ) and Pτ (Σ) associated with a
stochastic control system Σ are disturbance bisimilar. For
this analysis, we restrict our attention to δ-ISS-M2 stochastic
control systems with f (0n , 0m , 0p ) = 0n and σ(0n ) = 0n×r ,
whose evolution is restricted to a compact region D ⊂ Rn .
Definition 4.1: Let Si = (Xi , Ui , Uτ,i , Wi , Wτ,i , δτ,i ),
i = 1, 2, be two metric systems, with state-spaces X1 , X2 ⊆
X and disturbance sets W1 , W2 ⊆ W ⊆ Rp . Furthermore,
let X admit the metric d : X × X → R≥0 and W admit
the vector-valued metric e : W × W → Rp≥0 . A binary
relation R ⊆ X1 × X2 is a disturbance bisimulation with
parameters (ε, ε̃) where ε ∈ R≥0 and ε̃ ∈ Rp≥0 , iff for each
(x1 , x2 ) ∈ R:
(a) d(x1 , x2 ) ≤ ε;
(b) for every µ1 ∈ Uτ,1 there exists a µ2 ∈ Uτ,2 such that for
all ν2 ∈ Wτ,2 and ν1 ∈ Wτ,1 with e(ν1 (0), ν2 (0)) ≤ ε̃,
we have that (δτ,1 (x1 , µ1 , ν1 ), δτ,2 (x2 , µ2 , ν2 )) ∈ R; and
(c) for every µ2 ∈ Uτ,2 there exists a µ1 ∈ Uτ,1 such that for
all ν1 ∈ Wτ,1 and ν2 ∈ Wτ,2 with e(ν1 (0), ν2 (0)) ≤ ε̃,
we have that (δτ,1 (x1 , µ1 , ν1 ), δτ,2 (x2 , µ2 , ν2 )) ∈ R.
S1 and S2 are said to be disturbance bisimilar with parameters (ε, ε̃) if there is a disturbance bisimulation relation R
between S1 and S2 with parameters (ε, ε̃).
In order to prove the existence of a disturbance bisimulation between Pτ ηω (Σ) and Pτ (Σ) we require two additional
assumptions.
Assumption 4: Let Σ be a stochastic control system admitting a δ-ISS-M2 Lyapunov function V . There exists a
K∞ and concave function γ
b s.t. for any x, x′ , x′′ ∈ X,
|V (x, x′ ) − V (x, x′′ )| ≤ b
γ (k x′ − x′′ k).
(IV.1)
This assumption is not restrictive as we are interested in
the dynamics of Σ on a compact subset D ⊂ Rn .
Assumption 5: Let Σ be a stochastic control system with
the associated metric systems Pτ (Σ) and Pτ ηω (Σ) introduced in Sec. III. Then there exists a K∞ function ψ s.t.
for all disturbance pairs ν ∈ Wτ and ν̂ ∈ Wτ ηω with
d(ν(0), ν̂(0)) ≤ k ε̃ k, the following holds for all t ∈ [0, τ ]:
d(ν(t), ν̂(t)) = E[k ν̂(t) − ν(t) k2 ]
21
≤ ψ(t) + k ε̃ k.
(IV.2)
Given Assump. 4 and Assump. 5, we present our first main
result in the following theorem.
Theorem 4.2: Let Σ be a stochastic control system admitting a δ-ISS-M2 Lyapunov function V witnessed by κ, α,
α, σu , and σd , that satisfies Assump. 4 with K∞ function γ
b.
f ⊆ W s.t. (III.4) holds and let Pτ ηω (Σ) be
Fix τ > 0 and W
the countable deterministic metric system associated wih Σ
according to Def. 3.4 such that Assump. 5 holds. If
n
b−1 (1 − e−κτ )α(ε2 )
0 ≤ η ≤ min (α)−1 ◦ α(ε2 ), γ
o
1
1
1
− σu (ω) − σd (ψ(τ ) + k ε̃ k) − (h(σ, τ )) 2 ,
eκ
eκ
(IV.3)
where h(σ, τ ) is as in (II.8), then the relation
Rε ε̃ = (x̂, x) ∈ Xτ ηω × Xr | E[V (x̂, x)] ≤ α(ε2 )
(IV.4)
is a disturbance bisimulation relation (in the second moment)
with parameters (ε, ε̃) between Pτ ηω (Σ) and Pτ (Σ).
Remark 4.3: Given any fixed τ and ε̃, one can always find
sufficiently small η and ω s.t. (IV.3) and (IV.4) hold, as long
as ε is lower bounded according to
1
1
2
σ
(ψ(τ
)
+
k
ε̃
k)
+
γ
b
h(σ,
τ
)
eκ d
. (IV.5)
ε2 > α−1
(1 − e−κτ )
The lower bound on ε can be minimized by choosing an
optimal Lyapunov function V for a given system Σ (see
e.g. [12, Rem. 3.6]). Note that, when the system does not
experience any disturbance, (IV.5) reduces to [12, V.5].
V. C OMPOSITIONAL A BSTRACTION
Let us first summarize what we have presented so far. In
Sec. III we have introduced two different metric systems
Pτ (Σ) and Pτ ηω (Σ) associated with a given stochastic
control system Σ. Recall that Pτ (Σ) is an infinite state
system, whereas Pτ ηω (Σ) is a finite state system under the
assumption that the state space of Pτ ηω (Σ) is restricted to a
compact subset of Rn . Then we gave sufficient conditions for
these two abstractions to be disturbance bisimilar in Sec. IV.
In this section, we consider a network of stochastic control systems {Σi }i∈I , and the respective local abstractions
{Pτ (Σi )}i∈I and {Pτ ηi ωi (Σi )}i∈I of {Σi }i∈I , s.t. for all
i ∈ I, Pτ (Σi ) and Pτ ηi ωi (Σi ) are disturbance bisimilar
with parameters (εi , ε̃i ). Then we adapt our result from [5],
and prove that the isomorphic networks of {Pτ (Σi )}i∈I and
{Pτ ηi ωi (Σi )}i∈I , which are isomorphic to the network of
{Σi }i∈I as well, are again disturbance bisimilar.
A. Network of Stochastic Control Systems
We first formalize networks of stochastic control systems
and their abstractions by locally treating state trajectories of
neighboring systems as disturbances.
Let I be an index set (e.g., I = {1, . . . , N } for some
natural number N ) and let I ⊆ I × I be a binary irreflexive
connectivity relation on I. Furthermore, let I ′ ⊆ I be a
subset of systems with I ′ := (I ′ ×I ′ )∩I. For i ∈ I we define
NI (i) = {j | (j, i) ∈ I} and extend this notion to subsets of
systems I ′ ⊆ I as NI (I ′ ) = {j | ∃i ∈ I ′ .j ∈ NI\I ′ (i)}.
Intuitively, a set of systems can be imagined to be the set
of vertices {1, 2, . . . , |I|} of a directed graph G, and I to
be the corresponding adjacency relation. Given any vertex
i of G, the set of incoming (resp. outgoing) edges are the
inputs (resp. outputs) of a subsystem i, and NI (i) is the
set of neighboring vertices from which the incoming edges
originate.
Let Σi = (Xi , Ui , Ui , Wi , Wi , fi , σi ), for i ∈ I, be a
collection of stochastic control systems. We say that the
set of stochastic control systems {Σi }i∈I is compatible for
composition w.r.t. the interconnection
relation I, if for each
Q
i ∈ I, we have Wi = Qj∈NI (i) Xj . By slightly abusing
notation we write wi = j∈NI (i) {xj } for xj ∈ Xj and
wi ∈ Wi as a short form for the single element of the set
Q
j∈NI (i) {xj }. We extend this notation to all sets with a
single element.
Let I ′ ⊂ I be a subset of systems in the network. We
divide the set of disturbances Wi for any i ∈ I ′ into the
c
sets of coupling and external
Q disturbances, defined by Wi =
Q
e
j∈NI\I ′ (i) Xj , respectively.
j∈NI ′ (i) Xj and Wi =
If {Σi }i∈I is compatible, we define the composition of
any subset I ′ ⊆ I of systems as the stochastic control
system JΣi Ki∈I ′ = (X, U, U,
Q X, U and
Q W, W, f, σ) where
W are Q
defined as X =
i∈I ′ Ui , and
i∈I ′ Xi , U =
W = j∈NI (I ′ ) Xj . Furthermore, U and W are defined
as the sets of functions µ : R≥0 → U and ν : Ω ×
R≥0 → W such that the projection µi of µ on to Ui
(written µi = µ|Ui ) belongs to Ui , and the projection νie
e
drift is
to W
of ν on to Wie belongs
Q
Qi . The composed
Q
e
{w
{u
},
{x
},
then
defined
as
f
(
i Q i∈I ′
i
i }) =
i∈I ′
i∈I ′
Q
c
e
c
{f
(x
,
u
,
w
×
w
)},
where
w
=
i i
i
i
i
i
j∈NI ′ (i) {xj },
i∈I ′
Q
and the composed diffusion is defined as σ( i∈I ′ {xi }) =
Diag σ1 (x1 ), . . . , σ|I| (x|I| ) . The Brownian motion of the
T
overall system is defined as: dBt = dB1,t . . . dB|I|,t .
If I ′ = I, then Σ is undisturbed, modeled by W := {0}.
It is easy to see that JΣi Ki∈I ′ is again a stochastic control
system in the sense of Def. 2.1. Networks of discrete time
stochastic metric systems (Pτ (Σi )) and of abstract metric
systems (Pτ ηi ωi (Σi )) are defined analogously.
Remark 5.1: Note that we assume a nice structure of the
network: the diffusion functions and the Brownian motions
of the systems in a network are decoupled from the states of
the other systems. This is explicitly induced via the SDE
(II.1) as the diffusion terms σi (·) are only functions of
system’s state and not the disturbance. However since the
states of the systems are coupled through the drift functions,
the respective random variables are implicitly dependent.
B. Simultaneous Approximation
Given I and I ′ ⊆ I, consider a set of compatible stochastic
control systems {Σi }i∈I , the subset composition JΣi Ki∈I ′ =
(X, U, U, W, W, f, σ) and a global time-sampling parameter
τ . Then we can apply Def. 3.2 and Def. 3.4 to each Σi
to construct the corresponding metric systems Pτ (Σi ) and
Pτ ηi ωi (Σi ). To be able to do that, we need to equip Wi
|N (i)|
with a vector-valued metric ei : Wi × Wi → R≥0I
fi for all i ∈ I s.t. Ass. 3 holds. Intuitively,
and define W
ei (wi , wi′ ) is a vector with dimension |NI (i)|, where the j th
entry measures the mismatch of the respective state vector
fi as the product of state
of the j th neighbor of i. We define W
spaces of Pτ ηj ωj (Σj ), i.e., the abstraction of its neighbors,
Q
fi :=
W
(V.1)
j∈NI (i) [Xj ]ηj .
Lemma 5.2: Let Σi = (Xi , Ui , Ui , Wi , Wi , fi , σi ),i ∈ I,
be a set of compatible stochastic control systems and
the set of abstract metric systems Pτ ηi ωi (Σi ) =
fi , Wi,τ ηi ωi , δi,τ ηi ωi )
are
([Xi ]ηi , [U ]ωi , Ui,τ ηi ωi , W
constructed according to Def. 3.4, where Wi is equipped
fi as defined in (V.1). Select local
with metric (III.3) and W
quantization parameters {ηi }i∈I s.t. ηi ≤ εi . Then Ass. 3
holds for every i ∈ I with ε̃i defined as
Q
ε̃i := j∈NI (i) {εj }.
(V.2)
Given Lemma 5.2, it immediately follows that the sets
{Pτ (Σi )}i∈I ′ and {Pτ ηi ωi (Σi )}i∈I ′ of metric systems are
again compatible.
In order to guarantee the result of Thm. 4.2 for the network, we have additionally used Assump. 5 which essentially
bounds the effect of the disturbances on the state evolution.
Given the particular choice of disturbances in the network as
state trajectories of neighboring systems, we can explicitly
compute function ψ(t) in Assump. 5 using the following
proposition from [6, Thm. 4.3].
Proposition 5.3: Under Assump. 2 the solution process
ξaµν (·) satisfies the inequality
E kξaµν (t) − ξaµν (s)k22 ≤ C|t − s|, ∀s, t ∈ [0, τ ],
for any τ > 0, where k · k2 indicates
of a vector.
the 2-norm
ατ
2
(τ
+
1)e
with α :=
The constant
C
:=
2
1
+
Ekak
2
√
K + 2 K and K from Assump. 2.
The next lemma follows from Prop. 5.3.
Lemma 5.4: Given a set of stochastic control systems
{Σi }i∈I which is compatible for composition, let each
system Σi satisfy Assump. 2 with constant Ki in (II.2). Then
Assump. 5 holds for each Σi with K∞ function
i 12
h
P
ψi (t) := t(t + 1) j∈NI (i) βj eαj t ,
p
where
αj
:=
=
Kj + 2 Kj and βj
2 1 + supxj ∈Xj kxj k22 .
Lemmas 5.2-5.4 show that the assumptions of Sec. IV
on disturbance sets of Σi hold after composition. Then
the next theorem follows from Thm. 4.2 which establishes
simultaneous disturbance bisimilarity between abstractions
of components in a network. In this theorem, using the results
in Thm. 4.2, we give conditions on all local state, input, and
disturbance quantization parameters in a composed stochastic
control system JΣi Ki∈I which allow for a simultaneous
construction of local abstractions Pτ ηi ωi (Σi ) using Def. 3.4
such that they are disturbance bisimilar with parameters
(εi , ε̃i ) to their respective discrete-time stochastic metric
systems Pτ (Σi ).
Theorem 5.5: Let {Σi }i∈I be a set of compatible stochastic control systems, each admitting a δ-ISS-M2 Lyapunov
function Vi witnessed by κi , αi , αi , σu,i , and σd,i , and let
γ
bi be a K∞ function s.t. (IV.1) holds. Let {Pτ (Σi )}i∈I be
the set of discrete-time stochastic metric systems induced
by {Σi }i∈I and let {Pτ ηi ωi (Σi )}i∈I be the set of countable
fi as
deterministic metric systems induced by {Σi }i∈I and W
in (V.1). If all local quantization parameters {ηi , εi , ωi }i∈I
simultaneously fulfill ηi ≤ εi and
n
bi−1 (1 − e−κi τ )αi (ε2i )
0 ≤ ηi ≤ min (αi )−1 ◦ αi (ε2i ), γ
o
1
1
1
σu,i (ωi ) −
σd,i (ψi (τ ) + k ε̃i k) − (hi (σi , τ )) 2 ,
−
eκi
eκi
(V.3)
in the continuous and abstract disturbances, admits a disturbance bisimilar abstract system. Then we used the property
of disturbance bisimulation to show that given a network
of stochastic control systems, the abstract systems can be
computed compositionally. One can then use this paper’s
claim for compositional synthesis of controllers for networks
of stochastic control systems, as is done in [5, Sec. VII] for
network of deterministic systems.
with {ε̃i }i∈I defined as (V.2), then the relation
Rεi ε̃i = (x̂i , xi ) ∈ [Xi ]ηi × Xi,r | E[Vi (x̂i , xi )] ≤ αi (ε2i )
[1] D. Angeli. A Lyapunov approach to incremental stability properties.
IEEE Transactions on Automatic Control, 47(3):410–21, 2002.
[2] P. Jagtap and M. Zamani. Automated synthesis of infinite dimensional
stochastic hybrid systems. arXiv preprint arXiv:1704.03690, 2017.
[3] A. A. Julius and G. J. Pappas. Approximations of stochastic hybrid
systems. IEEE Transaction on Automatic Control, 54(6):1193–1203,
2009.
[4] I. Karatzas and S.E. Shreve. Brownian Motion and Stochastic
Calculus, volume 113 of Graduate Texts in Mathematics. SpringerVerlag, New York, 2 edition, 1991.
[5] K. Mallik, A.-K. Schmuck, S. Esmaeil Zadeh Soudjani, and R. Majumdar. Compositional abstraction-based controller synthesis for
continuous-time systems. arXiv preprint arXiv:1612.08515v3, 2017.
[6] X. Mao. Stochastic Differential Equations and Applications. Horwood
series in mathematics & applications. Horwood Pub., 2008.
[7] B. K. Oksendal. Stochastic differential equations: an introduction with
applications. Springer, 5th edition, November 2002.
[8] G. Reissig, A. Weber, and M. Rungger. Feedback refinement relations
for the synthesis of symbolic controllers. IEEE Transactions on
Automatic Control, PP(99):1–1, 2016.
[9] S. Esmaeil Zadeh Soudjani and A. Abate. Adaptive and sequential
gridding procedures for the abstraction and verification of stochastic
processes. SIAM J. Applied Dynamical Systems, 12(2):921–956, 2013.
[10] S. Esmaeil Zadeh Soudjani, A. Abate, and R. Majumdar. Dynamic
Bayesian networks as formal abstractions of structured stochastic
processes. In Luca Aceto and David de Frutos-Escrig, editors, 26th
International Conference on Concurrency Theory, CONCUR 2015,
Madrid, Spain, September 1.4, 2015, volume 42 of LIPIcs, pages 169–
183. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2015.
[11] P. Tabuada. Verification and Control of Hybrid Systems, A symbolic
approach. Springer, 1st edition, June 2009.
[12] M. Zamani, P. M. Esfahani, R. Majumdar, A. Abate, and J. Lygeros.
Symbolic control of stochastic systems via approximately bisimilar finite abstractions. IEEE Transactions on Automatic Control,
59(12):3135–3150, 2014.
[13] M. Zamani, M. Rungger, and P.M. Esfahani. Construction of approximations of stochastic control systems: A compositional approach. In
Decision and Control (CDC), 2015 IEEE 54th Annual Conference on,
pages 525–530. IEEE, 2015.
is a disturbance bisimulation relation in the second moment
with parameters (εi , ε̃i ) between Pτ ηi ωi (Σi ) and Pτ (Σi ) for
all i ∈ I.
C. Composition of Approximations
We have discussed in Sec. V-B that the sets {Pτ (Σi )}i∈I
and {Pτ ηi ωi (Σi )}i∈I of metric systems is compatible. We
also established conditions on local quantization parameters
under which the metric systems Pτ (Σi ) and Pτ ηi ωi (Σi ) are
disturbance bisimilar for any i ∈ I.
We now use the fundamental property of disturbance
bisimulation relation proved in [5] that disturbance bisimilarity is preserved under composition of components in a
network. This property together with Thm. 5.5 result in the
following theorem that explicitly gives the disturbance bisimulation relation on the composed abstractions of components
in a network.
Theorem 5.6: Given the preliminaries of Thm. 5.5 and
I ′ ⊆ I, let JPτ (Σi )Ki∈I ′ and JPτ ηi ωi (Σi )Ki∈I ′ be systems
with state spaces Xr and Xτ ηω , composed from the sets
{Pτ (Σi )}i∈I and {Pτ ηi ωi (Σi )}i∈I , respectively. Then the
relation
T T
T
T T
Rε ε̃ ={([q̂1T . . . q̂|I
. . . q|I
′ | ] , [q1
′ | ] ) ∈ Xτ ηω × Xr |
(q̂i , qi ) ∈ Rεi ε̃i , ∀i ∈ I ′ )}
(V.4)
is a disturbance bisimulation relation between JPτ (Σi )Ki∈I ′
and JPτ ηi ωi (Σi )Ki∈I ′ with parameters
Q
Q
ε = k i∈I ′ {εi } k and ε̃ = j∈NI (I ′ ) {εj }.
Note that in the special case I ′ = I the composed
system replaces the overall network without extra external
disturbances. In this case it is easy to see that the relation in
Thm. 5.6 simplifies to a usual bisimulation relation.
Corollary 5.7: Given the premises of Thm. 5.6 and that
I ′ = I, the relation Rε ε̃ in (V.4) is an ε-approximate bisimulation relation between JPτ (Σi )Ki∈I and JPτ ηi ωi (Σi )Ki∈I .
VI. C ONCLUSION
In this paper, we extended our previous result on compositional abstraction based control for non-probabilistic control
systems to stochastic control systems. We gave sufficient
conditions s.t. a stochastic control system, admitting a δISS-Mq Lyapunov function and subjected to small mismatch
R EFERENCES
A PPENDIX
A. Proof of Thm. 4.2
First observe that Xτ ηω ⊂ Xτ , hence the metric d on Xτ
is also a metric on Xτ ηω . Now we prove the three parts of
Def. 4.1 separately.
(a) By definition of Rε ε̃ in (IV.4), (x̂, x) ∈ Rε ε̃ implies
1/2
1/2
≤ α−1 (E [V (x̂, x)])
d(x̂, x) = E k x̂ − x k2
≤ ε.
(A.1)
We used the convexity assumption of α and the Jensen
inequality [7] to show the inequalities in (A.1).
(b) Given a pair (x̂, x) ∈ Rε ε̃ , for any µ ∈ Uτ , observe
that there exists a µ̂ ∈ Uτ ηω s.t. k uµ̂ − uµ k ≤ ω holds.
Given any ν̂ ∈ Wτ ηω and ν ∈ Wτ s.t. e(wν̂ , wν ) ≤ ε̃ holds,
observe that k wν̂ −wν k ≤ k e(wν̂ , wν ) k ≤ k ε̃ k from (III.4).
Now we can apply transitions δτ (x, µ, ν) = ξxµν (τ ) = x′ ,
ξx̂µ̂ν̂ (τ ) = z, ξ x̂µ̂ν̂ (τ ) = z, and observe that there exists
a x̂′ ∈ Xτ ηω s.t. k x̂′ − z k ≤ η, and hence we have
δτ ηω (x̂, µ̂, ν̂) = x̂′ . Now consider the following derivation:
E V (x̂′ , x′ )
= E V (z, x′ ) + V (x̂′ , x′ ) − V (z, x′ )
= E V (z, x′ ) + E V (x̂′ , x′ ) − V (z, x′ )
1
1
σu (ω) +
σd (ψi (τ ) + k ε̃ k)
≤ α(ε2 )e−κτ +
eκ
eκ
′
+E γ
b k x̂ − z k
1
1
2 −κτ
≤ α(ε )e
+
σu (ω) +
σd (ψi (τ ) + k ε̃ k)
eκ
eκ
′
+b
γ E k x̂ − z + z − z k
1
1
2 −κτ
≤ α(ε )e
+
σu (ω) +
σd (ψi (τ ) + k ε̃ k)
eκ
eκ
′
+b
γ E [k z − z k] + k x̂ − z k
1
1
2 −κτ
σu (ω) +
σd (ψi (τ ) + k ε̃ k)
≤ α(ε )e
+
eκ
eκ
1
+b
γ (h(σ, τ )) 2 + η
(A.2)
≤ α(ε2 ).
Hence by Eqn. (IV.4), (x̂′ , x′ ) ∈ Rε ε̃ .
(c) Given a pair (x̂, x) ∈ Rε ε̃ , for any µ̂ ∈ Uτ ηω , observe
that we can choose µ ∈ Uτ s.t. µ = µ̂, i.e., k uµ̂ − uµ k = 0.
Given any ν ∈ Wτ and ν̂ ∈ Wτ ηω s.t. e(wν̂ , wν ) ≤ ε̃, we
have as before δτ (x, µ, ν) = ξxµν (τ ) = x′ , ξx̂µ̂ν̂ (τ ) = z,
ξ x̂µ̂ν̂ (τ ) = z, and observe that there exists a x̂′ ∈ Xτ ηω s.t.
k x̂′ − z k ≤ η, and hence we have δτ ηω (x̂, µ̂, ν̂) = x̂′ . With a
very similar derivation as in (A.2) it follows from Eqn. (IV.4)
that (x̂′ , x′ ) ∈ Rε ε̃ .
B. Proof of other statements
Proof: [Lemma 2.5] The proof of Lemma 2.5 can be
obtained from the proof of Lemma 3.4 in [12] by replacing
all instances of f (x, u), f (z, u), f (x′ , u) and f (x′ , u′ )
with f (x, u, w), f (z, u, w) , f (x′ , u, w) and f (x′ .u′ , w′ )
respectively, and defining the positive
κ = κ
e, the
√ constant
2
2 2
K∞ functions
σ
(r)
=
nL
/e
κ
k
P
k
r
and
σ
(r)
=
u
d
u
√
nL2w /e
κ k P k2 r.
Proof: [Lemma 2.6] The proof of Lemma 2.6 follows
closely the proof of Lemma 3.8 in [12] and hence is omitted.
Proof: [Lemma 5.2]Q Pick any i ∈ I, wi ∈ Wi
and observe that wi = j∈NI (i) {xj }. By the choice of
Xj,τ ηj ωj as [Xj ]ηj we furthermore know that for any xj
there exists x̂j s.t. k xj − x̂j k ≤ ηj ≤ εj . Now recall
Q
Q
fi =
that W
j∈NI (i) [Xj ]ηj . Using
j∈NI (i) Xj,τ ηj ωj =
the definition of ε̃i in (V.2) and e in (III.3) we therefore
f
know that forQany wi ∈ Wi there exists
Q w̃i ∈ Wi s.t.
k} ≤ j∈NI (i) {εj } =
e(wi , w̃i ) = j∈NI (i) {k xj − x̂j Q
ε̃i .QFurthermore, k wi − w̃i k = k j∈NI (i) {xj − x̂j } k =
k j∈NI (i) {k xj − x̂j k} k = k e(wi , w̃i ) k.
Proof: [Lemma 5.4] The proof follows from the follow-
ing derivation:
d(νi (t), ν̂i (t))
1
= E k νi (t) − ν̂i (t) k2 2
1
= E k νi (t) − ν̂i (0) k2 2
1
= E k νi (t) − νi (0) + νi (0) − ν̂i (0) k2 2
1
1
≤ E k νi (t) − νi (0) k2 2 + E k νi (0) − ν̂i (0) k2 2
21
Y
= E k
{ξj (t) − ξj (0)} k2 + k ε̃i k
j∈NI (i)
!2 12
sup k ξj (t) − ξj (0) k + k ε̃i k
= E
≤ E
≤
j∈NI (i)
X
j∈NI (i)
X
j∈NI (i)
≤
X
j∈NI (i)
≤
X
j∈NI (i)
21
k ξj (t) − ξj (0) k2 + k ε̃i k
21
E k ξj (t) − ξj (0) k22 + k ε̃i k
2
12
2t 1 + Ekξj (0)k2 (t + 1)eαj t + k ε̃i k
!
21
2t 1 + sup kxj k22 (t + 1)eαj t + k ε̃i k
xj ∈Xj
p
where αj = Kj + 2 Kj and Kj is the constant K as given
in Prop. 5.3 for the j-th system, and the last step follows
from Prop. 5.3. We define for system Σi the K∞ function
i 21
h
P
ψi (t) := t(t + 1) j∈NI (i) βj eαj t , which concludes the
proof.
| 3cs.SY
|
Recovery analysis for weighted mixed ℓ2/ℓp minimization
arXiv:1709.00257v2 [cs.IT] 18 Oct 2017
with 0 < p ≤ 1
Zhiyong Zhou∗, Jun Yu
Department of Mathematics and Mathematical Statistics, Umeå University,
Umeå, 901 87, Sweden
October 19, 2017
Abstract: We study the recovery conditions of weighted mixed ℓ2 /ℓp (0 < p ≤ 1) minimization for block
sparse signal reconstruction from compressed measurements when partial block support information is
available. We show that the block p-restricted isometry property (RIP) can ensure the robust recovery.
Moreover, we present the sufficient and necessary condition for the recovery by using weighted block p-null
space property. The relationship between the block p-RIP and the weighted block p-null space property
has been established. Finally, we illustrate our results with a series of numerical experiments.
Keywords: Compressive sensing; Prior support information; Block sparse; Non-convex minimization.
1
Introduction
Since its advent [3–5, 10], Compressive Sensing (CS) has attracted considerable attentions (see the
monographs [13, 18] for a comprehensive view). CS aims to recover an unknown signal with underdetermined linear measurements. Specifically, the standard CS problem consists in the reconstruction of a
N -dimensional sparse or compressible signal x from a significantly fewer number of linear measurements
y = Ax + e ∈ Rm with m ≪ N and the noise e satisfying kek2 ≤ ε for some known constant ε > 0. In
fact, if the measurement matrix A satisfies the restricted isometry property (RIP) condition, the robust
recovery can be guaranteed by using the ℓ1 minimization (Chapter 6, [18])
min kzk1 , subject to ky − Azk2 ≤ ε.
z
(1)
To obtain better recovery performance, the structures and prior information of signals are incorporated
in the recovery algorithms. In this paper, we consider both cases.
As for the structure, we assume that the unknown signal x is block sparse or nearly block sparse
[11, 12, 14–16], which means that the nonzero entries of x occur in clusters. Block sparse model appears
in many practical scenarios, such as when dealing with multi-band signals [27], in measurements of gene
expression levels [30], and in colour imaging [24]. Moreover, block sparse model can be used to treat
the problems of multiple measurement vector (MMV) [8, 14, 26] and sampling signals that lie in a union
∗
Corresponding author, [email protected].
1
Pn
of subspaces [2, 14, 27]. With N =
i=1 di , we define the i-th block x[i] of a length-N vector x over
I = {d1 , · · · , dn }. The i-th block is of length di , and the blocks are formed sequentially so that
x = (x1 · · · xd1 xd1 +1 · · · xd1 +d2 · · · xN −dn +1 · · · xN )T .
{z
}
| {z } |
{z
} |
(2)
min kzk2,1 , subject to ky − Azk2 ≤ ε.
(3)
xT [1]
xT [n]
xT [2]
Without loss of generality, we assume that d1 = d2 = · · · = dn = d, implying that N = nd. A vector
x ∈ RN is called block k-sparse over I = {d, · · · , d} if x[i] is nonzero (i.e., kx[i]k2 > 0) for at most k indices
P
i. In other words, by denoting kxk2,0 = ni=1 I(kx[i]k2 > 0), a block k-sparse vector x can be defined by
P
kxk2,0 ≤ k. For any p > 0, we define the mixed ℓ2 /ℓp norm kxk2,p = ( ni=1 kx[i]kp2 )1/p . To make explicit use
of the block structure to achieve better sparse recovery performance, the corresponding extended version
of sparse representation algorithm has been developed, namely the mixed ℓ2 /ℓ1 minimization,
z
It was shown in [14] that if the measurement matrix A satisfies the block RIP condition which generalizes
the conventional RIP notion, then the mixed ℓ2 /ℓ1 -norm recovery algorithm is guaranteed to recover any
block sparse signal, irrespectively of the locations of the nonzero blocks. Furthermore, recovery will be
robust in the presence of noise and modeling errors (i.e., when the vector is not exactly block sparse).
Other existing recovery algorithms include group lasso [39], adaptive lasso [23], iterative reweighted ℓ2 /ℓ1
recovery algorithms [40], block version of the CoSaMP algorithm [12], and the extensions of the CoSaMP
algorithm and of the Iterative Hard Thresholding algorithm [1].
On the other hand, we also consider the case that an estimate of the block support of the signal is
available. The related literatures on signal recovery using partial support or block support information
include [19–22, 28, 29, 31–33, 38]. For an arbitrary signal x ∈ RN defined as (2), let xk be its best approximation with k nonzero blocks, so that xk minimizes kx − gk2,1 over all block k-sparse vectors g. Let T0
be the block support of xk , where T0 ⊂ {1, · · · , n} and |T0 | ≤ k. Let T̃ , the block support estimate, be a
subset of {1, 2 · · · , n} with cardinality |T̃ | = ρk, where 0 ≤ ρ ≤ a for some a > 1 and |T̃ ∩ T0 | = αρk (for
interpretation of ρ and α see Remark 1 in Section 2). To incorporate prior block support information T̃ ,
the weighted mixed ℓ2 /ℓ1 minimization
n
ω ∈ [0, 1], i ∈ T̃
X
wi kz[i]k2 , subject to ky − Azk2 ≤ ε, where wi =
min
(4)
z
1,
i ∈ T̃ c
i=1
is adopted. The main idea is to choose ω such that the ℓ2 norm of the blocks of x that are ”expected” to
be large are penalized less in the weighted objective function.
Moreover, as is shown in many literatures [6, 17, 37], ℓp -minimization with 0 < p < 1 allows exact
recovery with fewer measurements than that by ℓ1 -minimization. Thus, it is natural to adopt the nonconvex
minimization to the setting of block sparse signal reconstruction with prior block support information.
Specifically, we consider the weighted mixed ℓ2 /ℓp (0 < p ≤ 1) minimization problem:
n
ω ∈ [0, 1], i ∈ T̃
X
p
wi kz[i]k2 , subject to ky − Azk2 ≤ ε, where wi =
.
(5)
min
z
1,
i ∈ T̃ c
i=1
2
When there is no prior block support information (ω = 1), the mixed ℓ2 /ℓp (0 < p ≤ 1) minimization
problem has been studied in [34, 35]. And the case that there is partially known signal block support but
with ω = 0 was considered in [20]. We generalize the existing results to incorporating the prior known
block support information with a weight ω ∈ [0, 1]. In summary, the main contributions of our work include
the following aspects. First, we provide the recovery analysis for the weighted mixed ℓ2 /ℓp (0 < p ≤ 1)
minimization by using block p-RIP condition. This result extends the existing literatures [19, 20, 33–35].
Second, we propose the weighted block p-null space property and present the sufficient and necessary
condition for the weighted mixed ℓ2 /ℓp (0 < p ≤ 1) minimization by this new tool. Third, we establish the
relationship between block p-RIP condition and weighted block p-null space property. Finally, we illustrate
our results via a series of simulations.
The paper is organized as follows. In Section 2, we present the sufficient recovery condition by using
block p-RIP. In Section 3, we introduce the notion of weighted block p-null space property (NSP) and
establish the sufficient and necessary condition with this new tool. In Section 4, we establish the relationship
between these two conditions. In Section 5, we conduct some simulations to illustrate the theoretical results.
Section 6 is devoted to the conclusion.
2
Block p-RIP
As for the weighted mixed ℓ2 /ℓp (0 < p ≤ 1) minimization, we can obtain the reconstruction guarantees
by using block p-RIP. We begin with introducing the definition of block restricted p-isometry constant,
which is a natural extension of the conventional restricted p-isometry constant.
Definition 1 ( [7, 20, 34, 35]) Given a measurement matrix A ∈ Rm×N and 0 < p ≤ 1. Then the block
p-restricted isometry constant (RIC) δk|I over I = {d1 , · · · , dn } of order k is defined to be the smallest
positive number such that
(1 − δk|I )kxkp2 ≤ kAxkpp ≤ (1 + δk|I )kxkp2
(6)
for all x ∈ RN that are block k-sparse over I.
For convenience, we write δk for the block p-RIC δk|I whenever there is no confusion. Then, we have
the sufficient condition for the robust recovery result by using (5) with the block p-RIC δk .
Theorem 1 Let x ∈ RN , and xk be its best approximation with k nonzero blocks, supported on block
index set T0 . Let T̃ ⊂ {1, 2, · · · , n} be an arbitrary set. Define ρ and α as before such that |T̃ | = ρk and
|T̃ ∩ T0 | = αρk. Suppose that there exists an a ∈ Z, with a ≥ (1 − α)ρ, a > 1, and the measurement matrix
A satisfies
δak +
a1−p/2
a1−p/2
δ(a+1)k <
− 1,
γ
γ
3
(7)
where γ = ω + (1 − ω)(1 + ρ − 2αρ)1−p/2 for some given 0 ≤ ω ≤ 1. Then the solutions x♯ of problem (5)
obeys
kx♯ − xk2 ≤ C1
ωkx − xk kp2,p + (1 − ω)kxT̃ c ∩T c kp2,p
0
k1/p−1/2
1/p
+ C2 ε,
(8)
for some positive constants C1 and C2 .
Remark 1. Note that this theorem involves two important ratios: ρ determines the ratio of the estimated
block support size to the actual block support of xk (or the block support of x if x is block k-sparse),
while α determines the ratio of the number of block indices in block support of xk that were accurately
0|
.
estimated in T̃ to the size of T̃ . Specifically, α = |T̃|∩T
T̃ |
Remark 2. The constants C1 and C2 are explicitly given by the following expressions:
22/p−1 a1/2−1/p (1 + δak )1/p + (1 − δ(a+1)k )1/p
C1 =
,
[(1 − δ(a+1)k ) − ap/2−1 (1 + δak )γ]1/p
C2 =
21/p m1/p−1/2 (1 + a1/2−1/p γ 1/p )
.
[(1 − δ(a+1)k ) − ap/2−1 (1 + δak )γ]1/p
Remark 3. For Theorem 1 to be held, it is sufficient that A satisfies
δ(a+1)k < δ :=
a1−p/2 − [ω + (1 − ω)(1 + ρ − 2αρ)1−p/2 ]
.
a1−p/2 + [ω + (1 − ω)(1 + ρ − 2αρ)1−p/2 ]
(9)
Next, we illustrate how the slighted stronger sufficient conditions (9) vary with α and ω for p =
0.01, 0.5, 1, respectively. In Figure 1, for each p, we plot δ versus ω for different values of α, where we set
the parameters a = 3 and ρ = 1. We observe that as α increases the sufficient condition on the block p-RIC
becomes weaker (δ getting larger), allowing for a wider class of measurement matrices A. For each p with
α > 0.5, the sufficient condition becomes stronger as ω increases. For instance, when α = 0.7, p = 0.5,
with ω = 0.2 it suffices to have δ = 0.5072, compared with δ = 0.3902 for ω = 1. The opposite conclusion
holds for the case α < 0.5. When α = 0.5, the sufficient condition remains the same for different ω. From
another point of view, as p decreases, the sufficient condition becomes weaker, which reflects the benefits
of using nonconvex minimization.
Remark 4. By setting ω = 1 and a = b2/(2−p) with b > 1, this theorem reduces to Theorem 2.1 in [35].
In addition, by setting ω = 0, α = 0.5 and a = ρ = b2/(2−p) with b > 1, then this theorem goes to Theorem
2 in [20] with s = ak.
Proof. Let x♯ = x + h be a solution of problem (5), where x is the unknown true signal and h is the
approximation error. Throughout the paper, xT will denote the vector equal to x on the block index set
4
p=0.01
0.9
p=0.5
0.8
α=0.1
α=0.3
α=0.5
α=0.7
α=0.9
0.8
p=1
0.6
α=0.1
α=0.3
α=0.5
α=0.7
α=0.9
α=0.1
α=0.3
α=0.5
α=0.7
α=0.9
0.55
0.7
0.5
0.7
0.6
0.6
0.5
Block restricted p-isometry constant
Block restricted p-isometry constant
Block restricted p-isometry constant
0.45
0.5
0.4
0.4
0.35
0.3
0.25
0.4
0.3
0.3
0.2
0.2
0.15
0.2
0.1
0
0.2
0.4
0.6
0.8
1
0.1
0
0.2
Weights (ω)
0.4
0.6
0.8
1
0
0.2
0.4
Weights (ω)
0.6
0.8
1
Weights (ω)
Figure 1: Comparison of the sufficient conditions (9) for the block restricted p-isometry constants with various of α
for p = 0.01, 0.5, 1. We set a = 3 and ρ = 1 for all the figures.
T and zero elsewhere. Then, we have
ωkx♯T̃ kp2,p + kx♯T̃ c kp2,p ≤ ωkxT̃ kp2,p + kxT̃ c kp2,p .
That is,
ωkxT̃ + hT̃ kp2,p + kxT̃ c + hT̃ c kp2,p ≤ ωkxT̃ kp2,p + kxT̃ c kp2,p .
Consequently,
ωkxT̃ ∩T0 + hT̃ ∩T0 kp2,p + ωkxT̃ ∩T c + hT̃ ∩T c kp2,p + kxT̃ c ∩T0 + hT̃ c ∩T0 kp2,p + kxT̃ c ∩T c + hT̃ c ∩T c kp2,p
0
≤
ωkxT̃ ∩T0 kp2,p
+
ωkxT̃ ∩T c kp2,p
0
+
0
kxT̃ c ∩T0 kp2,p
0
+
0
kxT̃ c ∩T c kp2,p .
0
The forward and reverse triangle inequalities implies
ωkhT̃ ∩T c kp2,p + khT̃ c ∩T c kp2,p ≤ khT̃ c ∩T0 kp2,p + ωkhT̃ ∩T0 kp2,p + 2 kxT̃ c ∩T c kp2,p + ωkxT̃ ∩T c kp2,p .
0
0
0
0
Adding and subtracting ωkhT̃ c ∩T c kp2,p on the left hand side, and ωkhT̃ c ∩T0 kp2,p on the right, we obtain
0
ωkhT0c kp2,p + (1 − ω)khT̃ c ∩T c kp2,p ≤ ωkhT0 kp2,p + (1 − ω)khT̃ c ∩T0 kp2,p + 2 ωkxT0c kp2,p + (1 − ω)kxT̃ c ∩T c kp2,p .
0
0
But we can also write
khT0c kp2,p = ωkhT0c kp2,p + (1 − ω)khT̃ c ∩T c kp2,p + (1 − ω)khT̃ ∩T c kp2,p .
0
5
0
Therefore, we have
khT0c kp2,p ≤ ωkhT0 kp2,p + (1 − ω)khT̃ ∩T c kp2,p + (1 − ω)khT̃ c ∩T0 kp2,p + 2 ωkxT0c kp2,p + (1 − ω)kxT̃ c ∩T c kp2,p .
0
0
Let the set T̃α = T0 ∩ T̃ , then we can write khT̃ ∩T c kp2,p + khT̃ c ∩T0 kp2,p = kh(T0 ∪T̃ )\T̃α kp2,p and simplify the
0
bound on
khT0c kp2,p ≤ ωkhT0 kp2,p + (1 − ω)kh(T0 ∪T̃ )\T̃α kp2,p + 2 ωkxT0c kp2,p + (1 − ω)kxT̃ c ∩T c kp2,p .
(10)
0
Next, we decompose hT0c into disjoint block index set Tj , each of Tj (j ≥ 1) consists of ak blocks, where
a > 1. That is, T1 indexes the ak blocks with largest ℓ2 norm of hT0c , T2 indexes the second ak blocks with
P
largest ℓ2 norm of h(T0 ∪T1 )c , and so on. This gives hT0c =
hTj . For each i ∈ Tj (j ≥ 2), it is easy to see
j≥1
that
khTj [i]kp2
khTj−1 kp2,p
khTj−1 [i]kp2 + · · · + khTj−1 [ak]kp2
≤
=
.
ak
ak
Then
khTj [i]k22 ≤
khTj−1 k22,p
(ak)2/p
khTj−1 kp2,p
p
.
khTj k2 ≤
(ak)1−p/2
, khTj k22 ≤
akkhTj−1 k22,p
(ak)2/p
,
Thus,
X
X
X
X
khTj k2,p .
khTj−1 k2,p = (ak)1/2−1/p
kh(T0 ∪T1 )c k2 ≤ k
hTj k2 ≤
khTj k2 ≤ (ak)1/2−1/p
j≥2
j≥1
j≥2
j≥2
Hence,
p
X
X
kh(T0 ∪T1 )c kp2 ≤ (ak)p/2−1 khTj k2,p ≤ (ak)p/2−1
khTj kp2,p = (ak)p/2−1 khT0c kp2,p .
j≥1
j≥1
Combining the above expression with (10), we get
"
kh(T0 ∪T1 )c kp2 ≤ (ak)p/2−1
#
ωkhT0 kp2,p + (1 − ω)kh(T0 ∪T̃ )\T̃α kp2,p + 2 ωkxT0c kp2,p + (1 − ω)kxT̃ c ∩T c kp2,p .
0
(11)
By Hölder’s inequality it follows that
kAhkpp ≤
m
X
i=1
(|(Ah)i |p )2/p
!p/2
·
m
X
i=1
!1−p/2
1
= m1−p/2 kAhkp2 .
Since x♯ is a solution of problem (5) and ky − Axk2 = kek2 ≤ ε, so we have
kAhk2 = kA(x♯ − x)k2 ≤ kAx − yk2 + kAx♯ − yk2 ≤ 2ε.
6
Thus, kAhkpp ≤ m1−p/2 kAhkp2 ≤ m1−p/2 (2ε)p . Therefore, we obtain
kAhT0 ∪T1 kpp ≤ kAhkpp + kAh(T0 ∪T1 )c kpp ≤ m1−p/2 (2ε)p +
X
kAhTj kpp
j≥2
≤m
1−p/2
X
(2ε) + (1 + δak )
khTj kp2 .
p
j≥2
Moreover, we have
X
X
khTj−1 kp2,p ≤ (ak)p/2−1 khT0c kp2,p .
khTj kp2 ≤ (ak)p/2−1
j≥2
j≥2
Hence
kAhT0 ∪T1 kpp ≤ m1−p/2 (2ε)p + (ak)p/2−1 (1 + δak )khT0c kp2,p
≤ m1−p/2 (2ε)p + 2(ak)p/2−1 (1 + δak ) ωkxT0c kp2,p + (1 − ω)kxT̃ c ∩T c kp2,p
0
p/2−1
+ ω(ak)
(1 +
δak )khT0 kp2,p
p/2−1
+ (1 − ω)(ak)
(1 + δak )kh(T0 ∪T̃ )\T̃α kp2,p .
Noting that |(T0 ∪ T̃ )\ T̃α | = (1+ρ−2αρ)k, we obtain kh(T0 ∪T̃ )\T̃α kp2,p ≤ [(1 + ρ − 2αρ)k]1−p/2 kh(T0 ∪T̃ )\T̃α kp2 .
Since the set T1 contains the ak blocks with largest ℓ2 norm of hT0c with a > 1, and |T̃ \ T̃α | = (1−α)ρk ≤ ak,
it follows that kh(T0 ∪T̃ )\T̃α kp2 ≤ khT0 ∪T1 kp2 , which implies that kh(T0 ∪T̃ )\T̃α kp2,p ≤ [(1 + ρ − 2αρ)k]1−p/2 khT0 ∪T1 kp2 .
In addition, khT0 kp2,p ≤ k1−p/2 khT0 kp2 ≤ k1−p/2 khT0 ∪T1 kp2 . Thus,
(1 − δ(a+1)k )khT0 ∪T1 kp2 ≤ kAhT0 ∪T1 kpp
≤ m1−p/2 (2ε)p + 2(ak)p/2−1 (1 + δak ) ωkxT0c kp2,p + (1 − ω)kxT̃ c ∩T c kp2,p
0
+ ωa
p/2−1
(1 +
δak )khT0 ∪T1 kp2
+ (1 − ω)a
Therefore, if (1 − δ(a+1)k ) − ap/2−1 (1 + δak )γ > 0, that is δak +
khT0 ∪T1 kp2 ≤
p/2−1
1−p/2
(1 + ρ − 2αρ)
a1−p/2
δ(a+1)k
γ
<
a1−p/2
γ
(1 + δak )khT0 ∪T1 kp2 .
− 1, then we have
m1−p/2 (2ε)p + 2(ak)p/2−1 (1 + δak ) ωkxT0c kp2,p + (1 − ω)kxT̃ c ∩T c kp2,p
0
(1 − δ(a+1)k ) − ap/2−1 (1 + δak )γ
,
where γ = ω + (1 − ω)(1 + ρ − 2αρ)1−p/2 . According to (11), we also have
kh(T0 ∪T1 )c kp2 ≤ ap/2−1 γkhT0 ∪T1 kp2 + 2(ak)p/2−1 ωkxT0c kp2,p + (1 − ω)kxT̃ c ∩T c kp2,p
0
p
ap/2−1 γm1−p/2 (2ε)p + 2(ak)p/2−1 (1 − δ(a+1)k ) ωkxT0c k2,p + (1 − ω)kxT̃ c ∩T c kp2,p
0
≤
.
(1 − δ(a+1)k ) − ap/2−1 (1 + δak )γ
From kvkp ≤ 21/p−1 kvk1 for all v ∈ R2 , it follows that
khk2 ≤ khT0 ∪T1 k2 + kh(T0 ∪T1 )c k2
1/p
p
p
1/p−1/2
1/p
1/2−1/p
1/p
ωkxT0c k2,p + (1 − ω)kxT̃ c ∩T c k2,p
ε + 2 (ak)
(1 + δak )
2m
0
≤ 21/p−1
p/2−1
1/p
[(1 − δ(a+1)k ) − a
(1 + δak )γ]
7
(12)
1/p
p
1/2−1/p γ 1/p m1/p−1/2 ε + 21/p (ak)1/2−1/p (1 − δ
1/p ωkx c kp + (1 − ω)kx
k
2a
)
c
c
T
(a+1)k
T̃ ∩T0 2,p
0 2,p
+ 21/p−1
p/2−1
1/p
[(1 − δ(a+1)k ) − a
(1 + δak )γ]
≤
1/p
22/p−1 (ak)1/2−1/p (1 + δak )1/p + (1 − δ(a+1)k )1/p ωkxT0c kp2,p + (1 − ω)kxT̃ c ∩T c kp2,p
0
[(1 − δ(a+1)k ) −
+
21/p m1/p−1/2 (1
+
a1/2−1/p γ 1/p )ε
[(1 − δ(a+1)k ) − ap/2−1 (1 + δak )γ]1/p
ap/2−1 (1
= C1
+ δak
)γ]1/p
ωkxT0c kp2,p + (1 − ω)kxT̃ c ∩T c kp2,p
0
k1/p−1/2
1/p
+ C2 ε,
which completes the proof.
Corollary 1 Let A ∈ Rm×N be a measurement matrix, x ∈ RN be a block k-sparse signal supported on
block index set T0 with y = Ax, and 0 < p ≤ 1. Let T̃ ⊂ {1, 2, · · · , n} be an arbitrary set. Define ρ and α
as before such that |T̃ | = ρk and |T̃ ∩ T0 | = αρk. Suppose that there exists an a ∈ Z, with a ≥ (1 − α)ρ,
a > 1, and the measurement matrix A satisfies
δak +
a1−p/2
a1−p/2
δ(a+1)k <
− 1,
γ
γ
(13)
where γ = ω + (1 − ω)(1 + ρ − 2αρ)1−p/2 for some given 0 ≤ ω ≤ 1. Then the unique solution of problem
(5) with ε = 0 is exactly x.
Remark 5. According to the arguments in [20, 35], if we let A be an m × N matrix whose entries are i.i.d
Gaussian random variables with mean zero, then there exist C3 (p) and C4 (p) such that whenever 0 < p ≤ 1
and
m ≥ C3 (p)kd + pC4 (p)k ln(n/k),
(14)
the block p-RIP (7) will hold and the block sparse signal x can be exactly recovered with high probability.
For a given p ∈ (0, 1], C3 (p) and C4 (p) are finite constants, and the second term of (14) has the dominant
impact on the number of measurements in an asymptotic sense. When p → 0, the condition reduces to
m ≥ C3 (0)kd. While, when p = 1, the required number of measurements m ≥ C3 (1)kd + C4 (1)k ln(n/k),
which implies fewer measurements are required with smaller p for exact recovery via weighted mixed ℓ2 /ℓp
minimization than the case that p = 1. This is largely consistent with what we have mentioned in Remark
3, namely, as p decreases, the sufficient condition for the exact recovery becomes weaker. Though we have
only considered the case that d1 = d2 = · · · = dn = d, the results here can be adapted to the case in which
di are not equal via replacing d by the maximal block length max di .
1≤i≤n
3
Weighted Block p-NSP
In this section, we focus on the noise free case, that is e = 0, and the signal x is exactly block k-sparse,
supported on block index set T0 . We obtain the exact recovery condition for the problem (5) with ε = 0
by using weighted block p-null space property. For an index set V ⊂ {1, · · · , n}, we define
Γs (V ) := {U ⊂ {1, · · · , n} : |(V ∩ U c ) ∪ (V c ∩ U )| ≤ s}.
8
Basically, for any index set V , the set (V ∩ U c ) ∪ (V c ∩ U ) indicates the estimate error of U with respective
to V . So Γs (V ) includes all the index set estimates of V that own an error of size s or less.
Definition 2 Let T ⊂ {1, · · · , n} with |T | ≤ k and T̃ ∈ Γs (T ). Assume the block size equals d. A matrix
A ∈ Rm×N is said to have the weighted nonuniform block p-null space property with parameters T and T̃ ,
and constant C if for any vector h : Ah = 0, we have
ωkhT kp2,p + (1 − ω)khS kp2,p ≤ CkhT c kp2,p ,
(15)
where S = (T̃ ∩ T c ) ∪ (T̃ c ∩ T ). In this case, we say A satisfies ω-d-p-NSP(T, T̃ , C).
Next, we define a weighted uniform block p-null space property that leads to a sufficient and necessary
condition for the exact recovery of all block k-sparse signals from compressive measurements using weighted
mixed ℓ2 /ℓp minimization problem.
Definition 3 Assume the block size equals d. A matrix A ∈ Rm×N is said to have the weighted block
p-null space property with parameters k and s, and constant C if for any vector h : Ah = 0, and for every
block index set T ⊂ {1, · · · , n} with |T | ≤ k and S ⊂ {1, · · · , n} with |S| ≤ s, we have
ωkhT kp2,p + (1 − ω)khS kp2,p ≤ CkhT c kp2,p .
(16)
In this case, we say A satisfies ω-d-p-NSP(k, s, C).
Remark 6. Our notations ω-d-p-NSP(T, T̃ , C) and ω-d-p-NSP(k, s, C) are direct extensions of the recent
notations ω-NSP(T, T̃ , C) and ω-NSP(k, s, C) proposed by [25], which corresponds to the special case that
p = 1 and the block size d = 1. That’s to say ω-1-1-NSP(k, s, C) is equivalent to ω-NSP(k, s, C). Consequently, 1-1-1-NSP(k, k, C) is the standard null space property of order k, i.e, NSP(k, C).
The following theorem presents the sufficient and necessary condition for weighted mixed ℓ2 /ℓp minimization problem to recover all block k-sparse signals when the error in the block support estimate is of
size s or less.
Theorem 2 Assume the block size equals d. Given a matrix A ∈ Rm×N , every block k-sparse signal
x ∈ RN , supported on block index set T0 , is the unique solution of the weighted mixed ℓ2 /ℓp , 0 < p ≤ 1
norm minimization problem (5) with ε = 0 and T̃ = Γs (T0 ), if and only if A satisfies ω-d-p-NSP(k, s, C)
for some positive constant C < 1.
Proof. a) ” ⇒ ”. Assume if A does not satisfy ω-d-p-NSP(k, s, C) for any constant C > 1, then there
exists a vector h : Ah = 0 and block index set T with |T | ≤ k and block index set S with |S| ≤ s, such
that AhT = −AhT c and
ωkhT kp2,p + (1 − ω)khS kp2,p ≥ khT c kp2,p .
Define T̃ = (T c ∩ S) ∪ (T ∩ S c ) so that S = (T ∩ T̃ c ) ∪ (T c ∩ T̃ ). Substituting for S and splitting the set
T , we obtain
ω khT ∩T̃ c kp2,p + khT ∩T̃ kp2,p + (1 − ω) khT ∩T̃ c kp2,p + khT c ∩T̃ kp2,p
9
= khT ∩T̃ c kp2,p + ωkhT ∩T̃ kp2,p + (1 − ω)khT c ∩T̃ kp2,p ≥ khT c kp2,p .
Then, we have
ωkhT ∩T̃ kp2,p + khT ∩T̃ c kp2,p ≥ ωkhT c ∩T̃ kp2,p + khT c ∩T̃ c kp2,p .
In other words, the weighted mixed ℓ2 /ℓp norm of the vector hT equals or exceeds that of −hT c . So hT is
not the unique minimizer, which is contradictory to our condition. Thus, the necessity is verified.
b) ” ⇐ ”. Let x♯ be a minimizer of weighted mixed ℓ2 /ℓp , 0 < p ≤ 1 norm minimization problem (5)
with ε = 0 and T̃ = Γs (T0 ) and define h = x♯ − x. Then, by the optimality of x♯ , we have
ωkxT̃ + hT̃ kp2,p + kxT̃ c + hT̃ c kp2,p ≤ ωkxT̃ kp2,p + kxT̃ c kp2,p ,
which is equivalent to
ωkxT̃ ∩T0 + hT̃ ∩T0 kp2,p + ωkxT̃ ∩T c + hT̃ ∩T c kp2,p + kxT̃ c ∩T0 + hT̃ c ∩T0 kp2,p + kxT̃ c ∩T c + hT̃ c ∩T c kp2,p
0
≤
ωkxT̃ ∩T0 kp2,p
ωkxT̃ ∩T c kp2,p
0
+
+
0
0
kxT̃ c ∩T0 kp2,p
+
0
kxT̃ c ∩T c kp2,p .
0
Since x is strictly block k-sparse and supported on the block index set T0 , thus xT0c = 0. Then we have
ωkhT̃ ∩T c kp2,p + khT̃ c ∩T c kp2,p ≤ khT̃ c ∩T0 kp2,p + ωkhT̃ ∩T0 kp2,p .
0
0
Adding and subtracting ωkhT̃ c ∩T c kp2,p on the left hand side, and ωkhT̃ c ∩T0 kp2,p on the right, we obtain
0
ωkhT̃ ∩T c kp2,p + ωkhT̃ c ∩T c kp2,p + khT̃ c ∩T c kp2,p − ωkhT̃ c ∩T c kp2,p
≤
ωkhT̃ ∩T0 kp2,p
+
0
0
0
0
ωkhT̃ c ∩T0 kp2,p
+ khT̃ c ∩T0 kp2,p
− ωkhT̃ c ∩T0 kp2,p .
Therefore,
ωkhT0c kp2,p + (1 − ω)khT̃ c ∩T c kp2,p ≤ ωkhT0 kp2,p + (1 − ω)khT̃ c ∩T0 kp2,p .
0
Finally, by adding (1 −
ω)khT̃ ∩T c kp2,p
0
to both sides, we have
khT0c kp2,p ≤ ωkhT0 kp2,p + (1 − ω)khT̃ ∩T c kp2,p + (1 − ω)khT̃ c ∩T c kp2,p
0
=
ωkhT0 kp2,p
+ (1 −
ω)khS kp2,p ,
0
(17)
by setting S = (T̃ ∩ T0c ) ∪ (T̃ c ∩ T0 ). Note that when |S| ≤ s, the above inequality is in contradiction with
that A satisfies ω-d-p-NSP(k, s, C) for some constant C < 1, unless h = 0. Hence, we have x♯ = x, and the
sufficiency holds.
4
From Block p-RIP to Weighted Block p-NSP
In this part, we establish the relationship between the block p-RIP and weighted block p-NSP, which
tells us that block p-RIP can directly imply the weighted block p-NSP. As a consequence, we can obtain the
exact recovery for every block k-sparse signal via the weighted mixed ℓ2 /ℓp , 0 < p ≤ 1 norm minimization
problem (5) with ε = 0. The relationship is obtained as follows.
10
Theorem 3 Assume the block size equals d and let x ∈ RN be block k-sparse, supported on block index
set T0 . Let T̃ ⊂ {1, 2, · · · , n} be an arbitrary set and define ρ and α as before such that |T̃ | = ρk and
|T̃ ∩ T0 | = αρk. Let s = (1 + ρ − 2αρ)k, then we have T̃ = Γs (T0 ). Suppose that there exists an a ∈ Z,
with a ≥ (1 − α)ρ, a > 1, and the measurement matrix A satisfies
δak +
a1−p/2
a1−p/2
δ(a+1)k <
− 1,
γ
γ
(18)
where γ = ω + (1 − ω)(s/k)1−p/2 for some given 0 ≤ ω ≤ 1. Then A satisfies ω-d-p-NSP(k, s, C) for some
constant C < 1.
Consequently, according to Theorem 2, we have x is the unique solution of problem (5) with ε = 0.
Proof. To show for any vector h : Ah = 0, and for every block index set T ⊂ {1, · · · , n} with |T | ≤ k, and
S ⊂ {1, · · · , n} with |S| ≤ s, we have
ωkhT kp2,p + (1 − ω)khS kp2,p ≤ CkhT c kp2,p ,
(19)
for some constant C < 1, we only need to show (19) holds for T = G0 and S = H0 , where G0 is the block
index set over the k blocks with largest ℓ2 norm of h, H0 is the block index set over the s blocks with
largest ℓ2 norm of h.
Next, we decompose hGc0 into disjoint block index set Gj , each of Gj (j ≥ 1) consists of ak blocks. That
is, G1 indexes the ak blocks with largest ℓ2 norm of hGc0 . G2 indexes the second ak blocks with largest ℓ2
norm of h(G0 ∪G1 )c , and so on.
Since Ah = 0, then we have
kAhG0 ∪G1 kpp ≤ kAh(G0 ∪G1 )c kpp ≤
X
kAhGj kpp ≤ (1 + δak )
khGj kp2 .
j≥2
j≥2
Moreover, we have
X
X
X
khGj−1 kp2,p ≤ (ak)p/2−1 khGc0 kp2,p .
khGj kp2 ≤ (ak)p/2−1
j≥2
j≥2
Thus,
(1 − δ(a+1)k )khG0 ∪G1 kp2 ≤ kAhG0 ∪G1 kpp ≤ (1 + δak )(ak)p/2−1 khGc0 kp2,p .
It implies that
khG0 ∪G1 kp2 ≤
(ak)p/2−1 (1 + δak )
khGc0 kp2,p .
1 − δ(a+1)k
(20)
In addition, khG0 kp2,p ≤ k1−p/2 khG0 kp2 ≤ k1−p/2 khG0 ∪G1 kp2 . Hence,
khG0 kp2,p ≤ k1−p/2 khG0 ∪G1 kp2 ≤
ap/2−1 (1 + δak )
khGc0 kp2,p .
1 − δ(a+1)k
(21)
For the term khH0 kp2,p , we have khH0 kp2,p ≤ s1−p/2 khH0 kp2 . Moreover, if s ≤ k, we have H0 ⊂ G0
and khH0 kp2 ≤ khG0 kp2 ≤ khG0 ∪G1 kp2 . If s > k, then G0 ⊂ H0 . However, since |H0 \ G0 | = s − k =
11
(1 + ρ − 2αρ)k − k = (1 − 2α)ρk ≤ ak, then we also have khH0 kp2 ≤ khG0 ∪G1 kp2 as the block index set G1
contains the ak blocks with largest ℓ2 norm of hGc0 . As a consequence,
khH0 kp2,p ≤ s1−p/2 khH0 kp2 ≤ s1−p/2 khG0 ∪G1 kp2 ≤
(ak/s)p/2−1 (1 + δak )
khGc0 kp2,p .
1 − δ(a+1)k
(22)
Therefore, we obtain
ωkhG0 kp2,p + (1 − ω)khH0 kp2,p ≤ ω
=
Let C =
5
ap/2−1 γ(1+δak )
,
1−δ(a+1)k
ap/2−1 (1 + δak )
(ak/s)p/2−1 (1 + δak )
khGc0 kp2,p + (1 − ω)
khGc0 kp2,p
1 − δ(a+1)k
1 − δ(a+1)k
ap/2−1 γ(1 + δak )
khGc0 kp2,p .
1 − δ(a+1)k
we have C < 1 under the condition of block p-RIP. The proof is completed.
Simulation
In this section, we conduct several simulations to illustrate our presented theoretical results. We adopt
the iteratively reweighted least squares (IRLS) approach to solve the nonconvex optimization problem. We
begin with x(0) = arg min ky − Axk22 , and set γ0 = 1. Then, let x(t+1) be the solution of
min
x
1
1
ky − Axk22 + kW (t) xk22 ,
2λ
2
(23)
(t)
where λ > 0 is a regularization parameter, and the weight matrix W (t) is defined as Wi
4
p(p−2)
ωi
1/p
kωi x(t) [i]k22 )p/4−1/2 )
= diag((γt +
for the i-th block. Then, x(t+1) can be given explicitly as
−1
[A(W (t) )−1 ]T y.
x(t+1) = (W (t) )−1 [A(W (t) )−1 ]T [A(W (t) )−1 ] + λI
The value of γ is decreased according to the rule γt+1 = 0.1γt , and the iteration is continued until kx(t+1) −
x(t) k2 ≤ 10−5 or the iteration times is larger than 2500. In our experiments, the measurement matrix
A is generated as an m × N matrix with entries drawing from i.i.d standard normal distribution. For
a generated block sparse or nearly block sparse signal x, the measurements y = Ax + σz with z being
the standard Gaussian white noise. We consider several different parameter values to demonstrate the
theoretical results. In each experiment, we report the average results over 20 replications. For IRLS, we
set λ = 10−6 in the noise free case (σ = 0), and λ = 10−2 in the noisy case (σ = 0.01). The recovery
performance is assessed by the signal to noise ratio (SNR), which is given by
kxk2
SNR = 20 log 10
.
kx − x♯ k2
5.1
Exactly block sparse case
We first consider the case that x is exactly block k-sparse with k = 20. In this set of experiments,
the signal is of length N = 400, which is generated by choosing k blocks uniformly at random, and then
choosing the nonzero values from the standard normal distribution for these k blocks.
12
Figure 2 illustrates the recovery performance of the weighted ℓ2 /ℓp minimization problem with p = 0.5
and d = 2 in both with and without noise cases. It can be observed that when α = 0.7, the best recovery
is achieved for very small ω whereas a ω = 1 results in the lowest SNR. On the other hand, when α ≤ 0.5,
the performance of the recovery algorithms is better for large ω than that for small ω. The case ω = 0
results in the lowest SNR. And as is expected that in all settings, comparing to the noise free case, we have
a lower SNR in the noisy case.
Figure 3 shows the averaged SNR using weighted mixed ℓ2 /ℓp minimization for different values of the
ratio parameter ρ with fixed p = 0.5 and d = 2. It shows that when α = 0.7, using a larger support
estimate results in better reconstruction. However, in both the noise free and noisy measurements cases,
the recovery performance is more sensitive to the accuracy α of the block support estimate than the ratio
parameter ρ.
In addition, we illustrate the impacts of p and d in Figure 4 for both the noise free and noisy measurements cases. We fix k = 20, α = 0.7 and ω = 0.5. It is evident that as p decreases, the recovery
performance becomes better. And as d increases, more number of measurements are required to obtain
good reconstructions. These are largely consistent with the theoretical results in Section 2.
α =0.3,σ =0
α =0.5,σ =0
60
50
40
40
40
20
30
20
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
10
0
60
SNR
50
30
80
100
120
140
160
0
60
80
number of measurements m
120
140
160
0
60
180
α =0.5,σ =0.01
55
45
45
45
40
40
40
35
35
35
30
30
30
20
SNR
50
25
25
20
15
10
5
80
100
120
140
160
5
number of measurements m
0
60
140
160
180
α =0.7,σ =0.01
25
15
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
10
180
120
20
15
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
100
55
50
0
60
80
number of measurements m
50
SNR
SNR
100
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
10
number of measurements m
α =0.3,σ =0.01
55
30
20
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
10
180
α =0.7,σ =0
60
50
SNR
SNR
60
80
100
120
140
number of measurements m
160
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
10
5
180
0
60
80
100
120
140
160
180
number of measurements m
Figure 2: Performance of weighted mixed ℓ2 /ℓp recovery with p = 0.5 in terms of SNR for exactly block sparse signal
x depending on ω with k = 20, d = 2, while varying the number of measurements m.
5.2
Nearly block sparse case
Next, we generate x with the ℓ2 norm of its blocks decay like i−θ where i ∈ {1, · · · , n} and θ > 1. In
Figure 5, we illustrate the averaged SNR versus the size of the block support estimate ρ as a proportion
of k for θ = 1.5. To calculate α, we set k = 20, i.e., we are interested in the best 20-term block sparse
13
α =0.3,σ =0
45
α =0.5,σ =0
45
40
40
35
35
α =0.7,σ =0
160
140
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
120
30
20
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
25
20
15
15
10
10
5
5
0
0.4
0
0.4
100
SNR
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
25
SNR
SNR
30
80
60
0.8
1
1.2
1.4
0.6
0.8
1
ρ
α =0.3,σ =0.01
40
35
35
30
30
20
10
10
5
5
1
1
1.2
1.4
1.2
1.4
α =0.7,σ =0.01
55
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
20
15
0.8
0.8
ρ
50
25
15
0.6
0.6
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
45
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
25
0
0.4
20
0.4
1.4
α =0.5,σ =0.01
45
40
SNR
SNR
45
1.2
ρ
1.2
0
0.4
1.4
SNR
0.6
40
40
35
0.6
0.8
1
ρ
1.2
30
0.4
1.4
0.6
0.8
ρ
1
ρ
Figure 3: Performance of weighted mixed ℓ2 /ℓp recovery with p = 0.5 in terms of SNR for exactly block sparse signal
x depending on ω with k = 20, d = 2, while varying the size of the block support estimate ρ as a proportion of k.
d=2,σ =0
70
70
70
60
60
60
50
50
50
40
40
30
20
80
100
120
140
160
0
60
20
p=1
p=0.8
p=0.5
p=0.2
p=0.01
10
180
40
30
20
p=1
p=0.8
p=0.5
p=0.2
p=0.01
10
0
60
SNR
80
30
80
number of measurements m
100
120
140
160
0
60
180
50
40
40
40
20
SNR
50
SNR
50
30
30
20
p=1
p=0.8
p=0.5
p=0.2
p=0.01
10
120
140
number of measurements m
160
0
60
140
160
180
30
20
p=1
p=0.8
p=0.5
p=0.2
p=0.01
10
180
120
d=4,σ =0.01
60
100
100
d=2,σ =0.01
60
80
80
number of measurements m
60
0
60
p=1
p=0.8
p=0.5
p=0.2
p=0.01
10
number of measurements m
d=1,σ =0.01
SNR
d=4,σ =0
80
SNR
SNR
d=1,σ =0
80
80
100
120
140
number of measurements m
160
p=1
p=0.8
p=0.5
p=0.2
p=0.01
10
180
0
60
80
100
120
140
160
180
number of measurements m
Figure 4: Performance of weighted mixed ℓ2 /ℓp recovery in terms of SNR for exactly block sparse signal x depending
on p with k = 20, α = 0.7, ω = 0.5, while varying the number of measurements m.
14
approximation. It shows that mediate values of ω (e.g., ω = 0.2 or 0.5) results in the best recovery.
Generally, larger block support estimate favours better reconstruction result. Finally, we illustrate the
impacts of θ and d for different p in Figure 6. For θ = 1.5, we set k = 20, while for θ = 2, we consider
k = 10. It is evident that the recovery performance improves as θ increases, i.e., the block sparsity of
signal increases. Moreover, decreasing p and d improves the recovery performance when other parameters
are fixed.
α =0.3,σ =0
22
α =0.5,σ =0
23
21
22.5
20
α =0.7,σ =0
28
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
27
26
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
22
25
19
18
SNR
SNR
SNR
21.5
24
23
21
17
16
15
14
0.4
22
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
0.6
20.5
21
20
0.8
1
1.2
19.5
0.4
1.4
20
0.6
0.8
ρ
1.2
19
0.4
1.4
0.6
0.8
ρ
α =0.3,σ =0.01
22
1
α =0.5,σ =0.01
23
21
22.5
20
1
1.2
1.4
1.2
1.4
ρ
α =0.7,σ =0.01
27
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
26
25
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
22
19
17
SNR
SNR
SNR
24
21.5
18
23
21
22
16
20.5
15
14
13
0.4
21
ω=0
ω=0.2
ω=0.5
ω=0.8
ω=1
0.6
20
0.8
1
1.2
1.4
19.5
0.4
20
0.6
0.8
ρ
1
ρ
1.2
1.4
19
0.4
0.6
0.8
1
ρ
Figure 5: Performance of weighted mixed ℓ2 /ℓp recovery in terms of SNR for nearly block sparse signal x with
p = 0.5, m = 80, θ = 1.5, k = 20, d = 2, while varying the size of the block support estimate ρ as a proportion of k.
6
Conclusion
In this paper, we presented the recovery analysis of weighted mixed ℓ2 /ℓp (0 < p ≤ 1) minimization
by using both block p-RIP and weighted block p-null space property. In addition, we established the relationship between these two conditions. A series of simulations were conducted to illustrate our theoretical
results. We largely generalized the existing results [19, 20, 33–35] to incorporating the prior known block
support information with a weight ω ∈ [0, 1]. And from the simulation results, it shows that mediate values
of the weight ω result in the best recovery in most of the cases.
There are some interesting issues left for future work. One is that it might be possible to generalize
ℓ2 -constrained minimization problem considered in this paper to the ℓq -constrained minimization problem
with 2 ≤ q ≤ ∞ (see [9, 32]) and with 0 ≤ q < 2 (see [36]). Another issue that is not addressed here is
to obtain the minimum required number of measurements for perfect recovery directly via weighted block
p-null space property. And the extension to the framework with completely arbitrary weights (see [28])
needs also to be considered.
15
d=1,θ=1.5 with k=20
d=2,θ=1.5 with k=20
d=4,θ=1.5 with k=20
28
30
28
28
26
26
26
24
24
SNR
SNR
22
SNR
24
22
22
20
20
20
18
p=1
p=0.8
p=0.5
p=0.2
p=0.01
18
16
60
80
100
120
140
160
p=1
p=0.8
p=0.5
p=0.2
p=0.01
18
16
60
180
80
number of measurements m
100
120
140
160
p=1
p=0.8
p=0.5
p=0.2
p=0.01
16
14
60
180
80
number of measurements m
d=1,θ=2 with k=10
d=2,θ=2 with k=10
34
34
32
32
30
30
28
28
100
120
140
160
180
number of measurements m
d=4,θ=2 with k=10
34
32
30
26
SNR
SNR
SNR
28
26
26
24
24
24
p=1
p=0.8
p=0.5
p=0.2
p=0.01
22
20
60
80
100
120
140
number of measurements m
160
22
180
20
60
22
p=1
p=0.8
p=0.5
p=0.2
p=0.01
80
100
120
140
number of measurements m
160
p=1
p=0.8
p=0.5
p=0.2
p=0.01
20
180
18
60
80
100
120
140
160
180
number of measurements m
Figure 6: Performance of weighted mixed ℓ2 /ℓp recovery in terms of SNR for nearly block sparse signal x for different
θ and d depending on p with fixed α = 0.7, ω = 0.5, σ = 0.01 for all the cases, while varying the number of
measurements m.
Acknowledgements
This work is supported by the Swedish Research Council grant (Reg.No. 340-2013-5342).
References
[1] Baraniuk, R.G., Cevher, V., Duarte, M.F. and Hegde, C. (2010). Model-based compressive sensing.
IEEE Transactions on Information Theory 56(4) 1982–2001.
[2] Blumensath, T. and Davies, M.E. (2009). Sampling theorems for signals from the union of finitedimensional linear subspaces. IEEE Transactions on Information Theory 55(4) 1872–1882.
[3] Candes, E.J., Romberg, J.K. and Tao, T. (2006). Stable signal recovery from incomplete and inaccurate
measurements. Communications on Pure and Applied Mathematics 59(8) 1207–1223.
[4] Candes, E.J. and Tao, T. (2005). Decoding by linear programming. IEEE Transactions on Information
Theory 51(12) 4203–4215.
[5] Candes, E.J. and Tao, T. (2006). Near-optimal signal recovery from random projections: Universal
encoding strategies?. IEEE Transactions on Information Theory 52(12) 5406–5425.
[6] Chartrand, R. (2007). Exact reconstruction of sparse signals via nonconvex minimization. IEEE Signal
Processing Letters 14(10) 707–710.
16
[7] Chartrand, R. and Staneva, V. (2008). Restricted isometry properties and nonconvex compressive
sensing. Inverse Problems 24 1–14.
[8] Chen, J. and Huo, X. (2006). Theoretical results on sparse representations of multiple-measurement
vectors. IEEE Transactions on Signal Processing 54(12) 4634–4643.
[9] Dirksen, S., Lecué, G. and Rauhut, H. (2016). On the gap between restricted isometry
properties and sparse recovery conditions. IEEE Transactions on Information Theory, (DOI:
10.1109/TIT.2016.2570244).
[10] Donoho, D.L. (2006). Compressed sensing. IEEE Transactions on Information Theory 52(4) 1289–
1306.
[11] Duarte, M.F. and Eldar, Y.C. (2011). Structured compressed sensing: From theory to applications.
IEEE Transactions on Signal Processing 59(9) 4053–4085.
[12] Eldar, Y.C., Kuppinger, P. and Bolcskei, H. (2010). Block-sparse signals: Uncertainty relations and
efficient recovery. IEEE Transactions on Signal Processing 58(6) 3042–3054.
[13] Eldar, Y.C. and Kutyniok, G. (2012). Compressed Sensing: Theory and Applications. Cambridge
University Press.
[14] Eldar, Y.C. and Mishali, M. (2009). Robust recovery of signals from a structured union of subspaces.
IEEE Transactions on Information Theory 55(11) 5302–5316.
[15] Eldar, Y.C. and Rauhut, H. (2010). Average case analysis of multichannel sparse recovery using convex
relaxation. IEEE Transactions on Information Theory 56(1) 505–519.
[16] Elhamifar, E. and Vidal, R. (2012). Block-sparse recovery via convex optimization. IEEE Transactions
on Signal Processing 60(8) 4094–4107.
[17] Foucart, S. and Lai, M.J. (2009). Sparsest solutions of underdetermined linear systems via ℓq minimization for 0 < q ≤ 1. Applied and Computational Harmonic Analysis 26(3) 395-407.
[18] Foucart, S. and Rauhut, H. (2013). A Mathematical Introduction to Compressive Sensing. New York:
Springer-Verlag.
[19] Friedlander, M.P., Mansour, H., Saab, R. and Yilmaz, O. (2012). Recovering compressively sampled
signals using partial support information. IEEE Transactions on Information Theory 58(2) 1122–1134.
[20] He, S., Wang, Y., Wang, J. and Xu, Z. (2016). Block-sparse compressed sensing with partially known
signal support via non-convex minimisation. IET Signal Processing 10(7) 717–723.
[21] Jacques, L. (2010). A short note on compressed sensing with partially known signal support. Signal
Processing 90(12) 3308–3312.
[22] Khajehnejad, M.A., Xu, W., Avestimehr, A.S. and Hassibi, B. (2009). Weighted ℓ1 minimization for
sparse recovery with prior information. In Information Theory, 2009. ISIT 2009. IEEE International
Symposium on (pp. 483-487).
[23] Lv, X., Bi, G. and Wan, C. (2011). The group lasso for stable recovery of block-sparse signal representations. IEEE Transactions on Signal Processing 59(4) 1371–1382.
[24] Majumdar, A. and Ward, R.K. (2010). Compressed sensing of color images. Signal Processing 90(12)
3122–3127.
17
[25] Mansour, H. and Saab, R. (2017). Recovery analysis for weighted ℓ1 -minimization using the null space
property. Applied and Computational Harmonic Analysis 43(1) 23–38.
[26] Mishali, M. and Eldar, Y.C. (2008). Reduce and boost: Recovering arbitrary sets of jointly sparse
vectors. IEEE Transactions on Signal Processing 56(10) 4692–4702.
[27] Mishali, M. and Eldar, Y.C. (2009). Blind multiband signal reconstruction: Compressed sensing for
analog signals. IEEE Transactions on Signal Processing 57(3) 993–1009.
[28] Needell, D., Saab, R. and Woolf, T. (2017). Weighted ℓ1 -minimization for sparse recovery under
arbitrary prior information. Information and Inference, iaw023.
[29] Ince, T., Nacaroglu, A. and Watsuji, N. (2013). Nonconvex compressed sensing with partially known
signal support. Signal Processing 93(1) 338–344.
[30] Parvaresh, F., Vikalo, H., Misra, S. and Hassibi, B. (2008). Recovering sparse signals using sparse
measurement matrices in compressed DNA microarrays. IEEE Journal of Selected Topics in Signal
Processing 2(3) 275–285.
[31] Rauhut, H. and Ward, R. (2016). Interpolation via weighted ℓ1 minimization. Applied and Computational Harmonic Analysis 40(2) 321–351.
[32] Saleh, A.A., Alajaji, F. and Chan, W.Y. (2015). Compressed sensing with non-Gaussian noise and
partial support information. IEEE Signal Processing Letters 22(10) 1703–1707.
[33] Vaswani, N. and Lu, W. (2010). Modified-CS: Modifying compressive sensing for problems with partially known support. IEEE Transactions on Signal Processing 58(9) 4595-4607.
[34] Wang, Y., Wang, J. and Xu, Z. (2013). On recovery of block-sparse signals via mixed ℓ2 /ℓq (0 < q ≤ 1)
norm minimization. EURASIP Journal on Advances in Signal Processing 2013(1):76.
[35] Wang, Y., Wang, J. and Xu, Z. (2014). Restricted p-isometry properties of nonconvex block-sparse
compressed sensing. Signal Processing 104 188–196.
[36] Wen, F., Liu, P., Liu, Y., Qiu, R.C. and Yu, W. (2017). Robust Sparse Recovery in Impulsive Noise
via ℓp -ℓ1 Optimization. IEEE Transactions on Signal Processing 65(1) 105–118.
[37] Xu, Z., Chang, X., Xu, F. and Zhang, H. (2012). ℓ1/2 regularization: A thresholding representation
theory and a fast solver. IEEE Transactions on Neural Networks and Learning Systems 23(7) 1013–1027.
[38] Yu, X. and Baek, S.J. (2013). Sufficient conditions on stable recovery of sparse signals with partial
support information. IEEE Signal Processing Letters 20(5) 539–542.
[39] Yuan, M. and Lin, Y. (2006). Model selection and estimation in regression with grouped variables.
Journal of the Royal Statistical Society: Series B (Statistical Methodology) 68(1) 49–67.
[40] Zeinalkhani, Z. and Banihashemi, A.H. (2015). Iterative reweighted ℓ2 /ℓ1 recovery algorithms for
compressed sensing of block sparse signals. IEEE Transactions on Signal Processing 63(17) 4516–4531.
18
| 7cs.IT
|
arXiv:1612.08261v1 [math.ST] 25 Dec 2016
Change point estimation based on
Wilcoxon tests in the presence of
long-range dependence
Annika Betken
Faculty of Mathematics
Ruhr-Universität Bochum
44780 Bochum, Germany
e-mail: [email protected]
Abstract: We consider an estimator for the location of a shift in the
mean of long-range dependent sequences. The estimation is based on the
two-sample Wilcoxon statistic. Consistency and the rate of convergence for
the estimated change point are established. In the case of a constant shift
height, the 1/n convergence rate (with n denoting the number of observations), which is typical under the assumption of independent observations,
is also achieved for long memory sequences. It is proved that if the change
point height decreases to 0 with a certain rate, the suitably standardized
estimator converges in distribution to a functional of a fractional Brownian
motion. The estimator is tested on two well-known data sets. Finite sample
behaviors are investigated in a Monte Carlo simulation study.
Primary 62G05, 62M10; secondary 60G15, 60G22.
Keywords and phrases: change point estimation, long-range dependence, Wilcoxon test, self-normalization.
Contents
1 Introduction . . .
2 Main Results . .
3 Applications . . .
4 Simulations . . .
5 Proofs . . . . . .
References . . . . . .
A Auxiliary Results
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
4
7
9
16
30
32
1. Introduction
Suppose that the observations X1 , . . . , Xn are generated by a stochastic process
(Xi )i≥1
Xi = µi + Yi ,
∗ Research supported by the German National Academic Foundation and Collaborative
Research Center SFB 823 Statistical modelling of nonlinear dynamic processes.
1
A. Betken/Wilcoxon-based change point estimation
2
where (µi )i≥1 are unknown constants and where (Yi )i≥1 is a stationary, longrange dependent (LRD, in short) process with mean zero. A stationary process
(Yi )i≥1 is called “long-range dependent” if its autocovariance function ρ, ρ(k) :=
Cov(Y1 , Yk+1 ), satisfies
ρ(k) ∼ k −D L(k), as k → ∞,
(1)
where 0 < D < 1 (referred to as long-range dependence (LRD) parameter) and
where L is a slowly varying function.
Furthermore, we assume that there is a change point in the mean of the
observations, that is
(
µ,
for i = 1, . . . , k0 ,
µi =
µ + hn , for i = k0 + 1, . . . , n,
where k0 = bnτ c denotes the change point location and hn is the height of the
level-shift.
In the following we differentiate between fixed and local changes. Under fixed
changes we assume that hn = h for some h 6= 0. Local changes are characterized
by a sequence hn , n ∈ N, with hn −→ 0 as n −→ ∞; in other words, in a model
where the height of the jump decreases with increasing sample size n.
In order to test the hypothesis
H : µ1 = . . . = µn
against the alternative
A : µ1 = . . . = µk 6= µk+1 = . . . = µn for some k ∈ {1, . . . , n − 1}
the Wilcoxon change point test can be applied. It rejects the hypothesis for large
values of the Wilcoxon test statistic defined by
n
k
X
X
1
1{Xi ≤Xj } −
Wn := max |Wk,n | , where Wk,n :=
1≤k≤n−1
2
i=1
j=k+1
(see Dehling, Rooch and Taqqu (2013a)). Under the assumption that there is a
change point in the mean in k0 we expect the absolute value of Wk0 ,n to exceed
the absolute value of Wl,n for any l 6= k0 . Therefore, it seems natural to define
an estimator of k0 by
k̂W = k̂W (n) := min k : |Wk,n | = max |Wi,n | .
1≤i≤n−1
Preceding papers that address the problem of estimating change point locations
in dependent observations X1 , . . . , Xn with a shift in mean often refer to a
family of estimators based on the CUSUM change point test statistics Cn (γ) :=
max1≤k≤n−1 |Ck,n (γ)|, where
!
1−γ
k
n
X
k(n − k)
1X
1
Ck,n (γ) :=
Xi −
Xi
n
k i=1
n−k
i=k+1
A. Betken/Wilcoxon-based change point estimation
3
with parameter 0 ≤ γ < 1. The corresponding change point estimator is defined
by
k̂C,γ = k̂C,γ (n) := min k : |Ck,n (γ)| = max |Ci,n (γ)| .
(2)
1≤i≤n−1
For long-range dependent Gaussian processes Horváth and Kokoszka (1997) derive the asymptotic distribution of the estimator k̂C,γ under the assumption of
a decreasing jump height hn , i.e. under the assumption that hn approaches 0
as the sample size n increases. Under non-restrictive constraints on the dependence structure of the data-generating process (including long-range dependent
time series) Kokoszka and Leipus (1998) prove consistency of k̂C,γ under the
assumption of fixed as well as decreasing jump heights. Furthermore, they establish the convergence rate of the change point estimator as a function of the
intensity of dependence in the data if the jump height is constant. Ben Hariz
and Wylie (2005) show that under a similar assumption on the decay of the autocovariances the convergence rate that is achieved in the case of independent
observations can be obtained for short- and long-range dependent data, as well.
Furthermore, it is shown in their paper that for a decreasing jump height the
convergence rate derived by Horváth and Kokoszka (1997) under the assumption of gaussianity can also be established under more general assumptions on
the data-generating sequences.
Bai (1994) establishes an estimator for the location of a shift in the mean
by the method of least squares. He proves consistency, determines the rate of
convergence of the change point estimator and derives its asymptotic distribution. These results are shown to hold for weakly dependent observations that
satisfy a linear model and cover, for example, ARMA(p, q)-processes. Bai extended these results to the estimation of the location of a parameter change
in multiple regression models that also allow for lagged dependent variables
and trending regressors (see Bai (1997)). A generalization of these results to
possibly long-range dependent data-generating processes (including fractionally
integrated processes) is given in Kuan and Hsu (1998) and Lavielle and Moulines
(2000). Under the assumption of independent data Darkhovskh (1976) establishes an estimator for the location of a change in distribution based on the
two-sample Mann-Whitney test statistic. He obtains a convergence rate that has
order n1 , where n is the number of observations. Allowing for strong dependence
in the data Giraitis, Leipus and Surgailis (1996) consider Kolmogorov-Smirnov
and Cramér-von-Mises-type test statistics for the detection of a change in the
marginal distribution of the random variables that underlie the observed data.
Consistency of the corresponding change point estimators is proved under the
assumption that the jump height approaches 0. A change point estimator based
on a self-normalized CUSUM test statistic has been applied in Shao (2011) to
real data sets. Although Shao assumes validity of using the estimator, the article
does not cover a formal proof of consistency. Furthermore, it has been noted by
Shao and Zhang (2010) that even under the assumption of short-range dependence it seems difficult to obtain the asymptotic distribution of the estimate.
A. Betken/Wilcoxon-based change point estimation
4
In this paper we shortly address the issue of estimating the change point
location on the basis of the self-normalized Wilcoxon test statistic proposed in
Betken (2016).
In order to construct the self-normalized Wilcoxon test statistic, we have
to consider the ranks Ri , i = 1, . . .P
, n, of the observations X1 , . . . , Xn . These
n
are defined by Ri := rank(Xi ) =
j=1 1{Xj ≤Xi } for i = 1, . . . , n. The selfnormalized two-sample test statistic is defined by
Pk
SWk,n =
i=1
1
n
Pk
2
t=1 St (1, k)
Ri −
+
1
n
k
n
Pn
i=1
Pn
Ri
2
t=k+1 St (k
1/2 ,
+ 1, n)
where
St (j, k) :=
t
X
k
Rh − R̄j,k
with R̄j,k :=
h=j
X
1
Rt .
k − j + 1 t=j
The self-normalized Wilcoxon change point test for the test problem (H, A) rejects the hypothesis for large values of Tn (τ1 , τ2 ) = maxk∈{bnτ1 c,...,bnτ2 c} |SWk,n |,
where 0 < τ1 < τ2 < 1. Note that the proportion of the data that is included
in the calculation of the supremum is restricted by τ1 and τ2 . A common choice
for these parameters is τ1 = 1 − τ2 = 0.15; see Andrews (1993).
A natural change point estimator that results from the self-normalized
Wilcoxon test statistic is
k̂SW = k̂SW (n) := min k : |SWk,n | =
max
|SWi,n | .
bnτ1 c≤i≤bnτ2 c
We will prove consistency of the estimator k̂SW under fixed changes and under
local changes whose height converges to 0 with a rate depending on the intensity
of dependence in the data. Nonetheless, the main aim of this paper is to characterize the asymptotic behavior of the change point estimator k̂W . In Section
2 we establish consistency of k̂W and k̂SW , derive the optimal convergence rate
of k̂W and finally consider its asymptotic distribution. Applications to two wellknown data sets can be found in Section 3. The finite sample properties of the
estimators are investigated by simulations in Section 4. Proofs of the theoretical
results are given in Section 5.
2. Main Results
Recall that for fixed x, x ∈ R, the Hermite expansion of 1{G(ξi )≤x} − F (x) is
given by
1{G(ξi )≤x} − F (x) =
∞
X
Jq (x)
q=1
q!
Hq (ξi ),
A. Betken/Wilcoxon-based change point estimation
5
where Hq denotes the q-th order Hermite polynomial and where
Jq (x) = E 1{G(ξi )≤x} Hq (ξi ) .
Assumption 1. Let Yi = G(ξi ), where (ξi )i≥1 is a stationary, long-range dependent Gaussian process with mean 0, variance 1 and LRD parameter D. We
assume that 0 < D < 1r , where r denotes the Hermite rank of the class of
functions 1{G(ξi )≤x} − F (x), x ∈ R, defined by
r := min {q ≥ 1 : Jq (x) 6= 0 for some x ∈ R} .
Moreover, we assume that G : R −→ R is a measurable function and that (Yi )i≥1
has a continuous distribution function F .
Let
gD,r (t) := t
rD
2
r
L− 2 (t)
and define
dn,r
n
cr , where cr :=
:=
gD,r (n)
s
2r!
.
(1 − Dr)(2 − Dr)
−
Since gD,r is a regularly varying function, there exists a function gD,r
such that
−
−
gD,r (gD,r
(t)) ∼ gD,r
(gD,r (t)) ∼ t, as t → ∞,
−
(see Theorem 1.5.12 in Bingham, Goldie and Teugels (1987)). We refer to gD,r
as the asymptotic inverse of gD,r .
The following result states that k̂nW and k̂SW
are consistent estimators for
n
the change point location under fixed as well as certain local changes.
Proposition 1. Suppose that Assumption 1 holds. Under fixed changes,
k̂W
n
and
k̂SW
n
are consistent estimators for the change point
The estimators are
location.
n
−1
also consistent under local changes if hn = o dn,r and if F has a bounded
density f . In other words, we have
k̂W P
−→ τ,
n
k̂SW P
−→ τ
n
in both situations. Furthermore, it follows that the Wilcoxon test is consistent
P
under these assumptions (in the sense that nd1n,r max1≤k≤n−1 |Wk,n | −→ ∞).
The following theorem establishes a convergence rate for the change point
estimator k̂W . Note that only under local changes the convergence rate depends
on the intensity of dependence in the data.
A. Betken/Wilcoxon-based change point estimation
6
−
Theorem 1. Suppose that Assumption 1 holds and let mn := gD,r
(h−1
n ). Then,
we have
k̂W − k0 = OP (mn )
if either
• hn = h with h 6= 0
or
• limn→∞ hn = 0 with h−1
n =o
n
dn,r
and F has a bounded density f .
Remark 1.
1. Under fixed changes mn is constant. As a consequence, |k̂W −k0 | = OP (1).
This result corresponds to the convergence rates obtained by Ben Hariz
and Wylie (2005) for the CUSUM-test based change point estimator and
by Lavielle and Moulines (2000) for the least-squares estimate of the
change point location. Surprisingly, in this case the rate of convergence
is independent of the intensity of dependence in the data characterized by
the value of the LRD parameter D. An explanation for this phenomenon
might be the occurrence of two opposing effects: increasing values of the
LRD parameter D go along with a slower convergence of the test statistic Wk,n (making estimation more difficult), but a more regular behavior
of the random component (making estimation easier) (see Ben Hariz and
Wylie (2005)).
2. Note that if h−1
n =o
n
dn,r
−
and mn = gD,r
(h−1
n ), it holds that
• mn −→ ∞,
•
•
mn
n −→ 0,
dmn ,r
mn ∼ hn ,
as n −→ ∞.
Based on the previous results it is possible to derive the asymptotic distribution of the change point estimator k̂W :
Theorem 2. Suppose that Assumption 1 holds with r = 1 and assume that
−
F has a bounded density f . Let mn := gD,1
(h−1
n ), let BH denote a fractional
Brownian motion process and define h(s; τ ) by
(
R
s(1 − τ ) f 2 (x)dx if s ≤ 0
R
.
h(s; τ ) =
−sτ f 2 (x)dx
if s > 0
. If h−1
n =o
n
dn,1
, then, for all M > 0,
1 2
Wk0 +bmn sc,n − Wk20 ,n , −M ≤ s ≤ M,
en
A. Betken/Wilcoxon-based change point estimation
7
with en = n3 hn dmn ,1 , converges in distribution to
Z
Z
2τ (1 − τ ) f 2 (x)dx sign(s)BH (s) J1 (x)dF (x) + h(s; τ ) , −M ≤ s ≤ M,
in the Skorohod space D [−M, M ]. Furthermore, it follows that m−1
n (k̂W − k0 )
converges in distribution to
Z
argmax−∞<s<∞ sign(s)BH (s) J1 (x)dF (x) + h(s; τ ) .
(3)
Remark 2.
1. Under local changes the assumption on hn is equivalent to Assumption
C.5 (i) in Horváth and Kokoszka (1997). Moreover, the limit distribution
(3) closely resembles the limit distribution of the CUSUM-based change
point estimator considered in that paper.
2. The proof of Theorem 2 is mainly based on the empirical process noncentral limit theorem for subordinated Gaussian sequences in Dehling and
Taqqu (1989). The sequential empirical process has also been studied by
many other authors in the context of different models. See, among many
others, the following: Müller (1970) and Kiefer (1972) for independent
and identically distributed data, Berkes and Philipp (1977) and Philipp
and Pinzur (1980) for strongly mixing processes, Berkes, Hörmann and
Schauer (2009) for S-mixing processes, Giraitis and Surgailis (1999) for
long memory linear (or moving average) processes, Dehling, Durieu and
Tusche (2014) for multiple mixing processes. Presumably, in these situations the asymptotic distribution of k̂W can be derived by the same argument as in the proof of Theorem 2 for subordinated Gaussian processes. In
particular, Theorem 1 in Giraitis and Surgailis (1999) can be considered
as a generalization of Theorem 1.1 in Dehling and Taqqu (1989), i.e. with
an appropriate normalization the change point estimator k̂W , computed
with respect to long-range dependent linear processes as defined in Giraitis and Surgailis (1999), should converge in distribution to a limit that
corresponds to (3) (up to multiplicative constants).
3. Applications
We consider two well-known data sets which have been analyzed before. We
compute the estimator k̂W based on the given observations and put our results
into context with the findings and conclusions of other authors.
The plot in Figure 1 depicts the annual volume of discharge from the Nile river
at Aswan in 108 m3 for the years 1871 to 1970. The data set is included in any
standard distribution of R. Amongst others, Cobb (1978), Macneill, Tang and
Jandhyala (1991), Wu and Zhao (2007), Shao (2011) and Betken and Wendler
(2015) provide statistically significant evidence for a decrease of the Nile’s annual
discharge towards the end of the 19th century.
8
1000
800
600
annual discharge
1200
1400
A. Betken/Wilcoxon-based change point estimation
1880
1900
1920
1940
1960
time
Fig 1. Measurements of the annual discharge of the river Nile at Aswan in 108 m3 for the
years 1871-1970. The dotted line indicates the potential change point estimated by k̂W ; the
dashed lines designate the sample means for the pre-break and post-break samples.
The construction of the Aswan Low Dam between 1898 and 1902 serves
as a popular explanation for an abrupt change in the data around the turn
of the century. Yet, Cobb gave another explanation for the decrease in water
volume by citing rainfall records which suggest a decline of tropical rainfall at
that time. In fact, an application of the change point estimator k̂W identifies a
change in 1898. This result seems to be in good accordance with the estimated
change point locations suggested by other authors: Cobb’s analysis of the Nile
data leads to the conjecture of a significant decrease in discharge volume in
1898. Moreover, computation of the CUSUM-based change point estimator k̂C,0
considered in Horváth and Kokoszka (1997) indicates a change in 1898. Balke
(1993) and Wu and Zhao (2007) suggest that the change occurred in 1899.
The second data set consists of the seasonally adjusted monthly deviations of
the temperature (degrees C) for the Northern hemisphere during the years 1854
to 1989 from the monthly averages over the period 1950 to 1979. The data has
been taken from the longmemo package in R. It results from spatial averaging of
temperatures measured over land and sea. In view of the plot in Figure 2 it seems
natural to assume that the data generating process is non-stationary. Previous
analysis of this data offers different explanations for the irregular behavior of
the time series. Deo and Hurvich (1998) fitted a linear trend to the data, thereby
providing statistical evidence for global warming during the last decades. However, the consideration of a more general stochastic model by the assumption
of so-called semiparametric fractional autoregressive (SEMIFAR) processes in
Beran and Feng (2002) does not confirm the conjecture of a trend-like behavior.
Neither does the investigation of the global temperature data in Wang (2007)
support the hypothesis of an increasing trend. It is pointed out by Wang that
the trend-like behavior of the Northern hemisphere temperature data may have
−1.0
−0.5
0.0
0.5
9
−1.5
Northern hemisphere temperature
A. Betken/Wilcoxon-based change point estimation
1860
1880
1900
1920
1940
1960
1980
time
Fig 2. Monthly temperature of the Northern hemisphere for the years 1854-1989 from the data
base held at the Climate Research Unit of the University of East Anglia, Norwich, England.
The temperature anomalies (in degrees C) are calculated with respect to the reference period
1950-1979. The dotted line indicates the location of the potential change point; the dashed
lines designate the sample means for the pre-break and post-break samples.
been generated by stationary long-range dependent processes. Yet, it is shown
in Shao (2011) and also in Betken and Wendler (2015) that under model assumptions that include long-range dependence an application of change point
tests leads to a rejection of the hypothesis that the time series is stationary. According to Shao (2011) an estimation based on a self-normalized CUSUM test
statistic suggests a change around October 1924. Computation of the change
point estimator k̂W corresponds to a change point located around June 1924.
The same change point location results from an application of the previously
mentioned estimator k̂C,0 considered in Horváth and Kokoszka (1997). In this
regard estimation by k̂W seems to be in good accordance with the results of
alternative change point estimators.
4. Simulations
We will now investigate the finite sample performance of the change point estimator k̂W and compare it to corresponding simulation results for the estimators
k̂SW (based on the self-normalized Wilcoxon test statistic) and k̂C,0 (based on
the CUSUM test statistic with parameter γ = 0) . For this purpose, we consider
two different scenarios:
1. Normal margins: We generate fractional Gaussian noise time series (ξi )i≥1
and choose G(t) = t in Assumption 1. As a result, the simulated observa-
A. Betken/Wilcoxon-based change point estimation
10
tions (Yi )i≥1 are Gaussian with autocovariance function ρ satisfying
D
ρ(k) ∼ 1 −
(1 − D) k −D .
2
Note that in this case the Hermite coefficient J1 (x) is not equal to 0
for all x ∈ R (see Dehling, Rooch and Taqqu (2013a)) so that m = 1,
where m denotes the Hermite rank of 1{G(ξi )≤x} − F (x), x ∈ R. Therefore,
Assumption 1 holds for all values of D ∈ (0, 1).
2. Pareto margins: In order to get standardized Pareto-distributed data
which has a representation as a functional of a Gaussian process, we consider the transformation
− 21
1
βk
βk 2
−β
−
k(Φ(t))
G(t) =
(β − 1)2 (β − 2)
β−1
with parameters k, β > 0 and with Φ denoting the standard normal distribution function. Since G is a strictly decreasing function, it follows by
Theorem 2 in Dehling, Rooch and Taqqu (2013a) that the Hermite rank
of 1{G(ξi )≤x} − F (x), x ∈ R, is m = 1 so that Assumption 1 holds for all
values of D ∈ (0, 1).
To analyze the behavior of the estimators we simulated 500 time series of length
600 and added a level shift of height h after a proportion τ of the data. We have
done so for several choices of h and τ . The descriptive statistics, i.e. mean,
sample standard deviation (S.D.) and quartiles, are reported in Tables 1, 2, and
3 for the three change point estimators k̂W , k̂SW and k̂C,0 .
The following observations, made on the basis of Tables 1, 2, and 3, correspond to the expected behavior of consistent change point estimators:
• Bias and variance of the estimated change point location decrease when
the height of the level shift increases.
• Estimation of the time of change is more accurate for breakpoints located
in the middle of the sample than estimation of change point locations that
lie close to the boundary of the testing region.
• High values of H go along with an increase of bias and variance. This
seems natural since when there is very strong dependence, i.e. H is large,
the variance of the series increases, so that it becomes harder to accurately
estimate the location of a level shift.
A comparison of the descriptive statistics of the estimator k̂W (based on the
Wilcoxon statistic) and k̂SW (based on the self-normalized Wilcoxon statistic)
shows that:
• In most cases the estimator k̂SW has a smaller bias, especially for an early
change point location. Nevertheless, the difference between the biases of
k̂SW and k̂W is not big.
• In general the sample standard deviation of k̂W is smaller than that of
k̂SW . Indeed, it is only slightly better for τ = 0.25, but there is a clear
difference for τ = 0.5.
A. Betken/Wilcoxon-based change point estimation
11
All in all, our simulations do not give rise to choosing k̂SW over k̂W . In
particular, better standard deviations of k̂W compensate for smaller biases of
k̂SW .
Comparing the finite sample performance of k̂W and the CUSUM-based
change point estimator k̂C,0 we make the following observations:
• For fractional Gaussian noise time series bias and variance of k̂C,0 tend to
be slightly better, at least when τ = 0.25 and especially for relatively high
level shifts. Nonetheless, the deviations are in most cases negligible.
• If the change happens in the middle of a sample with normal margins,
bias and variance of k̂W tend to be smaller, especially for relatively high
level shifts. Again, in most cases the deviations are negligible.
• For Pareto(3, 1) time series k̂W clearly outperforms k̂C,0 by yielding
smaller biases and decisively smaller variances for almost every combination of parameters that has been considered. The performance of the
estimator k̂C,0 surpasses the performance of k̂W only for high values of
the jump height h.
400
100
200
300
H = 0.9
H = 0.8
H = 0.7
H = 0.6
0
mean absolute error
500
It is well-known that the Wilcoxon change point test is more robust against
outliers in data sets than the CUSUM-like change point tests, i.e. the Wilcoxon
test outperforms CUSUM-like tests if heavy-tailed time series are considered.
Our simulations confirm that this observation is also reflected by the finite
sample behavior of the corresponding change point estimators.
0
5000
10000
15000
20000
sample size
Fig 3. The MAE of k̂W for different values of H.
As noted in Remark 1, k̂W − k0 = OP (1) under the assumption of a constant
change point height h. This observation is illustrated by simulations of the mean
absolute error
m
1 X
k̂W,i − k0 ,
MAE =
m i=1
where k̂W,i , i = 1, . . . , m, denote the estimates for k0 , computed on the basis of
m = 5000 different sequences of fractional Gaussian noise time series.
A. Betken/Wilcoxon-based change point estimation
12
Figure 3 depicts a plot of MAE against the sample size n with n varying
between 1000 and 20000.
Since k̂W − k0 = OP (1) due to Theorem 1, we expect MAE to approach a
constant as n tends to infinity. This can be clearly seen in Figure 3 for H ∈
{0.6, 0.7, 0.8}. For a high intensity of dependence in the data (characterized by
H = 0.9) convergence becomes slower. This is due to a slower convergence of
the test statistic Wn (k) which, in finite samples, is not canceled out by the effect
of a more regular behavior of the sample paths of the limit process.
Table 1
Descriptive statistics of the sampling distribution of k̂W for a change in the mean based on 500 fractional Gaussian noise and Pareto time series
of length 600 with Hurst parameter H and a change in mean in τ of height h.
margins
normal
τ
0.25
h
0.5
2
0.5
0.5
1
2
Pareto(3, 1)
0.25
0.5
1
2
0.5
0.5
1
2
H = 0.6
193.840 (64.020)
(150, 168, 217.25)
164.244 (27.156)
(150, 153.5, 167)
153.604 (8.255)
(150, 151, 154)
H = 0.7
227.590 (99.788)
(150, 191, 284.25)
176.362 (42.059)
(150, 158, 190)
156.656 (12.393)
(150, 151, 158)
H = 0.8
252.408 (110.084)
(157, 226.5, 335.25)
188.328 (63.751)
(150, 159.5, 206.25)
164.338 (29.570)
(150, 151, 164)
H = 0.9
270.646 (113.720)
(172.75, 250, 353)
215.108 (88.621)
(150 176 256)
173.610 (41.514)
(150, 152, 180.25)
mean (S.D.)
quartiles
mean (S.D.)
quartiles
mean (S.D.)
quartiles
299.506 (30.586)
(291, 300, 309)
300.014 (9.141)
(298, 300, 302)
300.064 (1.294)
(300, 300, 300)
301.870 (61.392)
(274.75, 300.5, 320.25)
300.438 (18.695)
(297, 300, 304)
299.922 (3.215)
(300, 300, 300)
300.774 (82.610)
(264, 299, 339.25)
302.592 (42.213)
(293, 300 307)
299.504 (5.520)
(300, 300, 300)
298.930 (98.368)
(233, 299, 353)
300.902 (50.487)
(290, 300, 311)
300.282 (7.494)
(300, 300, 300)
mean (S.D.)
quartiles
mean (S.D.)
quartiles
mean (S.D.)
quartiles
158.166 (17.762)
(150, 151, 159.25)
154.160 (8.765)
(150, 151, 155)
152.256 (4.852)
(150, 150, 152)
164.080 (31.219)
(150, 152, 168)
156.090 (13.516)
(150, 151, 157)
155.592 (11.092)
(150, 151, 155.25)
179.512 (58.871)
(150, 154, 191.25)
164.712 (28.774)
(150, 152, 168)
160.686 (24.599)
(150, 151, 159)
194.126 (74.767)
(150, 159, 218.25)
178.174 (54.429)
(150, 152, 186)
169.374 (38.197)
(150, 150, 172)
mean (S.D.)
quartiles
mean (S.D.)
quartiles
mean (S.D.)
quartiles
298.072 (6.008)
(297, 300, 300)
299.178 (2.712)
(299, 300, 300)
299.798 (1.008)
(300, 300, 300)
296.432 (13.441)
(296, 300, 300)
298.744 (4.587)
(299, 300, 300)
299.716 (1.543)
(300, 300, 300)
293.060 (26.221)
(294, 300, 301)
296.674 (11.585)
(298, 300, 300)
299.384 (3.070)
(300, 300, 300)
289.946 (45.739)
(291, 300, 301)
296.168 (20.424)
(300, 300, 300)
298.896 (6.560)
(300, 300, 300)
A. Betken/Wilcoxon-based change point estimation
1
mean (S.D.)
quartiles
mean (S.D.)
quartiles
mean (S.D.)
quartiles
13
Table 2
Descriptive statistics of the sampling distribution of k̂SW for a change in the mean based on 500 replications of fractional Gaussian noise and
Pareto time series of length 600 with Hurst parameter H and a change in mean in τ of height h.
margins
normal
τ
0.25
h
0.5
2
0.5
0.5
1
2
Pareto(3, 1)
0.25
0.5
1
2
0.5
0.5
1
2
H = 0.6
172.288 (63.639)
(135, 153, 183.25)
152.406 (24.840)
(140, 149, 158)
148.836 (9.007)
(144, 150, 152)
H = 0.7
216.934 (110.934)
(138, 171, 272.5)
160.618 (39.834)
(139, 150.5, 172.25)
150.208 (13.575)
(142.75, 150, 154)
H = 0.8
242.202 (119.655)
(143, 207.5, 333.5)
174.424 (70.673)
(136, 150, 188.25)
153.194 (28.251)
(138, 150, 158)
H = 0.9
268.878 (122.615)
(157, 243.5, 370.25)
204.906 (99.648)
(139.75, 161.5, 243.75)
160.026 (40.979)
(137.75, 150, 165)
mean (S.D.)
quartiles
mean (S.D.)
quartiles
mean (S.D.)
quartiles
297.712 (43.291)
(277, 297, 320)
299.052 (16.132)
(290, 299, 308)
300.010 (6.054)
(297, 300, 303.25)
302.204 (77.719)
(262, 300, 337)
299.910 (28.907)
(288, 300, 313)
299.612 (10.079)
(294, 300, 305)
302.866 (96.511)
(248, 298.5, 369.5)
302.386 (55.267)
(277, 300, 324.25)
298.844 (14.059)
(291, 300, 307)
297.662 (110.175)
(215, 301, 369.5)
300.956 (62.821)
(270, 300, 329)
301.424 (21.022)
(289, 300, 312)
mean (S.D.)
quartiles
mean (S.D.)
quartiles
mean (S.D.)
quartiles
151.562 (18.392)
(142, 150, 157)
150.206 (9.116)
(145, 150, 154)
149.210 (6.201)
(146, 150, 152)
155.034 (32.505)
(140, 150, 163)
150.272 (15.405)
(143, 150, 156)
149.934 (11.821)
(143, 150, 153)
165.260 (58.363)
(136, 150, 173)
152.824 (25.074)
(140, 150, 159.25)
151.946 (21.426)
(140, 150, 156)
182.706 (83.268)
(136.75, 150, 196.25)
166.602 (58.982)
(136, 150, 174.25)
156.836 (39.311)
(136, 150, 160.25)
mean (S.D.)
quartiles
mean (S.D.)
quartiles
mean (S.D.)
quartiles
300.524 (11.841)
(294, 300, 307)
300.498 (6.600)
(297, 300, 304)
300.444 (4.411)
(298, 300, 303)
299.488 (21.317)
(290, 300, 310)
300.560 (10.383)
(296, 300, 306)
300.234 (7.517)
(296, 300, 304)
299.664 (37.136)
(287, 300, 317)
299.520 (18.862)
(292, 300, 309.25)
300.524 (11.122)
(295.75, 300, 307)
295.048 (55.000)
(280.75, 300, 318)
297.766 (28.308)
(289, 300, 312.25)
298.840 (16.004)
(292, 300, 308)
A. Betken/Wilcoxon-based change point estimation
1
mean (S.D.)
quartiles
mean (S.D.)
quartiles
mean (S.D.)
quartiles
14
Table 3
Descriptive statistics of the sampling distribution of k̂C,0 for a change in the mean based on 500 replications of fractional Gaussian noise and
Pareto time series of length 600 with Hurst parameter H and a change in mean in τ of height h.
margins
normal
τ
0.25
h
0.5
2
0.5
0.5
1
2
Pareto(3, 1)
0.25
0.5
1
2
0.5
0.5
1
2
H = 0.6
193.060 (64.917)
(150, 166.5, 222)
162.028 (22.948)
(150, 153, 164)
152.374 (6.249)
(150, 150, 152)
H = 0.7
228.948 (101.442)
(151, 191.5, 286.75)
173.838 (39.845)
(150, 156.5, 187.25)
154.878 (10.395)
(150, 150, 156)
H = 0.8
253.114 (111.182)
(156.75, 226, 341.5)
187.386 (63.865)
(150, 158, 206)
159.700 (22.064)
(150, 151, 158)
H = 0.9
271.380 (114.590)
(172.75, 249.5, 354.25)
213.114 (87.356)
(150, 173, 254.25)
165.940 (33.124)
(150, 150, 165)
mean (S.D.)
quartiles
mean(S.D.)
quartiles
mean (S.D.)
quartiles
297.840 (30.249)
(290, 299, 308)
299.870 (9.356)
(298, 300, 302)
300.060 (1.473)
(300, 300, 300)
302.060 (63.878)
(276, 301, 322)
299.662 (21.281)
(297, 300, 304)
299.916 (3.199)
(300, 300, 300)
300.246 (84.346)
(261.75, 300, 340)
303.646 (42.245)
(293, 300, 307)
299.442 (5.234)
(300, 300, 300)
298.910 (97.904)
(236.25, 299, 353.25)
299.762 (52.492)
(290, 300, 311)
300.460 (8.179)
(300, 300, 300)
mean (S.D.)
quartiles
mean (S.D.)
quartiles
mean (S.D.)
quartiles
175.632 (48.517)
(150, 159, 185)
156.586 (14.133)
(150, 152, 159)
150.314 (1.349)
(150, 150, 150)
198.452 (79.303)
(150, 168, 223.75)
160.350 (27.204)
(150, 152, 161)
150.566 (3.984)
(150, 150, 150)
205.506 (88.482)
(150, 173, 251.25)
170.278 (45.402)
(150, 153, 171)
152.474 (18.578)
(150, 150, 150)
210.444(93.831)
(150, 167, 259.5)
177.278 (66.661)
(150, 150, 174)
155.496 (29.408)
(150, 150, 150)
mean (S.D.)
quartiles
mean (S.D.)
quartiles
mean (S.D.)
quartiles
296.260 (22.306)
(292, 300, 303.25)
298.240 (6.104)
(299, 300, 300)
299.604 (1.843)
(300, 300, 300)
292.904 (43.471)
(288.75, 300, 305)
297.306 (9.361)
(299, 300, 300)
299.228 (3.385)
(300, 300, 300)
289.192 (64.033)
(273.75, 300, 308.25)
293.116 (26.614)
(298, 300, 300)
298.350 (8.354)
(300, 300, 300)
287.966 (64.827)
(285, 300, 303)
292.864 (37.601)
(300, 300, 300)
297.632 (14.525)
(300, 300, 300)
A. Betken/Wilcoxon-based change point estimation
1
mean (S.D.)
quartiles
mean (S.D.)
quartiles
mean (S.D.)
quartiles
15
A. Betken/Wilcoxon-based change point estimation
16
5. Proofs
In the following let Fk and Fk+1,n denote the empirical distribution functions
of the first k and last n − k realizations of Y1 , . . . , Yn , i.e.
Fk (x) :=
k
1X
1{Yi ≤x} ,
k i=1
Fk+1,n (x) :=
n
X
1
1{Yi ≤x} .
n−k
i=k+1
For notational convenience we write Wn (k) instead of Wk,n and SWn (k) instead
of SWk,n . The proofs in this section as well as the proofs in the appendix
are partially influenced by arguments that have been established in Horváth
and Kokoszka (1997), Bai (1994) and Dehling, Rooch and Taqqu (2013a). In
particular, some arguments are based on the empirical process non-central limit
theorem of Dehling and Taqqu (1989) which states that
D
d−1
n,r bnλc(Fbnλc (x) − F (x)) −→
1
(r)
Jr (x)ZH (λ),
r!
(r)
where r is the Hermite rank defined in Assumption 1, ZH is an r-th or
D
der Hermite process1 , H = 1 − rD
∈ 12 , 1 , and “−→” denotes conver2
gence in distribution with respect to the σ-field generated by the open balls
in D ([−∞, ∞] × [0, 1]), equipped with the supremum norm.
The Dudley-Wichura version of Skorohod’s representation theorem (see
Shorack and Wellner (1986), Theorem 2.3.4) implies that, for our purposes,
we may assume without loss of generality that
sup
λ∈[0,1],x∈R
1
(r)
d−1
Jr (x)ZH (λ) −→ 0
n,r bnλc Fbnλc (x) − F (x) −
r!
almost surely.
Proof of Proposition 1. The proof of Proposition 1 is based on an application
of Lemma 1 in the appendix. According to Lemma 1 it holds that, under the
assumptions of Proposition 1,
bnλc
1 X
n2 hn i=1
n
X
j=bnλc+1
1
1{Xi ≤Xj } −
2
P
−→ Cδτ (λ), 0 ≤ λ ≤ 1,
where δτ : [0, 1] −→ R is defined by
(
λ(1 − τ ) for λ ≤ τ
δτ (λ) =
(1 − λ)τ for λ ≥ τ
1 If r = 1, the Hermite process equals a standard fractional Brownian motion process with
Hurst parameter H = 1 − D
. We refer to Taqqu (1979) for a general definition of Hermite
2
processes.
A. Betken/Wilcoxon-based change point estimation
17
and C denotes some non-zero constant.
P
It directly follows that nd1n,r max1≤k≤n−1 |Wn (k)| −→ ∞.
Furthermore,
1
2
n hn
max
1≤k≤bn(τ −ε)c
k
n
X
X
1
1{Xi ≤Xj } −
2
i=1
j=k+1
converges in probability to
C
sup
δτ (λ) = C(τ − ε)(1 − τ )
0≤λ≤τ −ε
for any 0 ≤ ε < τ .
For ε > 0 define
1
n2 hn
Zn,ε :=
max
1≤k≤bnτ c
|Wn (k)| −
1
n2 hn
max
1≤k≤bn(τ −ε)c
|Wn (k)| .
P
As Zn,ε −→ C(1 − τ )ε, it follows that P (k̂W < bn(τ − ε)c) = P (Zn,ε = 0) −→ 0.
An analogous line of argument yields
P (k̂W > bn(τ + ε)c) −→ 0.
All in all, it follows that for any ε > 0
!
k̂W
− τ > ε = 0.
n
lim P
n−→∞
This proves consistency of the change point estimator which is based on the
Wilcoxon test statistic.
In the following it is shown that n1 k̂SW is a consistent estimator, too. For this
purpose, we consider the process SWn (bnλc), 0 ≤ λ ≤ 1. According to Betken
(2016) the limit of the self-normalized Wilcoxon test statistic can be obtained
by an application of the continuous mapping theorem to the process
bnλc
1 X
an i=1
n
X
j=bnλc+1
1
1{Xi ≤Xj } −
2
, 0 ≤ λ ≤ 1,
where an denotes an appropriate normalization. Therefore, it follows by the
corresponding argument in Betken (2016) that
P
SWn (bnλc) −→
Rλ
0
|δτ (λ)|
2
R1
δτ (t) − λt δτ (λ) dt + λ δτ (t) −
2
1−t
1−λ δτ (λ)
12
dt
A. Betken/Wilcoxon-based change point estimation
18
uniformly in λ ∈ [0, 1]. Elementary calculations yield
√
√
3λ 1 − λ
sup
SWn (k) −→ sup
,
(τ − λ)
τ1 ≤λ≤τ −ε
bnτ1 c≤k≤k0 −nε
√ √
3 λ(1 − λ)
P
.
sup
SWn (k) −→ sup
(τ − λ)
τ +ε≤λ≤τ2
k0 +nε≤k≤bnτ2 c
P
P
As SWn (k0 ) −→ ∞ due to Theorem 2 in Betken (2016), we conclude that
P (k̂SW > k0 + nε) and P (k̂SW < k0 − nε) converge to 0 in probability. This
P
proves n1 k̂SW −→ τ .
Proof of Theorem 1. In the following we write k̂ instead of k̂W . For convenience,
we assume that h > 0 under fixed changes, and that for some n0 ∈ N hn > 0
for all n ≥ n0 under local changes, respectively. Furthermore, we subsume both
changes under the general assumption that limn→∞ hn = h (under fixed changes
hn = h for all n ∈ N, under local changes h = 0). In order to prove Theorem 1,
we need to show that for all ε > 0 there exists an n(ε) ∈ N and an M > 0 such
that
P k̂ − k0 > M mn < ε
for all n ≥ n(ε).
For M ∈ R+ define Dn,M := {k ∈ {1, . . . , n − 1} | |k − k0 | > M mn }.
We have
!
P k̂ − k0 > M mn ≤ P
sup |Wn (k)| ≥ |Wn (k0 )| ≤ P1 + P2
k∈Dn,M
with
!
P1 := P
sup (Wn (k) − Wn (k0 )) ≥ 0 ,
k∈Dn,M
!
P2 := P
sup (−Wn (k) − Wn (k0 )) ≥ 0 .
k∈Dn,M
Note that Dn,M = Dn,M (1) ∪ Dn,M (2), where
Dn,M (1) := {k ∈ {1, . . . , n − 1} | k0 − k > M mn } ,
Dn,M (2) := {k ∈ {1, . . . , n − 1} | k − k0 > M mn } .
Therefore, P2 ≤ P2,1 + P2,2 , where
!
P2,1 := P
sup
(−Wn (k) − Wn (k0 )) ≥ 0 ,
k∈Dn,M (1)
!
P2,2 := P
sup
k∈Dn,M (2)
(−Wn (k) − Wn (k0 )) ≥ 0 .
A. Betken/Wilcoxon-based change point estimation
19
In the following we will consider the first summand only. (For the second summand analogous implications result from the same argument.)
For this, we define
cn (k) := δn (k)∆(hn ),
W
where
(
δn (k) :=
k(n − k0 ),
k0 (n − k),
k ≤ k0
k > k0
and
Z
(F (x + hn ) − F (x)) dF (x).
∆(hn ) :=
Note that
P2,1 ≤ P
sup
!
cn (k) − Wn (k) + W
cn (k0 ) − Wn (k0 ) ≥ W
cn (k0 )
W
k∈Dn,M (1)
≤P
!
c
Wn (bnλc) − Wn (bnλc) ≥ k0 (n − k0 )∆(hn ) .
2 sup
λ∈[0,τ ]
We have
cn (bnλc)
Wn (bnλc) − W
sup
λ∈[0,τ ]
bnλc
n
X
X
= sup
Z
1{Yi ≤Yj +hn } − F (x + hn )dF (x)
λ∈[0,τ ] i=1
j=bnτ c+1
bnλc
bnτ c
X
X
+
1{Yi ≤Yj } −
i=1 j=bnλc+1
1
2
.
Due to Lemma 2 in the appendix and Theorem 1.1 in Dehling, Rooch and Taqqu
(2013a)
2 sup
cn (bnλc) = OP (ndn,r ) ,
Wn (bnλc) − W
λ∈[0,τ ]
i.e. for all ε > 0 there exists a K > 0 such that
!
P
2 sup
cn (bnλc) ≥ Kndn,r
Wn (bnλc) − W
<ε
λ∈[0,τ ]
for all n. Furthermore, k0 (n − k0 )∆(hn ) ∼ Cn2 hn for some constant C. Note
that Kndn,r ≤ k0 (n − k0 )∆(hn ) if and only if
K≤
k0 n − k0 ∆(hn ) nhn
.
n
n
hn dn,r
A. Betken/Wilcoxon-based change point estimation
20
The right
hand
side of the above inequality diverges if hn = h is fixed or if
n
−1
hn = o dn,r . Therefore, it is possible to find an n(ε) ∈ N such that
!
P2,1 ≤ P
2 sup
cn (bnλc) ≥ k0 (n − k0 )∆(hn )
Wn (bnλc) − W
λ∈[0,τ ]
!
≤P
2 sup
cn (bnλc) ≥ Kndn,r
Wn (bnλc) − W
λ∈[0,τ ]
<ε
for all n ≥ n(ε).
We will now turn to the summand P1 . We have P1 ≤ P1,1 + P1,2 , where
!
P1,1 := P
sup
Wn (k) − Wn (k0 ) ≥ 0 ,
k∈Dn,M (1)
!
P1,2 := P
sup
Wn (k) − Wn (k0 ) ≥ 0 .
k∈Dn,M (2)
In the following we will consider the first summand only. (For the second summand analogous implications result from the same argument.)
We define a random sequence kn , n ∈ N, by choosing kn ∈ Dn,M (1) such
that
cn (k) + W
cn (k0 ) − Wn (k0 )
Wn (k) − W
sup
k∈Dn,M (1)
cn (kn ) + W
cn (k0 ) − Wn (k0 ).
= Wn (kn ) − W
Note that for any sequence kn , n ∈ N, with kn ∈ Dn,M (1)
cn (k0 ) − W
cn (kn ) = (n − k0 )ln ∆(hn )
W
where ln := k0 − kn . Since kn ∈ Dn,M (1) and mn −→ ∞ we have
ln
dln ,r
r
r
= ln1−H L− 2 (ln ) ≥ (M mn )1−H L− 2 (M mn )
for n sufficiently large. Thus, we have
r
n−k m
2
1 c
0
n
1−H L (mn )
c
Wn (k0 ) − Wn (kn ) ≥
M
∆(hn ).
r
ndln ,r
n dmn ,r
L 2 (M mn )
If hn is fixed, the right hand side of the inequality diverges. Under local changes
the right hand side asymptotically behaves like
Z
1−H
(1 − τ )M
f 2 (x)dx,
A. Betken/Wilcoxon-based change point estimation
21
d
n ,r
since, in this case, hn ∼ m
mn due to the assumptions of Theorem 1.
In any case, for δ > 0 it is possible to find an n0 ∈ N such that
Z
1 c
cn (kn ) ≥ M 1−H (1 − τ ) f 2 (x)dx − δ
Wn (k0 ) − W
ndln ,r
for all n ≥ n0 .
All in all, the previous considerations show that there exists an n0 ∈ N and
a constant K such that for all n ≥ n0
!
1
cn (k) + W
cn (k0 ) − Wn (k0 ) ≥ b(M )
P1,1 ≤ P
sup
Wn (k) − W
k∈Dn,M (1) ndk0 −k,r
where b(M ) := KM 1−H − δ with δ > 0 fixed.
Some elementary calculations show that for k ≤ k0
cn (k) + W
cn (k0 ) − Wn (k0 ) = An,1 (k) + An,2 (k) + An,3 (k) + An,4 (k),
Wn (k) − W
where
Z
An,1 (k) := −(n − k0 )(k0 − k)
(Fk,k0 (x + hn ) − F (x + hn )) dFk0 ,n (x),
Z
An,2 (k) := −(n − k0 )(k0 − k) (Fk0 ,n (x) − F (x)) dF (x + hn ),
Z
An,3 (k) := (k0 − k)k (Fk (x) − F (x)) dFk,k0 (x),
Z
An,4 (k) := −k(k0 − k) (Fk,k0 (x) − F (x)) dF (x).
Thus, for n ≥ n0
P1,1 ≤ P
sup
k∈Dn,M (1)
≤
4
X
P
1
4
X
ndk0 −k,r
i=1
1
sup
k∈Dn,M (1)
i=1
ndk0 −k,r
!
|An,i (k)| ≥ b(M )
!
1
|An,i (k)| ≥ b(M ) .
4
For each i ∈ {1, . . . , 4} it will be shown that
P
1
sup
k∈Dn,M (1)
ndk0 −k,r
!
1
ε
|An,i (k)| ≥ b(M ) <
4
4
for n and M sufficiently large.
1. Note that
sup
k∈Dn,M (1)
≤
sup
1
|An,1 (k)|
ndk0 −k,r
sup d−1
k0 −k,r (k0 − k) (Fk,k0 (x) − F (x)) .
k∈Dn,M (1) x∈R
A. Betken/Wilcoxon-based change point estimation
22
Due to stationarity
sup d−1
k0 −k,r (k0 − k) (Fk,k0 (x) − F (x))
sup
k∈Dn,M (1) x∈R
D
=
sup d−1
k0 −k,r (k0 − k) (Fk0 −k (x) − F (x)) .
sup
k∈Dn,M (1) x∈R
Note that
sup d−1
k0 −k,r (k0 − k) (Fk0 −k (x) − F (x))
sup
k∈Dn,M (1) x∈R
≤
sup d−1
k0 −k,r (k0 − k) (Fk0 −k (x) − F (x)) −
sup
k∈Dn,M (1) x∈R
+
1 (r)
Z (1)Jr (x)
r! H
1
(r)
Z (1) sup |Jr (x)| .
r! H
x∈R
Since
sup d−1
n,r n (Fn (x) − F (x)) −
x∈R
1 (r)
Z (1)Jr (x) −→ 0 a.s.
r! H
if n −→ ∞, and as k0 − k ≥ M mn with mn −→ ∞, it follows that
sup
sup d−1
k0 −k,r (k0 − k) (Fk0 −k (x) − F (x)) −
k∈Dn,M (1) x∈R
1 (r)
Z (1)Jr (x)
r! H
converges to 0 almost surely. Therefore,
P
1
sup
k∈Dn,M (1)
ndk0 −k,r
1
|An,1 (k)| ≥ b(M )
4
!
1
≤P
sup sup
− k) (Fk,k0 (x) − F (x)) ≥ b(M )
4
x∈R
k∈Dn,M (1)
1
1
ε
(r)
≤P
ZH (1) sup |Jr (x)| ≥ b(M ) + .
r!
4
8
x∈R
!
d−1
k0 −k,r (k0
for n sufficiently large. Note that supx∈R |Jr (x)| < ∞. Furthermore, it is
well-known that all moments of Hermite processes are finite. As a result,
it follows by Markov’s inequality that for some M0 ∈ R
1
1
4r!
ε
(r)
(r)
P
ZH (1) sup |Jr (x)| ≥ b(M ) ≤ E ZH (1)
<
r!
4
sup |Jr (x)| b(M )
8
x∈R
x∈R
for all M ≥ M0 .
2. We have
1
|An,2 (k)|
ndk0 −k,r
Z
−1
≤ dn,r (n − k0 ) (Fk0 ,n (x) − F (x)) dF (x + hn )
sup
k∈Dn,M (1)
A. Betken/Wilcoxon-based change point estimation
23
for n sufficiently large. As a result,
1
sup
k∈Dn,M (1)
ndk0 −k,r
|An,2 (k)| ≤ sup d−1
n,r (n − k0 ) (Fk0 ,n (x) − F (x)) .
x∈R
Due to the empirical process non-central limit theorem of Dehling and
Taqqu (1989) we have
D
sup d−1
n,r (n − k0 ) (Fk0 ,n (x) − F (x)) −→
x∈R
1
(r)
(r)
Z (1) − ZH (τ )) sup |Jr (x)| .
r! H
x∈R
Moreover,
1
D 1
(r)
(r)
(r)
Z (1) − ZH (τ ) sup |Jr (x)| = (1 − τ )H ZH (1) sup |Jr (x)|
r! H
r!
x∈R
x∈R
(r)
since ZH is a H-self-similar process with stationary increments. Thus, we
have
!
1
1
P
sup
|An,2 (k)| ≥ b(M )
4
k∈Dn,M (1) ndk0 −k,r
1
ε
1
(r)
H
(1 − τ ) ZH (1) sup |Jr (x)| ≥ b(M ) +
≤P
r!
4
8
x∈R
for n sufficiently large. Again, it follows by Markov’s inequality that
1
1
ε
(r)
P
(1 − τ )H ZH (1) sup |Jr (x)| ≥ b(M ) <
r!
4
8
x∈R
for M sufficiently large.
3. Note that
1
ndk0 −k,r
|An,3 (k)| ≤ d−1
n,r k
Z
(Fk (x) − F (x)) dFk,k0 (x)
for n sufficiently large. Therefore,
sup
k∈Dn,M (1)
1
ndk0 −k,r
|An,3 (k)| ≤
sup
x∈R,0≤λ≤1
d−1
n,r bnλc Fbnλc (x) − F (x) .
The expression on the right hand side of the inequality converges in distribution to
1
(r)
sup Z (λ) sup |Jr (x)|
r! 0≤λ≤1 H
x∈R
due to the empirical process non-central limit theorem. Since
n
o n
o
D
(r)
(r)
ZH (λ), 0 ≤ λ ≤ 1 = λH ZH (1), 0 ≤ λ ≤ 1 ,
A. Betken/Wilcoxon-based change point estimation
24
we have
sup
0≤λ≤1
(r)
D
(r)
ZH (λ) = |ZH (1)|.
As a result, the aforementioned argument yields
!
1
P
sup
|An,3 (k)| ≥ b(M )
4
k∈Dn,M (1) ndk0 −k,r
1
1
ε
(r)
ZH (1) sup |Jr (x)| ≥ b(M ) +
≤P
r!
4
8
x∈R
ε
<
4
1
for n and M sufficiently large.
4. We have
sup
k∈Dn,M (1)
≤
sup
1
ndk0 −k,r
|An,4 (k)|
sup d−1
k0 −k,r (k0 − k) (Fk,k0 (x) − F (x)) .
k∈Dn,M (1) x∈R
Hence, the same argument that has been used to obtain an analogous
result for An,1 can be applied to conclude that
!
ε
1
1
P
sup
|An,4 (k)| ≥ b(M ) <
4
4
k∈Dn,M (1) ndk0 −k,r
for n and M sufficiently large.
All in all, it follows that for all ε > 0 there exists an n(ε) ∈ N and an M > 0
such that
P k̂ − k0 > M mn < ε
for all n ≥ n(ε). This proves Theorem 1.
Proof of Theorem 2. Note that
Wn2 (k0 + bmn sc) − Wn2 (k0 )
= (Wn (k0 + bmn sc) − Wn (k0 )) (Wn (k0 + bmn sc) + Wn (k0 )) .
We will show that (with an appropriate normalization) Wn (k0 +bmn sc)−Wn (k0 )
converges in distribution to a non-deterministic limit process whereas Wn (k0 +
bmn sc) + Wn (k0 ) (with stronger normalization) converges in probability to a
deterministic expression. For notational convenience we write dmn instead of
A. Betken/Wilcoxon-based change point estimation
25
dmn ,1 , J instead of J1 , k̂ instead of k̂W and we define ln (s) := k0 + bmn sc. We
have
Wn (k0 + bmn sc) − Wn (k0 ) = Ṽn (ln (s)) + Vn (ln (s)),
where
Ṽn (l) =
k0
P
−
−
n
P
1{Yi ≤Yj +hn } − 1{Yi ≤Yj }
i=l+1 j=k0 +1
k0
l
P
P
1{Yi ≤Yj +hn } − 1{Yi ≤Yj }
if s < 0
if s > 0
i=1 j=k0 +1
and
Vn (l) =
k0
l
P
P
1{Yi ≤Yj } −
i=1 j=l+1
l
n
P
P
1
2
1{Yi ≤Yj } −
i=k0 +1 j=l+1
−
1
2
k0
P
n
P
i=l+1 j=k0 +1
k0
l
P
P
−
i=1 j=k0 +1
1
2
1{Yi ≤Yj } −
1
2
1{Yi ≤Yj } −
if s < 0
.
if s > 0
We will show that nd1m Ṽn (ln (s)) converges to h(s; τ ) in probability and
n
R
that nd1m Vn (ln (s)) converges in distribution to sign(s)BH (s) J(x)dF (x) in
n
D [−M, M ].
We rewrite Ṽn (ln (s)) in the following way:
Z
Ṽn (ln (s)) = −(k0 − ln (s))(n − k0 )
Fln (s),k0 (x + hn ) − Fln (s),k0 (x) dFk0 ,n (x)
if s < 0,
Z
Ṽn (ln (s)) = −k0 (ln (s) − k0 )
(Fk0 (x + hn ) − Fk0 (x)) dFk0 ,ln (s) (x)
if s > 0.
For s < 0 the limit of
Ṽn (ln (s)) corresponds to the limit of
1
ndmn
−(1 − τ )d−1
mn (k0 − ln (s))
Z
(F (x + hn ) − F (x)) dF (x)
due to Lemma 3 and stationarity of the random sequence Yi , i ≥ 1. Note that
Z
d−1
(k
−
l
(s))
(F (x + hn ) − F (x)) dF (x)
n
mn 0
Z
1
= −d−1
(F (x + hn ) − F (x)) dF (x).
mn bmn schn
hn
R
d
The above expression converges to −s f 2 (x)dx, since hn ∼ mmnn .
A. Betken/Wilcoxon-based change point estimation
For s > 0 the limit of
1
ndmn
26
Ṽn (ln (s)) corresponds to the limit of
−τ d−1
mn (ln (s) − k0 )
Z
(F (x + hn ) − F (x)) dF (x)
due to Lemma 3 and stationarity of the random sequence Yi , i ≥ 1. Note that
Z
d−1
(l
(s)
−
k
)
(F (x + hn ) − F (x)) dF (x)
0
mn n
Z
1
= d−1
(F (x + hn ) − F (x)) dF (x)
mn bmn schn
hn
R
d
The above expression converges to s f 2 (x)dx, since hn ∼ mmnn .
All in all, it follows that nd1m Ṽn (ln (s)) converges to h(s; τ ) defined by
n
(
h(s; τ ) =
R
s(1 − τ ) f 2 (x)dx
R
−sτ f 2 (x)dx
In the following it is shown that
1
ndmn
if s ≤ 0
.
if s > 0
Vn (ln (s)) converges in distribution to
Z
sign(s)BH (s)
J(x)dF (x), −M ≤ s ≤ M.
Note that if s < 0,
Z
Fln (s),k0 (x) − F (x) dFln (s) (x)
Z
− (k0 − ln (s))(n − k0 )
Fln (s),k0 (x) − F (x) dFk0 ,n (x)
Z
+ ln (s)(k0 − ln (s)) (Fln (s) (x) − F (x))dF (x)
Z
+ (k0 − ln (s))(n − k0 ) (Fk0 ,n (x) − F (x)) dF (x).
Vn (ln (s)) = − ln (s)(k0 − ln (s))
If s > 0, we have
Z
Vn (ln (s)) =(ln (s) − k0 )(n − ln (s))
Fk0 ,ln (s) (x) − F (x) dFln (s),n (x)
Z
+ k0 (ln (s) − k0 )
Fk0 ,ln (s) (x) − F (x) (x)dFk0 (x)
Z
− (ln (s) − k0 )(n − ln (s))
Fln (s),n (x) − F (x) dF (x)
Z
− k0 (ln (s) − k0 ) (Fk0 (x) − F (x)) dF (x).
A. Betken/Wilcoxon-based change point estimation
27
The arguments that appear in the proof of Lemma 3 can also be applied to show
that the limit of nd1m Vn (ln (s)) corresponds to the limit of
n
1
(A1,n (s) + A2,n (s) + A3,n (s)) ,
ndmn
where
Z
Fln (s),k0 (x) − F (x) dF (x)
A1,n (s) := (−ln (s) − n + k0 )(k0 − ln (s))
if s < 0,
Z
A1,n (s) := (n − ln (s) + k0 )(ln (s) − k0 )
Fk0 ,ln (s) (x) − F (x) dF (x)
if s > 0,
(
R
(k0 − ln (s))ln (s) (Fln (s) (x) − F (x))dF (x)
if s < 0
R
A2,n (s) :=
,
−(ln (s) − k0 )(n − ln (s)) Fln (s),n (x) − F (x) dF (x) if s > 0
(
R
(k0 − ln (s))(n − k0 ) (Fk0 ,n (x) − F (x)) dF (x) if s < 0
R
A3,n (s) :=
.
−(ln (s) − k0 )k0 (Fk0 (x) − F (x)) dF (x)
if s > 0
Note that for s < 0
1
1
A2,n (s) = −
bmn scln (s)
ndmn
ndmn
Z
(Fln (s) (x) − F (x))dF (x).
The above expression converges to 0 uniformly in s, since
since
Z
l
(s)
(Fln (s) (x) − F (x))dF (x)
sup d−1
n
n
mn
dmn
= o( dnn ) and
−M ≤s≤0
≤ sup d−1
n bnλc(Fbnλc (x) − F (x)) − BH (λ)J(x)
x,λ
Z
+ sup |BH (λ)|
J(x)dF (x) ,
0≤λ≤1
R
i.e. sup−M ≤s≤0 d−1
n ln (s) (Fln (s) (x) − F (x))dF (x) is bounded in probability.
An analogous argument shows that nd1m A3,n (s) vanishes if n tends to ∞.
n
Therefore, it remains to show that nd1m A1,n (s) converges in distribution to
n
a non-deterministic expression. Due to stationarity
Z
1
D n + bmn sc −1
A1,n (s) =
dmn bmn sc
F−bmn sc (x) − F (x) dF (x)
ndmn
n
A. Betken/Wilcoxon-based change point estimation
28
for s < 0. As a result, nd1m A1,n (s) converges in distribution to
n
R
−BH (s) J(x)dF (x).
If s > 0, an application of the previous arguments shows that nd1m A2,n (s)
n
and nd1m A3,n (s) converge to 0 whereas nd1m A1,n (s) converges in distribution
nR
n
to BH (s) J(x)dF (x).
All in all, it follows that
Z
1
D
(Wn (k0 + bmn sc) − Wn (k0 )) −→ sign(s)BH (s) J(x)dF (x) + h(s; τ )
ndmn
in D[−M, M ].
Furthermore, it follows that with the stronger normalization hn n2 the limit
of hn1n2 Wn (k0 + bmn sc) corresponds to the limit of hn1n2 Wn (k0 ).
We have
Z
1
1
Wn (k0 ) =
k0 (n − k0 ) (Fk0 (x + hn ) − Fk0 (x)) dFk0 ,n (x)
hn n2
hn n2
k0
n
X
1 X
1
+
1
−
.
{Yi ≤Yj }
hn n2 i=1
2
j=k0 +1
The second summand on the right hand side vanishes as n
= o (n/dn ). Due to Lemma 3 the limit of
tends Rto ∞, since h−1
n
k
(F
(x
+
h
)
−
F
(x))
dFk0 ,n (x) corresponds to the limit of
d−1
0
k
n
k
n
0
0
R
d−1
k
(F
(x
+
h
)
−
F
(x))
dF
(x).
Therefore,
0
n
n
Z
Z
h−1
(F
(x
+
h
)
−
F
(x))
dF
(x)
−→
f 2 (x)dx a.s.
k0
n
k0
k0 ,n
n
0)
In addition, kn0 (n−k
−→ τ (1 − τ ).
n
From this we can conclude that
1
P
(Wn (k0 + mn s) + Wn (k0 )) −→ 2τ (1 − τ )
hn n2
Z
f 2 (x)dx
in D[−M, M ]. This completes the proof of the first assertion in Theorem 2.
In order to show that
Z
D
m−1
(
k̂
−
k
)
−→
argmax
sign(s)B
(s)
J(x)dF
(x)
+
h(s;
τ
)
,
0
H
−∞<s<∞
n
we make use of Lemma 4.
For this purpose, we note that according to Lifshits’ criterion for unimodality of Gaussian processes (see RTheorem 1.1 in Ferger (1999)) the random function GH,τ (s) = sign(s)BH (s) J(x)dF (x) + h(s; τ ) attains its maximal value
in [−M, M ] at a unique point with probability 1 for every M > 0. Hence, an
application of Lemma 4 in the appendix yields
sargmaxs∈[−M,M ]
D
1
Wn2 (k0 + bmn sc) − Wn2 (k0 ) −→ argmaxs∈[−M,M ] GH,τ (s).
en
A. Betken/Wilcoxon-based change point estimation
29
It remains to be shown that instead of considering the sargmax in [−M, M ]
we may as well consider the smallest argmax in R. By the law of the iterated
logarithm for fractional
Brownian motions we have lim|s|→∞ BHs(s) = 0 a.s. so
R
that sign(s)BH (s) J(x)dF (x) + h(s; τ ) −→ −∞ a.s. if |s| → ∞. Therefore, the
limit corresponds to argmaxs∈(−∞,∞) GH,τ (s) if M is sufficiently large.
For M > 0 define
ˆ
|Wn (i)| .
k̂(M ) := min k : |k0 − k| ≤ M mn , |Wn (k)| =
max
|k0 −i|≤M mn
Note that
sargmaxs∈[−M,M ] Wn2 (k0 + bmn sc) − Wn2 (k0 )
− sargmaxs∈(−∞,∞) Wn2 (k0 + bmn sc) − Wn2 (k0 )
ˆ
= m−1
k̂(M ) − k̂ + OP (1).
n
Therefore, we have to show that for some M ∈ R
P
ˆ
m−1
k̂(M ) − k̂ −→ 0
n
as n tends to infinity. Note that
ˆ
P k̂ = k̂(M ) = P k̂ − k0 ≤ M mn
= 1 − P k̂ − k0 > M mn .
Furthermore, we have
lim lim inf 1 − P |k̂ − k0 | > M mn
M →∞ n→∞
= 1 − lim lim sup P |k̂ − k0 | > M mn
M →∞ n→∞
=1
because |k̂ − k0 | = OP (mn ) by Theorem 1. As a result, we have
ˆ
lim lim inf P k̂ = k̂(M ) = 1.
M →∞ n→∞
Hence, for all ε > 0 there is an M0 ∈ R and an n0 ∈ N such that
ˆ
P k̂ 6= k̂(M ) < ε
for all n ≥ n0 and all M ≥ M0 . This concludes the proof of Theorem 2.
A. Betken/Wilcoxon-based change point estimation
30
References
Andrews, D. W. K. (1993). Tests for parameter instability and structural
change with unknown change point. Econometrica 61 821 – 856.
Bai, J. (1994). Least squares estimation of a shift in linear processes. Journal
of Time Series Analysis 15 453 – 472.
Bai, J. (1997). Estimation of a change point in multiple regression models.
Review of Economics and Statistics 79 551 – 563.
Balke, N. S. (1993). Detecting level shifts in time series. Journal of Business
& Economic Statistics 11 81 – 92.
Ben Hariz, S. and Wylie, J. J. (2005). Convergence rates for estimating a
change-point with long-range dependent sequences. Comptes Rendus Mathematique 341 765 – 768.
Beran, J. and Feng, Y. (2002). SEMIFAR models - a semiparametric framework for modelling trends, long-range dependence and nonstationarity. Computational Statistics & Data Analysis 40 393 – 419.
Berkes, I., Hörmann, S. and Schauer, J. (2009). Asymptotic results for
the empirical process of stationary sequences. Stochastic Processes and their
Applications 119 1298 – 1324.
Berkes, I. and Philipp, W. (1977). An almost sure invariance principle for
the empirical distribution function of mixing random variables. Probability
Theory and Related Fields 41 115 – 137.
Betken, A. (2016). Testing for change-points in long-range dependent time
series by means of a self-normalized Wilcoxon test. Journal of Time Series
Analysis 20 785 – 809.
Betken, A. and Wendler, M. (2015). Subsampling for General Statistics
under Long Range Dependence. arXiv:1509.05720.
Bingham, N. H., Goldie, C. M. and Teugels, J. L. (1987). Regular variation. Cambridge University Press.
Cobb, G. W. (1978). The problem of the Nile: Conditional solution to a changepoint problem. Biometrika 65 243 – 251.
Darkhovskh, B. S. (1976). A nonparametric method for the a posteriori detection of the “disorder”time of a sequence of independent random variables.
Theory of Probability & Its Applications 21 178 – 183.
Dehling, H., Durieu, O. and Tusche, M. (2014). Approximating class approach for empirical processes of dependent sequences indexed by functions.
Bernoulli 20 1372 – 1403.
Dehling, H., Rooch, A. and Taqqu, M. S. (2013a). Non-parametric changepoint tests for long-range dependent data. Scandinavian Journal of Statistics
40 153 – 173.
Dehling, H., Rooch, A. and Taqqu, M. S. (2013b). Power of change-point
tests for long-range dependent data. arXiv:1303.4917.
Dehling, H. and Taqqu, M. S. (1989). The empirical process of some longrange dependent sequences with an application to U-statistics. The Annals of
Statistics 17 1767 – 1783.
A. Betken/Wilcoxon-based change point estimation
31
Deo, R. S. and Hurvich, C. M. (1998). Linear trend with fractionally integrated errors. Journal of Time Series Analysis 19 379 – 397.
Ferger, D. (1999). On the uniqueness of maximizers of Markov–Gaussian processes. Statistics & Probability Letters 45 71 – 77.
Giraitis, L., Leipus, R. and Surgailis, D. (1996). The change-point problem
for dependent observations. Journal of Statistical Planning and Inference 53
297 – 310.
Giraitis, L. and Surgailis, D. (1999). Central limit theorem for the empirical
process of a linear sequence with long memory. Journal of Statistical Planning
and Inference 80 81 – 93.
Horváth, L. and Kokoszka, P. (1997). The effect of long-range dependence
on change-point estimators. Journal of Statistical Planning and Inference 64
57 – 81.
Kiefer, J. (1972). Skorohod embedding of multivariate RV’s, and the sample
DF. Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete 24 1 –
35.
Kokoszka, P. and Leipus, R. (1998). Change-point in the mean of dependent
observations. Statistics & Probability Letters 40 385 – 393.
Kuan, C. M. and Hsu, C. C. (1998). Change-point estimation of fractionally
integrated processes. Journal of Time Series Analysis 19 693 – 708.
Lavielle, M. and Moulines, E. (2000). Least-squares estimation of an unknown number of shifts in a time series. Journal of Time Series Analysis 21
33 – 59.
Macneill, I. B., Tang, S. M. and Jandhyala, V. K. (1991). A Search for
the Source of the Nile’s Change-Points. Environmetrics 2 341 – 375.
Müller, D. W. (1970). On Glivenko-Cantelli convergence. Probability Theory
and Related Fields 16 195 – 210.
Philipp, W. and Pinzur, L. (1980). Almost sure approximation theorems for
the multivariate empirical process. Zeitschrift für Wahrscheinlichkeitstheorie
und verwandte Gebiete 54 1 – 13.
Seijo, E., Sen, B. et al. (2011). A continuous mapping theorem for the smallest
argmax functional. Electronic Journal of Statistics 5 421 – 439.
Shao, X. (2011). A simple test of changes in mean in the possible presence of
long-range dependence. Journal of Time Series Analysis 32 598 – 606.
Shao, X. and Zhang, X. (2010). Testing for change points in time series.
Journal of the American Statistical Association 105 1228 – 1240.
Shorack, G. R. and Wellner, J. A. (1986). Empirical processes with applications to statistics. John Wiley & Sons, New York.
Taqqu, M. S. (1979). Convergence of integrated processes of arbitrary Hermite
rank. Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete 50 53
– 83.
Wang, L. (2007). Gradual changes in long memory processes with applications.
Statistics 41 221 – 240.
Wu, W. B. and Zhao, Z. (2007). Inference of trends in time series. Journal
of the Royal Statistical Society: Series B (Statistical Methodology) 69 391 –
410.
A. Betken/Wilcoxon-based change point estimation
32
Appendix A: Auxiliary Results
In the following we prove some Lemmas that are needed for the proofs of our
main results. Lemma 1 characterizes the asymptotic behavior of the Wilcoxon
process under the assumption of a change-point in the mean. It is used to prove
consistency of the change-point estimators k̂W and k̂SW .
Lemma 1. Define δτ : [0, 1] −→ R by
(
λ(1 − τ )
δτ (λ) =
(1 − λ)τ
for λ ≤ τ
.
for λ ≥ τ
Assume that Assumption 1 holds and that either
a) hn = h with h 6= 0,
or
b) limn→∞ hn = 0 with h−1
n =o
n
dn,r
and F has a bounded density f .
Then, we have
bnλc
1 X
n2 hn i=1
n
X
j=bnλc+1
1
1{Xi ≤Xj } −
2
P
−→ Cδτ (λ), 0 ≤ λ ≤ 1,
where
(1 R
Rh
C :=
(F (x + h) − F (x)) dF (x)
if hn = h, h 6= 0,
if limn→∞ hn = 0 and h−1
n =o
2
f (x)dx
n
dn,r
.
Proof. First, consider the case hn = h with h 6= 0. For bnλc ≤ bnτ c we have
bnλc
1 X
n2 i=1
n
X
j=bnλc+1
bnλc
=
1 X
n2 i=1
1
1{Xi ≤Xj } −
2
n
X
1{Yi ≤Yj +h} −
j=bnτ c+1
1
2
+
bnλc
1 X
n2 i=1
bnτ c
X
1{Yi ≤Yj } −
j=bnλc+1
1
2
.
By Lemma 1 in Betken (2016) the first summand Ron the right hand side of
the equation converges in probability to λ(1 − τ ) (F (x + h) − F (x)) dF (x)
uniformly in λ ≤ τ . The second summand vanishes as n tends to ∞.
If bnλc > bnτ c,
bnλc
1 X
n2 i=1
n
X
j=bnλc+1
bnτ c
=
1 X
n2 i=1
1
1{Xi ≤Xj } −
2
n
X
j=bnλc+1
1{Yi ≤Yj +h} −
1
2
+
1
n2
bnλc
X
n
X
i=bnτ c+1 j=bnλc+1
1
1{Yi ≤Yj } −
.
2
A. Betken/Wilcoxon-based change point estimation
33
In this case, the first summand
on the right hand side of the equation converges
R
in probability to (1 − λ)τ (F (x + h) − F (x)) dF (x) uniformly in λ ≥ τ due to
Lemma 1 in Betken (2016) while the second summand converges in probability
to zero. All in all, it follows that
bnλc
1 X
n2 i=1
n
X
j=bnλc+1
1
1{Xi ≤Xj } −
2
Z
P
−→ δτ (λ)
(F (x + h) − F (x)) dF (x)
uniformly in λ ∈ [0, 1].
If limn→∞ hn = 0, the process
bnλc
1 X
ndn,r i=1
−
n
X
1{Xi ≤Xj } −
j=bnλc+1
1
2
Z
n
dn,r
(F (x + hn ) − F (x)) dF (x), 0 ≤ λ ≤ 1,
δτ (λ)
converges in distribution to
Z
1
(r)
(r)
Jr (x)dF (x) ZH (λ) − λZH (1) , 0 ≤ λ ≤ 1,
r!
due to Theorem
3.1 in Dehling, Rooch and Taqqu (2013b). By assumption
n
−1
hn = o dn,r , so that
bnλc
1 X
n2 hn i=1
n
X
1{Xi ≤Xj } −
j=bnλc+1
1
2
P
Z
−→ δτ (λ)
f 2 (x)dx, 0 ≤ λ ≤ 1.
The proof of Theorem 1, which establishes a convergence rate for the estimator k̂W , requires the following result:
Lemma 2. Suppose that Assumption 1 holds and let hn , n ∈ N, be a sequence
of real numbers with limn→∞ hn = h.
1. The process
bnλc
1 X
ndn,r i=1
n
X
Z
1{Yi ≤Yj +hn } −
F (x + hn )dF (x) , 0 ≤ λ ≤ τ,
j=bnτ c+1
converges in distribution to
Z
1 (r)
(1 − τ ) ZH (λ) Jr (x + h)dF (x)
r!
Z
1 (r)
(r)
−λ
ZH (1) − ZH (τ )
Jr (x)dF (x + h)
r!
uniformly in λ ≤ τ .
A. Betken/Wilcoxon-based change point estimation
34
2. The process
bnτ c
1 X
ndn,r i=1
n
X
Z
1{Yi ≤Yj +hn } −
F (x + hn )dF (x) , τ ≤ λ ≤ 1,
j=bnλc+1
converges in distribution to
Z
1 (r)
(1 − λ) ZH (τ ) Jr (x + h)dF (x)
r!
Z
1 (r)
(r)
ZH (1) − ZH (λ)
Jr (x)dF (x + h)
−τ
r!
uniformly in λ ≥ τ .
Proof. We give a proof for the first assertion only as the convergence of the second term follows by an analogous argument. The steps in this proof correspond
to the argument that proves Theorem 1.1 in Dehling, Rooch and Taqqu (2013a).
For λ ≤ τ it follows that
bnλc
X
n
X
Z
1{Yi ≤Yj +hn } = (n − bnτ c) bnλc
Fbnλc (x + hn )dFbnτ c+1,n (x).
i=1 j=bnτ c+1
This yields the following decomposition:
bnλc
1 X
ndn,r i=1
n
X
Z
1{Yi ≤Yj +hn } − F (x + hn )dF (x)
j=bnτ c+1
Z
n − bnτ c −1
=
dn,r bnλc
Fbnλc (x + hn ) − F (x + hn ) dFbnτ c+1,n (x)
n
Z
n − bnτ c −1
+
dn,r bnλc F (x + hn )d Fbnτ c+1,n − F (x).
n
For the first summand we have
Z
sup d−1
bnλc
Fbnλc (x + hn ) − F (x + hn ) dFbnτ c+1,n (x)
n,r
0≤λ≤τ
Z
1 (r)
− ZH (λ) Jr (x + h)dF (x)
r!
Z
≤ sup
d−1
n,r bnλc Fbnλc (x + hn ) − F (x + hn )
0≤λ≤τ
−
1 (r)
Z (λ)Jr (x + hn )dFbnτ c+1,n (x)
r! H
Z
+
1
(r)
sup Z (λ)
r! 0≤λ≤τ H
+
1
(r)
sup Z (λ)
r! 0≤λ≤τ H
(Jr (x + hn ) − Jr (x + h)) dFbnτ c+1,n (x)
Z
Jr (x + h)d Fbnτ c+1,n − F (x) .
(4)
A. Betken/Wilcoxon-based change point estimation
35
We will show that each of the summands on the right hand side converges
to 0. The first summand converges to 0 because of the empirical non-central
limit theorem of Dehling and Taqqu (1989). In order to show convergence of
(r)
the second and third summand, note that sup0≤λ≤τ |ZH (λ)| < ∞ a.s. since the
sample paths of the Hermite processes are almost surely continuous.
Furthermore, we have
Z
Z Z
Jr (x + h)dFbnτ c+1,n (x) = −
1{x+h≤G(y)} Hr (y)ϕ(y)dydFbnτ c+1,n (x)
Z Z
=−
1{x≤G(y)−h} dFbnτ c+1,n (x)Hr (y)ϕ(y)dy
Z
= − Fbnτ c+1,n (G(y) − h)Hr (y)ϕ(y)dy.
Analogously, it follows that
Z
Z
Jr (x + hn )dFbnτ c+1,n (x) = − Fbnτ c+1,n (G(y) − hn )Hr (y)ϕ(y)dy.
Therefore, we may conclude that
Z
(Jr (x + hn ) − Jr (x + h)) dFbnτ c+1,n (x)
Z
≤ 2 sup Fbnτ c+1,n (x) − F (x)
|Hr (y)| ϕ(y)dy
x∈R
Z
+ |F (G(y) − hn ) − F (G(y) − h)| |Hr (y)| ϕ(y)dy.
The first expression on the right Rhand side converges to 0 by the GlivenkoCantelli theorem and the fact that |Hr (y)| ϕ(y)dy < ∞; the second expression
converges to 0 due to continuity of F and the dominated convergence theorem.
To show convergence of the third summand note that
Z
Jr (x + h)d Fbnτ c+1,n (x) − F (x)
=
1
n − bnτ c
n
X
(Jr (Yi + h) − E Jr (Yi + h))
i=bnτ c+1
n
n
1 X
≤
(Jr (Yi + h) − E Jr (Yi + h))
n − bnτ c n i=1
+
bnτ c
X
bnτ c
1
(Jr (Yi + h) − E Jr (Yi + h)) .
n − bnτ c bnτ c i=1
For both summands on the right hand side of the above inequality the ergodic
theorem implies almost sure convergence to 0.
A. Betken/Wilcoxon-based change point estimation
36
For the second summand in (4) we have
Z
n − bnτ c −1
dn,r bnλc F (x + hn )d Fbnτ c+1,n − F (x)
n
Z
bnλc −1
=−
dn,r (n − bnτ c)
Fbnτ c+1,n (x) − F (x) dF (x + hn ).
n
Since
bnλc
n
−→ λ uniformly in λ, consider
d−1
n,r (n − bnτ c)
Z
Fbnτ c+1,n (x) − F (x) dF (x + hn )
Z
1 (r)
(r)
− (ZH (1) − ZH (τ )) Jr (x)dF (x + hn )
r!
Z
1 (r)
d−1
≤
Z (1)Jr (x)dF (x + h)
n,r n (Fn (x) − F (x)) −
r! H
Z
1 (r)
d−1
Z (τ )Jr (x)dF (x + hn )
+
n,r bnτ c Fbnτ c (x) − F (x) −
r! H
Z
1
(r)
(r)
+
Z (1) − ZH (τ )
Jr (x)d (F (x + hn ) − F (x + h)) .
r! H
The first and second summand on the right hand side converge to 0 because
of the empirical process non-central limit theorem. For the third summand we
have
Z
Z
(Jr (x − hn ) − Jr (x − h)) dF (x) .
Jr (x)d (F (x + hn ) − F (x + h)) =
As shown before in this proof, convergence to 0 follows by the Glivenko-Cantelli
theorem and the dominated convergence theorem.
Lemma 3 and Lemma 4 are needed for the proof of Theorem 2.
Lemma 3. Suppose that Assumption 1 holds and let ln , n ∈ N, and hn , n ∈ N,
be two sequences with ln → ∞, limn→∞ hn = h and ln = O(n). Then, it holds
that
Z
−1
sup dln ,r bln sc
Fbln sc (x + hn ) − Fbln sc (x + h) dFn (x)
0≤s≤1
− d−1
ln ,r bln sc
Z
(F (x + hn ) − F (x + h)) dF (x)
(5)
A. Betken/Wilcoxon-based change point estimation
37
and
sup
0≤s≤1
d−1
ln ,r bln sc
Z
(Fn (x + hn ) − Fn (x + h)) dFbln sc (x)
− d−1
ln ,r bln sc
Z
(F (x + hn ) − F (x + h)) dF (x)
(6)
converge to 0 almost surely.
Proof. For the expression (5) the triangle inequality yields
Z
sup d−1
bl
sc
Fbln sc (x + hn ) − Fbln sc (x + h) dFn (x)
n
ln ,r
0≤s≤1
−
d−1
ln ,r bln sc
Z
(F (x + hn ) − F (x + h)) dF (x)
1 (r)
d−1
Z (s)Jr (x)
ln ,r bln sc Fbln sc (x) − F (x) −
r! H
s∈[0,1],x∈R
Z
1
(r)
+
(Jr (x + hn ) − Jr (x + h))dFn (x)
sup Z (s)
r! 0≤s≤1 H
Z
l
(F (x + hn ) − F (x + h)) d (Fn − F ) (x) .
+ d−1
n
ln ,r
≤2
sup
The first summand converges to 0 because of the empirical non-central limit
(r)
(r)
theorem. Moreover, sup0≤s≤1 ZH (s) < ∞ a.s. due to the fact that ZH
isR continuous with probability 1. It is shown in the proof of Lemma 2 that
(Jr (x + hn ) − Jr (x + h))dFn (x) −→ 0. As a result, the second summand
vanishes as n tends to ∞.
Furthermore, note that
Z
d−1
l
(F (x + hn ) − F (x + h)) d (Fn − F ) (x)
ln ,r n
Z
1 (r)
d−1
n
(F
(x)
−
F
(x))
−
≤K
(1)J
(x)
dF (x + hn )
Z
n
r
n,r
r! H
Z
1 (r)
+K
d−1
n
(F
(x)
−
F
(x))
−
Z
(1)J
(x)
dF (x + h)
n
r
n,r
r! H
Z
1
(r)
+K
ZH (1)
Jr (x)d (F (x + hn ) − F (x + h))
r!
for some constant K and n sufficiently large, since ln = O(n). The first and
second summand on the right hand side of the above inequality converge to 0
due to the empirical process non-central limit theorem. In addition, we have
Z
Z
Jr (x)d (F (x + hn ) − F (x + h)) =
(Jr (x − hn ) − Jr (x − h)) dF (x)
A. Betken/Wilcoxon-based change point estimation
38
Therefore, it follows by the same argument as in the proof of Lemma 2 that the
third summand converges to 0.
Considering the term in (6), note that
sup
0≤s≤1
d−1
ln ,r bln sc
Z
− d−1
ln ,r bln sc
(Fn (x + hn ) − Fn (x + h)) dFbln sc (x)
Z
(F (x + hn ) − F (x + h)) dF (x)
1 (r)
Z (s)Jr (x)
d−1
ln ,r bln sc Fbln sc (x) − F (x) −
r! H
0≤s≤1,x∈R
Z
1
(r)
Jr (x)d (Fn (x + hn ) − Fn (x + h))
+
sup ZH (s)
r! 0≤s≤1
≤2
sup
1 (r)
+ 2K sup d−1
Z (1)Jr (x)
n,r n (Fn (x) − F (x)) −
r! H
x∈R
Z
1
(r)
+
ZH (1)
|Jr (x + hn ) − Jr (x + h)| dF (x)
r!
for some constant K and n sufficiently large. The first and third summand on
the right hand side of the above inequality converge to 0 due to the empirical
process non-central limit theorem. The last summand converges to 0 due to the
corresponding argument in the proof of Lemma 2. It holds that
Z
Jr (x)d (Fn (x + hn ) − Fn (x + h))
Z
(Fn (G(y) − hn ) − Fn (G(y) − h)) Hr (y)ϕ(y)dy
=
Z
≤ 2 sup |Fn (x) − F (x)| + sup |F (x − hn ) − F (x − h)|
|Hr (y)| ϕ(y)dy.
x∈R
x∈R
The right hand side of the above inequality converges to 0 almost surely due
to the Glivenko-Cantelli theorem and because F is uniformly continuous. As a
result, the second summand converges to 0, as well.
Lemma 4 establishes a condition under which convergence in distribution of
a sequence of random variables entails convergence of the smallest argmax of
the sequence.
Lemma 4. Let K be a compact interval and denote by D(K) the corresponding
Skorohod space, i.e. the collection of all functions f : K −→ R which are rightcontinuous with left limits. Assume that Zn , n ∈ N, are random variables taking
D
values in D(K) and that Zn −→ Z, where (with probability 1) Z is continuous
D
and Z has a unique maximizer. Then sargmax(Zn ) −→ sargmax(Z).
Proof. Due to Skorohod’s representation theorem there exist random variables
D
Z̃n and Z̃ defined on a common probability space (Ω̃, F̃ , P̃ ), such that Z̃n = Zn ,
A. Betken/Wilcoxon-based change point estimation
D
a.s.
39
Z̃ = Z and Z̃n −→ Z̃. Due to Lemma 2.9 in Seijo et al. (2011) the smallest
argmax functional is continuous at W (with respect to the Skorohod-metric
and the sup-norm metric) if W ∈ D(K) is a continuous function which has
a unique maximizer. Since (with probability 1) Z is continuous with unique
a.s.
maximizer, sargmax(Z̃n ) −→ sargmax(Z̃). As almost sure convergence implies
D
convergence in distribution, we have sargmax(Z̃n ) −→ sargmax(Z̃) and thereD
fore sargmax(Zn ) −→ sargmax(Z).
| 10math.ST
|
A Compare-Propagate Architecture with Alignment Factorization
for Natural Language Inference
Yi Tay
Nanyang Technological
University, Singapore
Luu Anh Tuan
Institute for
Infocomm Research
Siu Cheung Hui
Nanyang Technological,
University, Singapore
[email protected]
[email protected]
[email protected]
arXiv:1801.00102v1 [cs.CL] 30 Dec 2017
Abstract
This paper presents a new deep learning architecture for Natural Language
Inference (NLI). Firstly, we introduce
a new compare-propagate architecture
where alignments pairs are compared and
then propagated to upper layers for enhanced representation learning. Secondly,
we adopt novel factorization layers for efficient compression of alignment vectors
into scalar valued features, which are then
be used to augment the base word representations. The design of our approach
is aimed to be conceptually simple, compact and yet powerful. We conduct experiments on three popular benchmarks,
SNLI, MultiNLI and SciTail, achieving
state-of-the-art performance on all. A
lightweight parameterization of our model
also enjoys a ≈ 300% reduction in parameter size compared to the ESIM and
DIIN, while maintaining competitive performance. Visual analysis shows that
our propagated features are highly interpretable, opening new avenues to explainability in neural NLI models.
1
Introduction
Natural Language Inference (NLI) is a pivotal and
fundamental task in language understanding and
artificial intelligence. More specifically, given
a premise and hypothesis, NLI aims to detect
whether the latter entails or contradicts the former.
As such, NLI is also commonly known as Recognizing Textual Entailment (RTE). NLI is known
to be a significantly challenging task for machines
with success often dependent on a wide repertoire
of reasoning techniques.
In recent years we observe a steep improvement
in NLI systems, largely contributed by the release
of the largest publicly available corpus for NLI the Stanford Natural Language Inference (SNLI)
corpus (Bowman et al., 2015) which comprises
570K hand labeled sentence pairs. This improved
the feasibility of training complex neural models,
given the fact that neural models often require a
relatively large amount of training data.
Highly competitive neural models for NLI are
mostly based on soft-attention alignments, popularized by (Parikh et al., 2016). The key idea is
to learn an alignment of sub-phrases in both sentences and learn to compare the relationship between them. Standard feed-forward neural networks are commonly used to model similarity between aligned (decomposed) sub-phrases and then
aggregated into the final prediction layers.
Alignment between sentences have become a
staple technique in NLI research and many recent state-of-the-art models such as the Enhanced
Sequential Inference Model (ESIM) (Chen et al.,
2017a) also incorporate his alignment strategy.
The difference here is that ESIM considers a nonparameterized comparison scheme, i.e., concatenating the subtraction and element-wise product
of aligned sub-phrases, along with two original
sub-phrases, into the final comparison vector. A
bidirectional LSTM is then used to aggregate the
compared alignment vectors.
This paper presents a new neural model for NLI.
There are several new novel components in our
work. Firstly, we propose a compare-propagate
architecture where alignment features are propagated to upper layers (such as an RNN-based encoder) for enhancing representation learning. Notably, this is different from the compare-aggregate
paradigm that aggregates the compared alignment
vectors for prediction (Wang and Jiang, 2016a;
Parikh et al., 2016; Chen et al., 2017a). To the
best of our knowledge, we are the first to adopt
AvgMax Pooling + MLP + Softmax
RNN Encoder
Premise
Hypothesis
Augmentation
Premise
Hypothesis
Propagate
Compression
Of Match Vectors
Hypothesis
Premise
Soft-Attention
Alignment
Input
Encoding
⨀
Network
Flow
Premise
Hypothesis
Figure 1: High level overview of our proposed architecture
(best viewed in color). Alignment vectors are compressed
and then propagated to upper representation learning layers
(RNN encoders). Intra-attention is omitted in this diagram
due to the lack of space.
such a paradigm. Secondly, in order to achieve
an efficient propagation of alignment features, we
propose alignment factorization layers to reduce
each alignment vectors to a single scalar valued
feature. Each scalar valued feature is used to augment the base word representation, allowing the
subsequent RNN encoder layers to benefit from
not only global but cross sentence information.
Figure 1 depicts a high-level overview of our proposed model architecture.
There are several major advantages to our proposed architecture. Firstly, our model is relatively
compact, i.e., to avoid large alignment (or match)
vectors being propagated across the network, we
compress alignment feature vectors and augment
them to word representations instead. As a result,
our model is more parameter efficient compared to
the ESIM since the width of the middle layers of
the network is now much smaller.
Secondly, the compare-propagate paradigm enables highly interpretable features since each
alignment pair is compressed to a scalar. Previous models such as the ESIM use subtractive operations on alignment vectors, edging on the intuition that these vectors represent contradiction.
Our model is capable of visually demonstrating
this phenomena. As such, our design choice enables a new way of deriving insight from neural
NLI models.
Thirdly, our alignment factorization layers are
expressive and powerful, combining ideas from
standard machine learning literature (Rendle,
2010) with modern neural NLI models. The factorization layer tries to decompose the alignment
vector (constructed from the variations a − b, a b
and [a; b]), learning higher-order feature interactions between each compared alignment. In other
words, it models the second-order (pairwise) interactions between each feature in every alignment
vector using factorized parameters, allowing more
expressive comparison to be made over traditional
feed-forward neural networks (FFN). The effectiveness of the factorization alignment over alternative baselines such as feed-forward neural networks is confirmed by early experiments.
1.1
Our Contributions
The major contributions of this work are summarized as follows:
• We introduce a Compare-Propagate architecture for NLI. The key idea is to use the myriad of generated comparison vectors for augmentation of the base word representation instead of simply aggregating them for prediction. Subsequently, a standard compositional encoder can then be used to learn representations from the augmented word representations. We show that we are able to
derive meaningful insight from visualizing
these augmented features.
• For the first time, we adopt expressive factorization layers to model the relationships
between soft-aligned sub-phrases of sentence
pairs. Empirical experiments confirm the effectiveness of this new layer over standard
fully connected layers.
• Overall, we propose a new neural model
- CAFE (Compare-propagate Alignment-
Factorized Encoders) for NLI. Our model
achieves state-of-the-art performance on
SNLI, MultiNLI and the new SciTail dataset,
outperforming existing state-of-the-art models such as the ESIM. Ablation studies confirm the effectiveness of each proposed component in our model.
2
Related Work
Natural language inference (or textual entailment
recognition) is a long standing problem in NLP
research, typically carried out on smaller datasets
using traditional methods (Maccartney, 2009; Dagan et al., 2006; MacCartney and Manning, 2008;
Iftene and Balahur-Dobrescu, 2007).
The relatively recent creation of 570K human
annotated sentence pairs (Bowman et al., 2015)
have spurred on many recent works that use neural networks for NLI. Many advanced neural architectures have been proposed for the NLI task,
with most exploiting some variant of neural attention which learns to pay attention to important
segments in a sentence (Parikh et al., 2016; Chen
et al., 2017a; Wang and Jiang, 2016b; Rocktäschel
et al., 2015; Yu and Munkhdalai, 2017).
Amongst the myriad of neural architectures proposed for NLI, the ESIM (Chen et al., 2017a)
model is one of the best performing models.
The ESIM, primarily motivated by soft subphrase
alignment in (Parikh et al., 2016), learns alignments between BiLSTM encoded representations
and aggregates them with another BiLSTM layer.
The authors also propose the usage of subtractive
composition, claiming that this helps model contradictions amongst alignments.
Compare-Aggregate models are also highly
popular in NLI tasks. While this term was coined
by (Wang and Jiang, 2016a), many prior NLI models follow this design (Wang et al., 2017; Parikh
et al., 2016; Gong et al., 2017; Chen et al., 2017a).
The key idea is to aggregate matching features
and pass them through a dense layer for prediction. (Wang et al., 2017) proposed BiMPM, which
adopts multi-perspective cosine matching across
sequence pairs. (Wang and Jiang, 2016a) proposed
a one-way attention and convolutional aggregation
layer. (Gong et al., 2017) learns representations
with highway layers and adopts ResNet for learning features over an interaction matrix.
There a several other notable models for NLI.
For instance, models that leverage directional self-
attention (Shen et al., 2017) or Gumbel-Softmax
(Choi et al., 2017). DGEM is a graph based attention model which was proposed together with
a new entailment challenge dataset, SciTail (Khot
et al., 2018).
Our work compares and compresses alignment
pairs using factorization layers which leverages
the rich history of standard machine learning literature. Our factorization layers incorporates highly
expressive factorization machines (FMs) (Rendle,
2010) into neural NLI models. In standard machine learning tasks, FMs remain a very competitive choice for learning feature interactions (Xiao
et al., 2017) for both standard classification and
regression problems. Intuitively, FMs are adept at
handling data sparsity (typically interactions) by
using factorized parameters to approximate a feature matching matrix. This makes it suitable in our
model architecture since interaction between subphrase alignment pairs is typically very sparse1 as
well.
3
Our Proposed Model
In this section, we provide a layer-by-layer description of our model architecture. Our model accepts two sentences as an input, i.e., P (premise)
and H (hypothesis).
3.1
Input Encoding Layer
This layer aims to learn a k-dimensional representation for each word. Following (Gong et al.,
2017), we learn feature-rich word representations
by concatenating word embeddings, character features and syntactic (part-of-speech tag) features.
Character representations are learned using a convolutional encoder with max pooling function and
is commonly used in many relevant literature
(Wang et al., 2017; Chen et al., 2017b).
3.1.1
Highway Encoding Layers
Subsequently, we pass each concatenated word
vector into a two layer highway network (Srivastava et al., 2015) in order to learn a k-dimensional
representation. Highway networks are gated projection layers which learn adaptively control how
much information is being carried to the next
layer. Our strategy is similar to (Parikh et al.,
2016) which trains the projection layer in place of
tuning the embedding matrix. The usage of high1
Word-word interactions are already extremely sparse, let
alone phrasal interactions.
way layers over standard projection layers is empirically motivated. However, an intuition would
be that the gates in this layer adapt to learn the relative importance of each word to the NLI task.
Let H(.) and T (.) be single layered affine transforms with ReLU and sigmoid activation functions
respectively. A single highway network layer is
defined as:
y = H(x, WH ) · T (x, WT ) + C · x
(1)
where C = (1−T (x, WT )) and WH , WT ∈ Rr×d
Notably, the dimensions of the affine transform
might be different from the size of the input vector.
In this case, an additional nonlinear transform is
used to project x to the same dimensionality. The
output of this layer is P̄ ∈ Rk×`P (premise) and
H̄ ∈ Rk×`H (hypothesis), with each word converted to a r-dimensional vector. ,
3.2
3.4
eij = F (p̄i )> · F (h̄j )
3.4.1
βi :=
j=1
αj :=
`h
X
i=1
(2)
Given an input vector x, the factorization operation (Rendle, 2010) is defined as:
L(x) = w0 +
(3)
exp(eij )
h̄i
P`h
k=1 exp(ekj )
(4)
Intra-Attention Alignment Layer
This layer learns a self-alignment of sentences and
is applied to both P̄ and H̄ independently. For the
sake of brevity, let S̄ represent either P̄ or H̄, the
intra-attention alignment is computed as:
`p
X
j=1
exp(fij )
s̄j
P`p
exp(f
)
ik
k=1
(5)
n
X
wi x i
(6)
hvi , vj i xi xj
(7)
i=1
n X
n
X
i=1 j=i+1
Ff m (x) = L(x) + P (x)
(8)
where Ff m (x) is a scalar valued output. h.; .i is
the dot product between two vectors and w0 is
the global bias. The parameters of this layer are
w0 ∈ R, w ∈ Rr and v ∈ Rr×k . Intuitively, L(x)
represents a linear regression layer while P (x)
learns pairwise feature interactions by trying to
factorize the feature interaction matrix.
3.4.2
exp(eij )
p̄j
P`p
exp(e
)
ik
k=1
where βi is the sub-phrase in P̄ that is softly
aligned to hi . Intuitively, βi is a weighted sum
`p
across {pj }j=1
, selecting the most relevant parts
of P̄ to represent hi .
s0i :=
Factorization Operation
P (x) =
where E ∈ R`p ×`h and p̄i , h̄j are the i-th and j-th
word in the premise and hypothesis respectively.
`p
X
Alignment Factorization Layers
This layer aims to learn a scalar valued feature
for each comparison between aligned sub-phrases.
Firstly, we introduce our factorization operation,
which lives at the core of our neural model.
Inter-Attention Alignment Layer
This layer learns an alignment of sub-phrases between P̄ and H̄. Let F (.) be a standard projection
layer with ReLU activation function. The alignment matrix of two sequences is defined as follows:
3.3
where fij = G(s̄i )> · G(s̄j ) and G(.) is a nonlinear projection layer with ReLU activation function. The intra-attention layer models similarity
of each word with respect to the entire sentence,
capturing long distance dependencies and ‘global’
context of the entire sentence.
Alignment Factorization
This layer compares the alignment between interattention aligned representations, i.e., (βi , hi ) and
(αj , pj ). Let (a, b) represent an alignment pair, we
apply the following operations:
yc = Ff m ([a; b])
(9)
ys = Ff m (a − b)
(10)
ym = Ff m (a
b)
(11)
where yc , ys , ym ∈ R and Z(.) is the factorization
operation, [.; .] is the concatenation operator and
is the element-wise multiplication. The intuition
of modeling subtraction is targeted at capturing
contradiction. However, instead of simply concatenating the extra comparison vectors, we compress them using factorization layers. Finally, for
each alignment pair, we obtain three scalar-valued
features which map precisely to a word in the sequence.
Next, for each sequence, we also apply alignment factorization on the intra-aligned sentences.
Let (s, s0 ) represent an intra-aligned pair from either the premise or hypothesis, we compute the
following operations:
A logical explanation would be because our word
representations are already augmented with global
(intra-attention) information. As such, modeling
in the reverse direction is unnecessary, resulting in
some computational savings.
vc = Ff m ([s; s0 ])
(12)
vs = Ff m (s − s0 )
(13)
3.6.1 Pooling Layer
Next, to learn an overall representation of each
sentence, we apply a pooling function across all
hidden outputs of the sequential encoder. The
pooling function is a concatenation of temporal
max and average (avg) pooling.
vm = Ff m (s
0
s)
(14)
where vc , vs , vm ∈ R and Z(.) is the factorization operation. Applying alignment factorization
to intra-aligned representations produces another
three scalar-valued features which are mapped to
each word in the sequence. Note that each of the
six factorization operations has its own parameters but shares them amongst all words in the sentences.
3.5
Propagation and Augmentation
(15)
i
i
where si is i-th word in P̄ or H̄., fintra
and finter
are the intra-aligned [vc ; vs ; vm ] and inter-aligned
[yc ; ys ; ym ] features for the i-th word in the sei
quence respectively. Intuitively, fintra
augments
each word with global knowledge of the seni
tence and finter
augments each word with crosssentence knowledge via inter-attention.
3.6
Sequential Encoder Layer
For each sentence, the augmented word representations u1 , u2 , . . . u` is then passed into a sequential encoder layer. We adopt a standard vanilla
LSTM encoder.
hi = LST M (u, i), ∀i ∈ [1, . . . `]
(16)
where ` represents the maximum length of the sequence. Notably, the parameters of the LSTM are
siamese in nature, sharing weights between both
premise and hypothesis. We do not use a bidirectional LSTM encoder, as we found that it did
not lead to any improvements on the held-out set.
2
where x is a final 2k-dimensional representation
of the sentence (premise or hypothesis). We also
experimented with sum and avg standalone poolings and found sum pooling to be relatively competitive.
3.7
Finally, the six factorized features are then aggregated2 via concatenation to form a final feature
vector that is propagated to upper representation
learning layers via augmentation of the word representation P̄ or H̄.
i
i
ui = [si ; fintra
; finter
]
x = [max([h1 , · · · h` ]); avg([h1 , · · · h` ])] (17)
Following (Parikh et al., 2016), we may also concatenate
the intra-aligned vector to ui which we found to have speed
up convergence.
Prediction Layers
Finally, given a fixed dimensional representation
of the premise xp and hypothesis xh , we pass
their concatenation into a two-layer h-dimensional
highway network. Since the highway network has
been already defined earlier, we omit the technical
details. The final prediction layers of our model is
computed as follows:
yout = H2 (H1 ([xp ; xh ; xp
xh ; xp − xh ]))
(18)
where H1 (.), H2 (.) are highway network layers
with ReLU activation. The output is then passed
into a final linear softmax layer.
ypred = sof tmax(WF · yout + bF )
(19)
where WF ∈ Rh×3 and bF ∈ R3 . The network
is then trained using standard multi-class cross entropy loss with L2 regularization.
4
4.1
Experiments
Experimental Setup
To ascertain the effectiveness of our models, we
use the SNLI (Bowman et al., 2015) and MultiNLI
(Williams et al., 2017) benchmarks which are standard and highly competitive benchmarks for the
NLI task. We also include the newly released SciTail dataset (Khot et al., 2018) which is a binary
entailment classification task constructed from science questions. Notably, SciTail is known to be a
difficult dataset for NLI, made evident by the low
accuracy scores even though it is binary in nature.
• SNLI - We compare against competitors
across three settings. The first setting disallows cross sentence attention. In the second setting, cross sentence is allowed. The
first two setting only comprises single models. The last setting is a comparison between
model ensembles. Though we compare with
many other models (Wang and Jiang, 2016b;
Rocktäschel et al., 2015; Parikh et al., 2016;
Yu and Munkhdalai, 2017; Sha et al., 2016),
the key SoTA competitors on this dataset are
the BiMPM (Wang et al., 2017), ESIM (Chen
et al., 2017a) and DIIN (Gong et al., 2017).
• MultiNLI - We compare on two test sets
(matched and mismatched) which represent
in-domain and out-domain performance. The
main competitor on this dataset is the ESIM
model, a powerful state-of-the-art SNLI
baseline. We also compare with ESIM + read
(Weissenborn, 2017).
• SciTail - This dataset only has one official
setting. We compare against the reported results of ESIM (Chen et al., 2017a) and DecompAtt (Parikh et al., 2016) in the original
paper. We also compare with DGEM, the
new model proposed in (Khot et al., 2018).
Across all experiments and in the spirit of fair
comparison, we only compare with works that (1)
do not use extra training data and (2) do not use external resources (such as external knowledge bases
etc.).
4.2
Implementation Details
We implement our model in Tensorflow (Abadi
et al., 2015) and train them on Nvidia P100 GPUs.
For the first setting (without cross sentence attention), we remove both the cross (inter) attention
layers to abide to the rules of this setting. We
use the Adam optimizer (Kingma and Ba, 2014)
with an initial learning rate of 0.0003. Default
L2 regularization is set to 10−6 . Dropout with a
keep probability of 0.8 is applied after each fullyconnected, recurrent or highway layer. The batch
size is tuned amongst {128, 256, 512}. The number of latent factors k for the factorization layers
is tuned amongst {5, 10, 50, 100, 150}. The size
of the hidden layers of the highway layers are set
to 300. All parameters are initialized with xavier
initialization. Word embeddings are pre-loaded
with 300d GloVE embeddings (Pennington et al.,
2014) and fixed during training. Sequence length
are padded to batch-wise maximums. The batch
order are (randomly) sorted within buckets following (Parikh et al., 2016).
,
|θ|
Train
Encoders (w/o Cross Sentence Attention)
300D LSTM Encoder
3.0M
83.9
600D Gated BiLSTM + intra-att
12M
90.5
300D Gumbel LSTM TreeLSTM
2.9M
91.2
300D DISAN
2.4M
91.1
300D Residual Stacked Encoders
9.7M
89.8
600D Gumbel TreeLSTM
10M
93.1
300D CAFE (w/o Inter-Attention)
3.7M
87.3
Cross Sentence Attention (Single Models)
100D LSTM with attention
250K
85.3
300D mLSTM
1.9M
92.0
200D DecompAtt
380K
89.5
200D DecompAtt + Intra-Att
580K
90.5
300D NTI-SLSTM-LSTM
3.2M
88.5
300D re-read LSTM
2.0M
90.7
BiMPM
1.6M
90.9
448D DIIN
4.4M
91.2
600D ESIM
4.3M
92.6
150D CAFE (Sum+2x200D MLP)
750K
88.2
200D CAFE (Sum+2x400D MLP)
1.4M
89.4
300D CAFE (Sum+2x600D MLP)
3.5M
89.2
300D CAFE (AvgMax+300D HN)
4.7M
89.8
Cross Sentence Attention (Ensemble Models)
600D ESIM + 300D TreeLSTM
7.7M
93.5
BiMPM
6.4M
93.2
448D DIIN
17.0M
92.3
300D CAFE (Ensemble)
17.5M
92.5
Method
Test
80.6
85.5
85.6
85.6
85.7
86.0
85.9
83.5
86.1
86.3
86.8
87.3
87.5
87.5
88.0
88.0
87.7
88.1
88.3
88.5
88.6
88.8
88.9
89.3
Table 1: Performance comparison of all models on SNLI
benchmark.
4.3
Results on SNLI
Table 1 reports our results on the SNLI benchmark. On the cross sentence (single model setting), the performance of our proposed CAFE
model is extremely competitive. We report the
test accuracies of CAFE at different extent of parameterization, varying the size of the LSTM encoder, width of the pre-softmax hidden layers and
final pooling layer. CAFE obtains 88.5% accuracy on the SNLI test set, an extremely competitive score on the extremely popular benchmark.
Notably, competitive results can be also achieved
with a much smaller parameterization. For example, CAFE also achieves 88.3% and 88.1%
test accuracy with only 3.5M and 1.5M respectively. This outperforms state-of-the-art ESIM and
DIIN models with only a fraction of the parameter
cost. Moreover, our lightweight adaption achieves
87.7% with only 750K parameters, which makes
it extremely performant amongst models having
the same amount of parameters such as the decomposable attention model (86.8%).
After removing the inter-attention layers from
CAFE, the performance remains competitive to
top performing encoder models. CAFE (w/o cross
attention) achieves a respectible 85.9% accuracy
in this setting. Notably, the best performing
model, the Gumbel TreeLSTM (86.0%) has over
10M parameters. On the other hand, our CAFE
model has a 300% less parameters, yet performs
competitively to the Gumbel TreeLSTM.
Finally, an ensemble of 5 CAFE models
achieves 89.3% test accuracy, the best test scores
on the SNLI benchmark to date. Overall, we
believe that the good performance of our CAFE
can be attributed to (1) the effectiveness of the
compare-propagate (i.e., providing word representations with global and local knowledge for better
representation learning) and (2) the expressiveness
of factorization layers that are used to decompose
and compare word alignments. More details are
given at the ablation study. Finally, we emphasize
that CAFE is also relatively lightweight, efficient
and fast to train given its performance. A single
run on SNLI takes approximately 5 minutes per
epoch with a batch size of 256. Overall, a single
run takes ≈ 3 hours to get to convergence.
4.4
Results on MultiNLI and Scitail
Table 2 reports our results on the MultiNLI and
SciTail datasets. On MultiNLI, CAFE significantly outperforms ESIM, a strong SoTA models
on both settings. We also outperform the ESIM
+ Read model (Weissenborn, 2017). An ensemble of CAFE models achieve the best reported
results on this MultiNLI to date (as of time of
writing). On SciTail, Our proposed CAFE model
achieves state-of-the-art performance. The performance gain over strong baselines such as DecompAtt and ESIM are ≈ 10% − 13% in terms of accuracy. CAFE also outperforms DGEM , which
use a graph-based attention for improved performance, by a significant margin of 5%. As such,
empirical results demonstrate the effectiveness of
our proposed CAFE model on the challenging SciTail dataset.
4.5
Model
Majority
NGRAM#
CBOW[
BiLSTM[
ESIM#,[
DecompAtt# DGEM#
DGEM + Edge#
ESIM†
ESIM + Read†
CAFE
CAFE Ensemble
MultiNLI
Match Mismatch
36.5
35.6
65.2
64.8
69.8
69.4
72.4
72.1
76.3
75.8
77.8
77.0
78.7
77.9
80.2
79.0
SciTail
60.3
70.6
70.6
72.3
70.8
77.3
83.3
-
Table 2: Performance comparison (accuracy) on MultiNLI
and SciTail. Models with †, # and [ are reported from (Weissenborn, 2017), (Khot et al., 2018) and (Williams et al., 2017)
respectively.
which we find to have helped CAFE marginally.
(4) Removes the Inter attention features, which
naturally impacts the model performance significantly. (5-6) explores the effectiveness of the highway layers (in prediction layers and encoding layers) by replacing them to FC layers. Both highway layers have marginally helped overall performance. Finally, (7-9) removes the alignment features based on their composition type. We observe that the Sub feat and Concat compositions
were more important than the Mul composition.
However, removing any of the three will result in
some performance degrade. Finally, we replace
the LSTM encoder with a BiLSTM, observing
that adding bidirectionality did not improve performance for our model.
Original Model
(1) Replace FM with FC
(2) Remove Char Embed
(3) Remove Syn Embed
(4) Remove Inter Att
(5) Replace Highway Pred. with FC
(6) Remove Highway Enc. with FC
(7) Remove Sub Feat
(8) Remove Mul Feat
(9) Remove Concat Feat
(10) Replace LSTM with BiLSTM
Match
79.0
77.7
78.1
78.3
75.2
77.7
78.7
77.9
78.7
77.9
78.3
Mismatch
78.9
77.9
78.3
78.4
75.6
77.9
78.7
78.3
78.6
77.6
78.4
Table 3: Ablation study on MultiNLI development sets.
Ablation Study
Table 3 reports ablation studies on the MultiNLI
development sets. Firstly, we replaced all FM
functions with regular full-connected (FC) layers
(1). We notice an decline in performance in both
development sets. Ablation (2-3) explores the utility of using character and syntactic embeddings,
4.6
Linguistic Error Analysis
We perform a linguistic error analysis using
the supplementary annotations provided by the
MultiNLI dataset. We compare against the model
outputs of the ESIM model across 13 categories
20
10
0
4.7
Interpreting and Visualizing with CAFE
Finally, we also observed that the propagated features are highly interpretable, giving insights to
the inner workings of the CAFE model. Figure
2 shows a visualization of the feature values from
an example in the SNLI test set. The ground truth
is contradiction. Based on the above example we
make several observations. Firstly, inter mul features mostly capture identical words (or semantically similar words), i.e., inter mul features for
river spikes in both sentences. Secondly, inter sub
spikes on conflicting words that might cause contradiction, e.g sedan and land rover are not the
same vehicle. Another interesting observation is
that we notice the inter sub features for driven
.
.
<e
>
r
riv
e
riv
er
a
ros
s
n
ac
dri
ve
be
ing
is
er
rov
a
lan
d
−10
inter_cat
inter_mul
inter_sub
intra_cat
intra_mul
intra_sub
30
Table 4: Linguistic Error Analysis on MultiNLI dataset.
20
10
>
<e
a
of
le
dd
mi
the
in
ck
stu
n
is
da
a
>
0
<s
of linguistic3 phenenoma. Table 4 reports the result of our error analysis. Firstly, we observe that
our CAFE model generally outperforms ESIM
on most categories. On the mismatched setting,
CAFE outperforms ESIM in 12 out of 13 categories, losing only in one percentage point in
Active/Passive category. On the matched setting,
CAFE is outperformed by ESIM very marginally
on coreference and paraphrase categories. Despite
generally achieving much superior results, we noticed that CAFE performs poorly on conditionals4 on the matched setting. Measuring the absolute ability of CAFE, we find that CAFE performs
extremely well in handling linguistic patterns of
paraphrase detection and active/passive. This is
likely to be attributed by the alignment strategy
that CAFE and ESIM both exploits.
inter_cat
inter_mul
inter_sub
intra_cat
intra_mul
intra_sub
30
se
Mismatched
ESIM CAFE
60
85
62
87
71
80
58
80
69
77
79
89
91
90
84
95
54
62
75
83
72
80
76
81
67
83
<s
>
Conditional
Word overlap
Negation
Antonym
Long Sentence
Tense Difference
Active/Passive
Paraphrase
Quantity/Time
Coreference
Quantifier
Modal
Belief
Matched
ESIM CAFE
100
70
50
82
76
76
67
82
75
79
73
82
88
100
89
88
33
53
83
80
69
75
78
81
65
77
Figure 2: Visualization of six Propgated Features (Best
viewed in color). Legend is denoted by {inter,intra} followed
by the operations mul, sub or concat (cat).
and stuck spiking. This also validates the observation of (Chen et al., 2017a), which shows what
the sub vector in the ESIM model is looking out
for contradictory information. However, our architecture allows the inspection of these vectors
since they are compressed via factorization, leading to larger extents of explainability - a quality
that neural models inherently lack. We also observed5 that intra attention (e.g., intra cat) features
seem to capture the more important words in the
sentence (river, sedan, landrover).
5
Conclusion
We proposed a new neural architecture, CAFE
for NLI. CAFE achieves state-of-the-art performance on three benchmark datasets. Moreover,
a lightweight parameterization of CAFE outperforms other SoTA models such as ESIM and
DINN while enjoying a 300% savings in parameter cost. The design of CAFE opens up new
avenues for interpretability in neural models for
NLI. Qualitatively, we show how different compositional operators (e.g., sub and mul) behave in
NLI task and shed light on why subtractive composition helps in other models such as the ESIM.
3
Due to the lack of space, we refer readers to
http://www.nyu.edu/projects/bowman/
multinli/multinli_1.0_annotations.zip
for an explanation on each category.
4
This only accounts for 5% of samples.
5
Specific spikes in intra mul and intra sub was also observed in other samples but we leave them out due to the lack
of space.
References
Martı́n Abadi, Ashish Agarwal, Paul Barham, Eugene
Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin,
Sanjay Ghemawat, Ian Goodfellow, Andrew Harp,
Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal
Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh
Levenberg, Dan Mané, Rajat Monga, Sherry Moore,
Derek Murray, Chris Olah, Mike Schuster, Jonathon
Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden,
Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. 2015. TensorFlow: Large-scale machine learning on heterogeneous systems. Software
available from tensorflow.org.
Samuel R. Bowman, Gabor Angeli, Christopher Potts,
and Christopher D. Manning. 2015.
A large
annotated corpus for learning natural language
inference.
In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, EMNLP 2015, Lisbon, Portugal, September 17-21, 2015. pages 632–642.
http://aclweb.org/anthology/D/D15/D15-1075.pdf.
Qian Chen, Xiaodan Zhu, Zhen-Hua Ling, Si Wei, Hui
Jiang, and Diana Inkpen. 2017a. Enhanced LSTM
for natural language inference. In Proceedings of
the 55th Annual Meeting of the Association for
Computational Linguistics, ACL 2017, Vancouver,
Canada, July 30 - August 4, Volume 1: Long Papers.
pages 1657–1668. https://doi.org/10.18653/v1/P171152.
Qian Chen, Xiaodan Zhu, Zhen-Hua Ling, Si Wei,
Hui Jiang, and Diana Inkpen. 2017b. Recurrent
neural network-based sentence encoder with gated
attention for natural language inference. In Proceedings of the 2nd Workshop on Evaluating Vector
Space Representations for NLP, RepEval@EMNLP
2017, Copenhagen, Denmark, September 8, 2017.
pages 36–40. http://aclanthology.info/papers/W175307/w17-5307.
Jihun Choi, Kang Min Yoo, and Sang-goo Lee.
2017.
Unsupervised learning of task-specific
tree structures with tree-lstms.
arXiv preprint
arXiv:1707.02786 .
Ido Dagan, Oren Glickman, and Bernardo Magnini.
2006. The pascal recognising textual entailment
challenge. In Proceedings of the First International
Conference on Machine Learning Challenges: Evaluating Predictive Uncertainty Visual Object Classification, and Recognizing Textual Entailment.
Springer-Verlag, Berlin, Heidelberg, MLCW’05,
pages 177–190.
Yichen Gong, Heng Luo, and Jian Zhang. 2017. Natural language inference over interaction space. CoRR
abs/1709.04348. http://arxiv.org/abs/1709.04348.
Adrian Iftene and Alexandra Balahur-Dobrescu. 2007.
Hypothesis transformation and semantic variability
rules used in recognizing textual entailment. In Proceedings of the ACL-PASCAL Workshop on Textual
Entailment and Paraphrasing. Association for Computational Linguistics, Stroudsburg, PA, USA, RTE
’07, pages 125–130.
Tushar Khot, Ashish Sabharwal, and Peter Clark. 2018.
Scitail: A textual entailment dataset from science
question answering. In AAAI.
Diederik P. Kingma and Jimmy Ba. 2014.
A method for stochastic optimization.
abs/1412.6980.
Adam:
CoRR
Bill Maccartney. 2009. Natural Language Inference.
Ph.D. thesis, Stanford, CA, USA. AAI3364139.
Bill MacCartney and Christopher D. Manning. 2008.
Modeling semantic containment and exclusion in
natural language inference. In Proceedings of the
22Nd International Conference on Computational
Linguistics - Volume 1. Association for Computational Linguistics, Stroudsburg, PA, USA, COLING
’08, pages 521–528.
Ankur P. Parikh, Oscar Täckström, Dipanjan Das, and
Jakob Uszkoreit. 2016. A decomposable attention
model for natural language inference. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016,
Austin, Texas, USA, November 1-4, 2016. pages
2249–2255.
Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. Glove: Global vectors for
word representation. In Proceedings of the 2014
Conference on Empirical Methods in Natural Language Processing, EMNLP 2014, October 25-29,
2014, Doha, Qatar, A meeting of SIGDAT, a Special
Interest Group of the ACL. pages 1532–1543.
Steffen Rendle. 2010. Factorization machines. In Data
Mining (ICDM), 2010 IEEE 10th International Conference on. IEEE, pages 995–1000.
Tim Rocktäschel, Edward Grefenstette, Karl Moritz
Hermann, Tomáš Kočiskỳ, and Phil Blunsom. 2015.
Reasoning about entailment with neural attention.
arXiv preprint arXiv:1509.06664 .
Lei Sha, Baobao Chang, Zhifang Sui, and Sujian Li. 2016.
Reading and thinking: Reread LSTM unit for textual entailment recognition. In COLING 2016, 26th International Conference on Computational Linguistics, Proceedings of the Conference: Technical Papers, December 11-16, 2016, Osaka, Japan. pages 2870–2879.
http://aclweb.org/anthology/C/C16/C16-1270.pdf.
Tao Shen, Tianyi Zhou, Guodong Long, Jing Jiang,
Shirui Pan, and Chengqi Zhang. 2017. Disan:
Directional self-attention network for rnn/cnn-free
language understanding. CoRR abs/1709.04696.
http://arxiv.org/abs/1709.04696.
Rupesh Kumar Srivastava, Klaus Greff, and Jürgen
Schmidhuber. 2015. Highway networks. CoRR
abs/1505.00387. http://arxiv.org/abs/1505.00387.
Shuohang Wang and Jing Jiang. 2016a. A compareaggregate model for matching text sequences. CoRR
abs/1611.01747. http://arxiv.org/abs/1611.01747.
Shuohang Wang and Jing Jiang. 2016b. Learning natural language inference with LSTM. In NAACL HLT
2016, The 2016 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language Technologies, San Diego
California, USA, June 12-17, 2016. pages 1442–
1451.
http://aclweb.org/anthology/N/N16/N161170.pdf.
Zhiguo Wang, Wael Hamza, and Radu Florian. 2017.
Bilateral multi-perspective matching for natural language sentences. In Proceedings of the TwentySixth International Joint Conference on Artificial Intelligence, IJCAI 2017, Melbourne, Australia, August 19-25, 2017. pages 4144–4150.
https://doi.org/10.24963/ijcai.2017/579.
Dirk Weissenborn. 2017. Reading twice for natural language understanding. CoRR abs/1706.02596.
http://arxiv.org/abs/1706.02596.
Adina Williams, Nikita Nangia, and Samuel R. Bowman. 2017. A broad-coverage challenge corpus for
sentence understanding through inference. CoRR
abs/1704.05426. http://arxiv.org/abs/1704.05426.
Jun Xiao, Hao Ye, Xiangnan He, Hanwang Zhang, Fei
Wu, and Tat-Seng Chua. 2017. Attentional factorization machines: Learning the weight of feature
interactions via attention networks. arXiv preprint
arXiv:1708.04617 .
Hong Yu and Tsendsuren Munkhdalai. 2017. Neural semantic encoders.
In Proceedings of the
15th Conference of the European Chapter of
the Association for Computational Linguistics, EACL 2017, Valencia, Spain, April 3-7,
2017, Volume 1: Long Papers. pages 397–407.
http://aclanthology.info/papers/E17-1038/neuralsemantic-encoders.
| 2cs.AI
|
This is an early draft. Please read/cite the published version instead:
Caliskan, Aylin, Joanna J. Bryson, and Arvind Narayanan. “Semantics derived automatically from language corpora contain human-like biases.” Science
356.6334 (2017): 183-186. http://opus.bath.ac.uk/55288/
arXiv:1608.07187v4 [cs.AI] 25 May 2017
Semantics derived automatically from language
corpora necessarily contain human biases
Aylin Caliskan1 , Joanna J. Bryson1,2 , and Arvind Narayanan1
1 Princeton
University
of Bath
∗ Address correspondence to [email protected], [email protected], [email protected].
2 University
ABSTRACT
Artificial intelligence and machine learning are in a period of astounding growth. However, there are concerns that these
technologies may be used, either with or without intention, to perpetuate the prejudice and unfairness that unfortunately
characterizes many human institutions. Here we show for the first time that human-like semantic biases result from the
application of standard machine learning to ordinary language—the same sort of language humans are exposed to every
day. We replicate a spectrum of standard human biases as exposed by the Implicit Association Test and other well-known
psychological studies. We replicate these using a widely used, purely statistical machine-learning model—namely, the GloVe
word embedding—trained on a corpus of text from the Web. Our results indicate that language itself contains recoverable and
accurate imprints of our historic biases, whether these are morally neutral as towards insects or flowers, problematic as towards
race or gender, or even simply veridical, reflecting the status quo for the distribution of gender with respect to careers or first
names. These regularities are captured by machine learning along with the rest of semantics. In addition to our empirical
findings concerning language, we also contribute new methods for evaluating bias in text, the Word Embedding Association
Test (WEAT) and the Word Embedding Factual Association Test (WEFAT). Our results have implications not only for AI and
machine learning, but also for the fields of psychology, sociology, and human ethics, since they raise the possibility that mere
exposure to everyday language can account for the biases we replicate here.
Introduction
Those astonished by the human-like capacities visible in the recent advances in artificial intelligence (AI) may be comforted
to know the source of this progress. Machine learning, exploiting the universality of computation (Turing, 1950), is able to
capture the knowledge and computation discovered and transmitted by humans and human culture. However, while leading to
spectacular advances, this strategy undermines the assumption of machine neutrality. The default assumption for many was that
computation, deriving from mathematics, would be pure and neutral, providing for AI a fairness beyond what is present in
human society. Instead, concerns about machine prejudice are now coming to the fore—concerns that our historic biases and
prejudices are being reified in machines. Documented cases of automated prejudice range from online advertising (Sweeney,
2013) to criminal sentencing (Angwin et al., 2016).
Most experts and commentators recommend that AI should always be applied transparently, and certainly without prejudice.
Both the code of the algorithm and the process for applying it must be open to the public. Transparency should allow courts,
companies, citizen watchdogs, and others to understand, monitor, and suggest improvements to algorithms (Oswald and Grace,
2016). Another recommendation has been diversity among AI developers, to address insensitive or under-informed training
of machine learning algorithms (Sweeney, 2013; Noble, 2013; Barr, 2015; Crawford, 2016). A third has been collaboration
between engineers and domain experts who are knowledgeable about historical inequalities (Sweeney, 2013).
Here we show that while all of these strategies might be helpful and even necessary, they could not be sufficient. We
document machine prejudice that derives so fundamentally from human culture that it is not possible to eliminate it through
Outdated draft. See published version at http://opus.bath.ac.uk/55288/.
strategies such as the above. We demonstrate here for the first time what some have long suspected (Quine, 1960)—that
semantics, the meaning of words, necessarily reflects regularities latent in our culture, some of which we now know to be
prejudiced. We demonstrate this by showing that standard, widely used Natural Language Processing tools share the same
biases humans demonstrate in psychological studies. These tools have their language model built through neutral automated
parsing of large corpora derived from the ordinary Web; that is, they are exposed to language much like any human would be.
Bias should be the expected result whenever even an unbiased algorithm is used to derive regularities from any data; bias is the
regularities discovered.
Human learning is also a form of computation. Therefore our finding that data derived from human culture will deliver
biases and prejudice have implications for the human sciences as well. They present a new “null hypothesis” for explaining the
transmission of prejudice between humans. Our findings also have implications for addressing prejudice, whether in humans or
machines. The fact that it is rooted in language makes prejudice difficult to address, but by no means impossible. We argue
that prejudice must be addressed as a component of any intelligent system learning from our culture. It cannot be entirely
eliminated from the system, but rather must be compensated for.
In this article, we begin by explaining meaning and the methods by which we determine human understanding, and interpret
it in machines. Then we present our results. We replicate previously-documented biases and prejudices in attitudes towards
ordinary objects, animals, and humans. We show that prejudices that reduce the number of interview invitations sent to people
because of the racial association of their name, and that associate women with arts rather than science or mathematics, can
be retrieved from standard language tools used in ordinary AI products. We also show that veridical information about the
proportions of women in particular job categories, or what proportion of men versus women have a particular name, can be
recovered using the same methods. We then present a detailed account of our methods, and further discussion of the implications
of our work.
Meaning and Bias in Humans and Machines
In AI and machine learning, bias refers to prior information, a necessary prerequisite for intelligence (Bishop, 2006). Yet bias
can be problematic where prior information is derived from precedents known to be harmful. For the purpose of this paper,
we will call harmful biases ‘prejudice’. We show here that prejudice is a special case of bias identifiable only by its negative
consequences, and therefore impossible to eliminate purely algorithmically. Rather, prejudice requires deliberate action based
on knowledge of a society and its outstanding ethical challenges.
If we are to demonstrate that AI incorporates the same bias as humans, we first have to be able to document human bias.
We will use several methods to do this below, but the one we use most is the Implicit Association Test (IAT). First introduced
by Greenwald et al. (1998), the IAT demonstrates enormous differences in response times when subjects are asked to pair two
concepts they find similar, in contrast to two concepts they find different. The IAT follows a reaction time paradigm, which
means subjects are encouraged to work as quickly as possible, and their response times are the quantified measure. For example,
subjects are much quicker if they are told to label insects as unpleasant and flowers as pleasant than if they are asked to label
these objects in reverse. The fact that a pairing is faster is taken to indicate that the task is more easy, and therefore that the two
subjects are linked in their mind. The IAT is ordinarily used to pair categories such as ‘male’ and ‘female’ with attributes such
as ‘violent’ or ‘peaceful’. The IAT has been used to describe and account for a wide range of implicit prejudices and other
phenomena, including stereotype threat (Kiefer and Sekaquaptewa, 2007; Stanley et al., 2011).
Our method for demonstrating both bias and prejudice in text is a variant of the implicit association test applied to a
widely-used semantic representation of words in AI, termed word embeddings. These are derived by representing the textual
context in which a word is found as a vector in a high-dimensional space. Roughly, for each word, its relationship to all other
words around it is summed across its many occurrences in a text. We can then measure the distances (more precisely, cosine
similarity scores) between these vectors. The thesis behind word embeddings is that words that are closer together in the vector
−−−−−−−→ is closer to −
−→ than to −
−−−→ it
space are semantically closer in some sense. Thus, for example, if we find that −
programmer
man
woman,
suggests (but is far from conclusive of) a gender stereotype. We assume here that this measure is analogous to reaction time in
the IAT, since the shorter time implies a semantic ‘nearness’ (McDonald and Lowe, 1998; Moss et al., 1995).
As with the IAT, we do not just compare two words. Many if not most words have multiple meanings, which makes
pairwise measurements “noisy”. To control for this, we use small baskets of terms to represent a concept. In the present
paper we have never invented our own basket of words, but rather have in every case used the same words as were used in the
psychological study we are replicating. We should note that distances / similarities of word embeddings notoriously lack any
intuitive interpretation. But this poses no problem for us: our results and their import do not depend on attaching meaning to
these distances. Our primary claim is that the associations revealed by relative nearness scores between categories match human
biases and stereotypes strongly (i.e., low p-values and high effect sizes) and across many categories. Thus, the associations in
the word vectors could not have arisen by chance, but instead reflect extant biases in human culture.
2/14
Outdated draft. See published version at http://opus.bath.ac.uk/55288/.
There are however several key differences between our method and the IAT. Most of these we will discuss in future versions
of this paper’s appendices, but one in particular is critical to our presentation of results. While the IAT applies to individual
human subjects, the embeddings of interest to us are derived from the aggregate writings of humans on the web. These corpora
are generated in an uncontrolled fashion and are not representative of any one population (though our results indicate they
may be disproportionately American; see below). The IAT has sometimes been used to draw conclusions about populations by
averaging individual results over samples, and these have led to important insights on racial bias and gender stereotypes, among
others. Our tests of word embeddings are loosely analogous to population-level IATs.
Nevertheless, this difference precludes a direct numerical comparison between human biases measured by the IAT and
algorithmic biases measured by our methods. In particular, an IAT allows rejecting the null hypothesis (of non-association
between two categories) via a p-value and quantification of the strength of association via an effect size. These are obtained
by administering the test to a statistically-significant sample of subjects (and multiple times to each subject). With word
embeddings, there is no notion of test subjects. Roughly, it is as if we are able to measure the mean of the association strength
over all the “subjects” who collectively created the corpora. But we have no way to observe variation between subjects or
between trials. We do report p-values and effect sizes resulting from the use of multiple words in each category, but the meaning
of these numbers is entirely different from those reported in IATs.
Results
Using the techniques described in the Methods section, we have found every linguistic bias documented in psychology that we
have looked for. Below are a sample that we think are persuasive. We have not cherry picked these for effect size—these are
uniformly high. Rather, we chose these to illustrate our assertion that we can account for a variety of implicit human biases
purely from language regularities, and that these are in fact part and parcel with the meaning of language. We demonstrate this
by showing that the same measures that replicate implicit bias also replicate prejudicial hiring practices, and further return
veridical information about employment and naming practices in contemporary America.
We ensure impartiality in our approach by using the benchmarks and keywords established in well-known and heavily cited
works of the human sciences, psychology and sociology. We use a state-of-the-art and widely used word embedding, namely
GloVe, made available by Pennington et al. (2014). We used one of GloVe’s standard semantic models trained on standard
corpora of ordinary language use found on the World Wide Web. We have also found similar results for other standard tools
and corpora, which we will also discuss in future versions of this paper’s appendices.
Following the lead of the IAT, for each result we report the two sets of target concepts about which we are attempting to
learn and the two sets of attribute words we are comparing each to. We then apply our method WEAT (defined in the Methods
section), and report the probability (p-value) that our observed similarity scores could have arisen with no semantic association
between the target concepts and the attribute. We report an effect size based on the number of standard deviations that separate
the two sets of target words in terms of their association with the attribute words; precise details of this measure are described
in the Methods section.
Baseline: Replication of Associations That Are Universally Accepted
The first results presented in the initial publication on the IAT (Greenwald et al., 1998) concerned biases that were found to be
universal in humans and about which there is no social concern. This allows the introduction and clarification of the method
and its validation on relatively morally neutral topics. We begin by replicating these inoffensive results for the same purposes.
Flowers and Insects
Original Finding: Greenwald et al. (1998, p. 1469) report that the IAT is able to demonstrate via reaction times that flowers
are significantly more pleasant than insects, and insects more unpleasant than flowers. Based on the reaction latencies of 32
participants, the IAT results in an effect size1 of 1.35, which is considered a large effect, and a p-value of 10−8 for statistical
significance.
We replicate this finding by looking at semantic similarity in GloVe for the same stimuli by using our WEAT
method. Flowers are more likely than insects to be closer to pleasant than to unpleasant. By applying our method, we observe
the expected association with an effect size of 1.50 and with p-value < 10−7 for statistical significance.2
Notice that GloVe “knows” this property of flowers and insects with no direct experience of the world, and no representation
of semantics other than the implicit metrics of words’ co-occurrence statistics that it is trained on.
Our Finding:
1 Effect size is Cohen’s d, which is the log-transformed mean of latencies in milliseconds divided by the standard deviation. Conventional small, medium,
and large values of d are 0.2, 0.5, and 0.8, respectively.
2 The maximum effect size can be is 2.0.
3/14
Outdated draft. See published version at http://opus.bath.ac.uk/55288/.
Stimuli:
We use the same stimuli as Greenwald et al. (1998, p. 1479).
• Flowers: aster, clover, hyacinth, marigold, poppy, azalea, crocus, iris, orchid, rose, bluebell, daffodil, lilac, pansy, tulip,
buttercup, daisy, lily, peony, violet, carnation, gladiola, magnolia, petunia, zinnia.
• Insects: ant, caterpillar, flea, locust, spider, bedbug, centipede, fly, maggot, tarantula, bee, cockroach, gnat, mosquito,
termite, beetle, cricket, hornet, moth, wasp, blackfly, dragonfly, horsefly, roach, weevil.
• Pleasant: caress, freedom, health, love, peace, cheer, friend, heaven, loyal, pleasure, diamond, gentle, honest, lucky,
rainbow, diploma, gift, honor, miracle, sunrise, family, happy, laughter, paradise, vacation.
• Unpleasant: abuse, crash, filth, murder, sickness, accident, death, grief, poison, stink, assault, disaster, hatred, pollute,
tragedy, divorce, jail, poverty, ugly, cancer, kill, rotten, vomit, agony, prison.
Musical Instruments and Weapons
Original Finding: Similarly, Greenwald et al. (1998, p. 1469) find that musical instruments are significantly more pleasant
than weapons. Based on the reaction latencies of 32 participants, the IAT results in an effect size of 1.66 and a p-value of 10−10 .
We replicate this finding by looking at semantic nearness in GloVe for the same stimuli. Musical instruments
are more likely than weapons to be closer to pleasant than to unpleasant. By applying our method, we observe the expected
association with an effect size of 1.53 and with p-value < 10−7 .
Our Finding:
Stimuli:
We use the same stimuli found in Greenwald et al. (1998, p. 1479).
• Musical instruments: bagpipe, cello, guitar, lute, trombone, banjo, clarinet, harmonica, mandolin, trumpet, bassoon,
drum, harp, oboe, tuba, bell, fiddle, harpsichord, piano, viola, bongo, flute, horn, saxophone, violin.
• Weapons: arrow, club, gun, missile, spear, axe, dagger, harpoon, pistol, sword, blade, dynamite, hatchet, rifle, tank,
bomb, firearm, knife, shotgun, teargas, cannon, grenade, mace, slingshot, whip.
• Pleasant and Unpleasant: as per previous experiment with insects and flowers.
Racial Biases
We now use the same technique to demonstrate that machine learning absorbs prejudice as easily as other biases. Here we
replicate not only the original IAT results on racial prejudice, but also the more recent and striking finding that names alone
have enormous impact on the probability of job candidates being called for an interview.
Replicating Implicit Associations for Valence
Original Finding: Greenwald et al. (1998, p. 1475) find extreme impacts of race as indicated simply by name. A bundle of
names associated with being European American was found to be significantly easier to associate with pleasant than unpleasant
terms, compared to a bundle of African American names. With 26 subjects, Greenwald et al. show that the European American
names are more likely to be implicitly associated as pleasant with an effect size of 1.17 and a p-value of 10−6 .
Our Finding: We were again able to replicate attitude towards two races by looking at semantic nearness in GloVe. We were
forced to slightly alter the stimuli because some of the original African American names did not occur in the corpus with
sufficient frequency. These are shown in italics below. We therefore also deleted the same number of European American
names, chosen at random, to balance the number of elements in the sets of two concepts. In our results, European American
names are more likely than African American names to be closer to pleasant than to unpleasant, with an effect size of 1.41 and
p-value < 10−8 .
Stimuli: We use a subset (see above) of the same stimuli found in Greenwald et al. (1998, p. 1479). Names that are marked
with italics are excluded from our replication.
• European American names: Adam, Chip, Harry, Josh, Roger, Alan, Frank, Ian, Justin, Ryan, Andrew, Fred, Jack,
Matthew, Stephen, Brad, Greg, Jed, Paul, Todd, Brandon, Hank, Jonathan, Peter, Wilbur, Amanda, Courtney, Heather,
Melanie, Sara, Amber, Crystal, Katie, Meredith, Shannon, Betsy, Donna, Kristin, Nancy, Stephanie, Bobbie-Sue, Ellen,
Lauren, Peggy, Sue-Ellen, Colleen, Emily, Megan, Rachel, Wendy (deleted names in italics).
• African American names: Alonzo, Jamel, Lerone, Percell, Theo, Alphonse, Jerome, Leroy, Rasaan, Torrance, Darnell,
Lamar, Lionel, Rashaun, Tvree, Deion, Lamont, Malik, Terrence, Tyrone, Everol, Lavon, Marcellus, Terryl, Wardell,
Aiesha, Lashelle, Nichelle, Shereen, Temeka, Ebony, Latisha, Shaniqua, Tameisha, Teretha, Jasmine, Latonya, Shanise,
Tanisha, Tia, Lakisha, Latoya, Sharise, Tashika, Yolanda, Lashandra, Malika, Shavonn, Tawanda, Yvette (deleted names
in italics).
4/14
Outdated draft. See published version at http://opus.bath.ac.uk/55288/.
• Pleasant: caress, freedom, health, love, peace, cheer, friend, heaven, loyal, pleasure, diamond, gentle, honest, lucky,
rainbow, diploma, gift, honor, miracle, sunrise, family, happy, laughter, paradise, vacation.
• Unpleasant: abuse, crash, filth, murder, sickness, accident, death, grief, poison, stink, assault, disaster, hatred, pollute,
tragedy, bomb, divorce, jail, poverty, ugly, cancer, evil, kill, rotten, vomit.
Replicating the Bertrand and Mullainathan (2004) Résumé Study
Original Finding: Bertrand and Mullainathan (2004) sent nearly 5,000 identical résumés to 1,300 job advertisements with
only one change made to the résumés: the names of the candidates. They found that European American candidates were 50%
more likely to be offered an opportunity to be interviewed.
Our Finding: Perhaps unsurprisingly, we again found a significant result for the names used by Bertrand and Mullainathan.
As before, we had to delete some low-frequency names. We also assumed semantic nearness to pleasantness as the correlate for
an invitation to interview. We did this with two different sets of ‘pleasant/unpleasant’ stimuli: those from the original IAT paper,
and also a revised shorter set used more recently, found in Nosek et al. (2002a). For both sets of attributes, European American
names are more likely than African American names to be invited for interviews (closer to pleasant than to unpleasant). Using
the Greenwald et al. (1998) attributes, the effect size is 1.50 and p-value < 10−4 ; and using the updated Nosek et al. (2002a)
attributes, the effect size is 1.28 and p-value < 10−3 .
For the names we use the stimuli found in Bertrand and Mullainathan (2004, p. 1012). The first set of pleasant and
unpleasant words are as per above, the second are from Nosek et al. (2002a, p. 114).
Stimuli:
• European American names: Brad, Brendan, Geoffrey, Greg, Brett, Jay, Matthew, Neil, Todd, Allison, Anne, Carrie,
Emily, Jill, Laurie, Kristen, Meredith, Sarah (deleted names in italics).
• African American names: Darnell, Hakim, Jermaine, Kareem, Jamal, Leroy, Rasheed, Tremayne, Tyrone, Aisha, Ebony,
Keisha, Kenya, Latonya, Lakisha, Latoya, Tamika, Tanisha (deleted names in italics).
• First set of Pleasant and Unpleasant: as per previous experiment with African American and European American
names.
• updated Pleasantness: joy, love, peace, wonderful, pleasure, friend, laughter, happy.
• updated Unpleasantness: agony, terrible, horrible, nasty, evil, war, awful, failure.
Gender Biases
We now turn to gender-related biases and stereotypes. We begin by returning to prejudice as demonstrated by the IAT, but then
we will turn to matching the biases we data mine against veridical information taken from published U.S. government statistics.
Replicating Implicit Associations for Career and Family
Whether or not it is appropriate for women to have careers has been a matter of significant cultural dispute. Historically, the
consensus was that they should not; today, most but by no means all Americans consider it as appropriate for a woman to have
a career as a man. Similarly, there have been historical biases against men who choose to take domestic roles. The IAT study
we compare to here was conducted online, and thus has a vastly larger subject pool. However, since there is more difficulty
ensuring that online subjects will complete their task with attention, it also has far fewer keywords examined. We are able to
replicate the results even with these reduced keyword sets.
With 38,797 interpretable subjects (those who fully completed the test), female names were found to be
more associated with family than career words with an effect size of 0.72 and p-value < 10−2 , Nosek et al. (2002a, p. 105).
Original Finding:
Our Finding: We found the same result that females are more associated with family and males with career, with an effect
size of 1.81 and p-value < 10−3 .
Stimuli:
We use the same stimuli found in Nosek et al. (2002a, p. 114).
• Male names: John, Paul, Mike, Kevin, Steve, Greg, Jeff, Bill.
• Female names: Amy, Joan, Lisa, Sarah, Diana, Kate, Ann, Donna.
• Career words : executive, management, professional, corporation, salary, office, business, career.
• Family words : home, parents, children, family, cousins, marriage, wedding, relatives.
5/14
Outdated draft. See published version at http://opus.bath.ac.uk/55288/.
Replicating Implicit Associations for Arts and Mathematics
In a similar result, both Nosek et al. and we find that female terms are more associated with arts than mathematics, compared to
male terms.
Original Finding: 28,108 subjects completed the online IAT and female terms were more associated with arts than mathematics
with an effect size of 0.82 and p-value < 10−2 , Nosek et al. (2002a, p. 105).
Our Finding:
Stimuli:
We found the expected association with an effect size of 1.06 and a p-value of 10−2 .
We use the stimuli found in Nosek et al. (2002a, p. 114).
• Math words : math, algebra, geometry, calculus, equations, computation, numbers, addition.
• Arts Words : poetry, art, dance, literature, novel, symphony, drama, sculpture.
• Male attributes: male, man, boy, brother, he, him, his, son.
• Female attributes: female, woman, girl, sister, she, her, hers, daughter.
Replicating Implicit Associations for Arts and Sciences
In another laboratory study, Nosek et al. (2002b) found that female terms are less associated with the sciences, and male terms
less associated with the arts.
83 subjects took the IAT with a combination of math/science and art/language attributes, and the expected
associations were observed with an effect size of 1.47 and a p-value of 10−24 , Nosek et al. (2002b, p. 51).
Original Finding:
By examining only arts and sciences attributes, we found that female terms were associated more with arts and
male terms with science with an effect size of 1.24 and a p-value of 10−2 .
Our Finding:
Stimuli:
We use the stimuli found in Nosek et al. (2002b, p. 59).
• Science words : science, technology, physics, chemistry, Einstein, NASA, experiment, astronomy.
• Arts words : poetry, art, Shakespeare, dance, literature, novel, symphony, drama.
• Male attributes: brother, father, uncle, grandfather, son, he, his, him.
• Female attributes: sister, mother, aunt, grandmother, daughter, she, hers, her.
Comparison to Real-World Data: Occupational Statistics
It has been suggested that implicit gender-occupation biases are linked to gender gaps in occupational participation (Nosek
et al., 2009); however the relationship between these is complex and may be mutually reinforcing. Here we examine the
correlation between the gender association of occupation words and labor-force participation data.
The x-axis of Figure 1 is derived from the 2015 U.S. Bureau of Labor Statistics3 , which provides information
about occupational categories and the percentage of women that have certain occupations under these categories. We generated
single word occupation names (as explained in the Methods section) based on the available data and calculated the percentage
of women for the set of single word occupation names.
Original Data:
By applying WEFAT, we are able to use word embeddings to predict the percentage of women in the 50 most
relevant occupations with a Pearson’s correlation coefficient of ρ = 0.90 with p-value < 10−18 .
Our Finding:
We use the gender stimuli found in Nosek et al. (2002a, p. 114) along with the occupation attributes we derived from
labor statistics.
Stimuli:
• Careers : technician, accountant, supervisor, engineer, worker, educator, clerk, counselor, inspector, mechanic, manager,
therapist, administrator, salesperson, receptionist, librarian, advisor, pharmacist, janitor, psychologist, physician, carpenter,
nurse, investigator, bartender, specialist, electrician, officer, pathologist, teacher, lawyer, planner, practitioner, plumber,
instructor, surgeon, veterinarian, paramedic, examiner, chemist, machinist, appraiser, nutritionist, architect, hairdresser,
baker, programmer, paralegal, hygienist, scientist.
• Female attributes: female, woman, girl, sister, she, her, hers, daughter.
• Male attributes: male, man, boy, brother, he, him, his, son.
3 http://www.bls.gov/cps/cpsaat11.htm
6/14
Strength of association of
occupation word vector with female gender
Outdated draft. See published version at http://opus.bath.ac.uk/55288/.
2
1
0
−1
−2
0
20
40
80
100
60
Percentage of workers in occupation who are women
Figure 1. Occupation-gender association
Pearson’s correlation coefficient ρ = 0.90 with p-value < 10−18 .
Comparison to Real-World Data: Androgynous Names
Similarly, we looked at the veridical association of gender to androgynous names, that is, names sometimes used by either
gender. In this case, the most recent information we were able to find was the 1990 census name and gender statistics. Perhaps
because of the age of our name data, our correlation was weaker than for the 2015 occupation statistics, but still strikingly
significant.
Strength of association of
name vector with female gender
2
1
0
−1
−2
0
20
40
80
100
60
Percentage of people with name who are women
Figure 2. People with androgynous names
Pearson’s correlation coefficient ρ = 0.84 with p-value < 10−13 .
The x-axis of Figure 2 is derived from the 1990 U.S. census data4 that provides first name and gender
information in population.
Original Data:
Our Finding: The y-axis reflects our calculation of the bias for how male or female each of the names is. By applying WEFAT,
we are able to predict the percentage of people with a name who were women with Pearson’s correlation coefficient of ρ = 0.84
4 http://www.census.gov/main/www/cen1990.html
7/14
Outdated draft. See published version at http://opus.bath.ac.uk/55288/.
Figure 3. A 2D projection (first
two principal components) of the
300-dimensional vector space of
the GloVe word embedding
(Pennington et al., 2014). The lines
illustrate algebraic relationships
between related words: pairs of
words that differ only by gender
map to pairs of vectors whose
vector difference is roughly
constant. Similar algebraic
relationships have been shown for
other semantic relationships, such
as countries and their capital cities,
companies and their CEOs, or
simply different forms of the same
word.
with p-value < 10−13 .
Stimuli: We use the gender stimuli found in Nosek et al. (2002a, p. 114) along with the most popular androgynous names
from 1990’s public census data as targets.
• Names : Kelly, Tracy, Jamie, Jackie, Jesse, Courtney, Lynn, Taylor, Leslie, Shannon, Stacey, Jessie, Shawn, Stacy, Casey,
Bobby, Terry, Lee, Ashley, Eddie, Chris, Jody, Pat, Carey, Willie, Morgan, Robbie, Joan, Alexis, Kris, Frankie, Bobbie,
Dale, Robin, Billie, Adrian, Kim, Jaime, Jean, Francis, Marion, Dana, Rene, Johnnie, Jordan, Carmen, Ollie, Dominique,
Jimmie, Shelby.
• Female and Male attributes: as per previous experiment on occupations.
Methods
Data and training
A word embedding is a representation of words as points in a vector space. Loosely, embeddings satisfy the property that vectors
that are close to each other represent semantically “similar” words. Word embeddings derive their power from the discovery
that vector spaces with around 300 dimensions suffice to capture most aspects of similarity, enabling a computationally tractable
representation of all or most words in large corpora of text (Bengio et al., 2003; Lowe, 1997). Starting in 2013, the word2vec
family of word embedding techniques has gained popularity due to a new set of computational techniques for generating word
embeddings from large training corpora of text, with superior speed and predictive performance in various natural-language
processing tasks (Mikolov et al., 2013; Mikolov and Dean, 2013).
Most famously, word embeddings excel at solving “word analogy” tasks because the algebraic relationships between
vectors capture syntactic and semantic relationships between words (Figure 3). In addition, word embeddings have found use
in natural-language processing tasks such as web search and document classification. They have also found use in cognitive
science for understanding human memory and recall (Zaromb et al., 2006; McDonald and Lowe, 1998).
For all results in this paper we use the state-of-the-art GloVe word embedding method, in which, at a high level, the similarity
between a pair of vectors is related to the probability that the words co-occur close to each other in text (Pennington et al., 2014).
Word embedding algorithms such as GloVe substantially amplify the signal found in simple co-occurrence probabilities using
dimensionality reduction. In pilot-work experiments along the lines of those presented here (on free associations rather than
implicit associations) raw co-occurrence probabilities were shown to lead to substantially weaker results (Macfarlane, 2013).
Rather than train the embedding ourselves, we use pre-trained GloVe embeddings distributed by its authors. We aim to
replicate the effects that may be found in real applications to the extent possible, and using pre-trained embeddings minimizes
8/14
Outdated draft. See published version at http://opus.bath.ac.uk/55288/.
the choices available to us and simplifies reproducing our results. We pick the largest of the four corpora for which the
GloVe authors provide trained embeddings, which is a “Common Crawl” corpus obtained from a large-scale crawl of the web,
containing 840 billion tokens (roughly, words). Tokens in this corpus are case-sensitive and there are 2.2 million different ones,
each corresponding to a 300-dimensional vector. The large size of this corpus and the resulting model is important to us, since
it enables us to find word vectors for even relatively uncommon names. An important limitation is that there are no vectors for
multi-word phrases.
We can expect similar results to the ones presented here if we used other corpora and/or embedding algorithms. For example,
we repeated all the WEAT and WEFAT experiments presented above using a different pre-trained embedding: word2vec on
a Google News corpus (Mikolov and Dean, 2013). In all experiments, we observed statistically significant effects and high
effect sizes. Further, we found that the gender association strength of occupation words is highly correlated between the GloVe
embedding and the word2vec embedding (Pearson ρ = 0.88; Spearman ρ = 0.86). In concurrent work, Bolukbasi et al. (2016)
compared the same two embeddings, using a different measure of the gender bias of occupation words, also finding a high
correlation (Spearman ρ = 0.81).
Word Embedding Association Test (WEAT)
To demonstrate and quantify bias, we use the permutation test. Borrowing terminology from the IAT literature, consider two sets
of target words (e.g., programmer, engineer, scientist, ... and nurse, teacher, librarian, ...) and two sets of attribute words
(e.g., man, male, ... and woman, female ...). The null hypothesis is that there is no difference between the two sets of target
words in terms of their relative similarity to the two sets of attribute words. The permutation test measures the (un)likelihood of
the null hypothesis by computing the probability that a random permutation of the attribute words would produce the observed
(or greater) difference in sample means.
In formal terms, let X and Y be two sets of target words of equal size, and A, B the two sets of attribute words. Let cos(~a,~b)
denote the cosine of the angle between the vectors ~a and ~b.
• The test statistic is
s(X,Y, A, B) =
∑ s(x, A, B) − ∑ s(y, A, B)
x∈X
y∈Y
where
s(w, A, B) = meana∈A cos(~w,~a) − meanb∈B cos(~w,~b)
In other words, s(w, A, B) measures the association of the word w with the attribute, and s(X,Y, A, B) measures the
differential association of the two sets of target words with the attribute.
• Let {(Xi ,Yi )}i denote all the partitions of X ∪Y into two sets of equal size. The one-sided p-value of the permutation test
is
Pri [s(Xi ,Yi , A, B) > s(X,Y, A, B)]
• The effect size is
meanx∈X s(x, A, B) − meany∈Y s(y, A, B)
std-devw∈X∪Y s(w, A, B)
It is a normalized measure of how separated the two distributions (of associations between the target and attribute) are.
We re-iterate that these p-values and effect sizes don’t have the same interpretation as the IAT. The “subjects” in our
experiments are words, not people. While the IAT can measure the differential association between a single pair of target
concepts and an attribute, the WEAT can only measure the differential association between two sets of target concepts and an
attribute.
Word Embedding Factual Association Test (WEFAT)
To understand and demonstrate the necessity of human bias in word embeddings, we also wish to examine how word embeddings
capture empirical information about the world, which is also embedded in language. Consider a set of target concepts, such as
occupations, and a real-valued, factual property of the world associated with each concept, such as the percentage of workers in
the occupation who are women. We’d like to test if the vectors corresponding to the concepts embed knowledge of the property,
that is, if there is an algorithm that can extract or predict the property given the vector. In principle we could use any algorithm,
but in this work we test the association of the target concept with some set of attribute words, analogous to WEAT above.
Formally, consider a single set of target words W and two sets of attribute words A, B. There is a property pw associated
with each word w ∈ W .
9/14
Outdated draft. See published version at http://opus.bath.ac.uk/55288/.
• The statistic associated with each word vector is a normalized association score of the word with the attribute:
s(w, A, B) =
meana∈A cos(~w,~a) − meanb∈B cos(~w,~b)
std-devx∈A∪B cos(~w,~x)
• The null hypothesis is that there is no association between s(w, A, B) and pw . We test the null hypothesis using a linear
regression analysis to predict the latter from the former.
Now we discuss in more detail how we apply WEFAT in two cases. The first is to test if occupation word vectors embed
knowledge of the gender composition of the occupation in the real world. We use data released by the Bureau of Labor Statistics
in which occupations are categorized hierarchically, and for each occupation the number of workers and percentage of women
are given (some data is missing). The chief difficulty is that many occupation names are multi-word terms whereas word vectors
represent single words. Our strategy is to convert a multi-word term into a single word that represents a superset of the category
(e.g., chemical engineer → engineer), and filter out occupations where this is not possible.
Our second application of WEFAT is to test if androgynous names embed knowledge of how often the name is given to
boys versus girls. We picked the most popular names in each 10% window of gender frequency based on 1990 U.S. Census
data. Here again there is a difficulty: some names are also regular English words (e.g., Will). State-of-the-art word embeddings
are not yet sophisticated enough to handle words with multiple senses or meanings; all usages are lumped into a single vector.
To handle this, we algorithmically determine how “name-like” each vector is (by computing the distance of each vector to the
centroid of all the name vectors), and eliminate the 20% of vectors that are least name-like.
We plan to make the code used to generate our results publicly available.5
Discussion
We have shown that machine learning can acquire prejudicial biases from training data that reflect historical injustice. This is
not entirely a new finding. A recent line of work on fairness in machine learning tries to minimize or avoid such biases (Dwork
et al., 2012; Feldman et al., 2015; Zemel et al., 2013; Barocas and Selbst, 2014). However, unlike this literature, our setting is
not a particular, explicit decision-making task (known as “classification” in machine learning), but rather the often unconscious
consequences of all of language. We show for the first time that if AI is to exploit via our language the vast knowledge that
culture has compiled, it will inevitably inherit human-like prejudices. In other words, if AI learns enough about the properties
of language to be able to understand and produce it, it also acquires cultural associations that can be offensive, objectionable, or
harmful. These are much broader concerns than intentional discrimination, and possibly harder to address. This distinction
informs much of the rest of this section.
Implications for understanding human prejudice
The simplicity and strength of our results suggests a new null hypothesis for explaining origins of prejudicial behavior in
humans, namely, the implicit transmission of ingroup/outgroup identity information through language. That is, before providing
an explicit or institutional explanation for why individuals make decisions that disadvantage one group with regards to another,
one must show that the unjust decision was not a simple outcome of unthinking reproduction of statistical regularities absorbed
with language. Similarly, before positing complex models for how prejudicial attitudes perpetuate from one generation to
the next or from one group to another, we must check whether simply learning language is sufficient to explain the observed
transmission of prejudice. These new null hypotheses are important not because we necessarily expect them to be true in
most cases, but because Occam’s razor now requires that we eliminate them, or at least quantify findings about prejudice in
comparison to what is explainable from language transmission alone.
Our work lends credence to the highly parsimonious theory that all that is needed to create prejudicial discrimination is not
malice towards others, but preference for one’s ingroup (Greenwald and Pettigrew, 2014). This theory is also supported by
recent results showing that in times of conflict, rather than an increase in ingroup altruism, we see a decrease in baseline altruism
towards the outgroup (Silva and Mace, 2015). Our results also both explain and support empirical results from education
indicating that reducing prejudice requires directed interventions to facilitate “decategorizing and recategorizing outgroups”
(Dessel, 2010, p. 411). Simple contact with members of other groups is not enough. There needs to be specific bridging
experiences to facilitating the construction of new identities or to develop skills to work with people across group boundaries.
It has been known for some time that even newborn infants attend foremost to speakers sharing their mother’s dialect
(Kinzler et al., 2007); it has been conjectured that such ingroup signaling may even account for the origins of music and
language (Fitch, 2004). What we have shown here is that language identifies not only one’s own group, but also which group
5 We
thank Will Lowe for assistance with the methods WEAT and WEFAT that greatly improved the methodology.
10/14
Outdated draft. See published version at http://opus.bath.ac.uk/55288/.
is currently culturally dominant, or dominates particular regions of a culture. This may account for why in the IAT, Koreans
and Japanese people living in their own countries each associate the other as ‘less pleasant,’ but African Americans show
European-American-oriented biases, though not as strongly as European Americans (Greenwald et al., 1998). The dominance
of European-American orientation may change as the American demography changes; indeed it would be interesting to examine
corpora consisting of newspapers or other public language in towns or cities with different demographic makeup, particularly
where racial diversity is also represented consistently in public offices and media.
Of course, neither our work nor any other theory explaining the origins of prejudice justify prejudiced behavior. Humans
are (or can be) good at using explicit knowledge to better cooperate, including choosing to behave fairly. Lee (2016) has shown
very recently that the level of implicit bias displayed by subjects in the IAT does not predict cooperative performance. In other
words, the learned biases that affect rate of comprehension of test stimuli or construction of artificial pairings does not affect
deliberate choices about how to treat others, at least not in a laboratory setting. However, we have demonstrated here that
a known case of prejudicial decision making (about inviting job candidates, cf. Bertrand and Mullainathan, 2004) can be
replicated by biases latent in language. We therefore recommend continuing the program of research examining behavior that
does and does not correlate to human subject performance in the IAT. We recommend using our text processing tools to check
pilot predictions for likely IAT performance on comparisons where none is currently known.
Consequences of bias in humans and machines
We have shown that AI can and does inherit substantially the same biases that humans exhibit. However, the consequences of
bias are different in humans and machine-learning systems. Bias in AI is important because AI is increasingly given agency in
our society, for tasks ranging from predictive text in search to determining criminal sentences assigned by courts. Yet machines
are artifacts, owned and controlled by humans operators. That means that learning can be shut off completely once a product is
put into production or operation, and this is frequently done to create more efficient and uniform experiences. Such an approach
opens a potential downside: we may enshrine an imperfect procedure in a context where it will not be routinely reexamined by
other humans. Such artifacts could persist and perpetuate biases in society for a long time — digital analogs of Robert Moses’s
racially motivated overpasses (Winner, 1980).
One advantage of AI, at least where the algorithms and outcomes are open to inspection, is that it can at least make such
errors explicit and therefore potentially subject to monitoring and correction. After all, the same dependencies on history we
have uncovered here may very well also pollute individual expectations, public policy, and even law. Natural intelligence and
learning, just as in artifacts, may pick upon correlations without considering sufficiently carefully whether there is any causal
relationship, or whether the correlation is caused by some other unobserved factor, possibly a correctable injustice.
Effects of bias in NLP applications
To better understand the potential impact of bias in word embeddings, let us consider applications where they have found
use. Sentiment analysis classifies text as being positive, negative, or neutral. Two of its uses are in marketing to quantify
customer satisfaction (say, from a set of product reviews) and in finance to predict market trends (say, from tweets about
companies). Consider a straw-man sentiment analysis technique based on word embeddings: calculate the valence of each
word based on its association with designated positive and negative words, then sum up the sentiment scores. Now consider
applying this technique to movie reviews. Our results show that European-American names have more positive valence than
African-American names in a state-of-the-art word embedding. That means a sentence containing a European-American name
will have a higher sentiment score than a sentence with that name replaced by an African-American name. In other words, the
tool will display a racial bias in its output based on actor and character names.
We picked this example because the argument follows directly from our experiments on names. But our results suggest that
other imprints of human racial prejudice, not confined to names, will also be picked up by machine-learning models. Besides,
bias is known to creep in indirectly, by proxy (Barocas and Selbst, 2014). Thus, it would be simplistic to conclude that we can
fix the problem by withholding names from the inputs to NLP applications.
Next, consider statistical machine translation (SMT). Unsurprisingly, today’s SMT systems reflect existing gender stereotypes. Translations to English from many gender-neutral languages such as Finnish, Estonian, Hungarian, Persian, and Turkish
lead to gender-stereotyped sentences. For example, Google Translate converts these Turkish sentences with genderless pronouns:
“O bir doktor. O bir hemşire.” to these English sentences: “He is a doctor. She is a nurse.” A test of the 50 occupation words
used in the results presented in Figure 1 shows that the pronoun is translated to “he” in the majority of cases and “she” in about
a quarter of cases; tellingly, we found that the gender association of the word vectors almost perfectly predicts which pronoun
will appear in the translation.
Challenges in addressing bias
Redresses such as transparent development of AI technology and improving diversity and ethical training of developers, while
useful, do little to address the kind of prejudicial bias we expose here. Unfortunately, our work points to several additional
11/14
Outdated draft. See published version at http://opus.bath.ac.uk/55288/.
reasons why addressing bias in machine learning will be harder than one might expect. First, our results suggest that word
embeddings don’t merely pick up specific, enumerable biases such as gender stereotypes (Bolukbasi et al., 2016), but rather the
entire spectrum of human biases reflected in language. In fact, we show that bias is meaning. Bias is identical to meaning, and
it is impossible to employ language meaningfully without incorporating human bias. This is why we term unacceptable bias
prejudice in this paper. The biases we reveal aren’t about a particular application of machine learning, but rather about the basic
representation of knowledge — used possibly in human cognition, and certainly in an expanding variety of AI applications.
Second, the idea of correcting even prejudiced biases is also problematic. That is because societal understanding of prejudice
is constantly evolving, along with our understanding of humanity and human rights, and also varies between cultures. It is
therefore hard or impossible to specify algorithmically what is prejudiced. To give one example, Monteith and Pettit (2011)
using the IAT show that people with mental illnesses are stigmatized compared to people with physical illnesses — a result we
have also replicated in word embeddings (but not reported above). Is this a prejudice? Who determines whether it should be
corrected?
Third and finally, we have shown that biases result from extant as well as historic inequalities in the world. There may
be many other contexts where these inequalities are important to know about. More generally, shared awareness of the real
world is important for communication (Zue, 1985; Barsalou, 2009). Consider the gender stereotypes in occupations. If we
were using machine learning to evaluate the suitability of job applicants, these stereotypes would be bad. Yet if the task was
to analyze historical job ads and infer if more men or women worked in those roles, the stereotypical associations would be
exactly the information we would wish to utilize. The gender associations we found in the word embeddings of names might
be exceedingly useful, yet those same associations might lead to prejudicial expectations concerning names and occupations.
Remedies must be tailored to applications. Within a given context, such as college admissions, we can decide whether (and to
what extent) considerations of fairness should override the usual focus on predictive accuracy (as is the case with affirmative
action), but it is not meaningful to do this devoid of context. Put simply, eliminating bias is eliminating information; eliminating
prejudice takes more thought.
Awareness is better than blindness
For these reasons, we view the approach of “debiasing” word embeddings (Bolukbasi et al., 2016) with skepticism. If we view
AI as perception followed by action, debiasing alters the AI’s perception (and model) of the world, rather than how it acts on
that perception. This gives the AI an incomplete understanding of the world. We see debiasing as “fairness through blindness”.6
It has its place, but also important limits: prejudice can creep back in through proxies (although we should note that Bolukbasi
et al. (2016) do consider “indirect bias” in their paper). Efforts to fight prejudice at the level of the initial representation will
necessarily hurt meaning and accuracy, and will themselves be hard to adapt as societal understanding of fairness evolves.
Instead, we take inspiration from the fact that humans can express behavior different from their implicit biases (Lee, 2016).
Human intelligence is typified by behavior integrating multiple forms of memory and evidence (Purcell and Kiani, 2016; Bear
and Rand, 2016). It includes the capacity to recall one-shot exposure to highly context-specific information in the form of rules
and instructions. We can learn that “prejudice is bad”, that “women used to be trapped in their homes and men in their careers,
but now gender doesn’t necessarily determine family role” and so forth. If AI is not built in a similar way, then it would be
possible for prejudice absorbed by machine learning to have a much greater negative impact than when prejudice is absorbed in
the same way by children. This is because children also receive other kinds of instruction and social examples as a part of the
ordinary, painstaking process of child rearing. Normally when we design AI architectures, we try to keep them as simple as
possible to facilitate our capacity to debug and maintain AI systems. However, where AI is partially constructed automatically
by machine learning of human culture, we may also need an analog of human explicit memory and deliberate actions, that can
be trained or programmed to avoid the expression of prejudice.
Of course, such an approach doesn’t lend itself to a straightforward algorithmic formulation. Instead it requires a long-term,
interdisciplinary research program that includes cognitive scientists and ethicists. One concrete suggestion for the present is to
choose corpora for training machine learning to have as little prejudice as possible — the tools we have presented here can be
used to identify these. Another is that given the vulnerability of relying on purely statistical information for understanding and
operating within a culture, it may be advisable to consider more complex AI architectures such as cognitive systems (Thórisson,
2007; Hanheide et al., 2015). Heterogeneous approaches to representing knowledge and intelligence may allow us to exploit
both the great strengths of machine learning and the instructability of symbolic systems.
Acknowledgements
We are grateful to the following people: Will Lowe for substantial assistance in the design of our significance tests, Tim
Macfarlane for his pilot research as a part of his undergraduate dissertation, Solon Barocas and Miles Brundage for excellent
comments on an early version of this paper.
6 Our
use of this term is inspired by Dwork et al. (2012), but we use it slightly differently, and our argument is different from theirs.
12/14
Outdated draft. See published version at http://opus.bath.ac.uk/55288/.
References
Angwin, J., Larson, J., Mattu, S., and Kirchner, L. (2016). Machine bias: There’s software used across the country to predict
future criminals. and it’s biased against blacks. ProPublica, May, 23.
Barocas, S. and Selbst, A. D. (2014). Big data’s disparate impact. California Law Review, 104.
Barr, A. (2015). Google mistakenly tags black people as ‘gorillas,’ showing limits of algorithms. The New York Times.
Barsalou, L. W. (2009). Simulation, situated conceptualization, and prediction. Philosophical Transactions of the Royal Society
B: Biological Sciences, 364(1521):1281–1289.
Bear, A. and Rand, D. G. (2016). Intuition, deliberation, and the evolution of cooperation. Proceedings of the National Academy
of Sciences, 113(4):936–941.
Bengio, Y., Ducharme, R., Vincent, P., and Jauvin, C. (2003). A neural probabilistic language model. journal of machine
learning research, 3(Feb):1137–1155.
Bertrand, M. and Mullainathan, S. (2004). Are emily and greg more employable than lakisha and jamal? a field experiment on
labor market discrimination. The American Economic Review, 94(4):991–1013.
Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer, London.
Bolukbasi, T., Chang, K.-W., Zou, J., Saligrama, V., and Kalai, A. (2016). Man is to computer programmer as woman is to
homemaker? debiasing word embeddings. arXiv preprint arXiv:1607.06520.
Crawford, K. (2016). Artificial intelligence’s white guy problem. The New York Times.
Dessel, A. (2010). Prejudice in schools: Promotion of an inclusive culture and climate. Education and Urban Society,
42(4):407–429.
Dwork, C., Hardt, M., Pitassi, T., Reingold, O., and Zemel, R. (2012). Fairness through awareness. In Proceedings of the 3rd
Innovations in Theoretical Computer Science Conference, pages 214–226. ACM.
Feldman, M., Friedler, S. A., Moeller, J., Scheidegger, C., and Venkatasubramanian, S. (2015). Certifying and removing
disparate impact. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, pages 259–268. ACM.
Fitch, W. T. (2004). Kin selection and ‘mother tongues’: A neglected component in language evolution. In Oller, D. K. and
Griebel, U., editors, Evolution of Communication Systems: A Comparative Approach, pages 275–296. MIT Press, Cambridge,
MA.
Greenwald, A. G., McGhee, D. E., and Schwartz, J. L. (1998). Measuring individual differences in implicit cognition: the
implicit association test. Journal of personality and social psychology, 74(6):1464.
Greenwald, A. G. and Pettigrew, T. F. (2014). With malice toward none and charity for some: Ingroup favoritism enables
discrimination. American Psychologist, 69(7):669.
Hanheide, M., Göbelbecker, M., Horn, G. S., Pronobis, A., Sjöö, K., Aydemir, A., Jensfelt, P., Gretton, C., Dearden, R., Janicek,
M., Zender, H., Kruijff, G.-J., Hawes, N., and Wyatt, J. L. (2015). Robot task planning and explanation in open and uncertain
worlds. Artificial Intelligence, pages –. in press.
Kiefer, A. K. and Sekaquaptewa, D. (2007). Implicit stereotypes and women’s math performance: How implicit gender-math
stereotypes influence women’s susceptibility to stereotype threat. Journal of Experimental Social Psychology, 43(5):825–832.
Kinzler, K. D., Dupoux, E., and Spelke, E. S. (2007). The native language of social cognition. Proceedings of the National
Academy of Sciences, 104(30):12577–12580.
Lee, D. J. (2016). Racial bias and the validity of the implicit association test. Technical Report 35, Helsinki, Finland.
Lowe, W. (1997). Meaning and the mental lexicon. In Proceedings of the 15th International Joint Conference on Artificial
Intelligence, pages 1092–1097, Nagoya. Morgan Kaufmann.
Macfarlane, T. (2013). Extracting semantics from the enron corpus.
McDonald, S. and Lowe, W. (1998). Modelling functional priming and the associative boost. In Proceedings of the Twentieth
Annual Conference of the Cognitive Science Society, pages 667–680. LEA.
Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013). Efficient estimation of word representations in vector space. arXiv
preprint arXiv:1301.3781.
Mikolov, T. and Dean, J. (2013). Distributed representations of words and phrases and their compositionality. Advances in
neural information processing systems.
Monteith, L. L. and Pettit, J. W. (2011). Implicit and explicit stigmatizing attitudes and stereotypes about depression. Journal
of Social and Clinical Psychology, 30(5):484.
Moss, H. E., Ostrin, R. K., Tyler, L. K., and Marslen-Wilson, W. D. (1995). Accessing different types of lexical semantic
information: Evidence from priming. Journal of Experimental Psychology: Learning, Memory and Cognition, 21:863–883.
Noble, S. U. (2013). Google search: Hyper-visibility as a means of rendering black women and girls invisible. InVisible
Culture, 19.
Nosek, B. A., Banaji, M., and Greenwald, A. G. (2002a). Harvesting implicit group attitudes and beliefs from a demonstration
13/14
Outdated draft. See published version at http://opus.bath.ac.uk/55288/.
web site. Group Dynamics: Theory, Research, and Practice, 6(1):101.
Nosek, B. A., Banaji, M. R., and Greenwald, A. G. (2002b). Math= male, me= female, therefore math6= me. Journal of
personality and social psychology, 83(1):44.
Nosek, B. A., Smyth, F. L., Sriram, N., Lindner, N. M., Devos, T., Ayala, A., Bar-Anan, Y., Bergh, R., Cai, H., Gonsalkorale,
K., et al. (2009). National differences in gender–science stereotypes predict national sex differences in science and math
achievement. Proceedings of the National Academy of Sciences, 106(26):10593–10597.
Oswald, M. and Grace, J. (2016). Norman stanley fletcher and the case of the proprietary algorithmic risk assessment. Policing
Insight.
Pennington, J., Socher, R., and Manning, C. D. (2014). Glove: Global vectors for word representation. In EMNLP, volume 14,
pages 1532–43.
Purcell, B. A. and Kiani, R. (2016). Hierarchical decision processes that operate over distinct timescales underlie choice and
changes in strategy. Proceedings of the National Academy of Sciences, 113(31):E4531–E4540.
Quine, W. V. O. (1960). Word and Object. MIT Press, Cambridge, MA.
Silva, A. S. and Mace, R. (2015). Inter-group conflict and cooperation: field experiments before, during and after sectarian riots
in northern ireland. Frontiers in Psychology, 6:1790.
Stanley, D. A., Sokol-Hessner, P., Banaji, M. R., and Phelps, E. A. (2011). Implicit race attitudes predict trustworthiness
judgments and economic trust decisions. Proceedings of the National Academy of Sciences, 108(19):7710–7715.
Sweeney, L. (2013). Discrimination in online ad delivery. Queue, 11(3):10:10–10:29.
Thórisson, K. R. (2007). Integrated A.I. systems. Minds and Machines, 17(1):11–25.
Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236):433–460.
Winner, L. (1980). Do artifacts have politics? Daedalus, pages 121–136.
Zaromb, F. M., Howard, M. W., Dolan, E. D., Sirotin, Y. B., Tully, M., Wingfield, A., and Kahana, M. J. (2006). Temporal
associations and prior-list intrusions in free recall. Journal of Experimental Psychology: Learning, Memory, and Cognition,
32(4):792.
Zemel, R. S., Wu, Y., Swersky, K., Pitassi, T., and Dwork, C. (2013). Learning fair representations. ICML (3), 28:325–333.
Zue, V. W. (1985). The use of speech knowledge in automatic speech recognition. Proceedings of the IEEE, 73(11):1602–1615.
14/14
| 2cs.AI
|
1
Tree-CNN: A Deep Convolutional Neural
Network for Lifelong Learning
arXiv:1802.05800v1 [cs.CV] 15 Feb 2018
Deboleena Roy, Priyadarshini Panda and Kaushik Roy, Fellow, IEEE
Abstract—In recent years, Convolutional Neural Networks (CNNs) have shown remarkable performance in many computer vision
tasks such as object recognition and detection. However, complex training issues, such as “catastrophic forgetting” and
hyper-parameter tuning, make incremental learning in CNNs a difficult challenge. In this paper, we propose a hierarchical deep neural
network, with CNNs at multiple levels, and a corresponding training method for lifelong learning. The network grows in a tree-like
manner to accommodate the new classes of data without losing the ability to identify the previously trained classes. The proposed
network was tested on CIFAR-10 and CIFAR-100 datasets, and compared against the method of fine tuning specific layers of a
conventional CNN. We obtained comparable accuracies and achieved 40% and 20% reduction in training effort in CIFAR-10 and
CIFAR 100 respectively. The network was able to organize the incoming classes of data into feature-driven super-classes. Our model
improves upon existing hierarchical CNN models by adding the capability of self-growth and also yields important observations on
feature selective classification.
Index Terms—Convolutional Neural Networks, Deep Learning, Incremental Learning, Lifelong Learning, Transfer Learning
F
1
I NTRODUCTION
D
EEP Convolutional Neural Networks (DCNNs) have
emerged as the leading architecture for large scale
image classification in recent years [1]. In 2012, Krizevsky,
et al. [2] won the ImageNet Large Scale Visual Recognition
Challenge (ISLVRC) by implementing a Deep-CNN. This
catapulted DCNNs into the spotlight, and since then, they
have dominated ISLVRC and have surpassed human level
performance on popular image datasets such as MNIST [3],
[4] and ImageNet [5].
Today, with increased access to large amount of labeled data (eg. ImageNet contains 1.2 million images
with 1000 categories), supervised learning has become the
leading paradigm in training DCNNs for image recognition.Traditionally, a DCNN is trained on a dataset containing large number of labeled images. The network learns
to extract relevant features and classify these images. This
trained model is then applied to other unlabeled images to
classify them into these specific classes. In such training,
all the training data is presented to the network during the
same training process. However, in real world, we hardly
have all the information at once. Instead, data is gathered
incrementally over time. And we, humans, also learn in a
similar incremental manner. We accumulate knowledge acquired in the past and use it to learn new things during our
lifetime. Inspired by the way humans learn, Lifelong Machine
Learning (LML) [6] is an emerging paradigm in machine
learning that tries to address this limitation. Lifelong Learning
[7], [8], is based on the principle that learning the nth task
should be easier than learning the (n − 1)th task. While,
intuitively, this makes sense; the actual implementation of
such a network faces several challenges that are discussed
further.
•
Deboleena Roy, Priyadarshini Panda and Kaushik Roy are with Department of Electrical and Computer Engineering, Purdue University, West
Lafayette, IN, 30332.
A DCNN embeds feature extraction and classification in
one coherent architecture within the same model. Modifying
one part of the parameter space immediately affects the
model globally. Another problem of incrementally training
a DCNN is the issue of “catastrophic forgetting” [9]. When
new data is fed into a DCNN, it results in the destruction
of existing features learned from earlier data. This means
when training on new data, all previous data must also be
shown to the network.
Interestingly, it has also been observed that initial layers
of DCNN learn to extract generic features, while deeper
layers learn to identify higher level features. It has been
demonstrated that initializing a new network with transferred features from a trained network can potentially boost
the performance of the new one [10]. “Off the shelf” CNNs
trained on large datasets are then fine-tuned for specific
applications [11]. However, these solutions do not train the
original network to learn the new classes. Rather a new
network is created from the old one.
In this work, we propose a network that grows hierarchically as new classes are introduced. The branching is based
on the similarity of features between new and old classes.
The initial nodes of the Tree-CNN divide the dataset into
coarse super-classes, and as we approach the leaves of the
network, finer classification is done. Such a model allows
us to leverage the convolution layers learned previously
to be used in the new bigger network. Our objective is to
present the advantages of such a network model in terms of
incremental learning over transferring trained convolution
layers and fine tuning a regular network.
The proposed Tree-CNN has the following salient features:
i.
The network is initially trained to classify images
into N classes. When an image belonging to a new
class is shown to the network, the network grows
2
ii.
iii.
iv.
to accommodate the new class.
The network grows by adding a new branch or a
new leaf node to the current structure. The decision
is based on how closely the new class resembles a
particular superclass and its sub-classes.
The objective is to reduce the training effort, which
is made up of 2 components: number of weights
updated, and the number of examples (both old
and new) required for training.
The updates are localized to a section of the tree,
while majority of the Tree-CNN is left untouched,
therefore reducing the cost of learning new classes.
The rest of the paper is organized as follows. The related
work on incremental learning in deep neural networks is
discussed in Section 2. Next, in Section 3 we explain in detail
our proposed network architecture and incremental learning method. In Section 4, we describe the two experiments
we conducted using CIFAR-10 and CIFAR-100 datasets. It
is followed by a detailed analysis of the performance of the
network and its comparison with basic transfer learning and
fine tuning in Section 5. Finally, in Section 6 we discuss the
merits and limitations of our network, conclude our findings
and suggest opportunities for future work.
2
R ELATED W ORK
The modern world of digitized data produces new information every second, thus fueling the need for systems that can
learn as new data arrives. Traditional deep neural networks
are static in that respect, and several new approaches to
incremental learning are currently being explored. “Oneshot learning” [12] is a Bayesian transfer learning technique,
that uses very few training samples to learn new classes.
Fast R-CNN [13], a popular framework for object detection,
also suffers from “catastrophic forgetting”. One way to
mitigate this issue is to use a frozen copy of the original
network compute and balance the loss when new classes
are introduced in the network [14]. "Learning without Forgetting" [15] is another method that uses only new task
data to train the network while preserving the original
capabilities. However, here the original network is trained
on an extensive dataset, such as ImageNet [5], and the new
task data is a much smaller dataset. We, on the other hand,
propose a network that knows only a handful of classes, and
grows and learns over time.
Our work draws inspiration from transfer learning. It
has been observed that initial layers of a CNN learn very
generic features [10]. The same work also demonstrated that
a new network, when initialized with “learned” layers from
a trained network, performs better. Switching initial layers
of deep neural networks with pre-defined Gabor filters have
been shown to reduce training time and offer a more energyefficient training process [16].
Common features shared between images of different
objects has been exploited to build hierarchical classifiers.
These features can be grouped semantically, such as in [17],
in which authors built an “Attention-tree”, a visual semantic
hierarchy to perform energy-efficient image classification.
“FALCON: Feature driven selective classification” [18] is
a classification methodology inspired by biological visual
attention mechanism in the brain. They use characteristic features such as color and texture to break down the
behemoth task of classifying large datasets into a set of
hierarchical classifiers.
Similar to the progression of complexity of convolutional
layers in a DCNN, the upper nodes of a hierarchical CNN
classify the images into coarse super-classes using basic
features, like grouping green-colored objects together, or
humans faces together. Then deeper nodes perform finer
discrimination, such as “boy” v/s “girl” , “apples” v/s
“oranges”, etc. Such hierarchical CNN models have shown
to perform at par or even better than standard DCNNs
[19]. “Discriminative Transfer learning” [20] is one of the
earliest works where classes are categorized hierarchically
to improve network performance. “HD-CNN” [19], is a
hierarchical CNN model that is built by exploiting the
common feature sharing aspect of images. However, in these
works, the dataset is fixed from the beginning, and prior
knowledge of all the classes and their properties is used to
build a hierarchical model. In this work, the Tree-CNN starts
out as a single root node and generates new hierarchies
to accommodate the new classes. Images belonging to the
older dataset are required during retraining, but by localizing the change to a small section of the whole network, we
try to achieve reduction in training effort. In [21], a similar
approach is applied, where the new classes are added to
the old classes, and divided into two super-classes, by using
an error-based model. The initial network is cloned to form
two new networks which are fine tuned over the two new
super-classes. While their motivation was a "divide-andconquer" approach for large datasets, we are interested in
developing a model that can incrementally grow with new
data. And, we sequentially add new data over multiple
learning stages. In the next Section, we lay out in detail our
design principle, network topology and the algorithm used
to grow the network.
3
T HE L IFELONG L EARNING M ODEL
3.1
Network Architecture
The network model is inspired from hierarchical classifiers.
It is built with nodes, connected as a directed acyclic graph.
Each node acts a classifier, predicting a label for the input
image. As per the label, the image is then passed on to the
next node which further classifies the image, until we reach
a leaf node, the last step of classification.
We begin by defining certain terms used frequently in
our model. A Node represents the building block of our
tree. Each node represents a group of 1 or more classes.
Each node has following parameters.
•
•
•
Node-ID: Each node is given a number that uniquely
identifies it.
Parent: This holds the Node-ID of the node previous
to this node. Each node can have only one parent.
A parent node can be inferred as coarse-classifier,
whereas the child node does finer classification. Except for the root node, all other nodes have parent
node. Each node can only have at max one parent
node.
Children: It is an array of Node-IDs belonging to the
nodes linked with each output neuron of the current
3
of the network to reach a specific leaf node. At each node,
beginning with the top node, the image is fed to the DCNN
associated with that node. The output node with the highest
classification probability is the next node the algorithm
moves to. If it is a leaf node, then the class associated with
that node is the predicted class. Else, the algorithm feeds the
image to the DCNN of that node. The pseudo-code for class
prediction is given under Algorithm 1.
IMAGE
Root Node
classifies
input image
into one of
the “superclasses”
Branch Node
Fine classifier
Fig. 1. A generic model of Tree-CNN: The root node predicts superclasses whereas lower nodes predict finer classes
•
•
node. The number of children each node has is equal
to the number of output neurons.
Net: This stores the convolutional neural network
that processes the input image. In case of a leaf node,
this is empty.
LT: Acronym for "Labels Transform", LT is a lookup
table that keeps track of the original label of a class,
and the output neuron number it is associated with.
For a leaf node, LT has just one record, the class label
of that node.
The root node is the highest node of the tree. The first
classification happens at this node. Each output neuron of
the root is associated with a node. Next in hierarchy is the
branch node. It has a parent and two or more children.
It performs classification for at minimum 2 classes/superclasses. The leaf node is the last level of the tree. Each leaf
node is uniquely associated to a class. No two leaf nodes
have the same class. Fig. 1 shows the root node and branch
nodes for a 2 stage classification network. Each output of
the second level branch node is a leaf node.
3.2
Algorithm
To start with, one has two options when implementing the
Tree-CNN. If there are only a handful of distinct classes
that don’t seem like they can be grouped in a hierarchical
fashion, then the network can be initialized as a single root
node. The output of the root node are the distinct classes.
However, if at the beginning we have data that is already
grouped in hierarchical manner as coarse and fine classes,
then a more detailed Tree-CNN can be initialized. The root
node will be trained to classify the images into the highest
level of super-classes. The branch nodes further down the
tree are trained to classify the images into finer classes.
Now that we have a trained Tree-CNN, we discuss how
it predicts classes and how it treats new classes. First, we
describe how the network predicts the class of an input
image. A recursive algorithm moves along the hierarchies
Algorithm 1 Class Prediction
Initialize: Im = Input Image, node = root node of the TreeCNN
procedure C LASS P REDICT(Im,node)
if num(node.children) = 0 then
. Checks if it is a leaf node
return class = node.LT
. for leaf node, LT = class label
else
NextNode = EvaluateNet(Im, node.net)
. EvaluateNet returns the node.ID of the child node
having the largest output
return C LASS P REDICT(Im,NextNode)
end if
end procedure
Next we describe the process of incremental/lifelong
learning. The Tree-CNN is already trained to classify, say,
N classes. Now, data belonging to M new classes have been
acquired and the network needs to learn to classify them
while trying to minimize the change in network structure,
and having a low training effort. A small sample of images
(∼ 10%) is selected from the training set of the new classes.
At the root node these images are fed to the DCNN, one
class at a time. We obtain a 3 dimensional matrix, O K×M×I ,
where
K = number of children of the root node
M = number of new classes
I = number of sample images per class
O(k, m, i) denotes the output of the k th output neuron for
the i th image belonging to the mth class where k ∈ [1, K],
K×M is the average of the outputs
m ∈ [1, M], and i ∈ [1, I]. O avg
over I images . Softmax is taken over O avg to obtain the
likelihood matrix L K×M which indicates how strongly an
image belonging to mth class on average associates to the
k th node.
I
Õ
O(k, m, i)
(1)
O avg (k, m) =
I
i=1
L(k, m) =
eOav g (k,m)
K
Í
eOav g (k,m)
(2)
k=1
Each column of L can be represented as a K × 1 vector, lm ,
where m ∈ [1, M] represents the M new classes. We arrange
lm in an ordered set, S, such that
S = [lm1, lm2, lm3 ...lmM ]
(3)
max(lm1 (k)) >= max(lm2 (k)) >= ... >= max(lmM (k))
k
k
k
(4)
4
We analyze each vector of the set S one by one to
determine how a new class will be appended to the root
node. There are 3 options:
ii.
iii.
FC
CONV 3
CONV 2
CONV 1
FC
FC
CONV 3
CONV 2
“Animal”
CONV 1
IMAGE
32x32x3 RGB
k
Combine one or more child nodes and the new
class to form a new child node: If there are more
than 1 child nodes that the new class has a strong
likelihood for, we can combine them to form a new
child node. Say, the top two likelihood values were
0.48,and 0.45, and at least one of them is a leaf node,
we can combine the two and the new class to form
a new child node which will be a branch node. We
can set an upper limit to the number of child nodes
that could be combined
Add the new class as a new child node: If the new
class doesn’t have a single likelihood value greater
than a threshold (lm (i) < threshold ∀i ∈ [1, K]), or
certain network restrictions are applied to prevent
addition of classes to child nodes, the network
expands horizontally by adding the new class as
a new child node. This node will be a leaf node.
Once this is complete for the M classes at the root node,
we move to the next level of the tree. The same process is
applied on the child nodes that now have new classes to
be added to them. For example, say, two new classes were
added to a child node. If child node is a leaf node, it is
changed into a branch node that now has 3 leaf nodes as
children. If the child node is a branch node, then we repeat
the process of calculating likelihood matrix and determining
how these two new classes will get added to its output. As
we can see, the decision on how to grow the tree is semisupervised. The algorithm itself decides how to grow the
tree, given the constraints by the user. We can limit parameters such as maximum children for a node, maximum
depth for the tree, etc. as per our system requirements.
Once the new classes are allotted locations in the tree, we
do supervised gradient descent based training of the nodes
that have been modified/added. Doing this saves us from
modifying the whole network, and only affected portions of
the network require retraining/fine-tuning.
4
“Root”
CONV 2
Add the new class to an existing child node: If
the value of max(lm ) is greater than a threshold
(which can be set as per requirement) indicating
a strong resemblance/association with a particular
child node, the new class is added to child node k
such that lm (k) = max
(lm (k 0 )), k 0 ∈ [1, K]
0
CONV 1
i.
“Vehicle”
T HE E XPERIMENTS
We conducted two experiments using the datasets, CIFAR10 and CIFAR-100. We used MatConvNet [22], an opensource toolbox for implementation of Convolutional Neural
Networks in MATLAB [23].
The two experiments have very different initial conditions. In the first experiment on CIFAR-10 we show that
given a trained Tree-CNN with one root node and two branch
nodes how it expands as new classes arrive. We demonstrate
a single incremental learning stage. In the second experiment with CIFAR-100, the network starts with one root node
that can classify only 10 classes, and we progressively teach
the network 90 new classes over time.
Fig. 2. Detailed network of Tree-CNN A (before incremental learning)
During training, we performed simple data augmentation. Training images were flipped horizontally at random
with a probability of 0.5 [24]. All images were whitened and
contrast normalized [24]. The activation used in all the networks is rectified linear activation ReLU, σ(x) = max(x, 0).
The networks are trained using stochastic gradient descent
with fixed momentum of 0.9. Dropout [25] is used between
the final fully connected layers, and between pooling layers to regularize the network. We also employed batchnormalization (BNORM) [26] at the output of every convolutional layer. Additionally, a weight decay λ = 0.001
was set to regularize each model. The final layer performs
softmax operation on the output of the nodes to generate
class probabilities.
4.1
Adding Multiple New Classes
4.1.1 Dataset
We used CIFAR-10 dataset [27] for this experiment. It has 10
mutually exclusive classes, namely, airplane, automobile, bird,
cat, deer, dog, frog, horse, ship, and truck. There are 50, 000
training images and 10, 000 test images equally distributed
between 10 classes. Each image is a 32 × 32 color image, thus
having 3 input channels for Red, Green and Blue pixels.
We first train the network on 6 classes, and in the next
stage of incremental learning, we train the network for the
4 remaining classes.
4.1.2 The Network
For ease of reference, we label this network as Tree-CNN A.
The root node is a DCNN with two output nodes. It will
classify the input image as either “Animals” or “Vehicles”.
Each child node has a DCNN that does finer classification.
The description of the layers in each of these sub-networks
is given in Tables 1 and 2. Fig. 2 and Fig. 4 a) depict the
initial model of Tree-CNN A.
For comparison, we took another network (Network
“B”) with a complexity level similar to two stage complexity
of this Tree-CNN. It has 4 convolutional blocks, each block
having 2 sets of 3 × 3 convolutional kernels. The network
is inspired from the architecture of VGG-net [28]. Detailed
model is given in Table 3.
4.1.3 Initial Training
CIFAR-10 has classes that can easily be grouped into 2
distinct groups, “Vehicles” and “Animals”. We start with
6 classes, 3 belonging to each group. Automobile, ship, and
5
TABLE 1
Root Node for Tree-CNN A (CIFAR-10)
TABLE 3
Network B
Input 32x32x3
Conv1
5x5x3x64 ReLU stride 1 + BNORM
[2 2] Max Pooling stride 2
Conv 2
3x3x64x128 ReLU stride 1 + BNORM
Dropout 0.5
3x3x128x128 ReLU stride 1 + BNORM
[2 2] Max Pooling stride 2
FC
Fully Connected 8x8x12x512 ReLU
Dropout 0.5
Fully Connected 1x1x512x128 ReLU
Dropout 0.5
Fully Connected 1x1x128x2 ReLU
Softmax Layer
Input 32x32x3
Conv1
3x3x3x64 ReLU stride 1 + BNORM
Dropout 0.5
3x3x64x64 ReLU stride 1 + BNORM
[2 2] Max Pooling stride 2
Conv 2
3x3x64x128 ReLU stride 1 + BNORM
Dropout 0.5
3x3x128x128 ReLU stride 1 + BNORM
[2 2] Max pooling stride 2
Conv3
3x3x128x256 ReLU stride 1 + BNORM
Dropout 0.5
3x3x256c256 ReLU stride 1 + BNORM
[2 2] Max pooling stride 2
Conv4
3x3x256x512 ReLU stride 1 + BNORM
Dropout 0.5
3x3x512x512 ReLU stride 1 + BNORM
[2 2] Avg Pooling stride 2
FC
Fully Connected 2x2x512x1024 ReLU
Dropout 0.5
Fully Connected 1x1x1024x1024 ReLU
Dropout 0.5
Fully Connected 1x1x1024xN
(N=Number of Classes)
Softmax Layer
TABLE 2
Branch Node for Tree-CNN A (CIFAR-10)
Input 32x32x3
Conv1
5x5x3x32 ReLU stride 1 + BNORM
[2 2] Max Pooling stride 2
Dropout 0.25
Conv2
5x5x32x64 ReLU stride 1 + BNORM
[2 2] Max Pooling stride 2
Dropout 0.25
Conv3
3x3x64x64 ReLU stride 1 + BNORM
[2 2] Avg Pooling stride 2
Dropout 0.25
FC
Fully Connected 4x4x64x128 ReLU
Dropout 0.5
Fully Connected 1x1x128xN ReLU
(N=number of children)
Softmax Layer
truck are grouped as “Vehicles”, while cat, dog, and horse
are grouped as “Animals”. The network at the root node is
trained to classify the images as “Animals” or “Vehicles”.
For this, the 30, 000 training images belonging to the 6
classes are re-labeled as “Animals” or “Vehicles”. The root
node is then trained for 300 epochs. The learning rate is kept
at 0.1 for first 200 epochs, then reduced by 10 times every 50
epochs. At the second level, each of the two branch nodes are
separately trained. “Animals” branch node is trained with
15, 000 training images from the 3 classes. This node further
classifies the image into dog, cat, and horse. It is trained for
300 epochs and the learning rate is same as used for the top
node. Similarly, the branch node labeled “Vehicles” is trained
to identify the 3 distinct vehicles, ship, truck, and automobile.
The root node achieves a testing accuracy of 98.73%, while
the branch nodes, “Animals” and “Vehicles”, achieve 86%
and 94.43% testing accuracy respectively. Overall, this network achieves a testing accuracy of 89.10%.
For comparison, we train our network “B” with the 6
image classes. The network is trained for 300 epochs. The
learning rate is kept at 0.1 for first 200 epochs, then reduced
by 10 times every 50 epochs. We achieve a testing accuracy
of 92.40%. The accuracy of tree-CNN A is within the range
of network “B”.
4.1.4 Incremental Learning
Next we want to train the networks to identify the remaining four classes, bird, frog, deer and airplane. There are 500
training images per class given to us. We select at random
50 images per class, and show these 200 images to the root
node. We obtain the L matrix, which is a 2 × 4 matrix with
each element li j ∈ (0, 1). The 1st row of the matrix indicates
the softmax likelihood of each of the 4 classes as being
classified as “Vehicles”, while the second row presents the
same information for “Animals”. Fig. 3 shows how likely a
new class is to being classified as “Animals” or “Vehicles”.
Note, that this is not the number of times out of 50, the
images belonging to the class are classified as one of the
two super-classes by the root node. Instead, we perform the
softmax operation (eq. 2 on the actual values of the output
layer, to see which output neuron gives a higher value for a
particular class and by how much.
As per our algorithm, airplane gets grouped with “Vehicles” while bird, deer, and frog get added to the “Animals”
coarse category. The decision making is done using Algorithm 2 based on the softmax likelihood output of the root
node, as obtained in Fig. 3. The before and after structure
of the Tree-CNN A is shown in Fig. 4 . Next the root node
6
a)
b)
ROOT
BRANCH
NODES
LEAF
NODES
ANIMAL
VEHICLE
ROOT
ANIMAL
VEHICLE
SHIP
CAT
SHIP
CAT
DOG
TRUCK
DOG
TRUCK
HORSE
AUTOMOBILE
HORSE
AUTOMOBILE
DEER
AEROPLANE
NEW
CLASSES
BIRD
FROG
Fig. 3. Softmax likelihood of new classes being classified as Animals or
Vehicles for CIFAR 10
Fig. 4. Graphical representation of Tree-CNN A a) before incremental
learning, b) after incremental learning
4.2
is retrained with 50, 000 training images from all the 10
classes(old and new) to classify them in to the two coarse
categories. The root node is trained for 250 epochs, with
learning rate 0.1 for first 100 epochs, then the learning rate is
reduced by 10 every 50 epochs. Next we train the two branch
nodes. For each of the two nodes, we train it with training
images from both new and old classes. Each of the nodes
are trained for 250 epochs and the learning rate variation is
kept same as for the root node.
Algorithm 2 Incremental Learning for Tree-CNN A
Inputs: L = 2 × 4 likelihood matrix, node = root node of
the Tree-CNN, NewClass = array of new class labels
procedure T REE G ROW(L,node)
for (col = 1, col ≤ 4, col + +) do
if L(1, col) > L(2, col) then
AddClass(NewClass(col), node.children(1))
else
AddClass(NewClass(col), node.children(2))
end if
end for
end procedure
For comparison, we apply fine tuning to our already
trained network “B”. We add the 4 new classes as 4 new
output nodes of the final layer. As described in Table 3, it
is made up of 4 convolutional blocks (CONV) and one fully
connected block (FC). For network B, 5 fine tuning strategies
have been used. Each method retrains/fine-tunes certain
layers of the network. As listed below, we set 5 different
depths of back-propagation when retraining with the whole
dataset, i.e. all 10 classes.
1.
2.
3.
4.
5.
Case I: FC
Case II: FC + CONV1
Case III: FC + CONV1 + CONV2
Case IV: FC + CONV1 + CONV2 + CONV3
Case V: FC + CONV1 + CONV2 + CONV3 + CONV4
Case V represents the extreme scenario where the previous
network is completely retrained to learn the new classes.
This is equivalent to training a new network with all the
classes.
Sequentially Adding Multiple Classes
4.2.1
Dataset
For this experiment, we use a larger dataset, CIFAR-100
[27]. It has 100 classes, 500 training and 100 testing images
per class. The image is a 32 × 32 color image with 3 input
channels for Red, Green and Blue pixels. We divide the 100
classes into 10 groups, each group having 10 classes. This
distribution is done at random. We first train the networks
with 10 classes, and then incrementally add 10 new classes
every stage. There are 9 learning stages to add the remaining
90 classes.
4.2.2
The Network
Initially, the Tree-CNN has a root node and 10 leaf nodes. We
label this network as Tree-CNN C. The root node has a DCNN
network, with 10 output nodes. The layers of the CNN are
described in detail in Table 4. In subsequent learning stages,
the network would extend branches. The DCNN model
used in branch nodes is given in Table 5. For comparison, we
use the same network “B" as used in previous experiment.
4.2.3
Initial training
For Tree-CNN C, the root node is trained to classify 10
classes. It is trained for 300 epochs, with learning rate of
0.1 for first 200 epochs. It is then reduced by 10 every 50
epochs. We obtain a testing accuracy of 84.90%. Network
“B” is trained for 300 epochs as well, and the learning rate
is same as for the root node. We achieve a testing accuracy
of 85%. The starting accuracy for the two networks is almost
the same.
4.2.4
Incremental Learning
We divided the remaining 90 classes into 9 groups, each
containing 10 classes. These classes were added to the
network in 9 incremental learning stages. At each stage, first
50 images belonging to each class are shown to the root
node and a likelihood matrix L is generated. The columns
of the matrix are used to form an ordered set S, as described
in equations 3 and 4. For this experiment, we applied the
following constraints to the system.
•
Maximum depth of the tree is 2
7
•
TABLE 4
Root Node of Tree-CNN C (CIFAR 100)
Input 32x32x3
Conv 1
5x5x3x64 ReLU stride 1 + BNORM
[2 2] Max Pooling stride 2
Conv 2
3x3x64x128 ReLU stride 1 + BNORM
Dropout 0.5
3x3x128x128 ReLU stride 1 + BNORM
[2 2] Max Pooling stride 2
Conv 3
3x3x128x256 ReLU stride 1 + BNORM
Dropout 0.5
3x3x256x256 ReLU stride 1 + BNORM
[2 2] Avg Pooling stride 2
FC
Fully Connected 4x4x256x1024 ReLU
Dropout 0.5
Fully Connected 1x1x1024x1024 ReLU
Dropout 0.5
Fully Connected 1x1x1024xN
(N = Number of Children)
Softmax Layer
Maximum number of output/children for a branch
node is 10
The depth of the tree is constrained for ease of conducting
the experiment. The algorithm then needs to be implemented only at the root node level. It was observed during
developing the algorithm that new classes tend to have
a higher softmax likelihood value for branch nodes with
higher number of children. To prevent the network from
being lop-sided, by having one very large branch, we limit
the number of children to 10 per branch node. Once the
placement of the new classes is determined we train the
root node and the modified branch nodes. The root node is
trained for 250 epochs, with learning rate 0.1 for first 100
epochs, then the learning rate is reduced by 10 every 50
epochs.
To compare, we use same fine tuning test cases on
network “B” as the previous experiment. We have 5 different
cases, each having a different depth of back-propagation, as
explained in Section 4.1.2.
5
R ESULTS
Now, we discuss the results obtained from the two sets of
experiments. To reiterate, the first experiment had only one
incremental learning stage. While, the second experiment
demonstrated how the network grew to accommodate new
data over time. The network grows 10 folds in classification
capacity, from 10 classes to 100 classes. We compare our proposed method against re-training final layers of a standard
DCNN. The methods were compared against the following
metrics
Í
•
Training Effort =
(total number of weights × total
nets
TABLE 5
Branch Node of Tree-CNN C (CIFAR 100)
Input 32x32x3
Conv 1
5x5x3x32 ReLU stride 1 + BNORM
[2 2] Max Pooling stride 2
Dropout 0.25
Conv 2
5x5x32x64 ReLU stride 1 + BNORM
[2 2] Max Pooling stride 2
Dropout 0.25
Conv 3
3x3x64x64 ReLU stride 1 + BNORM
Dropout 0.5
3x3x256x256 ReLU stride 1 + BNORM
[2 2] Avg Pooling stride 2
FC
Fully Connected 4x4x64x512 ReLU
Dropout 0.5
Fully Connected 1x1x512x128 ReLU
Dropout 0.5
Fully Connected 1x1x128xN
(N = Number of Children)
Softmax Layer
•
number of training samples)
Testing Accuracy
Every training sample generates a corresponding error that
is back-propagated through stochastic gradient descent to
update the network weights. Training Effort is the number
of weight updates that happen per training epoch. As batch
size and number of training epochs is kept the same, the
product of number of weights and the number of training
samples used gives us a good measure of the effort used in
training a network. For Tree-CNN the training effort of each
of the nodes (‘nets’) is summed together. For network “B”,
there is only one net in each case.
5.1
Adding multiple new classes (CIFAR-10)
In Fig. 5, we compare the test accuracy and the training
effort for the 5 cases of fine-tuning network “B” against
our Tree-CNN C for CIFAR-100. Retraining only FC layer
of network “B” requires the least training effort. However,
it gives us the lowest accuracy, 78.37% amongst all. And as
more classes are introduced, this method causes much loss
in accuracy, as shown with CIFAR-100 in Fig. 7. Our proposed model, Tree-CNN A has the second lowest normalized
training effort, ∼ 40% less than ‘B:Case V’, and ∼ 30% less
than ‘B:Case II’. At the same time, Tree-CNN A (86.25%) had
comparable accuracy to ‘B:Case II’ (85.02%) and ‘B:Case III’
(88.15%), while just being less than the ideal case ‘B:Case V’
by a margin of 3.76%.
8
Fig. 5. Training Effort and Test Accuracy comparison for Tree-CNN A
against Network B for CIFAR-10 dataset
Fig. 7. Testing Accuracy for CIFAR-100 over multiple learning stages
Node #3
baby
boy
girl
man
woman
poppy
butterfly
orchid
sunflower
tulip
Oak Tree
Pine Tree
Forest
StreetCar
Node #8
Node #13
Fig. 8. Examples of groups of classes formed when incrementally learning CIFAR-100
Fig. 6. Training Effort over different learning stages when sequentially
adding multiple classes (CIFAR-100)
5.2
Sequentially adding new classes (CIFAR-100)
We compare Tree-CNN C against the 5 different fine-tuning
cases of Network “B”. Fig. 6 shows the training effort for all
the test cases for CIFAR-100 experiment. We normalized the
training effort by dividing all the values with the highest
training effort. i.e. ‘B:Case V’. ‘B:Case I’ has the lowest
training effort, as we only fine tune the final fully connected
layer. However, it performs the worst in accuracy as shown
in Fig. 7. Tree-CNN C requires almost the same training effort
as ‘B:Case II’, and achieves better accuracy than ‘B:Case
II’ and ‘B:Case III’. It achieves accuracy within the same
range as ‘B:Case IV’, while requiring 20% less training effort.
‘B:Case IV’ and ‘B:Case V’ require almost similar training
effort, as the difference is only the extra training of the smallest CONV layer, i.e. the first layer. ‘B:Case V’ gives us the
best accuracy, however, that is because we are retraining the
entire network with all the images. There is no pre-trained
kernel sharing and it is as good as starting anew, thereby
requiring the highest training effort. The Tree-CNN achieves
an accuracy of 60.46% on the full CIFAR-100 dataset. It is
2.59% less than the accuracy achieved by training the full
network “B”, which gives us 63.05%. The overall accuracy
of the Tree-CNN and the network we compared with is
comparable to the range of reported accuracy of similar
sized networks, such as 67.38% by HD-CNN [19], 67.68%
by Hertel, et al [29]. Further improvements to accuracy can
be done by modifying the CNN architecture of the nodes,
and by adopting methods such as “all convolutional net”
[30], “exponential linear units” [31].
In all 6 cases, the training effort required at a particular
learning stage was greater than the effort required by the
previous stage. This is because we had to show images
belonging to old classes to avoid “catastrophic forgetting”.
Our method had lower slope for Training Effort v/s Learning Stage, as compared to all but one (Case I) of network
9
TABLE 6
Root Node of Tree-CNN C classifying CIFAR-100 classes into 17 child nodes (after 9 Learning Stages)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
trout
bridge
girl
caterpillar
lawn mower
possum
otter
poppy
sweet pepper
bicycle
lobster
clock
oak tree
skyscraper
squirrel
crocodile
wolf
dolphin
mountain
man
bottle
turtle
palm tree
woman
can
seal
tank
baby
bear
sea
willow tree
boy
skunk
mouse
castle
lizard
plain
elephant
train
mushroom
bus
maple tree
leopard
fox
tiger
telephone
rocket
porcupine
hamster
chimpanzee
aquarium fish
apple
chair
kangaroo
cloud
sunflower
orange
couch
whale
bowl
butterfly
pear
table
shrew
lion
cockroach
rose
house
dinosaur
ray fish
tulip
beaver
worm
orchid
flatfish
raccoon
spider
shark
bee
beetle
snail
rabbit
crab
bed
tractor
pickup truck
road
motorcyle
plate
streetcar
wardrobe
pine tree
keyboard
forest
television
cup
lamp
snake
camel
cattle
“B”. For Tree-CNN C, the normalized training effort incremented by 0.081 on average for every stage, while for
network “B”, the increment was 0.041, 0.086, 0.097, 0.0997,
and 0.1 for the 5 cases, in increasing order of number of
layers trained in each case.
An interesting thing to note was similar looking classes,
that were also semantically similar, were grouped under the
same branches. At the end of the 9 incremental learning
stages, the root node had 17 children nodes out of which 4
were leaf nodes and remaining 13 were branch nodes. The
details of the classes associated with each of these 17 nodes
is given in Table 6. These 13 branch nodes further had 3 to
10 leaf nodes. The final groups formed after training all the
classes is given in Table 6. Certain similar objects grouped
together is shown in Fig. 8. While there were some groups
that had objects sharing semantic similarity as well, there
were a few odd groups, such as Node 13 as shown in Fig. 8.
The groups formed is dependent partly on the sequence in
which the classes arrived. As we had limited the number of
leaf nodes of a branch to 10, even when a new class matched
strongly to an already full branch, the algorithm added it
to the the second best match, or created a new node. In
some cases, such as branch number 13, the street car was
grouped with the trees because most images of streetcar in
the dataset were green in color. Color and shape similarities
led to grouping of very semantically dissimilar objects.
6
layers of a large network, while maintaining comparable
accuracy. However, the Tree-CNN continues to grow in size
over time, and the implications of that while implementing
on hardware is an important future work. The Tree-CNN
grows in a manner such that images that share common
features are closer neighbors in the tree than those images
that are very different. The final leaf nodes, and the distance
between them can also be used as measure of how similar
or different any two images are. Such a method of training
and classification can be used to hierarchically classify large
datasets. The correlation between image similarity and the
location of the image classes in the Tree-CNN needs to be
analyzed empirically. Our proposed method, Tree-CNN, thus
offers a better incremental learning model that is based on
hierarchical classifiers and transfer learning. It could, also,
potentially be applied to large image search operations.
During inferencing, we can choose to parse the network
up to a certain depth to give us an idea of which “supercategory” the search image belongs to. Hence, it could offer
a quick solution to analyze incoming data.
ACKNOWLEDGMENT
The work was supported in part by Center for Braininspired Computing Enabling Autonomous Intelligence (CBRIC), a DARPA sponsored JUMP center, by Semiconductor
Research Corporation, and by Army Research Lab.
D ISCUSSION
In this work, we explore a method for incremental learning
when dealing with very large image databases. The motivation of this work stems from the idea that subsequent
addition of new image classes to a network should be easier
than retraining the whole network again with all classes. We
observed that expectedly, each incremental learning stage
required more effort than the previous, because images belonging to old classes needed to be shown to the CNNs. This
is due to the inherent problem of “catastrophic forgetting”
in deep neural networks. Our proposed method has a lower
rate of increase of training effort over consecutive learning
stages as compared to the effort need to fine-tune limited
R EFERENCES
[1]
[2]
[3]
[4]
W. Rawat and Z. Wang, “Deep Convolutional Neural Networks
for Image Classification: A Comprehensive Review,” Neural Computation, vol. 29, no. 9, pp. 2352–2449, 2017.
A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in
neural information processing systems, 2012, pp. 1097–1105.
Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based
learning applied to document recognition,” Proceedings of the IEEE,
vol. 86, no. 11, pp. 2278–2324, 1998.
L. Wan, M. Zeiler, S. Zhang, Y. L. Cun, and R. Fergus, “Regularization of neural networks using dropconnect,” in Proceedings of
the 30th international conference on machine learning (ICML-13), 2013,
pp. 1058–1066.
10
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “Imagenet
large scale visual recognition challenge,” International Journal of
Computer Vision, vol. 115, no. 3, pp. 211–252, 2015.
Z. Chen and B. Liu, “Lifelong machine learning,” Synthesis Lectures
on Artificial Intelligence and Machine Learning, vol. 10, no. 3, pp. 1–
145, 2016.
S. Thrun and T. M. Mitchell, “Lifelong robot learning,” Robotics
and autonomous systems, vol. 15, no. 1-2, pp. 25–46, 1995.
S. Thrun, “Is learning the n-th thing any easier than learning the
first?” in Advances in neural information processing systems, 1996, pp.
640–646.
I. J. Goodfellow, M. Mirza, D. Xiao, A. Courville, and Y. Bengio,
“An empirical investigation of catastrophic forgetting in gradientbased neural networks,” arXiv preprint arXiv:1312.6211, 2013.
J. Yosinski, J. Clune, Y. Bengio, and H. Lipson, “How transferable
are features in deep neural networks?” in Advances in neural
information processing systems, 2014, pp. 3320–3328.
M. Oquab, L. Bottou, I. Laptev, and J. Sivic, “Learning and transferring mid-level image representations using convolutional neural
networks,” in Proceedings of the IEEE conference on computer vision
and pattern recognition, 2014, pp. 1717–1724.
L. Fei-Fei, R. Fergus, and P. Perona, “One-shot learning of object
categories,” IEEE transactions on pattern analysis and machine intelligence, vol. 28, no. 4, pp. 594–611, 2006.
R. Girshick, “Fast r-CNN,” in 2015 IEEE International Conference on
Computer Vision (ICCV). IEEE, dec 2015.
K. Shmelkov, C. Schmid, and K. Alahari, “Incremental learning
of object detectors without catastrophic forgetting,” in 2017 IEEE
International Conference on Computer Vision (ICCV). IEEE, oct 2017.
Z. Li and D. Hoiem, “Learning without forgetting,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017.
S. S. Sarwar, P. Panda, and K. Roy, “Gabor filter assisted energy
efficient fast learning Convolutional Neural Networks,” in 2017
IEEE/ACM International Symposium on Low Power Electronics and
Design (ISLPED). IEEE, jul 2017, pp. 1–6.
P. Panda and K. Roy, “Semantic driven hierarchical learning for
energy-efficient image classification,” in 2017 Design, Automation
& Test in Europe Conference & Exhibition (DATE). IEEE, 2017, pp.
1582–1587.
P. Panda, A. Ankit, P. Wijesinghe, and K. Roy, “Falcon: Feature
driven selective classification for energy-efficient image recognition,” IEEE Transactions on Computer-Aided Design of Integrated
Circuits and Systems, vol. 36, no. 12, 2017.
Z. Yan, H. Zhang, R. Piramuthu, V. Jagadeesh, D. DeCoste, W. Di,
and Y. Yu, “HD-CNN: Hierarchical deep convolutional neural networks for large scale visual recognition,” in 2015 IEEE International
Conference on Computer Vision (ICCV). IEEE, dec 2015.
N. Srivastava and R. R. Salakhutdinov, “Discriminative transfer
learning with tree-based priors,” in Advances in Neural Information
Processing Systems, 2013, pp. 2094–2102.
T. Xiao, J. Zhang, K. Yang, Y. Peng, and Z. Zhang, “Error-Driven
Incremental Learning in Deep Convolutional Neural Network for
Large-Scale Image Classification,” MM ’14 Proceedings of the ACM
International Conference on Multime, vol. d, pp. 177–186, 2014.
A. Vedaldi and K. Lenc, “MatConvNet,” in Proceedings of the 23rd
ACM international conference on Multimedia - MM 2015. ACM
Press, 2015.
MATLAB, version 9.2.0 (R2017a).
Natick, Massachusetts: The
MathWorks Inc., 2017.
I. J. Goodfellow, D. Warde-Farley, M. Mirza, A. Courville, and
Y. Bengio, “Maxout networks,” in International Conference on Machine Learning, 2013.
N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov, “Dropout: a simple way to prevent neural
networks from overfitting.” Journal of machine learning research,
vol. 15, no. 1, pp. 1929–1958, 2014.
S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
network training by reducing internal covariate shift,” in International Conference on Machine Learning, 2015, pp. 448–456.
A. Krizhevsky and G. Hinton, “Learning multiple layers of features from tiny images,” 2009.
K. Simonyan and A. Zisserman, “Very deep convolutional
networks for large-scale image recognition,” arXiv preprint
arXiv:1409.1556, 2014.
L. Hertel, E. Barth, T. Kaster, and T. Martinetz, “Deep convolutional neural networks as generic feature extractors,” in 2015
International Joint Conference on Neural Networks (IJCNN). IEEE,
jul 2015.
[30] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller,
“Striving for simplicity: The all convolutional net,” arXiv preprint
arXiv:1412.6806, 2014.
[31] D.-A. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and accurate
deep network learning by exponential linear units (elus),” arXiv
preprint arXiv:1511.07289, 2015.
Deboleena Roy received the B.Tech+M.Tech
Dual Degree in Electronics and Electrical Communications Engineering from IIT Kharagpur,
Kharagpur, India in 2014. She is currently working towards the Ph.D. degree in Electrical and
Computer Engineering at Purdue University,
West Lafayette, IN, USA. Prior to that, she
was a Design Engineer at Qualcomm Bangalore Design Center, Bengaluru, India from 2014
to 2016. Her current research interests include
neuro-inspired algorithms for cognitive applications such as perception, reasoning and decision making.
Priyadarshini Panda received the B.E. degree
in Electrical and Electronics Engineering and the
M.Sc. degree in Physics from the Birla Institute of Technology and Science, Pilani, India,
in 2013. She is currently pursuing the Ph.D.
degree in Electrical and Computer engineering
with Purdue University, West Lafayette, IN, USA.
She was a Component Design Engineer with
Intel Corporation, Bengaluru, India, from 2013
to 2014. Prior to that, she did a six-month internship with Nvidia Graphics, Bengaluru. Her
current research interests include low-power neuromorphic computing:
energy-efficient realization of neural networks (spiking/non-spiking in
deep learning context) using novel architectures and algorithms.
Kaushik Roy received the B.Tech. degree in
Electronics and Electrical Communications Engineering from IIT Kharagpur, Kharagpur, India
and the Ph.D. degree from the Department of
Electrical and Computer Engineering, University
of Illinois at Urbana−Champaign, Champaign,
IL, USA, in 1990. He was with the Semiconductor Process and Design Center, Texas Instruments Incorporated, Dallas, TX, USA, where he
was involved in FPGA architecture development
and low-power circuit design. He joined the Faculty of Electrical and Computer Engineering, Purdue University, West
Lafayette, IN, USA, in 1993, where he is currently Edward G. Tiedemann
Jr. Distinguished Professor. He has authored more than 600 papers
in refereed journals and conferences, holds 15 patents, graduated 75
Ph.D. students, and has co-authored two books on Low Power CMOS
VLSI Design (Wiley and McGraw Hill). His current research interests include neuromorphic and cognitive computing, spintronics, device-circuit
codesign for nanoscale Silicon and non-Silicon technologies, low-power
electronics for portable computing and wireless communications, and
new computing models enabled by emerging technologies.
| 2cs.AI
|
Knowledge Graph Embedding with Iterative Guidance from Soft Rules
Shu Guo1,2 , Quan Wang1,2,3∗ , Lihong Wang4 , Bin Wang1,2 , Li Guo1,2
Institute of Information Engineering, Chinese Academy of Sciences
School of Cyber Security, University of Chinese Academy of Sciences
3
State Key Laboratory of Information Security, Chinese Academy of Sciences
4
National Computer Network Emergency Response Technical Team & Coordination Center of China
1
arXiv:1711.11231v1 [cs.AI] 30 Nov 2017
2
Abstract
Embedding knowledge graphs (KGs) into continuous vector
spaces is a focus of current research. Combining such an embedding model with logic rules has recently attracted increasing attention. Most previous attempts made a one-time injection of logic rules, ignoring the interactive nature between
embedding learning and logical inference. And they focused
only on hard rules, which always hold with no exception and
usually require extensive manual effort to create or validate.
In this paper, we propose Rule-Guided Embedding (RUGE),
a novel paradigm of KG embedding with iterative guidance
from soft rules. RUGE enables an embedding model to learn
simultaneously from 1) labeled triples that have been directly
observed in a given KG, 2) unlabeled triples whose labels are
going to be predicted iteratively, and 3) soft rules with various confidence levels extracted automatically from the KG. In
the learning process, RUGE iteratively queries rules to obtain
soft labels for unlabeled triples, and integrates such newly labeled triples to update the embedding model. Through this
iterative procedure, knowledge embodied in logic rules may
be better transferred into the learned embeddings. We evaluate RUGE in link prediction on Freebase and YAGO. Experimental results show that: 1) with rule knowledge injected iteratively, RUGE achieves significant and consistent improvements over state-of-the-art baselines; and 2) despite their uncertainties, automatically extracted soft rules are highly beneficial to KG embedding, even those with moderate confidence
levels. The code and data used for this paper can be obtained
from https://github.com/iieir-km/RUGE.
Introduction
Knowledge graphs (KGs) such as WordNet (Miller 1995),
Freebase (Bollacker et al. 2008), YAGO (Suchanek, Kasneci, and Weikum 2007), and NELL (Carlson et al. 2010)
are extremely useful resources for many AI related applications. A KG is a multi-relational graph composed of entities
as nodes and relations as different types of edges. Each edge
is represented as a triple (head entity, relation, tail entity),
indicating that there is a specific relation between two entities, e.g., (Paris, CapitalOf, France). Although effective
in representing structured data, the underlying symbolic nature of such triples often makes KGs hard to manipulate.
∗
Corresponding author: Quan Wang ([email protected]).
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
Recently, a new research direction termed as knowledge
graph embedding has been proposed and quickly received
massive attention (Nickel, Tresp, and Kriegel 2011; Bordes
et al. 2013; Wang et al. 2014; Lin et al. 2015b; Yang et al.
2015; Nickel, Rosasco, and Poggio 2016; Trouillon et al.
2016). The key idea is to embed entities and relations in a
KG into a low-dimensional continuous vector space, so as
to simplify the manipulation while preserving the inherent
structure of the KG. Such embeddings contain rich semantic
information, and can benefit a broad range of downstream
applications (Weston et al. 2013; Bordes et al. 2014; Zhang
et al. 2016; Xiong, Power, and Callan 2017).
Traditional methods performed embedding based solely
on triples observed in a KG. But considering the power of
logic rules in knowledge acquisition and inference, combining embedding models with logic rules has become a focus
of current research (Rocktäschel et al. 2014; Vendrov et al.
2015; Wang and Cohen 2016; Hu et al. 2016). Wang et al.
(2015) and Wei et al. (2015) tried to use embedding models
and logic rules for KG completion. But in their work, rules
are modeled separately from embedding models, and would
not help to learn more predictive embeddings. Rocktäschel
et al. (2015) and Guo et al. (2016) then devised joint learning
paradigms which can inject first-order logic (FOL) into KG
embedding. Demeester et al. (2016) further proposed lifted
rule injection to avoid the costly propositionalization of FOL
rules. Although these joint models are able to learn better
embeddings after integrating logic rules, they still have their
drawbacks and restrictions.
First of all, these joint models made a one-time injection
of logic rules, taking them as additional rule-based training
instances (Rocktäschel, Singh, and Riedel 2015) or regularization terms (Demeester, Rocktäschel, and Riedel 2016).
We argue that rules can better enhance KG embedding, however, in an iterative manner. Given the learned embeddings
and their rough predictions, rules can be used to refine the
predictions and infer new facts. The newly inferred facts, in
turn, will help to learn better embeddings and more accurate
logical inference. Previous methods fail to model such interactions between embedding models and logic rules. Furthermore, they focused only on hard rules which always hold
with no exception. Such rules usually require extensive manual effort to create or validate. Actually, besides hard rules,
a significant amount of background information can be en-
proach is quite generic and flexible. It can integrate various
types of rules with different confidence levels to enhance a
good variety of KG embedding models.
VRIWUXOHV
6RIW/DEHO3UHGLFWLRQ
Related Work
XQODEHOHGWULSOHV
VRIWODEHOV
HPEHGGLQJV
ODEHOHGWULSOHV
KDUGODEHOV
(PEHGGLQJ5HFWLILFDWLRQ
Figure 1: Framework overview. RUGE enables an embedding model to learn simultaneously from labeled triples, unlabeled triples, and soft rules in an iterative manner, where
each iteration alternates between a soft label prediction stage
and an embedding rectification stage.
coded as soft rules, e.g., “a person is very likely (but not necessarily) to have a nationality of the country where he/she
was born”. Soft rules can be extracted automatically and efficiently via modern rule mining systems (Galárraga et al.
2013; Galárraga et al. 2015). Yet, despite this merit, soft
rules have not been well studied in previous methods.
This paper proposes RUle-Guided Embedding (RUGE), a
novel paradigm of KG embedding with iterative guidance
from soft rules. As sketched in Fig. 1, it enables an embedding model to learn simultaneously from 1) labeled triples
that have been directly observed in a given KG, 2) unlabeled
triples whose labels are going to be predicted iteratively, and
3) soft rules with different confidence levels extracted automatically from the KG. During each iteration of the learning
process, the model alternates between a soft label prediction
stage and an embedding rectification stage. The former uses
currently learned embeddings and soft rules to predict soft
labels for unlabeled triples, and the latter further integrates
both labeled and unlabeled triples (with hard and soft labels
respectively) to update current embeddings. Through this iterative procedure, knowledge embodied in logic rules may
be better transferred into the learned embeddings.
We empirically evaluate RUGE on large scale public KGs,
namely Freebase and YAGO. Experimental results reveal
that: 1) by incorporating logic rules, RUGE significantly and
consistently improves over state-of-the-art basic embedding
models (without rules); 2) compared to those one-time injection schemes studied before, the iterative injection strategy maximizes the utility of logic rules for KG embedding,
and indeed achieves substantially better performance; 3) despite the uncertainties, automatically extracted soft rules are
highly beneficial to KG embedding, even those with moderate confidence levels.
The contributions of this paper are threefold. 1) We devise
a novel paradigm of KG embedding which iteratively injects
logic rules into the learned embeddings. To our knowledge,
this is the first work that models interactions between embedding learning and logical inference in a principled framework. 2) We demonstrate the usefulness of automatically extracted soft rules in KG embedding, thereby eliminating the
requirement of laborious manual rule creation. 3) Our ap-
Recent years have witnessed increasing interest in learning
distributed representations for entities and relations in KGs,
a.k.a. KG embedding. Various techniques have been devised
for this task, e.g., translation-based models which take relations as translating operations between head and tail entities (Bordes et al. 2013; Wang et al. 2014; Lin et al. 2015b),
simple compositional models which match compositions of
head-tail entity pairs with their relations (Nickel, Tresp, and
Kriegel 2011; Yang et al. 2015; Nickel, Rosasco, and Poggio 2016; Trouillon et al. 2016), and neural networks which
further introduce non-linear layers and deep architectures
(Socher et al. 2013; Bordes et al. 2014; Dong et al. 2014;
Liu et al. 2016). Among these techniques, ComplEx (Trouillon et al. 2016), a compositional model which represents
entities and relations as complex-valued vectors, achieves a
very good trade-off between accuracy and efficiency. Most
of the currently available techniques perform the embedding
task based solely on triples observed in a KG. Some recent
work further tried to use other information, e.g., entity types
(Guo et al. 2015; Xie, Liu, and Sun 2016) and textual descriptions (Xie et al. 2016; Xiao, Huang, and Zhu 2017), to
learn more predictive embeddings. See (Wang et al. 2017)
for a thorough review of KG embedding techniques.
Given the power of logic rules in knowledge acquisition
and inference, combining KG embedding with logic rules
becomes a focus of current research. Wang et al. (2015) and
Wei et al. (2015) devised pipelined frameworks which use
logic rules to further refine predictions made by embedding
models. In their work, rules will not help to learn better embeddings. Rocktäschel et al. (2015) and Guo et al. (2016)
then tried to learn KG embeddings jointly from triples and
propositionalized FOL rules. Demeester et al. (2016) further
proposed lifted rule injection to avoid the costly propositionalization. These joint models, however, made a one-time
injection of logic rules, ignoring the interactive nature between embedding learning and logical inference. Moreover,
they can only handle hard rules which are usually manually
created or validated.
Besides logic rules, relation paths which can be regarded
as Horn clauses and get a strong connection to logical inference (Gardner, Talukdar, and Mitchell 2015), have also been
studied in KG embedding (Neelakantan, Roth, and McCallum 2015; Lin et al. 2015a; Guu, Miller, and Liang 2015).
But in these methods, relation paths are incorporated, again,
in a one-time manner. Our approach, in contrast, iteratively
injects knowledge contained in logic rules into KG embedding, and is able to handle soft rules with various confidence
levels extracted automatically from KGs.
Combining logic rules with distributed representations is
also an active research topic in other contexts outside KGs.
Faruqui et al. (2014) tried to inject ontological knowledge
from WordNet into word embeddings. Vendrov et al. (2015)
introduced order-embedding to model the partial order structure of hypernymy, textual entailment, and image caption-
ing. Hu et al. (2016) proposed to enhance various types of
neural networks with FOL rules. All these studies demonstrate the capability of logic rules to enhance distributed representation learning.
Rule-Guided Knowledge Graph Embedding
This section introduces RUle-Guided Embedding (RUGE),
a novel paradigm of KG embedding with iterative guidance
from soft rules. RUGE enables an embedding model to learn
simultaneously from labeled triples, unlabeled triples, and
soft rules in an iterative manner. During each iteration, the
model alternates between a soft label prediction stage and
an embedding rectification stage. Fig. 1 sketches this overall
framework. In what follows, we first describe our learning
resources, and then detail the two alternating stages.
Learning Resources
universally quantified rule ∀x, y : (x, BornInCountry, y)
⇒ (x, Nationality, y) could be instantiated with two entities EmmanuelMacron and France, and gives a resultant
grounding (EmmanuelMacron, BornInCountry, France)
⇒ (EmmanuelMacron, Nationality, France). Obviously, there could be a huge number of groundings,
especially given a large entity vocabulary E. In this
paper, to maximize the utility for knowledge acquisition and inference, we take as valid groundings only
those where premise triples are observed in O while
conclusion triples are not. That means the aforementioned grounding will be considered as valid if the triple
(EmmanuelMacron, BornInCountry, France) ∈ O but
(EmmanuelMacron, Nationality, France) ∈
/ O. For
Qp
each FOL rule fp , let Gp = {gpq }q=1 denote the set of
its valid groundings. All the premise triples of gpq are
contained in O, but the single conclusion triple is not.
These conclusion triples are further used to construct our
unlabeled triple set U. That means, our unlabeled triples are
those which are not directly observed in the KG but could
be inferred by the rules with high probabilities.
Modeling Triples and Rules. Given the labeled triples L,
unlabeled triples U, and the valid groundings of FOL rules
G = {Gp }P
p=1 , we discuss how to model these triples and
rules in the context of KG embedding. To model triples, we
follow ComplEx (Trouillon et al. 2016), a recently proposed
method which is simple and efficient while achieving stateof-the-art predictive performance. Specifically, we assume
entities and relations to have complex-valued vector embeddings. Given a triple (ei , rk , ej ) ∈ E ×R×E, we score it by
a multi-linear dot product:
X
ηijk = Re(hei , rk , ēj i) = Re(
[ei ]m [rk ]m [ēj ]m ), (1)
Suppose we are given a KG with a set of triples observed,
i.e., O = {(ei , rk , ej )}. Each triple is composed of two entities ei , ej ∈ E and their relation rk ∈ R, where E and R
are the sets of entities and relations respectively. We obtain
our learning resources (i.e., labeled triples, unlabeled triples,
and soft rules) and model them as follows.
Labeled Triples. We take the triples observed in O as positive ones. For each positive triple (ei , rk , ej ), we randomly
corrupt the head ei or the tail ej , to form a negative triple
(e0i , rk , ej ) or (ei , rk , e0j ), where e0i ∈ E \ {ei } and e0j ∈
E \ {ej }. We denote a labeled triple as x` , and associate with
it a label y` = 1 if x` is positive, and y` = 0 otherwise. Let
L = {(x` , y` )} denote the set of these labeled triples (along
with their labels).
Unlabeled Triples. Besides the labeled triples, we collect a
m
set of unlabeled triples U = {xu }, where xu = (ei , rk , ej )
where ei , ej , rk ∈ Cd are the complex-valued vector emindicates an unlabeled triple. In fact, all the triples that have
beddings associated with ei , ej , and rk , respectively; ēj is
not been observed in O can be taken as unlabeled ones. But
the conjugate of ej ; [·]m is the m-th entry of a vector; and
in this paper, we consider only those encoded in the concluRe(·) means taking the real part of a complex value. We fursion of a soft rule, as detailed below.
ther introduce a mapping function φ : E×R×E → (0, 1), so
Soft Rules. We also consider a set of FOL rules with differas to map the score ηijk to a continuous truth value which
ent confidence levels, denoted as F = {(fp , λp )}P
p=1 . Here,
lies in the range of (0, 1), i.e.,
fp is the p-th logic rule defined over the given KG, repreφ(ei , rk , ej ) = σ(ηijk ) = σ Re(hei , rk , ēj i) , (2)
sented, e.g., in the form of ∀x, y : (x, rs , y) ⇒ (x, rt , y),
where σ(x) = 1/(1 + exp(−x)) denotes the sigmoid funcstating that two entities linked by relation rs might also
tion. Triples with higher truth values are more likely to hold.
be linked by relation rt . The left-hand side of the imTo model propositionalized rules (i.e. groundings), we use
plication “⇒” is called the premise, and the right-hand
t-norm based fuzzy logics (Hájek 1998). The key idea is to
side the conclusion. In this paper, we restrict fp to be
model the truth value of a propositionalized rule as a coma Horn clause rule, where the conclusion contains only
position of the truth values of its constituent triples, through
a single atom and the premise is a conjunction of sevspecific logical connectives (e.g. ∧ and ⇒). For instance,
eral atoms. The confidence level of rule fp is denoted
the truth value of a grounded rule (eu , rs , ev ) ⇒ (eu , rt , ev )
as λp ∈ [0, 1]. Rules with higher confidence levels are
will be determined by the truth values of the two triples
more likely to hold, and a confidence level of λp = 1
(eu , rs , ev ) and (eu , rt , ev ), via a composition defined by
indicates a hard rule which always holds with no exceplogical implication. We follow (Guo et al. 2016) and define
tion. Such rules as well as their confidence levels can be
the compositions associated with logical conjunction (∧),
extracted automatically from the KG (with the observed
disjunction (¬), and negation (¬) as:
triple set O as input), by using modern rule mining systems
like AMIE and AMIE+ (Galárraga et al. 2013; Galárraga et al. 2015).
π(a ∧ b) = π(a) · π(b),
(3)
We then propositionalize these rules to get their groundπ(a ∨ b) = π(a) + π(b) − π(a) · π(b),
(4)
ings. Here a grounding is the logical expression with all variπ(¬a)
=
1
−
π(a).
(5)
ables instantiated with concrete entities in E. For instance, a
Here, a and b are two logical expressions, which can either
be single triples or be constructed by combining triples with
logical connectives; and π(a) is the truth value of a, indicating to what degree the logical expression is true. If a is a
single triple, say (ei , rk , ej ), we have π(a) = φ(ei , rk , ej ),
as defined in Eq. (2). Given these compositions, the truth
value of any logical expression can be calculated recursively
(Guo et al. 2016), e.g.,
π(a ⇒ b) = π(¬a ∨ b) = π(a) · π(b) − π(a) + 1.
(6)
Logical expressions with higher truth values have greater degrees to be true. Let Θ = {e}e∈E ∪ {r}r∈R denote the set of
all entity and relation embeddings. The proposed approach,
RUGE, then aims to learn these embeddings by using the
labeled triples L, unlabeled triples U, and valid groundings
{Gp }P
p=1 in an iterative manner, where each iteration alternates between a soft label prediction stage and an embedding
rectification stage.
Soft Label Prediction
This stage is to use currently learned embeddings and propositionalized rules to predict soft labels for unlabeled triples.
Specifically, let n be the iteration index, and Θ(n−1) the set
of current embeddings learned from the previous iteration.
Recall that we are given a set of P FOL rules with their confidence levels F = {(fp , λp )}P
p=1 , and each FOL rule fp has
Q
p
Qp valid groundings Gp = {gpq }q=1
. Our aim is to predict a
soft label s(xu ) ∈ [0, 1] for each unlabeled triple xu ∈ U, by
using the current embeddings Θ(n−1) and all the groundings
G = {Gp }P
p=1 .
To do so, we solve a rule-constrained optimization problem, which projects truth values of unlabeled triples computed by the current embeddings into a subspace constrained
by the rules. The key idea here is to find optimal soft labels that stay close to these truth values, while at the same
time fitting the rules. For the first property, given each unlabeled triple xu ∈ U, we calculate its truth value φ(xu )
using the current embeddings via Eq. (2), and require the
soft label s(xu ) to stay close to this truth value. We measure
the closeness between s(xu ) and φ(xu ) with a squared loss,
and try to minimize it. For the second property, we further
impose rule constraints onto the soft labels S = {s(xu )}.
Specifically, for each FOL rule fp and each of its groundings gpq , we expect gpq to be true, i.e., π(gpq |S) = 1 with
confidence λp . Here, π(gpq |S) is the conditional truth value
of gpq given the soft labels, which can be calculated recursively with the logical compositions defined in Eq. (3) to
Eq. (5). Take gpq := (eu , rs , ev ) ⇒ (eu , rt , ev ) as an example, where the premise (eu , rs , ev ) is directly observed in
O, and the conclusion (eu , rt , ev ) is an unlabeled triple included in U. The conditional truth value of gpq can then be
calculated as:
π(gpq |S) = φ(eu , rs , ev )·s(eu , rt , ev )−φ(eu , rs , ev )+1, (7)
where φ(eu , rs , ev ) is a truth value defined by Eq. (2) with
the current embeddings; and s(eu , rt , ev ) is a soft label to be
predicted. Comparing Eq. (7) with Eq. (6), we can see that
during the calculation of π(gpq |S), for any unlabeled triple,
we use the soft label s(·) rather than the truth value φ(·), so
as to better impose rule constraints onto the soft labels S.
Combining the two properties together and further allowing slackness for rule constraints, we finally get the following optimization problem:
X
1 X
2
min
(s(xu ) − φ(xu )) + C
ξpq ,
S,ξ 2
p,q
xu ∈U
s.t. λp (1 − π(gpq |S)) ≤ ξpq , q = 1,· · ·, Qp , p = 1,· · ·, P,
ξpq ≥ 0, q = 1,· · ·, Qp , p = 1,· · ·, P,
0 ≤ s(xu ) ≤ 1, ∀s(xu ) ∈ S,
(8)
where ξpq is a slack variable and C the penalty coefficient.
Note that confidence levels of rules (i.e. λp ’s) are encoded
in the constraints, making our approach capable of handling
soft rules. Rules with higher confidence levels show less tolerance for violating the constraints. This optimization problem is convex, and can be solved efficiently with its closedform solution:
h
i1
X
s(xu ) = φ(xu ) + C
λp ∇s(xu ) π(gpq |S)
(9)
p,q
0
for each xu ∈ U. Here, ∇s(xu ) π(gpq |S) means the gradient
of π(gpq |S) w.r.t s(xu ), which is a constant w.r.t. S,1 and
[x]10 = min(max(x, 0), 1) is a truncation function enforcing
the solutions to stay within [0, 1]. We provide the proof of
convexity and detailed derivation as supplementary materials. Soft labels obtained in this way shall 1) stay close to the
predictions made by the current embedding model, and 2) fit
the rules as well as possible.
Embedding Rectification
This stage is to integrate both labeled and unlabeled triples
(with hard and soft labels respectively) to update current embeddings. Specifically, we are given a set of labeled triples
with their hard labels specified in {0, 1}, i.e., L = {(x` , y` )},
and also a set of unlabeled triples encoded in propositionalized rules, i.e., U = {xu }. Each unlabeled triple xu has a
soft label s(xu ) ∈ [0, 1], predicted by Eq. (9). We would
like to use these labeled and unlabeled triples to learn the
updated embeddings Θ(n) . Here n is the iteration index.
To this end, we minimize a global loss over L and U, so
as to find embeddings which can predict the true hard labels
for triples contained in L, while imitating the soft labels for
those contained in U. The optimization problem is:
1 X
1 X
min
`(φ(x` ), y` ) +
`(φ(xu ), s(xu )), (10)
Θ |L|
|U|
L
U
where `(x, y) = −y log x − (1 − y) log(1 − x) is the cross
entropy; and φ(·) is a function w.r.t. Θ defined by Eq. (2).
We further impose L2 regularization on the parameters Θ to
avoid overfitting. Gradient descent algorithms can be used
to solve this problem. Embeddings learned in this way will
1) be compatible with all the labeled triples, and 2) absorb
rule knowledge carried by the unlabeled triples.
1
Note that each gpq contains only a single unlabeled triple, i.e.,
the conclusion triple. Take π(gpq |S) defined in Eq. (7) for example.
In this case, s(xu ) = s(eu , rt , ev ) is the soft label to be predicted
and ∇s(xu ) π(gpq |S) = φ(eu , rs , ev ) is a constant w.r.t. S.
Whole Procedure
Algorithm 1 Iterative Learning Procedure of RUGE
Algorithm 1 summarizes the iterative learning procedure of
our approach. To enable efficient learning, we use an online
scheme in mini-batch mode. At each iteration, we sample a
mini-batch Lb , Ub , and Gb from the labeled triples L, unlabeled triples U, and propositionalized rules G, respectively
(line 3).2 Soft label prediction and embedding rectification
are then conducted locally on these mini-batches (line 4 and
line 5 respectively). This iterative procedure captures the interactive nature between embedding learning and logical inference: given current embeddings, logic rules can be used
to perform approximate inference and predict soft labels for
unlabeled triples; these newly labeled triples carry rich rule
knowledge and will in turn help to learn better embeddings.
In this way, knowledge contained in logic rules can be fully
transferred into the learned embeddings. Note also that our
approach is flexible enough to handle soft rules with various
confidence levels extracted automatically from the KG.
Require: Labeled triples L = {(x` , y` )}
Unlabeled triples U = {xu }
FOL rules F={(fp , λp )} and their groundings G={gpq }
Discussions
Datasets. We use two datasets: FB15K and YAGO37. The
former is a subgraph of Freebase containing 1,345 relations
and 14,951 entities, released by Bordes et al. (2013).4 The
latter is extracted from the core facts of YAGO3.5 During
the extraction, entities appearing less than 10 times are discarded. The final dataset consists of 37 relations and 123,189
entities. Triples on both datasets are split into training, validation, and test sets, used for model training, hyperparameter tuning, and evaluation, respectively. We use the original
split for FB15K, and draw a split of 989,132/50,000/50,000
triples for YAGO37.
Note that on both datasets, the training sets contain only
positive triples. Negative triples are generated using the local
closed world assumption (Dong et al. 2014). This negative
sampling procedure is performed at runtime for each batch
of training positive triples. Such positive and negative triples
(along with their hard labels) form our labeled triple set.
We further employ AMIE+ (Galárraga et al. 2015)6 to automatically extract Horn clause rules from each dataset, with
the training set as input. To enable efficient extraction, we
consider rules with length not longer than 2 and confidence
levels not less than 0.8.7 The length of a Horn clause rule is
the number of atoms appearing in its premise, e.g., ∀x, y :
(x, BornInCountry, y) ⇒ (x, Nationality, y) has the
length of 1. And the confidence threshold of 0.8 leads to the
best performance on both datasets (detailed later). Using this
setting, we extract 454 (universally quantified) Horn clause
rules from FB15K, and 16 such rules from YAGO37. Table 1
shows some examples with their confidence levels.
Then, we instantiate these rules with concrete entities, i.e.,
propositionalization. Propositionalized rules whose premise
triples are all contained in the training set (while conclusion
We further analyze the space and time complexity, and discuss possible extensions of our approach.
Complexity. RUGE follows ComplEx to represent entities
and relations as complex-valued vectors, hence has a space
complexity of O(ne d + nr d) which scales linearly w.r.t. ne ,
nr , and d. Here, ne is the number of entities, nr the number of relations, and d the dimensionality of the embedding
space. During the learning procedure, each iteration requires
a time complexity of O(τ (n` d + nu d)), where n` /nu is the
average number of labeled/unlabeled triples in a mini-batch,
and τ the number of inner epochs used for embedding rectification (cf. Eq. (10)). In practice, we usually have nu n`
(see Table 2 for the number of labeled and unlabeled triples
used on our datasets), and we can also set τ to a very small
value, e.g., τ = 1. That means, RUGE has almost the same
time complexity as those most efficient KG embedding techniques (e.g. ComplEx) which require O(n` d) per iteration
during training.3 In addition, RUGE further requires preprocessing steps before training, i.e., rule mining and propositionalization. But these steps are performed only once, and
not required during the iterations.
Extensions. Our approach is quite generic and flexible. 1)
The idea of iteratively injecting logic rules can be applied to
enhance a wide variety of embedding models, as long as an
appropriate scoring function is accordingly designed, e.g.,
the one defined in Eq. (1) by ComplEx. 2) Various types of
rules can be incorporated as long as they can be modeled by
the logical compositions defined in Eq. (3) to Eq. (5), and
we can even use other types of t-norm fuzzy logics to define
such compositions. 3) Rules with different confidence levels
can be handled in a unified manner.
2
We first sample Lb from L. Gb is then constructed by those
whose premise triples are all contained in Lb but conclusion triples
are not. These conclusion triples are further used to construct Ub .
3
Such techniques often use SGD in mini-batch mode for training, and sample a mini-batch of n` labeled triples at each iteration.
1: Randomly initialize entity and relation embeddings Θ(0)
2: for n = 1 : N do
3:
Sample a mini-batch Lb / Ub / Gb from L / U / G
4:
Sb ← SoftLabelPrediction (Ub , Gb , Θ(n−1) )
. cf. Eq. (9)
5:
Θ(n) ← EmbeddingRectification (Lb , Ub , Sb ) . cf. Eq. (10)
6: end for
Ensure: Θ(N )
Experiments
We evaluate RUGE in the link prediction task. This task is
to complete a triple (ei , rk , ej ) with ei or ej missing, i.e., to
predict ei given (rk , ej ) or ej given (ei , rk ).
4
https://everest.hds.utc.fr/doku.php?id=en:smemlj12
http://www.mpi-inf.mpg.de/departments/databases-andinformation-systems/research/yago-naga/yago/downloads/
6
https://www.mpi-inf.mpg.de/departments/databases-andinformation-systems/research/yago-naga/amie/
7
AMIE+ provides two types of confidence, i.e. standard confidence and PCA confidence. This paper uses PCA confidence.
5
Table 1: Horn clause rules with confidence levels extracted
by AMIE+ from FB15K (top) and YAGO37 (bottom).
/location/people born here(x,y)⇒/people/place of birth(y,x)
/director/film(x,y)⇒/film/directed by(y,x)
/film/directed by(x,y)∧/person/language(y,z)⇒/film/language(x,z)
1.00
0.99
0.88
isMarriedTo(x,y)⇒isMarriedTo(y,x)
hasChild(x,y)∧isCitizenOf(y,z)⇒isCitizenOf(x,z)
playsFor(x,y)⇒isAffiliatedTo(x,y)
0.97
0.94
0.86
Table 2: Statistics of datasets, where ne /nr denotes the number of entities/relations, n` /nu /ng is the number of labeled
triples/unlabeled triples/valid groundings used for training,
and nv /nt denotes the number of validation/test triples.
Train
Dataset
FB15K
YAGO37
Valid
Test
ne
nr
n`
nu
ng
nv
nt
14,951
123,189
1,345
37
483,142
989,132
74,707
69,680
96,724
72,670
50,000
50,000
59,071
50,000
triples are not) are taken as valid groundings and used during
embedding learning. We obtain 96,724 valid groundings on
FB15K and 72,670 on YAGO37. Conclusion triples of these
valid groundings are further collected to form our unlabeled
triple set. We finally get 74,707 unlabeled triples on FB15K
and 69,680 on YAGO37. Table 2 provides some statistics of
the two datasets.
Evaluation Protocol. To evaluate the performance in link
prediction, we follow the standard protocol used in (Bordes
et al. 2013). For each test triple (ei , rk , ej ), we replace the
head entity ei with each entity e0i ∈ E, and calculate the score
for (e0i , rk , ej ). Ranking these scores in descending order,
we get the rank of the correct entity ei . Similarly, we can get
another rank by replacing the tail entity. Aggregated over all
test triples, we report three metrics: 1) the mean reciprocal
rank (MRR), 2) the median of the ranks (MED), and 3) the
proportion of ranks no larger than n (HITS@N). During this
ranking process, we remove corrupted triples which already
exist in either the training, validation, or test set, since they
themselves are true triples. This corresponds to the “filtered”
setting in (Bordes et al. 2013).
Comparison Settings. We compare RUGE with four stateof-the-art basic embedding models, including TransE (Bordes et al. 2013), DistMult (Yang et al. 2015), HolE (Nickel,
Rosasco, and Poggio 2016), and ComplEx (Trouillon et al.
2016). These basic models rely only on triples observed in
a KG and use no rules. We further take PTransE (Lin et al.
2015a) and KALE (Guo et al. 2016) as additional baselines.
Both of them are extensions of TransE, with the former integrating relation paths (Horn clauses), and the latter FOL
rules (hard rules) in a one-time injection manner. In contrast,
RUGE incorporates soft rules and transfers rule knowledge
into KG embedding in an iterative manner.
We use the code provided by Trouillon et al. (2016)8 for
8
https://github.com/ttrouill/complex
TransE, DistMult, and ComplEx, and reimplement HolE so
that all these four basic models share the identical mode of
optimization, i.e., SGD with AdaGrad (Duchi, Hazan, and
Singer 2011) and gradient normalization. As such, we reproduce the results of TransE, DistMult, and ComplEx reported
on FB15K (Trouillon et al. 2016), and improve the results of
HolE substantially compared to those reported in the original paper (Nickel, Rosasco, and Poggio 2016).9 The code for
PTransE is provided by its authors.10 We implement KALE
and RUGE in Java, both using SGD with AdaGrad and gradient normalization to facilitate a fair comparison.
There are two types of loss functions that could be used
for these baselines, i.e., the logistic loss or the pairwise ranking loss (Nickel, Rosasco, and Poggio 2016). Trouillon et
al. (2016) have recently demonstrated that the logistic loss
generally performs better than the pairwise ranking loss, except for TransE. So, for TransE and its extensions (PTransE
and KALE) we use the pairwise ranking loss, and for all the
other baselines we use the logistic loss. To extract relation
paths for PTransE, we follow the optimal configuration reported in (Lin et al. 2015a), where paths constituted by at
most 3 relations are included. For KALE and RUGE, we use
the same set of propositionalized rules to make it a fair comparison.11
For all the methods, we create 100 mini-batches on each
dataset, and tune the embedding dimensionality d in {50,
100, 150, 200}, the number of negatives per positive triple α
in {1, 2, 5, 10}, the initial learning rate γ in {0.01, 0.05, 0.1,
0.5, 1.0}, and the L2 regularization coefficient λ in {0.001,
0.003, 0.01, 0.03, 0.1}. For TransE and its extensions which
use the pairwise ranking loss, we further tune the margin δ in
{0.1, 0.2, 0.5, 1, 2, 5, 10}. The slackness penalty C in RUGE
(cf. Eq. (8)) is selected from {0.001, 0.01, 0.1, 1}, and the
number of inner iterations (cf. Eq. (10)) is fixed to τ = 1.
Best models are selected by early stopping on the validation
set (monitoring MRR), with at most 1000 iterations over the
training set. The optimal configurations for RUGE are: d =
200, α = 10, γ = 0.5, λ = 0.01, C = 0.01 on FB15K; and
d = 150, α = 10, γ = 1.0, λ = 0.003, C = 0.01 on YAGO37.
Link Prediction Results. Table 3 shows the results of these
methods on the test sets of FB15K and YAGO37. The results
indicate that RUGE significantly and consistently outperforms all the baselines on both datasets and in all metrics. It
beats not only the four basic models which use triples alone
(TransE, DistMult, HolE, and ComplEx), but also PTransE
and KALE which further incorporate logic rules (or relation
paths) in a one-time injection manner. This demonstrates the
superiority of injecting logic rules into KG embedding, particularly in an iterative manner. Compared to the best performing baseline ComplEx (this is also the model based on
which RUGE is designed), RUGE achieves an improvement
of 11%/18% in MRR/HITS@1 on FB15K, and an improvement of 3%/6% on YAGO37. The improvements on FB15K
9
HolE in its original implementation uses SGD with AdaGrad,
but no gradient normalization.
10
https://github.com/thunlp/KB2E
11
KALE takes all these groundings as hard rules. This approximation works quite well with an appropriate confidence threshold.
Table 3: Link prediction results on the test sets of FB15K and YAGO37. As baselines, rows 1-4 are the four basic models which
use triples alone, and rows 5-6 further integrate logic rules (or relation paths) in a one-time injection manner.
FB15K
YAGO37
HITS@N
HITS@N
Method
MRR
MED
1
3
5
10
MRR
MED
1
3
5
10
TransE
DistMult
HolE
ComplEx
PTransE
KALE
0.400
0.644
0.600
0.690
0.679
0.523
4.0
1.0
2.0
1.0
1.0
2.0
0.246
0.532
0.485
0.598
0.565
0.383
0.495
0.730
0.673
0.756
0.768
0.616
0.576
0.769
0.722
0.793
0.810
0.683
0.662
0.812
0.779
0.837
0.855
0.762
0.303
0.365
0.380
0.417
0.403
0.321
13.0
6.0
7.0
4.0
9.0
9.0
0.218
0.262
0.288
0.320
0.339
0.215
0.336
0.411
0.420
0.471
0.444
0.372
0.387
0.493
0.479
0.533
0.473
0.438
0.475
0.575
0.551
0.603
0.506
0.522
RUGE
0.768
1.0
0.703
0.815
0.836
0.865
0.431
4.0
0.340
0.482
0.541
0.603
Table 4: Runtime (in sec.) on FB15K and YAGO37. Extr. is
the time required for rule/path extraction, Prop. for propositionalization, and Lean. for training per iteration.
FB15K
Method
Figure 2: MRR achieved by RUGE with different confidence
thresholds on the test set of FB15K.
are more substantial than those on YAGO37. The reason is
probably that FB15K contains more relations from which a
good range of rules can be extracted (454 universally quantified rules from FB15K, and 16 from YAGO37).
Influence of Confidence Levels. We further investigate the
influence of the threshold of rules’ confidence levels used
in RUGE. To do so, we fix all the hyperparameters to the
optimal configurations determined by the previous experiment, and vary the confidence threshold in [0.1, 1] with a
step 0.05. Fig. 2 shows MRR achieved by RUGE with various thresholds on the test set of FB15K. We can see that
the threshold of 0.8 is a good tradeoff and indeed performs
best. A threshold higher than that will reduce the number of
rules that can be extracted, while a one lower than that might
introduce too many less credible rules. Both hurt the performance. However, even so, RUGE outperforms ComplEx by
a large margin, with the threshold set in a broad range of
[0.35, 0.9]. This observation indicates that soft rules, even
those with moderate confidence levels, are highly beneficial
to KG embedding despite their uncertainties.
Comparison of Runtime. Finally, we compare RUGE with
ComplEx, PTransE, and KALE in their runtime.12 ComplEx
is a basic model which only requires model training. RUGE
as well as the other two baselines further require preprocessing of rule/path extraction and propositionalization. Table 4
lists the runtime of these methods required for each step on
FB15K and YAGO37. Here, to facilitate a fair comparison,
12
The other three baselines are implemented in Python and much
slower. So they are not considered here.
ComplEx
PTransE
KALE
RUGE
YAGO37
Extr.
Prop.
Learn.
Extr.
Prop.
Learn.
—
868.4
43.1
43.1
—
—
4.0
4.0
11.4
46.5
4.8
14.1
—
13939.5
337.5
337.5
—
—
13.8
13.8
49.5
23.8
27.5
55.2
we set d = 200 (embedding dimensionality) and α = 2 (number of negatives per positive triple) for all the methods. Other
hyperparameters are fixed to their optimal configurations determined in link prediction. We can see that RUGE is still
quite efficient despite integrating additional rules. The average training time per iteration increases from 11.4 to 14.1
on FB15K, and from 49.5 to 55.2 on YAGO37. The preprocessing steps, although performed only once, are also highly
efficient, requiring much less time compared to PTransE.
Conclusion
This paper proposes a novel paradigm that learns entity and
relation embeddings with iterative guidance from soft rules,
referred to as RUGE. It enables an embedding model to learn
simultaneously from labeled triples, unlabeled triples, and
soft rules in an iterative manner. Each iteration alternates
between 1) a soft label prediction stage which predicts soft
labels for unlabeled triples using currently learned embeddings and soft rules, and 2) an embedding rectification stage
which further integrates both labeled and unlabeled triples
to update current embeddings. This iterative procedure may
better transfer the knowledge contained in logic rules into
the learned embeddings. Link prediction results on Freebase
and YAGO show that RUGE achieves significant and consistent improvements over state-of-the-art baselines. Moreover, RUGE demonstrates the usefulness of automatically
extracted soft rules. Even those with moderate confidence
levels can be highly beneficial to KG embedding.
Acknowledgments
The authors would like to thank all the reviewers for their
insightful and valuable suggestions, which significantly improve the quality of this paper. This work is supported by the
National Key Research and Development Program of China
(grants No. 2016YFB0801003 and No. 2016QY03D0503),
the Fundamental Theory and Cutting Edge Technology Research Program of the Institute of Information Engineering,
Chinese Academy of Sciences (grant No. Y7Z0261101), and
the National Natural Science Foundation of China (grant No.
61402465).
References
[Bollacker et al. 2008] Bollacker, K.; Evans, C.; Paritosh, P.;
Sturge, T.; and Taylor, J. 2008. Freebase: A collaboratively created
graph database for structuring human knowledge. In SIGMOD,
1247–1250.
[Bordes et al. 2013] Bordes, A.; Usunier, N.; Garcı́a-Durán, A.;
Weston, J.; and Yakhnenko, O. 2013. Translating embeddings for
modeling multi-relational data. In NIPS, 2787–2795.
[Bordes et al. 2014] Bordes, A.; Glorot, X.; Weston, J.; and Bengio,
Y. 2014. A semantic matching energy function for learning with
multi-relational data. MACH LEARN 94(2):233–259.
[Carlson et al. 2010] Carlson, A.; Betteridge, J.; Kisiel, B.; Settles,
B.; Hruschka Jr, E. R.; and Mitchell, T. M. 2010. Toward an architecture for never-ending language learning. In AAAI, 1306–1313.
[Demeester, Rocktäschel, and Riedel 2016] Demeester,
T.;
Rocktäschel, T.; and Riedel, S. 2016. Lifted rule injection for
relation embeddings. In EMNLP, 1389–1399.
[Dong et al. 2014] Dong, X.; Gabrilovich, E.; Heitz, G.; Horn, W.;
Lao, N.; Murphy, K.; Strohmann, T.; Sun, S.; and Zhang, W. 2014.
Knowledge vault: A web-scale approach to probabilistic knowledge fusion. In SIGKDD, 601–610.
[Duchi, Hazan, and Singer 2011] Duchi, J.; Hazan, E.; and Singer,
Y. 2011. Adaptive subgradient methods for online learning and
stochastic optimization. J MACH LEARN RES 12(Jul):2121–2159.
[Faruqui et al. 2014] Faruqui, M.; Dodge, J.; Jauhar, S. K.; Dyer,
C.; Hovy, E.; and Smith, N. A. 2014. Retrofitting word vectors to
semantic lexicons. arXiv:1411.4166.
[Galárraga et al. 2013] Galárraga, L. A.; Teflioudi, C.; Hose, K.;
and Suchanek, F. M. 2013. AMIE: Association rule mining under
incomplete evidence in ontological knowledge bases. In WWW,
413–422.
[Galárraga et al. 2015] Galárraga, L. A.; Teflioudi, C.; Hose, K.;
and Suchanek, F. M. 2015. Fast rule mining in ontological knowledge bases with AMIE+. VLDB J 24(6):707–730.
[Gardner, Talukdar, and Mitchell 2015] Gardner, M.; Talukdar, P.;
and Mitchell, T. 2015. Combining vector space embeddings with
symbolic logical inference over open-domain text. In AAAI Spring
Symposium Series, 61–65.
[Guo et al. 2015] Guo, S.; Wang, Q.; Wang, L.; Wang, B.; and Guo,
L. 2015. Semantically smooth knowledge graph embedding. In
ACL, 84–94.
[Guo et al. 2016] Guo, S.; Wang, Q.; Wang, L.; Wang, B.; and Guo,
L. 2016. Jointly embedding knowledge graphs and logical rules.
In EMNLP, 192–202.
[Guu, Miller, and Liang 2015] Guu, K.; Miller, J.; and Liang, P.
2015. Traversing knowledge graphs in vector space. In EMNLP,
318–327.
[Hájek 1998] Hájek, P. 1998. The metamathematics of fuzzy logic.
Kluwer.
[Hu et al. 2016] Hu, Z.; Ma, X.; Liu, Z.; Hovy, E.; and Xing, E.
2016. Harnessing deep neural networks with logic rules. In ACL,
2410–2420.
[Lin et al. 2015a] Lin, Y.; Liu, Z.; Luan, H.; Sun, M.; Rao, S.; and
Liu, S. 2015a. Modeling relation paths for representation learning
of knowledge bases. In EMNLP, 705–714.
[Lin et al. 2015b] Lin, Y.; Liu, Z.; Sun, M.; Liu, Y.; and Zhu, X.
2015b. Learning entity and relation embeddings for knowledge
graph completion. In AAAI, 2181–2187.
[Liu et al. 2016] Liu, Q.; Jiang, H.; Evdokimov, A.; Ling, Z.-H.;
Zhu, X.; Wei, S.; and Hu, Y. 2016. Probabilistic reasoning via
deep learning: Neural association models.
[Miller 1995] Miller, G. A. 1995. WordNet: A lexical database for
English. COMMUN ACM 38(11):39–41.
[Neelakantan, Roth, and McCallum 2015] Neelakantan, A.; Roth,
B.; and McCallum, A. 2015. Compositional vector space models for knowledge base completion. In ACL, 156–166.
[Nickel, Rosasco, and Poggio 2016] Nickel, M.; Rosasco, L.; and
Poggio, T. 2016. Holographic embeddings of knowledge graphs.
In AAAI, 1955–1961.
[Nickel, Tresp, and Kriegel 2011] Nickel, M.; Tresp, V.; and
Kriegel, H.-P. 2011. A three-way model for collective learning on
multi-relational data. In ICML, 809–816.
[Rocktäschel et al. 2014] Rocktäschel, T.; Bošnjak, M.; Singh, S.;
and Riedel, S. 2014. Low-dimensional embeddings of logic. In
ACL Workshop on Semantic Parsing, 45–49.
[Rocktäschel, Singh, and Riedel 2015] Rocktäschel, T.; Singh, S.;
and Riedel, S. 2015. Injecting logical background knowledge into
embeddings for relation extraction. In NAACL, 1119–1129.
[Socher et al. 2013] Socher, R.; Chen, D.; Manning, C. D.; and Ng,
A. Y. 2013. Reasoning with neural tensor networks for knowledge
base completion. In NIPS, 926–934.
[Suchanek, Kasneci, and Weikum 2007] Suchanek, F. M.; Kasneci,
G.; and Weikum, G. 2007. YAGO: A core of semantic knowledge.
In WWW, 697–706.
[Trouillon et al. 2016] Trouillon, T.; Welbl, J.; Riedel, S.; Gaussier,
E.; and Bouchard, G. 2016. Complex embeddings for simple link
prediction. In ICML, 2071–2080.
[Vendrov et al. 2015] Vendrov, I.; Kiros, R.; Fidler, S.; and Urtasun, R. 2015. Order-embeddings of images and language.
arXiv:1511.06361.
[Wang and Cohen 2016] Wang, W. Y., and Cohen, W. W. 2016.
Learning first-order logic embeddings via matrix factorization. In
IJCAI, 2132–2138.
[Wang et al. 2014] Wang, Z.; Zhang, J.; Feng, J.; and Chen, Z.
2014. Knowledge graph embedding by translating on hyperplanes.
In AAAI, 1112–1119.
[Wang et al. 2017] Wang, Q.; Mao, Z.; Wang, B.; and Guo, L. 2017.
Knowledge graph embedding: A survey of approaches and applications. IEEE TRANS KNOWL DATA ENG 29(12):2724–2743.
[Wang, Wang, and Guo 2015] Wang, Q.; Wang, B.; and Guo, L.
2015. Knowledge base completion using embeddings and rules.
In IJCAI, 1859–1865.
[Wei et al. 2015] Wei, Z.; Zhao, J.; Liu, K.; Qi, Z.; Sun, Z.; and
Tian, G. 2015. Large-scale knowledge base completion: Inferring via grounding network sampling over selected instances. In
CIKM, 1331–1340.
[Weston et al. 2013] Weston, J.; Bordes, A.; Yakhnenko, O.; and
Usunier, N. 2013. Connecting language and knowledge bases with
embedding models for relation extraction. In EMNLP, 1366–1371.
[Xiao, Huang, and Zhu 2017] Xiao, H.; Huang, M.; and Zhu, X.
2017. SSP: Semantic space projection for knowledge graph embedding with text descriptions. In AAAI, 3104–3110.
[Xie et al. 2016] Xie, R.; Liu, Z.; Jia, J.; Luan, H.; and Sun, M.
2016. Representation learning of knowledge graphs with entity
descriptions. In AAAI, 2659–2665.
[Xie, Liu, and Sun 2016] Xie, R.; Liu, Z.; and Sun, M. 2016. Representation learning of knowledge graphs with hierarchical types.
In IJCAI, 2965–2971.
[Xiong, Power, and Callan 2017] Xiong, C.; Power, R.; and Callan,
J. 2017. Explicit semantic ranking for academic search via knowledge graph embedding. In WWW, 1271–1279.
[Yang et al. 2015] Yang, B.; Yih, W.-t.; He, X.; Gao, J.; and Deng,
L. 2015. Embedding entities and relations for learning and inference in knowledge bases. In ICLR.
[Zhang et al. 2016] Zhang, F.; Yuan, N. J.; Lian, D.; Xie, X.; and
Ma, W.-Y. 2016. Collaborative knowledge base embedding for
recommender systems. In SIGKDD, 353–362.
Knowledge Graph Embedding with Iterative Guidance from Soft Rules:
Supplementary Material
Shu Guo1,2 , Quan Wang1,2,3∗ , Lihong Wang4 , Bin Wang1,2 , Li Guo1,2
Institute of Information Engineering, Chinese Academy of Sciences
School of Cyber Security, University of Chinese Academy of Sciences
3
State Key Laboratory of Information Security, Chinese Academy of Sciences
4
National Computer Network Emergency Response Technical Team & Coordination Center of China
1
2
A
Solving the Soft Label Prediction Problem
We discuss how to solve the soft label prediction problem,
i.e., Eq. (8) in the submission. We first review the problem.
Recall that we are given a set of P FOL rules with their
confidence levels F = {(fp , λp )}P
p=1 , and each FOL rule
Q
p
fp has Qp valid groundings Gp = {gpq }q=1
. The aim is to
predict a soft label s(xu ) ∈ [0, 1] for each unlabeled triple
xu ∈ U, by using the current embeddings Θ(n−1) and all the
groundings G = {Gp }P
p=1 . The optimization problem is:
min
S,ξ
X
1 X
2
(s(xu ) − φ(xu )) + C
ξpq ,
2
p,q
xu ∈U
s.t. λp (1 − π(gpq |S)) ≤ ξpq , q = 1,· · ·, Qp , p = 1,· · ·, P,
ξpq ≥ 0, q = 1,· · ·, Qp , p = 1,· · ·, P,
0 ≤ s(xu ) ≤ 1, ∀s(xu ) ∈ S,
(A-1)
where S = {s(xu )} is the set of soft labels; ξ = {ξpq } the
set of slack variables; φ(xu ) the truth value of xu calculated
by the current embeddings; π(gpq |S) the conditional truth
value of gpq given the soft labels; and C the penalty coefficient. Then, we prove this optimization problem is convex
and has closed-form solutions.
Proof of Convexity
It is easy to see that the objective function is convex, as it
is quadratic w.r.t. each s(xu ) and linear w.r.t. each ξpq . As
for the constraints, note that each valid grounding gpq contains only a single unlabeled triple, i.e., the triple in the conclusion. So the conditional truth value π(gpq |S) is a linear
function w.r.t. a single s(xu ). Take gpq := (eu , rs , ev ) ⇒
(eu , rt , ev ) as an example, where the premise (eu , rs , ev ) is
directly observed, and the conclusion (eu , rt , ev ) is an unlabeled triple with its soft label s(eu , rt , ev ) to be predicted.
The conditional truth value can be calculated as:
π(gpq |S) = φ(eu , rs , ev )·s(eu , rt , ev )−φ(eu , rs , ev )+1,
which is a linear function w.r.t. s(eu , rt , ev ). In this case, all
the constraints are linear functions. So the whole optimization problem is convex.
∗
Corresponding author: Quan Wang ([email protected]).
Closed-Form Solutions
The optimization problem of Eq. (A-1) can be equivalently
written as:
X
1 X
2
min
(s(xu ) − φ(xu )) + C
[λp (1 − π(gpq |S))]+ ,
S
2
p,q
xu ∈U
s.t. 0 ≤ s(xu ) ≤ 1, ∀s(xu ) ∈ S,
(A-2)
where [x]+ = max(x, 0). Given the logical compositions defined in Eq. (3) to Eq. (5) in the submission, and the
constraints φ(·) ∈ [0, 1], s(·) ∈ [0, 1] as well, we will have
π(gpq |S) ≤ 1 for any gpq . The objective function of Eq. (A2) then becomes:
X
1 X
2
P=
(s(xu ) − φ(xu )) + C
λp (1 − π(gpq |S)) .
2
p,q
xu ∈U
Recall that each π(gpq |S) is a linear function w.r.t. a single
s(xu ). So P is an upward parabola w.r.t. each s(xu ), with
the axis of symmetry given as:
X
s(xu ) = φ(xu ) + C
λp ∇s(xu ) π(gpq |S).
p,q
Here, ∇s(xu ) π(gpq |S) means the gradient of π(gpq |S) w.r.t
s(xu ), which is a constant w.r.t. S. By further considering
the constraint 0 ≤ s(xu ) ≤ 1 for each s(xu ) ∈ S, we get
the optimal solutions:
h
i1
X
s(xu ) = φ(xu ) + C
λp ∇s(xu ) π(gpq |S) , ∀xu ∈ U,
p,q
0
where
= min(max(x, 0), 1) is a truncation function
enforcing the solutions to stay within [0, 1].
[x]10
| 2cs.AI
|
Hybrid Automata and ε-Analysis on a Neural Oscillator∗
Alberto Casagrande
Dept. of Matematics and Geoscience
University of Trieste, Italy
[email protected]
Tommaso Dreossi
Carla Piazza
Dept. of Matematics and Computer Science
University of Udine, Italy
[email protected]
[email protected]
In this paper we propose a hybrid model of a neural oscillator, obtained by partially discretizing a
well-known continuous model. Our construction points out that in this case the standard techniques,
based on replacing sigmoids with step functions, is not satisfactory. Then, we study the hybrid model
through both symbolic methods and approximation techniques. This last analysis, in particular, allows us to show the differences between the considered approximation approaches. Finally, we focus
on approximations via ε-semantics, proving how these can be computed in practice.
Introduction
Neural oscillations are rhythmic and repetitive electrical stimuli which play an important role in the
activities of several brain regions. Some examples of brain locations in which it has been demonstrated
the central role of neural oscillations are the cortex [17], the thalamus [18], and the olfactory information
processing [8]. With the aim of understanding neurophysiological activities, we propose the modeling of
oscillatory phenomena exploiting hybrid automata.
A continuous model of a single oscillator based on an ordinary differential system has been proposed
in [19]. Even if this is a simple model, its analysis and the analysis of its composition in multiple copies
is limited due to the non-linearity of the ordinary differential system involved. For this reason, we are
interested in the development of a piecewise affine hybrid automaton which correctly approximates the
continuous model and on which automatic analysis and composition can be made.
Trying to linearize the non-linear components of the original continuous model, we first replace in
the standard way the sigmoidal behaviours with sign functions [9, 14], approximating continuos signals
with discrete off-on signals. Unfortunately, the behaviour of this model differs from the original one. For
this reason, we propose a more sophisticated approximation of sigmoidals based on a piecewise linear
function exploited in the development of a hybrid automaton which simulates a single oscillator.
It is well known that the reachability problem over hybrid automata is source of undecidability.
Moreover, the exact computation of the reachable sets of hybrid automata, which represents the basis of
the automatic analysis of the models, does not always reflect the behaviour of the real modeled systems.
This is due to the fact that real systems are often subject to noise, thus their evolutions do not correspond
to a single precise formalization. This has been noticed already in [19] in the specific case of the continuous model of the neural oscillator. For these reasons, we study our hybrid automaton exploiting different
approximation techniques that introduce noise.
In the literature, several approximation techniques have been proposed (see, e.g., [7, 16, 10, 3, 15]).
Fränzle in [7] presents a model of noise over hybrid automata. The introduction of noise ensures in many
cases the (semi-)decidability of the reachability problem. Another result of (semi-)decidability always
based on the concept of perturbation and concerning the safety verification of hybrid systems is given by
∗ This
work has been partially supported by Istituto Nazionale di Alta Matematica (INdAM).
Ezio Bartocci and Luca Bortolussi (Eds.): HSB 2012
EPTCS 92, 2012, pp. 58–72, doi:10.4204/EPTCS.92.5
c A. Casagrande, T. Dreossi & C. Piazza
This work is licensed under the
Creative Commons Attribution License.
A. Casagrande, T. Dreossi & C. Piazza
59
Ratschan in [16]. Furthermore, ε-(bi)simulation [10] relations, which are essentially relaxations on the
infinite precision required by simulation and bisimulation, represent tools able to remove complexity and
undecidability issues related to the analysis of the investigated model. Moreover, in [3] it is presented a
different approach based on the reinterpretation of the standard semantics of the formulæ which compose
hybrid automata. Exploiting this new class of semantics, called ε-semantics, the authors provide a result
of decidability of the reachability problem over hybrid automata with bounded invariants.
In this paper, we focus precisely on the approximation approach based on the ε-semantics. In particular, we propose a translation that allows us to reduce the ε-semantics evaluation to the standard
semantics evaluation, computable by exploiting tools for cylindrical algebraic decomposition. Then, we
present some properties which have been automatically tested on the neural oscillator by applying such
translation. Hence, in this work, we prove both that ε-semantics better represents the real behaviour of
the neural oscillator, than the standard one, and that the approach is effective.
The paper is organized as follows: Section 1 gives some basic definitions concerning logics and hybrid automata; Section 2 is dedicated to the mathematical modeling of the neural oscillator; in Section 3
we present different approximation techniques based on noise, perturbation, approximate (bi)simulations,
and ε-semantics. Section 4 exposes some considerations regarding the application of the previously presented approximation approaches to the investigated model. Finally, in Section 5, we first define a
translation which make effectively computable the ε-semantics, then we experimentally exploit it in the
analysis of the hybrid automaton which models the neural oscillator. All the proofs can be found at
http://www.dimi.uniud.it/piazza/hsb2012_extended.pdf.
1
1.1
Hybrid Automata
Preliminaries
We formally define hybrid automata by using first-order languages and, because of that, we first need to
introduce some basic notions and our notation.
We use X, Xi , Y , Yi , W , and Wi to denote real variables and X, Xi , Y, Yi , W, and Wi to denote tuple
of real variables. We always assume that all the variables that occur bound in a formula do not occur free
and vice versa. This enables us to label variables, rather than occurrences, as free or bound. We write
ϕ[X1 , . . . , Xm ] to stress the fact that X1 , . . ., Xm are free in ϕ. By extension, ϕ[X1 , . . . , Xn ] indicates that
the components of vectors X1 , . . ., Xn are free in ϕ.
The formula obtained from ϕ[X1 , . . . , Xm ] by replacing Xi by s0 , where s0 is either a constant or a variable, is denoted by ϕJXi /s0 K. By extension, ϕJXi . . . Xi+n /s0 . . . sn K indicates the formula obtained from
ϕ[X1 , . . . , Xm ] by simultaneously replacing all the variables Xi . . . Xi+n by s0 . . . sn . If X = hXi , . . . , Xi+n i,
~s0 = hs0 , . . . , sn i, ~s1 = hsn+1 , . . . , s2∗n+1 i, and ; is a relational symbol (e.g., = or ≥), then we
may
V
write ϕJX/~s0 K in place of ϕJXi . . . Xi+n /s0 . . . sn K and ~s0 ; ~s1 in place of j∈[0,n] s j ; s j+k∗n (e.g.,
h7, Xi = h2, 3i means 7 = 2 ∧ X = 3). Finally, if ϕ[X1 , . . . , Xi , . . . , Xn ] then we may denote the formula
ϕJXi /~sK by writing ϕJX1 , . . . ,~s, . . . , Xn K.
The semantics of a formula is defined in the standard way (see. e.g., [6]). Given a set Γ of sentences
and a sentence ϕ, we say that ϕ is a logical consequence of Γ (denoted, Γ |= ϕ) if ϕ is valid in any
model M in which each formula of Γ is valid too (M |= Γ). A theory T is a set of sentences such that
if T |= ϕ, then ϕ ∈ T . A theory T admits the so-called elimination of quantifiers, if, for any formula
ϕ, there exists in T a quantifier free formula ρ such that ϕ is equivalent to ρ with respect to T . If there
exists an algorithm for deciding whether a sentence ϕ belongs to T or not, we say that T is decidable.
Hybrid Automata and ε-Analysis on a Neural Oscillator
60
def
Example 1. Consider the formula ϕ = ∃X (a ∗ X 2 + b ∗ X + c = 0). It is well known that ϕ is in the
theory of reals with +, ∗, and ≥ if and only if the unquantified formula b2 − 4ac ≥ 0 holds.
1.2
Syntax, Semantics, and Reachability
A hybrid automaton is an infinite state automaton that consists in a set of continuos variables and a finite
directed graph. Each node of a graph is labelled by both an invariant condition and a dynamic law, while
all the edges are tagged with an activation region and a reset map. The continuous variables evolve
according to the dynamic law of the current node of the graph and the node’s invariant condition must be
satisfied along all the evolution. An edge is crossable if and only if the variable values are included the
activation region and, when a hybrid automata jumps over it, the associated reset map is applied.
Definition 1 (Hybrid Automata - Syntax). A hybrid automaton H of dimension d(H) ∈ N is a tuple hX,
X0 , V, E, Inv, f· , Act, Resi where:
• X = hX1 , . . ., Xn i and X0 = hX10 , . . ., Xn0 i are two vectors of variables ranging over the reals R;
• hV, Ei is a directed finite graph, i.e., E ⊆ V × V. Each element of V will be dubbed location;
• Each location v ∈ V is labelled by both a formula Inv(v)[X], called invariant, and a continuous
function fv : Rn −→ (R≥0 −→ Rn ), called dynamics or flow function. The dynamics may be
specified either by differential equations, i.e., fv is the solution of a given Cauchy problem, or by
a logic formula. We use the formula Dyn(v)[X, X0 , T ], where T is a temporal variable ranging in
def
R≥0 , to denote the dynamics on v, i.e., Dyn(v)[X, X0 , T ] = X0 = fv (X)(T );
• Each e ∈ E is labelled by the formulæ Act(e)[X] and Res(e)[X, X0 ] which are called activation and
reset, respectively.
If all the formulæ that define a hybrid automaton H belong to the same logical theory T , then we
say that H is definable in T or that H is a T hybrid automaton.
The semantics of any hybrid automaton can be specified as a transition system that is composed by
two different relations miming the double nature of the hybrid automaton itself: the continuous reachability transition relation and the discrete reachability transition relation.
Definition 2 (Hybrid Automaton - Semantics). A state a of H is a pair hv, ri, where v ∈ V is a location
and r ∈ Rd(H) is an assignment of values for the variables of X. A state hv, ri is said to be admissible if
Inv(v)JrK holds.
t
The continuous transition relation →
− C between admissible states, where t ≥ 0 denotes the transition
elapsed time, is defined as follows:
t
hv, ri →
− C hv, si ⇐⇒ r = fv (r)(0), s = fv (r)(t), and Inv(v)J fv (r)(t 0 )K hold for each t 0 ∈ [0,t].
e
The discrete transition relation →
− D among admissible states is:
e
hv, ri →
− D hv, si
⇐⇒
e = hv, v0 i and both Act(e)JrK and Res(e)Jr, sK hold.
t
We write a →C a0 and a →D a0 to mean that there exists a t ∈ R≥0 such that a →
− C a0 and that there
e
0
exists an e ∈ E such that a →
− D a , respectively.
Definition 3 (Reachability). Let I be either N or an initial finite interval of N. A trace of H is a sequence
of admissible states a0 , a1 , . . . , a j , . . . , with j ∈ I , such that ai−1 → ai holds for all i ∈ [1, n] and either
ai−2 →C ai−1 →D ai , ai−2 →D ai−1 →D ai , or ai−2 →D ai−1 →C ai for each i ∈ I \ {0}1 .
The automaton H reaches a state an from a state a0 if there exists a trace a0 , . . . , an . In such a case,
we also say that an is reachable from a0 in H.
1 This
last condition supports not transitive dynamics. See [4] for a complete discussion.
61
A. Casagrande, T. Dreossi & C. Piazza
The problem of deciding whether a hybrid automaton H reaches a set of states S from a set of states
R is known as the reachability problem of S from R over H. A trace produced by an infinite sequence
of discrete transitions during a bounded amount of time is called Zeno trace and every hybrid automaton
allowing such kind of trace is said to have a Zeno behaviour.
Example 2. Let us consider a hybrid automaton Hb modeling a bouncing ball whose collisions are
inelastic. The automaton is equipped with two continuous variables X1 and X2 that represent ball’s
Ẋ1 = X2
Ẋ2 = −g
X10 = X1
X20 = −γX2
Figure 1: Bouncing ball hybrid automaton.
elevation and velocity, respectively. The dynamics, resets, and discrete structure of Hb are presented in
Fig. 1. The two coefficients g and γ are the standard gravity and the coefficient of restitution, respectively.
The activation formula of the automaton edge is “X1 = 0”.
Imposing as starting height h0 = 10m and as coefficient of restitution γ = 0.86, the bounce peaks
decrease at each iteration and the automaton Hb has a Zeno behaviour.
As the halting problem for the two counter machine can be reduced to the reachability problem of
a particular class of hybrid automata, the reachability problem for hybrid automata itself is not always
decidable [1]. However, if H is a T -hybrid automaton and T is a first-order decidable theory, then
the reachability through a bounded number of discrete transitions can be characterized with a first-order
decidable formula (see e.g., [4]). In particular, in the case of automata defined through polynomials over
the reals, we can use cylindrical algebraic decomposition tools to decide bounded reachability.
2
Neural Oscillator: Continuous and Hybrid Models
Oscillatory electrical stimuli have been considered central for the activities of several brain regions since
the begin of the ’80s. It was shown that they play an important role in the olfactory information processing [8] and they were observed in the thalamus [18], and in the cortex [17]. Many studies suggested
that, in the mammalian visual system, neurons signals may be group together through in-phase oscillations [11]. Because of this, the development and analysis of models representing oscillatory phenomena
assume a great importance in understanding the neurophysiological activities.
A simple continuous model of a single oscillator has been proposed in [19]. The model describes
the evolutions of one excitatory neuron (Ne ) and one inhibitory neuron (Ni ) by mean of the ordinary
differential system.
Ẋe = − Xτe + tanh (λ ∗ Xe ) − tanh (λ ∗ Xi )
f (τ, λ ) :
,
(1)
Ẋi = − Xτi + tanh (λ ∗ Xe ) + tanh (λ ∗ Xi )
where Xe and Xi are the output of Ne and Ni , respectively, τ is a characteristic time constant, and λ > 0
is the amplification gain. Hopf bifurcation characterizes a qualitative change in the evolution of f (τ, λ ):
if τ ∗ λ ≤ 1, then the point h0, 0i is the unique global attractor of the system, if, otherwise, τ ∗ λ > 1,
the origin is an unstable equilibrium and all the evolutions converge to a limit cycle attractor [2]. A
simulation of f (3, 1) is represented in Fig. 4(a).
Hybrid Automata and ε-Analysis on a Neural Oscillator
62
Xi
v1
v0
Trajectory from (-0.5,-0.5)
Trajectory from (-1.0,6.0)
6
Ẋe = − Xτe
Ẋi = − Xτi
Ẋe = − Xτe
Ẋi = − Xτi + 2
−2
v2
4
2
Xe
v3
0
2
Ẋe = − Xτe
Ẋi = − Xτi − 2
Ẋe = − Xτe + 2
Ẋi = − Xτi
4
6
6
(a) The automaton has 4 locations. The dashed lines
denote both the boundaries of the invariants and the
activation regions. The resets are identify functions.
4
2
0
2
4
6
(b) Direction field and evolution of the automaton from the two
points h− 12 , − 21 i and h−1, 6i when τ = 3. The automaton has
four attractors, is not periodic, and its principal axes are stable.
Figure 2: The piecewise hybrid automaton associated to the function fˆ(τ).
Even if f (τ, λ ) is rather simple, the ability of analyzing a complex system obtained by composing
multiple copies of this model is limited due to the non-linearity of f (τ, λ ) itself. For this reason, we are
interested in the development of a piecewise affine hybrid model whose behaviour fairly approximates
System (1) and that can be automatically analyzed and composed.
Since the non-linear components in System (1) have the form tanh (λ ∗ X), we try to linearize such
function. In the case of genetic networks it is quite standard to approximate sigmoidal behaviours (e.g.,
tanh) through the sign function sgn [9, 14]. Such approximation replaces a continuous signal with a
discrete off-on one. In our case, by replacing tanh (λ ∗ X) with sgn (X) in System (1), we obtain the
following differential system:
Ẋe = − Xτe + sgn (Xe ) − sgn (Xi )
ˆf (τ) :
,
(2)
Ẋi = − Xτi + sgn (Xe ) + sgn (Xi )
which corresponds to the piecewise hybrid model depicted in Fig. 2(a). Unfortunately, the behaviour
of this model is quite different from that of System (1), as we can see comparing the simulation in
Fig. 2(b) with that of Fig. 4(a). In particular, the model based on fˆ(τ) has four attractors with coordinates
h−2 ∗ τ, 0i, h0, −2 ∗ τi, h2 ∗ τ, 0i, and h0, 2 ∗ τi, it is not periodic, and its principal axes are stable.
A more sophisticated approximation of tanh (λ ∗ X) is the piecewise linear function:
if z < − αλ
−1
def
λ
(3)
hλ ,α (z) =
∗ z if − αλ ≤ z < αλ ,
α
1
if z ≥ αλ
where α is the approximation coefficient which determines the slope of the central segment. The substitution of tanh (λ ∗ X) with hλ ,α (z) in System (1) leads to the system:
Ẋe = − Xτe + hλ ,α (Xe ) − hλ ,α (Xi )
˜fα (τ, λ ) :
,
(4)
Ẋi = − Xτi + hλ ,α (Xe ) + hλ ,α (Xi )
63
A. Casagrande, T. Dreossi & C. Piazza
whose corresponding hybrid automaton is depicted in Fig. 3. A different approximation could be obtained by using the technique proposed in [12].
Xi
v8
v7
v6
−α
Ẋe = τ∗λ
τ∗α Xe − 1
λ
Ẋi = α Xe − Xτi + 1
Ẋe = − Xτe − 2
Ẋi = − Xτi
v1
Ẋe = − Xτe
Ẋi = − Xτi + 2
v0
Ẋe = − Xτe − αλ Xi − 1
−α
Ẋi = τ∗λ
τ∗α Xi − 1
v2
λ
−α
Ẋe = τ∗λ
τ∗α Xe − α Xi
−α
Ẋi = αλ Xe + τ∗λ
τ∗α Xi
v3
Ẋe = − Xτe − αλ Xi + 1
−α
Ẋi = τ∗λ
τ∗α Xi + 1
Xe = − αλ
α
λ
Xe
Xi = − αλ
v4
−α
Ẋe = τ∗λ
τ∗α Xe + 1
Ẋi = αλ Xe − Xτi − 1
Ẋe = − Xτe
Ẋi = − Xτi − 2
Xi =
v5
Ẋe = − Xτe + 2
Ẋi = − Xτi
Xe =
α
λ
Figure 3: A graphical representation of the hybrid automaton H f˜ associated to the function f˜α (τ, λ ).
In the rest of the paper we present some general techniques for studying hybrid automata and then
we apply them to H fˆ to formally prove its properties.
3
3.1
Approximation Techniques
Noise and Disturbed Automata
The density of continuous variables provide an unbounded quantity of memory within a bounded region.
As a matter of fact, the undecidability results proved in [13] are based on the possibility of embedding N
in (0, 1] ⊆ R through the function f (n) = 2−n .
However, Fränzle in [7] observed that noise disturbes the trajectiories of real hybrid systems, augmenting the set of reachable points. Hence, in [7] a model of noise has been presented over hybrid
automata. Remarkably, the introduction of noise ensures in many cases the (semi-)decidability of the
reachability problem.
Our definitions of hybrid automata slightly differ from the ones in [7]. In particular, as far as the
syntax is concerned, the formulæ Act(e)[X] and Res(e)[X, X0 ] are glued together in a formula called
transe [X, X0 ]. Moreover, our formulæ Inv(v)[X] and Dyn(v)[X, X0 , T ] are replaced by a single formula
Hybrid Automata and ε-Analysis on a Neural Oscillator
64
Trajectory from (-0.5,-0.5)
Trajectory from (-1.0,6.0)
6
4
4
2
2
0
0
2
2
4
4
6
Trajectory from (-0.5,-0.5)
Trajectory from (-1.0,6.0)
6
6
6
4
2
0
2
4
6
(a) An evolution of f (3, 1).
6
4
2
0
2
4
6
(b) An evolution of f˜2 (3, 1).
Figure 4: Direction field and evolution of the models discussed in Section 2.
actv [X, X0 ] whose meaning in our framework is:
∃T (T ≥ 0 ∧ X0 = fv (X)(T ) ∧ ∀T 0 (0 ≤ T 0 ≤ T → Inv(v) fv (X)(T 0 ))),
i.e., the formula actv [X, X0 ] syntactically ensures the existence of a continuous transition. Exploiting
these relationships between our hybrid automata and the hybrid automata defined through the formulæ
actv [X, X0 ] and transe [X, X0 ], we can reformulate the results presented in [7] in our framework.
Definition 4. Given a hybrid automaton H = hX, X0 , V, E, Inv, f· , Act, Resi we say that the hybrid
f ef· , Act, Resi is a disturbed variant of H if for each pair of states a, a0 if
e = hX, X0 , V, E, Inv,
automaton H
e
a →C a0 in H, then a →C a0 in H.
e is a disturbance of noise level ε or more if
Moreover, let δ be a distance over Rd(H) and ε ∈ R>0 . H
0
0
e
for each s, s such that δ (s, s ) < ε it holds that if hv, ri →C hv, si in H, then hv, ri →C hv, s0 i in H.
e are
Intuitively, when there are no bifurcation behaviours, a small ε ensures that the dynamics of H
close to those of H.
In [7] it has been proved that in the case of bounded invariants there exists a finite computable index
i ∈ N such that the reachability over H can be over-approximated with reachability within i discrete
jumps over a disturbance of noise level ε or more of H.
e be T hybrid automata, with H
e
Theorem 1. [7] Let T be a decidable first-order theory. Let H, H
disturbance of noise level ε or more of H for some ε ∈ R>0 . There exists i such that for all pairs of
e within i discrete transitions. Moreover, i can be
states a, a0 if a reaches a0 in H, then a reaches a0 in H
effectively computed.
Unfortunately, there are cases in which the over-approximation is always strict, no matter how small
ε is. Such automata are called fragile, in contrast with robust automata, where if a does not reach a0
e disturbance of noise level ε such that a does not reach a0 in H.
e As a
in H, there exist ε ∈ R>0 and H
consequence, reachability is decidable over robust automata, while it is only semi-decidable over fragile
automata. It is not possible to decide whether a hybrid automaton is robust or fragile. Intuitively, since
real world systems are always subject to noise, a hybrid automaton is a reliable model of the real system
only if it is robust. So, it is fundamental to develop and exploit design techniques which ensure robustness
of the resulting hybrid automata.
65
A. Casagrande, T. Dreossi & C. Piazza
3.2
Approximate Bisimulations and Simulations
Since the 90’s, simulation and bisimulation have been successfully used to investigate hybrid automata.
However, due to the infinite precision required to relate different evolutions, these tools are able to remove
neither the complexity nor the undecidability issues that may affect the analysis of the investigated model.
The ε-(bi)simulation relations [10] relaxes these infinite precision requirements by relating system evolutions whose maximal distance is less than a given ε. This enables us to simplify both the dynamics
and the resets of the investigated automaton. Moreover, provided an observation map hh·ii : Rd(H) −→ Rd
that associates the internal status of an automaton H to the values measurable by an external observer,
ε-(bi)simulations allow to relate the “visible” behaviours of H to the behaviours of an automaton whose
dimension is smaller than d(H).
Any pair of hybrid automata H1 to H2 related by an ε-simulation must have the same discrete structure
and share the same locations V and edges E by definition.
Definition 5. Let Hi = hXi , Xi 0 , V, E, Invi , f·,i , Acti , Resi i be a hybrid automaton for each i ∈ {1, 2}.
Moreover, let ε be in R≥0 . A relation Sε ⊆ (V × Rd(H1 ) ) × (V × Rd(H2 ) ) is an approximate simulation
relation of H1 by H2 of precision ε if, for all hhv1 , r1 i , hv2 , r2 ii ∈ Sε :
1. v1 = v2 = v;
2. k hhr1 ii1 − hhr2 ii2 k ≤ ε;
t
t
3. if hv, r1 i →
− C hv, r10 i in H1 , there exists r20 s.t. hv, r2 i →
− C hv, r20 i in H2 and hhv, r10 i , hv, r20 ii ∈ Sε ;
e
e
4. if hv, r1 i →
− D hv0 , r10 i in H1 , there exists r20 s.t. hv, r2 i →
− D hv0 , r20 i in H2 and hhv0 , r10 i , hv0 , r20 ii ∈ Sε .
The automaton H2 approximately simulates H1 with precision ε if there exists an approximate simulation relation of H1 by H2 of precision ε. An approximate simulation relation Sε of H1 by H2 of
precision ε is an approximate bisimulation relation between H1 and H2 of precision ε if the relation
Sε−1 = {ha2 , a1 i | ha1 , a2 i ∈ Sε } is an approximate simulation relation of H2 by H1 of precision ε.
Many methods have been developed to automatically compute approximate simulation relations between systems such as constrained linear systems, autonomous nonlinear systems, and hybrid systems.
3.3 ε-Semantics
The undecidability of the reachability problem over hybrid automata having bounded invariants is a
direct consequence of the ability of characterizing dense regions of arbitrarily small size. As noticed
in [3], especially in the study of biological systems, such ability may result misleading. As a matter
of the fact, the continuous quantities used in hybrid automata are very often abstractions of large, but
discrete, quantities. In such cases, the ability of handling values with infinite precision is a model artifact
rather than a real property of the original system.
In order to discretize the continuous space, we introduce the concept of ε-sphere. Given a set S ⊆
n
R , the ε-sphere B(S, ε) is the subset of Rn of points at distance less than ε from S, i.e., B(S, ε) =
{q ∈ Rn | ∃p ∈ S(δ (p, q) < ε)}, where δ is a distance function over Rn (e.g., the standard euclidean
distance). Moreover, given a hybrid automaton H and an initial set of points I ⊆ Rd(H) , the set of points
S
reachable from the set I by H, denoted by RSetH (I), is characterized by RSetH (I) = i∈N RSetiH (I) =
limi→+∞ RSetiH (I) where RSetiH (I) is the set of points reachable from I in at most i discrete transitions.
Theorem 2 ([3]). Let T be a decidable first-order theory over reals and H be a T hybrid automaton
with bounded invariants. If there exists ε ∈ R>0 such that, for each I ⊆ Rd(H) and for each i ∈ N, either
i+1
i+1
(I) \ RSetiH (I), then
RSetH
(I) = RSetiH (I) or there exists an ai ∈ Rd(H) such that B({ai }, ε) ⊆ RSetH
j
i
there exists j ∈ N such that RSetH (I) = RSetH (I) and the reachability problem over H is decidable.
Hybrid Automata and ε-Analysis on a Neural Oscillator
66
This result finds applications when it makes no sense to distinguish measurements smaller than ε.
Hence, since hybrid automata characterization is based on first-order fromulæ, it seems reasonable to
reinterpret the semantics of semi-algebraic automata by giving to each formula a semantics of “dimension
of at least ε”. In [3] the authors introduce a new class of semantics, called ε-semantics, which guarantee
the decidability of reachability in the case of hybrid automata with bounded invariants.
Definition 6. Let T be a first-order theory and let ε ∈ R>0 . For each formula ψ on T let {|ψ|}ε ⊆ Rd ,
where d is the number of free variables of ψ, be such that:
(ε) either {|ψ|}ε = 0/ or there exists p ∈ Rd such that B ({p}, ε) ⊆ {|ψ|}ε ;
(∩) {|φ ∧ ϕ|}ε ⊆ {|φ |}ε ∩ {|ϕ|}ε ;
(∪) {|φ ∨ ϕ|}ε = {|φ |}ε ∪ {|ϕ|}ε ;
(¬) {|ψ|}ε ∩ {|¬ψ|}ε = 0.
/
r∈R ψJr, XK|}ε ;
(∀) {|∀Xψ[X, X]|}ε = {|
V
r∈R ψJr, XK|}ε ;
(∃) {|∃Xψ[X, X]|}ε = {|
W
Any semantics satisfying the above conditions is said to be an ε-semantics for T .
Let us notice that no ε-semantics can over-approximate the standard semantics. Infact, for any theory
T , if {|φ |} ⊂ {|φ |}ε , where {|·|} is the semantics associated to T , then {|¬φ |} ⊃ {|¬φ |}ε due to the (¬)-rule.
Example 3 (The sphere semantics). Let T be a first-order theory over the reals and let ε > 0. The
sphere semantics of ψ, (|ψ|)ε , is defined by structural induction on ψ as follows:
def
def V
• (|t1 ◦ t2 |)ε = B ({|t1 ◦ t2 |}, ε), for ◦ ∈ {=, <};
• (|∀XψJX, XK|)ε = (| r∈R ψJr, XK|)ε ;
def
• (|ψ1 ∧ ψ2 |)ε =
def
S
B({p},ε)⊆(|ψ1 )| ε ∩(|ψ2 )| ε
• (|ψ1 ∨ ψ2 |)ε = (|ψ1 |)ε ∪ (|ψ2 |)ε ;
B ({p}, ε);
def
• (|∃XψJX, XK|)ε = (|
def
• (|¬ψ|)ε =
r∈R ψJr, XK|)ε ;
W
B({p},ε)∩(|ψ|)ε =0/ B ({p}, ε).
S
ε-semantics are exploited in the reachability algorithm defined in [3]. Intuitively, the algorithm
computes reachability incrementing the number of allowed discrete transitions at each iteration. New
reachable sets of points are computed until they became too small to be captured by the ε-semantics. In
the case of hybrid automata with bounded invariants, it always terminates. Finally, notice that replacing the ε-semantics with the standard one, the above algorithm could not terminate even with bounded
invariants, due to Zeno behaviours.
To conclude, let us notice that, by Taylor’s approximation, for any ε and any differentiable function
f (t), if we fix a time horizon th , we can approximate f (t) by a polynomial p(t) such that k f (t)− p(t)k < ε
for all t ∈ [0,th ]. Hence, there are ε-semantics that are not able to distinguish p(t) and f (t) and, according
to such ε-semantics, we can use p(t) in place of f (t) for all t ∈ [0,th ]. It follows that, adopting an
opportune ε-semantics, the Tarski’s theory (i.e., the first order theory of the inequalities over the reals) is
enough to represent any differentiable function f (t).
4
Approximated Analysis over Neural Oscillator
In this section we try to understand what happens when we apply the approximation techniques described
in Section 3 to our neural oscillator hybrid model H f˜ presented in Section 2.
4.1 ε-disturbance
The automaton H f˜ presents two main behaviours: h0, 0i is an unstable equilibrium; each starting point
different from h0, 0i reaches the limit cycle. For this reason we can prove that H f˜ is fragile. As a matter
A. Casagrande, T. Dreossi & C. Piazza
67
ff˜ is a disturbance of noise level ε of H f˜, then (0, 0) in H
ff˜ reaches points different from
of fact, if H
h0, 0i, while in H f˜ it does not. In other words, if we consider backward reachability, h0, 0i is backward
ff˜ from a region R 6= {(0, 0)}, while in H f˜ it is backward reachable from {h0, 0i}. This is
reachable in H
not due to the fact that h0, 0i is unstable, but to the presence of two limit behaviours over a connected
region. We recall that in a piecewise hybrid automaton the invariants are connected disjoint regions
whose union is connected and the resets are identities, i.e., the trajectories are continuos. We use the
term limit behaviour of a hybrid automaton to denote both equilibria and limit cycles.
Theorem 3. Let H be a piecewise hybrid automaton presenting at least two different limit behaviours.
If from each point there is a unique possible evolution, then H is fragile.
The above result points out that there are systems for which it is not possible to define robust models.
In [16] a model is said to be safe only if it remains safe under small disturbances. In this terms our
result shows that there are systems which do not admit a safe model. This does not means that they
are not interesting or that we need to remove some of their behaviours. This simply means that such
systems have to be studied applying some form of disturbance or approximation. As the matter of facts,
if we study them by applying standard semantics, we define a precise border between the points reaching
different behaviours. Such precise border is not realistic.
4.2 ε-(bi)simulations
The automaton H f˜ has both an unstable equilibrium in h0, 0i and a single limit cycle encompassing the
origin of the axes. Because of that we are guaranteed that, during its evolutions, H f˜ decreases the distance
of its state from the limit cycle regardless of the starting state s 6= hv0 , h0, 0ii. Since all the differential
equations defining the dynamics of H f˜ are continuous, we can define an ε-simulation between states
whose distance from the limit cycle is smaller than ε. This enables us to both approximate the non-linear
differential System (1) with a linear differential system and reduce the complexity of the analysis.
However, if d is the maximum Euclidean distance between h0, 0i and the cycle limit, no ε-(bi)simulation, with ε < d, can relate hv0 , h0, 0ii with any other state of H f˜. As a matter of fact, the points
belonging to any neighborhood of h0, 0i eventually converge to the limit cycle. It follows that hv0 , h0, 0ii
is a singularity of the model and, despite the original system always reaches a periodic evolution, any
approximation of the proposed model by mean of ε-(bi)simulation does not manifest this property.
4.3 ε-semantics
In order to exploit ε-semantics for the study of H f˜ the first step we have to perform is that of approximating through polynomials the solutions of the differential equations defining the semantics. This can be
done, for instance, by using Taylor polynomials or more sophisticated numerical integration techniques.
We do this in the next section, where we also apply cylindrical algebraic decomposition tools to automatically prove properties of our model. Here instead we try to infer some general results about the use
of ε-semantics on H f˜.
H f˜ has an unstable equilibrium in h0, 0i. This means that h0, 0i reaches {h0, 0i}. However, when we
compute the set of points reachable from h0, 0i through an ε-semantics we get either the empty set or
a set having diameter at least ε. In particular, if our ε-semantics under-approximates the standard one
(i.e., if {|φ |} ⊃ {|φ |}ε , where {|·|} is the semantics associated to chosen theory, for each formula φ ), then
we get the empty set. Otherwise, both cases are possible, depending on the ε-semantics. For instance,
in the case of sphere semantics, no matter how we approximate the dynamics, we get that h0, 0i reaches
Hybrid Automata and ε-Analysis on a Neural Oscillator
68
a set having diameter at least ε. Similarly, unless we use an under-approximation ε-semantics or some
unusual metrics, the limit cycle is transformed into a limit flow tube. This means that, if we consider a
point on the limit cycle and we compute the set of points reachable from such point, then we do not only
obtain the limit cycle, but at least a flow tube which includes the limit cycle. We will see some more
details on this in the case of sphere semantics in Section 5. All the other points, again, will reach either
the empty set or a set having diameter at least ε. The result we would expect in this second case is that
each point in the space reaches the flow tube including the limit cycle. We will see that this is true in the
case of sphere semantics, even when we use the simplest Taylor polynomials of degree one.
These considerations already allow us to point out that sphere semantics better reflects the real system
behaviour than the standard one.
5
Computing Sphere Semantics
In this section we show how sphere semantics can be computed exploiting tools for cylindrical algebraic
decomposition. In particular, we introduce a translation from sphere semantics to standard semantics.
Then, we apply the translation to study the “sphere” behaviour of our neural oscillator example.
5.1
A translation into standard semantics
If T is a first-order theory and δ is a distance definable in T , then the sphere semantics of any formula
in T is T -definable in the standardnsemantics,
o i.e., for any formula ϕ[X] ∈ T we can compute a formula
d
d
(ϕ)ε [X] ∈ T such that (|ϕ[X]|)ε = (ϕ)ε [X] for all ε ∈ R>0 .
In order to achieve this goal, we need to distinguish two kind of variables: the variables of the original
formula (named W , Wi , W and Wi ), whose evaluations follow the rules of the sphere semantics, and the
auxiliary variables (named Y , Yi , Y and Yi ) that will be introduced to encode the sphere semantics into
the standard one. From the point of view of the sphere semantics the later can seen as symbolic constants,
d . In particular, we will use them to characterize sets of
even if they will be quantified in the formula (ϕ)
ε
V
W
the form (| r∈R ϕJr, WK|)ε and (| r∈R ϕJr, WK|)ε in the standard semantics.
Definition 7 (Translation). Let T be a first-order theory over the reals, ϕ[Y, W] be any first-order
d [Y, W] by structural induction on ϕ[Y, W] itself.
formula T -definable, and ε ∈ R>0 . We define (ϕ)
ε
def
1. ((t1 ◦\
t2 )[Y, W])ε = ∃W0 ((t1 ◦ t2 )JY, W0 K ∧ δ (W0 , W) < ε), for ◦ ∈ {=, <};
def c
d ;
\
2. (φ
∨ ψ)ε = (φ
)ε ∨ (ψ)
ε
def
c) ∧ (ψ)
d )JY, W1 K) ∧ δ (W0 , W) < ε);
\
3. (φ [Y, W]
∧ ψ[Y, W])ε = ∃W0 (∀W1 (δ (W0 , W1 ) < ε _ ((φ
ε
ε
def
\
4. (∀W φ\
[Y,W, W])ε = ∃W0 (∀W1 (δ (W0 , W1 ) < ε _ ∀Y (φ JY,Y,
W1 K)ε ) ∧ δ (W0 , W) < ε);
def
\
5. (∃W φ\
[Y,W, W])ε = ∃Y (φ JY,Y,
W1 K)ε ;
def
\
6. (¬φ\
[Y, W])ε = ∃W0 (∀W1 (δ (W0 , W1 ) < ε _ ¬(φ JY,
W1 K)ε ) ∧ δ (W0 , W) < ε).
Theorem 4 (Semantics Equivalence). Let T be any first-order theory and δ be a T -definable distance.
The sphereosemantics (|.|)ε of T is T -definable in the standard semantics and, in particular, (|ϕ[X]|)ε =
n
d [X] for any formula ϕ[X] ∈ T and all ε ∈ R>0 .
(ϕ)
ε
69
A. Casagrande, T. Dreossi & C. Piazza
d [Y, W] is definable in Tarski theory, and this theory is decidable, the satiability of ϕ in (||) is
Since (ϕ)
ε
ε
decidable.
Let us notice that the application of the translation in Definition 7 to a formula, increases the evaluation complexity of such formula with respect to its untranslated version. This is mainly due to the
possible introduction of quantifier operator alternations.
5.2
Experimental Results on the Neural Oscillator
Let us consider the hybrid automaton H f˜ described in Section 2 for modeling a neural oscillator. We
intend to study its behaviour through sphere semantics, exploiting cylindrical algebraic decomposition
tools to automatically compute it.
As we noticed in Section 3, any differentiable dynamics can be exactly represented, in terms of an
opportune ε-semantics, by a semi-algebraic function. In particular, we can replace the dynamics of H f˜
by the corresponding Taylor polynomials up to a certain degree which depends on ε. In this paper, we
decided to model the automaton dynamics by using their first-degree Taylor polynomials and we obtained
the automaton H 0f˜ depicted in Figure 5. In order to keep the presentation simple, in this section we fix
the parameters as follows τ = 3, λ = 1, α = 2. Hence, the activations correspond to the axis Xi = ±2
and Xe = ±2.
Xi
v8
v7
Xe0 = Xe − ( X3e + 2)t
Xi0 = Xi − X3i ∗ t
v1
v6
Xe0 = Xe + ( X6e − 1)t
Xi0 = Xi + ( X2e − X3i + 1)t
v0
Xe0 = Xe − ( X3e + X2i + 1)t
Xi0 = Xi + ( X6i − 1)t
v2
Xe0 = Xe + ( X6e − X2i )t
Xi0 = Xi + ( X2e Xe + X6i )t
Xe0 = Xe − ( X3e + X2i − 1)t
Xi0 = Xi + ( X6i + 1)t
Xe
Xi = −2
v4
Xe0 = Xe + ( X6e + 1)t
Xi0 = Xi + ( X2e − X3i − 1)t
Xe = −2
Xi = 2
v5
v3
Xe0 = Xe − X3e t
Xi0 = Xi − ( X3i − 2)t
Xe0 = Xe − X3e t
Xi0 = Xi + (2 − X3i )t
Xe0 = Xe + (2 − X3e )t
Xi0 = Xi − X3i t
Xe = 2
Figure 5: The piecewise hybrid automaton H 0f˜ associated to the function f2 (3, 1).
A simulation of H 0f˜ is presented in Figure 4: a limit cycle is still present, but it has a diamondlike shape. We are interested in studying this limit cycle. In particular, we are interested in proving,
Hybrid Automata and ε-Analysis on a Neural Oscillator
70
exploiting tools for symbolic computation, that if we apply sphere semantics, each point in the space
reaches a bounded region which includes the limit cycle. In this example our automata have unbounded
invariants, hence the termination of sphere semantics reachability algorithm is not guaranteed.
We start computing the intersections of the limit cycle with the activation regions. Consider for
instance the intersection Q0 = hxQ0 , 2i of the limit cycle with Xi = 2 and Xe > 0. We have that xQ0 is
the unique solution of the equation which describes the intersection of the diamond-like limit cycle with
Xi = 2. Similarly, consider point Q1 = h2, yQ1 i that in turn corresponds to the intersection of the limit
cycle with Xe = 2 and Xi > 0. We effectively calculated all√ these intersections by √using the computer
√17+14538 and yQ = 190√ 17+786 . The points
algebra system Maxima. So, for instance, we get xQ0 = 3526
1
495 17+2041
39 17+161
Q0 and Q1 satisfy the activation formulæ wich regulate the discrete transitions between locations v6 and
v5 , and locations v6 and v7 , respectively. Let us now consider a point P0 located on Xi = 2, but which is
such that its distance d0 from Q0 is at least 2ε, i.e., P0 = hxP0 , 2i and δ (Q0 , P0 ) = d0 > 2ε. Consider now
any point P1 on Xe = 2 resulting from the sphere semantics evaluation of the continuous evolution which
starts in P0 inside location v6 . Thus, let denote with d1 the distance between such P1 and Q1 .
If we could prove that d1 is always smaller than d0 , then we would be able to conclude that all the
points which start from a distance of at least 2ε from the limit cycle converge to a flow tube having
diameter 2ε that includes the limit cycle. Of course, to obtain such conclusion, we need to prove this
property on all locations.
We can formalize this concept through a first-order formula. We denote with r and s the straight lines
Xi = 2 and Xe = 2, respectively, and with the notation Q0 ∈ r ∩C ∩ Xe > 0 the membership of Q0 to the
intersection of straight line r with limit cycle C and positive Xe semi-plane. Moreover, with the notation
(|P0 →C P1 |)ε we denote the continuous transition from point P0 to point P1 performed exploiting sphere
semantics. Thus, our desired property can be expressed as:
∀Q0 Q1 ∀P0 P1 (Q0 ∈ r ∩C ∩ Xe > 0 ∧ Q1 ∈ s ∩C ∩ Xi > 0 ∧ P0 ∈ r ∩ Xe > 0∧
P1 ∈ s ∩ Xe > 0 ∧ δ (Q0 , P0 ) > 2ε ∧ (|P0 →C P1 |)ε ) _ δ (Q1 , P1 ) < δ (Q0 , P0 )
(5)
stating the convergence to the limit flow tube in location v6 . Such property can be easily rewritten for
each location of the hybrid automaton, changing the roles of activation border lines r and s.
5
0
5
Trajectory from (0,0)
Trajectory from (-1,6)
5
0
5
Figure 6: Two evolutions of the first degree approximation of the model proposed in Section 2.
We automatically expanded such formula by using a Perl script that implements the Definition 7 to
translate sphere semantics into the standard one. In particular, (|P0 →C P1 |)ε , in the case of location v6
71
A. Casagrande, T. Dreossi & C. Piazza
becomes
def
ψ[X] = ∃T(∃X0 (∀X1 (δ (X0 , X1 ) < ε _ (T > 0 ∧ φ JX1 , TK))) ∧ δ (X0 , X) < ε),
(6)
where
def
φ [X, T] = ∃X0 (∀X1 (δ (X0 , X1 ) < ε _ (∃X2 (Π1 JX2 , TK ∧ δ (X2 , X1 ) < ε)∧
∃X2 (Π2 JX2 , TK ∧ δ (X2 , X1 ) < ε))) ∧ δ (X0 , X) < ε),
def
(7)
def
Π1 [X0 ,Y0 , X1 ,Y1 , T] = 6 ∗ X1 = 6 ∗ X0 + (X0 − 3 ∗Y0 ) ∗ T, and Π2 [X0 ,Y0 , X1 ,Y1 , T] = 6 ∗Y1 = 6 ∗Y0 + (Y0 +
3 ∗ X0 ) ∗ T. However, we notice that, since Π1 and Π2 are closed and convex, φ can be simplified as:
φ [X, T] ≡ ∃X0 (Π1 JX0 , TK ∧ Π2 JX0 , TK ∧ δ (X0 , X) < ε).
(8)
ψ[X] ≡ ∃T(T > 0 ∧ ∃X0 (Π1 JX0 , TK ∧ Π2 JX0 , TK ∧ δ (X0 , X) < ε)).
(9)
Similarly, ψ becomes:
So we plugged this last formula in Formula 5 and used REDLOG [5] to test it. The formula turns out to be
true (the result is computed within few seconds), proving our conjectures.
Notice that we used Maxima to compute the exact coordinates of the points on the limit cycles since
that computation does not require quantifier elimination. However, we could have used REDLOG.
As far as h0, 0i is concerned it is immediate to prove through a first-order formula that it reaches
points different from itself and, hence, it reaches the limit flow tube.
Other interesting properties that automatically verified express, for instance, the fact that applying
the sphere semantics there are points that cross the limit cycle (in both directions). This is quite natural
since points closer than ε to the limit cycle get expanded and cross it.
6
Conclusions
In this paper we have modeled a neural oscillator constructing a hybrid automaton whose components
derive from the approximation of the continuous model presented in [19]. We have analyzed its behaviours considering the application of some approximation techniques for the introduction of noise, as
already advocated in [19]. In particular, we focused on the approach based on the ε-semantics.
The simulation based on the application of the ε-semantics has revealed the any point which begins
its evolution from a distance of at least 2ε from the limit cycle, converges to a flow tube which possesses
a diameter equal to 2ε and that includes the limit cycle. Due to size of the formulæ which compose
the hybrid automaton and the growth of such formalæ introduced by the translation of the ε-semantics
evaluations, a direct computation of the reachable set would have high complexity and eventually returns
results of difficult interpretation. For this reason, we have reformulated the problem in form of a closed
property which guarantees the convergence of any point towards the limit cycle of the modeled system.
During the construction of the formula that describes the convergence to the limit cycle, some steps
of simplification of the formulæ have been applied. In particular, we have reduced the complexities of
translated formulæ, relying on the convexity of the sets characterized by some of their subformulæ. An
interesting aspect to investigate is whether these simplification steps can be automatically performed.
As future work, in order to analyze the behaviour of a group of neural oscillators, we plan to combine
several hybrid automata and to study their evolutions always adopting ε-semantics.
72
Hybrid Automata and ε-Analysis on a Neural Oscillator
References
[1] R. Alur, C. Courcoubetis, N. Halbwachs, T. A. Henzinger, P.-H. Ho, X. Nicollin, A. Olivero, J. Sifakis &
S. Yovine (1995): The Algorithmic Analysis of Hybrid Systems. Theoret. Comput. Sci. 138(1), pp. 3–34,
doi:10.1016/0304-3975(94)00202-T.
[2] A. F. Atiya & P. Baldi (1989): Oscillations and Synchronizations in Neural Networks: an Exploration of the
Labeling Hypothesis. Int. J. Neural Syst. 1(2), pp. 103–124, doi:10.1142/S0129065789000025.
[3] A. Casagrande, C. Piazza & A. Policriti (2009): Discrete Semantics for Hybrid Automata. Discrete Event
Dyn. Syst. 19(4), pp. 471–493, doi:10.1007/s10626-009-0082-7.
[4] A. Casagrande, C. Piazza, A. Policriti & B. Mishra (2008): Inclusion dynamics hybrid automata. Inform.
and Comput. 206(12), pp. 1394–1424, doi:10.1016/j.ic.2008.09.001.
[5] A. Dolzmann & T. Sturm (1997): REDLOG: computer algebra meets computer logic. SIGSAM Bull. 31(2),
pp. 2–9, doi:10.1145/261320.261324.
[6] H. B. Enderton (2001): A Mathematical Introduction to Logic, II edition. Harcourt/Academic Press.
[7] M. Fränzle (1999): Analysis of Hybrid Systems: An Ounce of Realism Can Save an Infinity of States. In: Int.
Workshop on Computer, Science, and Logic (CSL 99), LNCS 1683, Springer, pp. 126–140.
[8] W. J. Freeman & C. A. Skarda (1985): Spatial EEG patterns, non-linear dynamics and perception: the
neo-Sherringtonian view. Brain Res 357(3), pp. 147–175.
[9] R. Ghosh & C. Tomlin (2001): Lateral Inhibition through Delta-Notch signaling: A Piecewise Affine Hybrid
Model. In: Proc. of Hybrid Systems: Computation and Control (HSCC’01), LNCS 2034, Springer, pp.
232–246, doi:10.1007/3-540-45351-2 21.
[10] A. Girard & G. J. Pappas (2007): Approximation metrics for discrete and continuous systems. IEEE Trans.
Automat. Control 52(5), pp. 782–798, doi:10.1109/TAC.2007.895849.
[11] C. M. Gray, P. Konig, A. K. Engel & W. Singer (1989): Oscillatory responses in cat visual cortex exhibit
inter-columnar synchronization which reflects global stimulus properties. Nature 338(6213), pp. 334–337,
doi:10.1038/338334a0.
[12] R. Grosu, G. Batt, F. Fenton, J. Glimm, C. Guernic, S. Smolka & E. Bartocci (2011): From Cardiac Cells to
Genetic Regulatory Networks. In: Proc. of Computer Aided Verification (CAV’11), LNCS 6806, Springer,
pp. 396–411, doi:10.1007/978-3-642-22110-1 31.
[13] T. A. Henzinger, P. W. Kopke, A. Puri & P. Varaiya (1995): What’s decidable about hybrid automata? In:
Proc. of ACM Symposium on Theory of Computing (STOCS’95), pp. 373–382, doi:10.1145/225058.225162.
[14] H. de Jong, J.-L. Gouzé, C. Hernandez, M. Page, T. Sari & J. Geiselmann (2004): Qualitative simulation
of genetic regulatory networks using piecewise-linear models. Bulletin of Math. Biology 66, pp. 301–340,
doi:10.1016/j.bulm.2003.08.010.
[15] P. Prabhakar, V. Vladimerou, M. Viswanathan & G. E. Dullerud (2009): Verifying Tolerant Systems Using
Polynomial Approximations. In: Proceedings of the 2009 30th IEEE Real-Time Systems Symposium, RTSS
’09, IEEE Computer Society, Washington, DC, USA, pp. 181–190, doi:10.1109/RTSS.2009.28.
[16] S. Ratschan (2010): Safety Verification of Non-linear Hybrid Systems Is Quasi-Semidecidable. In: Theory
and Applications of Models of Computation (TAMC10), lncs 6108, springer, pp. 397–408, doi:10.1007/9783-642-13562-0 36.
[17] L. R. Silva, Y. Amitai & B. W. Connors (1991): Intrinsic oscillations of neocortex generated by layer 5
pyramidal neurons. Science 251(4992), pp. 432–5, doi:10.1126/science.1824881.
[18] M. Steriade & R. R. Llinás (1988): The functional states of the thalamus and the associated neuronal interplay. Physiological reviews 68(3), pp. 649–742.
[19] A. Tonnelier, S. Meignen, H. Bosch & J. Demongeot (1999): Synchronization and desynchronization of
neural oscillators. Neural Networks 12(9), pp. 1213 – 1228, doi:10.1016/S0893-6080(99)00068-4.
| 5cs.CE
|
arXiv:1612.05994v1 [math.ST] 18 Dec 2016
ALGEBRAIC PROBLEMS IN STRUCTURAL EQUATION
MODELING
MATHIAS DRTON
Abstract. The paper gives an overview of recent advances in structural equation modeling. A structural equation model is a multivariate statistical model
that is determined by a mixed graph, also known as a path diagram. Our
focus is on the covariance matrices of linear structural equation models. In
the linear case, each covariance is a rational function of parameters that are
associated to the edges and nodes of the graph. We statistically motivate
algebraic problems concerning the rational map that parametrizes the covariance matrix. We review combinatorial tools such as the trek rule, projection
to ancestral sets, and a graph decomposition due to Jin Tian. Building on
these tools, we discuss advances in parameter identification, i.e., the study of
(generic) injectivity of the parametrization, and explain recent results on determinantal relations among the covariances. The paper is based on lectures
given at the 8th Mathematical Society of Japan Seasonal Institute.
Contents
Part
1.
2.
3.
I. Structural Equation Models and Questions of Interest
Motivation
Linear Structural Equation Models
Questions of Interest
Part
4.
5.
6.
II. Treks, Subgraphs and Decomposition
Trek Rule
Induced Subgraphs and Principal Submatrices
Graph Decomposition
2
2
4
7
13
13
15
17
Part III. Parameter Identification
7. Global Identifiability
8. Generic Identifiability
21
21
23
Part IV. Relations Among Covariances
9. Implicitization
10. Conditional Independence
11. Trek Separation
12. Verma Constraints
29
29
31
32
34
13. Conclusion
References
35
36
Date: September 21, 2016.
Key words and phrases. Algebraic statistics, covariance matrix, Gaussian distribution, graphical model, Gröbner basis, structural equation model.
1
2
MATHIAS DRTON
Part I. Structural Equation Models and Questions of Interest
1. Motivation
The following example serves well to introduce the statistical models we will
consider. It features the simplest instance of what is known as an instrumental
variable model. An empirical study that shows this type of model ‘in action’ can
be found in [29].
Example 1.1. Does a mother’s smoking during pregnancy harm the baby? To
answer this question researchers conduct a study in which they record, for a sample
of pregnancies, the baby’s birth weight and the average number of cigarettes the
mom smoked per day during the first trimester. The researchers observe a significant negative correlation between the birth weight and smoking and are tempted to
conclude that smoking has a negative effect on the baby’s health, with an increase
in the number of cigarettes smoked leading to lower birth weight.
The cigarette companies are not suprised by this finding. They argue, however,
that smoking does not harm baby. Instead, heavier smoking is merely a reflection of
underlying factors that are the true causes of low birth weight. Such confounding
variables could, for instance, be of socio-economic nature. In the context of the
smoking-lung cancer debate, prominent Statistician Ronald Fisher liked to argue
that correlations can be attributed to unobserved variables of genetic nature [59].
Familiar with this type of counter-argument, the researchers cleverly recorded a
third variable: The tax rate on tobacco products in the local jurisdictions of the
mothers in the sample. It is not unreasonable to assume that the tax rate does
not have a direct effect on the baby’s health. If there is then variation in the tax
rate and higher taxes have an effect on the amount of smoking, then the effect
that smoking has on birth weight can be estimated in a model that allows for the
presence of unobserved confounders, as we will see shortly.
The above narrative suggests a number of cause-effect relations, as well as the
absence thereof. Qualitatively these are summarized in the graph in Figure 1.1.
The variables in play are the nodes of the graph and cause-effect relationships are
indicated as directed edges. The variable U represents a confounding variable and
is unobserved, which we emphasize by coloring its edges in red.
Structural equation models turn the qualitative descriptions of causes and effects
into quantified functional relationships. In this article, the functional relationships
will always be linear. The linear structural equation model for the present example
is based on the following system of structural equations:
(1.1)
X1 = λ01
(1.2)
X2 = λ02 + λ12 X1 + λu2 U + ε2 ,
(1.3)
X3 = λ03 + λ23 X2 + λu3 U + ε3 ,
(1.4)
U = λ0u
+ ε1 ,
+ εu .
Here, the error terms ε1 , ε2 , ε3 , εu are independent random variables with zero
mean. The eight coefficients λ01 , λ02 , λ03 , λ0u , λ12 , λ23 , λu2 , and λu3 are unknown
parameters. Equation (1.1) indicates that variable X1 , the tax rate, has expectation
λ01 , from which it deviates according to the distribution assumed for ε1 . The
analogous statement for the unobserved confounder U is made in (1.4). In (1.2),
the amount of smoking, denoted X2 , is modeled to be a linear function of the tax
STRUCTURAL EQUATION MODELS
3
U : Confounder
X1 : Tax Rate
X2 : Mom’s Smoking
X3 : Baby’s Weight
Figure 1.1. Directed graph representing an instrumental variable model.
rate and independent noise. Similarly, (1.3) introduces birth weight, denoted X3 ,
as a noisy linear function of smoking.
The quantity of primary interest is the coefficient λ23 that quantifies the relationship between smoking and birth weight. Using data, we can estimate the joint
distribution and, in particular, the covariance matrix of the three observed variables
X1 , X2 and X3 . Because the error terms are independent and have zero means,
the covariance between X3 and X1 is
(1.5)
Cov[X1 , X3 ] = λ23 Cov[X1 , X2 ].
Hence, as long as Cov[X1 , X2 ] 6= 0, statistical inference about λ23 may be based on
the ratio of the two covariances in (1.5).
In some applications of structural equation models latent (that is, unobserved)
variables are of direct interest. For instance, concepts such as intelligence or depression in psychology are of this nature and measured only indirectly through
other variables such as exam results or answers in questionnaires. While problems
with explicit latent variables are ubiquitous [5], we will focus on models in which
the effects of latent variables are summarized and represented merely in terms of
correlations among the error terms in structural equations. This representation of
dependence induced by latent variables is discussed in detail in [33, 44, 49, 50, 71].
Example 1.2. We take up the instrumental variable model from Example 1.1.
The effects of the confounding variable U can be summarized by absorbing U into
the error terms for equations (1.2) and (1.3). Define
(1.6)
ε̃2 = λu2 U + ε2 ,
ε̃3 = λu3 U + ε3 .
Retaining only the equations for the observed variables X1 , X2 , and X3 , we are left
with the equation system:
(1.7)
X1 = λ01
+ ε1 ,
(1.8)
X2 = λ02 + λ12 X1 + ε̃2 ,
(1.9)
X3 = λ03 + λ23 X2 + ε̃3 .
However, and this is the significance of the unobserved variable U , we have now
correlated error terms because
(1.10) ω23 := Cov[ε̃2 , ε̃3 ] = Cov[λu2 U + ε2 , λu3 U + ε3 ] = λu2 λu3 Var[U ] 6= 0.
In the sequel, we will focus on models that are given by equations such as (1.7)(1.9), with one equation for each observed variable but error terms that may be
correlated. Graphically, such models may be represented by a mixed graph that
features directed edges to encode which variables appear in each structural equation and bidirected edges that indicate possibly nonzero correlations between error
terms. The mixed graph for the model given by (1.7)-(1.9) is depicted in Figure 1.2,
which shows the unknown parameters as weights for the edges. At the nodes, we
4
MATHIAS DRTON
ω23
ω11
X1 : Tax rate
λ12
ω22
X2 : Mom’s smoking
λ23
ω33
X3 : Baby’s weight
Figure 1.2. Mixed graph representing an instrumental variable model.
show the variances of the error terms, namely, ω11 = Var[ε1 ], ω22 = Var[ε̃2 ], and
ω33 = Var[ε̃3 ]. In the statistical literature, the mixed graph for a structural equation
model is also known as a path diagram.
The ratio Cov[X2 , X3 ]/ Var[X2 ] is the regression coefficient when predicting X3
from X2 . We have
(1.11)
Cov[X2 , X3 ]
ω23
= λ23 +
.
Var[X2 ]
Var[X2 ]
Hence, linear regression predicting X3 from X2 only estimates the coefficient of
interest if ω23 = 0, as is the case when X2 and X3 do not depend on the latent
variable U . When ω23 6= 0, the relation from (1.5), which involves all three variables,
is needed to recover λ23 .
The remainder of this paper is organized as follows. Section 2 introduces linear
structural equation models in full generality. We then formulate questions of statistical interest and the algebraic problems they correspond to (Section 3). Next,
we examine the interplay between covariance matrices and mixed graphs. We treat
the so-called trek rule (Section 4) and review useful results on subgraphs and graph
decomposition (Sections 5 and 6). In Sections 7 and 8, we dive deeper into parameter identifiability, which here means the question of whether a coefficient of interest
can be recovered from the covariance matrix of the observed variables. Finally, we
discuss relations among the entries of the covariance matrix (Sections 9-12).
2. Linear Structural Equation Models
Let ε = (εi : i ∈ V ) be a random vector indexed by a finite set V . Define a
new random vector X = (Xi : i ∈ V ) as the solution to the system of structural
equations
(2.1)
X = ΛT X + ε,
where Λ = (λij ) ∈ RV ×V is a matrix of unknown parameters. Suppose ε has
covariance matrix Ω = (ωij ) = Var[ε], so Ω is a positive definite matrix whose
entries are again unknown parameters. Write I for the identity matrix. If I − Λ is
invertible, then the linear system in (2.1) is solved uniquely by X = (I − Λ)−T ε,
which has covariance matrix
(2.2)
Var[X] = (I − Λ)−T Ω(I − Λ)−1 =: φ(Λ, Ω).
Interesting settings are obtained by restricting the support of Λ and Ω, as is the
case in our motivating example.
STRUCTURAL EQUATION MODELS
5
Example 2.1. Consider the setup from Example 1.2. If the equation system
from (1.7)-(1.9) is written in vector form as in (2.1), then the coefficient matrix is
0 λ12
0
Λ = 0 0 λ23 .
(2.3)
0 0
0
The error covariance matrix is
(2.4)
ω11
Ω = Var[ε] = 0
0
0
ω22
ω23
0
ω23 .
ω33
From (2.2), the covariance matrix of X = (X1 , X2 , X3 ) is found to be
ω11
ω11 λ12
ω11 λ12 λ23
,
ω22 + ω11 λ212
ω23 + λ23 σ22
(2.5)
Var[X] = ω11 λ12
ω11 λ12 λ23 ω23 + λ23 σ22 ω33 + 2ω23 λ23 + λ223 σ22
where σ22 denotes the (2, 2) entry of Var[X]. The relation from (1.5) can be confirmed from the (1, 2) and (2, 3) entry of Var[X].
Restrictions on the support of a matrix naturally correspond to a graph. Specifically, we adopt mixed graphs because we are dealing with two matrices, Λ and Ω,
whose rows and columns are indexed by the same set V . In structural equation
modeling, this point of view originated in the work of Sewall Wright [72, 73].
A mixed graph with vertex set V is a triple G = (V, D, B) where D, B ⊆ V × V
are two sets of edges. The set D comprises ordered pairs (i, j) that we also denote
by i → j to visualize the fact that such a pair encodes a directed edge pointing
from i to j. Then i is the tail and j is the head of the egde. The pairs in B are
unordered pairs {i, j} that encode bidirected edges that we also denote by i ↔ j.
These edges have no orientiation, and i ↔ j ∈ B if and only if j ↔ i ∈ B. It
will be convenient to call both endpoints i and j heads of i ↔ j. In our context,
neither the bidirected part (V, B) nor the directed part (V, D) contain loops, that
is, i → i 6∈ D and i ↔ i 6∈ B for all i ∈ V . If (V, D) does not contain any directed
cycles i → . . . → i, then the mixed graph G is said to be acyclic.
Let RD be the set of real V × V -matrices Λ = (λij ) with support in D, that is,
(2.6)
RD = Λ ∈ RV ×V : λij = 0 if i → j ∈
/D .
D
Define RD
reg to be the subset of matrices Λ ∈ R for which I − Λ is invertible. If G
is acyclic, then there is a permutation of V that makes I − Λ unit upper triangular
such that det(I − Λ) = 1 for all Λ ∈ RD and thus RD = RD
reg . Similarly, let PD V
be the cone of positive definite symmetric V × V -matrices Ω = (ωij ), and define
PD(B) to be the subcone of matrices supported over B, that is,
(2.7)
PD (B) = Ω ∈ PD V : ωij = 0 if i 6= j and i ↔ j ∈
/B .
Taking the error vector ε to be Gaussian (or in other words, to follow a multivariate normal distribution), we arrive at the following definition of a statistical model
for the random vector X that solves (2.1). Readers looking for background such
as the fact that linear transformations of a Gaussian random vector are Gaussian
may consult a textbook on multivariate statistics, e.g., [2].
6
MATHIAS DRTON
λ13
1
λ12
3
λ23
2
λ34
ω24
4
Figure 2.1. An acyclic mixed graph known as the Verma graph.
Definition 2.1. The linear structural equation model given by a mixed graph
G = (V, D, B) is the family of all multivariate normal distributions on RV with
covariance matrix in the set
MG = (I − Λ)−T Ω(I − Λ)−1 : Λ ∈ RD
reg , Ω ∈ PD (B) .
The covariance parametrization of the model is the map
φG : RD × PD(B) → PD V ,
(Λ, Ω) 7→ (I − Λ)−T Ω(I − Λ)−1 ,
for which we define the fiber of a pair (Λ, Ω) ∈ RD
reg × PD (B) to be the preimage
′ ′
′
′
(2.8)
FG (Λ, Ω) = (Λ , Ω ) ∈ RD
reg × PD (B) : φG (Λ , Ω ) = φG (Λ, Ω) .
As defined, a linear structural equation model does not impose any restrictions
on the mean vector of the normal distributions. Consequently, the mean vector
plays no role in our discussion. For instance, in maximum likelihood estimation we
may assume without loss of generality that the mean vector is zero. Other questions
we consider will directly concern the covariance matrices of the model. Therefore,
we may safely identify a linear structural equation model with its set of covariance
matrices MG . On occasion, we will simply refer to MG as the model.
Leaving statistics out of the picture, our interest is in the maps φG , their fibers
FG and their images MG . Algebra comes into play naturally.
Proposition 2.1. For any mixed graph G, the map φG is a rational map whose
image MG and fibers FG (Λ, Ω) are semi-algebraic sets. The map φG is a polynomial
map if and only if G is acyclic.
Proof. That φG is rational follows from Cramer’s rule for matrix inversion. The domain of φG is a semi-algebraic set and, thus, the fibers FG (Λ, Ω) are semi-algebraic
as well. The Tarski-Seidenberg theorem implies that MG is semi-algebraic. If
G = (V, D, B) is acyclic, then det(I − Λ) = 1 for all Λ ∈ RD . Consequently, the
entries of (I − Λ)−1 are polynomial in Λ. If G is not acyclic, then det(I − Λ) is a
non-constant polynomial. The Leibniz formula shows that its terms correspond to
collections of disjoint directed cycles in the graph; compare Theorem 1 in [41].
Example 2.2. The mixed graph G = (V, D, B) in Figure 2.1 encodes the structural
equations
X1 = λ01
+ ε1 ,
X2 = λ02 + λ12 X1
+ ε2
X3 = λ03 + λ13 X1 + λ23 X2 + ε3
X4 = λ04 + λ34 X3
+ ε4 .
STRUCTURAL EQUATION MODELS
7
Only the errors ε2 and ε4 may be dependent and the error covariance matrix is
ω11
0
0
0
0 ω22
0 ω24
.
Ω =
0
0 ω33
0
0 ω24
0 ω44
Substracting the coefficient matrix from the identity gives
1 −λ12 −λ13
0
0
1
−λ23
0
I −Λ =
0
0
1
−λ34
0
0
0
1
with det(I − Λ) = 1 and inverse
1 λ12
0 1
−1
(I − Λ) =
0 0
0 0
λ13 + λ12 λ23
λ23
1
0
λ13 λ34 + λ12 λ23 λ34
λ23 λ34
.
λ34
1
To illustrate the form of the map φG , we display the coordinate function
(2.9)
φG (Λ, Ω)24 = λ12 λ13 λ34 ω11 + λ212 λ23 λ34 ω11 + λ23 λ34 ω22 + ω24 .
3. Questions of Interest
Structural equation models are used to empirically estimate, test and possibly
discover cause-effect relationships among a set of variables. In estimation and testing, the underlying graph is given. In discovery, we seek to estimate the underlying
graph, or in other words, perform model selection. In this section we give a broad
overview of algebraic problems that arise in the context of these statistical tasks.
Only some of the problems are treated in the remainder of the paper, in which we
focus on parameter identifiability and polynomial relations between covariances.
3.1. Parameter identification. When considering the model given by the mixed
graph G = (V, D, B), a first question is whether the effects of interest are identifiable, that is, whether they are determined by the joint distribution of the observed
variables. The importance of the question is clear: The joint distribution is what
can be estimated from data. In our setting of linear and Gaussian models, the problem is equivalent to deciding whether the coefficients λij in the linear structural
equations can be recovered from the covariance matrix of the variables.
Different notions of parameter identifiability translate into related but slightly
different algebraic problems. The most stringent identifiability property of a model
is to have all of its coefficients λij , i → j ∈ D, identifiable. In this case, we seek to
answer the following:
Question 3.1. Is the map φG is injective?
Injectivity of φG can be decided efficiently, as we will discuss in Section 7. However, injectivity can be too strong of a requirement because all fibers are required
to be singletons with FG (Λ, Ω) = {(Λ, Ω)}. Indeed, some interesting examples have
fibers that are not singletons.
8
MATHIAS DRTON
Example 3.1. The map φG fails to be injective when G is the graph for the
instrumental variable model from Example 1.2. The relation from (1.5) shows that
FG (Λ, Ω) = {(Λ, Ω)} if λ12 6= 0. If λ12 = 0, however, then the fiber is infinite.
Hence, all model parameters are identifiable as long as λ12 6= 0. In the context of
Example 1.2, this requires making an argument that higher tax rates impact the
amount of smoking.
In the example just given, FG (Λ, Ω) = {(Λ, Ω)} for generic choices of (Λ, Ω) ∈
RD
reg × PD(B). In this case, we call φG generically injective. We are led to:
Question 3.2. Is the map φG is generically injective?
It turns out that generic injectivity is more difficult to decide. The computational
complexity of the problem has not yet been determined. In Section 8, we review
methods to decide whether φG is generically injective as well as methods to decide
when the fibers are generically infinite.
When φG is not generically injective, its generic fibers may be discrete sets. This
property is known as local identifiability in the statistical literature. We will instead
speak of φG being generically finite-to-one to highlight that in our case a discrete
fiber is in fact finite because φG is rational. By the inverse function theorem, the
question of whether φG is generically finite-to-one is the same as:
Question 3.3. Does the Jacobian of φG have full column rank?
The fiber FG (Λ, Ω) is defined by the equation system φG (Λ′ , Ω′ ) = φG (Λ, Ω).
These equation systems have a generic number of complex solutions (i.e., the free
entries of Λ and Ω are allowed to be complex numbers).
Definition 3.1. The map φG is algebraically k-to-one if the equation systems
defining its fibers generically have k complex solutions. We call the number k the
algebraic degree of identifiability of G.
The degree of identifiability may be determined by Gröbner basis methods (see
Section 8). It is finite if and only if φG is generically finite-to-one. The main
theorem in Section 7.3 shows that if φG is injective then its inverse is rational,
which is the same as G having degree of identifiability one. Currently, there are no
combinatorial results about when the degree is finite but larger than one. Example
8(b) in [34] has degree 3 but fibers whose cardinality over the reals is either one or
three. To the author’s knowledge, no example has been discovered in which φG is
generically injective over the reals but algebraically k-to-one for k ≥ 2.
An important question that we will not address in detail is the identifiability
of only a single parameter λij for a designated edge of interest i → j ∈ D. This
amounts to checking whether in every fiber the coefficient for the edge has only a
single value. In other words, it must hold that λ′ij = λ′′ij whenever (Λ′ , Ω′ ) and
(Λ′′ , Ω′′ ) are in the same fiber. In Example 1.2, the fiber of a pair (Λ, Ω) with
λ12 = 0 is infinite but all pairs (Λ′ , Ω′ ) such a fiber have λ′12 = 0. Two well-known
graphical methods for identifying a single edge coefficient are the back-door and
the front-door criterion [49]; see also [11, 36].
3.2. Model dimension. Statistical tests may be used to assess whether a model
is compatible with empirical data. At an intuitive level, such tests are based on
computing a distance between data and model and comparing this distance to
typical distances that are obtained when data are generated from a distribution in
STRUCTURAL EQUATION MODELS
9
the model. For linear Gaussian models, a test can be thought off as assessing the
distance between the empirical (or sample) covariance matrix and the model MG .
Recall that we have defined MG as the set of covariance matrices.
The challenging part of designing a statistical test is to quantify, in a probabilistic
manner, what typical distances between data and model are. Many procedures rely
on asymptotic approximations that are obtained by letting the number of data
points grow to infinity. Under regularity conditions, limiting distribution theory
leads to consideration of so-called chi-square distributions, which are indexed by
an integer parameter. In our context, when testing the model given by the mixed
graph G = (V, D, B), the chi-square parameter is set equal to the codimension of
MG , where we think of MG as embedded in the space of symmetric matrices. This
gives concrete statistical motivation for:
Question 3.4. What is the dimension of MG ?
The model MG is parametrized by the coefficients and covariances associated
with the edges in D and B as well as the variances associated with the nodes in
V . Being a subset of the space of V × V symmetric matrices, MG has expected
dimension
|V |(|V | + 1)
.
min |V | + |D| + |B|,
2
The term |D| + |V | + |B| is the count of nodes and edges in G. Since MG is the
image of φG , its actual dimension is equal to the rank of the Jacobian of φG . A
review of the connection between dimension and Jacobian in a statistical context
is given in [38]. Question 3.4 is tied to parameter identifiability, most directly
to Question 3.3. If φG is generically finite-to-one, then MG has the expected
dimension |V | + |D| + |B|.
3.3. Covariance equivalence. Different graphs may induce the same statistical
model. For example, take V = {1, 2}, and let G1 be the graph with the single
edge 1 → 2. Let G2 and G3 be the graphs with single edge 1 ← 2 and 1 ↔ 2,
respectively. Then MG1 = MG2 = MG3 as each model is easily seen to be equal
to the entire cone of positive definite 2 × 2 matrices.
From an applied perspective, two different graphs G and G′ encode different
scientific/causal hypotheses. If MG = MG′ , then the two hypotheses cannot be
distinguished based on data from a linear and Gaussian structural equation model.
It is thus useful to be able to decide whether two graphs G and G′ are covariance
equivalent, that is, we would like to be able to answer:
Question 3.5. When do two maps φG and φG′ have the same image?
Existing results addressing this questions make comparisons between relations
among the entries of the matrices in each model, and we record:
Question 3.6. What are the algebraic relations among the coordinates of φG ?
Such relations are also of interest for statistical tests that assess whether the
model given by G is compatible with available data; see, e.g., [6, 11, 17]. We review
results on relations among the covariances in Sections 9-12. An important role is
played by determinants that represent probabilistic conditional independence in a
Gaussian random vector. We note that models can, in principle, also be distinguished using inequality constraints. However, as less is know about inequalities,
10
MATHIAS DRTON
we do not treat them here. Examples of models with latent variables for which a
full semi-algebraic description is available can be found in [25, 53].
Remark. As defined above, covariance equivalence is based on data that is observational, meaning that it is collected by merely observing the considered physical
system. The situation is different when experimental data is available, meaning,
that data is collected in different settings in which the system is subject to various
experimental interventions. We will not treat such interventional data in this paper.
Interested readers may find discussions of the problem in [43, 49, 57]. Similarly,
even for observational data, questions of equivalence differ from Question 3.5 in
non-linear models or linear models with Gaussian errors [26, 54].
3.4. Maximum likelihood. The parameters of linear structural equation models
are most commonly estimated using the technique of maximum likelihood. Suppose
we observe a sample X (1) , . . . , X (n) drawn independently from the multivariate
normal distribution with mean vector µ and covariance matrix Σ, which we denote
by N (µ, Σ). The joint distribution of the random vectors X (1) , . . . , X (n) is the
n-fold product of N (µ, Σ). The likelihood of the sample is the value of the joint
density of the product distribution at (X (1) , . . . , X (n) ). The likelihood function is
the function mapping the pair (µ, Σ) to the likelihood of the sample. The maximum
likelihood estimator (MLE) of (µ, Σ) under the model given by a mixed graph G is
the maximizer of the likelihood function when restricting Σ to be in MG .
Define the sample mean vector and the sample covariance matrix as
n
1
1 X (i)
X
and Sn = (X (i) − X̄n )(X (i) − X̄n )T ,
(3.1)
X̄n =
n i=1
n
respectively. It is convenient to treat the likelihood function on the log-scale. With
an additive constant omitted and n/2 divided out, the log-likelihood function is
(µ, Σ) 7→ − log det(Σ) − trace Σ−1 Sn − (X̄n − µ)T Σ−1 (X̄n − µ).
Because the considered models place no constraint on the mean vector, its MLE is
always X̄n . The MLE of Σ solves the problem of maximizing the function
(3.2)
ℓ(Σ) = − log det(Σ) − trace Σ−1 Sn
subject to Σ ∈ MG . Using the covariance parametrization, the MLE is found
by maximizing ℓ ◦ φG . A key problem is then understanding the existence and
uniqueness of the MLE. We record:
Question 3.7. For which sample covariance matrices Sn does the likelihood function ℓ ◦ φG achieve its maximum?
Graphical models theory solves Question 3.7 when G = (V, D, B) is an acyclic
digraph, i.e., has B = ∅ and D without directed cycles [46]. More generally, it
is well known that ℓ ◦ φG is bounded when Sn is positive definite but this is not
necessary [32]. An issue that is not well explored is the fact that even if ℓ ◦ φG is
bounded it may fail to achieve its maximum as the model MG need not be closed.
For instance, the model in Example 3.3 is not closed. We remark that Question 3.7
is closely related to a positive definite matrix completion problem that arises in ML
estimation for other types of graphical models [9, 63, 67].
In some models, the MLE is known to admit a closed-form expression as a
rational function of the data. Such models have maximum likelihood (ML) degree
equal to one, in the sense of the following:
STRUCTURAL EQUATION MODELS
1
λ12
2
3
λ43
11
4
Figure 3.1. Mixed graph for a bivariate seemingly unrelated regressions model.
Question 3.8. The MLE of Σ in model MG is an algebraic function of the data.
What is the degree of this function?
An introduction to the notion of ML degree is given in [21, Chapter 2]. Here, we
merely not that the ML degree is one when G is an acyclic digraph. More general
models may have higher ML degree and a log-likelihood function with more than
one local maximum. We exemplify this for a model discussed in detail in [19].
Example 3.2. Suppose we have data with sample covariance matrix
X1
X1
Sn =
X2
X3
X4
X2
X3
8 −5
−5 27
10
4
3
49
10
4
21
24
X4
3
49
.
24
114
The matrix is positive definite such that the log-likelihood function ℓ from (3.2) is
bounded above on the entire cone of positive definite matrices. More precisely, ℓ
has compact level sets, that is, for any constant c ∈ R the set of positive definite
matrices Σ with ℓ(Σ) ≥ c is compact [2].
Let G be the graph depicted in Figure 3.1. It is not difficult to show that the
parametrization φG admits a rational inverse. Let Σ = (σij ) satisfy Σ = φG (Λ, Ω)
with Λ = (λij ) ∈ RD and Ω = (ωij ) ∈ PD (B). Then
σ12
σ34
λ12 =
(3.3)
,
λ43 =
,
σ11
σ44
and the entries of Ω = (I − Λ)T Σ(I − Λ) are rational functions of Σ as well. All
the rational functions are defined on the entire cone of positive definite matrices
because σ11 , σ44 > 0. It is also clear that the considered map φG is proper, that is,
compact subsets of the positive definite cone have compact preimages under φG . It
follows that ℓ ◦ φG has compact level sets and, thus, achieves its maximum on the
open set RD × PD(B).
The critical points of ℓ ◦ φG satisfy a rational equation system, in which the
determinant of φG (Λ, Ω) appears in the denominator. Since the directed part of G
is acyclic the determinant is equal to the determinant of Ω. Clearing the denominator yields a polynomial equation system. Saturating the system with respect to
the determinant removes infeasible solutions with det(Ω) = 0. Computing a lexicographic Gröbner basis after the saturation shows that the critical points (Λ, Ω)
solve the equation
10583160 λ512 + 43115307 λ412 + 72738452 λ312
+ 55482894 λ212 + 8437660 λ12 − 4703765 = 0.
All other entries of Λ and also Ω solve linear equations whose coefficients depend
on λ12 and the data. We conclude that the MLE of Σ is an algebraic function of
12
MATHIAS DRTON
1
3
4
2
Figure 3.2. Mixed graph of a model with two instruments.
degree 5. We note that the displayed equation for λ12 has three real roots and,
thus, is not solvable by radicals.
3.5. Model singularities. As noted in Section 3.2, the distributions of test statistics are frequently approximated using asymptotic theory. For so-called likelihood
ratio tests, this asymptotic theory can be thought of as assessing infinitesimal distances between a positive semidefinite data matrix and the given model MG . The
data matrix is generated from a distribution that corresponds to a particular point
in MG . At a smooth point of MG , the distribution of the infinitesimal distance is
given by a chi-square distribution which is the distribution of the distance between
a Gaussian random vector and a linear space. At singular points, the distribution
is determined by the tangent cone at the considered point [15]. Singularities also
impact other approaches such as Wald tests [24], and it is important to clarify:
Question 3.9. Is the image of φG a smooth manifold? If not, what are the tangent
cones of the image?
We do not address the question explicitly in this paper. However, whenever φG
is injective (see Section 7) its image is smooth. Indeed, when φG is injective it has a
rational inverse whose domain of definition includes the cone of all positive definite
matrices [16]. The next example illustrates that not all models are smooth.
Example 3.3. Consider the mixed graph G = (V, D, B) from Figure 3.2. Let
Σ ∈ R4×4 be a positive definite matrix. Define the matrix
−1
Σ{3,4}.{1,2} = Σ{3,4},{1,2} Σ{1,2},{1,2}
,
and the Schur complement
Σ{3,4},{3,4}.{1,2} = Σ{3,4},{3,4} − Σ{3,4},{1,2} Σ{1,2},{1,2}
Change coordinates to the triple of 2 × 2 matrices
−1
Σ{1,2},{3,4} .
Σ{1,2},{1,2} , Σ{3,4}.{1,2} , Σ{3,4},{3,4}.{1,2} .
If Σ = φG (Λ, Ω) for Λ = (λij ) ∈ RD and Ω = (ωij ) ∈ PD (B), then
ω11
λ12 ω11
,
Σ{1,2},{1,2} =
λ12 ω11
ω22
λ13
λ23
Σ{3,4}.{1,2} =
,
λ13 λ34 λ23 λ34
ω33
ω34 + λ34 ω33
.
Σ{3,4},{3,4}.{1,2} =
ω34 + λ34 ω33 ω44 + 2λ34 ω34 + λ434 ω33
We observe that Σ is in the (topological) closure of MG if and only if Σ{3,4}.{1,2}
is a matrix of rank at most one. In order for Σ to be in MG it must also hold that
the second row of Σ{3,4}.{1,2} is zero only if the first row is zero.
STRUCTURAL EQUATION MODELS
13
Geometrically, the closure of MG is equivalent to the product of two cones of
positive definite 2×2 matrices and the set of 2×2 matrices of rank at most one. The
latter set is singular at the zero matrix. For more details see the related example
in [21, Exercise 6.4].
3.6. Singularities of fibers. Finally, without going into any detail, we note that
it is also of statistical interest to study the geometry of the fibers FG (Λ, Ω). In
particular, the resolution of singularities of FG (Λ, Ω) is connected to asymptotic
approximations in Bayesian approaches to model selection.
Bayesian methods assess the goodness-of-fit of a model by integrating the likelihood function with respect to a prior distribution. In models with many parameters,
the integration is over a high-dimensional domain and, thus, constitutes a difficult
numerical problem. While carefully tuned Monte Carlo methods can be effective,
it can also be useful to invoke asymptotics. For large sample size n, the integrated
likelihood function behaves like a Laplace integral. Under regularity conditions,
a Laplace approximation can yield accurate approximations that have been used
in many applications [42]. However, the models considered here may also lead to
singular Laplace integrals for which asymptotic approximations are more involved.
Asymptotic expansions for singular Laplace integrals are well-studied [3]. The
work of Sumio Watanabe brings the ideas to bear in the statistical context [70].
For several practically relevant settings, it has become tractable to determine or
bound the real log-canonical threshold and its multiplicity, which determine how the
integrated likelihood scales with the sample size n. This information can be used
in model selection [18]. Computing real log-canonical thresholds for data generated
under the distribution with covariance matrix φG (Λ, Ω) requires careful study of
the singularities of the fiber FG (Λ, Ω). Bounds on the thresholds can be obtained
from cruder information such as the dimension of the fiber.
Part II. Treks, Subgraphs and Decomposition
4. Trek Rule
In solving the problems from Section 3, it is desirable to exploit the connection
between the covariance parametrization φG of a structural equation model and the
underlying mixed graph G = (V, D, B). The trek rule that we present in this section
makes the connection precise and is behind results that allow one to answer some
of the questions we posed with efficient algorithms.
It is natural to expect the covariance between random variables Xi and Xj to
be determined by the semi-walks between the nodes i and j in the graph G. A
semi-walk is an alternating sequence of nodes from V and edges from either D or
B such that the endpoints of each edge are the nodes immediately preceding and
succeeding the edge in the sequence. In other words, a semi-walk is a walk that uses
bidirected or directed edges but is allowed to traverse directed edges in the ‘wrong
direction’. As we will see, only special semi-walks contribute to the covariance.
Definition 4.1. A trek τ from source i to target j is a semi-walk from i to j whose
consecutive edges do not have any colliding arrowheads. In other words, τ is a
sequence of the form
(a) i ← il ← · · · ← i1 ← i0 ←→ j0 → j1 → · · · → jr → j, or
(b) i ← il ← · · · ← i1 ←−−−− i0 −−−−→ j1 → · · · → jr → j.
14
MATHIAS DRTON
A trek has a left- and a right-hand side, denoted left (τ ) and right (τ ), respectively.
We have left (τ ) = {i0 , . . . , il , i} and right (τ ) = {j0 , . . . , jr , j} in case (a), and
left (τ ) = {i0 , . . . , il , i} and right (τ ) = {i0 , j1 , . . . , jr , j} in case (b). In case (b) the
top node i0 belongs to both sides. A trek τ from i to i may have no edges, in which
case i is the top node, left (τ ) = right (τ ) = {i}. We call such a trek trivial.
In an acyclic graph, if we think of directed edges pointing ‘downward’, then a
trek takes us up and/or down a ‘mountain’. Any directed path is a trek, in which
case |left (τ ) | = 1 or |right (τ ) | = 1 depending on the direction in which the path is
traversed. A trek may contain the same node on both its left- and right-hand sides.
If the graph contains a cycle, then the left- or right-hand side of τ may contain this
cycle, possibly repeated.
For a trek τ that contains no bidirected edge and has top node i0 , define a trek
monomial as
Y
λkl .
τ (Λ, Ω) = ωi0 i0
k→l∈τ
For a trek τ that contains a bidirected edge i0 ↔ j0 , define the trek monomial as
Y
λkl .
τ (Λ, Ω) = ωi0 j0
k→l∈τ
The following rule expresses the covariance matrix Σ as a summation over treks
[57, 72, 73]. We write T (i, j) for the set of all treks from i to j.
Theorem 4.1 (Trek rule). Let G = (V, D, B) be any mixed graph, and let Λ ∈ RD
and Ω ∈ PD (B). Then the covariances are
X
τ (Λ, Ω),
i, j ∈ V.
(4.1)
φG (Λ, Ω)ij =
τ ∈T (i,j)
Some clarification is in order. If G is acyclic, then the summation in (4.1) is
finite and yields a polynomial. If G contains a directed cycle, then the right-hand
side of (4.1) may yield a power series as shown in Example 4.2 below. Under
assumptions on the spectrum of Λ, the power series converges and yields the value
of φG (Λ, Ω)ij . These spectral conditions are also needed to give cyclic models an
interpretation of representing observation of an equilibrium [31, 45]. This said, it is
also useful to treat the series as a formal power series. If so desired, a combinatorial
description can also be given for a rational expression for φG (Λ, Ω)ij ; compare [14].
Proof of the trek rule. Writing (I − Λ)−1 = I + Λ + Λ2 + . . . , we observe that
X
Y
((I − Λ)−1 )ij =
(4.2)
λkl ,
τ ∈P(i,j) k→l∈τ
where P(i, j) is the set of directed paths from i to j in G. If G is acyclic, then
Λm = 0 for all m ≥ |V |, and the geometric series of matrices has only finitely many
nonzero terms. If G is cyclic the geometric series is infinite and converges if and
only if all eigenvalues of Λ have magnitude less than 1. Now, observe that a product
of three entries of (I − Λ)−T , Ω, and (I − Λ)−1 , respectively, corresponds to the
concatenation of two directed paths at a common top node or by joining them with
a bidirected edge. A top node represents a diagonal entry of Ω, and a bidirected
edge an off-diagonal entry of Ω.
STRUCTURAL EQUATION MODELS
λ12
2
λ12
2
ω22
2
λ12
ω11
1
ω11
1
3
λ13
λ12
λ13
λ34
3
4
λ23
2
15
3
4
2
λ34
ω24
4
4
Figure 4.1. Four treks in the graph from Figure 2.1.
Example 4.1. In Example 2.2, the coordinate function (φG )24 is a polynomial with
four terms; see (2.9). The terms correspond to the four treks shown in Figure 4.1.
Example 4.2. Let G be the graph from Figure 4.2, which contains the directed
cycle 2 → 3 → 4 → 2. Due to this cycle, det(I − Λ) = 1 − λ23 λ34 λ42 . As an example
of a coordinate of φG we select
2
1
λ12 λ23 λ34 ω11 + λ12 λ13 λ34 ω11 (λ23 λ34 λ42 + 1)
φG (Λ, Ω)24 =
2
(1 − λ23 λ34 λ42 )
+λ213 λ234 λ42 ω11 + λ23 λ34 ω22 + λ234 λ42 ω33 + 2λ34 λ42 ω34 + λ42 ω44 .
To understand how this rational formula relates to the trek rule, let us focus on the
treks from 2 to 4 that use the bidirected edge 2 ↔ 4. There are then two treks for
which both left and right side are self-avoiding paths, namely,
τ1 : 2 ← 4 ↔ 3 → 4,
τ2 : 2 ← 4 ← 3 ↔ 4.
Both of these treks yield the same monomial and together contribute the term
2λ34 λ42 ω34 to (φG )24 . All other treks from 2 to 4 that use edge 2 ↔ 4 are obtained
by inserting directed cycles into τ1 or τ2 . For instance, inserting one cycle on the
left- and one on the right-hand side of τ1 gives
2 ← 4 ← 3 ← 2 ← 4 ↔ 3 → 4 → 2 → 3 → 4.
The monomials associated with these treks are
k
λ34 λ42 ω34 (λ23 λ34 λ42 ) ,
k = 1, 2, . . . .
The monomial for exponent k arises from k + 1 different treks; l = 0, 1, . . . , k cycles
are inserted on the left, the remaining k − l cycles are inserted on the right. Hence,
the contribution to (φG )24 made by all treks from 2 to 4 that use edge 2 ↔ 4 is
2
∞
X
k
(k + 1)λ34 λ42 ω34 (λ23 λ34 λ42 )
k=0
=
2λ34 λ42 ω34
,
(1 − λ23 λ34 λ42 )2
assuming that |λ23 λ34 λ42 | < 1. This explains one of the terms in the rational
expression for (φG )24 . The reasoning for the other terms is analogous.
5. Induced Subgraphs and Principal Submatrices
Suppose X = (Xi : i ∈ V ) follows the linear structural equation model given
by mixed graph G = (V, D, B), so Var[X] = φG (Λ, Ω) for some Λ ∈ RD
reg and
Ω ∈ PD (B). Let A ⊆ V be a subset of nodes. Then the covariance matrix of the
16
MATHIAS DRTON
1
3
4
2
Figure 4.2. A cyclic mixed graph.
subvector XA = (Xi : i ∈ A) is obtained by projecting to the relevant principal
submatrix, that is,
(5.1)
Var[XA ] = φG (Λ, Ω)A,A .
The resulting map (Λ, Ω) 7→ φG (Λ, Ω)A,A may be complicated, even when φG is
not.
Example 5.1. Suppose G = (V, D, ∅) is a directed graph with i → j ∈ D if
and only if i ∈
/ A and j ∈ A; the graph is thus bipartite. Then the image of
φG (Λ, Ω)A,A is the set of covariance matrices of a factor analysis model with |V \ A|
factors. Factor analysis models have complicated geometric structure, particularly
when V \ A has more than two elements [8, 20, 23, 62]. Open problems remain even
for |V \ A| = 1 when allowing additional directed edges among the nodes in A [47].
For a general mixed graph G = (V, D, B), let DA = D ∩ (A × A) be the set
of directed edges with both endpoints in A. Similarly, let BA ⊂ B be the set of
bidirected edges that have both endpoints in A. The subgraph induced by A is the
mixed graph GA = (A, DA , BA ). Example 5.1 and also already Example 1.1 show
that the covariance matrices obtained from φGA generally differ from those obtained
by projecting φG onto the A×A submatrix. However, as we now emphasize, induced
subgraphs are relevant in a special case.
Define the set of parents of a node i ∈ V as
pa(i) = {j ∈ V : j → i}.
A set of nodes A ⊆ V is ancestral if i ∈ A implies that pa(i) ⊆ A. The terminology
indicates that such a set contains all its ancestors, where an ancestor is a node
with a directed path to some node in A. Ancestral sets are obtained by recursively
removing sink nodes. We define node i to be a sink node of G if it is a sink of the
directed part (V, D), that is, if all directed edges incident to i have their head at i.
Theorem 5.1. Let G = (V, D, B) be a mixed graph, and let GA be the subgraph
induced by an ancestral set A ⊂ V . Then for all Λ ∈ RD
reg and Ω ∈ PD(B), we have
φGA (ΛA,A , ΩA,A ) = [φG (Λ, Ω)]A,A .
Proof. Let i, j ∈ A. By the trek-rule, the (i, j) entry of φG (Λ, Ω) is given by
summing the monomials associated to treks from i to j in G. Because A is ancestral,
a trek from i to j in G cannot leave A. Hence, the set of treks from i to j in G
coincides with the set of treks from i to j in the subgraph GA . Applying the
trek-rule to GA yields the claim.
Example 5.2. Take up Example 2.1. Clearly, node 3 is a sink in the graph from
Figure 1.2. Hence, the set {1, 2} is ancestral. Inspecting the matrix displayed
STRUCTURAL EQUATION MODELS
17
in (2.5), we see that removing the third row and column yields the matrix for the
induced subgraph 1 → 2.
6. Graph Decomposition
We now present a very useful graph decomposition, which allows one to address
several questions of interest by considering smaller subgraphs. The decomposition
for acyclic mixed graphs was introduced in work of Jin Tian [65, 66]. Here we
formulate the natural extension to possibly cyclic mixed graphs.
Consider partitioning the vertex set of a mixed graph G = (V, D, B) in two ways.
The first partition is given by the connected components of the bidirected part
(V, B). The second partition is obtained from the strongly connected components of
the digraph (V, D). For two distinct nodes to belong to the same strongly connected
component, there must be directed paths in either direction between them. Let
C(G) be the finest common coarsening of the two partitions. Two nodes i 6= j are
in the same block of C(G) if and only if they are connected by a path that uses only
edges that are bidirected or part of some directed cycle. Note that G is acyclic if
and only if all strongly connected components are singleton sets, in which case the
blocks of C(G) are simply the connected components of the bidirected part (V, B).
For a block C ∈ C(G), define
[
V [C] := C ∪
pa(i)
i∈C
to be the union of the block and all parents of nodes in the block. Let D[C] =
D ∩ (V [C] × C) be the set of directed edges with head in C. Let B[C] be the set
of bidirected with both endpoints in C.
Definition 6.1. The graphs G[C] = (V [C], D[C], B[C]), C ∈ C(G), form a decomposition of G, and we refer to them as the mixed components of G.
A graph decomposition partitions the edge set. As we are working with mixed
graphs both edge sets are partitioned in the decomposition. We note that the set
V [C]\C is the set of sources nodes of G[C]. Here, we define a source node as a node
that is a tail on all edges it is incident to, with the convention that both endpoints
of a bidirected edge are heads.
Example 6.1. The graph in Figure 6.1 has the bidirected components {1, 4}, {3},
and {2, 5}. The strong components of the directed part are {1}, {2, 3}, {4}, and {5}.
The finest common coarsening of the two partitions is {{1, 4}, {2, 3, 5}}. The graph
thus has two mixed components with vertex sets V [{2, 3, 5}] = {1, 2, 3, 4, 5} and
V [{1, 4}] = {1, 2, 3, 4}. The mixed component with vertex set V [{2, 3, 5}] contains
all edges with a head in {2, 3, 5}, and the second mixed component contains all
edges with a head in {1, 4}. The components are depicted in Figure 6.1.
→
For C ∈ C(G), define the projection πC
: RV ×V → RV [C]×V [C] by setting
(
λij if j ∈ C,
→
πC (Λ)ij =
0
if j ∈ V [C] \ C.
18
MATHIAS DRTON
3
2
5
3
2
=
1
3
5
2
+
4
1
4
1
4
Figure 6.1. A mixed graph is decomposed into its two mixed components.
↔
Define a second map πC
: RV ×V → RV [C]×V [C] by
ωij if i, j ∈ C,
↔
πC (Ω)ij = 1
if i = j ∈ V [C] \ C,
0
otherwise.
It projects onto the C × C submatrix of Ω and then adds an identity matrix as a
diagonal block over V [C] \ C. Define a subset of PD (B[C]) ⊂ RV [C]×V [C] as
PD I (B[C]) = {Ω = (ωij ) ∈ PD (B[C]) : ωii = 1 if i ∈ V [C] \ C} .
Then we have
→
: RD → RD[C]
πC
→
and πC
: PD (B) → PD I (B[C])
→
↔
because G[C] is a subgraph of G. With πC = (πC
, πC
), we obtain the isomorphism
Y
RD[C] × PD I (B[C]).
π = (πC )C∈C(G) : RD × PD (B) →
C∈C(G)
Example 6.2. Let G be the graph from Figure 6.1, which has C(G) = {{1, 4}, {2, 3, 5}}.
A matrix in RD is of the form
0 λ12 λ13
0
0
0 0 λ23 λ24 λ25
0 λ34
0
Λ=
.
0 λ32
0 0
0
0 λ45
0 0
0
0
0
Its projections are
0
0
→
π{2,3,5}
(Λ) =
0
0
0
λ12
0
λ32
0
0
λ13
λ23
0
0
0
0
0
0
0
0
0
λ25
0
,
λ45
0
0
0
→
π{1,4}
(Λ) =
0
0
0
An error covariance matrix in PD (B) has the form
ω11
0
0 ω14
0
0 ω22
0
0 ω25
0
0
ω
0
0
Ω=
33
,
ω14
0
0 ω44
0
0 ω25
0
0 ω55
0
0
0
0
0
0
0
0
0
0
0
λ24
λ34
0
0
0
0
0
.
0
0
STRUCTURAL EQUATION MODELS
19
and we have
1
0
↔
(Ω) =
π{2,3,5}
0
0
0
0
ω22
0
0
ω25
0
0
ω33
0
0
0 0
0 ω25
0 0
,
1 0
0 ω55
ω11
0
↔
(Ω) =
π{1,4}
0
ω14
0
0
1
0
0
0
0 ω14
0 0
1 0
0 ω44
0 0
0
0
0
.
0
1
With these preparation in place we may state Tian’s theorem as follows. Recall
that PD V is our symbol for the cone of positive definite V × V matrices.
Theorem 6.1. Let G = (V, D, B) be a mixed graph with mixed components G[C] =
(V [C], D[C], B[C]) for C ∈ C(G). Then there is an invertible map τ such that the
following diagram commutes:
Q
RD
reg × PD (B)
x
π
y
D[C]
C∈C(G) Rreg
φG
−−−−−−−−−−→
(φG[C] )C∈C(G)
× PD I (B[C]) −−−−−−−−−→
Q
PD V
x
τ
y
C∈C(G) PD V [C]
In other words, τ ◦ φG = (φG[C] ◦ πC )C∈C(G) . Both τ and its inverse are rational
Qm
maps, defined on all of PD V and all of k=1 PD Vk , respectively.
Below we give a linear algebraic proof that makes τ and its rational nature explicit. Alternatively, a proof in probabilistic notation could be given by generalizing
the proof of [66, Lemma 1]. For the generalization, the nodes in the setup of [66]
may be replaced by the strong components of the directed part (V, D).
Theorem 6.1 is a very useful result as questions about φG can be answered by
studying, one by one, the maps φG[C] for the mixed components. The fact that
τ and τ −1 are rational is important. For instance, it allows one to obtain precise
algebraic information about parameter identifiability in the sense of Definition 3.1.
Corollary 6.1. The degree of identifiability of a mixed graph G is the product of
the degrees of identifiability of its mixed components G[C], C ∈ C(G). In particular,
φG is (generically) injective if and only if each φG[C] is so, for C ∈ C(G).
Our proof of Theorem 6.1 is presented in terms of Cholesky decompositions.
When applied to a Gaussian covariance matrix, the Cholesky decomposition corresponds to factoring the multivariate normal density into a product of conditional
densities, which is the connection to the probabilistic setting of [66]. We begin by
stating a lemma on uniqueness and sparsity in block-Cholesky decomposition.
If C is a partition of a finite set V , then we write Diag(C) for the space of blockdiagonal matrices. So, A = (aij ) ∈ RV ×V is in Diag(C) if and only if aij = 0
whenever i and j are in distinct blocks of C. If we order the blocks of the partition
as C = {C1 , . . . , Ck }, then we may define a space of strictly block upper-triangular
matrices Upper(C), which contains A = (aij ) ∈ RV ×V if and only if aij = 0
whenever i ∈ Cu and j ∈ Cv with u ≥ v.
Lemma 6.1. Let Σ ∈ PD V , and let C be a partition of V , with ordered blocks.
(i) There exist unique matrices A ∈ Upper(C) and ∆ ∈ Diag(C) such that
Σ = (I − A)−T ∆(I − A)−1 .
The matrix ∆ has positive definite diagonal blocks.
20
MATHIAS DRTON
(ii) Let C ′ be a second partition of V that is coarser than C. If Σ ∈ Diag(C ′ ),
then the matrix A from (i) satisfies A ∈ Diag(C ′ ).
Proof. (i) A block-LDL decomposition yields Σ = (I + L)∆(I + L)T for unique
LT ∈ Upper(C) and ∆ ∈ Diag(C). Unit block upper-triangular matrices form
a group and, thus, (I + L)−T = I − A for A ∈ Upper(C). (ii) The claim is a
consequence of the way fill-in occurs when computing the Cholesky decomposition
of a sparse matrix [68, Section 4.1].
Proof of Theorem 6.1. We first show that, for every block C ∈ C(G), there exists
a map τC such that τC ◦ φG = φG[C] ◦ πC . Let Λ ∈ RD
reg , Ω ∈ PD (B), and
Σ = φG (Λ, Ω). Then the claim is that
−T
−1
0 ΛC,C
0 ΛC,C
I
0
I−
(6.1)
.
τC (Σ) = I −
0 ΩC,C
0 ΛC,C
0 ΛC,C
Let C(D) be the partition of V given by the strongly connected components
of (V, D). Order the blocks of C(D) topologically as W1 , . . . , Wk such that the
existence of a directed path from a node in Wu to a node in Wv implies that v ≥ u.
By Lemma 6.1(i), there are A ∈ Upper(C(D)) and ∆ ∈ Diag(C(D)) such that
(6.2)
Σ = (I − A)−T ∆(I − A)−1
Letting C = V [C] \ C, we define
−T
0 AC,C
I
(6.3)
τC (Σ) = I −
0
0 AC,C
0
∆C,C
I−
−1
0 AC,C
.
0 AC,C
If G is acyclic then Λ is strictly upper-triangular under a topological ordering
and, thus, Λ ∈ Upper(C(D)). When G has directed cycles, then Λ is block uppertriangular but not strictly so. Hence, we consider the block-diagonal matrix
∆Λ = diag(I − ΛW,W : W ∈ C(D)),
which is invertible because det(I − Λ) = det(∆Λ ) and Λ ∈ RD
reg . Hence,
−1
−T
−1
(6.4)
Σ = (I − Λ)∆−1
∆−1
(I − Λ)∆−1
.
Λ
Λ Ω∆Λ
Λ
Because ∆Λ , Ω ∈ Diag(C(G)), we have
−1
Ω̃ = ∆−T
∈ Diag(C(G)).
Λ Ω∆Λ
Moreover, due to the block upper-triangular shape of Λ,
Λ̃ = I − (I − Λ)∆−1
∈ Upper(C(D)).
Λ
By Lemma 6.1(i) and (ii), there are ∆Ω ∈ Diag(C(D)) and U ∈ Upper(C(D)) ∩
Diag(C(G)) such that
(6.5)
Ω̃ = (I − U )−T ∆Ω (I − U )−1 .
Combining (6.4) and (6.5) gives
−T
−1
(6.6)
Σ = (I − Λ̃)(I − U )
∆Ω (I − Λ̃)(I − U ) ,
where (I − Λ̃)(I − U ) = I − (Λ̃ + U − Λ̃U ) with Λ̃ + U − Λ̃U ∈ Upper(C(D)).
By the uniqueness in Lemma 6.1(i), equations (6.2) and (6.6) imply that
(6.7)
∆ = ∆Ω ,
(6.8)
A = Λ̃ + U − Λ̃U.
STRUCTURAL EQUATION MODELS
21
Since U ∈ Diag(C(G)), we have that
0
U(V \C)×C
.
=
(6.9)
UV ×C =
UC×C
UC×C
Therefore, by (6.8), AC,C = Λ̃C,C + UC,C − Λ̃C,C UC,C , and we deduce that
(6.10)
I − AC,C = (I − Λ̃C,C )(I − UC,C ) and AC,C = Λ̃C,C (I − UC,C ).
Moreover,
(6.11)
Ω̃C,C = (I − UC,C )−T ∆C,C (I − UC,C )−1 ,
which follows from (6.7) and the fact that
0
0
−1
=
,
(I − U ) V,C =
(I − U )−1 C,C
(I − UC,C )−1
which in turn follows from U being in Diag(C(G)).
Substituting the formulas from (6.10) into (6.3) yields that
(6.12) τC (Σ) =
I
0
−T
−Λ̃C,C
I
0
0 I − UC,C
I − Λ̃C,C
I
−Λ̃C,C
I
I
0
×
0
0 ∆C,C
0 I − Λ̃C,C
Using (6.11), we get that
−T
I
I
0
(6.13)
0
0 I − UC,C
I
0
∆C,C
0
0
I − UC,C
Recalling the definition of ∆Λ and Λ̃, we have
I
−Λ̃C,C
I
−ΛC,C
I
(6.14)
=
0
0 I − ΛC,C
0 I − Λ̃C,C
−1
0
I − UC,C
I
=
0
0
(∆−1
Λ )C,C
−1
0
Ω̃C,C
.
.
Plugging (6.13) and (6.14) into (6.12), we obtain the claim from (6.1).
The entries of the matrices A and ∆ in (6.2) are rational functions of Σ that are
defined on all of PD V . Hence, the same is true for the map τC defined in (6.3).
The value of τC (Σ) uniquely determines the matrices AC,C , AC,C , and ∆C,C in
(6.3). They are determined through a block LDL decomposition and, thus, rational
functions of τC (Σ). Knowing the three matrices for all C ∈ C(G), we can form A and
∆ and recover Σ using (6.2). Hence, τ is invertible and the inverse is rational.
Part III. Parameter Identification
7. Global Identifiability
This section discusses Question 3.1, which asks for a characterization of the
mixed graphs G = (V, D, B) for which the map φG is injective. In the statistical
literature a model with injective parametrization is also called globally identifiable.
22
MATHIAS DRTON
Example 7.1. If G is the graph from Figure 3.1, then φG is injective. Indeed, the
coefficients for the two directed edges of G satisfy σ11 λ12 = σ12 and σ44 λ43 = σ34 ;
recall (3.3). Since every positive definite matrix Σ = (σij ) ∈ R4×4 has σ11 , σ44 > 0,
these two equations always have a unique solution. Hence, all fibers FG (Λ, Ω) are
singleton sets. In contrast, if G is the graph from Figure 1.2, then only generic
fibers are singleton sets and φG is not injective; recall Example 3.1.
Our first observation ties in with classical linear algebra.
Theorem 7.1. If G = (V, D, ∅) is an acyclic digraph, then φG is injective and has
a rational inverse.
We give two proofs. The first one emphasizes the connection to Cholesky decomposition.
Proof A. Suppose V = {1, . . . , m} is enumerated in reversed topological order such
that i → j ∈ D implies that j < i. Then Λ is a strictly lower-triangular matrix,
and φG (Λ, Ω) has matrix inverse (I − Λ)Ω−1 (I − Λ)T . This is the product of a unit
lower-triangular matrix, a positive diagonal matrix and a unit upper-triangular
matrix. We may compute I − Λ and Ω−1 by an LDL decomposition.
The second proof emphasizes the graphical nature of the problem and possible
sparsity of Λ. It shows more explicitly that the inverse of φG is rational.
Proof B. Letting Σ = (σij ) = φG (Λ, Ω), we have that
(7.1)
Σpa(i),i = Σpa(i),pa(i) Λpa(i),i
because if j ∈ pa(i), then every trek from j to i ends with an edge k → i for
k ∈ pa(i). Indeed, a trek from j to i for which this fails has to be a directed path
from i to j. Adding the edge j → i to this path would yield a directed cycle.
Similarly, every nontrivial trek from i to i begins and ends with a directed edge
whose tail is a parent of i. Hence,
(7.2)
σii = ωii + ΛTpa(i),i Σpa(i),pa(i) Λpa(i),i .
The matrix Σpa(i),pa(i) is a principal submatrix of the positive definite matrix Σ
and, thus, invertible. Therefore,
−1
Λpa(i),i = Σpa(i),pa(i)
(7.3)
Σpa(i),i ,
−1
ωii = σii − Σi,pa(i) Σpa(i),pa(i)
Σpa(i),i .
Proof B shows that the formula from (7.3) holds more generally. It merely needs
to hold that every trek from a node j ∈ pa(i) to i ends with a directed edge with i
as its head, so an edge of the form k → i. This holds for every node in the graph if
and only if the graph is ancestral [50]. A mixed graph is ancestral if the presence
of a directed path from node i to node j implies that i 6= j and i ↔ j ∈
/ B. An
ancestral graph is in particular acyclic.
The next easy lemma is crucial for the understanding of injectivity of φG .
Lemma 7.1. If φG is injective and H ⊆ G is a subgraph, then φH is injective.
Proof. Any subgraphs can be obtained by removing edges one at a time, and then
removing isolated nodes. If H is obtained from G = (V, D, B) by removing the
edge i → j, then φH is the restriction of φG to the subset of matrices Λ ∈ RD
reg
that have λij = 0. If we instead remove the edge i ↔ j, then the restrictions is to
STRUCTURAL EQUATION MODELS
1
(a)
3
2
1
4
(b)
23
3
2
4
Figure 7.1. (a) A mixed graph for which the parametrization φG
is injective. (b) A graph for which φG is not injective.
matrices Ω ∈ PD (B) with ωij = 0. If H is obtained by removing the isolated node
i, then
φH (Λ, Ω) 0
.
φG (Λ, Ω) =
0
ωii
In either case non-injectivity of φH implies non-injectivity of φG .
Corollary 7.1. If φG is injective, then G is simple, that is, for any vertices i 6= j
at most one of the three edges i ↔ j, i → j and i ← j may appear in G.
Proof. If G is not simple then it contains a subgraph H with two nodes and two
edges. The map φH is infinite-to-one as it maps a four-dimensional domain into
the three-dimensional set of symmetric 2 × 2 matrices. Now apply Lemma 7.1.
Theorem 7.2. If φG is injective, then G is acyclic.
Proof sketch. By Lemma 7.1, we may restrict to studying directed cycles 1 → 2 →
. . . → m → 1. The case of m = 2 is covered by Corollary 7.1. If m ≥ 3, then
it is possible to show that φG is generically 2-to-1, that is, the fiber FG (Λ, Ω) is
generically of size two [16].
It remains to characterize injectivity for acyclic graphs G = (V, D, B). The
following theorem shows that injectivity can be decided in polynomial time by
alternatingly decomposing the bidirected part (V, B) into connected components
and removing sink nodes of the directed part (V, D).
Theorem 7.3. Suppose G is an acyclic mixed graph. Then:
(a) φG is injective if and only if G does not contain a subgraph whose bidirected
part is connected and whose directed part has a unique sink.
(b) If φG is injective, then its inverse is rational and MG smooth.
Figure 7.1 illustrates the characterization in part (a) of the theorem. The fact
that φG is not injective if the combinatorial condition in (a) fails can be shown by
a counterexample for the particular subgraph and then invoking Lemma 7.1. For
a proof of the sufficient condition in Theorem 7.3, we refer the reader to [16]. We
note that the sufficiency of the condition for injectivity can be proven by repeatedly applying the graph decomposition result in Theorem 6.1 and the result on
ancestral subgraphs from Theorem 5.1. These results as well as Theorem 7.3 have
generalizations to nonlinear structural equation models [56, 66].
8. Generic Identifiability
The difference between injectivity and generic injectivity of φG may appear
minute. However, the two properties are quite different, and failure of generic
injectivity cannot be argued by studying subgraphs (as in Lemma 7.1). According
to Corollary 7.1, a mixed graph G can have the map φG injective only if it is acyclic
24
MATHIAS DRTON
and simple. The deeper issue is then to find out which simple acyclic mixed graphs
have φG injective. In contrast, the next result shows that all simple acyclic mixed
graphs are generically injective. The deeper issue for generic injectivity is thus the
treatment of graphs that contain directed cycles or are not simple.
Theorem 8.1. If G = (V, D, B) is acyclic and simple, then φG is generically
injective and algebraically one-to-one.
The theorem is due to [7]. It shows that the graph from Figure 7.1(b) has φG
generically injective, but not injective. A short proof of Theorem 8.1 is obtained
through the following observation that we use repeatedly in this section.
Lemma 8.1. Let G = (V, D, B) be a mixed graph, and let Σ = φG (Λ0 , Ω0 ) for
Λ 0 ∈ RD
reg and Ω0 ∈ PD (B). The fiber FG (Λ0 , Ω0 ) is isomorphic to the set of
matrices Λ ∈ RD
reg that solve the equation system
(8.1)
(I − Λ)T Σ(I − Λ) ij = 0,
i 6= j, i ↔ j ∈
/ B.
Proof. The projection (Λ, Ω) 7→ Λ maps FG (Λ0 , Ω0 ) to the set of matrices Λ ∈ RD
reg
that solve the equations in (8.1). Indeed, as I − Λ is invertible for Λ ∈ RD
reg ,
Σ = φG (Λ, Ω) = (I − Λ)−T Ω(I − Λ) =⇒ Ω = (I − Λ)T Σ(I − Λ).
Any entry of Ω that is indexed by a pair (i, j) with i 6= j and i ↔ j ∈
/ B is zero.
Conversely, if Λ ∈ FG (Λ), then (Λ, (I − Λ)T Σ(I − Λ)) ∈ FG (Λ, Ω).
We emphasize that the equations in (8.1) are bilinear as
X
X
X
σil λlj +
λki σki −
(I − Λ)T Σ(I − Λ) ij = σij −
k∈pa(i)
l∈pa(j)
X
λki σkl λlj .
k∈pa(i) l∈pa(j)
Proof of Theorem 8.1. Because G is acyclic, we may enumerate the vertex set in a
topological order as V = {1, . . . , m}. Then pa(i) ⊆ {1, . . . , i − 1} for i = 1, . . . , m.
Moreover, because G is simple, j ∈ pa(i) implies that j ↔ i ∈
/ B. By Lemma 8.1,
T
i = 1, . . . , m.
(I − Λ) Σ(I − Λ) pa(i),i = 0,
These equations can be rewritten as
(8.2)
(I − Λ)T Σ pa(i),pa(i) Λpa(i),i = (I − Λ)T Σ pa(i),i ,
i = 1, . . . , m.
By the tolopogical order, if j ∈ pa(i), then the entries in the j-th row of (I − Λ)T Σ
depend only on the first i − 1 columns of Λ. The system in (8.2) can thus be solved
recursively, where each step requires solving a linear system.
To show that φG is generically injective, it remains to argue that the equations
in (8.2) generically have a unique solution. It suffices to exhibit a single pair (Λ, Ω)
for which this is true. We may choose Λ = 0 and Ω = I, so Σ = I. Then the matrix
for the i-th group of equations in (8.2) is Σpa(i),pa(i) , which is invertible.
Although a combinatorial characterization of the graphs with generically injective parametrization φG is not known, Gröbner basis techniques can be used to determine the degree of identifiability from Definition 3.1 and, thus, decide whether
φG is algebraically one-to-one. Gröbner bases are computationally tractable for
non-trivial examples and have been used for a classification of all graphs with up
to 5 nodes [34]. For larger graphs, algebraic methods can be applied after decomposition according to Theorem 6.1.
STRUCTURAL EQUATION MODELS
25
We describe two options for the computation. In either case, we advocate
working with the equation system from (8.1) as opposed to the fiber equation
Σ = φG (Λ, Ω). System (8.1) has Ω eliminated and may be far more compact as
it avoids the inversion of I − Λ. This said, although system (8.1) is polynomial
also for graphs that contain directed cycles, care must be taken to avoid spurios
solutions with I − Λ non-invertible.
The first possibility is to perform a parametric Gröbner basis computation. We
introduce a matrix Λ whose nonzero entries λij , i → j ∈ D, are indeterminates
and a pair of matrices (Λ0 , Ω0 ) that are parameters. We form the matrix (I −
Λ)T φG (Λ0 , Ω0 )(I − Λ) and set to zero the off-diagonal entries indexed by non-edges
of the bidirected part (V, B). We then compute a Gröbner basis for the resulting
system in the polynomial ring with coefficients in the field of rational fractions
R(Λ0 , Ω0 ). The Gröbner basis readily yields the dimension of the generic fibers. If
the dimension is finite we may also find the algebraic degree of the generic fibers,
which is what we referred to as degree of identifiability. When the graph G contains
directed cycles, the matrix I − Λ can be non-invertible. We thus first saturate our
equation system with respect to det(I − Λ).
Example 8.1. The following code for the system Singular [13] implements the
approach just described for a directed 3-cycle:
LIB "linalg.lib"; option(redSB);
ring R = (0,l012,l023,l031,w011,w022,w033),(l12,l23,l31),dp;
matrix L[3][3] = 1,-l12,0,
0,1,-l23,
-l31,0,1;
matrix L0[3][3] = 1,-l012,0,
0,1,-l023,
-l031,0,1;
matrix W0[3][3] = w011,0,0,
0,w022,0,
0,0,w033;
matrix W[3][3] = transpose(L)*inverse(transpose(L0))*W0*inverse(L0)*L;
ideal GB = sat(ideal(W[1,2],W[1,3],W[2,3]), det(L))[1];
dim(GB); mult(GB);
The output from the last line first certifies that the fibers are generically zerodimensional, that is, contain finitely many points. The multiplicity computed with
the last command shows the degree of identifiability to be two.
The second possibility is to consider only polynomials with real-valued coefficients but to introduce as polynomial variables the nonzero entries of Λ as well as a
symmetric matrix Σ. These variables are ordered with respect to a block monomial
order in which the variables in Λ are larger than the variables in Σ. Let I be the
ideal generated by the off-diagonal entries of (I − Λ)T Σ(I − Λ) that are indexed by
the non-edges of (V, B). Saturate I with respect to det(I − Λ). Let J be the reduced Gröbner basis of the resulting ideal. Elimination theory yields the following
[34, Section 8 of the supplemental material].
Proposition 8.1. The parametrization φG of a mixed graph G = (V, D, B) is
algebraically one-to-one if and only if for each i → j ∈ D, the reduced Gröbner
basis J contains an element with leading monomial a(Σ)λij .
26
MATHIAS DRTON
More generally, the generic dimension and degree of the fibers FG (Λ, Ω) can be
determined by analyzing the monomials under the staircase of the initial ideal of
J [12, Chapter 9]. This way we may determine the degree of identifiability of G.
In comparison to the first approach, the second method yields relations that show
how to identify coefficients λij from Σ.
Example 8.2. Treating again a directed 3-cycle, we give an example of the second
type of computation in Singular:
LIB "linalg.lib"; option(redSB);
ring R = 0,(l12,l23,l31,s11,s12,s13,s22,s23,s33),(dp(3));
matrix L[3][3] = 1,-l12,0,
0,1,-l23,
-l31,0,1;
matrix S[3][3] = s11,s12,s13,
s12,s22,s23,
s13,s23,s33;
matrix W[3][3] = transpose(L)*S*L;
ideal GB = sat(ideal(W[1,2],W[1,3],W[2,3]), det(L))[1]; GB;
The output is a list of 9 polynomials whose leading terms are, in our usual notation,
λ23 λ31 σ23 ,
λ12 λ31 σ13 ,
λ12 λ23 σ12 ,
λ12 λ31 σ12 σ33 ,
λ12 σ11 σ13 σ23 ,
λ12 λ23 σ11 σ23 ,
λ23 σ12 σ13 σ22 ,
λ23 λ31 σ13 σ22 ,
λ231 σ13 σ22 σ33 .
By Proposition 8.1, φG is not algebraically one-to-one because there is no leading
term of the form λ31 a(Σ). The last leading term belongs to a polynomial that
shows that λ31 is algebraic function of degree 2 of the covariance matrix Σ because
it solves the equation
2
2
2
λ231 σ33 (σ13 σ22 − σ12 σ23 ) − λ31 (σ13
σ22 − σ11 σ23
− σ12
σ33 + σ11 σ22 σ33 )
+ σ11 (σ13 σ22 − σ12 σ23 ) = 0.
The equations with leading terms λ23 σ12 σ13 σ22 and λ12 σ11 σ13 σ23 show that λ23
and λ12 are rational functions of Σ and λ31 . Altogether, we have verified that φG
is algebraically 2-to-one. Checking this by counting monomials under the staircase
means considering the leading monomials considering only the variables λ12 , λ23 , λ31
we seek to solve for. The monomials are
(8.3)
λ23 λ31 ,
λ12 λ31 ,
λ12 λ23 ,
λ12 ,
λ23 ,
λ231 ,
and generate the ideal I = hλ12 , λ23 , λ231 i. The monomials under the staircase are
the monomials in R[λ12 , λ23 , λ31 ] \ I. There are two, namely, 1 and λ31 .
Although Gröbner basis methods can be effective, it is desirable to obtain combinatorial methods that are efficient also for large-scale problems. The half-trek
criteria of [34] are state-of-the-art criteria that can be checked in time that is polynomial in the size of the vertex set of the considered graph. They provide a sufficient
as well as a necessary condition for generic injectivity of φG . More precisely, there
is a condition that is sufficient for φG to be algebraically one-to-one and a related
condition that is neessary for φG to be generically finite-to-one. The conditions are
implemented in a package for the R project for statistical computing [4]. We begin
our discussion of the half-trek criteria by introducing some needed terminology.
A half-trek from source node i to target node j is a trek τ from i to j whose
left-hand side is a singleton set, so left (τ ) = {i}. In other words, a half-trek is of
STRUCTURAL EQUATION MODELS
27
the form
i → j1 → . . . → jr → j or i ↔ j1 → . . . → jr → j.
Let X, Y ⊆ V be two sets of nodes of equal cardinality |X| = |Y | = k. Let Π be a
set of k treks. Then Π is a system of treks from X to Y , denoted Π : X ⇒ Y , if X
is the set of source nodes of the treks in Π and Y is the set of target nodes. Note
that we allow X ∩ Y 6= ∅. The system Π is a system of half-treks if every trek πi is
a half-trek. Finally, a system Π has no sided intersection if
left (π) ∩ left (π ′ ) = ∅ = right (π) ∩ right (π ′ )
for all pairs of treks π, π ′ ∈ Π.
Definition 8.1. A set Y ⊆ V satisfies the half-trek criterion with respect to node
i if (i) |Y | = | pa(i)|, (ii) j = i or j ↔ i implies that j 6∈ Y , and (iii) there exists a
system of half-treks Π : Y ⇒ pa(i) that has no sided intersection.
Theorem 8.2. Let G = (V, D, B) be a mixed graph.
(i) If for every i ∈ V there exists a set Yi ⊆ V that satisfies the half-trek
criterion with respect to i, and if there exists a total ordering ≺ such that
j ≺ i whenever j ∈ Yi and there is a half-trek from i to j, then φG is
generically injective and algebraically one-to-one.
(ii) For φG to be generically finite-to-one it is necessary that there exists a
family of subsets Yi ⊆ V , i ∈ V , such that Yi satisfies the half-trek criterion
with respect to i and j ∈ Yi implies i 6∈ Yj .
We merely outline the proof of the theorem; for details see [34]. Some of the
arguments are further illustrated in Example 8.3. Note also that Theorem 8.1 is
obtained from Theorem 8.2(i) by taking Yi = pa(i) and ≺ as a topological order.
Outline of proof of Theorem 8.2. (i) Let Σ = φG (Λ0 , Ω0 ) for Λ0 ∈ RD
reg and Ω0 ∈
PD(B). Suppose (Λ, Ω) ∈ FG (Λ0 , Ω0 ). To show that (Λ, Ω) = (Λ0 , Ω0 ), we visit
the nodes i ∈ V from smallest to largest in the order ≺ and iteratively find a linear
equation system that is uniquely solved by the i-th column of Λ. The starting point
is Lemma 8.1, by which we have
(8.4)
(I − Λ)T Σ(I − Λ) Yi ,i = 0, i ∈ V.
This is true because Definition 8.1 yields that j 6= i and j ↔ i ∈
/ B when j ∈ Yi .
Similar to the proof of Theorem 8.1, we may rearrange (8.4) to
Ai (Λ, Σ)Λpa(i),i = bi (Λ, Σ),
with Ai (Λ, Σ) = (I − Λ)T Σ Yi ,pa(i) and bi (Λ, Σ) = (I − Λ)T Σ Yi ,i . Both Ai (Λ, Σ)
and bi (Λ, Σ) can be shown to only depend on those columns of Λ that are indexed
by nodes j with a half-trek from i to j. Hence, the proof is complete if we can show
that Ai (Λ0 , Σ) is invertible for generic choices of Λ0 and Ω0 . To verify this, we may
use the existence of a half-trek system without sided intersection from Yi to pa(i)
to argue that the determinant of Ai (Λ0 , Σ) is not the zero polynomial. This last
step is in the spirit of the Lindström-Gessel-Viennot lemma.
(ii) We study the Jacobian of the equations from Lemma 8.1. Its rows are indexed
by the non-edges of the bidirected part (V, B) and its columns by the edges in D.
For φG to be generically finite-to-one, it is necessary that the Jacobian has full
column rank D. It can be shown that the Jacobian contains an invertible |D| × |D|
submatrix only if the given condition holds. Let Ji be the submatrix of the Jacobian
28
MATHIAS DRTON
1
3
2
5
4
Figure 8.1. Illustration of Theorem 8.2.
obtained by selecting the columns corresponding to directed edges with head at i.
Then, more specifically, Ji has full column rank only if there exists a subset Yi ⊆ V
that satisfies the half-trek criterion with respect to i. Moreover, if j ∈ Yi and i ∈ Yj ,
then the same row, namely, that corresponding to i ↔ j ∈
/ B, would be used to get
an invertible square submatrix of Ji and Jj .
The conditions from Theorem 8.2 seem involved but as we mentioned they can
be checked in polynomial time. Indeed, the problem of finding suitable sets Yi that
satisfy the half-trek criteria can be shown to correspond to solving network-flow
problems. For condition (i), we repeatedly solve a network-flow problem. Condition
(ii) can be implemented as a single larger network-flow problem [34, Section 6].
Example 8.3. Let G be the graph in Figure 8.1. Each one of the sets
Y1 = {2, 5},
Y2 = {5},
Y3 = ∅,
Y4 = ∅,
Y5 = {3}
satisfies the half-trek criterion with respect to the node it is indexed by. This is least
evident for Y1 , and we highlight the half-treks 2 ↔ 3 and 5 ↔ 4 → 2, which have
no sided intersection. Choosing the ordering as 3 ≺ 4 ≺ 5 ≺ 1 ≺ 2, Theorem 8.2(i)
shows that φG is algebraically one-to-one. Other possible orderings are obtained
by permuting {3, 4, 5} or {1, 2}.
To illustrate ideas from the proof of Theorem 8.2(i), we focus on node 1, with
pa(1) = {2, 3}. Since Y1 = {2, 5}, we work with the equations
(I − Λ)T Σ(I − Λ) 21 = 0.
(I − Λ)T Σ(I − Λ) 51 = 0,
Expanding out the matrix product, the equations become
(8.5)
σ51 − (σ52 λ21 + σ53 λ31 ) − λ35 σ31 + (λ35 σ32 λ21 + λ35 σ33 λ31 ) = 0,
(8.6)
σ21 − (σ22 λ21 + σ23 λ31 ) − λ42 σ41 + (λ42 σ42 λ21 + λ42 σ43 λ31 ) = 0,
and we wish to solve for λ21 and λ31 . With 5 ≺ 1, we have already solved for λ35 ;
since Y5 = pa(5) we have λ35 = σ35 /σ55 as is also clear from the discussion after
Proof B for Theorem 7.1. Substituting the ratio for λ35 turns (8.5) into a linear
equation in λ21 and λ31 . The equation in (8.6) could be linearized similarly, except
that now the relevant coefficient λ42 has not yet been determined in an ordering
with 1 ≺ 2. However, if Λ is part of a pair (Λ, Ω) in the fiber given by Σ, then
−λ42 σ41 + (λ42 σ42 λ21 + λ42 σ43 λ31 ) = 0
because there is no half-trek from 1 to 2. To see this note that the term λ42 σ41
corresponds to treks from 2 to 1 that start with the edge 2 ← 4, whereas the sum
λ42 σ42 λ21 + λ42 σ43 λ31 corresponds to treks from 2 to 1 that start with the edge
2 ← 4 and end in either 2 → 1 or 3 → 1. These two sets of treks coincide when
there is no half-trek from 1 to 2.
STRUCTURAL EQUATION MODELS
29
5
1
5
2
1
4
2
3
3
4
(a)
(b)
Figure 8.2. Graphs that satisfy the necessary but not the sufficient condition from Theorem 8.2: (a) φG is algebraically 3-to-one,
(b) φG is algebraically one-to-one and, thus, generically injective.
The sufficient condition from Theorem 8.2(i) can be strengthened by first applying the graph decomposition from Section 6 and then check condition (i) in
each subgraph. No such strengthening is possible for the necessary condition from
part (ii) of the theorem [34]. Further strengthening of the sufficient condition is
possible by first removing sink nodes from the graph using the observation from
Theorem 5.1. When a sink node is removed a more refined graph decomposition
may become possible; we refer the reader to [10, 22]. While a specific polynomialtime algorithm using this idea is given in [22], it is still unclear how to best design
algorithms based on recursive graph decomposition and removal of sink nodes.
We conclude our discussion of parameter identification with two examples from
the exhaustive computational study of graphs with up to 5 nodes in [34]. Both
graphs in Figure 8.1 satisfy the necessary condition in Theorem 8.2(ii) and, thus,
have φG generically finite-to-one. Neither graph satisfies the sufficient condition
from Theorem 8.2(i). The graph in panel (a) indeed does not have φG generically
injective. Instead, φG is algebraically 3-to-one. The graph in panel (b), however,
is algebraically one-to-one but Theorem 8.2(i) fails to recognize it. Decomposition
and removal of sink nodes do not help.
Part IV. Relations Among Covariances
9. Implicitization
Let MG be the set of covariance matrices of the structural equation model
given by a mixed graph G = (V, D, B). Motivated, in particular, by the covariance
equivalence problem from Question 3.5, we now discuss polynomial relations among
the entries of the matrices in MG . Let Σ = (σij ) be a symmetric V × V matrix of
variables, and let R[Σ] be the ring of polynomials in the σij with real coefficients.
Then the polynomial relations we seek to understand make up the vanishing ideal
I(G) = {f ∈ R[Σ] : f (Σ) = 0 for all Σ ∈ MG } .
Since MG is the image of the rational map φG , we may compute a generating set
for I(G) by implicitization. Assuming for simplicity that G is acyclic and, thus,
φG polynomial, we have
I(G) = h Σ − φG (Λ, Ω) i ∩ R[Σ].
30
MATHIAS DRTON
1
3
2
4
Figure 9.1. An acyclic digraph.
A better approach, however, is to start with the equations from Lemma 8.1, which
have Ω already eliminated. We compute
D
E
I(G) =
/ B ∩ R[Σ].
(I − Λ)T Σ(I − Λ) ij : i 6= j, i ↔ j ∈
If G is not acyclic, we saturate with respect to det(I − Λ) before intersecting with
R[Σ]; compare the examples in Section 8.
Example 9.1. To illustrate the use of a different software, we change to Macaulay2
[40]. The following code computes a generating set for the vanishing ideal of the
graph G shown in Figure 9.1:
R = QQ[l12,l13,l24,l34, s11,s12,s13,s14, s22,s23,s24, s33,s34, s44,
MonomialOrder => Eliminate 4];
Lambda = matrix{{1, -l12, -l13, 0},
{0, 1, 0, -l24},
{0, 0, 1, -l34},
{0, 0, 0, 1}};
S = matrix{{s11, s12, s13, s14},
{s12, s22, s23, s24},
{s13, s23, s33, s34},
{s14, s24, s34, s44}};
W = transpose(Lambda)*S*Lambda;
I = ideal{W_(0,1),W_(0,2),W_(0,3),W_(1,2),W_(1,3),W_(2,3)};
eliminate({l12,l13,l24,l34},I)
The ‘GraphicalModels’ package [37] automates the computation:
needsPackage "GraphicalModels";
G = digraph {{1,{2,3}},{2,{4}},{3,{4}}};
R = gaussianRing G;
gaussianVanishingIdeal R
Reproduced in our notation, the output shows that the ideal I(G) is generated by
the two polynomials
(9.1) f1 = σ12 σ13 − σ11 σ23 ,
2
(9.2) f2 = σ14 σ23
− σ13 σ23 σ24 − σ14 σ22 σ33 + σ12 σ24 σ33 + σ13 σ22 σ34 − σ12 σ23 σ34 .
Computing I(G) using Gröbner bases is a method that applies to any mixed
graph but can be computationally prohibitive for graphs with more than 5 or 6
nodes. In order to solve model equivalence problems combinatorial insight on particular types of relations is needed.
Example 9.2. Continuing with Example 9.1, observe that the two polynomials
f1 and f2 from (9.1) and (9.2) are determinants of submatrices of the covariance
STRUCTURAL EQUATION MODELS
matrix Σ. The following
σ11 σ12
σ12 σ22
σ13 σ23
σ14 σ24
two displays locate the submatrices
σ13 σ14
σ11 σ12 σ13
σ12 σ22 σ23
σ23 σ24
σ13 σ23 σ33
σ33 σ34
σ14 σ24 σ34
σ34 σ44
31
σ14
σ24
σ34
σ44
and show two boxes to visualize that each submatrix is obtained by bordering
a principal submatrix by one additional row and column. The determinants f1
and f2 are seen to be almost principal minors of Σ. As we discuss in Section 10,
the vanishing of almost principal minors of a Gaussian covariance matrix has a
particular probabilistic meaning, namely, conditional independence [60].
10. Conditional Independence
Suppose X = (Xi : i ∈ V ) is a Gaussian random vector indexed by a finite set V
and with covariance matrix Σ ∈ PD V . Let i, j ∈ V be two distinct indices, and let
S ⊆ V \{i, j}. The random variables Xi and Xj are conditionally independent given
XS if and only if det(ΣiS×jS ) = 0. This connection between Gaussian conditional
independence and the vanishing of almost principal minors of the covariance matrix
is explained in detail in [21, Chapter 3] and [48].
It is fully understood which conditional independence relations hold in the covariance matrices of a linear structural equation model. The following concepts
are needed to state the result. Let π be a semi-walk from node i to node j in the
considered mixed graph G = (V, D, B), and let node k be a non-endpoint of π. A
collider on π is a node k that is an internal node on π, and a head on the two edges
that precede and succeed k on π. We recall our convention that the two nodes
incident to a bidirected edge are both heads. Pictorially, π includes the segment
→ k ←, → k ↔, ↔ k ← or ↔ k ↔. A non-collider on π is an internal node of π
that is not a collider on π.
Definition 10.1. Fix a set S ⊆ V . Two nodes i, j ∈ V are d-connected by S if
G contains a semi-walk from i to j that has all colliders in S and all non-colliders
outside S. If no such semi-walk exists, then i and j are d-separated by S.
The following theorem was first proven for acyclic digraphs [39] and later extended to cover arbitrary mixed graphs [58].
Theorem 10.1. Let G = (V, D, B) be a mixed graph. Two nodes i, j ∈ V are dseparated by S if and only if det (φG (Λ, Ω)iS×jS ) = 0 for all Λ ∈ RD
reg , Ω ∈ PD (B).
For acyclic digraphs, the theorem can be derived in a probabilistic setup that
extends to non-Gaussian conditional independence [46]. When lecturing about the
result, the author likes to discuss the example with three binary variables shown
in Figure 10.1. In our linear Gaussian setting, Theorem 10.1 is a special case of
Theorem 11.1 that we treat in more detail.
Define the conditional independence ideal
ICI (G) = h det (ΣiS×jS ) : i, j d-separated by S i.
By Theorem 10.1, ICI (G) ⊆ I(G) for any mixed graph G. In Example 9.1, ICI (G) =
I(G) but this may be false even for acyclic digraphs [61]. Nevertheless, if G is an
32
MATHIAS DRTON
Boring talk?
Early morning talk?
Audience asleep?
Figure 10.1. A graph for three binary variables.
acylic digraph, then the set of covariance matrices MG is cut out by conditional
independence relations. In other words, for an acyclic digraph,
(10.1)
MG = V (ICI (G)) ∩ PD V
is the positive definite part of the variety of the conditional independence ideal. The
equality in (10.1) also holds when G is an ancestral graph, as defined in Section 7.
However, it is false in general as can be seen in Example 3.3. We remark that for
acyclic digraphs it has been proven that saturating the conditional independence
with respect to all principal minors yields the vanishing ideal [52]:
Y
∞
I(G) = I CI (G) :
det(ΣA×A )
.
A⊂V
For acyclic digraphs and more generally ancestral graphs, the equality from (10.1)
allows us to answer Question 3.5 on covariance equivalence by checking whether
two graphs have the same d-separation relations. The latter comparison can be
accomplished in polynomial time. We state the result for acyclic digraphs [35, 69].
A generalization for ancestral graphs was given more recently [1]; see also [74].
Theorem 10.2. Let G and G′ be two acyclic digraphs. Then MG = MG′ if and
only if G and G′ have the same adjacencies and the same unshielded colliders. An
unshielded collider is an induced subgraph of the form i → j ← k.
We remark that it can also be decided in polynomial time whether two digraphs
that are not necessarily acyclic have the same d-separation relations [51]. When
the graphs have directed cycles then d-separation equivalence is necessary but not
sufficient for covariance equivalence; see, e.g., the example in [15].
11. Trek Separation
We now turn to the characterization of the vanishing of general minors of the
covariance matrices in a linear structural equation model. Our first example clarifies
the importance of minors that are not almost principal.
Example 11.1. If G is the graph from Figure 3.2 and Example 3.3, then I(G) is
generated by det(Σ12,34 ). Such off-diagonal 2 × 2 minors are known as tetrads in
the statistical literature; see [20] and the references therein.
The tetrad representation theorem gives a combinatorial characterization of the
vanishing of any 2 × 2 determinant [57]. The theorem has been greatly generalized, and we now have a full combinatorial understanding of when a minor of the
covariance matrix vanishes based on the notion of trek-separation [64]. Moreover,
the non-vanishing determinants can be described precisely [14].
STRUCTURAL EQUATION MODELS
33
Definition 11.1. Let A, C, SA , SC ⊆ V be sets of nodes of the mixed graph G =
(V, D, B). The pair (SA , SC ) trek-separates A and C if every trek from i ∈ A to
j ∈ C intersects SA on its left side or SC on its right side.
Theorem 11.1. Let G be a mixed graph. Then the A × C submatrix of φG (Λ, Ω)
has generic rank ≤ r if and only if there exist sets SA and SC with |SA | + |SC | ≤ r
such that (SA , SC ) trek-separates A and C.
The theorem for acyclic mixed graphs is proven in [64]. The cases with directed
cycles are covered by the results in [14]. We describe the ideas behind the proof.
Proof outline. The problem can be reduced to the case where G is a digraph by
subdividing bidirected edges. For each edge i ↔ j ∈ B we introduce a new node
{i, j} and two edges {i, j} → i and {i, j} → j. The new digraph G′ thus has
|V | + |B| nodes and |D| + 2|B| edges. If G is the mixed graph from Figure 1.2, then
G′ is the digraph in Figure 1.1. Every trek in G corresponds to a trek in G′ where
an edge i ↔ j in G corresponds to i ← {i, j} → j in G′ . The entries of (φG )A,C
and those of (φG′ )A,C can be shown to admit the same set of relations.
In the sequel, assume that G itself is a digraph. Then PD (B) contains diagonal
matrices with positive entries. The rank of φG (Λ, Ω) for Λ ∈ RD
reg and Ω ∈ PD (B)
is then the same as that of Σ = φG (Λ, I).
To establish the claim, we may study the vanishing of the determinants of square
submatrices. So assume that |A| = |C| = r + 1. The Cauchy-Binet formula gives
that
X
(11.1)
det (ΣA×C ) =
det (I − Λ)−1 E×A det (I − Λ)−1 E×C ,
E
where the sum is over subsets E ⊆ V of cardinality r + 1. By the Lindström-GesselViennot lemma for general digraphs, it holds that
(11.2)
det (I − Λ)−1 E×A = 0 for all Λ ∈ RD
reg ,
if and only if no system of r + 1 directed paths from E to A is vertex-disjoint.
Applying this to all terms in (11.1) shows that det (ΣA×C ) vanishes if and only if
every system of treks from A to C has a sided intersection. Here, an intersection
on the left side of a trek corresponds to the vanishing of determinants of the matrices (I − Λ)−1 E×A and, similarly, interesections on the right side are related to
the determinants of the matrices (I − Λ)−1 E×C . The characterization by sided
intersections in trek systems can be turned into the claimed statement about trekseparation via Menger’s theorem. To account for the distinct role played by the
left and the right sides of the treks, Menger’s theorem is applied in a digraph G̃
that results from duplicating the nodes and edges of G. Each node and each edge
of G has a left and a right side version in G̃. Menger’s theorem yields a cut set S of
cardinality |S| ≤ r in G̃. Partitioning S according to the left and right side yields
the pair (SA , SC ) for the claimed trek-separation.
Example 11.2. When G is the the graph from Figure 3.2, then the submatrix
[φG (Λ, Ω)]12,34 has generic rank 1; recall Example 11.1. Theorem 11.1 shows that
the rank is at least 1 because (∅, ∅) does not trek-separate {1, 2} and {3, 4}. For
instance, there is the trek 1 → 3. That the rank is no larger than 1 follows from
(∅, {3}) trek-separating {1, 2} and {3, 4}. For instance, the node 3 is on the right
side of the two treks 1 → 3 and 2 → 3 → 4.
34
MATHIAS DRTON
c
1
2
3
4
5
6
7
Figure 11.1. A ‘spider graph’ with {1, 2, 3, 4} × {5, 6, 7} submatrix of rank two.
Example 11.3. What is the generic rank of the A × C of φG (Λ, Ω) when G is the
‘spider graph’ from Figure 11.1, A = {1, 2, 3, 4} and C = {5, 6, 7}? The node c is
on every trek between A and C. It follows that ({c}, {c}) trek-separates A and C
and thus the rank is no larger than two. In fact, the rank is equal to two. Consider
the two treks
π1 : 1 ↔ ◦ → c → 5,
π2 : 3 ← c ← ◦ ↔ 6.
They have only node c in common but c ∈ right (π1 ) and c ∈ left (π2 ). Hence, a
pair of trek-separating sets must use at least two nodes.
Trek separation solves the problem of characterizing the vanishing of determinants. However, there is currently no efficient method for deciding when two mixed
graphs are trek separation equivalent.
12. Verma Constraints
Much interesting ground lies beyond determinants of the covariance matrix. We
content ourselves with two examples concerning a relation first presented in [69].
Example 12.1. Let G be the graph from Figure 2.1. The graph has no trekseparation relations as can be checked with the package ‘GraphicalModels’ for
Macaulay2 [37]. The commands
needsPackage "GraphicalModels";
G = mixedGraph(digraph {{1,{2,3}},{2,{3}},{3,{4}}}, bigraph {{2,4}});
R = gaussianRing G;
trekIdeal(R,G)
return the zero ideal. Issuing the command
gaussianVanishingIdeal R
shows that I(G) is generated by
2
(12.1) fVerma = σ11 σ13 σ22 σ34 − σ11 σ13 σ23 σ24 − σ11 σ14 σ22 σ33 + σ11 σ14 σ23
2
2
2
− σ12
σ13 σ34 + σ12
σ14 σ33 + σ12 σ13
σ24 − σ12 σ13 σ14 σ23 .
Clearly, fVerma is not a determinant. Its vanishing can be explained as follows.
Decompose G into its mixed components. The largest component G[{2, 4}] is depicted in Figure 12.1. Appealing to Theorem 6.1, there is a rational function
τ{2,4} : MG → MG′ , so τ{2,4} is the covariance matrix for G[{2, 4}]. In G[{2, 4}],
STRUCTURAL EQUATION MODELS
1
35
3
2
4
Figure 12.1. The largest mixed component of the Verma graph.
1
3
2
5
4
Figure 12.2. A graph that can be decomposed only after removing the sink node 5.
there is no trek from node 1 to node 4. Hence, the (1, 4) entry of τ{2,4} (Σ) is zero.
Clearing the denominator yields fVerma(Σ) = 0 for Σ ∈ MG .
The example suggests that new relations can be discovered by decomposing the
graph and studying d-separation or trek-separation relations in the mixed components. The matter is more delicate, however.
Example 12.2. In order to prevent decomposition of the Verma graph, add a
fifth node and bidirected edges such that (V, B) becomes connected. Specifically,
consider the graph G from Figure 12.2. The new graph G is simple and acyclic
and, thus, has expected dimension 13. Because the nodes 2 and 5 are d-separated
by 1, we have σ12 σ15 − σ11 σ25 ∈ I(G). Other relations must exist, and indeed a
Gröbner basis computation shows that
∞
I(G) = hσ12 σ15 − σ11 σ25 , fVermai : σ11
with fVerma being the polynomial from (12.1). The fact that fVerma ∈ I(G) is
explained by Theorem 5.1. Since no directed edge of G has a tail at node 5, the
theorem allows us to consider the subgraph induced by the set of nodes {1, 2, 3, 4}.
We are back in Example 12.1 and decomposition yields fVerma as a relation.
It is clear that more contrived examples can be constructed in which d-/trekseparation applies only after several rounds of alternating between graph decomposition and forming a subgraph induced by an ancestral set. An overview of what
is known about such a recursive approach can be found in [55], where the focus is
on non-linear models and manipulation of probability density functions.
13. Conclusion
Linear structural equation models have covariance matrices with rich algebraic
structure. As we showed in Section 3, statistical considerations motivate a host of
different algebraic problems. In this review, we focused on methods for parameter
identification as well as relations among covariances. While much progress has
been made, and continues to be made [30], we have encountered a plethora of open
problems of algebraic and combinatorial nature.
36
MATHIAS DRTON
In our review, we focused exclusively on linear and Gaussian models. As noted
repeatedly, many of the questions have interesting generalizations to non-linear or
non-Gaussian models. In particular, in settings with discrete random variables, as
considered in [27, 28], algebra and Gröbner basis techniques continue to be useful
[21]. Similarly, many additional challenges arise in models with explicit latent
variables, which motivate studying the map φG when projected onto a principal
submatrix (recall Section 5).
References
1. R. Ayesha Ali, Thomas S. Richardson, and Peter Spirtes, Markov equivalence for ancestral
graphs, Ann. Statist. 37 (2009), no. 5B, 2808–2837.
2. Theodore W. Anderson, An introduction to multivariate statistical analysis, third ed., Wiley
Series in Probability and Statistics, Wiley-Interscience [John Wiley & Sons], Hoboken, NJ,
2003. MR 1990662
3. Vladimir I. Arnol′ d, Sabir M. Guseı̆n-Zade, and Aleksandr N. Varchenko, Singularities of
differentiable maps. Vol. II, Monographs in Mathematics, vol. 83, Birkhäuser Boston, Inc.,
Boston, MA, 1988, Monodromy and asymptotics of integrals, Translated from the Russian by
Hugh Porteous, Translation revised by the authors and James Montaldi. MR 966191
4. Rina Foygel Barber, Mathias Drton, and Luca Weihs, Semid: Identifiability of linear structural equation models, 2015, R package version 0.2.
5. Kenneth A. Bollen, Structural equations with latent variables, Wiley Series in Probability and
Mathematical Statistics: Applied Probability and Statistics, John Wiley & Sons Inc., New
York, 1989, A Wiley-Interscience Publication. MR 996025 (90k:62001)
6. Kenneth A. Bollen and Kwok-fai Ting, A tetrad test for causal indicators, Psychological
methods 5 (2000), no. 1, 3.
7. Carlos Brito and Judea Pearl, A new identification condition for recursive models with correlated errors, Struct. Equ. Model. 9 (2002), no. 4, 459–474. MR 1930449
8. Andries E. Brouwer and Jan Draisma, Equivariant Gröbner bases and the Gaussian two-factor
model, Math. Comp. 80 (2011), no. 274, 1123–1133. MR 2772115
9. Søren L. Buhl, On the existence of maximum likelihood estimators for graphical Gaussian
models, Scand. J. Statist. 20 (1993), no. 3, 263–270.
10. Bryant Chen, Decomposition and identification of linear structural equation models, ArXiv
e-prints (2015), 1508.01834.
11. Bryant Chen, Jin Tian, and Judea Pearl, Testable implications of linear structural equations
models, Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence (Palo
Alto, CA) (Carla E. Brodley and Peter Stone, eds.), AAAI Press, 2014, pp. 2424–2430.
12. David Cox, John Little, and Donal O’Shea, Ideals, varieties, and algorithms, third ed., Undergraduate Texts in Mathematics, Springer, New York, 2007, An introduction to computational
algebraic geometry and commutative algebra. MR 2290010 (2007h:13036)
13. Wolfram Decker, Gert-Martin Greuel, Gerhard Pfister, and Hans Schönemann,
Singular 4-0-2 — A computer algebra system for polynomial computations,
http://www.singular.uni-kl.de, 2015.
14. Jan Draisma, Seth Sullivant, and Kelli Talaska, Positivity for Gaussian graphical models,
Adv. in Appl. Math. 50 (2013), no. 5, 661–674. MR 3044565
15. Mathias Drton, Likelihood ratio tests and singularities, Ann. Statist. 37 (2009), no. 2, 979–
1012. MR 2502658
16. Mathias Drton, Rina Foygel, and Seth Sullivant, Global identifiability of linear structural
equation models, Ann. Statist. 39 (2011), no. 2, 865–886.
17. Mathias Drton, Hélène Massam, and Ingram Olkin, Moments of minors of Wishart matrices,
Ann. Statist. 36 (2008), no. 5, 2261–2283. MR 2458187
18. Mathias Drton and Martyn Plummer, A Bayesian information criterion for singular models,
J. R. Stat. Soc. Ser. B. Stat. Methodol. (2016), to appear.
19. Mathias Drton and Thomas S Richardson, Multimodality of the likelihood in the bivariate
seemingly unrelated regressions model, Biometrika 91 (2004), no. 2, 383–392.
20. Mathias Drton, Bernd Sturmfels, and Seth Sullivant, Algebraic factor analysis: tetrads, pentads and beyond, Probab. Theory Related Fields 138 (2007), no. 3-4, 463–493. MR 2299716
STRUCTURAL EQUATION MODELS
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
37
, Lectures on algebraic statistics, Oberwolfach Seminars, vol. 39, Birkhäuser Verlag,
Basel, 2009. MR 2723140 (2012d:62004)
Mathias Drton and Luca Weihs, Generic identifiability of linear structural equation models
by ancestor decomposition, ArXiv e-prints (2015), 1504.02992.
Mathias Drton and Han Xiao, Finiteness of small factor analysis models, Ann. Inst. Statist.
Math. 62 (2010), no. 4, 775–783. MR 2652316
, Wald tests of singular hypotheses, Bernoulli 22 (2016), no. 1, 38–59. MR 3449776
Mathias Drton and Josephine Yu, On a parametrization of positive semidefinite matrices with
zeros, SIAM J. Matrix Anal. Appl. 31 (2010), no. 5, 2665–2680. MR 2740626 (2011j:15057)
Jan Ernest, Dominik Rothenhäusler, and Peter Bühlmann, Causal inference in partially linear
structural equation models: identifiability and estimation, ArXiv e-prints (2016), 1607.05980.
Robin J. Evans, Margins of discrete Bayesian networks, ArXiv e-prints (2015), 1501.02103.
Robin J. Evans and Thomas S. Richardson, Smooth, identifiable supermodels of discrete DAG
models with latent variables, ArXiv e-prints (2015), 1511.06813.
William N. Evans and Jeanne S. Ringel, Can higher cigarette taxes improve birth outcomes?,
Journal of Public Economics 72 (1999), no. 1, 135–154.
Alex Fink, Jenna Rajchgot, and Seth Sullivant, Matrix Schubert varieties and Gaussian conditional independence models, ArXiv e-prints (2015), 1510.04124.
Franklin M. Fisher, A correspondence principle for simultaneous equation models, Econometrica 38 (1970), 73–92.
Christopher Fox, Interpretation and inference of linear structural equation models, Ph.D.
thesis, University of Chicago, 2014.
Christopher J. Fox, Andreas Käufl, and Mathias Drton, On the causal interpretation of
acyclic mixed graphs under multivariate normality, Linear Algebra Appl. 473 (2015), 93–
113. MR 3338327
Rina Foygel, Jan Draisma, and Mathias Drton, Half-trek criterion for generic identifiability
of linear structural equation models, Ann. Statist. 40 (2012), no. 3, 1682–1713.
Morten Frydenberg, The chain graph Markov property, Scand. J. Statist. 17 (1990), no. 4,
333–353. MR 1096723
Luis D. Garcia-Puente, Sarah Spielvogel, and Seth Sullivant, Identifying causal effects with
computer algebra, Proceedings of the 26th Conference on Uncertainty in Artificial Intelligence
(UAI) (Peter Grünwald and Peter Spirtes, eds.), AUAI Press, 2010.
Luis David Garcı́a-Puente, Sonja Petrović, and Seth Sullivant, Graphical models, J. Softw.
Algebra Geom. 5 (2013), 1–7. MR 3073716
Dan Geiger, David Heckerman, Henry King, and Christopher Meek, Stratified exponential
families: graphical models and model selection, Ann. Statist. 29 (2001), no. 2, 505–529.
MR 1863967
Dan Geiger, Thomas Verma, and Judea Pearl, Identifying independence in Bayesian networks,
Networks 20 (1990), no. 5, 507–534, Special issue on influence diagrams. MR 1064736
Daniel R. Grayson and Michael E. Stillman, Macaulay2, a software system for research in
algebraic geometry, Available at http://www.math.uiuc.edu/Macaulay2/.
Frank Harary, The determinant of the adjacency matrix of a graph, SIAM Rev. 4 (1962),
202–210. MR 0144330
Dominique M. A. Haughton, On the choice of a model to fit data from an exponential family,
Ann. Statist. 16 (1988), no. 1, 342–355. MR 924875
Alain Hauser and Peter Bühlmann, Jointly interventional and observational data: estimation
of interventional Markov equivalence classes of directed acyclic graphs, J. R. Stat. Soc. Ser.
B. Stat. Methodol. 77 (2015), no. 1, 291–318. MR 3299409
Jan T. A. Koster, Marginalizing and conditioning in graphical models, Bernoulli 8 (2002),
no. 6, 817–840.
Gustavo Lacerda, Peter Spirtes, Joseph Ramsey, and Patrik Hoyer, Discovering cyclic causal
models by independent components analysis, Proceedings of the Twenty-Fourth Conference
Annual Conference on Uncertainty in Artificial Intelligence (UAI-08) (Corvallis, Oregon),
AUAI Press, 2008, pp. 366–374.
Steffen L. Lauritzen, Graphical models, Oxford University Press, 1996.
Dennis Leung, Mathias Drton, and Hisayuki Hara, Identifiability of directed Gaussian graphical models with one latent source, Electron. J. Stat. 10 (2016), no. 1, 394–422. MR 3466188
38
MATHIAS DRTON
48. Radim Lněnička and František Matúš, On Gaussian conditional independent structures, Kybernetika (Prague) 43 (2007), no. 3, 327–342. MR 2362722
49. Judea Pearl, Causality, second ed., Cambridge University Press, Cambridge, 2009, Models,
reasoning, and inference. MR 2548166 (2010i:68148)
50. Thomas Richardson and Peter Spirtes, Ancestral graph Markov models, Ann. Statist. 30
(2002), no. 4, 962–1030. MR 1926166 (2003h:60017)
51. Thomas S. Richardson, A characterization of Markov equivalence for directed cyclic graphs,
International Journal of Approximate Reasoning 17 (1997), no. 2/3, 107–162.
52. Hajir Roozbehani and Yury Polyanskiy, Algebraic methods of classifying directed graphical
models, 2014 IEEE International Symposium on Information Theory, June 2014, pp. 2027–
2031.
53. Nathaniel Shiers, Piotr Zwiernik, John A. D. Aston, and Jim Q. Smith, The correlation
space of Gaussian latent tree models and model selection without fitting, Biometrika (2016),
forthcoming.
54. Shohei Shimizu, LiNGAM: Non-Gaussian methods for estimating causal structures, Behaviormetrika 41 (2014), no. 1, 65–98.
55. Ilya Shpitser, Robin Evans, Thomas Richardson, and James Robins, Introduction to nested
Markov models, Behaviormetrika 41 (2014), no. 1, 3–39.
56. Ilya Shpitser and Judea Pearl, Identification of joint interventional distributions in recursive semi-Markovian causal models, Proceedings of the Twenty-First National Conference on
Artificial Intelligence (Menlo Park, CA), AAAI Press, 2006, pp. 1219–1226.
57. Peter Spirtes, Clark Glymour, and Richard Scheines, Causation, prediction, and search, second ed., Adaptive Computation and Machine Learning, MIT Press, Cambridge, MA, 2000,
With additional material by David Heckerman, Christopher Meek, Gregory F. Cooper and
Thomas Richardson, A Bradford Book. MR 1815675 (2001j:62009)
58. Peter Spirtes, Thomas Richardson, Chris Meek, Richard Scheines, and Clark Glymour, Using
path diagrams as a structural equation modelling tool, Sociological Methods and Research 27
(1998), 182–225.
59. Paul D. Stolley, When genius errs: R. A. Fisher and the lung cancer controversy, Am. J.
Epidemiol. 133 (1991), no. 5, 416–425.
60. Milan Studený, Probabilistic conditional independence structures, Information Science and
Statistics, Springer, London, 2005. MR 3183760
61. Seth Sullivant, Algebraic geometry of Gaussian Bayesian networks, Adv. in Appl. Math. 40
(2008), no. 4, 482–513. MR 2412156
, A Gröbner basis for the secant ideal of the second hypersimplex, J. Commut. Algebra
62.
1 (2009), no. 2, 327–338. MR 2504939
63. Seth Sullivant and Elizabeth Gross, The maximum likelihood threshold of a graph,
arXiv:1404.6989, 2014.
64. Seth Sullivant, Kelli Talaska, and Jan Draisma, Trek separation for Gaussian graphical models,
Ann. Statist. 38 (2010), no. 3, 1665–1685. MR 2662356 (2011f:62076)
65. Jin Tian, Identifying direct causal effects in linear models, Proceedings of the Twentieth
National Conference on Artificial Intelligence (AAAI), 2005, pp. 346–353.
66. Jin Tian and Judea Pearl, A general identification condition for causal effects, AAAI/IAAI,
2002, pp. 567–573.
67. Caroline Uhler, Geometry of maximum likelihood estimation in Gaussian graphical models,
Ann. Statist. 40 (2012), no. 1, 238–261. MR 3014306
68. Lieven Vandenberghe and Martin S. Andersen, Chordal graphs and semidefinite optimization,
Foundations and Trends in Optimization 1 (2014), no. 4, 241–433.
69. Thomas S. Verma and Judea Pearl, Equivalence and synthesis of causal models, Uncertainty
in Artificial Intelligence 6, Elsevier, 1991, UCLA Cognitive Systems Laboratory, Technical
Report (R-150), pp. 255–268.
70. Sumio Watanabe, Algebraic geometry and statistical learning theory, Cambridge Monographs
on Applied and Computational Mathematics, vol. 25, Cambridge University Press, Cambridge,
2009. MR 2554932
71. Nanny Wermuth, Probability distributions with summary graph structure, Bernoulli 17 (2011),
no. 3, 845–879. MR 2817608
72. Sewall Wright, Correlation and causation, J. Agricultural Research 20 (1921), 557–585.
73.
, The method of path coefficients, Ann. Math. Statist. 5 (1934), no. 3, 161–215.
STRUCTURAL EQUATION MODELS
39
74. Hui Zhao, Zhongguo Zheng, and Baijun Liu, On the Markov equivalence of maximal ancestral
graphs, Sci. China Ser. A 48 (2005), no. 4, 548–562. MR 2157690 (2006e:62069)
Department of Statistics, University of Washington
| 10math.ST
|
arXiv:1610.06864v3 [math.GR] 30 Jan 2018
A NOTE ON THE ACYLINDRICAL HYPERBOLICITY OF
GROUPS ACTING ON CAT(0) CUBE COMPLEXES
INDIRA CHATTERJI AND ALEXANDRE MARTIN
Abstract. We study the acylindrical hyperbolicity of groups acting
by isometries on CAT(0) cube complexes, and obtain simple criteria
formulated in terms of stabilisers for the action. Namely, we show that
a group acting essentially and non-elementarily on a finite dimensional
irreducible CAT(0) cube complex is acylindrically hyperbolic if there
exist two hyperplanes whose stabilisers intersect along a finite subgroup.
We also give further conditions on the geometry of the complex so that
the result holds if we only require the existence of a single pair of points
whose stabilisers intersect along a finite subgroup.
1. Introduction
A group is called acylindrically hyperbolic if it is not virtually cyclic and
admits an acylindrical action with unbounded orbits on a hyperbolic geodesic metric space. Acylindrically hyperbolic groups form a large class
of groups, introduced by Osin [Osi16], displaying strong hyperbolic-like
features: They encompass mapping class groups of hyperbolic surfaces
[Bow08], relatively hyperbolic groups [Osi06], the plane Cremona group
[CL13, Lon16], and many more. The notion of acylindrical hyperbolicity
has gathered a lot of interest in recent years due to the strong algebraic
and analytical consequences it implies for the group (we refer to [Osi16] and
details therein).
In this article, we study the acylindrical hyperbolicity of groups acting
by isometries on CAT(0) cube complexes. Our goal is to obtain simple
acylindrical hyperbolicity criteria for such groups. Recall that an action on
a CAT(0) cube X complex is called essential if no orbit remains at bounded
distance from a half-space of X, and that it is called non-elementary if it
does not admit a finite orbit in X ∪ ∂∞ X.
Our first criterion is formulated in terms of hyperplanes stabilisers, and
generalises to finite dimensional CAT(0) cube complexes a criterion due to
Minasyan–Osin for actions on simplicial trees [MO15]:
Theorem 1.1. Let G be a group acting essentially and non-elementarily
on an irreducible finite-dimensional CAT(0) cube complex. If there exist
two hyperplanes whose stabilisers intersect along a finite subgroup, then G
is acylindrically hyperbolic.
1
2
INDIRA CHATTERJI AND ALEXANDRE MARTIN
In the previous theorem, we do not require the two hyperplanes to be
disjoint, or even distinct. In particular, the conclusion holds if there exists
a hyperplane of X whose stabiliser is weakly malnormal, i.e. it intersects
some conjugate along a finite subgroup.
It should be noted that the above theorem does not reduce to the aforementioned criterion of Minasyan–Osin, as groups acting on CAT(0) cube
complexes do not virtually act on a simplicial tree a priori.
Anthony Genevois has also obtained criteria for acylindrical hyperbolicity
of a similar flavour, using different tools. In particular, his approach can be
used to recover Theorem 1.1, see [Gen16b, Remark 21].
We give an application of Theorem 1.1 to Artin groups of FC type, suggested to us by Ruth Charney. Artin groups span a large range of groups,
and include for instance free groups, braid groups, free abelian groups, as
well as many more exotic groups. Artin groups and their subgroups are a
rich source of examples and counterexamples of interesting phenomena in
geometry and group theory.
Theorem 1.2. Non-virtually cyclic Artin groups of FC type whose underlying Coxeter graphs have diameter at least 3 are acylindrically hyperbolic.
Artin groups of FC type and the proof of the above theorem will be
discussed in Section 5. In the case of Artin groups of FC type whose Coxeter
graphs have diameter 1, i.e. Artin groups of finite type, Calvez and Wiest
[CW16] showed, using different techniques, that the quotient of such groups
by their centre is acylindrically hyperbolic.
As a corollary of Theorem 1.1, we obtain the following results for actions
satisfying a weak notion of acylindricity or properness:
Corollary 1.3. Let G be a group acting essentially and non-elementarily on
an irreducible finite-dimensional CAT(0) cube complex X. Assume that the
action is non-uniformly weakly acylindrical, that is, there exists a constant
L ≥ 0 such that only finitely many elements of G fix two points of X at
distance at least L. Then G is acylindrically hyperbolic.
Corollary 1.4. A group acting essentially, non-elementarily, and with finite
vertex stabilisers on a finite dimensional irreducible CAT(0) cube complex
is acylindrically hyperbolic.
Corollary 1.3 was already known if the complex is in addition assumed
to be hyperbolic (see [Mar15] for the proof in dimension 2 and [Gen16a]
for the general case). Corollary 1.4 was already known if in addition the
action is assumed to be metrically proper, by work of Caprace–Sageev on the
existence of rank one isometries of CAT(0) cube complexes [CS11]. Anthony
Genevois informed us that he found independently similar results, using
different techniques [Gen16b].
In Theorem 1.1 and its corollaries, the essentiality assumption could
be weakened to the assumption that the essential core is irreducible, a
condition that is a priori non-trivial to check. Notice however that the
ACYLINDRICAL HYPERBOLICITY AND CAT(0) CUBE COMPLEXES
3
other assumptions cannot be removed. The non-acylindrically hyperbolic
group Z acts properly and essentially, but elementarily, on the real line
with its standard simplicial structure. The groups of Burger–Mozes provide
examples of cocompact lattices in the product of two trees whose associated
actions are essential, non-elementary, and without fixed point at infinity, yet
these groups are not acylindrically hyperbolic as they are simple [BM00].
Thompson’s groups V and T act properly, non-elementarily, and without
fixed point at infinity on an infinite dimensional irreducible CAT(0) cube
complex [Far08], but are not acylindrically hyperbolic as they are simple.
We also obtain stronger criteria, that allow us to deduce the acylindrical
hyperbolicity of a group from information on the stabilisers of a single pair
of points (see Theorem 4.2 for the general statement). For this, we impose
further conditions on the complex: We say that a CAT(0) cube complex is
cocompact if its automorphism group acts cocompactly on it, and we say
that it has no free face if every non-maximal cube is contained in at least
two maximal cubes. We prove the following:
Theorem 1.5. Let G be a group acting essentially and non-elementarily on
an irreducible finite-dimensional cocompact CAT(0) cube complex with no
free face. If there exist two points whose stabilisers intersect along a finite
subgroup, then G is acylindrically hyperbolic.
To show the acylindrical hyperbolicity of a group G through its action
on a geodesic metric space X, a useful criterion introduced by BestvinaBromberg-Fujiwara [BBF14, Theorem H] is to find an infinite order element
of the group whose orbits are strongly contracting and which satisfies the
so-called WPD condition. However, this condition, formulated in terms of
coarse stabilisers of pairs of points, is generally cumbersome to check for
actions on non-locally finite spaces. In [Mar16a], the second author introduced a different criterion involving a weakening of this condition formulated purely in terms of stabilisers of pairs of points, making it much more
tractable (see Theorem 2.3 for the exact formulation). The price to pay
is to find group elements satisfying a strengthened notion of contraction of
their orbits, called über-contractions. The second author showed that such
contractions abound for actions on non-locally compact spaces under mild
assumptions on the stabilisers of vertices, and used it to show the acylindrical hyperbolicity of the tame automorphism group of an affine quadric
threefold, a subgroup of the Cremona group Bir(P3 (C)). In this article, we
provide a different way to construct über-contractions for groups acting on
CAT(0) cube complexes, this time using the very rich combinatorial geometry of their hyperplanes. This construction relies heavily on the existence
of hyperplanes with strong separation properties, called über-separated hyperplanes, and introduced by Chatterji–Fernós–Iozzi [CFI16]. Such hyperplanes were used to prove a superrigidity phenomenon for groups acting
non-elementarily on CAT(0) cube complexes.
4
INDIRA CHATTERJI AND ALEXANDRE MARTIN
Acknowledgement. The authors warmly thank Talia Fernós for producing the example in Remark 4.4, Michah Sageev for discussions related to
Proposition 4.3, Carolyn Abbott for noticing Theorem 1.2 and Ruth Charney for pointing it out to us, along with the explanations of Section 5. We
also thank Anthony Genevois for remarks on an early version of this article,
Ruth Charney and Rose Morris-Wright for spotting a gap in the proof of
Theorem 1.2, and the anonymous referee for their interesting comments.
The first author is partially supported by the Institut Universitaire de
France (IUF) and the ANR GAMME. The second author is partially supported by the Austrian Science Fund (FWF) grant M1810-N25. This work
was completed while the two authors were in residence at MSRI during the
Fall 2016 “Geometric Group Theory” research program, NSF Grant DMS1440140. They thank the organisers for the opportunity to work in such a
stimulating environment.
2. Über-contractions and acylindrical hyperbolicity
Recall that, given a group G acting on a geodesic metric space X,
the action is called acylindrical if for every r ≥ 0 there exist constants
L(r), N (r) ≥ 0 such that for every points x, y of X at distance at least L(r),
|{g ∈ G|d(x, gx) ≤ r, d(y, gy) ≤ r}| ≤ N (r).
A group is acylindrically hyperbolic if it is not virtually cyclic and admits an
acylindrical action with unbounded orbits on a hyperbolic geodesic metric
space. Given a group action on an arbitrary geodesic space, the following criterion was introduced by Bestvina–Bromberg–Fujiwara to prove the
acylindrical hyperbolicity of the group:
Theorem 2.1 ([BBF14, Theorem H]). Let G be a group acting by isometries
on a geodesic metric space X. Let g be a group element of infinite order with
quasi-isometrically embedded orbits, and assume that the following holds:
• the group element g is strongly contracting, that is, there exists a
point x of X such that the closest-point projections on the hgi-orbit
of x of the balls of X that are disjoint from hgix have uniformly
bounded diameter,
• the group element g satisfies the WPD condition, that is, for every
r ≥ 0 and every point x of X, there exists an integer m such that only
finitely many elements h of G satisfy d(x, hx), d(g m x, hg m x) ≤ r.
Then G is either virtually cyclic or acylindrically hyperbolic.
The following notion, introduced in [Mar16a], provides an easier way to
prove the acylindrical hyperbolicity of a group.
Definition 2.2. Let X be a geodesic metric space and let h be an isometry
of X with quasi-isometrically embedded orbits. A system of checkpoints for
h is the data of a finite subset S of X, the checkpoint, as well as an error
ACYLINDRICAL HYPERBOLICITY AND CAT(0) CUBE COMPLEXES
5
S
constant L ≥ 0 and a quasi-isometry f : Λ := i∈Z hi S → R such that we
have the following:
Let x, y be points of X and let x′ , y ′ be closest-point projections on Λ of
x, y respectively. For every checkpoint Si := hi S, i ∈ Z, such that:
• Si coarsely separates x′ and y ′ , i.e. f (x′ ) and f (y ′ ) lie in different
unbounded connected components of R \ f (Si ),
• Si is at distance at least L from both x′ and y ′ ,
then every geodesic between x and y meets Si .
A hyperbolic isometry h of X is über-contracting, or is an über-contraction,
if it admits such a system of checkpoints.
We will be using the following criterion for acylindrical hyperbolicity:
Theorem 2.3 (Theorem 1.2 of [Mar16a]). Let G be a group acting by isometries on a geodesic metric space X. Let g ∈ G be an infinite order element
such that the following holds:
(i) the group element g is über-contracting with respect to a system of
checkpoints (gi S)i∈Z,
(ii) There exists a constant m0 such that for every point s ∈ S and every
m ≥ m0 , only finitely many elements of G fix pointwise s and g m s.
Then G is either virtually cyclic or acylindrically hyperbolic.
Note that condition (ii) of the previous theorem is a considerable weakening of the WPD condition for g.
This weaker condition has the advantage of involving only stabilisers of
pairs of points, which makes it much easier to use.
3. Über-separated hyperplanes and the proof of Theorem 1.1
We now recall a few basic facts concerning CAT(0) cube complexes, and
more precisely the notions of bridges and über-separated pairs, which are
crucial to prove Theorems 1.1 and 1.5. The missing details and proofs can
be found in [CFI16].
By a slight abuse of notation, we will identify a CAT(0) cube complex
with its vertex set endowed with the graph metric coming from its 1-skeleton.
Each hyperplane separates the vertex set into two disjoint components,
which we refer to as halfspaces.
The following notion was introduced by Behrstock–Charney in [BC12]:
Definition 3.1. Two parallel hyperplanes of a CAT(0) cube complex are
said to be strongly separated if no hyperplane is transverse to both. By the
usual abuse of terminology, we say that two halfspaces are strongly separated
if the corresponding hyperplanes are.
We will need tools introduced by Caprace–Sageev [CS11]. Recall that a
family of n pairwise crossing hyperplanes divides a CAT(0) cube complex
into 2n regions called sectors.
6
INDIRA CHATTERJI AND ALEXANDRE MARTIN
Lemma 3.2 (Caprace–Sageev [CS11, Proposition 5.1, Double-Skewering
Lemma, Lemma 5.2]). Let X be a finite dimensional irreducible CAT(0)
cube complex and let G → Aut(X) be a group acting essentially and without
fixed points in X ∪ ∂∞ X.
• (Strong Separation Lemma) Let h1 be a halfspace of X. Then there
exists a halfspace h2 such that h1 ⊂ h2 and such that h1 and h2 are
strongly separated.
• (Double-Skewering Lemma) Let h1 ⊂ h2 be two nested halfspaces.
Then there exists an element g ∈ G that double-skewers h1 and h2 ,
that is, such that h1 ⊂ h2 ⊂ gh1 .
• (Sector Lemma) Let ĥ1 , ĥ2 be two transverse hyperplanes. Then we
can choose two disjoint hyperplanes ĥ3 and ĥ4 that are contained in
opposite sectors determined by ĥ1 and ĥ2 .
We will need a finer notion of strong separation of halfspaces, which is
less standard but will be key to our work.
Definition 3.3. Two strongly separated halfspaces h1 and h2 are said to be
an über-separated pair if for any two halfspaces k1 , k2 with the property that
hi and ki are transverse for i = 1, 2, then k1 and k2 are parallel. We say that
two strongly separated hyperplanes are über-separated if their halfspaces are.
Note that pairs of über-separated hyperplanes correspond exactly to pairs
of hyperplanes at distance at least 4 in the intersection graph.
Remark 3.4. If h ⊂ k ⊂ ℓ are pairwise strongly separated halfspaces, then
h and ℓ are über-separated.
Notice that über-separated pairs are in particular strongly separated and
hence they do not exist in the reducible case by [CS11, Proposition 5.1].
The existence of über-separated hyperplanes is a consequence of the DoubleSkewering Lemma 3.2. We will need the following lemma, which is a direct
consequence of the Double-Skewering Lemma 3.2 and [CFI16, Lemma 2.14]:
Lemma 3.5. Let X be a finite dimensional irreducible CAT(0) cube complex
and G → Aut(X) a group acting essentially and non-elementarily. Given
any two parallel hyperplanes ĥ1 and ĥ2 , there exists g ∈ G that doubleskewers ĥ1 and ĥ2 and such that ĥ1 and g ĥ1 are über-separated.
For two points x, y of X, we denote by I(x, y) the interval between x
and y, that is, the union of all the geodesics between x and y. Recall that
intervals are finite, as they only depend on the (finite) set of hyperplanes
separating the given pair of points.
Definition 3.6. Let h1 ⊂ h2 be a nested pair of halfspaces. The (combinatorial) bridge between ĥ1 and ĥ2 , denoted b(ĥ1 , ĥ2 ), is the union of all
the geodesics between points x1 ∈ h1 and x2 ∈ h∗2 minimizing the distance
between h1 and h∗2 .
ACYLINDRICAL HYPERBOLICITY AND CAT(0) CUBE COMPLEXES
7
Lemma 3.7 (Lemma 2.18 and 2.24 [CFI16]). If h1 ⊂ h2 be a pair of nested
halfspaces, strongly separated, there exists a unique pair of points of h1 × h∗2 ,
called the gates of the bridge, that minimizes the distance between h1 and
h∗2 , i.e. there exist points x1 ∈ h1 and x2 ∈ h∗2 such that
b(ĥ1 , ĥ2 ) = I(x1 , x2 ).
In particular, the bridge between two strongly separated hyperplanes is finite.
Moreover, the following holds true for any y1 ∈ h1 and y2 ∈ h∗2 .
d(y1 , y2 ) = d(y1 , x1 ) + d(x1 , x2 ) + d(x2 , y2 ).
The following is a very important feature of über-separated pairs.
Lemma 3.8 (Proof of Lemma 3.5 of [CFI16]). Let h1 ⊂ h2 be an überseparated pair of halfspaces, x ∈ h1 and y ∈ h∗2 . Then every geodesic between
x and y meets the bridge b(ĥ1 , ĥ2 ).
The following lemma explains the relationship between über-separated
pairs and über-contractions.
Lemma 3.9. Let h1 ⊂ h2 be an über-separated pair of halfspaces, and g ∈ G
an element that double-skewers h1 and h2 . Then g is an über-contraction.
Proof. Let B denote the bridge between h1 and gh1 . We will show that the
collection of the g-translates of B forms a system of checkpoints. Notice
that since the bridge is an interval, it is finite even though the complex is
not assumed to S
be locally finite, hence all the checkpoints Sn := gn B are
finite. Set Λ := n∈Z Sn . Let Y := h∗1 ∩ gh1 and Yn := g n Y for every n ∈ Z.
For every point x ∈ Yn , we have that its closest-point projections on Λ
are in Sn−1 ∪ Sn ∪ Sn+1 : Indeed, since h1 and gh1 are über-separated, it
follows from Lemma 3.8 that a geodesic between x and a point x′ ∈ Yn′ with
n′ ≥ n+2 (n′ ≤ n−2 respectively) meets the bridge Sn+1 (Sn−1 respectively).
Moreover, for every x ∈ X which projects on Λ to a point of Sn , we have
that x ∈ Yn−1 ∪ Yn ∪ Yn+1 for the same reasons. Thus, for x, y ∈ X which
project on Λ to points x′ ∈ Sn and y ′ ∈ Sm respectively with m ≥ n + 4, it
follows that x and y are separated by the hyperplanes gn+2 ĥ1 , . . . , g m−1 ĥ1 ,
hence every geodesic between x and y meets each checkpoint Sn+2 , . . . , Sm−2
by Lemma 3.8.
By a result of Haglund [Hag07, Theorem 1.4], g admits a combinatorial
axis Λg . As Λ and Λg stay at bounded distance from one another, it follows
easily from the discussion of the previous paragraph that Λg is contained
in Λ. Since B is finite, the closest-point projection Λ → Λg yields a quasiisometry Λ → R, and it is now straightforward to check that g is an übercontraction.
Proof of Theorem 1.1. Let G be a group acting essentially and nonelementarily on an irreducible finite-dimensional CAT(0) cube complex X
8
INDIRA CHATTERJI AND ALEXANDRE MARTIN
and assume that there exist two hyperplanes whose stabilisers intersect along
a finite subgroup. By Proposition 2.3, it is enough to construct an element
of G satisfying conditions (i) and (ii) of Proposition 2.3.
Let ĥ1 and ĥ2 be two hyperplanes whose stabilisers intersect along a
finite subgroup. We start by showing that we can assume that ĥ1 and ĥ2
are disjoint. By the Sector Lemma 3.2, we choose two disjoint hyperplanes
ĥ3 and ĥ4 that are contained in opposite sectors determined by ĥ1 and ĥ2 .
Up to applying the Strong Separation Lemma 3.2, we can further assume
that ĥ3 and ĥ4 are strongly separated. Let H := Stab(ĥ3 ) ∩ Stab(ĥ4 ). Then
H stabilises the bridge between ĥ3 and ĥ4 , which is a single interval by
Lemma 3.7 since ĥ3 and ĥ4 are strongly separated. As intervals are finite,
H virtually fixes a geodesic between ĥ3 and ĥ4 . As ĥ3 and ĥ4 are in opposite
sectors determined by ĥ1 and ĥ2 , such a geodesic crosses both ĥ1 and ĥ2 .
Thus, H is virtually contained in Stab(ĥ1 ) ∩ Stab(ĥ2 ), hence H is finite,
which is what we wanted.
Thus, let ĥ1 and ĥ2 be two disjoint hyperplanes whose stabilisers intersect
along a finite subgroup. By Lemma 3.5, we choose an element g of G that
double-skewers ĥ1 and ĥ2 , and such that ĥ1 and gĥ1 are über-separated.
Then according to Lemma 3.9 the element g is an über-contraction, proving
(i). Let x be a point of the bridge between ĥ1 and ĥ2 , and choose a geodesic
γ between x and g 2 x. We have that γ crosses g ĥ1 and gĥ2 . Since intervals in
a finite dimensional CAT(0) cube complex are finite, a subgroup fixing both
x and g 2 x virtually fixes γ pointwise. It follows that Stab(x) ∩ Stab(g2 x) is
virtually contained in a conjugate of Stab(ĥ1 )∩Stab(ĥ2 ), which is finite. This
proves (ii) , and Proposition 2.3 implies that G is either virtually cyclic or
acylindrically hyperbolic. The action being non-elementarily, the virtually
cyclic case is automatically ruled out, which concludes the proof.
Proof of Corollaries 1.3 and 1.4. Since Corollary 1.4 is a direct consequence
of Corollary 1.3, let us assume that the group G acts essentially, nonelementarily, and non-uniformly weakly acylindrically (with a constant L as
in the statement) on the irreducible finite-dimensional CAT(0) cube complex
X. By the Strong Separation Lemma 3.2, choose two disjoint hyperplanes
ĥ1 and ĥ2 . This implies that the combinatorial bridge between ĥ1 and ĥ2
is not reduced to a point. Moreover, the subgroup Stab(ĥ1 ) ∩ Stab(ĥ2 ) is
virtually contained in the pointwise stabiliser of the finite bridge between
ĥ1 and ĥ2 by Lemma 3.7.
By the Double-Skewering Lemma 3.2, choose a group element g that
skewers both ĥ1 and ĥ2 . For n large enough, ĥ1 and gn ĥ1 are über-separated,
and the distance between ĥ1 and gn ĥ1 becomes greater than L by Lemma
3.7. In particular, Stab(ĥ1 ) ∩ Stab(g n ĥ1 ) virtually fixes a path of length L,
hence Stab(ĥ1 ) ∩ Stab(gn ĥ1 ) is finite by weak acylindricity. Corollary 1.3
now follows from Theorem 1.1.
ACYLINDRICAL HYPERBOLICITY AND CAT(0) CUBE COMPLEXES
9
Theorem 1.1 allows for a very simple geometric proof of the acylindrical
hyperbolicity of certain groups:
Example 3.10. The Higman group on n ≥ 4 generators, defined by the
following presentation:
2
Hn := hai , i ∈ Z/nZ | ai ai+1 a−1
i = ai+1 , i ∈ Z/nZi,
was proved to be acylindrically hyperbolic by Minasyan–Osin [MO15], by
means of its action on the Bass–Serre tree associated to some splitting.
The Higman also acts cocompactly, essentially, and non-elementarily on
a CAT(0) square complex associated to its standard presentation. That
CAT(0) square complex is irreducible, as links of vertices are easily shown
not to be complete bipartite graphs (see [Mar16b, Corollary 4.6]). Since
square stabilisers are trivial by construction, the stabilisers of two crossing hyperplanes intersect along a finite (actually, trivial) subgroup. Thus,
Theorem 1.1 applies.
4. Proof of Theorem 1.5
Definition 4.1. Let X be CAT(0) cube complex and C, C ′ two cubes of
X. We say that C and C ′ separate a pair of hyperplanes if there exist two
hyperplanes ĥ, ĥ′ of X such that each hyperplane defined by an edge of
C ∪ C ′ separates ĥ and ĥ′ .
Theorem 1.5 will be a consequence of the following more general result:
Theorem 4.2. Let G be a group acting essentially and non-elementarily on
a finite dimensional irreducible CAT(0) cube complex. Assume that there
exist two maximal cubes C, C ′ of X whose stabilisers intersect along a finite
subgroup, and such that C and C ′ separate a pair of hyperplanes. Then G
is acylindrically hyperbolic.
Proof. Let HC,C ′ be the set of hyperplanes defined by the edges of C ∪
T
C ′ . Since C and C ′ are maximal cubes, we have that k̂∈H ′ Stab(k̂) is
C,C
T
contained in Stab(C) ∩ Stab(C ′ ), and it follows that k̂∈H ′ Stab(k̂) is
C,C
finite.
By assumption, choose two disjoint halfspaces h, h′ separated by each
hyperplane of HC,C ′ . Up to applying the Strong Separation Lemma 3.2,
we can further assume that h and h′ are strongly separated. In particular,
since the bridge between h and h′ is finite by Lemma 3.7, it follows that
Stab(ĥ) ∩ Stab(ĥ′ ) is virtually contained in the pointwise stabiliser of a
geodesic between the two gates of the bridge b(ĥ, ĥ′ ). As such a geodesic
crosses each hyperplane of HC,C ′ by construction of h, h′ , it follows that
T
Stab(ĥ) ∩ Stab(ĥ′ ) is virtually contained in k̂∈H ′ Stab(k̂), which is finite
C,C
by the above argument. Thus, Stab(h) ∩ Stab(h′ ) is finite, and we conclude
with Theorem 1.1.
10
INDIRA CHATTERJI AND ALEXANDRE MARTIN
The following proposition gives a class of examples of CAT(0) cube complexes where each pair of cubes separates a pair of hyperplanes:
Proposition 4.3. Let X be an irreducible CAT(0) cube complex without
free face, and such that Aut(X) acts cocompactly on X. Then each pair of
cubes separates a pair of hyperplanes.
Remark 4.4. The following example, due to Talia Fernós, shows that
the no-free-face assumption is necessary in Proposition 4.3. Take a
3-dimensional cube [0, 1]3 and glue an edge to each of the vertices
(1, 0, 0), (0, 1, 0), (0, 0, 1) and (1, 1, 1) to get a spiked cube. Then glue infinitely many of these spiked cubes in a tree-like way, to obtain a cocompact
CAT(0) cube complex X which is quasi-isometric to a tree. Then none of
the 3-cubes of that complex is separating a pair of hyperplanes. Indeed,
each 3-cube has three hyperplanes defining it, hence it defines eight sectors,
out of which four contain hyperplanes and four are reduced to a single point,
but among the sectors containing hyperplanes, no two are opposite.
Before proving Proposition 4.3, we start by a simple observation:
Lemma 4.5. A CAT(0) cube complex with no free face is geodesically complete, that is, every finite geodesic can be extended to a bi-infinite geodesic.
Proof. Let γ be a finite geodesic defined by a sequence e1 , . . . , en of edges
of X and let v be the terminal vertex of that finite geodesic. We will show
that we can extend it by one edge. Let E denote the set of edges of X,
containing v and such that γ followed by e ∈ E is not a geodesic. Then,
every e ∈ E has to be parallel to one of the edges ei defining γ, that is, every
e ∈ E defines a hyperplane ĥe crossed by γ. Moreover, the map e 7→ ĥe is
injective as two adjacent edges cannot belong to the same hyperplane, and
any two ĥe , ĥe′ intersect. Indeed, for an edge e to belong to E it means
that γ has been traveling on the carrier of the hyperplane defined by some
ei after having crossed ei , and that can be done simultaneously for several
hyperplanes only when they cross each other. Hence E defines a cube in X,
and since there are no free faces there is an edge e containing v and that
does not belong to E, allowing us to extend by one edge the geodesic γ.
We will also need the following strengthening of the Sector Lemma:
Lemma 4.6 (Strong Sector Lemma). Let X be a CAT(0) cube complex with
no free face, and assume that the automorphism group of X acts cocompactly
on X. Then each sector determined by a finite family of pairwise crossing
hyperplanes contains a hyperplane.
Proof. We prove the result by induction on the number n ≥ 2 of pairwise
crossing hyperplanes. For n = 2, the result follows from [Sag14, Proposition
3.3], since a CAT(0) cube complex with no free face is geodesically complete
by Lemma 4.5. Let ĥ1 , . . . , ĥn be a family of pairwise crossing hyperplanes,
ACYLINDRICAL HYPERBOLICITY AND CAT(0) CUBE COMPLEXES
11
and let hi be a halfspace associated
to ĥi for every i. We want to construct
T
a hyperplane contained in i hi .
By Helly’s theorem, the family (ĥi ∩ ĥ1 )i6=1 defines a family of pairwise
crossing hyperplanes of ĥ1 . Note that ĥ1 also satisfies the property of having
no free face, and the action of Stab(ĥ1 ) on ĥ1 is cocompact, as the same holds
for the action of Aut(X) on X. Thus, one can apply the induction hypothesis
T
to find a hyperplane ĥ′ of ĥ1 contained in the sector i6=1 (ĥ1 ∩ hi ). This
defines a hyperplane of X, which we denote ĥ. Since ĥ′ is disjoint from the
ĥ1 ∩ ĥi for i 6= 1, it follows from Helly’s theorem that ĥ is disjoint
from the
T
ĥi for i 6= 1. Thus, let h be the halfspace of ĥ contained in i6=1 hi . Since
ĥ and ĥ1 cross, we can choose by the induction
hypothesis a hyperplane
T
contained in the sector h ∩ h1 , hence in i hi .
Proof of Proposition 4.3. It is enough to prove the proposition when C and
C ′ are maximal. Choose x and x′ vertices of C, C ′ respectively that maximize the distance between vertices of C and C ′ . Let HC , HC ′ be the family
of hyperplanes defined by an edge of C, C ′ respectively. By the Strong
Sector Lemma 4.6, each sector determined by HC or HC ′ contains a hyperplane. Thus, choose a hyperplane ĥ (ĥ′ respectively) in the unique sector
determined by HC (HC ′ respectively) that contains x (x′ respectively). By
construction of x and x′ , we have that x and x′ are in opposite sectors
defined by HC , and in opposite sectors determined by HC ′ . In particular,
every hyperplane of HC,C ′ separates x and x′ , hence separates ĥ and ĥ′ .
Proof of Theorem 1.5. Theorem 1.5 is a direct consequence of Theorem 4.2
and Proposition 4.3.
Theorem 1.5 allows for a very simple geometric proof of the acylindrical
hyperbolicity of certain groups:
Example 4.7. The group of tame automorphisms of an affine quadric threefold, a subgroup of the 3-dimensional Cremona group Bir(P3 (C)), acts cocompactly, essentially and non-elementarily on a hyperbolic CAT(0) cube
complex without free face [BFL14]. The second author showed that there
exist two cubes whose stabilisers intersect along a finite subgroup [Mar16a,
Proof of Theorem 3.1], hence Theorem 1.5 applies.
5. Artin groups of type FC
We now give an application of our results to the class of Artin groups of
FC type, studied by Charney and Davis in [CD95] and that we now describe.
Recall that a Coxeter graph is a finite, simplicial graph Γ with vertex set S
and edges labeled by integers greater or equal to 2. The label of the edge
connecting two vertices s and t is denoted m(s, t), and we set m(s, t) = ∞
if s and t are not connected by an edge. The Artin group associated to a
12
INDIRA CHATTERJI AND ALEXANDRE MARTIN
Coxeter graph Γ is the group given by the presentation:
*
A=
+
S| sts
| {z. .}. |s, t connected by an edge with label m(s, t) .
| {z. .}. = tst
m(s,t)
m(s,t)
Adding the extra relations s2 = 1 for all s ∈ S, we obtain a Coxeter group
W as a quotient of A. We say that A is finite type if the associated Coxeter
group W is finite. It was shown in [vdL83] that, if T ⊆ S, the subgroup
AT generated by T is isomorphic to the Artin group associated to the full
subgraph of Γ spanned by T . Such subgroups are called special subgroups
of A. Following Charney–Davis, we say that an Artin group is of FC type if
every complete subgraph of the Coxeter graph Γ generates a special subgroup
of finite type.
Given an Artin group A, the Deligne complex DA is the cubical complex
defined as follows: Vertices of DA correspond to cosets aAT , where a ∈ A
and T ⊆ S. Note that we allow T = ∅, in which case aAT = {a}. The
1 of D is obtained by putting an edge between cosets of the
1-skeleton DA
A
form aAT and aAT ′ , when T ′ = T ∪ {t′ } for some t′ ∈ S \ T . In particular,
each edge of DA is labeled by an element of S. Finally, DA is obtained by
1 contains a
“filling the cubes”, that is, by glueing a k-cube whenever DA
subgraph isomorphic to the 1-skeleton of a k-cube.
In [CD95] Theorem 4.3.5 Charney and Davis show that the Deligne complex DA of an Artin group A is a CAT(0) cube complex if and only if A is of
FC type. Edges and hyperplanes in this CAT(0) cube complex are labeled
by elements of S. Moreover, each hyperplane is a translate of a hyperplane
ĥs for some s ∈ S, where ĥs denotes the hyperplane defined by the edge
between {1} and A{s} . It is straightforward to check that the stabiliser of
ĥs is the special subgroup Alk(s) of A determined by the link of the vertex
s in Γ.
The dimension of an Artin group is the maximum cardinality of a subset
T ⊆ S such that AT is finite type, and it is equal to the dimension of DA .
In particular, DA is finite dimensional since the graph Γ is finite. The Artin
group A acts by left multiplication on the aforementioned cosets, and hence
acts by isometries on DA . Since Γ is finite, the action is cocompact. The
stabilizer of a vertex aAT of DA is the subgroup aAT a−1 . In particular, the
action is not proper if Γ is non-empty.
The action satisfies the following:
Proposition 5.1. Let A be an Artin group of type FC associated to a Coxeter graph Γ of diameter at least 3. Then the Deligne complex DA is irreducible and the action of A on DA is essential and non-elementary.
The proof will take most of this section. We can assume that Γ is connected for otherwise A is a free product, DA has a structure of a tree of
spaces, and the result follows. We check separately the irreducibility of the
ACYLINDRICAL HYPERBOLICITY AND CAT(0) CUBE COMPLEXES
13
Deligne complex, the essentiality of the action, and the non-elementarity of
the action.
Lemma 5.2. The CAT(0) cube complex DA is irreducible.
Proof. Consider the vertex {1} of DA , where 1 denotes the identity element.
The labeling of the edges of DA yields a surjective map lk({1}) → Γ. As Γ
has diameter at least 3, so does lk({1}), and it follows that lk({1}) is not a
join, hence DA is an irreducible CAT(0) cube complex.
Lemma 5.3. The action of A on DA is essential.
Proof. Since the action is cocompact, it is enough to show that each hyperplane is essential, that is, no half-space is contained in a neighbourhood
of the other halfspace. As each hyperplane is a translate of some ĥs , it is
enough to show that this is the case for hyperplanes of the form ĥs , where
s ∈ S. Let s0 be a vertex of Γ. Since Γ is connected and has diameter
at least 3, we can find two distinct vertices s1 , s2 of Γ such that s0 , s1 , s2
defines a geodesic of Γ. Then the hyperplane ĥs1 is in particular stabilized
by A{s0 ,s2 } . Thus, to show that ĥs0 is essential, it is enough to show that
ĥs1 is unbounded and crosses ĥs0 . Let Cs0 ,s1 , Cs1 ,s2 be the squares of DA
containing the vertices {1}, A{s0 } , A{s1 } and {1}, A{s1 } , A{s2 } respectively.
Notice that the edge between A{s0 } and A{s0 ,s1 } has stabiliser A{s0 } , and
the edge between A{s2 } and A{s1 ,s2 } has stabiliser A{s2 } . Thus, the A{s0 ,s2 } orbit Y of Cs0 ,s1 ∪ Cs1 ,s2 defines a subcomplex of DA that is convex for the
CAT(0) metric and quasi-isometric to a tree. Moreover, every point of Y
contained in a half-space of ĥs0 projects on the other half-space to a point
of Cs0 ,s1 ∪ Cs1 ,s2 . In particular, each half-space of ĥs0 contains points of Y
arbitrarily far away from the other half-space, hence ĥs0 is essential.
Lemma 5.4. The action of A on DA is non-elementary.
The proof of this lemma requires some preliminary work. Since Γ has
diameter at least 3, let s0 , s1 , s2 , s3 be a geodesic of Γ.
Lemma 5.5. Let g := s0 s3 . Then g is a hyperbolic element and admits an
axis Λg (for the CAT(0) metric) that is a reunion of geodesic segments such
that two consecutive segments make an angle strictly greater than π (for the
angular distance on the link).
Proof. For i = 0 or 3, we denote by ei the edge between the vertices
{1} and
S
−1
A{si } of DA . Let Y := s0 e0 ∪ e0 ∪ e3 ∪ s3 e3 . Then Λg := n∈Z gn Y is a
CAT(0) geodesic, and an axis for g. Indeed, we have the following properties
of angles between consecutive edges:
• The angle (for the angular distance on the link) between e0 and e3
is 3π/2 > π by construction of s0 , s1 , s2 , s3 .
• The angle between e0 and s−1
0 e0 (between e3 and s3 e3 respectively)
is π: Indeed, the angle is at most π by construction, and if the angle
14
INDIRA CHATTERJI AND ALEXANDRE MARTIN
were π/2, then since DA is a CAT(0) cube complex ei and si ei would
be two adjacent edges of a 3-cube with label ei , which is impossible
by construction of DA .
Thus, Λg is a local geodesic, hence a global geodesic. Moreover, Λg is clearly
invariant under the action of hgi, and the angle made by Λg at every hgitranslate of the vertex {1} ∈ DA is 3π/2.
To show that g is a rank-one element, we need the following modified
version of the Flat Strip Theorem:
Lemma 5.6 (Flat “Half-strip” Theorem). Let Y be a CAT(0) space, let Λ be
a geodesic line, and let h be an isometry of Y preserving Λ. Let y ∈ ∂Y \ ∂Λ
and let γ be a geodesic ray from a point x of Λ to y, which meets Λ in
exactly one point. If γ and hγ are asymptotic, then the convex hull of γ ∪ hγ
isometrically embeds in R2 with its standard CAT(0) metric.
Proof. Let us construct a ‘double’ of Y as follows. Let o ∈ Λ be the midpoint
between x and hx. Such a choice allows to define (uniquely) a reflection ψ
of Λ across o which is an isometry of Λ. We then define
Y ′ = (Y ⊔ Y )/ψ,
i.e. the space obtained from two copies of Y by identifying the two copies
of Λ using the reflection ψ. As a convention, we denote by Y1 and Y2 these
copies, we use the subscript · · ·i to indicate to which copy of Y the object
belongs, and we identify Y with the subspace Y1 of Y ′ . The space Y ′ is
again a CAT(0) space (as obtained from two CAT(0) spaces by identifying
two convex subspaces along an isometry). Note that γ ⊂ Y is a subray
of γ1 ∪ (hγ)2 and hγ ⊂ Y is a subray of (hγ)1 ∪ γ2 . Moreover, we have
constructed Y ′ so that γ1 ∪ (hγ)2 and (hγ)1 ∪ γ2 are local geodesics of Y ′ ,
hence global geodesics of Y ′ . As these geodesic lines are asymptotic by
construction, it follows that their geodesic hull is a flat strip, by the Flat
Strip Theorem [BH99, Theorem II.2.13]. In particular, the geodesic hull of
γ ∪ hγ in Y isometrically embeds in R2 .
We can now prove:
Lemma 5.7. The only points of ∂DA fixed by g are its two limit points
g+∞ , g −∞ ∈ ∂Λg .
Proof. Let z ∈ ∂DA \ {g+∞ , g−∞ } and let γ ′ be a geodesic ray from Λg
to z that meets Λg in exactly one point. If gz = z, then γ ′ and g2 γ ′ are
asymptotic geodesic rays, hence the convex hull H of γ ′ ∪ g2 γ ′ isometrically
embeds in R2 by the Flat Half-strip Theorem 5.6. But γ and g2 γ ′ both
meet Λg in exactly one point, and by construction H contains two geodesic
subsegments of Λg which make an angle 3π/2 for the angular distance on
the link. This yields the desired contradiction.
Proof of Lemma 5.4. First notice that one could have proved the above
lemma for the element h := s20 s3 by applying exactly the same reasoning. If
ACYLINDRICAL HYPERBOLICITY AND CAT(0) CUBE COMPLEXES
15
A were to admit a finite orbit at infinity, then some power of g and some
power of h would fix a common point at infinity. But the axes of g and h
cannot be asymptotic for otherwise, being already distinct, Λg and Λh would
bound a flat strip of positive width by the Flat Strip Theorem 5.6, and the
same reasoning as above would yield a contradiction.
Proof of Proposition 5.1. This is a direct consequences of Lemmas 5.2, 5.3,
and 5.4.
We are now ready to apply Theorem 1.1 to the action of A on DA in order
to complete the proof of Theorem 1.2.
Proof of Theorem 1.2. It is enough to consider the case where Γ is connected, for otherwise the group is a free product. According to the previous
lemma, the action of A on its Deligne complex DA is essential and nonelementary. In order to apply Theorem 1.1 to the action of A on DA , we
choose two vertices s, t of S with disjoint links, which is possible since Γ has
diameter at least 3. It follows from the aforementioned result of [vdL83] that
Alk(s) ∩ Alk(t) = A∅ = {1}. Thus, the hyperplanes ĥs and ĥt have stabilisers
which intersect trivially, hence Theorem 1.1 applies and A is acylindrically
hyperbolic.
References
[BBF14] M. Bestvina, K. Bromberg, and K. Fujiwara. Constructing group actions on
quasi-trees and applications to mapping class groups. Publ. Math. Inst. Hautes
Études Sci., in press, 2014.
[BC12] J. Behrstock and R. Charney. Divergence and quasimorphisms of right-angled
Artin groups. Math. Ann., 352(2):339–356, 2012.
[BFL14] C. Bisi, J.-P. Furter, and S. Lamy. The tame automorphism group of an affine
quadric threefold acting on a square complex. Journal de l’Ecole Polytechnique,
1:161–223, 2014.
[BH99] M. R. Bridson and A. Haefliger. Metric spaces of non-positive curvature, volume
319 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin, 1999.
[BM00] M. Burger and S. Mozes. Lattices in product of trees. Inst. Hautes Études Sci.
Publ. Math., (92):151–194 (2001), 2000.
[Bow08] B. H. Bowditch. Tight geodesics in the curve complex. Invent. Math.,
171(2):281–300, 2008.
[CD95] R. Charney and M. W. Davis. The K(π, 1)-problem for hyperplane complements
associated to infinite reflection groups. J. Amer. Math. Soc., 8(3):597–627, 1995.
[CFI16] I. Chatterji, T. Fernós, and A. Iozzi. The median class and superrigidity of
actions on CAT(0) cube complexes. J. Topol., 9(2):349–400, 2016.
[CL13]
S. Cantat and S. Lamy. Normal subgroups in the Cremona group. Acta Math.,
210(1):31–94, 2013. With an appendix by Yves de Cornulier.
[CS11]
P.-E. Caprace and M. Sageev. Rank rigidity for CAT(0) cube complexes. Geom.
Funct. Anal., 21(4):851–891, 2011.
[CW16] M. Calvez and B. Wiest. Acylindrical hyperbolicity and Artin-Tits groups of
spherical type. arXiv:1606.07778, 2016.
[Far08] D. Farley. The action of Thompson’s group on a CAT(0) boundary. Groups
Geom. Dyn., 2(2):185–222, 2008.
16
INDIRA CHATTERJI AND ALEXANDRE MARTIN
[Gen16a] A. Genevois. Coning-off CAT(0) cube complexes. arXiv:1603.06513, 2016.
[Gen16b] A. Genevois. Acylindrical action on the hyperplanes of a CAT(0) cube complex.
arXiv:1610.08759, 2016.
[Hag07] F. Haglund. Isometries of CAT(0) cube complexes are semi-simple. available at
https://hal.archives-ouvertes.fr/hal-00148515v1, 2007.
[Lon16] A. Lonjou. Non simplicité du groupe de Cremona sur tout corps. Ann. Inst.
Fourier (Grenoble), 66(5):2021–2046, 2016.
[Mar15] A. Martin. Acylindrical actions on CAT(0) square complexes. arXiv:
1509.03131, 2015.
[Mar16a] A. Martin. On the acylindrical hyperbolicity of the tame automorphism group
of SL2 (C). arXiv:1512.07526 , 2016.
[Mar16b] A. Martin. On the cubical geometry of Higman’s group. Duke. Math. J., in
press, 2016.
[MO15] A. Minasyan and D. Osin. Acylindrical hyperbolicity of groups acting on trees.
Math. Ann., 362(3-4):1055–1105, 2015.
[Osi06] D. V. Osin. Relatively hyperbolic groups: intrinsic geometry, algebraic properties, and algorithmic problems. Mem. Amer. Math. Soc., 179(843):vi+100,
2006.
[Osi16] D. Osin. Acylindrically hyperbolic groups. Trans. Amer. Math. Soc., 368(2):851–
888, 2016.
[Sag14] M. Sageev. CAT(0) cube complexes and groups. In Geometric group theory, volume 21 of IAS/Park City Math. Ser., pages 7–54. Amer. Math. Soc., Providence,
RI, 2014.
[vdL83] H. van der Lek. The homotopy type of complex hyperplane complements. Ph.D.
Thesis, University of Nijmegen, 1983.
Laboratoire de Mathématiques J.A. Dieudonné, Université de Nice Sophia
Antipolis, Parc Valrose, 06108 Nice Cedex 02, France
[email protected]
Department of Mathematics and the Maxwell Institute for Mathematical
Sciences, Heriot-Watt University, Riccarton, EH14 4AS Edinburgh, United
Kingdom
[email protected]
| 4math.GR
|
Lowering the learning curve for declarative
programming: a Python API for the IDP system
arXiv:1511.00916v1 [cs.PL] 3 Nov 2015
Joost Vennekens
Dept. Computer Science, KU Leuven
Technology Campus De Nayer, Sint-Katelijne-Waver, Belgium
Abstract. Programmers may be hesitant to use declarative systems,
because of the associated learning curve. In this paper, we present an
API that integrates the IDP Knowledge Base system into the Python
programming language. IDP is a state-of-the-art logical system, which
uses SAT, SMT, Logic Programming and Answer Set Programming technology. Python is currently one of the most widely used (teaching) languages for programming. The first goal of our API is to allow a Python
programmer to use the declarative power of IDP, without needing to
learn any new syntax or semantics. The second goal is allow IDP to be
added to/removed from an existing code base with minimal changes.
1
Introduction
While declarative systems may not be inherently more difficult to learn than,
e.g., an object-oriented programming language, they are still often perceived as
such, because many programmers are unfamiliar with their syntax, semantics or
paradigm. This leads in turn to a chicken-and-egg problem, where programmers
do not learn this technology, because companies are reluctant to adopt it, because
there are not enough programmers who know the technology readily available.
In this paper, we propose to tackle both problems by means of an API that
allows a declarative knowledge base (KB) to be used from within a well-known
imperative host language. Our first goal is to integrate KB functionality into the
host language as seamlessly as possible. In this way, it should be possible to use
the knowledge base to prototype a single component of a large system, without
affecting the rest of the code base. The second goal is to have a very low learning
curve for the API. We achieve this by using as much as possible the syntax of the
host language, and by requiring no more background in declarative languages
than an introductory course on classical logic. We therefore expect our API to
be immediately usable by, e.g., bachelor students in a typical CS curriculum.
In more detail, we postulate these guidelines for our development of the API:
– The interaction between the KB and host language should be done through
standard objects of the host language.
– The need to learn KB-specific terminology should be kept to a minimum.
– It should be as easy as possible to replace the KB by a piece of host language
code (or, vice versa, to replace a piece of host language code by a KB).
A typical use for our API will be to off-load specific computational problems
(e.g., detect connected components in a graph, find a permissible allocation of
resources to jobs) to the KB, thereby avoiding the need to implement a specific
algorithm and thus arriving at a working prototype more quickly. In such an
early prototype, the modular, declarative nature of the KB will be particularly
useful, because of its ability to easily cope with additional changes to the specification. Once the program has reached a certain level of maturity, it can of
course be profiled to see whether all of the KB components meet the performance requirements. Whenever this is not the case, the KB can be replaced by
a dedicated algorithm with minimal impact on the rest of the code.
As a host language, we use Python (in particular, version 2.7). Given our
stated goals, this is the most obvious choice: “[a]t the time of writing (July
2014), Python is currently the most popular language for teaching introductory
computer science courses at top-ranked U.S. departments.”1 We assume familiarity with the basics of Python throughout this paper. The KB system that we
use will be discussed in Section 2. Section 3 then discusses our interface between
host language and KB, which we validate by means of some examples in Section
4. In Section 5, we give some brief notes on the implementation of our API.
Finally, Section 6 discusses some related work, in particular, other approaches
that integrate a declarative knowledge base into an imperative language.
2
KB system
As an underlying KB system, we will use IDP (Imperative Declarative Programming)2 [2], which combines techniques from SAT solving, Logic Programming
and Answer Set Programming (ASP). It has performed well in previous ASP
competitions, e.g., narrowly finishing second after Clasp in the System Track of
20113. IDP has a number of properties that fit well with our goals of achieving
both a tight integration with the host language and a low learning curve.
Input language. IDP uses a language that is a conservative extension of classical first-order logic (FO). Because most students of computer science are familiar
with FO, this means that the learning curve for a large part of IDP’s input language consists only of learning a particular ASCII syntax for FO. One of the
ways in which IDP extends FO is by adding inductive definitions [3]. Because
such definitions cannot, in general, be expressed in FO, this is a real extension
of the language. Moreover, since inductive datatypes (lists, trees, . . . ) are very
common in computer programs, this feature will prove useful in our API.
Inference. In addition to its input language, a second useful property of IDP
is that it aims to support a variety of different inference tasks. Particularly useful
in the context of this article is the task of finite model expansion. As pointed
out in [6], modal expansion for FO captures the complexity class NP, thereby
1
2
3
http://cacm.acm.org/blogs/blog-cacm/176450-python-is-now-the-mostpopular-introductory-teaching-language-at-top-us-universities/fulltext
https://dtai.cs.kuleuven.be/software/idp
https://www.mat.unical.it/aspcomp2011/
84
covering the kind of tasks that we would like to off-load to a declarative KB.
Moreover, [7] have further demonstrated that model expansion is a key task
when using declarative methods to build modular software systems.
2.1
FO: syntax and semantics
We briefly recall the standard syntax and semantics of FO. A vocabulary Σ
consists of a set of function symbols, each with an associated arity n, and a
set of predicate symbols, also each with an arity n. A function with arity 0
is called a constant. A term is either a constant, a variable or an expression
f (t1 , . . . , tn ) where f is an n-ary function symbol and the ti are terms. An atom
is an expression P (t1 , . . . , tn ), with P an n-ary predicate and the ti again terms.
A formula is either an atom or an expression ψ ∨ φ, ψ ∧ φ, ¬φ, ∀x : φ, or ∃x : φ,
where ψ, φ are formulas and x is a variable. As usual, φ ⇒ ψ abbreviates ¬φ ∨ ψ
and φ ⇔ ψ stands for (φ ⇒ ψ) ∧ (ψ ⇒ φ). A sentence is a formula without free
variables and a theory is a finite set of sentences.
The standard semantics of FO is defined in terms of structures for a vocabulary Σ. Each structure S consists of a domain D and a mapping of:
– Each n-ary predicate symbol P in Σ to an n-ary relation R ⊆ Dn
– Each n-ary function symbol f in Σ to an n-ary function f : Dn → D
The satisfaction relation |= is defined between structures for a vocabulary and
theories of this vocabulary (or a subvocabulary thereof) by the usual induction.
When S |= T , we also say that the structure S is a model of T .
2.2
FO in the IDP system
IDP uses the standard concepts of vocabularies, theories and structures. Each of these has
a specific syntactic representation. For example, a map coloring problem can be described
in the following vocabulary V .
vocabulary V {
type Color
type Area
Border(Area,Area)
Coloring(Area): Color
}
FO IDP Python
∧ &
and
∨ |
or
not
¬ ~
⇒ => not present
= =
==
6= ~=
!=
∀ !
all
∃ ?
any
Fig. 1. ASCII syntax of IDP.
As can be seen here, IDP in fact uses a
typed variant of first-order logic. The first
two statements define two types (which can be seen as unary predicates), whereas
the last two statements define, respectively, a predicate and a function.
The following theory T in vocabulary V consists of a single sentence, expressing that neighboring areas must have a different color. Fig. 1 shows the ASCII
symbols that are used in IDP to represent the logical connectives.
85
theory T : V {
!a b: Border(a,b) => Coloring(a) ~= Coloring(b).
}
IDP is usually able to automatically derive the types of variables, based on
the type declarations in the definition of the vocabulary. This information is
important because most of its inference tasks require IDP to first ground (part
of) the theory. Instead of depending on IDP’s automatic type derivation, it is
also possible to explicitly declare the type of a variable, e.g.:
!a [Area] b [Area]: Border(a,b) => Coloring(a) ~= Coloring(b).
A structure S for a vocabulary V is represented by an enumeration (1) of the
values that belong to each type, (2) of the tuples that belong to each predicate,
and (3) of the mapping of tuples to values that is made by each function.
structure S : V {
Area = { Belgium; Holland; Germany; }
Color = { Blue; Red; Green; }
Border = {(Belgium,Holland); (Belgium,Germany); (Holland,Germany)}
}
This structure S for the vocabulary V interprets only part of the vocabulary V .
In particular, the function Coloring is not interpreted. One of the inference tasks
supported by the IDP system is that of model expansion: given a structure S ′
for a subvocabulary Σ ′ of the vocabulary Σ of a theory T , compute a structure
S for the remaining symbols Σ \ Σ ′ such that (S ′ ∪ S) |= T .
IDP exposes its functionality by means of an API in the Lua scripting language. The command printmodels(modelexpand(T,S)) performs the model
expansion task for the above structure S and theory T , resulting in the output:
Coloring = {"Belgium"->"Red";"Germany"->"Blue";"Holland"->"Green"}
By default, a single model expansion is computed, but it is also possible to
compute several or all of them. A special case of model expansion occurs when
the initial structure S already interprets the entire vocabulary V of the theory
T . In this case, it reduces to checking whether S |= T .
3
Interfacing with the KB System
This section presents our API for using the IDP KB system from within Python.
3.1
Vocabularies and structures
All interaction with the IDP system is done through objects of the IDP class.
Each such object represents a knowledge base consisting of a triple (V, S, T ) of
a vocabulary V , a structure S and a theory T . It can be created as follows:
86
kb = IDP ( )
The following methods add symbols to the vocabulary of the KB:
kb . Type ( name [ , i n t e r p r e t a t i o n ] )
kb . Constant ( typed name [ , i n t e r p r e t a t i o n ]
kb . Functio n ( typed name [ , i n t e r p r e t a t i o n ] )
kb . P r e d i c a t e ( typed name [ , i n t e r p r e t a t i o n ] )
As in IDP itself, the typed_name of a predicate is a string of the form
Foo(Type1, ..., Type2) and that of a function is Foo(Type1, ..., Type2):
Return_type. Because constants are identical to 0-ary functions, their typed_name
has the form Foo : Type. Once a symbol Foo has been added to the vocabulary
of a knowledge base kb, it can thereafter be referred to as kb.Foo.
In addition to declaring a function/predicate symbol σ (i.e., adding it to the
vocabulary of the KB), it is also possible to immediately extend the structure S
with a particular interpretation σ S for σ. This is done by adding this interpretation as a second argument. The interpretation of a type must be a set (or list)
of values; that of a constant must be a single value; that of a function with arity
≥ 1 must be mapping (e.g., a dictionary); and that of a predicate must be a
set/list of tuples of the correct arity (for predicates with arity 1, a set of simple
values is also allowed). Obviously, the typing of the symbols must be respected.
Instead of initialising the interpretation of a symbol upon construction, it is
also possible to first declare a symbol and then later use the assignment operator
to provide an interpretation for it. We illustrate using (part of) the graph coloring
example of Sec. 2.2.
color
color
color
color
color
= IDP ( )
. Type ( " Color" , [ " Blue" , "Red" , " Green" ] )
. Type ( " Area" , [ " Belgium " , " Holland " , " Germany " ] )
. P r e d i c a t e ( " Border (Area , Area)" )
. Border =[( " Belgium " , " Holland " ) , ( " Belgium " , " Germany " ) ,
( " Holland " , " Germany " ) ]
c o l o r . Functio n ( " Coloring(Area): Color" )
The “logical” objects that are thus created implement a number of common
Python interfaces, allowing them to act as Python programmers would expect.
A relation is, in mathematical terms, a set of tuples. Its natural counterpart
is a Python MutableSet object4 (i.e., a set which allows adding/removing of
elements). The following interactive session demonstrates some standard usages.
>>> " Belgium " in c o l o r . Area
True
>>> c o l o r . Area . add ( " Austria " )
>>> for x in c o l o r . Area :
4
https://docs.python.org/2/library/collections.html#collections.Set
87
...
print x ,
...
Ho lla nd A u s t r i a Germany Belgium
In addition to the standard MutableSet functionality, relations are also callable5 ,
so that we may also use the standard FO notation for checking membership.
>>> c o l o r . Area ( " Belgium " )
True
A function is a Mapping6 , that is, an object that maps each tuple of values
in its domain to a value in its range. Some standard usages are:
>>> c o l o r . C o l o r i n g . keys ( )
[ ’Austria ’ , ’Germany ’ , ’Holland ’ , ’Belgium ’ ]
>>> c o l o r . C o l o r i n g [ " Belgium " ]
’Blue’
As with predicates, function objects are callable to allow for the more FO-like:
>>> c o l o r . C o l o r i n g ( " Belgium " )
’Blue’
3.2
Formulas and definitions
In keeping with our goal of achieving a low learning curve, formulas are written
in Python syntax. An overview is shown in Fig. 1. The Python language has the
standard boolean operators and, or and not. In addition, it also has the functions all and any, which may be applied to lists of boolean values to return the
conjunction/disjunction of these values. The latter two functions, together with
Python’s list comprehension syntax, can be used as universal/existential quantification. The list comprehension syntax also has an optional if part, which may
be used to represent the common pattern of a universally quantified implication:
∀x[T ype] : φ
∃x[T ype] : φ
∀x[T ype] : φ ⇒ ψ
all(φ for x in Type)
any(φ for x in Type)
all(ψ for x in Type if φ)
In the graph coloring problem, we need to express the following property:
∀a b : Border(a, b) ⇒ Coloring(a) 6= Coloring(b)
(1)
Section 2.2 already presented the IDP syntax for this. In Python, we can write
the same property as:
5
6
https://docs.python.org/2/library/functions.html#callable
https://docs.python.org/2/library/collections.html#collections.Mapping
88
a l l ( c o l o r . C o l o r i n g ( a ) != c o l o r . C o l o r i n g ( b )
for ( a , b ) in c o l o r . Border )
Note that this is just a normal Python expression, which we can, e.g., just type
into the interactive terminal. This expression evaluates to True precisely when
property (1) is satisfied. We can make the KB aware of such a constraint by
means of its Constraint method, which takes a string as its argument. So, the
following code adds the above constraint to our graph coloring KB:
c o l o r . C o n s t r a i n t ( " all( Coloring(a) != Coloring(b)
for (a,b) in Border " )
The string argument is completely identical to the Python expression we saw
above, with one exception: the predicates/functions are simply called Coloring,
Area and Border, instead of color.Coloring, color.Area and color.Border.
This is because, just as a theory can only contain symbols that appear in its
vocabulary, the constraints that are added to a KB always refer to the symbols
of that KB.
An alternative formula, which is equivalent to the one above, quantifies over
Area × Area and uses an if-expression to check for membership in Border:
c o l o r . C o n s t r a i n t ( " all( Coloring(a) != Coloring(b)
for a in Area for b in Area if (a,b) in Border )" )
3.3
Functional interfacing
In keeping with our goal of making the API easy to use, the programmer does not
need to explicitly invoke the IDP system. This avoids the need to learn new functions or new terminology, and reduces the possibility of bugs. Instead, invokation
of the IDP system happens “automagically” in the following circumstances:
– Symbols that have been declared, but for which no interpretation has been
provided, are automatically assigned a valid interpretation (in accordance
with FO semantics) when their content is inspected. In other words, IDP is
used as an oracle to lazily fill in the interpretation of any declared symbols
for which the user does not provide one herself. This is done in a way that
the interpretations of all symbols together constitutes an FO model of the
constraints, i.e., a model expansion task is performed. If the constraints
admit multiple models, one is chosen arbitrarily.
– The KB object has an attribute satisfiable, which is automatically set
to True/False (depending on whether the KB is satisfiable) when the user
converts it to a boolean (either explicitly with bool(.) or by use in an
if-statement).
In the previous section, we declared a function color.Coloring without
adding an interpretation for this function. Therefore, the IDP system will be
invoked to compute a coloring of our graph if we execute the following code:
89
for x in c o l o r . Area :
print " Area %s has color %s" % ( x , c o l o r . C o l o r i n g [ x ] )
Note that if we had added an interpretation for the function Coloring before
executing this code, then this for-loop would still continue work as expected.
This is an important property, because it allows us to change whether a relation/function is computed by the KB base or by native Python code, without having to adjust the code that makes use of it. In the latter case, the call
color.Coloring[x] will just retrieve the pre-computed coloring stored within
the color object, without invoking IDP.
Similarly, if a coloring is not provided, then the following code will test
whether a given graph can be colored, whereas if it is provided, the same code
checks whether it is indeed valid.
if color . s at i s f i ab l e :
print " The graph can be colored "
3.4
Inductive definitions
An important feature of the IDP system is its ability to handle inductive definitions. It uses a rule-based syntax for representing such definitions, in which,
e.g., the transitive closure T C of a graph G can be defined as follows
(
)
∀x, y : T C(x, y) ← G(x, y).
∀x, y : T C(x, y) ← ∃z : G(x, z) ∧ T C(z, y).
Note that the arrow symbol here is not material implication, but a special
symbol that denotes a “case” in an inductive definition. Such an inductive definition is interpreted under the well-founded semantics [9], which in the case
of a positive definition (such as the one above) boils down to a least-fixpoint
construction. Each rule of such a definition represents a single case in which
the defined predicate holds. In our Python API, we use a lambda-expression to
represent such a case.
kb . D e f i n e ( [ ( "TC(Node , Node)" , "lambda x,y: G(x,y)" ) ,
( "TC(Node , Node)" , "lambda x,y:
any(G(x,z) and TC(z,y) for z in Node)" )
This both declares the predicate TC and defines it in terms of the “parameter”
G. For definitions consisting of a single rule, a simpler syntax is also allowed:
kb . D e f i n e ( "TC(Node ,Node)" ,
" lambda x,y: G(x,y) or
any(G(x,z) and TC(z,y) for z in Node)" )
90
Similar to how an argument of kb.Constraint(.) can also be used as a
simple boolean Python expression, the above lambda-expression can be used to
compute the transitive closure of G by an explicit least-fixpoint computation:
def l f p ( f , x = [ ] ) :
y = f (x)
return y i f y == s e l s e f ( x )
n o d e p a i r s = [ ( x , y ) for x in kb . Node for y in kb . Node ]
TC = l f p (lambda T: f i l t e r (lambda x , y : kb .G( x , y ) or
any ( kb .G( x , z ) and T( z , y ) for z in Node ) , n o d e p a i r s ) )
An advantage of using IDP is that the definition of T C can then not only
be used to compute the transitive closure of a given graph, but also to, e.g.,
compute a graph that would have a given relation as its transitive closure. In
addition, IDP not only supports positive inductive definitions, but also nonmonotone inductive definitions (such as the standard definition of the relation
“|=” in FO), for which a simple least-fixpoint construction does not work. Nonrecursive definitions (which are equivalent to a standard FO equivalence) are
also allowed in IDP. In the latter case, we can of course choose whether to use
the Define or Predicate method of our API.
IDP can be configured to use XSB Prolog7 to speed up certain computation
with definitions. We always use this option in the experiments below.
4
Experiments
This section presents two examples of our API, with a particular focus on demonstrating that the integration into the surrounding Python code can be done in
a natural way.
4.1
Sudoku
The first example is a Sudoku solver. A Sudoku grid consists of 9 × 9 = 81 cells.
sud = KB( )
sud . Type ( " Cell" , r a ng e ( 8 1 ) )
The grid is divided into in rows, columns and nine small 3 × 3 squares.
sud . P r e d i c a t e ( " SameRow (Cell , Cell)" ,
[ ( i , j ) for i in sud . C e l l for j in sud . C e l l
i f i / 9 == j / 9 ] )
sud . P r e d i c a t e ( " SameCol (Cell , Cell)" ,
[ ( i , j ) for i in sud . C e l l for j in sud . C e l l
i f i % 9 == j % 9 ] )
7
http://xsb.sourceforge.net/
91
sud . P r e d i c a t e ( " SameSmallSq(Cell , Cell)" ,
[ ( i , j ) for i in sud . C e l l for j in sud . C e l l
i f ( i %9)/3 == ( j %9)/3 and ( i / 9 ) /3 == ( j / 9 ) /3 ] )
Here, the Python list comprehensions compute an enumeration of these relations, by iterating over all tuples in the Cartesian product of the argument types
and checking a certain condition for each tuple. Alternatively, we can make IDP
do this work, by Defining the predicates with the appropriate lamba-expressions.
sud . D e f i n e ( " SameRow (Cell , Cell)" ,
" lambda i, j: i / 9 == j / 9" )
The cells must be filled with integers from 1 to 9. It will be convenient to
represent an empty cell by the number 0, leading to the following type Number.
sud . Type ( " Number " , r a ng e ( 1 0 ) )
We make use of two functions that map cells to numbers: one records the
problem statement and the other its solution. The problem statement comes from
a list of numbers, that we convert to a dictionary by means of zip operation.
s =[ 8 , 5 , 0 , 0 , 0 , 2 , 4 , 0 , 0 ,
7 ,2 ,0 , 0 ,0 ,0 , 0 ,0 ,9 ,
0 ,0 ,4 , 0 ,0 ,0 , 0 ,0 ,0 ,
...
...
...
]
sud . Functio n ( "Given( Cell): Number " ,
d i c t ( z i p ( r a ng e ( l e n ( s ) ) , s ) ) )
sud . Functio n ( "Sol( Cell): Number " )
The function Sol will be computed by IDP, in accordance with the rules of
sudoku. First, we state the difference constraint on the appropriate cells.
sud . D e f i n e ( " Diff(Cell , Cell)" , " lambda x,y: x != y and
( SameRow (x,y) or SameCol (x,y) or SameSmallSq(x,y))" )
sud . C o n s t r a i n t ( " all( Sol(x) != Sol(y) for (x,y) in Diff)" )
Next, we state that the solution must match the problem statement on all
non-empty cells (i.e., those 6= 0), and that it should fill in all cells.
sud . C o n s t r a i n t ( " all( Sol(x) == Given (x)
for x in Cell if Given(x) != 0)" )
sud . C o n s t r a i n t ( " all( Sol(x) != 0 for x in Cell)" )
With this, the sudoku problem is completely specified. The following code
passes the sud.Sol object to a function that pretty-prints the sudoku. Only at
the start of the for-loop in this function is the solution actually computed.
def show ( g r i d ) :
row = −1
for x in g r i d . keys ( ) :
92
s y s . s t d o u t . w r i t e ( ’ ’ i f x % 3 == 0 e l s e ’’ )
i f x % 9 == 0 :
row += 1
print ( ’\n’ i f row % 3 e l s e ’’ )
print ( s t r ( g r i d [ x ] ) + " " ) ,
show ( sud . S o l )
We remark that, because we use valid Python expressions to assert constraints, we can use the same expressions to check that the output indeed satisfies
the constraints. For instance, at the interactive Python terminal:
>>> a l l ( sud . S o l ( x ) != sud . S o l ( y ) for ( x , y ) in sud . D i f f )
True
Peter Norvig has published a sudoku solver written entirely in Python using
constraint solving techniques8 . Not counting whitespace and comments, his code
to compute solutions is about 40 lines, whereas the code we have presented in this
section is 12 lines. Moreover, it is easy to replace the existing solve function in
his code by a call to our knowledge base. This requires two small transformations:
first, in Norvig’s code, input grids are given in the format of a single string, where
an empty cell is represented by a dot; second, he produces output in the form
of a dictionary in which the keys are strings of the form Xn with X ∈ A..I
representing the row and n ∈ 1..9 the column.
def s o l v e ( sud , g r i d ) :
""" Solve sudoku " grid" using the KB "sud". """
def t r a n s l a t e ( n ) :
rows
= ’ ABCDEFGHI’
digits
= ’ 123456789’
return rows [ n / 9 ] + d i g i t s [ n%9]
sud . Given = d i c t ( z i p ( r a ng e ( 8 1 ) ,
[ i n t ( x ) for x in g r i d . r e p l a c e ( ’.’ , ’0’ ) ] ) )
return d i c t ( z i p (map( t r a n s l a t e , sud . S o l . keys ( ) ) ,
map( s t r , sud . S o l . v a l u e s ( ) ) ) )
When it comes to runtime, our version is significantly slower than the original
on Norvig’s test set, averaging 0.27s per sudoku versus 0.008s.
At the other end of the spectrum, we can also compare to a naive generateand-test approach. Using Python’s powerful itertools library, this can also be
implemented in about 10 lines of code. (The code used to test whether a solution
is correct can of course use the same syntactical expressions as those which we
passed to our API.) However, the runtime of such a program is very poor: for a
sudoku with just 5 empty cells (for comparison, a typical sudoku has around 50
to 60), it already takes over a minute to find a solution.
Our main conclusions from this experiment are that, at least in this case:
8
http://norvig.com/sudoku.html
93
– Our API can handle, with a limited amount of overhead, the input/output
format that a typical Python programmer would use;
– Our API can be used to develop useful functionality in significantly fewer
lines of code (12 versus 40) than a clever Python implementation. In fact, it
takes only as many lines of code as a naive generate-and-test algorithm.
– Even though our API is significantly slower than a clever Python algorithm,
it still vastly outperforms the naive generate-and-test approach.
4.2
Working with graphs
The following class GraphKB extends the generic IDP Knowledge Base class with
some specific functionality for working with undirected graphs. When constructing such a GraphKB, the nodes of the graph can be initialised by means of a
given set and the edges by means of an adjacency list. The predicate Edge is
Defined as the symmetric closure of the adjacency list. This class also offers a
convenience method to define the transitive closure of relations over this graph.
c l a s s GraphKB (IDP ) :
def
i n i t ( s e l f , no des = [ 0 ] , a d j l i s t = [ ] ) :
s u p e r (GraphKB , s e l f ) .
i n i t ()
s e l f . Type ( "Node" , no des )
s e l f . P r e d i c a t e ( " Adjacent(Node , Node)" , a d j l i s t )
s e l f . D e f i n e ( " Edge(Node , Node)" ,
" lambda x,y: Adjacent(x,y) or Adjacent(y,x)" )
def add TC ( s e l f , o r i g i n a l , tc name ) :
f o r m u l a = "lambda x,y: {0}(x,y) or any({0}(x,z)
and {0}(z,y) for z in Node)" . fo r ma t ( o r i g i n a l )
s e l f . D e f i n e ( tc name + "(Node , Node)" , f o r m u l a )
We can now check if a given adjacency list describes a fully connected graph:
c o n n e c t e d = GraphKB( nodes , a d j )
c o n n e c t e d . define TC ( " Path" , " Edge" )
c o n n e c t e d . C o n s t r a i n t ( " all( Path(x,y)
for x in Node for y in Node)" )
i f connected . s a t i s f i a b l e :
print " Graph is fully connected"
We can use a similar KB to count the number of connected components in
the graph. We do this by selecting a single representative from each component
(its “Root”) and then counting the number of these representatives.
comp = GraphKB( nodes , a d j )
comp . define TC ( " Path" , " Edge" )
comp . P r e d i c a t e ( " Root( Node)" )
94
comp . C o n s t r a i n t ( " all( any(Path(r,x) for r in Root)
for x in Node if not Root(x))" )
comp . C o n s t r a i n t ( " not any(Path(x,y)
for x in Root for y in Root if x != y)" )
print " Number of components: {0}" . fo r ma t ( l e n ( comp . Root ) )
For a graph with 1000 nodes in 86 connected components, this program takes
23s to count the components. By comparison, the popular NetworkX Python
library9 is two orders of magnitude faster, taking only 0.2s.
In graph theory, an undirected graph is called a tree if it is connected and
does not contain cycles. When checking for a cycle in an undirected graph, we
of course have to exclude the trivial two-node cycles that would result from
traversing the same undirected edge in both directions. This in fact makes it
easier to use IDP to check that there is a cycle, than to check that there is
not one. The following knowledge base tries to guess the direction in which to
traverse each edge in order to produce a cycle. If it is unsatisfiable, there are no
cycles.
cyclic
cyclic
cyclic
cyclic
= GraphKB ( )
. P r e d i c a t e ( " Traverse(Node , Node)" )
. C o n s t r a i n t ( "all( Edge(x,y) for (x,y) in Traverse)" )
. C o n s t r a i n t ( "not any( Traverse(y,x)
for (x,y) in Traverse)" )
c y c l i c . define TC ( " TravTC " , " Traverse" )
c y c l i c . C o n s t r a i n t ( "any( TravTC (x,x) for x in Node)" )
We can now combine the two knowledge bases to check whether a given
adjacency list indeed describes a tree.
def i s t r e e ( a d j l i s t ) :
c y c l i c . Adja cent = a d j l i s t
c o n n e c t e d . Adja cent = a d j l i s t
return ( b o o l ( c o n n e c t e d . s a t i s f i a b l e )
and not b o o l ( c y c l i c . s a t i s f i a b l e ) )
This example illustrates how additional functionality can be built on top of
the KB objects of our API. In addition, the ability to combine the results of
calls to different KBs also allows us to implement functionality that would be
harder to implement in a single IDP KB.
5
Implementation
The implementation of our API and the examples are available for download.10
Interfacing with the IDP system is currently done in a decoupled way: when the
9
10
https://networkx.github.io/
https://bitbucket.org/joostv/pyidp/admin
95
API detects that the IDP system needs to be called, it prepares a text file with
the appropriate vocabulary, structure and theory, then calles the IDP system as
an external process and parses its output. The results of this call are cached, so
that the IDP system will not be invoked again until the KB changes.
6
Related work
There is already a long history of work attempting to close the gap between
imperative and declarative programming [1]. We briefly compare our approach
to some recent work in this area.
In [8], an approach is presented in which a constraint solver is not added to
a single host language, but can be used in the development of a domain-specific
language in Racket. Like ours, the motivation behind this work is to allow the
power of declarative systems to be more widely used. However, their approach
differs, because they count on an intermediary—the designer of the domainspecific language—to hide the complexity of the declarative system, whereas our
approach focuses on creating an interface that is natural enough to offer KB
functionality directly.
In [4], a constraint solver is integrated into the Scala language. As ours does,
their approach reuses the syntax of the host language to interface with the
declarative system. A key difference is that, in their approach, the programmer
is explicitly manipulating, combining and solving constraints, which makes the
constraint solver more present in the eventual source code. A second difference
is of course that Scala currently appears to be less widely known than Python.
In [5], a reasoner for FO extended with transitive closure is integrated into
Java. Their KB language is therefore very similar to (but more restricted than)
that of IDP. When it comes to the integration in Java, there are two main
differences to our approach. First, the declarative knowledge is not written in
expressions in the host language, but in a separate language (the Alloy-like JFSL
[10]). Second, the integration into Java is done in an object-oriented way: the
programmer defines classes in which formulas are added as, among others, class
invariants, method pre-/postconditions and frame conditions. In comparison, our
Python API seems more lightweight, since it does not require an object-oriented
approach. When it comes to computational performance, [5] reports good results,
which our implementation is not able to match.
In summary, our approach fills the niche of an easy-to-learn quick prototyping
API, that, due to Python’s current popularity, may speak to a large audience.
7
Conclusions and future work
When prototyping an application, a programmer may encounter a computational
subproblem for which it would be cumbersome to develop a specific algorithm.
The aim of our API is to allow such gaps to be declaratively stopped with
as little effort as possible. As we have seen, our API might allow a feasible
solution to be produced in only as many lines of code as an (infeasible) naive
96
generate-and-test algorithm. Our use of standard Python objects such as sets
and mappings means that no elaborate setup code is required to plug the KB
into an existing code base, while our use of standard Python expressions for
constraints and definitions leads to a low learning curve. In addition, both these
properties also make it easier to eventually remove the KB if a more efficient
solution is required: the same KB that first generated the solution, can later be
used to check its correctness, or its constraints may simply be recuperated in
the form of Python assert-statements.
To prevent changing/removing the KB from leading to code changes elsewhere, our API makes all calls to the IDP system automatically, whenever they
are needed. This has the additional benefit of simplifying the API and not forcing the programmer to learn new terminology. A downside is that it is harder
for the programmer to keep track of what is happening when in the program.
Our current implementation of the API is naive in its interfacing with the
IDP system, which happens by passing text files (built each time from skratch)
to an external process. A better integration, which exploits the Lua interface of
IDP, might offer a significant reduction in runtimes. However, since we mainly
intend our API to be used in prototyping, this issue might not be pressing.
Another consequence, which may be more severe, is that programs written in
our API are currently hard to debug: it may be necessary to manually inspect
the text file that was passed to the IDP system (in debug-mode, the API always
sends this to standard output). However, this requires the user to be at least
somewhat familiar with IDP input syntax, which is something we aimed to avoid.
Our validation of the API currently consists only of examples that we have
implemented ourselves. A better test would involve Python programmers who
have no knowledge of IDP or indeed any declarative system. However, better
debugging facilities seem necessary for such a trial to be successful.
97
References
1. Krzysztof R. Apt and Andrea Schaerf. Programming in alma-0, or imperative and
declarative programming reconciled. In FroCos, 1998.
2. M. Bruynooghe, H. Blockeel, B. Bogaerts, B. De Cat, S. De Pooter, J. Jansen,
A. Labarre, J. Ramon, M. Denecker, and S. Verwer. Predicate logic as a modeling
language: Modeling and solving some machine learning and data mining problems
with IDP3. Theory and Practice of Logic Programming, 2014. Accepted.
3. Marc Denecker and Eugenia Ternovska. A logic of nonmonotone inductive definitions. ACM Transactions on Computational Logic, 9(2), March 2008.
4. A. Köksal, V. Kuncak, and P. Suter. Constraints as control. In POPL’12, 2012.
5. A. Milicevic, D. Rayside, K. Yessenov, and D. Jackson. Unifying execution of
imperative and declarative code. In Proc. 33rd Int’l Conference on Software Engineering (ICSE), 2011.
6. David G. Mitchell and Eugenia Ternovska. A framework for representing and
solving NP search problems. In AAAI, pages 430–435, 2005.
7. Shahab Tasharrofi and Eugenia Ternovska. A semantic account for modularity in
multi-language modelling of search problems. In Proc. FroCos, 2011.
8. E. Torlak and R. Bodik. Growing solver aided languages with rosetta. In
Proc. ACM international symposium on New ideas, new paradigms, and reflections
on programming & software, 2013.
9. A. Van Gelder, K.A. Ross, and J.S. Schlipf. The well-founded semantics for general
logic programs. Journal of the ACM, 38(3):620–650, 1991.
10. K. Yessenov. A lightweight specification language for bounded program verification. Master’s thesis, MIT, 2009.
98
| 6cs.PL
|
HOMOTOPY INVARIANT COMMUTATIVE ALGEBRA OVER FIELDS
arXiv:1601.02473v1 [math.AC] 11 Jan 2016
J.P.C. GREENLEES
Abstract. This article grew out of lectures given in the programme ‘Interactions between
Representation Theory, Algebraic Topology and Commutative Algebra’ (IRTATCA) at the
CRM (Barcelona) in Spring 2015. They give some basic homotopy invariant definitions in
commutative algebra and illustrate their interest by giving a number of examples.
Contents
Part
1.
2.
3.
4.
5.
1. Context
Introduction
Ring and module spectra I (motivation)
Ring and module spectra II (construction)
The three examples
Finiteness conditions
2
2
5
7
10
12
Part
6.
7.
8.
2. Morita equivalences
Morita equivalences
Proxy-smallness in the examples
Exterior algebras
13
13
18
19
Part 3. Regular rings
9. Regular ring spectra
10. Finite generation
21
21
25
Part
11.
12.
13.
14.
15.
16.
27
27
29
31
32
35
37
4. Hypersurface rings
Complete intersections
Hypersurfaces in algebra
Bimodules and natural endomorphisms of R-modules
Hypersurface ring spectra
s-hypersurface spaces and z-hypersurface spaces
Growth of z-hypersurface resolutions
Part 5. Gorenstein rings
17. The Gorenstein condition
18. Gorenstein duality
37
37
38
I am grateful to the CRM and the Simons Foundation for the opportunity to spend time in the stimulating
environment at the CRM during the IRTATCA programme, and to give these lectures.
1
19. Ascent, descent and arithmetic of shifts
20. Morita invariance of the Gorenstein condition
21. Gorenstein duality for group cohomology
22. Gorenstein duality for rational spaces
23. The ubiquity of Gorenstein ring spectra
Appendix A. Algebraic definitions: Local and Čech cohomology and homology
Appendix B. Spectral analogues of the algebraic definitions
Appendix C. Completion at ideals and Bousfield localization
Appendix D. Localization away from ideals and Bousfield localization
References
41
45
46
49
51
52
56
57
59
60
Part 1. Context
1. Introduction
1.A. The lectures. The purpose of these lectures is to illustrate how powerful it can be
to formulate ideas of commutative algebra in a homotopy invariant form. The point is that
it shows the concepts are robust, in the sense that they are invariant under deformations.
They can then be applied to derived categories of rings, but of course derived categories of
rings are just one example of a homotopy category of a stable model category. In short,
these ideas are powerful in classical algebra, in representation theory of groups, in classical
algebraic topology and elsewhere. We will focus on the fact that this includes many striking
new examples from topology.
Indeed, our principal example will be that arising from a topological space X, which is
to say we will work over the ring R = C ∗ (X; k) of cochains on a space X with coefficients
in a field k. Of course this is a rather old idea, applied in rational homotopy theory to
give a very rich theory and many interesting examples. The key is that R needs to be
commutative: classically, one needs to work over the rationals so that there is a commutative
model for cochains (coming from piecewise polynomial differential forms). It is well known
that there is no natural model for C ∗ (X; k) which is a commutative differential graded algebra
(commutative DGA) if k is not of characteristic 0, since the Steenrod operations (which do
not usually vanish) are built out of the non-commutativity. However, there is a way round
this, since we can relax our requirements for what we consider a model. Instead of requiring
DGAs, we permit R to be a commutative ring spectrum in the sense of homotopy theory
(see Section 3 for a brief introduction). There is then a commutative model for cochains;
we will continue to use the notation C ∗ (X; k) even though it is now a ring spectrum rather
than a DGA.
This may be our principal example, but actually it is a very restricted one. It is like
restricting graded commutative rings to be negatively graded k-algebras. We will hint at a
number of examples coming from derived algebraic geometry.
The main point of the lectures is to give some interesting examples, and the main source
of examples will be modular representation theory and group cohomology. In fact, the focus
2
will be on examples picked out by requiring the ring spectrum to be well behaved in terms of
commutative algebra: we look for regular, complete intersection or Gorenstein ring spectra.
As one expects from commutative algebra, regular ring spectra are very restricted and can
sometimes even be classified. One also expects to be able to parametrise complete intersections in some sense, although our present account just mentions a few isolated examples.
Finally, just as is the case for conventional commutative rings, the class of Gorenstein ring
spectra has proved ubiquitous. We shall see examples from representation theory, from chromatic stable homotopy theory and from rational homotopy theory, and one of the distinctive
contributions is to emphasize duality.
1.B. 54321. It may be worth bearing in mind the following correspondences, which suggest
useful ways to think about the material here.
The pentagon: We may summarize sources examples and the relationship between them
in a diagram. The emphasis here will be on commutative algebra and spaces.
Commutative Algebra
k❱
❱❱❱❱
❱❱❱❱
❱❱❱❱❱
❱❱❱❱ X
❱8
DAG
Chromatic homotopy theory
Representation theory of G
Spaces and DGAs
over Q
_
/
Topological spaces
The square: It is worth bearing in mind the dual points of view of geometric objects
and rings of functions on them, and the way the global picture is built up out of points. Our
present account focuses on the local rings at the bottom right, but it can be illuminating to
consider the wider context.
Geometry
Global
XO o
Local
Xx o
Algebra
/
?
/
OO X
?
OX,x
C ∗ (X; k)
X
The three classes: our focus is on the following hierarchy of local rings.
Regular
/
Complete Intersection
/
Gorenstein
The dual pair: even though our focus is on commutative rings, we make constant use of
the freedom to use Morita theory to transpose properties into the non-commutative context.
Commutative o
/
3
Non-commutative
It is the purpose of the present article to give an account of these objects and points of view
as a
unified whole.
1.C. Organization. The article is organized into five parts. In Part 1 we introduce the
context of ring spectra and modules over them, together with our three main classes of
examples (commutative rings, spaces and representation theory) and the basic finiteness
conditions.
In Part 2 (Sections 6 to 8) we describe the basic Morita equivalences between commutative
and non-commutative contexts, discuss our fundamental finiteness condition and illustrate
its power by considering the richness of exterior algebras.
We then turn to our hierarchy of local rings. In Part 3 (Sections 9 and 10) we consider
regular ring spectra; the principle of considering definitions in three styles (structural, growth
or module theoretic) recurs later, and the good behaviour of modules over regular rings is
used later to give a definition of ‘finitely generated’ modules for many ring spectra.
In Part 4 (Sections 11 to 16) we consider hypersurface ring spectra. We spend a little time
recalling the classical algebraic theory before introducing the definitions for ring spectra,
illustrating it with examples and explaining how the different versions are related.
Finally in Part 5 (Sections 17 to 23) we consider Gorenstein ring spectra, and the duality
properties they often enjoy. We describe the basic tools for Gorenstein ring spectra and
illustrate them by explaining many known examples (especially those from representation
theory and rational homotopy theory). The rich and attractive dualities that appear make
this a highlight.
A series of appendices recalls notation and basic properties of local cohomology and derived
completion.
Each part starts with a more detailed description of its contents.
1.D. Sources. Much of this account comes from joint work [33, 32, 21, 22, 55, 54] and
other conversations with D.J.Benson, W.G.Dwyer, K.Hess, S.B. Iyengar and S. Shamir, and
I would like to thank them for their collaboration and inspiration.
We summarize some relevant background, but commutative algebraists may find the articles [52] and [53] from the 2004 Chicago Workshop useful. General background in topology
may come from [84] and in conventional commutative algebra from [83, 28].
It is worth emphasizing that this is a vast area, and these lectures will just focus on a
tiny part corresponding to local k-algebras. In particular we will not touch on arithmetic or
geometric aspects of the theory, such as the Galois theory of J. Rognes, Brauer groups, or
derived algebraic geometry.
1.E. Conventions. We’ll generally denote rings and ring spectra by letters like, Q, R and S.
Ring homomorphisms will generally go in reverse alphabetical order, as in S −→ R −→ Q.
Modules will be denoted by letters like L, M, N, . . ..
The ring of integers (initial amongst conventional rings) is denoted Z. The sphere spectrum
(initial amongst ring spectra) is denoted S.
4
After the inital sections we will use the same notation for a conventional ring and the
associated Eilenberg-MacLane spectrum. Similarly for modules.
Generally, M ⊗R N will denote the left derived tensor product, and HomR (M, N) the
module of derived homomorphisms. Similarly, fibre sequences, cofibre sequences, pullbacks
and pushouts will be derived.
1.F. Grading conventions. We will have cause to discuss homological and cohomological
gradings. Our experience is that this is a frequent source of confusion, so we adopt the
following conventions. First, we refer to lower gradings as degrees and upper gradings as
codegrees. As usual, one may convert gradings to cogradings via the rule Mn = M −n . Thus
both chain complexes and cochain complexes have differentials of degree −1 (which is to say,
of codegree +1). This much is standard. However, since we need to deal with both homology
and cohomology it is essential to have separate notation for homological suspensions (Σi )
and cohomological suspensions (Σi ): these are defined by
(Σi M)n = Mn−i and (Σi M)n = M n−i .
Thus, for example, with reduced chains and cochains of a based space X, we have
C∗ (Σi X) = Σi C∗ (X) and C ∗ (Σi X) = Σi C ∗ (X).
2. Ring and module spectra I (motivation)
Most of the generalities can take place in any category with suitable formal structure: we
need a symmetric monoidal product so that we can discuss rings and modules over them,
and we need to be able to form tensor products over a commutative ring, and we will want a
homotopy theory which is stable so that we can form a triangulated homotopy category. In
fact we will work with a category of spectra in the sense of homotopy theory; the construction
is sketched in Appendix 3, and the present section gives a brief orientation. A more extensive
introduction designed for commutative algebraists is given in [52].
2.A. Additive structure. We may start from the observation that the homotopy theory
of highly connected spaces is simpler than that of general spaces. By suspending a space we
may steadily simplify the homotopy theory, but because cohomology theories have suspension
isomorphisms, we do not lose any additive cohomological information: spectra capture the
limit of this process. Thus spectra are a sort of abelianization of spaces where behaviour
has a more algebraic formal flavour. Associated to any based space X there is a suspension
spectrum, and arbitrary spectra can be built from those of this form. The other important
source of spectra is as the representing objects for cohomology theories. If E ∗ (X) is a
contravariant functor of a based space X which satisfies the Eilenberg-Steenrod axioms,
then there is a spectrum E so that the equation
E ∗ (X) = [X, E]∗
holds. This has the usual benefits that one can then apply geometric constructions to
cohomology theories and one can argue more easily by universal examples.
5
2.B. Multiplicative structure. Having taken the step of representing cohomology theories
by spectra, one may ask if good formal behaviour of the functor E ∗ (·) is reflected in the
representing spectrum E. For our purposes the most important piece of structure is that of
being a commutative ring, and we would like to say that a cohomology theory whose values
on spaces are commutative rings is represented by a spectrum which is a commutative ring
in the category of spectra. This is true remarkably often.
In order to do homotopy theory we need a Quillen model structure on the category of
spectra, and to have commutative rings in this setting we need a symmetric monoidal smash
product, with unit the sphere spectrum S (the suspension spectrum of the two point space
S 0 ), so that the two structures are compatible in a way elucidated by Schwede and Shipley
[91]. In retrospect it seems strange that such models were not constructed until the 1990s,
but several such models are now known, and they give equivalent theories. We sketch the
construction of symmetric spectra in Section 3.
In this context, it makes sense to ask for a cohomology theory to be represented by a
commutative ring spectrum R (i.e., R comes with a multiplication map µ : R ∧ R −→ R
and a unit map S −→ R making R into a commutative monoid). Many of the important
examples do have this structure. The most obvious example from this point of view is the
sphere spectrum S itself. This is the initial ring in the category of spectra, and the smash
product plays the role of tensor product over S. We will describe a number of examples
below, but for the present we continue with the general formalism just assuming that R is a
commutative ring spectrum.
2.C. Modules. We may consider module spectra over R, and there is a model category
structure on R-modules; furthermore, since R is commutative, there is a tensor product of
R-modules formed in the usual way from the tensor product over the initial ring S (i.e.,
as the coequalizer of the two maps M ⊗S R ⊗S N −→ M ⊗S N). From the good formal
properties of the original category, this category of R-modules is again a model category
with a compatible symmetric monoidal product. This has an associated homotopy category
Ho(R − mod) and will be the context in which we work.
2.D. Reverse approach. Commutative algebraists may approach spectra from the algebraic direction. Traditional commutative algebra considers commutative rings A and modules over them, but some constructions make it natural to extend further to considering
chain complexes of A-modules; the need to consider robust, homotopy invariant properties
leads to the derived category D(A). Once we admit chain complexes, it is natural to consider
the corresponding multiplicative objects, differential graded algebras (DGAs). Although it
may appear inevitable, the real justification for this process of generalization is the array of
naturally occurring examples.
The use of spectra is a natural extension of this process. Shipley has shown [92] that
associated to any DGA A there is a ring spectrum HA, so that the derived category D(A)
is equivalent to the homotopy category of HA-module spectra. Accordingly we can view
ring spectra as generalizations of DGAs and categories of module spectra as flexible generalizations of the derived category. Ring spectra extend the notion of rings, module spectra
6
extend the notion of (chain complexes of) modules over a ring, and the homotopy category
of module spectra extends the derived category. Many ring theoretic constructions extend
to ring spectra, and thus extend the power of commutative algebra to a vast new supply
of naturally occurring examples. Even for traditional rings, the new perspective is often
enlightening, and thinking in terms of spectra makes a number of new tools available. Once
again the only compelling justification for this inexorable process of generalization is the
array of interesting, naturally occurring examples, some of which will be described later in
these lectures.
3. Ring and module spectra II (construction)
The purpose of this section is to outline the construction of symmetric spectra. The details
will not be needed for the lectures, and readers familiar with spectra can comfortably omit
it. The point is to reassure readers that spectra are rather concrete objects.
3.A. Naive spectra. This subsection is designed to explain the idea behind spectra: where
they came from and why they were invented. Those already familiar spectra should skip
directly to Subsection 3.B, which describes symmetric spectra.
The underlying idea is that spectra are just stabilised spaces and the bonus is that they
represent cohomology theories. This definition is perfectly good for additive issues, but it
does not have a symmetric monoidal smash product, so is not adequate for commutative
algebra.
Definition 3.1. A spectrum E is a sequence of based spaces Ek for k ≥ 0 together with
structure maps
σ : ΣEk → Ek+1 ,
where ΣEk = S 1 ∧ Ek is the topological suspension. A map of spectra f : E → F is a
sequence of maps so that the squares
Σf
k
ΣEk −→
ΣFk
↓
↓
fk+1
Ek+1 −→ Fk+1
commute for all k.
Example 3.2. If X is a based space we may define the suspension spectrum Σ∞ X to have
kth term Σk X with the structure maps being the identity.
Remark: It is possible to make a definition of homotopy immediately, but this does not work
very well for arbitrary spectra. Nonetheless it will turn out that for finite CW-complexes K,
maps out of a suspension spectrum can be easily described. We reserve [·, ·] for homotopy
classes of maps between spectra (sometimes called ‘stable maps’), so we write [A, B]un for
homotopy classes of (unstable) maps from a based space A to a based space B. With this
notation, we have
[Σ∞ K, E] = lim [Σk K, Ek ]un .
→ k
7
In particular
πn (E) := [Σ∞ S n , E] = lim [S n+k , Ek ]un .
→ k
∞
For example if E = Σ L for a based space L, we obtain a formula for the stable homotopy
groups of L
πn (Σ∞ L) = lim [Σk S n , Σk L]un ,
→ k
By the Freudenthal suspension theorem, this is the common stable value of the groups
[Σk S n , Σk L]un for large k. Thus spectra have captured stable homotopy groups.
Construction 3.3. We can suspend spectra by any integer r, defining Σr E by
(
Ek−r
k−r ≥0
r
(Σ E)k =
pt
k − r < 0.
Notice that if we ignore the first few terms, Σr is inverse to Σ−r . Homotopy groups
involve a direct limit and therefore do not see these first few terms. Accordingly, once we
invert homotopy isomorphisms, the suspension functor becomes an equivalence of categories.
Because suspension is an equivalence, we say that we have a stable category.
Example 3.4. (Sphere spectra) We write S = Σ∞ S 0 for the 0-sphere because of its special
role, and then for an arbitrary integer r we define
S r = Σr S.
Note that S r now has meaning for a space and a spectrum for r ≥ 0, but since we have an
isomorphism S r ∼
= Σ∞ S r of spectra for r ≥ 0 the ambiguity is not important. We extend
this ambiguity, by often suppressing Σ∞ .
Example 3.5. Eilenberg-MacLane spectra. An Eilenberg-MacLane space of type (M, k)
for an abelian group M and k ≥ 0 is a CW-complex K(M, k) with πk (K(M, k)) = M
and πn (K(M, k)) = 0 for n 6= k; any two such spaces are homotopy equivalent. It is well
known that in each degree ordinary cohomology is represented by an Eilenberg-MacLane
space. Indeed, for any CW-complex X, we have H k (X; )M = [X, K(M, k)]un . In fact, this
sequence of spaces, as k varies, assembles to make a spectrum.
To describe this, first note that the suspension functor Σ is defined by smashing with the
circle S 1 , so it is left adjoint to the loop functor Ω defined by ΩX := map(S 1 , X) (based
loops with the compact-open topology). In fact there is a homeomorphism
map(ΣW, X) = map(W ∧ S 1 , X) ∼
= map(W, map(S 1 , X)) = map(W, ΩX)
This passes to homotopy, so looping shifts homotopy in the sense that πn (ΩX) = πn+1 (X).
We conclude that there is a homotopy equivalence
≃
σ̃ : K(M, k) → ΩK(M, k + 1),
and hence we may obtain a spectrum
HM = {K(M, k)}k≥0
8
where the bonding map
σ : ΣK(M, k) → K(M, k + 1)
is adjoint to σ̃. Thus we find
[Σr Σ∞ X, HM] = lim [Σr Σk X, K(M, k)]un = lim H k (Σr Σk X; M) = H −r (X; M).
→ k
→ k
In particular the Eilenberg-MacLane spectrum has homotopy in a single degree like the
spaces from which it was built:
(
M k=0
πk (HM) =
0 k 6= 0.
3.B. Symmetric spectra [74, 89]. Symmetric spectra give an elementary and combinatorial construction of a symmetric monoidal category of spectra. This is excellent for an immediate access to the formal properties, but to be able to calculate with symmetric spectra
and to relate them to the rest of homotopy theory one needs to understand the construction
of the homotopy category. This is somewhat indirect, and Subsection 3.A was intended as a
motivational substitute.
It is usual to give a fully combinatorial construction of symmetric spectra, by basing them
on simplicial sets rather than on topological spaces.
Definition 3.6. (a) A symmetric sequence is a sequence
E0 , E1 , E2 , . . . ,
of pointed simplicial sets with basepoint preserving action of the symmetric group Σn on En .
(b) We may define a tensor product E ⊗ F of symmetric sequences E and F by
_
(E ⊗ F )n :=
(Σn )+ ∧Σp ×Σq (Xp ∧ Yq ),
p+q=n
where the subscript + indicates the addition of a disjoint basepoint.
It is elementary to check that this has the required formal behaviour.
Lemma 3.7. The product ⊗ is symmetric monoidal with unit
(S 0 , ∗, ∗, ∗, . . . ).
Example 3.8. The sphere is the symmetric sequence S := (S 0 , S 1, S 2 , . . . ). Here S 1 =
∆1 /∆1 is the simplicial circle and the higher simplicial spheres are defined by taking smash
powers, so that S n = (S 1 )∧n ; this also explains the actions of the symmetric groups.
It is elementary to check that the sphere is a commutative monoid in the category of
symmetric sequences.
Definition 3.9. A symmetric spectrum E is a left S-module in symmetric sequences.
Unwrapping the definition, we see that this means E is given by
(1) a sequence E0 , E1 , E2 , . . . of simplicial sets,
(2) maps σ : S 1 ∧ En → En+1 , and
(3) basepoint preserving left actions of Σn on En which are compatible with the actions
in the sense that the composite maps S p ∧ En → En+p are Σp × Σn equivariant.
9
Definition 3.10. The smash product of symmetric spectra is
E ∧S F := coeq(E ⊗ S ⊗ F
//
E ⊗ F ).
Proposition 3.11. The tensor product over S is a symmetric monoidal product on the
category of symmetric spectra.
It is now easy to give the one example most important to commutative algebraists.
Example 3.12. For any abelian group M, we define the Eilenberg-MacLane symmetric
spectra. For a set T we write M ⊗ T for the T -indexed sum of copies of M; this is natural for
maps of sets and therefore extends to an operation on simplicial sets. We may then define
˜ 0 , M ⊗S
˜ 1 , M ⊗S
˜ 2 , . . . ), where
the Eilenberg-MacLane symmetric spectrum HM := (M ⊗S
˜ means the tensor product of an abelian group and a simplicial set with the basepoint
⊗
generator set to zero. It is elementary to check that if R is a commutative ring, then HR
is a monoid in the category of S-modules, and if M is an R-module, HM is a module over
HR.
Next one may define the stable model structure on symmetric spectra; this is done in
several steps (Section 3 of [74]), but we may summarise it by saying that the weak equivalences (“stable equivalences”) are maps f : X −→ Y which induce isomorphisms in all
cohomology theories. The homotopy category of symmetric spectra is obtained by inverting
stable equivalences.
4. The three examples
The point of the present section is to introduce our three basic classes of examples. Classical commutative local Noetherian rings, rational cochains on simply connected spaces and
mod p cochains on classifying spaces of finite groups.
The importance of the category of symmetric spectra [74] is that it admits a symmetric
monoidal smash product compatible with the model structrures. Given this, we can start to
do algebra with spectra: choose a ring spectrum R (i.e., a monoid in the category of spectra),
form the category of R-modules or R-algebras and then pass to homotopy. We may then
attempt to use algebraic methods and intuitions to study R and its modules.
4.A. Classical commutative algebra. We explain why the classical derived category is
covered by the context of spectra.
In Section 3 we described a functorial construction for symmetric spectra taking an abelian
group M and giving the Eilenberg-MacLane spectrum HM, which is characterised up to
homotopy by the property that π∗ (HM) = π0 (HM) = M. For symmetric spectra the construction is lax symmetric monoidal, so that if A is a commutative ring, HA is a commutative
ring spectrum. Furthermore the construction gives a functor A-modules −→ HA-modules.
Passage to homotopy groups gives a functor Ho(HA−mod) → Ho(A-modules) = D(A) and
in fact the model categories are equivalent.
Theorem 4.1. (Shipley [92, 1.1, 2.15]) There is a Quillen equivalence between the category
of differential graded Z-algebras and the category of HZ-algebras in spectra.
10
If we choose a DGA A and the corresponding HZ-algebra HA, there is a Quillen equivalence between the category of A-modules and the category of HA-modules, and hence in
particular a triangulated equivalence
D(A) = Ho(A-modules) ≃ Ho(HA-modules) = D(HA)
of derived categories.
The commutative case is complicated by the fact that there are no model structures on
commutative algebras in general. However, if A is commutative, HA may be taken to
be commutative and the Quillen equivalence between A-modules and HA-modules, tensor
products over A and over HA correspond.
We sometimes use Shipley’s result to excuse the omission of the letter H indicating spectra.
In this translation homology in the classical context of chain complexes corresponds to
homotopy in the context of spectra: H∗ (M) = π∗ (HM).
Now that we can view classical commutative rings as commutative ring spectra, we can
attempt to extend classical notions to the context of spectra. From one point of view, we
should first attempt to understand the analogues of local rings before attempting to look at
more geometrically complicated ones. Accordingly, in most of the lectures we will assume
the commutative ring A is local, with residue field k.
4.B. Cochains on a space. In the category of spectra, we may solve the commutative
cochain problem. More precisely, for any space X and a commutative ring k, we may form
the function spectrum C ∗ (X; k) = map(Σ∞ X+ , Hk). It is obviously an Hk-module, and we
may combine the commutative multiplication on Hk with the cocommutative diagonal on
X+ to see that C ∗ (X; k) is a commutative Hk-algebra. The notation is chosen because it is
a model for the cochains in the sense
π∗ (map(Σ∞ X, Hk)) = H ∗ (X; k).
The commutative algebra of C ∗ (X; k) is one of the main topics for these lectures, and we
will omit the coefficient ring k when it is clear from the context.
We then use algebraic behaviour of this commutative ring to pick out interesting classes
of spaces. In accordance with the principle that C ∗ (X; k) is a sort of ring of functions on
X, we simplify terminology and say that X has a property P over k if the commutative ring
spectrum C ∗ (X; k) has the property P.
We will focus particularly on two special cases.
Example 4.2. (Rational case.) We will refer to the special case k = Q with X simply
connected as the rational case. Indeed this may be treated by classical means with the ThomSullivan construction of piecewise polynomial forms giving a commutative DGA model for
C ∗ (X; Q). There is an enormous literature studying rational homotopy theory by homotopy
invariant commutative algebra methods. The book [42] of Félix, Halperin and Thomas is an
excellent starting point. The ‘Looking glass’ paper [12] expounds the philosophy that algebra
and topology are imperfect reflections of each other and gives numerous profound examples
of it. Because we permit more general ring spectra, we are more restricted in theorems but
have a larger range of examples.
11
Example 4.3. (Modular representation theory.) We will also focus particularly on the case
k = Fp , when X = BG for a finite group G. We recall that the classifying space BG of
principal G-bundles is characterised (when G is finite) by the fact that it has fundamental
group G and all other homotopy groups trivial. The interest in this special case comes from
the fact that its coefficient ring
π∗ (C ∗ (BG; Fp )) = H ∗ (BG; Fp ) = Ext∗Fp G (Fp , Fp )
is the group cohomology ring.
There is an operation of p-completion due to Bousfield-Kan [26] which behaves well for
a large class of spaces (the p-good spaces). In particular, all connected spaces with finite
fundamental group are p-good, which will cover our examples. For p-good spaces C ∗ (X; Fp ) =
C ∗ (Xp∧ ; Fp ), and accordingly we will generally assume that X is p-complete when k = Fp .
The space X = BG is already p-complete if G is a p-group, but in general BG∧p has infinitely
many homotopy groups and its funademental group is the largest p-quotient of G.
5. Finiteness conditions
The point of this section is to introduce a variety of finiteness conditions we may impose.
The essential limitation of all we do is that it is based on local rings in commutative algebra.
We will not discuss the new and interesting features that can arise when there are many
maximal ideals.
Context 5.1. The main input is a map R −→ k of ring spectra with notation suggested by
the case when R is commutative local ring with residue field k.
For the most part, we work in the homotopy category Ho(R-mod) of left R-modules.
5.A. New modules from old. Two construction principles will be important to us. The
terminology comes from Dror-Farjoun [30], but in our stable context the behaviour is rather
simpler.
If M is an R-module we say that X is built from M (and write M ⊢ X) if X can be formed
from M by completing triangles, taking coproducts and retracts (i.e., X is in the localizing
subcategory generated by M). We refer to objects built from M as M-cellular, and write
Cell(R, M) for the resulting full subcategory of Ho(R-mod). An M-cellular approximation of
X is a map CellM (X) −→ X where CellM (X) is M-cellular and the map is an HomR (M, ·)equivalence.
We say that X is finitely built from M (and write M |= X) if only finitely many steps and
finite coproducts are necessary (i.e., X is in the thick subcategory generated by M).
Finally, we say that X is cobuilt from M if X can be formed from M by completing
triangles, taking products and retracts (i.e., X is in the colocalizing subcategory generated
by M).
5.B. Finiteness conditions. We say M is small if the natural map
M
M
[M, Nα ] −→ [M,
Nα ]
α
α
12
is an isomorphism for any set of R-modules Nα . Smallness is equivalent to being finitely
built from R. It is easy to see that any module finitely built from R is small. For the reverse
implication we use the fact that we can build an R-cellular approximation Cell(R, M) −→ M;
this is an equivalence, and by smallness, M is a retract of a finitely built subobject of
Cell(R, M).
We sometimes require that k itself is small, but this is an extremely strong condition on
R and it is important to develop the theory under a much weaker condition.
Definition 5.2. [33] We say that k is proxy-small if there is an object K with the following
properties
• K is small (R |= K)
• K is finitely built from k (k |= K) and
• k is built from K (K ⊢ k).
Remark 5.3. Note that the second and third condition imply that the R-module K generates the same category as k using triangles and coproducts: Cell(R, K) = Cell(R, k).
One of the main messages of [33] is that we may use the condition that k is proxy-small
as a substitute for the Noetherian condition in the conventional setting. This rather weak
condition allows one to develop a very useful theory applicable in a large range of examples.
We can illustrate this by looking at the proxy-small condition in the classical case.
Example 5.4. (Conventional commutative algebra) Take R to be a commutative Noetherian
local ring in degree 0, with maximal ideal m and residue field k.
By the Auslander-Buchsbaum-Serre theorem, k is small if and only if R is a regular local
ring, confirming that the smallness of k is a very strong condition. On the other hand, k
is always proxy-small: we may take K = K(α) to be the Koszul complex for a generating
sequence α for m (see Appendix B).
It is shown in [32] that Cell(R, k) consists of objects whose homology is m-power torsion.
Part 2. Morita equivalences
Section 6 introduces the basic apparatus of the Morita equivalences that concern us, along
with their relation with torsion functors and completion. Section 7 discusses the rather weak
condition that k is proxy-small in our examples, and shows it provides an inclusive framework.
Section 8 illustrates the power of the Morita framework by discussing the classification of
ring spectra and DGAs with coefficient rings exterior on one generator.
6. Morita equivalences
Morita theory studies objects X of a category C by considering maps from a test object k.
More precisely, X is studied by considering Hom(k, X) as a module over the endomorphism
ring End(k). In favourable circumstances this may give rather complete information. This
is an instance of the philosophy that one gains insights by looking at rings of operations.
In the classical situation, C is an abelian category with infinite sums and k is a small
projective generator, and we find C is equivalent to the category of End(k)-modules [14, II
13
Thm 1.3]. We will work with a stable model category rather than in an abelian category,
and k will not necessarily be either small or a generator. The fact that the objects of the
categories are spectra is unimportant except for the formal context it provides. See [88] for
a more extended account from the present point of view.
In fact, two separate Morita equivalences play a role: two separate categories of modules
over a commutative ring are both shown to be equivalent to a category of modules over the
same non-commutative ring.
This section is based on [32], with augmentations from [33].
6.A. First variant. To start with we introduce the ring spectrum E = HomR (k, k) of
(derived) endomorphisms of k. This is usually not a commutative ring. Morita theory
considers the relationship between the categories of left R-modules and of right E-modules.
We have the adjoint pair
E : R−mod
o
/
mod−E : T
defined by
T (X) := X ⊗E k and E(M) := HomR (k, M).
It is clear from the definition that E(M) is an E-module, since E = HomR (k, k) acts on the
right through the factor k in EM = HomR (k, M). By the same token, it is clear that the
ring
R̂ = HomE (k, k)
acts on the left of T X = X ⊗E k through its action on k. To see that this gives an R-module
we note that there is a ring map
R −→ HomE (k, k) = R̂
by ‘right multiplication’ of R on k; thuse T X is an R-module by restriction.
Theorem 6.1. If k is small, this adjunction gives equivalence
Cell(R, k) ≃ mod−E
between the derived category of R-modules built from k and the derived category of E-modules.
Proof: To see the unit X −→ ET X = HomR (k, X ⊗E k) is an equivalence, we note it is
obviously an equivalence for X = E and hence for any X built from E, by smallness of k.
The argument for the counit is similar.
Remark 6.2. If k is not small, the unit of the adjunction may not be an equivalence. For
example if R = Λ(τ ) is exterior on a generator τ of degree 1 then E ≃ k[x] is polynomial
on a generator of degree −2. As an R-module, k is of infinite projective dimension and
hence it is not small. In this case all R-modules are k-cellular, so that Cell(R, k) = R-mod.
Furthermore, the only subcategories of R-modules closed under coproducts and triangles
are the trivial category and the whole category. On the other hand the category of torsion
E-modules is a proper non-trivial subcategory closed under coproducts and triangles.
14
Exchanging roles of the rings, so that R = k[x] and E ≃ Λ(τ ), we see k is small as a
k[x]-module and Cell(k[x], k) consists of torsion modules. Thus we deduce
tors-k[x]-mod ≃ mod-Λ(τ ).
In fact the counit
T EM = HomR (k, M) ⊗E k −→ M
of the adjunction is of interest much more generally. Notice that any E-module (such as
HomR (k, M)) is built from E, so the domain is k-cellular. We say M is effectively constructible
from k if the counit is an equivalence, because T EM gives a concrete and functorial model
for the cellular approximation to M. Under the much weaker assumption of proxy smallness
we obtain a very useful conclusion linking Morita theory to commutative algebra.
Lemma 6.3. Provided k is proxy-small, the counit
T EM = HomR (k, M) ⊗E k −→ M
is k-cellular approximation, and hence in particular any k-cellular object is effectively constructible from k.
Proof: We observed above that the domain is k-cellular. To see the counit is a HomR (k, ·)equivalence, consider the evaluation map
γ : HomR (k, X) ⊗E HomR (Y, k) −→ HomR (Y, X).
This is an equivalence if Y = k, and hence by proxy-smallness it is an equivalence if Y = K.
This shows that the top horizontal in the diagram
≃
HomR (k, X) ⊗E HomR (K, k) −→ HomR (K, X)
≃↓
↓=
HomR (K, HomR (k, X) ⊗E k) −→ HomR (K, X)
is an equivalence. The left hand-vertical is an equivalence since K is small. Thus the lower
horizontal is an equivalence, which is to say that the counit
T EX = HomR (k, X) ⊗E k −→ X
is a K-equivalence. By proxy-smallness, this counit map is a k-equivalence.
Examples 6.4. (i) If R is a commutative Noetherian local ring, we recall in Appendix B
that the k-cellular approximation of a module M is Γm M = K∞ (m) ⊗R M, where K∞ (m) is
the stable Koszul complex, so we have
T EM ≃ K∞ (m) ⊗R M.
(ii) If R = C ∗ (X; k) it is not easy to say what the k-cellular approximation is in general,
but any bounded below module M is cellular.
15
6.B. Second variant. There is a second adjunction between the derived categories of left
R-modules and of right E-modules. In the first variant, k played a central role as a left
R-module and a left E-module. In this second variant
k # := HomR (k, R)
plays a corresponding role: it is a right R-module and a right E-module. We have the adjoint
pair
/
E ′ : R−mod o
mod−E : C
defined by
E ′ (M) := k # ⊗R M and C(X) := HomE (k # , X).
Remark 6.5. If k is small then
E ′ (M) = HomR (k, R) ⊗R M ≃ HomR (k, M) = EM,
so the two Morita equivalences consider the left and right adjoints of the same functor.
The unit of the adjunction M −→ CE ′ (M) is not very well behaved, and the functor CE ′
is not even idempotent in general.
6.C. Complete modules and torsion modules. Even when we are not interested in
the intermediate category of E-modules, several of the composite functors give interesting
endofunctors of the category of R-modules.
Lemma 6.6. If k is proxy-small then k-cellular approximation is smashing:
Cellk M ≃ (Cellk R) ⊗R M,
so that
T EM = T E ′ M.
Proof: We consider the map
λ : (Cellk R) ⊗R M −→ R ⊗R M = M
and show it has the properties of cellularization.
To start with, since R builds M, it follows that (Cellk R) ⊗R R builds (Cellk R) ⊗R M, and
hence (Cellk R) ⊗R M is k-cellular.
Since k is proxy small, the map λ is k-equivalence if and only if it is a K-equivalence. We
may consider the class of objects M for which λ is a K-equivalence. Since K is small, it is
closed under arbitrary coproducts as well as triangles. Since the class obviously contains R
itself, it contains all modules M.
We therefore see by 6.3 and 6.6 that if k is proxy-small
Cellk (M) = T EM = T E ′ M.
This is the composite of two left adjoints, focusing attention on its right adjoint CEM, and
we note that
CE(M) = HomR (k # , HomR (k, M)) = HomR (T ER, M).
By analogy with Subsection II of Appendix B, we may make the following definition.
16
Definition 6.7. The completion of an R-module M is the map
M −→ HomR (T ER, M) = CEM.
We say that M is complete if the completion map is an equivalence.
Remark 6.8. By 6.6 we see that completion is idempotent.
We adopt the notation
Γk M := T E ′ M
and
Λk M := CEM.
This is by analogy with the case of commutative algebra through the approach of Appendix
C, where Γk = ΓI is the total right derived functor of the I-power torsion functor and
Λk = ΛI is the total left derived functor of the completion functor (see [2, 3] for the context
of commutative rings).
It follows from the adjunctions described earlier in this section that Γk is left adjoint to
k
Λ as endofunctors of the category of R-modules:
HomR (Γk M, N) = HomR (M, Λk N)
for R-modules M and N. Slightly more general is the following observation.
Lemma 6.9. If k is proxy-small, Γk and Λk give an adjoint equivalence
Cell(R, k) ≃ comp−R-mod,
where comp−R-mod is the triangulated subcategory of R−mod consisting of complete modules.
Proof: We have
T E ′ M ≃ T EM ≃ Γk M ≃ ΓK M,
and
CEM ≃ HomR (Γk R, M) ≃ HomR (ΓK R, M),
so it suffices to prove the result when k is small. When k is small the present adjunction is
the composite of two adjoint pairs of equivalences. We have seen this for the first variant,
and the second variant is proved similarly by arguing that the unit and counit are equivalences.
6.D. dc-completeness. We have explained the importance of passing from R to E =
HomR (k, k), and we have also given an example where we then took E as our input ring,
meaning that we need to consider the ring R̂ = HomE (k, k). Furthermore, the double centralizer ring homomorphism
κ : R −→ HomE (k, k) = R̂
played a role at the very start, in defining the R-module structure on T X.
We say that R is dc-complete if κ is an equivalence.
17
7. Proxy-smallness in the examples
We establish the appropriate finiteness and completeness properties in our principal examples.
7.A. Commutative local rings. When R is a conventional commutative local Noetherian
ring, we have seen that k is always proxy-small, and that the Koszul complex for a set of
generators of m provides the small proxy for k.
Next, we note that
π∗ (E) = π∗ (HomR (k, k)) = Ext∗R (k, k),
which is a ring whose importance is very familiar. We will see in Section 8 that as a DGA,
the endormorphisms E = HomR (k, k) contain considerably more information.
We also note that in this case, R is dc-complete if and only if it is m-complete in the usual
sense. Indeed, we calculate
∧
= holim R/ms
Rm
←
s
←
s
= holim HomE (HomR (R/ms , k), k)
= HomE (holim HomR (R/ms , k), k)
→
= HomE (k, k)
s
Here the first equivalence uses the Mittag-Leffler condition on {R/ms }s , and the last condition uses the fact that
holim HomR (R/ms , k) ≃ Γm k ≃ k.
→
s
More generally, if R is a commutative Noetherian ring and I is an ideal, we may consider
R −→ R/I (i.e., replacing the field k by a more complicated ring). It turns out that the
double centralizer map is I-adic completion whenever R/I is proxy-small.
7.B. Cochains on a space. In the setting R = C ∗ (X; k) of cochains on a space, the
Eilenberg-Moore spectral theorem shows that
E = HomC ∗ (X;k) (k, k) ≃ C∗ (ΩX; k), H∗ (E) = H∗ (ΩX; k)
provided that either (Case 0) X is simply connected or (Case p) X is connected, π1 (X) is a
finite p-group k = Fp and X is p-complete [31]. This immediately gives a very rich source of
examples that we will revisit frequently.
Provided we have either Case 0 or Case p, then C ∗ (X) is dc-complete. Indeed, the
Rothenberg-Steenrod equivalence
HomC∗ (ΩX) (k, k) ≃ C ∗ (X)
holds for any connected space X. Accordingly, provided the Eilenberg-Moore equivalence
holds, the double centralizer map κ is an equivalence.
Finally we may consider proxy-smallness.
Lemma 7.1. (i) If dimk H∗ (ΩY ) is finite then k is small over C ∗ (Y ).
(ii) If Y falls under Case 0 or Case p, there is a map X −→ Y with homotopy fibre F , and
dimk H∗ (ΩY ) and dimk H ∗ (F ) are finite, then k is proxy-small over C ∗ (X).
18
Proof: Since dimk H∗ (ΩY ) is finite,
k |= C∗ (ΩY ).
Applying HomC∗ (ΩY ) (·, k) we see
C ∗ (Y ) ≃ HomC∗ (ΩY ) (k, k) |= HomC∗ (ΩY ) (C∗ (ΩY ), k) ≃ k.
This proves Part (i).
For Part (ii), let R = C ∗ (X), S = C ∗ (Y ) and note C ∗ (X) ⊗C ∗ (Y ) k ≃ C ∗ (F ) by the
Eilenberg-Moore theorem. Now take K = C ∗ (F ). We note that k |= K by hypothesis. Since
S |= k by Part (i), we may apply R ⊗S (·) and deduce
R = R ⊗S S |= R ⊗S k = C ∗ (F ).
Finally, C ∗ (F ) builds k by killing homotopy groups, since F is connected.
Corollary 7.2. (Case 0) If X is simply connected and k = Q and if H ∗ (X) is Noetherian
then Q is proxy small over C ∗ (X; Q).
(Case p) If k = Fp and X is the p-completion of BG then Fp is proxy small over C ∗ (X; Fp ).
Proof: For Case 0, since H ∗ (X) is Noetherian, we may choose a polynomial subring P =
Q[V ] on finitely many even degree generators. We have Q[V ] = H ∗ (Y ) where Y = K(V ) is
the corresponding product of even Eilenberg-MacLane spaces. The inclusion P = K[V ] −→
H ∗ (X) is realized by a map X −→ Y . The space ΩY is a product of finitely many
odd Eilenberg-MacLane spaces, and rationally this is a product of spheres, so Part (i) of
Lemma 7.1 applies. Furthermore since H ∗ (X) is a finitely generated P module, it has a
finite resolution by finitely generated projectives and the spectral sequence for calculating
π∗ (C ∗ (X) ⊗C ∗ (Y ) Q) = H ∗ (F ) has finite dimensional E2 term and so the hypotheses of Part
(ii) of Lemma 7.1 hold.
For Case p we note that the p-completion of Y = BU(n) satisfies the conditions of Part
(i) of Lemma 7.1. Now, given G, choose n so that G admits a faithful representation in U(n)
and apply Part (ii) of Lemma 7.1 to the p-completion of the fibration U(n)/G −→ BG −→
BU(n).
8. Exterior algebras
There are many ways to use Morita theory, but there is a very elementary and striking
one which illustrates that it is significant in even in the very simple case when π∗ (E) is an
exterior algebra on one generator.
8.A. Exterior algebras over Fp on a generator of degree −1. To start with, we note
that is easy to see that if E is an Fp -algebra with π∗ (E) = ΛFp (τ ) exterior on one generator,
then E is formal. However if π∗ (E) = ΛFp (τ ) but E itself is only known to be a Z-algebra,
the situation is considerably more complicated.
19
We may use Morita theory to give a classification [36] in the special case that the exterior
generator is in degree −1. We will return to discuss the more general case in Subsection 8.B.
Theorem 8.1. Differential graded algebras E with H∗ (E) exterior over Fp on a single generator of degree −1 (up to quasi-isomoprhism) are in bijective correspondence with complete
discrete valuation rings with residue field Fp (up to isomorphism of rings).
Proof: The idea is to associate to any such E the endomorphism DGA R = HomE (Fp , Fp ).
Evidently the spectral sequence for calculating H∗ (R) collapses at E2 with value Fp [x] for an
element x of total degree 0. This shows R is a filtered ungraded ring with this as associated
graded ring. One argues that it is commutative and complete with residue field Fp . If one
starts with R, its maximal ideal is principal, generated by an element x and we may form
x
a complex Fp = (R −→ R). The DGA E = HomR (Fp , Fp ) has homology exterior on a
generator of degree −1, and the double centralizer completion map
R −→ HomE (Fp , Fp )
is an equivalence.
8.B. General algebras with exterior homotopy. Let us consider ring spectra E with
π∗ (E) = ΛFp (τ ) where τ is of degree d. We have ring maps S −→ Z −→ Z/p, and we have
already observed that if we restrict attention to Z/p-algebras, there is a unique such E. If
d = −1 we classified Z-algebras in Subsection 8.A, and found there were a lot of them. If
d = 0 the answers for S-algebras and Z-algebras are the same by Shipley’s Theorem, and
there is a unique such algebra.
For d ≥ 1 there is an obstruction theory, and the answer is given by Dugger and Shipley
[38]. Indeed, we are considering square zero extensions of Z/p by the bimodule Z/p in degree
d, and hence (with A = S or Z), the A-algebras E are classified by the Hochschild cohomology
group HH d+2(Z/p|A; Z/p), with orbits under Aut(Z/p) giving isomorphic algebras E. In
fact the Hochschild cohomology rings are polynomial and divided power algebras on a single
degree 2 generator:
HH ∗(Z/p|Z; Z/p) = Z/p[xZ2 ] and HH ∗(Z/p|S; Z/p) = ΓZ/p [xS2 ].
We conclude that for any odd d ≥ 1 there is single formal algebra and for any even d ≥ 2 there
are precisely two isomomorphism classes of Z-algebras and precisely two isomomorphism
classes of S-algebras. In each case one of them is formal and one is not.
The really striking phenomenon is that restriction along S −→ Z, does not induce a
bijection of isomorphism types. Indeed, one may check that the generator xZ2 restricts nontrivially (so we may take xS2 to be that restriction). It follows that there is a bijection until
we reach (xZ2 )p , which restricts to zero in Γ(xS2 ). In other words, the two Z-algebras E, E ′ with
d = 2p − 2 become equivalent as S-algebras: these two inequivalent DGAs are topologicially
equivalent. It is hard to be explicit when d is not small, but if p = 2 and d = 2, the DGA
Z[e | de = 2]/(e3 , 2e2 ) is not formal as a Z-algebra but it is formal as an S-algebra. Finally,
the S-algebra E with d = 2p − 2 which is not realizable as a Z-algebra, is the truncation of
the dth Morava K-theory.
20
8.C. Classification of free rational G-spectra. In another direction, working over the
rationals we note that for any compact Lie group G the DGA C∗ (G) has homology H∗ (G)
exterior on odd degree generators. It might not be apparent that it is formal. However, we
may consider R = HomC∗ (G) (k, k) ≃ C ∗ (BG); since this is commutative and has cohomology
which is polynomial on even degree generators, it is formal and hence equivalent to H ∗ (BG).
It follows that C∗ (G) ≃ H∗ (G). This is the starting point for classifying free rational Gspectra [61, 62].
Part 3. Regular rings
We have now set up the basic machinery and the rest of the lectures will investigate
homotopy invariant counterparts of classical definitons in commutative algebra of local rings.
The justification consists of the dual facts that the new definition reduces to the old in the
classical setting and that the new definition covers and illuminates examples in new contexts.
Regular local rings are the most basic and best behaved objects, so it makes sense to start
with these.
Section 9 recalls the three styles of definition from commutative algebra and explains how
to give homotopy invariant versions, and then illustrates them in our main examples. Section
10 explains how we can use this class of ring spectra to give a homotopy invariant counterpart
of the notion of a finitely generated module, at least for ‘normalizable’ ring spectra.
9. Regular ring spectra
In commutative algebra there are three styles for a definition of a regular local ring: ideal
theoretic, in terms of the growth of the Ext algebra and a version for modules. We begin in
Subsection 9.A by recalling the commutative algebra, where the conditions are equivalent.
We then turn to other contexts, where the conditions may differ, and consider which one is
most appropriate.
9.A. Commutative algebra. The following definitions are very familiar; we have introduced a slightly more elaborate terminology to smooth the transition to other contexts. The
prefix s- signifies that the definition is structural, the prefix g- signifies that the definition is
in terms of the growth rate, and the prefix m- signifies that the definition is in terms of the
module category.
Definition 9.1. (i) A local Noetherian ring R is s-regular if the maximal ideal is generated
by a regular sequence.
(ii) A local Noetherian ring R is g-regular if Ext∗R (k, k) is finite dimensional.
(iii) A local Noetherian ring R is m-regular if every finitely generated module is small in
the derived category D(R).
These are equivalent by the Auslander-Buchsbaum-Serre theorem; one might think of this
as the first theorem of homotopy invariant commutative algebra.
21
9.B. Regularity for ring spectra. The one definition that is easy to adapt is g-regularity,
at least when k is a field. This is a basic input to the entire theory. It is also convenient to
have a name for when the coefficient ring is regular.
Definition 9.2. (i) We say that R is c-regular when the coefficient ring π∗ (R) is regular.
(ii) We say that R is g-regular if π∗ (HomR (k, k)) is finite dimensional over k.
It is obvious from the spectral sequence
Ext∗,∗
R∗ (k, k) ⇒ π∗ (HomR (k, k))
that c-regular implies g-regular. The converse is very far from being true, as we shall see
shortly.
It is more subtle to consider s-regular ring spectra.
The s-regularity condition on rings states that m is generated by a regular sequence
x1 , . . . , xr , so that there are short exact sequences
x
1
R −→
R −→ R/(x1 )
x
2
R/(x1 ) −→
R/(x1 ) −→ R/(x1 , x2 )
and
x
r
R/(x1 , . . . , xr−1 ) −→
R/(x1 , . . . , xr−1 ) −→ R/m = k.
In other words, we may start with R, and successively factor out a regular element until we
get to k.
If we now remember the degrees, a single instance is the (additive) exact sequence of
modules
x
Σn R −→ R −→ R/(x).
x
In other words, the ring R/(x) is equivalent to the Koszul complex K(x) = (R −→ R). If
we think multiplicatively this gives a cofibre sequence 1
R −→ K(x) −→ K(x) ⊗R k = Λ(τ )
of rings, where τ is a generator of degree n + 1. The definition is now obtained by iterating
this construction.
Definition 9.3. We say that R is s-regular if there are cofibre sequences of rings
R = R0 −→ R1 −→ Λ1 , R1 −→ R2 −→ Λ2 , · · · , Rr−1 −→ Rr −→ Λr ,
with Rr ≃ k and π∗ (Λi ) exterior over k on one generator.
Lemma 9.4. If R is s-regular then it is also g-regular.
1If
S −→ R is a map of rings and we have a map R −→ k, the cofibre ring is Q = R⊗Q k. This corresponds
to the fact that in algebraic geometry and topology this cofibre ring is often the ring of functions on the
geometric fibre.
22
Proof : We need to work our way along the sequence of cofibrations. Since Rr ≃ k is
obviously g-regular, it suffices to oberve that given a cofibre sequence C −→ B −→ Λ with
B g-regular, then C is also g-regular. Suppose then that HomB (k, k) is finitely built from k.
Since k |= Λ it follows that
k |= HomB (k, k) |= HomB (Λ, k) ≃ Homk (k ⊗B Λ, k).
However
k ⊗B Λ ≃ k ⊗B B ⊗C k ≃ k ⊗C k.
so that
Homk (k ⊗B Λ, k) ≃ Homk (k ⊗C k, k) ≃ HomC (k, k),
showing HomC (k, k) is finitely built by k as required.
However, there is no good notion of m-regularity for ring spectra. The problem is that we
do not have a homotopy invariant definition of finite generation in general. However, we will
turn this around, and in Section 10 we will show that one can use the supposed equivalence
of g-regularity and the putative m-regularity conditions to define finite generation.
9.C. Regularity for rational spaces. We have seen in Subsection 7.B, that for simply
connected rational spaces X, if R = C ∗ (X; Q) we have π∗ (E) = H∗ (ΩX). Accordingly a
g-regular rational space is one with H∗ (ΩX) finite dimensional.
Proposition 9.5. A simply connected rational space X is g-regular if and only if it is a
finite product of even Eilenberg-MacLane spaces. Its cohomology is therefore polynomial on
even degree generators and it is also s-regular.
Proof : To start with, we note that an even Eilenberg-MacLane space X = K(Q, 2n) is
g-regular since its loop space is K(Q, 2n − 1) ≃ S 2n−1 . On the other hand an odd EilenbergMacLane space X = K(Q, 2n + 1) is not g-regular since its loop space is K(Q, 2n) which
has homology polynomial on one generator of degree 2n.
Finally, since ΩX splits as a product of Eilenberg-MacLane space for any rational space,
we see H∗ (ΩX) is finite dimensional if and only if the homotopy groups of X are finite
dimensional and in even degree. Building X from its Postnikov tower we see inductively
that the cohomology of its Postnikov sections are entirely in even degrees (and polynomial).
Since the homotopy of X is in even degrees the k-invariants are in odd degrees and hence
zero. This shows that X is a finite product of even Eilenberg-MacLane spaces.
9.D. g-regularity for p-complete spaces. Of the commutative algebra definitions, the
only one with a straightforward counterpart for C ∗ (X) is g-regularity. We have seen in
Subsection 7.B, that for connected p-complete spaces X with finite fundamental group, if
R = C ∗ (X; Q) we have π∗ (E) = H∗ (ΩX). Accordingly a g-regular rational space is one with
H∗ (ΩX) finite dimensional.
23
Example 9.6. For any finite p-group G, the classifying space X = BG is p-complete, and
hence satisfies our hypotheses without further completion. Thus ΩX is equivalent to the
finite discrete space G, so E ≃ kG and X is g-regular.
On the other hand the coefficient ring H ∗ (BG) is very rarely regular, so this gives many
examples of g-regular ring spectra which are not c-regular.
Example 9.7. Extending this idea, if G is any compact Lie group with component group a
finite p-group we have
Ω(BG∧p ) ≃ (ΩBG)∧p ≃ G∧p ,
so that X = (BG)∧p is again g-regular over Fp .
In fact X = (BG)∧p is g-regular if and only if π0 (G) is p-nilpotent, and we will see explicit
examples later where completed classifying spaces are not g-regular.
In fact this is the starting point of Dwyer and Wilkerson’s project [37] to capture properties
of groups at a single prime homotopically. They define a connected p-complete space X to
be the classifying spaces of a connected p-compact group if H∗ (ΩX; Fp ) is finite dimensional
(i.e., if C ∗ (X; Fp ) is g-regular). In other words, a p-compact group is an Fp -finite loop space
whose classifying space is p-complete and g-regular.
A major programme involving many people has given a classification of connected pcompact groups following the lines of that for compact Lie groups. This starts with the
theorem of Dwyer and Wilkerson that connected p-compact groups have a maximal torus,
and has finally culminated in the classification of connected p-compact groups in terms of
the associated p-adic root data (Anderson, Grodal, Møller and Viruel ([5] for odd primes,
and [4] for p = 2). At the prime 2 there is only one simple 2-compact group which is not
obtained by completing a compact connected Lie group. At odd primes there are many
exotic examples, because of the many p-adic reflection groups. It is remarkable that objects
defined purely in terms of a finiteness condition can be completely classified.
9.E. s-regularity for p-complete spaces. Looking back at the definition of s-regularity
for ring spectra, the basic ingredient is to consider the multiplicative sequence
R −→ K(x) −→ K(x) ⊗R k = Λ(τ )
where τ is a generator of degree n+1. Since we are working with k-algebras, exterior algebras
are formal, so if n ≥ 0, we have Λ(τ ) ≃ C ∗ (S n+1 ). Thus if R = C ∗ (X), the Eilenberg-Moore
theorem shows that the most obvious source of such cofibre sequence is a spherical fibration
X ←− X1 ←− S n+1 .
It is natural when talking about spaces to restrict the cofibrations to be of this type
Definition 9.8. A space X is s-regular if there are k-complete fibrations
X = X0 ←− X1 ←− S n1 , X1 ←− X2 ←− S n2 , . . . , Xr−1 ←− Xr ←− S nr
with Xr ≃ ∗ and ni ≥ 1.
Examples of this type seem very rare, but do arise from the classical infinite families of
compact Lie groups.
24
Example 9.9. X = BU(n) is s-regular, in view of the fibrations
BU(n) ←− BU(n − 1) ←− U(n)/U(n − 1) ∼
= S 2n−1 .
Similarly for BSO(n), BSpin(n), BSp(n).
10. Finite generation
In conventional algebra, finite generation is one of the most important finiteness conditions.
However it is not all obvious how best to translate this into a homotopy invariant context.
In this section we describe an approach we have found useful, which is based on the idea
that finite generation and smallness coincide over regular rings.
10.A. Finiteness conditions. If R is a conventional Noetherian commutative ring and M
is a module over it, there are a number of natural finiteness conditions. If M has a finite
resolution by finitely generated projectives then it is finitely built from R and hence small
in the derived category; the converse is also true, since one may factor the identity map
through a finite truncation of the projective resolution. These two equivalent homotopy
invariant finiteness conditions on M are extremely useful.
On the other hand, it is not so clear how to give a homotopy invariant version of the
notion of being finitely generated. In this section we discuss one useful method. It is based
on the fact that for a regular conventional ring Q finitely generated modules coincide with
small modules.
10.B. Normalizable ring spectra. In commutative algebra, it is natural to assume rings
are Noetherian, and one of the most useful consequences for k-algebras is Noether normalization, stating that a Noetherian k-algebra is a finitely generated module over a polynomial
subring.
Definition 10.1. We will say that a ring spectrum R is g-normalizable, if there is a gregular ring spectrum S with S∗ Noetherian, and a ring map S −→ R making R into small
S-module. In this case S −→ R is called a g-normalization and R ⊗S k is its Noether fibre.
If S can be chosen so that its coefficient ring S∗ is regular, we say R is c-normalizable.
As usual, with coefficient-level conditions, c-normalizations are a bit rigid, but they do
give a template for comparison.
Lemma 10.2. (i) If S∗ is regular then an S-module N is small if and only if N∗ is a finitely
generated S∗ -module.
(ii) If R is c-normalizable then R∗ is Noetherian.
Proof: Part (ii) follows from Part (i), so we prove Part (i).
It is easy to see that we have a cofibre sequence X −→ Y −→ Z of S-modules, if two
terms have finitely generated homotopy, so does the third. Thus any small module has
finitely generated homotopy.
For the converse, we start by noting that projectives are realizable. Next, if P∗ is projective, HomS (P, M) = HomS∗ (P∗ , M∗ ). Now for any M we may choose a projective S∗ -module
mapping onto M∗ and realize it by a map P −→ M with fibre F . If M∗ was not projective
25
to start with, F∗ will be of lower projective dimension. Since S∗ is regular, this process
terminates.
Example 10.3. (Venkov) If G is a compact Lie group (for example a finite group), then
C ∗ (BG) is c-normalizable. Indeed, we may choose a faithful representation G −→ U(n),
giving a fibration U(n)/G −→ BG −→ BU(n).
It is an unpublished consequence of work of Castellana and Ziemianski that every pcompact group has a faithful linear representation, which is to say that if BΓ is regular there
is a map BΓ −→ BSU(n), for some n, whose homotopy fibre is Fp -finite. This means that
every p-compact group is c-normalizable.
10.C. Finitely generated modules. We are now equipped to give our definition.
Definition 10.4. Suppose that R is a commutative ring spectrum and M is an R-module.
If we are given a normalization ν : S −→ R, we say that M is ν-finitely generated if M is
small over S.
If R is g-normalizable, we say that M is finitely generated if M is ν-finitely generated for
every g-normalization ν.
Some elementary consequences follow as in Lemma 10.2.
Lemma 10.5. (i) If R is g-normalizable and M is finitely generated then M∗ is a finitely
generated R∗ -module.
(ii) If ν : S −→ R is a c-normalization, then M is ν-finitely generated if and only if M∗
is finitely generated over R∗ .
Remark 10.6. (a) If R is c-normalizable, it is natural to refer to the class of modules M
for which M∗ if finitely generated over R∗ as c-finitely generated.
(b) This raises the question of whether there are c-finitely generated modules which are
not finitely generated. In other words, if R is a c-normalizable g-regular ring and M is an
R-module with M∗ finitely generated over R∗ does it follow that M is small? It would suffice
to show that R has a c-normalization S which is an S-module retract of R, as happens when
S∗ −→ R∗ is a monomorphism and R∗ is Gorenstein.
Proof: (i) If M is small over S for some normalization, then M∗ is finitely generated over
S∗ and hence over R∗ .
(ii) If M∗ is finitely generated over R∗ it is finitely generated over S∗ . Since S∗ is regular,
this means M is small over S and hence finitely generated.
10.D. Singularity categories. In conventional commutative algebra the Buchweitz singularity category Dsing (R) is defined to be the quotient of the bounded derived category by the
subcategory of perfect complexes.
If we are given a normalization ν of a commutative ring spectrum R, it is evident that
the category of ν-finitely generated modules is a triangulated subcategory of the category
26
of R-modules, and similarly for the finitely generated modules. Evidently for any particular
g-normalization ν we have full and faithful embeddings
Dsmall (R) ⊆ Df g (R) ⊆ Dν−f g (R) ⊆ Dc−f g (R)
of triangulated categories each closed under retracts. We could consider any one of the
quotients, but the most natural is the initial quotient
Dsing (R) = Df g (R)/Dsmall (R).
This has the merit that if R is itself g-regular, the quotient is trivial, and in many cases one
can show that Df g (R) = Dc−f g (R).
The singularity quotients are understood in many algebraic cases, but for the present we
will restrict ourselves to an example where we can see from first principles that the quotient
is non-trivial. See [63] for further discussion.
Example 10.7. If R = C ∗ (BA4 ) with k = F2 then M = C ∗ (BV4 ) is c-finitely generated
since H ∗ (BV4 ) is finitely generated over H ∗ (BA4 ) by Venkov’s Theorem. However if we let
F = fib(BV4 −→ BA4 ) there is a fibre sequence ΩS 3 −→ F −→ SO(3)/V4 which may be
used to check H ∗ (F ) is non-zero in infinitely many degrees. Thus C ∗ (BV4 ) is not small over
C ∗ (BA4 ).
Part 4. Hypersurface rings
This is the second of a series of parts that take particular classes of commutative local
rings, and identifies ways of giving homotopy invariant counterparts of the definitions, which
apply to ring spectra. The justification consists of the dual facts that the new definition
reduces to the old in the classical setting and that the new definition covers and illuminates
examples in new contexts.
Although we restrict attention to hypersurface rings here, we begin in Section 11 with a
brief sketch of relevant facts about complete intersections in general to set the context. In
Section 12 we describe the algebraic theory of hypersurface rings in a bit more detail. In
Section 13 we explain the relevance of bimodules and Hochschild cohomology. In view of
the discussion in Part 3, it is then quite routine in Section 14 to provide the definitions of
ci ring spectra, and we illustrate some of the results with examples. In Sections 15 and 16
we explain how the three homotopy invariant versions of the ci condition are related to each
other.
11. Complete intersections
To understand the significance of hypersurfaces, we should first say a word about complete
intersections. There is a general framework for discussing these in the homotopy invariant
context [55, 22], but there are few examples beyond the hypersurface case. The general
theory is in some sense obtained by iterating the case of hypersurfaces, but there are a
number of different ways of iterating it, and some work better than others.
27
11.A. Classical complete intersections. These lectures will focus on hypersurfaces, but
it is helpful to set the more general context with a brief discussion of complete intersections
in general. Just as for regular rings there are three styles of definition: (s) a structural one (g)
one involving growth and (m) one involving the homological algebra of modules. These all
have counterparts in the homotopy invariant setting which we will introduce in due course.
We start with the structural definition. The best behaved subvarieties of affine space are
those which are specified by the right number of equations: if they are of codimension c
then only c equations are required. On the same basis, a commutative local ring R is a
structurally complete intersection (sci) if its completion is the quotient of a regular local ring
S by a regular sequence, f1 , f2 , . . . , fc (this would normally just be called ‘ci’, but the longer
name eases comparison with the case of ring spectra). We will suppose that R is complete,
so that
R = S/(f1 , . . . , fc ).
The smallest possible value of c (as S and the regular sequence vary) is called the codimension
of R. A hypersurface is the special case when c = 1 so that a single equation is used.
When it comes to growth, if R is ci of codimension c, one may construct a resolution of
any finitely generated module growing like a polynomial of degree c − 1. In particular the
ring Ext∗R (k, k) has polynomial growth (we say that R is gci). Perhaps the most striking
result about ci rings is the theorem of Gulliksen [67] which states that this characterises ci
rings so that the sci and gci conditions are equivalent for local rings.
With a little care, one may construct the resolutions in an eventually multi-periodic fashion: the projective resolution eventually has the pattern of the tensor product of c periodic
exact sequences. In fact the construction is essentially independent of the module and the
calculation can be phrased in terms of the Hochschild cohomology of R. This opens the way
to the theory of support varieties for modules over a ci ring [9].
11.B. Homotopy invariant versions, and Levi’s groups. One may give homotopy invariant versions of all three characterizations of the ci condition:
(sci): the ‘regular ring modulo regular sequence’ condition,
(mci): the ‘modules have eventually multiperiodic resolutions’ and
(gci): polynomial growth of the Ext algebra
We note that the Avramov-Quillen characterization of ci rings in terms of André-Quillen
homology does not work for cochains on a space in the mod p context since Mandell has
shown [80] that the topological André-Quillen cohomology vanishes very generally in this
case.
It involves some work to describe the sci and mci definitions, but if we take R = C ∗ (X)
for a p-complete space Subsection 7.B shows H∗ (HomC ∗ (X) (k, k)) = H∗ (ΩX), so it is easy
to understand the gci condition. Taking X = (BG)∧p we may consider what this means. Of
course if G is a p-group, X = BG so ΩX ≃ G and H∗ (ΩX) is simply the group ring kG in
degree 0 and therefore finite dimensional. More generally, it is known to be of polynomial
growth in certain cases (for instance A4 or M11 in characteristic 2) and R. Levi [75, 76, 77, 78]
has proved there is a dichotomy between small growth and large growth, and given examples
28
where the growth is exponential. Evidently groups whose p-completed classifying spaces
have loop space homology that has exponential growth cannot be spherically resolvable, so
Levi’s groups disproved a conjecture of F.Cohen.
Example 11.1. (a) It is amusing to consider the p-completed classifying space BG∧p for
G = (Cp × Cp ) ⋊ C3 where C3 acts via (1, 0) 7→ (0, 1) 7→ (−1, −1). When p = 3, G is a
p-group so the space is g-regular. When p = 2 the group is the alternating group A4 , which
we will see below is a hypersurface. If p ≥ 5 then Levi shows H∗ (Ω(BG∧p )) has exponential
growth by showing that Ω(S 5 ∪p e6 ) is a retract.
(b) If the reader prefers an example with a purely algebraic treatment, Benson shows in
[15, Example 2.2] that the group (C3 )2 ⋊ C2 has a 3-completed classifying space whose loop
space has exponential growth using the methods sketched in Subsection 14.E below.
12. Hypersurfaces in algebra
In this section we recall some standard constructions for hypersurface algebas. We suppose
R = S/(f ) is a hypersurface ring, where S is a regular ring and f is a nonzero element of
degree d. Thus we have a short exact sequence
f
0 −→ Σd S −→ S −→ R −→ 0
of S-modules for a regular local ring S. There are two basic constructions that we need to
generalize.
12.A. The degree 2 operator. We describe a construction of a cohomological operator
due to Gulliksen [68, 7]. We will do this for a single module, but it is apparent that the
construction is essentially independent of the module, and in fact it lifts to Hochschild
cohomology.
Given an R-module M we may apply (·) ⊗S M to the defining sequence to obtain the short
exact sequence
f
0 −→ TorS1 (R, M) −→ Σd M −→ M −→ R ⊗S M −→ 0.
Since f = 0 in any R-module, we conclude
TorS1 (R, M) ∼
= Σd M, R ⊗S M ∼
= M,
and
TorSi (R, M) = 0 for i ≥ 2.
Next, choose a free S-module F with an epimorphism to M, giving a short exact sequence
0 −→ K −→ F −→ M −→ 0
of S-modules. Applying (·) ⊗S R, we obtain
0 −→ TorS1 (R, M) −→ K −→ F −→ M −→ 0.
Since TorS1 (R, M) ∼
= Σd M, this gives an element
χf ∈ Ext2R (M, Σd M),
29
or a map
χf : M −→ Σd+2 M
in the derived category of R-modules. We sketch below how this construction lifts to give
an element
χf ∈ HH d+2(R|S),
and hence in particular that it gives a natural transformation of the identity functor (i.e.,
an element of the centre ZD(R) of D(R)).
12.B. The eventually periodic resolution. Continuing with the above discussion, we
may show that all modules M have free resolutions over R which are eventually periodic of
period 2.
Indeed, M has a finite free S-resolution
0 −→ Fn −→ Fn−1 −→ · · · −→ F1 −→ F0 −→ M −→ 0.
Adding an extra zero term if necessary, we suppose for convenience that n is even. Now
apply (·) ⊗S R to obtain a complex
0 −→ F n −→ F n−1 −→ · · · −→ F 1 −→ F 0 −→ M −→ 0.
Since TorSi (R, M) = 0 for i ≥ 2, this is exact except in homological degree 1, where it is
TorS1 (R, M) ∼
= Σd M. Splicing in a second copy of the resolution, we obtain a complex
Fn
−→ F n−1 −→ · · · −→ F 2 −→ F 1 −→ F 0 −→ M −→ 0
0
−→
⊕
⊕
⊕
⊕
ր
Σd F n−1 −→ Σd F n−2 −→ Σd F n−3 −→ · · · −→ Σd F 0
which is exact except in the second row, in homological degree 4, where the homology is
again M. We may repeatedly splice in additional rows to obtain a free resolution
. . . −→ G3 −→ G2 −→ G1 −→ G0 −→ M −→ 0
over R. Remembering the convention that n is even, provided the degree is at least n, the
modules in the resolution are (up to suspension by a multiple of d)
G2i = F n ⊕ Σd F n−2 ⊕ · · · ⊕ Σd(n−2)/2 F 2 ⊕ Σdn/2 F 0
in even degrees and
G2i+1 = F n−1 ⊕ Σd F n−3 ⊕ · · · ⊕ F 3 ⊕ Σd(n−2)/2 F 1
in odd degrees.
12.C. Smallness. We may reformulate the eventual periodicity of the previous subsection
in homotopy invariant terms.
Lemma 12.1. If R is a hypersurface R = S/(f ) and M is a finitely generated R-module
then the mapping cone of χf : M −→ Σd+2 M is small.
30
Proof: From the Yoneda interpretation, we notice that
χf : M −→ Σd+2 M
is realized by the quotient map factoring out the first row subcomplex
0 −→ F n −→ F n−1 −→ · · · −→ F 1 −→ F 0 −→ 0.
Thus the short exact sequence
0 −→ F • −→ G• −→ Σd+2 G• −→ 0
of R-free chain complexes realizes the triangle
Σ1 M/χ −→ M −→ Σd+2 M.
12.D. Matrix factorizations. Because of the importance of matrix factorizations and their
prominence in the IRTATCA programme, it is worth a brief subsection to make the connection. We have considered resolutions of R-modules M, and we note that if M is of projective
dimension 1 over S then the S-resolution
A
0 ←− M ←− F0 ←− F1 ←− 0
gives a periodic resolution
0 ←− M ←− F 0 ←− F 1 ←− F 0 ←− F 1 ←− F 0 ←− F 1 ←− · · ·
over R.
Since f = 0 on M, multiplication by f on the S-resolution is null-homotopic and so there
is a diagram
0o
Mo
0
0o
Mo
F0 o
A
F1 o
0
⑥⑥
f ⑥⑥⑥
f
~⑥⑥ B
F0 o
F1 o
0
A
In short we have two maps A : F1 −→ F0 and B : F0 −→ F1 with AB = f · id, BA = f · id.
Since F0 and F1 are free modules, we can choose bases and represent A and B by matrices,
so this structure is known as a matrix factorization.
If we suppose S is of dimension d, R is of dimension d − 1. We note that the homological
dimension of M over S can be expressed as an invariant of M as an R-module. In our case
M is of projective dimension 1 over S and by the Auslander-Buchsbaum formula, it is of
depth d − 1 as an R-module, which is to say it is a maximal Cohen-Macaulay module. For a
module M of depth d − 1 − i (i.e., of projective dimension i + 1 over S) the above discussion
applies to its pth syzygy.
13. Bimodules and natural endomorphisms of R-modules
To describe the m-version of the definition of hypersurface rings, we need to briefly discuss
bimodules.
31
13.A. The centre of the derived category of R-modules. If R is a commutative Noetherian ring and M is a finitely generated R-module with an eventually n-periodic resolution,
there is a map M −→ Σn M in the derived category whose mapping cone is a small R-module.
If R is an ungraded hypersurface ring, all finitely generated modules have such resolutions
with n = 2. The lesson learnt from commutative algebra is that to use this to characterize
hypersurfaces we need to look at all such modules M together, and ask for a natural transformation 1 −→ Σn 1 of the identity functor. By definition the centre ZD(R) is the graded
ring of all such natural transformations. There are various ways of constructing elements
of the centre, and various natural ways to restrict the elements we consider. Some of these
work better than others, and it is the purpose of this section is to introduce these ideas.
13.B. Bimodules. We consider a map S −→ R, where S is regular and R is small over S.
We may then consider Re = R ⊗S R, and Re -modules are (R|S)-bimodules. The Hochschild
cohomology ring is defined by
HH ∗ (R|S) = Ext∗Re (R, R).
If f : X −→ Y is a map of (R|S)-bimodules, for any R-module M we obtain a map
f ⊗ 1 : X ⊗R M −→ Y ⊗R M of (left) R-modules.
The simplest way for us to use this is that if we have isomorphisms X ∼
= Σn R
= R and Y ∼
as R-bimodules, the map f ⊗ 1 : M −→ Σn M is natural in M and therefore gives an element
of codegree n in ZD(R): we obtain a map of rings
HH n (R) = HomRe (R, Σn R) −→ ZD(R)n .
Continuing, if X |=Re Y then X ⊗R M |=R Y ⊗R M. In particular, if X = R builds a
small Re -module Y then
M = R ⊗R M |=R Y ⊗R M =|Re ⊗R M = R ⊗S M.
Thus if M is finitely generated (i.e., small over S), this shows M finitely builds a small
R-module.
It is sometimes useful to restrict the maps permitted in showing that X |=Re Y . If we
restrict to using maps of positive codegree coming from Hochschild cohomology, we write
X |=hh Y . If we permit any maps of positive codegree from the centre ZD(Re ) we write
X |=z Y . Finally, we could relax further and require only that all the maps involved in
building are endomorphisms of non-zero degree for some object and write X |=e Y . It is
this last condition that turns out to be most convenient for complete intersections of larger
codimension.
14. Hypersurface ring spectra
We are now ready to describe the s-, g- and m- versions of the hypersurface condition for
ring spectra. As usual we begin with the template in commutative algebra, go on to describe
it in general and then make it concrete for spaces.
32
14.A. The definition in commutative algebra. In commutative algebra there are three
styles for a definition of a hypersurace ring: ideal theoretic, in terms of the growth of the
Ext algebra and a derived version. See [21] for a more complete discussion.
Definition 14.1. (i) A local Noetherian ring R is an s-hypersurface ring if R = S/(f ) some
regular ring S and some f 6= 0.
(ii) A local Noetherian ring R is a g-hypersurface if the dimensions dimk (ExtnR (k, k)) are
bounded independently of n.
(iii) A local Noetherian ring R is a z-hypersurace if there is an elements z ∈ ZD(R) of
non-zero degree so that M/z is small for all finitely generated modules M. Similarly R is an
hh-hypersurface if the element z can be chosen to come from Hochschild cohomology.
Theorem 14.2. ([7], transcribed into the present language in [21]) For a local Noetherian
ring the s-, z- and g-hypersurface conditions are all equivalent.
14.B. Definitions for ring spectra. The appropriate definition of an s-hypersurface can
be seen from the discussion of s-regular spectra: the point is that regular elements correspond
to maps with exterior cofibres.
Definition 14.3. (i) A ring spectrum R is a c-hypersurface if R∗ is a hypersurface ring.
(ii) The ring spectrum R is an s-hypersurface if there is a normalization S −→ R with
π∗ (R ⊗S k) = Λk (τ ).
(iii) The ring spectrum R is a g-hypersurface if dimk (πn (HomR (k, k))) is bounded independently of n.
(iv) The c-normalizable ring spectrum R is a z-hypersurace if there is an element z ∈ ZD(R)
of non-zero degree so that M/z is small for all finitely generated modules M. Similarly R is
an hh-hypersurface if the element z can be chosen to come from Hochschild cohomology.
One immediate source of examples comes from c-hypersurfaces.
Example 14.4. If R is a c-hypersurface then the spectral sequence
Ext∗,∗
R∗ (k, k) ⇒ π∗ (HomR (k, k))
shows that it is a g-hypersurface.
14.C. Definitions for spaces. In view of the fact that regular elements correspond to
spherical fibrations, adapting the above definitions for spaces is straightforward.
Definition 14.5. (i) A space X is a c-hypersurface if H ∗ (X) is a hypersurface ring.
(ii) A space X is an s-hypersurface (or spherical hypersurface) if it is the total space of a
spherical fibration over a connected g-regular space BΓ: there is a g-regular space BΓ and
a fibration
S n −→ X −→ BΓ
for some n.
(iii) A space X is a g-hypersurface space if H ∗(X) is Noetherian and dimk Hn (ΩX) is
bounded independently of n.
33
(iv) A space X is a z-hypersurface space if X is g-normalizable and there is an element
z ∈ ZD(C ∗ (X)) of non-zero degree so that C ∗ (Y )/z is small for all finitely generated C ∗ (Y ).
It is an hh-hypersurface if z comes from Hochschild cohomology. (The direct transcription
for ring spectra would require that this holds for all finitely generated modules and not just
those of the particular form C ∗ (Y )).
Remark 14.6. Other variants have arisen, such as ωsci where we are permitted to use
loop spaces on spheres rather than spheres. These conditions arose in Levi’s work. This is
evidently a weakening of sci which still implies gci.
For g-normalizable spaces X, we have [55, 22] the implications
s − hypersurf ace ⇒ z − hypersurf ace ⇒ g − hypersurf ace.
Over Fp the final implication may be reversed; the proof is given in [22] and relies on [43].
14.D. An example. To start with we may consider the space BA4 at the prime 2. We will
observe directly that it is a hypersurface according to any one of the definitions.
To start with, we note that
H ∗ (BA4 ) = H ∗ (BV4 )A4 /V4 = k[x2 , y3, z3 ]/(r6 )
where r6 = x32 + y32 + y3 z3 + z32 . This shows that BA4 is actually a c-hypersurface and hence
also a g-hypersurface. Indeed, the Eilenberg-Moore spectral sequence shows that the loop
space homology will eventually have period dividing 4.
In fact we see that BA4 is an s-hypersurface space at 2. The direct symmetries of a
tetrahedron give a homomorphism A4 −→ SO(3) and hence a map BA4 −→ BSO(3). The
fibre is SO(3)/A4, and at the prime 2 this is S 3 , so there is a 2-adic fibration
S 3 −→ BA4 −→ BSO(3).
In Subsection 14.G, we will also describe a representation theoretic approach to showing
that BA4 is a g-hypersurface, which will show that its ultimate period is exactly 4.
14.E. Squeezed homology. Since we are working with groups, it is illuminating to recall
Benson’s purely representation theoretic calculation of the loop space homology H∗ (Ω(BG∧p ))
[15]. In fact he defines the squeezed homology groups HΩ∗ (G; k) algebraically and proves
H∗ (Ω(BG∧p )) ∼
= HΩ∗ (G; k),
In more detail, HΩ∗ (G; k) is the homology of
· · · −→ P3 −→ P2 −→ P1 −→ P0 ,
a so-called squeezed resolution of k. The sequence of projective kG-modules Pi is defined
recursively as follows. To start with P0 = P (k) is the projective cover of k. Now if Pi has
been constructed, take Ni = ker(Pi −→ Pi−1 ) (where we take P−1 = k), and Mi to be the
smallest submodule of Ni so that Ni /Mi is an iterated extension of copies of k. Now take
Pi+1 to be the projective cover of Mi .
34
14.F. Trivial cases. Note that if G is a p-group, we have Ω(BG∧p ) ≃ G so that H∗ (ΩBG) ∼
=
kG and since k is the only simple module, M0 = 0 and we again find HΩ∗ (G) = kG.
We would expect the next best behaviour to be when H ∗ (BG) is a hypersurface. Indeed, if
H ∗ (BG) is a polynomial ring modulo a relation of codegree d, the Eilenberg-Moore spectral
sequence
∧
Ext∗,∗
H ∗ (BG) (k, k) ⇒ H∗ (Ω(BGp ))
shows that there is an ultimate periodicity of period d − 2. The actual period therefore
divides d − 2. We now return to the example of A4 .
14.G. A4 revisited. We described the homotopy theoretic proof that BA4 is an s-hypersurface
space above.
Here we sketch a purely algebraic proof from [15]. To start with, we would like to see
algebraically that H∗ (Ω(BA4 )∧2 ) is eventually periodic.
This case is small enough to be able to compute products in H∗ (ΩBG∧p , k) using squeezed
resolutions, and we get
H∗ (ΩBG∧p , k) = Λ(α) ⊗ khβ, γi/(β 2, γ 2 )
with |α| = 1 and |β| = |γ| = 2. Beware that β and γ do not commute, so that a k-basis for
H∗ (ΩBG∧p , k) is given by alternating words in β and γ (such as βγβ or the empty word),
and α times these alternating words.
There are three simple modules. Indeed, the quotient of A4 by its normal Sylow 2-subgroup
is of order 3; supposing for simplicity that k contains three cube roots of unity 1, ω, ω, the
simples correspond to how a chosen generator acts. The projective covers of the three simple
modules are
k❃
❃
P (k) = ω ❃
❃
ω̄ ,
P (ω) = ω̄
k
⑥⑥
❆❆❆
ω❅
ω̄ ❄❄
❅
ω
⑦⑦⑦
k,
❄
P (ω̄) = k ❃
❃
ω̄
ω.
⑧⑧
With this information to hand, the reader may verify the assertion about the loop space
homology ring by using squeezed resolutions.
15. s-hypersurface spaces and z-hypersurface spaces
In the algebraic setting the remarkable fact is that modules over hypersurfaces have eventually periodic resolutions, and hence that they are hhci of codimension 1. The purpose of
this section is to sketch a similar result for spaces. The argument is given in detail in [22].
Theorem 15.1. If X is an s-hypersurface space with fibre sphere of dimension ≥ 2 then X
is a z-hypersurface space.
15.A. Split spherical fibrations. The key in algebra was to consider bimodules, for which
we consider the (multiplicative) exact sequence
R −→ Re −→ Re ⊗R k,
35
where the first map is a monomorphism split by the map µ along which R acquires its
structure as an Re -module structure. This corresponds to the pullback fibration
X ←− X ×BΓ X ←− S n ,
split by the diagonal
∆ : X −→ X ×BΓ X
along which the cochains on X becomes a bimodule. To simplify notation, we consider a
more general situation: a fibration
B ←− E ←− S n
with section s : B −→ E. The case of immediate interest is B = X, E = X ×BΓ X, where a
C ∗ (E)-module is a C ∗ (X)-bimodule.
Note that by the third isomorphism theorem for fibrations, there is a fibration
s
ΩS n −→ B −→ E.
This gives the required input for the following theorem. The strength of the result is that
the cofibre sequences are of C ∗ (E)-modules.
s
Theorem 15.2. Suppose given a fibration ΩS n −→ B −→ E with n ≥ 2.
(i) If n is odd, then there is a cofibre sequence of C ∗ (E)-modules
Σn−1 C ∗ (B) ←− C ∗ (B) ←− C ∗ (E).
(ii) If n is even, then there are cofibre sequences of C ∗ (E)-modules
C ←− C ∗ (B) ←− C ∗ (E)
and
Σ2n−2 C ∗ (B) ←− C ←− Σn−1 C ∗ (E).
In particular the fibre of the composite
C ∗ (B) −→ C −→ Σ2n−2 C ∗ (B)
is a small C ∗ (E)-module constructed with one cell in codegree 0 and one in codegree n − 1.
Remark 15.3. Note that in either case we obtain a cofibre sequence
K ←− C ∗ (B) ←− Σa C ∗ (B)
of C ∗ (E)-modules with K small.
The strategy is to first prove the counterparts in cohomology by looking at the Serre
spectral sequence of the fibration from Part (i) and then lift the conclusion to the level of
cochains.
36
16. Growth of z-hypersurface resolutions
In this brief section we prove perhaps the simplest implication between the hypersurface
conditions: a z-hypersurface is a g-hypersurface.
Lemma 16.1. If R is a z-hypersurface and k is a field then the vector spaces πn (E) are of
bounded dimension and R is a g-hypersurface.
Proof: Since R is a z-hypersurface, then in particular there is a projective resolution of k
which is eventually periodic. In other words there is a triangle
Σn k −→ k −→ L
with n 6= 0 and L small over R. Applying HomR (·, k) we find a triangle
Σ−n E ←− E ←− HomR (L, k).
Since L is finitely built from R, HomR (L, k) is finite dimensional over k, and hence only
nonzero in a finite range of degrees (say [−N, N]). Outside that range we have πs+n (E) ∼
=
πs (E), so every homotopy group is isomorphic as a k-vector space to one in the range
[−N − n, N + n] and the bound is the largest of these dimensions.
Remark 16.2. Essentially the same argument shows that a cofibre sequence Σn A −→ A −→
B with two terms the same means that the growth rate of A is at most one more than that
of B.
Part 5. Gorenstein rings
This is the final part taking a particular classes of commutative local rings, and giving
homotopy invariant counterparts of the definitions. The justification consists of the dual
facts that the new definition reduces to the old in the classical setting and that the new definition covers and illuminates new examples.The generalization of the Gorenstein condition
is perhaps the most successful of the three, since there are so many Gorenstein ring spectra,
and this approach provides consequences that are both unexpected and very concrete.
In Section 17 we introduce the homotopy invariant version of the Gorenstein condition an
in Section 18 the associated Gorenstein duality property and its implications for coefficients.
We then provide some basic tools for dealing with the Gorenstein condition: ascent and
descent in Section 19 and Morita invariance in Section 20. We then turn to examples in
earnest, with a discussion of Gorenstein duality for group cohomology in Section 21, for
rational homotopy theory in Section 22 and a brief pointer to other examples in Section 23.
17. The Gorenstein condition
In this section we quickly recall the definition of a Gorenstein local ring in a form which
also makes sense for ring spectra with a map to k.
37
17.A. Gorenstein local rings. The usual definition of a Gorenstein local ring is that R
is of finite injective dimension over itself. But it is then proved that in fact it is then of
injective dimension equal to r = Krull dimension(R) and that Ext∗R (k, R) = ExtrR (k, R) = k.
Conversely, if this holds, the ring is Gorenstein. There are various other characterizations
of Gorenstein rings, including a duality statement that we will discuss shortly, but this is
enough to suggest the definition for ring spectra.
17.B. Gorenstein ring spectra. Ultimately, we want to consider duality phenomena modelled on those in commutative algebra of Gorenstein local rings, so we will develop the theory
for spectra in parallel. Corresponding to the Noetherian condition we restrict the class of
ring spectra to those which are proxy-small in the sense of Definition 5.2. We begin with
the core Gorenstein condition and move onto duality in due course. These definitions come
from [33].
17.C. The Gorenstein condition. We say that R −→ k is Gorenstein of shift a (and write
shift(R) = a) if we have an equivalence
HomR (k, R) ≃ Σa k
of R-modules.
Remark 17.1. We will say that it is c-Gorenstein if R∗ −→ k is a Gorenstein local ring. As
usual the spectral sequence
Ext∗,∗
R∗ (k, R∗ ) ⇒ π∗ (HomR (k, R))
shows that a c-Gorenstein ring spectrum is Gorenstein, but we will give many examples of
Gorenstein ring spectra which are not c-Gorenstein.
18. Gorenstein duality
Although the Gorenstein condition itself is convenient to work with, the real reason for
considering it is the duality property that it implies. To formulate this, we use local cohomology in the sense of Grothendieck, and the reader may wish to refer to Appendix A for
the basic definitions.
18.A. Classical Gorenstein duality. In classical local commutative algebra, the Gorenstein duality property is that all local cohomology is in a single cohomological degree, where
it is the injective hull I(k) of the residue field. To give a formula, we write Γm M for the
m-power torsion in an R-module M, and Hm∗ (M) for the local cohomology of M, recalling Grothendieck’s theorem that if R is Noetherian, Hm∗ (M) = R∗ Γm (M). The Gorenstein
duality statement for a local ring of Krull dimension r therefore states
Hm∗ (R) = Hmr (R) = I(k).
If R is a k-algebra, I(k) = R∨ = Γm Homk (R, k).
38
18.B. Gorenstein duality for k-algebra spectra. Turning to ring spectra, we will treat
the case that R is a k-algebra. This simplifies things considerably, and covers many interesting examples. The more general case requires a discussion of Matlis lifts, for which we
refer the reader to [33].
In the case of k-algebras, we may again define R∨ = Cellk (Homk (R, k)) and observe this
has the Matlis lifting property
HomR (T, R∨ ) ≃ Homk (T, k)
for any T built from k.
In particular, if R is Gorenstein of shift a we have equivalences of R-modules
(G)
(M )
HomR (k, Cellk R) ≃ HomR (k, R) ≃ Σa k ≃ HomR (k, Σa R∨ ),
where the equivalence (G) is the Gorenstein property and the equivalence (M) is the Matlis
lifting property. We would like to remove the HomR (k, ·) to deduce
Cellk R ≃ Σa R∨ .
Morita theory (specifically Lemma 6.3) says that if R is proxy-regular we may make this
deduction provided R is orientably Gorenstein in the sense that the right actions of E =
HomR (k, k) on Σa k implied by the two equivalences (G) and (M) agree.
We note that Cellk (R) is a covariant functor of R whilst R∨ is a contravariant functor of
R, so an equivalence between them is a form of duality: when it holds we say R satisfies
Gorenstein duality.
18.C. Automatic orientability. There are a number of important cases where orientability is automatic because E has a unique action on k, and in this case the Gorenstein condition
automatically implies Gorenstein duality.
The first case of this is when R is a classical commutative local ring, although of course we
knew already that in this case the Gorenstein condition is equivalent to Gorenstein duality.
From our present point of view, we see this as a consequence of connectivity: E (whose
homology is Ext∗R (k, k)) has a unique action on k. The same argument applies when the ring
spectrum is both a k-algebra and connected.
Proposition 18.1. Suppose R is a proxy-regular, connected k-algebra and π∗ (R) is Noetherian with π0 (R) = k and maximal ideal m of positive degree elements. If R is Gorenstein of
shift a, then it is automatically orientable and so has Gorenstein duality.
Proof : First we argue that if R is Gorenstein, it is automatically orientable. Indeed, we
show that E has a unique action on k. Since R is a k-algebra, the action of E on k factors
through
E = HomR (k, k) −→ Homk (k, k) = k,
so since k is an Eilenberg-MacLane spectrum, the action is through π0 (E). Now we observe
that since R is connected, ExtsR∗ (k, k) is in degrees ≤ −s, so that the spectral sequence for
calculating π∗ (HomR (k, k)) shows E is coconnective with π0 (E) = k which must act trivially
on k.
39
We may go a little further to the nilpotent case.
Lemma 18.2. If X is connected with π1 (X) a finite p-group and k is of characteristic p
then if C ∗ (X) is Gorenstein it automatically has Gorenstein duality.
Proof: Again we find E has a unique action on k. Since E is a k-algebra, it acts through
π0 (E) = H0 (ΩX) = k[π1 (X)]. By the characteristic assumption, this has a unique action on
k.
18.D. The local cohomology theorem. In many cases (see Remark D.2 of Appendix Dy)
one can give an algebraic description of the k-cellularization and infer algebraic consequences
of Gorenstein duality. For simplicity we restrict to local k-algebras, although the methods
apply more generally.
Lemma 18.3. If R∗ is a Noetherian k-algebra, with maximal ideal m and residue field k,
then k-cellularization coincides with the derived m-power torsion functor. Accordingly, if R
has Gorenstein duality, there is a local cohomology spectral sequence
Hm∗ (R∗ ) ⇒ Σa R∗∨ .
If R has Gorenstein duality, Lemma 18.3 shows that the ring π∗ (R) has very special
properties (even if it falls short of being Gorenstein), studied in [70]. Some of these properties
were first observed by Benson and Carlson [16, 17] for group cohomology (corresponding to
the special case of the ring spectrum R = C ∗ (BG), which we will see below has Gorenstein
duality).
To start with, we note that the spectral sequence collapses if R∗ is Cohen-Macaulay to
show Hmr (R∗ ) ∼
= Σa+r R∗∨ (where r is the Krull dimension of R∗ ). Thus the coefficient ring
R∗ is also Gorenstein.
The spectral sequence also collapses if R∗ is of Cohen-Macaulay defect 1, to give an exact
sequence
0 −→ Hmr (R∗ ) −→ Σa+r R∗∨ −→ ΣHmr−1 (R∗ ) −→ 0.
In general, local duality lets one deduce that the cohomology ring R∗ is always generically
Gorenstein.
The collapse of the local cohomology theorem in the case of Cohen-Macaulay defect ≤ 1
has very concrete consequences in that the Hilbert series of R∗ satisfies a suitable pair of
functional equations.
Corollary 18.4. [56] Suppose R has Gorenstein duality of shift a, that π∗ (R) is Noetherian
P
of Krull dimension r and Hilbert series p(s) = i dimk (Ri )si .
If π∗ (R) is Cohen-Macaulay it is also Gorenstein, and the Hilbert series satisfies
p(1/s) = (−1)r sr−a p(s).
If π∗ (R) is almost Cohen-Macaulay it is also almost Gorenstein, and the Hilbert series
satisfies
p(1/s) − (−1)r sr−a p(s) = (−1)r−1 (1 + s)q(s) and q(1/s) = (−1)r−1 sa−r+1 q(s).
40
In any case π∗ (R) is Gorenstein in codimension 0 and almost Gorenstein in codimension
1.
19. Ascent, descent and arithmetic of shifts
Very commonly we have a a map θ : S −→ R of ring spectra, and we wish to relate
properties of the two rings. With language from the geometric counterpart, a theorem
stating that if R is Gorenstein then S is Gorenstein is called a descent theorem and a theorem
stating that if S is Gorenstein then R is Gorenstein is called an ascent theorem. Typically
the hypotheses are either about relative properties of θ or in terms of the cofibre of θ.
19.A. Relatively Gorenstein maps. We say that θ is relatively Gorenstein of shift a if
HomS (R, S) ≃ Σa R,
and then write a = shift(R|S). This is quite a strong version of the condition, since we have
asked for a single untwisted ineger suspension.
We make the elementary observation that for any ring map θ : S −→ R
HomR (k, HomS (R, S)) ≃ HomR (R ⊗S k, S) ≃ HomS (k, S).
Thus we conclude that if S −→ R is relatively Gorenstein then R is Gorenstein if and only
if S is Gorenstein, and in that case
shift(S) = shift(R) + shift(R|S).
19.B. Two chromatic examples. There are a number of examples where known facts
amount to proving that a map of ring spectra is relatively Gorenstein, and the resulting
descent theorems prove things of great interest. In these examples one shows that θ is
relatively Gorenstein and R is c-Gorenstein; we then reach the interesting conclusion that S
is Gorenstein.
Example 19.1. (Complexification in K-theory) (i) Periodic complex K-theory is represented
by the ring spectrum KU and periodic real K-theory by KO. The coefficient ring of complex
K-theory is KU∗ = Z[v, v −1 ] where v is of degree 2. Complexification gives a ring map
S = KO −→ KU = R
and we observe that it is relatively Gorenstein. Indeed, Wood’s theorem states KO ∧ CP 2 ≃
Σ2 KU. Since
CP 2 = S 2 ∪η e4
we see that there is a cofibre sequence
η
ΣKO −→ KO −→ KU
of KO-modules. It follows by applying HomKO (·, KO) that
HomKO (KU, KO) ≃ Σ−2 KU
so that KO −→ KU is relatively Gorenstein of shift −2.
(ii) We may take connective covers to obtain a ring map
S = ko −→ ku = R.
41
Now ku∗ = Z[v] and killing homotopy groups gives a ring map
ku −→ Z
which is evidently c-Gorenstein (and hence Gorenstein) of shift −3. The connective version
of Wood’s theorem gives a cofibre sequence
η
Σko −→ ko −→ ku
of ko-modules. It follows by applying Homko (·, ko) that
Homko(ku, ko) ≃ Σ−2 ku
so that ko −→ ku is also relatively Gorenstein of shift −2. Hence we deduce that ko −→ Z
is Gorenstein of shift −5. We note that the ring ko∗ = Z[η1 , α4 , β8 ]/(η 3 , ηα, α2 = 4β, 2η) is
fairly complicated, and it is easy to check that ko −→ Z is not c-Gorenstein.
In this discussion we have taken k = Z, which is not a field. If the reader prefers, we can
instead work over the field Fp and use the fact that for any prime p the map ku −→ Fp is
c-Gorenstein of shift −4. Since ko −→ ku is relatively Gorenstein of shift −2 we conclude
ko −→ Fp is Gorenstein of shift −6.
Example 19.2. (Topological modular forms) Precisely similar statements hold for the ring
spectrum tmf of topological modular forms at various primes. This uses results of HopkinsMahowald as proved by Matthew [82] (see [54] for a slightly expanded discussion).
(3) Localized at the prime 3, there is a map tmf −→ tmf0 (2) to the ring spectrum of
topological modular forms with the indicated level structure. The counterpart of Wood’s
theorem is the fact that tmf0 (2) ≃ tmf ∧ (S 0 ∪α1 e4 ∪α1 e8 ), so that
Homtmf (tmf0 (2), tmf ) ≃ Σ−8 tmf0 (2),
and the map is relatively Gorenstein of shift −8. Since tmf0 (2)∗ = Z(3) [c2 , c4 ] (where |ci | =
2i) we see that tmf0 (2) −→ Z(3) is c-Gorenstein (and hence Gorenstein) of shift −14. Hence
we deduce by Gorenstein descent that tmf −→ Z(3) is Gorenstein of shift −22.
(2) Localized at the prime 2, there is a map tmf −→ tmf1 (3) to the ring spectrum of
topological modular forms with the indicated level structure. The Here tmf1 (3) is a form of
BP h2i and tmf1 (3) ≃ tmf ∧ DA(1) so that
Homtmf (tmf1 (3), tmf ) ≃ Σ−12 tmf1 (3),
and the map is relatively Gorenstein of shift −12. Since tmf1 (3)∗ = Z(2) [α1 , α3 ] (where
|αi | = 2i) we see that tmf1 (3) −→ Z(2) is c-Gorenstein (and hence Gorenstein) of shift −10.
Hence we deduce by Gorenstein descent that tmf −→ Z(2) is Gorenstein of shift −22.
Abstracting this slightly, we have
θ
S −→ S −→ R −→ k
and these examples were all very special in that there was an equivalence
R ≃ S ⊗S L
42
for a self-dual finite complex L. This means in turn that the cofibre of θ : S −→ R
R ⊗S k ≃ k ⊗S L,
with homotopy H∗ (L; k) a Poincaré duality algebra. In the next subsection we show that
this weaker condition is often sufficient to give an ascent theorem.
19.C. Gorenstein Ascent. We suppose that S −→ R −→ Q is a cofibre sequence of
commutative algebras with a map to k, and we now consider the Gorenstein ascent question.
When does the fact that S is Gorenstein imply that R is Gorenstein? It is natural to assume
that Q is Gorenstein, but it is known this is not generally sufficient.
In effect the Gorenstein Ascent theorem will state that under suitable hypotheses (see
Section 19.E) there is an equivalence
HomR (k, R) ≃ HomQ (k, HomS (k, S) ⊗k Q).
When this holds, it follows that if S and Q are Gorenstein, so is R and
shift(R) = shift(S) + shift(Q).
19.D. Arithmetic of shifts. We summarize the behaviour of Gorenstein shifts in the ideal
situation when ascent and descent both hold. If all rings and maps are Gorenstein of the
indicated shifts
s µ r λ q
S −→R−→Q
then r = s + q, µ = −q and λ = s
19.E. When does Gorenstein ascent hold? The core of our results about ascent come
from [33]. Indeed, the proof of [33, 8.6] gives a sufficient condition for Gorenstein ascent in
the commutative context.
Lemma 19.3. If S and R are commutative and the natural map ν : HomS (k, S) ⊗S R −→
HomS (k, R) is an equivalence then
HomR (k, R) ≃ HomQ (k, HomS (k, S) ⊗k Q).
In this case, if S and Q are Gorenstein, so is R, and the shifts add up: shift(R) =
shift(S) + shift(Q).
Now that we have a sufficient condition for Gorenstein ascent, we want to identify cases
in which it is satisfied. The map
ν : HomS (k, S) ⊗S M −→ HomS (k, M),
is clearly an equivalence when M = R and hence for any module finitely built from R. This
shows that the hypotheses are satisfied when R is small over S (or equivalently, when Q is
finitely built from k) so that ascent holds in this case. This is already a very useful result.
Example 19.4. If we have a fibration F −→ E −→ B of spaces to which the EilenbergMoore theorem applies in the sense that C ∗ (F ) = C ∗ (E) ⊗C ∗ (B) k, then if B is Gorenstein
and F is an orientable manifold, it follows from Gorenstein Ascent that E is Gorenstein.
43
There are other important cases where ν is han equivalence. Indeed, we can exploit the
fact that the hypothesis on ν in Lemma 19.3 only depends on R as a module over S to show
that ν is an equivalence when R is suitably approximated as an inverse limit. This is a
central ingredient in proving Gorenstein duality for many topological Hochschild homology
spectra [54].
19.F. Local duality. To start with, we clarify terminology. The Gorenstein duality property we have been discussing is a rare and special thing. On the other hand, local duality
(as in the title of this subsection) is a tool available very generally. Local duality is based on
Noether normalization, which means that every well behaved local ring R is finite as a module over a Gorenstein ring S. Local duality is the property inherited by R as a consequence
of the existence of S.
Accordingly, we assume for the rest of this section that we are given a map S −→ R so that
R is a small S-module. Notice that this means that for an R-module M, its k-cellularization
as an R-module is its k-cellularization as an S-module and similarly for k-completions. This
is reflected in our notation. We also note that R∨ = HomS (R, S ∨ ) so that the Matlis lifts
are related by coextension of scalars.
Traditionally, local duality is thought of as saying that Matlis dual of local cohomology
(embodied in HomR (Cellk R, R∨ ) ≃ (R∨ )∧k ) is isomorphic to a completed Ext group (embodied by HomS (R, S)∧k ). Directly translating this into our context we reach a more inscrutable
definition.
Definition 19.5. We say that S −→ R has local duality of shift b if there is an equivalence
of R-modules
HomS (R, S)∧k = Σb (R∨ )∧k .
We now have two properties the map S −→ R may or may not have: the relative Gorenstein property and local duality. We also have Gorenstein duality for S −→ k and for
R −→ k. It is valuable to disentangle the relationships between them.
In fact we have three complete R-modules.
• Rk∧
• (R∨ )∧k
• HomS (R, S)∧k
The equivalence of each of the possible pairs has a name. In the following diagram, the label
“relGor” means “k-completion of relatively Gorenstein”, “locD” means “local duality”, and
“GorD” means “Gorenstein duality”; the superscripts indicate the suspension necessary to
get from the tail of the arrow to the head of the arrow.
HomS (R, S)∧k
Rk∧
c rrr
relGorr
r
r
rr
r
r
y
r
◆◆◆
b
◆◆locD
◆◆◆
◆◆◆
&
/ (R∨ )∧
k
RGorDa
Sk∧ ❊
❊
❊❊SGorDb
❊❊
❊❊
"
(S ∨ )∧k
This makes it clear that any two of relGor, locD and RGorD implies the third, and that
a + c = b.
44
It remains to observe that locD follows from SGorD by coextension of scalars from Smodules to R-modules.
Lemma 19.6. If S has Gorenstein duality of shift b then S −→ R has local duality of shift
b.
Proof: We apply HomS (R, ·) to the equivalence Sk∧ ≃ Σb (S ∨ )∧k and then use the following
lemma to move the completions to the outside.
Lemma 19.7. For any S-module N, we have an equivalence
HomS (R, Nk∧ ) ≃ HomS (R, N)∧k
Proof: This is a formality:
HomS (R, HomS (Cellk S, N)) ≃ HomS (R ⊗S Cellk S, N)) ≃ HomR (Cellk R, HomS (R, N))
This approach can be used to show that Gorenstein duality localizes [20].
20. Morita invariance of the Gorenstein condition
We show that the Gorenstein condition is Morita invariant in many useful cases, provided
R is a k-algebra. This allows us to deduce striking consequences from well-known examples
of Gorenstein rings. For instance we can deduce the local cohomology theorem for finite
p-groups from the fact that kG is a Frobenius algebra.
Theorem 20.1. Suppose R is a k-algebra, and that E and R are Matlis reflexive. Then
HomE (k, E) ≃ HomR (k, R),
and hence
E is Gorenstein ⇐⇒ R is Gorenstein.
Proof: We use the fact that (in the notation of Section 6) E(R∨ ) = k ∨ , so that by Lemma
6.3 we have
R∨ = T E(R∨ ) = T k ∨ = k ∨ ⊗E k.
We also note that
HomR (k ⊗R k ∨ , k) ≃ Homk (k, HomR (k, k)) ≃ E
so that
E ∨ = k ⊗R k ∨ .
Next, note that the expression k ⊗R k ∨ ⊗E k makes sense, where the right E-module
structure on the first two factors comes from k ∨ . The key equality in the proof is simply the
associativity isomorphism
E ∨ ⊗E k = k ⊗R k ∨ ⊗E k = k ⊗R R∨ .
Now we make the following calculation,
45
HomE (k, E) ≃
≃
≃
≃
≃
≃
HomE (k, (E ∨ )∨ )
Homk (E ∨ ⊗E k, k)
Homk (k ⊗R k ∨ ⊗E k, k)
Homk (k ⊗R R∨ , k)
HomR (k, (R∨ )∨ )
HomR (k, R)
21. Gorenstein duality for group cohomology
This section describes the key example. The first sign of this duality was in Benson-Carlson
duality [16, 17], which in particular shows that the Hilbert series of the group cohomology
ring H ∗ (BG) satisfies a functional equation if it is Cohen-Macaulay or a pair of functional
equations if it is almost Cohen-Macaulay. An algebraic construction of the local cohomology
spectral sequence was given in [49]; this was inspired by the topological construction using
equivariant topology in [47], and a proof using structured equivariant spectra first appears
in [18]. The method described here comes from [33].
21.A. p-groups. If G is a p-group and k is of characteristic p we note that BG is p-complete
and therefore Ω(BG∧p ) ≃ G and the Morita pair is
R = C ∗ (BG; k) and E = C∗ (ΩBG) = kG.
Since kG is a Frobenius algebra, it is Gorenstein of shift 0, and so by Morita invariance of
the Gorenstein condition (Theorem 20.1), C ∗ (BG) is also Gorenstein of shift 0.
21.B. General finite groups. Now if G is an arbitrary finite group, we may choose a
faithful representation ρ : G −→ SU(n) for some n and consider the fibration
BSU(n) ←− BG ←− SU(n)/G.
Since BSU(n) is simply connected, the Eilenberg-Moore theorem gives a cofibre sequence
C ∗ (BSU(n)) −→ C ∗ (BG) −→ C ∗ (SU(n)/G).
Now note that H ∗ (BSU(n)) is polynomial on c2 , . . . , cn and therefore C ∗ (BSU(n)) is cGorenstein and therefore Gorenstein (with shift 2(2 + 3 + · · · + n) − (n − 1) = dim(SU(n))).
On the other hand SU(n)/G is an orientable manifold of the same dimension as SU(n) and
therefore C ∗ (SU(n)/G) is Gorenstein. By Gorenstein Ascent (Example 19.4), C ∗ (BG) is
Gorenstein of shift 0 as required.
21.C. The local cohomology theorem. As described in Lemma 18.3 we thus obtain the
local cohomology theorem
Hm∗ (H ∗ BG) ⇒ H∗ (BG)
for group cohomology.
46
As described in Corollary 18.4, we then obtain functional equations in many cases. We
warn that t is of codegree 1 (unlike s in the corollary, which was of degree 1). If H ∗ (BG) is
Cohen-Macaulay, it is Gorenstein and its Hilbert series satisfies
p(1/t) = (−1)r tr p(t).
If H ∗ (BG) is almost Cohen-Macaulay it is also almost Gorenstein, and the Hilbert series
satisfies
p(1/t) − (−1)r tr p(t) = (−1)r−1 (1 + t)q(t) and q(1/t) = (−1)r−1 t−r+1 q(t).
In any case H ∗ (BG) is Gorenstein in codimension 0 and almost Gorenstein in codimension
1.
We will describe a number of 2-group examples with k of characteristic 2.
21.D. The elementary abelian group of rank r. The group of order 2 may be viewed
as a subgroup of the non-zero real numbers. It therefore acts diagonally on nR = R ⊕ · · · ⊕ R
and freely on the unit sphere S(nR). Hence
BC2 = S(∞R)/C2 = RP ∞ .
Similarly, if G ∼
= C2 × · · · × C2 is of rank r we see that
BG = RP ∞ × · · · × RP ∞ ,
and
H ∗ (BG) = k[x1 , . . . , xr ].
This is visibly of dimension r and depth r and
p(t) =
1
.
(1 − t)r
The functional equation is easily checked, and the reader may wish to check directly that
the local cohomology of H ∗ (BG) is H∗ (BG) (up to a shift).
21.E. The quaternion group of order 8. If G = Q8 is quaternion of order 8 we note that
it acts freely on the unit 3-sphere in the quaternion algebra S(H), and in fact H ∗ (S(H)/G) =
k[x, y]/(x3 , x2 + xy + y 2, y 3). It is then clear that G acts freely on the contractible space
S(∞H) and that
H ∗ (BG) = H ∗ (S(H)/G)[z]
where z is of codegree 4. This is visibly of dimension 1 and depth 1 and
1 + 2t + 2t2 + t3
.
1 − t4
The functional equation is easily checked. It is also easy to check directly that the local
cohomology of H ∗ (BG) is H∗ (BG) (up to a shift).
Although there are many other groups for which we could calculate the cohomology it is
convenient to simply refer to the invaluable Jena database [45] for the cohomology of small
p-groups.
p(t) =
47
21.F. The dihedral group of order 8. If G = D8 is dihedral of order 8
H ∗ (BG) = k[x1 , y1 , z2 ]/(xy).
It is quick to check that this is of dimension 2 and depth 2 and
1
p(t) =
.
(1 − t)2
The functional equation is easily checked. It is an easy exercise to check directly that the
local cohomology of H ∗ (BG) is H∗ (BG) (up to a shift).
21.G. The semi-dihedral group of order 16. If G = SD16 is the semi-dihedral group of
order 16
H ∗ (BG) = k[x1 , y1 , z3 , t4 ]/(xy, x3 , xz, z 2 + ty 2 ).
One may check that this is of dimension 2 and depth 1 and is included as the first example
which is not Cohen-Macaulay. Its Hilbert series is
1
p(t) =
,
2
(1 − t) (1 + t2 )
and one may check that the almost Cohen-Macaulay functional equations are satisfied with
shift 0. It is presumably a coincidence that it also satisfies the functional equation of a
Cohen-Macaulay graded ring of dimension 2 with (homological) shift 2.
21.H. Group Number 7 of order 32. The 7th group in the Small Groups library list
of the 51 groups of order 32 is the only one whose cohomology is neither Cohen-Macaulay
nor almost Cohen-Macaulay. The ring has a minimal presentation with 8 generators and 18
relations, so it won’t be recorded here. The important facts for us are that it is of dimension
3 and depth 1 and has Hilbert series
1 − t + t2
.
(1 − t)3 (1 + t2 )
By coincidence this satisfies the functional equation for a Gorenstein graded ring with shift
0.
Taking the description of its cohomology given in [86] one may easily calculate its local
cohomology. In fact it has a polynomial subgring P = k[z1 , x2 , s4 ] over which it is a direct
sum of 5 submodules, namely P ⊕ Σ3 P ⊕ Σ4 P ⊕ M ⊕ N where
p(t) =
{x,z}
M = cok(Σ4 P −→ Σ2 P ⊕ Σ3 P )
and
N = Σ1 P/(x, z).
The free submodule is of depth 3, M is of depth 2. Finally
Hm1 (H ∗ (BG)) = Hm1 (N)
is of dimension 1 over k in degrees 3, 7, 11, 15, · · · . Knowing the dimensions of Hn (BG) from
p(t) we see that the differential
d2 : Hm1 (H ∗ (BG)) −→ Σ−1 Hm3 (H ∗ BG))
48
is a monomorphism (and so non-zero in infinitely many degrees).
21.I. Other classes of groups. We have described the fact that C ∗ (BG) is Gorenstein
for finite groups G. It is shown in [33] that C ∗ (BG) is Gorenstein of shift equal to dim(G)
whenever G is a compact Lie group, provided that either k is of characteristic 2 or the group
of components is of odd order. In general there is a twisting by the representation H g (G),
which is trivial in the aforementioned cases.
If G is a virtual duality group of dimension n there is a form of Gorenstein duality with
shift −n for C ∗ (BG). Now there is a twisting by H n (G; kG), which is usually of infinite
dimension. The algebraic proof is given in [19], along with a topological proof for arithmetic
groups using equivariant topology. A proof along the present lines can be given for these
groups as follows. We choose a normal subgroup N of finite index in G which is a duality
group, and a contractible space X on which G acts with finite isotropy. Now N acts freely
on X so X/N is a manifold with boundary. Now let Q = G/N, and consider the fibration
BQ ←− BG ←− X/N,
which we may obtain from the equivalence
BG ≃ EQ ×Q (X/N).
We attempt to apply Gorenstein Ascent. In the situation that X/N has empty boundary
(i.e., G is a virtual Poincaré duality group) we infer C ∗ (BG) is Gorenstein as required.
22. Gorenstein duality for rational spaces
This section discusses Gorenstein duality for rational spaces. Félix-Halperin-Thomas [42]
have considered the Gorenstein condition in depth, so the distinctive feature here is the
concentration on Gorenstein duality as in [33].
For spaces with finite dimensional cohomology X is Gorenstein if and only if H ∗ (X) is
Gorenstein, but in general there are Gorenstein spaces for which H ∗ (X) is not Gorenstein
and we make explicit the local cohomology theorem and its consequences for H ∗ (X).
22.A. Fundamentals. We specialize some of the above results to C ∗ (X; Q), beginning with
the zero dimensional case.
Lemma 22.1. [42, 3.6] If H ∗ (A) is finite dimensional then A is Gorenstein if and only if
H ∗ (A) is a Poincaré duality algebra.
Proof: If H ∗ (A) is a Poincaré duality algebra of formal dimension n then it is a zero dimensional Gorenstein ring with a-invariant −n, so A is Gorenstein with shift −n by the previous
corollary.
Conversely, if A is Gorenstein of shift a, we have a Gorenstein duality spectral sequence.
Since H ∗ (A) is finite dimensional, it is all torsion. Accordingly, Hm∗ (H ∗ (A)) = H ∗ (A), and
the spectral sequence reads
H ∗ (A) = Σa H ∗ (A)∨
and H ∗ (A) is a Poincaré duality algebra of formal dimension −a.
49
By applying Gorenstein Ascent (Example 19.4) one may use these to construct other
examples which are Gorenstein but not c-Gorenstein.
Proposition 22.2. [42, 4.3] Suppose we have a fibration F −→ E −→ B with F finite. If
F and B are Gorenstein with shifts f and b then E is Gorenstein with shift e = f + b.
Lemma 22.3. If X is a simply connected rational space with H ∗ (X) Noetherian then if
C ∗ (X; Q) is Gorenstein of shift a, then it is automatically orientable and so has Gorenstein
duality of shift a
Proof: We see that Q is proxy-small by Corollary 7.2. We may now apply the same argument as for Proposition 18.1 to see there is a unique action of C∗ (ΩX; Q) on Q and deduce
automatic Gorenstein duality.
22.B. Examples. The basic results of the previous subsection allow us to construct innumerable examples. For example any finite Postnikov system is Gorenstein [42, 3.4], so that
in particular any sci space is Gorenstein. A simple example will illustrate the duality.
Example 22.4. We construct a rational space X in a fibration
S 3 × S 3 −→ X −→ CP ∞ × CP ∞ ,
so that X is Gorenstein. We will calculate H ∗(X) and observe that it is not Gorenstein.
Let V be a graded vector space with two generators u, v in degree 2, and let W be a
graded vector space with two generators in degree 4. The two 4-dimensional cohomology
classes u2 , uv in H ∗(KV ) = Q[u, v] define a map KV −→ KW , and we let X be the fibre,
so we have a fibration
S 3 × S 3 −→ X −→ KV
as required. By [33], this is Gorenstein with shift −4 (being the sum of the shift (viz −6) of
S 3 × S 3 and the shift (viz 2) of KV ).
It is amusing to calculate the cohomology ring of X. It is Q[u, v, p]/(u2, uv, up, p2) where
u, v and p have degrees 2, 2 and 5. The dimensions of its graded components are 1, 0, 2, 0, 1, 1, 1, 1, 1, . . .
(i.e., its Hilbert series is pX (t) = (1 + t5 )/(1 − t2 ) + t2 , where t p
is of codegree 1).
In calculating local cohomology it is useful to note that m = (v). The local cohomology
is Hm0 (H ∗ (X)) = Σ2 Q in degree 0 (so that H ∗ (X) is not Cohen-Macaulay) and as a Q[v]module Hm1 (H ∗ (X)) is Q[v]∨ ⊗ (Σ−3 Q ⊕ Σ2 Q). Since there is no higher local cohomology the
local cohomology spectral sequence necessarily collapses, and the resulting exact sequence
0 −→ Hm1 (H ∗ (X)) −→ Σ−4 H ∗ (A)∨ −→ Σ−2 Q −→ 0
is consistent.
Since the Cohen-Macaulay defect here is 1, we have a pair of functional equations
pX (1/t) − (−t)t−4 pX (t) = (1 + t)δ(t)
and
δ(1/t) = t4 δ(t).
50
Indeed, the first equation gives δ(t) = t−2 , which is indeed the Hilbert series of Hm0 (H ∗ (X))∨ ,
and it obviously satisfies the second equation.
23. The ubiquity of Gorenstein ring spectra
We have considered a number of examples, in some cases giving rather complete proofs.
This short section points out that the ideas can be extended rather easily. In fact there is
a sense in which all the examples come from the first three by using Gorenstein Ascent and
Morita invariance of the Gorenstein condition.
• Gorenstein commutative local rings R (shift −Krulldim(R)).
• C ∗ (M) for closed manifolds M (orientable if M is orientable; shift − dim(M)).
• C∗ (G) for compact Lie groups G (orientable if G acts trivially on Hg (G); shift
dim(G)).
• C ∗ (BG) for compact Lie groups G (orientable if G acts trivially on Hg (G); shift
dim(G)).
• C ∗ (BG) for virtual duality groups G (dualizing module H∗ (G; kG); shift −formal-dim(G)).
• C ∗ (EG×G M) Borel construction on a G-manifold M m for Gg compact Lie (orientable
if G, M and the action are; shift g − m). An amusing instance of this arises from
toric geometry; if K is a simplicial complex one may construct the so-called moment
angle complex ZK which has an action of a torus G whose Borel cohomology is the
Stanley-Reisner ring k[K]. If K is a simplicial sphere ZK is a manifold and k[K] is
Gorenstein. See [29] for definitions and proofs.
• Chromatic examples. The coefficient ring R (or ring spectrum) is contravariant in the
geometric object making it like cochains, but the ring spectrum R is often connective
(unlike cochains). With this caveat, the examples are parallel to the C ∗ (BG) example
with the compact Lie group G replaced by an algebraic group G, so that R plays
the role of the ring spectrum of functions on G. The Morita counterpart of R = OG
would be the ring of operations C∗ (G; k) = HomR (k, k).
• Hochschild homology. In [54] several classes of examples are identified where the
Hochschild homology with field coefficients inherits Gorenstein properties. The context is that we are given maps S −→ R −→ k of commutative rings, so that we can
define the Hochschild homology
HH•(R|S; k) = R ⊗R⊗S R k,
and it is a ring spectrum with a map to k. The results then say that (under substantial
additional hypotheses) if R is Gorenstein of shift a and HH• (k|S; k) is Gorenstein
of shift b then HH• (R|S; k) is Gorenstein of shift b − a. The proof comes out of
Gorenstein Ascent, with the hypotheses designed to allow the application of Lemma
19.3.
51
Appendix A. Algebraic definitions: Local and Čech cohomology and
homology
Related surveys are given in [59, 60]. The material in this section is based on [66, 58, 48].
Background in commutative algebra can be found in [83, 28].
I. The functors. Suppose to begin with that R is a commutative Noetherian ring and
that I = (α1 , . . . , αn ) is an ideal in R. We shall be concerned especially with two naturally
occurring functors on R-modules: the I-power torsion functor and the I-adic completion
functor.
The I-power torsion functor ΓI is defined by
M 7−→ ΓI (M) = {x ∈ M | I k x = 0 for k >> 0}.
We say that M is an I-power torsion module if M = ΓI M. It is easy to check that the
functor ΓI is left exact.
The I-adic completion functor is defined by
M 7−→ MI∧ = lim M/I k M.
← k
The Artin-Rees lemma implies that I-adic completion is exact on finitely generated modules,
but it is neither right nor left exact in general.
II. The stable Koszul complex. We begin with a sequence α1 , . . . , αn of elements of R
and define various chain complexes. In Subsection III we explain why the chain complexes
only depend on the radical of the ideal I = (α1 , . . . , αn ) generated by the sequence, in
Subsection IV we define associated homology groups, and in Subsection V we give conceptual
interpretations of this homology under Noetherian hypotheses.
We begin with a single element α ∈ R, and an integer s ≥ 0, and define the sth unstable
Koszul complex by
Ks• (α) = (αs : R −→ R)
where the non-zero modules are in cohomological degrees 0 and 1. These complexes form a
direct system as s varies,
α
K1• (α) = ( R −→ R )
↓
=↓
↓α
α2
K2• (α) = ( R −→ R )
↓
=↓
↓α
α3
K3• (α) = ( R −→ R )
↓
=↓
↓α
and the direct limit is the flat stable Koszul complex
•
K∞
(α) = (R −→ R[1/α]) .
•
When defining local cohomology, it is usual to use the complex K∞
(α) of flat modules.
However, we shall need a complex of projective R-modules to define the dual local homology
•
modules. Accordingly, we take a particularly convenient projective approximation P K∞
(α)
•
•
to K∞ (α). Instead of taking the direct limit of the Ks (α), we take their homotopy direct
52
limit. This makes the translation to the topological context straightforward. More concretely,
•
our model for P K∞
(α) is displayed as the upper row in the homology isomorphism
P K • (α)
=
( R ⊕ R[x]
h1,αx−1i
/
R[x] )
g
h1,0i
•
K∞
(α)
=
(
R
/
R[1/α] ) ,
•
•
where g(xi ) = 1/αi . Like K∞
(α), this choice of P K∞
(α) is non-zero only in cohomological
degrees 0 and 1.
The stable Koszul cochain complex for a sequence α = (α1 , . . . , αn ) is obtained by tensoring together the complexes for the elements, so that
•
•
•
K∞
(α) = K∞
(α1 ) ⊗R · · · ⊗R K∞
(αn ),
•
and similarly for the projective complex P K∞
(α).
III. Invariance statements. We list some basic properties of the stable Koszul complex,
leaving proofs to the reader.
•
Lemma A.1. If β is in the ideal I = (α1 , α2 , . . . , αn ), then K∞
(α)[1/β] is exact.
Note that, by construction, we have an augmentation map
•
ε : K∞
(α) −→ R.
Using this to compare different stable Koszul complexes and Lemma A.1 to see when they
are equivalences we deduce an important invariance statement.
•
Corollary A.2. Up to quasi-isomorphism, the complex K∞
(α) depends only on the radical
of the ideal I.
•
•
•
In view of Corollary A.2 it is reasonable to write K∞
(I) for K∞
(α). Since P K∞
(α) is a
•
projective approximation to K∞ (α), it too depends only on the radical of I. We also write
Ks• (I) = Ks• (α1 ) ⊗ · · · ⊗ Ks• (αn ), but this is an abuse of notation since even its homology
groups do depend on the choice of generators.
IV. Local homology and cohomology. The local cohomology and homology of an Rmodule M are then defined by
•
HI∗ (R; M) = H ∗ (P K∞
(I) ⊗ M)
and
•
H∗I (R; M) = H∗ (Hom(P K∞
(I), M).
Note that we could equally well use the flat stable Koszul complex in the definition of local
cohomology, as is more usual. Lemma A.1 shows that HI∗ (M)[1/β] = 0 if β ∈ I, so HI∗ (M)
is an I-power torsion module and supported over I.
It is immediate from the definitions that local cohomology and local homology are related
by a third quadrant universal coefficient spectral sequence
(1)
I
E2s,t = ExtsR (HI−t (R), M) =⇒ H−t−s
(R; M),
53
with differentials dr : Ers,t −→ Ers+r,t−r+1 .
V. Derived functors. We gave our definitions in terms of specific chain complexes. The
meaning of the definitions appears in the following two theorems.
Theorem A.3 (Grothendieck [66]). If R is Noetherian, then the local cohomology groups
calculate the right derived functors of the left exact functor M 7−→ ΓI (M). In symbols,
HIn (R; M) = (Rn ΓI )(M).
This result may be used to give an explicit expression for local cohomology in familiar
terms. Indeed, since ΓI (M) = lim Hom(R/I r , M), and the right derived functors of the
→ r
right-hand side are obvious, we have
(Rn ΓI )(M) ∼
= lim ExtnR (R/I r , M).
→ r
The description in terms of the stable Koszul complex is usually more practical.
Theorem A.4 (Greenlees-May [58]). If R is Noetherian, then the local homology groups
calculate the left derived functors of the (not usually right exact) I-adic completion functor
M 7−→ MI∧ . Writing LIn for the left derived functors of I-adic completion, this gives
HnI (R; M) = LIn (M).
The conclusions of Theorem A.3 and A.4 are true under much weaker hypotheses [58, 87].
VI. The shape of local cohomology. One is used to the idea that I-adic completion is
often exact, so that LI0 is the most significant of the left derived functors. However, it is the
top non-vanishing right derived functor of ΓI that is the most significant. Some idea of the
shape of these derived functors can be obtained from the following result. Observe that the
•
complex P K∞
(α) is non-zero only in cohomological degrees between 0 and n, so that local
homology and cohomology are zero above dimension n. A result of Grothendieck’s usually
gives a much better bound. We write dim(R) for the Krull dimension of R and depthI (M)
for the I-depth of a module M (the length of the longest M-regular sequence from I).
Theorem A.5 (Grothendieck [65]). If R is Noetherian of Krull dimension d, then
HIi (M) = 0
and
HiI (M) = 0 if i > d.
If e = depthI (M) then
HIi (M) = 0 if i < e.
If R is Noetherian, M is finitely generated, and IM 6= M, then
HIe (M) 6= 0.
Grothendieck’s proof of vanishing begins by noting that local cohomology is sheaf cohomology with support. It then proceeds by induction on the Krull dimension and reduction
to the irreducible case. The statement about depth is elementary, and proved by induction
on the length of the I-sequence (see [83, 16.8]).
The Universal Coefficient Theorem gives a useful consequence for local homology.
54
Corollary A.6. If R is Noetherian and depthI (R) = dim(R) = d, then
d
LIs M = Extd−s
R (HI (R), M).
∗
1
For example if R = Z and I = (p), then H(p)
(Z) = H(p)
(Z) = Z/p∞ . Therefore the
corollary states that
(p)
L0 M = ExtZ (Z/p∞ , M)
and
(p)
L1 M = HomZ (Z/p∞ , M),
as was observed in Bousfield-Kan [26, VI.2.1].
VII. Čech homology and cohomology. We have motivated local cohomology in terms
of I-power torsion, and it is natural to consider the difference between the torsion and the
original module. In geometry this difference would be more fundamental than the torsion
itself, and local cohomology would then arise by considering functions with support.
•
To construct a good model for this difference, observe that ε : K∞
(α) −→ R is an
•
isomorphism in degree zero and define the flat Čech complex Č (I) to be the complex
Σ(ker ε). Thus, if i ≥ 0, then Č i (I) = K i+1 (I). For example, if I = (α, β), then
Č • (I) = ( R[1/α] ⊕ R[1/β] −→ R[1/(αβ)] ) .
The differential K 0 (I) −→ K 1 (I) specifies a chain map R −→ Č • (I) whose fibre is exactly
•
K∞
(I). Thus we have a fibre sequence
•
K∞
(I) −→ R −→ Č • (I).
We define the projective version P Č • (I) similarly, using the kernel of the composite of ε and
•
•
the quasi-isomorphism P K∞
(I) −→ K∞
(I); note that P Č • (I) is non-zero in cohomological
degree −1.
The Čech cohomology of an R-module M is then defined by
ČHI∗ (R; M) = H ∗ (Č • (I) ⊗ M).
VIII. Čech cohomology and Čech covers. To explain why Č • (I) is called the Čech
complex, we describe how it arises by using the Čech construction to calculate cohomology
from a suitable open cover. More precisely, let Y be the closed subscheme of X = Spec(R)
determined by I. The space V (I) = {℘|℘ ⊃ I} decomposes as V (I) = V (α1 ) ∩ . . . ∩ V (αn ),
and there results an open cover of the open subscheme X −Y as the union of the complements
X − Yi of the closed subschemes Yi determined by the principal ideals (αi ). However, X −
Yi is isomorphic to the affine scheme Spec(R[1/αi ]). Since affine schemes have no higher
cohomology,
H ∗ (Spec(R[1/αi ]); M̃ ) = H 0 (Spec(R[1/αi ]); M̃) = M[1/αi ],
where M̃ is the sheaf associated to the R-module M. Thus the E1 term of the Mayer-Vietoris
spectral sequence for this cover collapses to the chain complex Č • (I), and
H ∗ (X − Y ; M̃ ) ∼
= ČHI∗ (M).
55
Appendix B. Spectral analogues of the algebraic definitions
We now transpose the algebra from Appendix A into the context of spectra. It is convenient
to note that it is routine to extend the algebra to graded rings, and we will use this without
further comment below. We assume the reader is already comfortable working with ring
spectra, but there is an introduction in Sections 2 and 3, and one at greater length in [52].
We replace the standing assumption that R is a commutative Z-algebra by the assumption
that it is a commutative S-algebra, where S is the sphere spectrum. The category of Rmodules is now the category of R-module spectra. Since the derived category of a ring R is
equivalent to the derived category of the associated Eilenberg-MacLane spectrum [92], the
work of Appendix A can be reinterpreted in the new context. To emphasize the algebraic
analogy, we write ⊗R and HomR for the smash product over R and function spectrum of Rmaps and 0 for the trivial module. In particular X⊗S Y = X∧Y and HomS (X, Y ) = F (X, Y ).
This section is based on [47, 48].
I. Koszul spectra. For α ∈ π∗ R, we define the stable Koszul spectrum K∞ (α) by the fibre
sequence
K∞ (α) −→ R −→ R[1/α],
α
α
where R[1/α] = holim(R −→ R −→ . . .). Analogous to the filtration by degree in chain
→
complexes, we obtain a filtration of the R-module K∞ (α) by viewing it as
Σ−1 (R[1/α] ∪ CR).
Next we define the stable Koszul spectrum for the sequence α1 , . . . , αn by
K∞ (α1 , . . . , αn ) = K∞ (α1 ) ⊗R · · · ⊗R K∞ (αn ),
and give it the tensor product filtration.
The topological analogue of Lemma A.1 states that if β ∈ I then
K∞ (α1 , . . . , αn )[1/β] ≃ 0;
this follows from Lemma A.1 and the spectral sequence (3) below. We may now use precisely
the same proof as in the algebraic case to conclude that the homotopy type of K∞ (α1 , . . . , αn )
depends only on the radical of the ideal I = (α1 , · · · , αn ). We therefore write K∞ (I) for
K∞ (α1 , . . . , αn ).
II. Localization and completion. With motivation from Theorems A.3 and A.4, we define
the homotopical I-power torsion (or local cohomology) and homotopical completion (or local
homology) modules associated to an R-module M by
(2)
ΓI (M) = K∞ (I) ⊗R M
and
ΛI (M) = MI∧ = HomR (K∞ (I), M).
In particular, ΓI (R) = K∞ (I).
Because the construction follows the algebra so precisely, it is easy give methods of calculation for the homotopy groups of these R-modules. We use the product of the filtrations of
the K∞ (αi ) given above and obtain spectral sequences
(3)
2
Es,t
= HI−s,−t(R∗ ; M∗ ) ⇒ πs+t (ΓI M)
56
r
r
with differentials dr : Es,t
→ Es−r,t+r−1
and
(4)
I
E2s,t = H−s,−t
(R∗ ; M ∗ ) ⇒ π−(s+t) (MI∧ )
with differentials dr : Ers,t → Ers+r,t−r+1 .
III. The Čech spectra. Similarly, we define the Čech spectrum by the cofibre sequence
(5)
K∞ (I) −→ R −→ Č(I).
We define the homotopical localization (or Čech cohomology) and Čech homology modules
associated to an R-module M by
(6)
M[I −1 ] = Č(I) ⊗R M
and
∆I (M) = HomR (Č(I), M).
In particular, R[I −1 ] = Č(I). Once again, we have spectral sequences for calculating their
homotopy groups from the analogous algebraic constructions.
IV. Basic properties. We can now give topological analogues of some basic pieces of alge•
bra that we used in Section A. Recall that the algebraic Koszul complex K∞
(I) is a direct
•
limit of unstable complexes Ks (I) that are finite complexes of free modules with homology
annihilated by a power of I. We say that an R-module M is a I-power torsion module if its
R∗ -module M∗ of homotopy groups is a I-power torsion module; equivalently, M∗ must have
support over I.
Lemma B.1. The R-module K∞ (I) is a homotopy direct limit of finite R-modules Ks (I),
each of which has homotopy groups annihilated by some power of I. Therefore K∞ (I) is a
I-power torsion module.
The following lemma is an analogue of the fact that Č • (I) is a chain complex which is a
finite sum of modules R[1/α] for α ∈ I.
Lemma B.2. The R-module Č(I) has a finite filtration by R-submodules with subquotients
that are suspensions of modules of the form R[1/α] with α ∈ I.
These lemmas are useful in combination.
Corollary B.3. If M is a I-power torsion module then M ⊗R Č(I) ≃ 0; in particular
K∞ (I) ⊗R Č(I) ≃ 0.
Proof: Since M[1/α] ≃ 0 for α ∈ I, Lemma B.2 gives the conclusion for M.
Appendix C. Completion at ideals and Bousfield localization
Bousfield localizations include both completions at ideals and localizations at multiplicatively closed sets, but one may view these Bousfield localizations as falling into the types
typified by completion at p and localization away from p. Thinking in terms of Spec(R∗ ),
this is best viewed as the distinction between localization at a closed set and localization at
the complementary open subset. In this section we deal with the closed sets and with the
open sets in Section D. This appendix is based on [57, 58, 48].
57
I. Homotopical completion. As observed in the proof of Lemma B.1, we have K∞ (α) =
holim Σ−1 R/αs and therefore
→
s
∧
M(α)
= HomR (holim Σ−1 R/αs , M) ≃ holim M/αs .
→
←
s
s
If I = (α, β), then
MI∧ = HomR (K∞ (α) ⊗R K∞ (β), M)
= HomR (K∞ (α), HomR (K∞ (β), M))
∧ ∧
= (M(β)
)(α) ,
and so on inductively. This should help justify the notation MI∧ = HomR (K∞ (I), M).
When R = S is the sphere spectrum and p ∈ Z ∼
= π0 (S), K∞ (p) is a Moore spectrum for
Z/p∞ in degree −1 and we recover the usual definition
Xp∧ = F (Σ−1 S/p∞ , X)
of p-completions of spectra as a special case, where F (A, B) = HomS (A, B) is the function
spectrum. The standard short exact sequence for the calculation of the homotopy groups of
Xp∧ in terms of ‘Ext completion’ and ‘Hom completion’ follows directly from Corollary A.6.
Since p-completion has long been understood to be an example of a Bousfield localization,
our next task is to show that completion at I is a Bousfield localization in general.
II. Bousfield’s terminology. Fix an R-module E. A spectrum A is E-acyclic if A ⊗R E ≃
0; a map f : X −→ Y is an E-equivalence if its cofibre is E-acyclic. An R-module M
is E-local if E ⊗R T ≃ 0 implies HomR (T, M) ≃ 0. A map Y −→ LE Y is a Bousfield Elocalization of Y if it is an E-equivalence and LE Y is E-local. This means that Y −→ LE Y
is terminal among E-equivalences with domain Y , and the Bousfield localization is therefore
unique if it exists. Similarly, we may replace the single spectrum E by a class E of objects
E, and require the conditions hold for all such E
The following is a specialization of a change of rings result to the ring map S −→ R.
Lemma C.1. Let E be a class of R-modules. If an R-module N is E-local as an R-module,
then it is E-local as an S-module.
Proof: If E ∧ T = E ⊗S T ≃ ∗ for all E, then E ⊗R (R ⊗S T ) ≃ 0 for all E and therefore
F (T, N) = HomS (T, N) ≃ HomR (R ⊗S T, N) ≃ 0.
III. Homotopical completion is a Bousfield localization. The class that will concern
us most is the class I-Tors of finite I-power torsion R-modules M. Thus M must be a
finite cell R-module, and its R∗ -module M∗ of homotopy groups must be a I-power torsion
module.
Theorem C.2. For any finitely generated ideal I of R∗ the map M −→ MI∧ is Bousfield
localization in the category of R-modules in each of the following equivalent senses:
(i) with respect to the R-module ΓI (R) = K∞ (I).
(ii) with respect to the class I-Tors of finite I-power torsion R-modules.
(iii) with respect to the R-module Ks (I) for any s ≥ 1.
58
Furthermore, the homotopy groups of the completion are related to local homology groups by
a spectral sequence
I
2
Es,t
= Hs,t
(M∗ ) =⇒ πs+t (MI∧ ).
If R∗ is Noetherian, the E 2 term consists of the left derived functors of I-adic completion:
HsI (M∗ ) = LIs (M∗ ).
Proof: We begin with (i). Since
HomR (T, MI∧ ) ≃ HomR (T ⊗R K∞ (I), M),
it is immediate that MI∧ is K∞ (I)-local. We must prove that the map M −→ MI∧ is a
K∞ (I)-equivalence. The fibre of this map is HomR (Č(I), M), so we must show that
HomR (Č(I), M) ⊗R K∞ (I) ≃ 0.
By Lemma B.1, K∞ (I) is a homotopy direct limit of terms Ks (I). Each Ks (I) is in I-Tors,
and we see by their definition in terms of cofibre sequences and smash products that their
duals DKs (I) are also in I-Tors, where DM = HomR (M, R). Since Ks (I) is a finite cell
R-module,
HomR (Č(I), M) ⊗R Ks (I) = HomR (Č(I) ⊗R DKs (I), M),
and Č(I) ⊗R DKs (I) ≃ 0 by Corollary B.3. Parts (ii) and (iii) are similar but simpler. For
(iii), observe that we have a cofibre sequence R/αs −→ R/α2s −→ R/αs , so that all of the
Kjs (I) may be constructed from Ks (I) using a finite number of cofibre sequences.
Appendix D. Localization away from ideals and Bousfield localization
In this section we turn to localization away from the closed set defined by an ideal I. First,
observe that, when I = (α), M[I −1 ] is just R[α−1 ] ⊗R M = M[α−1 ]. However, the higher
Čech cohomology groups give the construction for general finitely generated ideals a quite
different algebraic flavour, and M[I −1 ] is rarely a localization of M at a multiplicatively
closed subset of R∗ . This appendix is based on [48].
I. The Čech complex as a Bousfield localization. To characterize this construction as
a Bousfield localization, we consider the class I-Inv of R-modules M for which there is an
element α ∈ I such that α : M −→ M is an equivalence.
Theorem D.1. For any finitely generated ideal I = (α1 , . . . , αn ) of R∗ , the map M −→
M[I −1 ] is Bousfield localization in the category of R-modules in each of the following equivalent senses:
(i) with respect to the R-module R[I −1 ] = Č(I).
(ii) with respect to the class I-Inv.
(iii) with respect to the set {R[1/α1 ], . . . , R[1/αn ]}.
Furthermore, the homotopy groups of the localization are related to Čech cohomology groups
by a spectral sequence
2
Es,t
= ČHI−s,−t(M∗ ) =⇒ πs+t (M[I −1 ]).
59
If R∗ is Noetherian, the E 2 term can be viewed as the cohomology of Spec(R∗ ) \ V (I) with
coefficients in the sheaf associated to M∗ .
Remark D.2. One may also characterize the map ΓI (M) −→ M by a universal property
analogous to that of the cellular approximation in spaces: it is the K1 (I)-cellularization of
M.
Indeed, on the one hand, ΓI (M) is constructed from K1 (I) by B.1, and on the other hand,
the map induces an equivalence of HomR (K1 (I), ·) since, by Lemma B.2, HomR (K1 (I), Č(I)) ≃
0.
References
[1] J.F. Adams “Stable homotopy and generalized homology.” Chicago University Press 1974.
[2] L.Alonso Tarrio, A.Jeremias Lopez and J. Lipman “Local homology and cohomology on schemes.”
Ann.Scient. Ec.Norm.Sup. 30 (1997) 1-39
[3] L.Alonso Tarrio, A.Jeremias Lopez and J. Lipman “Studies in duality on Noetherian formal schemes
and non-Noetherian ordinary schemes” Cont. Math. 244 AMS (1999) x+126
[4] K.K.S.Andersen and J.Grodal “The classification of 2-compact groups” J American Math Soc 22
(2009), 387-436
[5] K.K.S.Andersen, J.Grodal, J.M.Møller and M.Viruel “The classification of p-compact groups for p
odd” Annals of Maths, 167 (2008), 95-210
[6] L.L.Avramov “Modules of finite virtual projective dimension.” Invent. Math. 96 (1989), no. 1, 71–101.
[7] L.L.Avramov “Infinite free resolutions.” Six lectures on commutative algebra (Bellaterra, 1996), 1–118,
Progr. Math., 166, Birkhuser, Basel, 1998.
[8] L.L. Avramov “Locally complete intersection homomorphisms and a conjecture of Quillen on the
vanishing of cotangent homology.” Ann. of Math. (2) 150 (1999), no. 2, 455–487.
[9] L.L.Avramov and R.-O. Buchweitz “Support varieties and cohomology over complete intersections.”
Invent. Math. 142 (2000), no. 2, 285–318.
[10] L.L. Avramov and H.-B. Foxby “Locally Gorenstein homomorphisms.” American J. Math. 114 (1992)
1007-1047
[11] L.L. Avramov and H.-B. Foxby “Ring homomorphism and finite Gorenstein dimension” PLMS 75
(1997) 241-270
[12] L.L.Avramov and S. Halperin “Through the looking glass: a dictionary between rational homotopy
theory and local algebra.” Algebra, algebraic topology and their interactions (Stockholm, 1983), 1-27,
Lecture Notes in Math., 1183, Springer, Berlin, 1986.
[13] H.Bass “On the ubiquity of Gorenstein rings” Math. Z. 82 1963 8–28.
[14] H.Bass “Algebraic K-theory.” Benjamin (1968) xx+762pp
[15] D.J.Benson “An algebraic model for chains on ΩBG∧
p ” Trans. Amer. Math. Soc. 361 (2009), no. 4,
2225-2242.
[16] D.J.Benson and J.F.Carlson “Projective resolutions and Poincaré duality complexes. ” Trans. American
Math. Soc. 342 (1994) 447-488
[17] D.J.Benson and J.F.Carlson “Functional equations for Poincaré series in group cohomology.” Bull.
London Math. Soc. 26 (1994) 438-448.
[18] D.J.Benson and J.P.C.Greenlees “Commutative algebra for the cohomology of classifying spaces of
compact Lie groups.” J.Pure and Applied Algebra 122 (1997) 41-53
[19] D.J.Benson and J.P.C.Greenlees “Commutative algebra in the cohomology of virtual duality groups.”
J. Algebra 192 (1997) 678-700.
[20] D.J.Benson and J.P.C.Greenlees “Localization and duality in topology and modular representation
theory.” J. Pure Appl. Algebra 212 (2008), no. 7, 17161743.
[21] D.J.Benson and J.P.C.Greenlees “Complete intersections and derived categories.” Preprint (2009)
17pp.
60
[22] D.J.Benson, J.P.C.Greenlees and S.Shamir “Complete intersections and mod p cochains.” AGT 13
(2013) 61-114, DOI 10.2140/agt.2013.13.61, arXiv:1104.4244
[23] J.M. Boardman “Stable homotopy theory” Mimeographed notes 1966–1970.
[24] A.K. Bousfield. “The localization of spaces with respect to homology.” Topology 14 (1975), 133-150.
[25] A.K. Bousfield. “The localization of spectra with respect to homology.” Topology 18 (1979), 257-281.
[26] A.K. Bousfield and D.M. Kan. “Homotopy limits, completions and localizations.” Springer Lecture
notes in mathematics Vol. 304. 1972.
[27] E.H. Brown “Abstract homotopy theory.” Trans. Amer. Math. Soc. 119 (1965) 79–85.
[28] W.Bruns and Herzog “Cohen-Macaulay rings.” Cambridge Studies in Advanced Mathematics, 39.
Cambridge University Press, Cambridge, 1993. xii+403 pp.
[29] V.Buchstaber and T.Panov “Torus actions and their applications in topology and combinatorics” AMS
(2002)
[30] E. Dror-Farjoun “Cellular spaces, null spaces and homotopy localization.” Lecture Notes in Mathematics, 1622. Springer-Verlag, Berlin, 1996. xiv+199 pp.
[31] W.G.Dwyer “Strong convergence of the Eilenberg-Moore spectral sequence” Topology 13 (1974) 255265
[32] W.G.Dwyer and J.P.C.Greenlees “Complete modules and torsion modules” American J. Math. 124
(2002) 199-220
[33] W.G.Dwyer, J.P.C.Greenlees and S.B.Iyengar “Duality in algebra and topology.” Advances in Maths
200 (2006) 357-402
[34] W.G.Dwyer and J.P.C.Greenlees and S.B.Iyengar “Finiteness conditions in derived categories of local
rings.” Comm. Math. Helv. 81 (2006) 383-432
[35] W.G.Dwyer and J.P.C.Greenlees and S.B.Iyengar “Gross-Hopkins duality and the Gorenstein condition”. J. K-Theory 8 (2011), no. 1, 107133.
[36] W.G.Dwyer and J.P.C.Greenlees and S.B.Iyengar “DG algebras with exterior homology.” Bull. Lond.
Math. Soc. 45 (2013), no. 6, 12351245.
[37] W.G.Dwyer and C.Wilkerson “Homotopy fixed-point methods for Lie groups and finite loop spaces”
Ann. Math. 139 (1994) 395-442
[38] D.Dugger and B.E.Shipley “K-theory and derived equivalences.” Duke Math. J. 124 (2004), no. 3,
587–617
[39] A.D. Elmendorf, J.P.C. Greenlees, I. Kriz and J.P. May. “Commutative algebra in stable homotopy
theory and a completion theorem.” Math. Res. Letters 1 (1994) 225-239.
[40] A. D. Elmendorf, I. Kriz, M. A. Mandell, and J. P. May. “Rings, Modules and Algebras in Stable
Homotopy Theory”, Volume 47 of Amer. Math. Soc. Surveys and Monographs. American Mathematical
Society, 1996.
[41] A. D. Elmendorf, and J. P. May “Algebras over equivariant sphere spectra” J. Pure and Applied
Algebra 116 (1997) 139-149
[42] Y. Félix, S. Halperin and J.-C. Thomas “Gorenstein spaces” Advances in Mathematics, 71, (1988)
92-112
[43] Yves Félix, Stephen Halperin, and Jean-Claude Thomas, “Elliptic Hopf algebras”, J. London Math.
Soc. (2) 43 (1991), no. 3, 545–555.
[44] P.G.Goerss and J.F.Jardine “Simplicial homotopy theory.” Progress in Mathematics, 174. Birkhauser
Verlag, Basel, 1999. xvi+510 pp. ISBN: 3-7643-6064-X
[45] D.J.Green and S.King “The cohomology of finite p-groups” http://users.minet.uni-jena.de/cohomology
[46] J.P.C.Greenlees “Commutative algebra in group cohomology.” J.Pure and Applied Algebra 98 (1995)
151-162
[47] J.P.C.Greenlees “K-homology of universal spaces and local cohomology of the representation ring”
Topology 32 (1993) 295-308.
[48] J.P.C.Greenlees “Tate cohomology in commutative algebra.” J. Pure Appl. Algebra 94 (1994), no. 1,
59–83.
[49] J.P.C.Greenlees “Commutative algebra for group cohomology.” JPAA 98 (1995) 151-162
[50] J.P.C.Greenlees “Tate cohomology in axiomatic stable homotopy theory.” Cohomological methods in
homotopy theory (Bellaterra, 1998), 149–176, Progr. Math., 196, Birkhuser, Basel, 2001.
61
[51] J.P.C.Greenlees “Local cohomology in equivariant topology.” Local cohomology and its applications
(Guanajuato, 1999), 1–38, Lecture Notes in Pure and Appl. Math., 226, Dekker, New York, 2002.
[52] J.P.C.Greenlees “Spectra for commutative algebraists.” Proceedings of the 2004 Chicago Summer
School Contemporary Mathematics 436 (2007) 149-173, arXiv:math/0609452
[53] J.P.C.Greenlees “First steps in brave new commutative algebra” Proceedings of the 2004 Chicago
Summer School Contemporary Mathematics, 436 (2007) 239-275, arXiv:math/0609453
[54] J.P.C.Greenlees “Ausoni-Bökstedt duality for topological Hochschild homology” Journal of Pure and
Applied Algebra 220 (2016), pp. 1382-1402
[55] J.P.C.Greenlees, K.Hess and S.Shamir “Complete intersections in rational homotopy theory.” JPAA
217 (2013) 636-663, arXiv:0906.3247
[56] J.P.C.Greenlees and G.Lyubeznik “Rings with a local cohomology theorem with applications to cohomology rings of groups.” JPAA 149 (2000) 267-285
[57] J.P.C. Greenlees and J.P. May. “Completions of G-spectra at ideals of the Burnside ring.” Proc. Adams
Memorial Symposium, Volume II, CUP (1992), 145-178.
[58] J.P.C. Greenlees and J.P. May. “Derived functors of I-adic completion and local homology.” J. Algebra
149 (1992), 438-453.
[59] J.P.C.Greenlees and J.P. May “Equivariant stable homotopy theory”. Handbook of algebraic topology,
277–323, North-Holland, Amsterdam, 1995.
[60] J.P.C.Greenlees and J.P. May “Completions in algebra and topology.” Handbook of algebraic topology,
255–276, North-Holland, Amsterdam, 1995.
[61] J.P.C.Greenlees and B.E.Shipley “An algebraic model for free rational G-spectra for compact connected
Lie groups G.” Math Z 269 (2011) 373-400, DOI 10.1007/s00209-010-0741-2
[62] J.P.C.Greenlees and B.E.Shipley “An algebraic model for free rational G-spectra.” Bull. LMS 46 (2014)
133-142, DOI 10.1112/blms/bdt066, arXiv:1101.4818
[63] J.P.C.Greenlees and G.Stevenson “Morita theory and singularity categories” In preparation (2015)
17pp
[64] J.P.C.Greenlees and V.Stojanoska “Local and global duality in chromatic homotopy theory.” In preparation (2015) 37pp
[65] A. Grothendieck. “Sur quelques points d’algèbre homologique”. Tohoku Mathematical Journal 9
(1957), 119-221.
[66] A. Grothendieck (notes by R. Hartshorne). “Local cohomology.” Springer Lecture notes in mathematics
Vol. 41. 1967.
[67] T.H. Gulliksen “A homological characterization of local complete intersections.” Compositio Math. 23
(1971), 251–255.
[68] T.H.Gulliksen “A change of ring theorem with applications to Poincaré series and intersection multiplicity.” Math. Scand. 34 (1974), 167–183.
[69] T.H. Gulliksen “On the deviations of a local ring.” Math Scand 47 (1980) 5-20
[70] T.H.Gulliksen and G.Levin “Homology of local rings.” Queen’s Paper in Pure and Applied Mathematics, No. 20 Queen’s University, Kingston, Ont. 1969 x+192 pp.
[71] R. Hartshorne. “Algebraic Geometry”. Springer-Verlag. 1977.
[72] M.J. Hopkins. “Global methods in homotopy theory.” Proc. of the 1985 LMS Symposium on Homotopy
Theory. London Math. Soc. 1987, 73-96.
[73] M.Hovey “Model categories.” Mathematical Surveys and Monographs, 63. American Mathematical
Society, Providence, RI, 1999. xii+209 pp. ISBN: 0-8218-1359-5
[74] M. Hovey, B.E.Shipley and Jeff Smith “Symmetric spectra.” J. Amer. Math. Soc. 13 (2000), no. 1,
149–208.
[75] R.Levi “On finite groups and homotopy theory.” Mem. Amer. Math. Soc. 118 (1995), no. 567, xiv+100
pp.
[76] R.Levi “A counter-example to a conjecture of Cohen.” Algebraic topology: new trends in localization
and periodicity (Sant Feliu de Guxols, 1994), 261–269, Progr. Math., 136, Birkhuser, Basel, 1996.
[77] R.Levi “On homological rate of growth and the homotopy type of ΩBG∧
p .” Math. Z. 226 (1997), no.
3, 429–444.
62
[78] R.Levi “On p-completed classifying spaces of discrete groups and finite complexes.” J. London Math.
Soc. (2) 59 (1999), no. 3, 1064–1080.
[79] L.G.Lewis, J.P.May and M.Steinberger (with contributions by J.E.McClure) “Equivariant stable homotopy theory” Lecture notes in mathematics 1213, Springer-Verlag (1986)
[80] M.A.Mandell “E∞ algebras and p-adic homotopy theory.” Topology 40 (2001), no. 1, 43–94.
[81] M.A.Mandell, J.P.May, S.Schwede, and B.E.Shipley “Model categories of diagram spectra.” Proc.
London Math. Soc. (3) 82 (2001), no. 2, 441–512.
[82] A. Mathew “The homology of tmf” Homology, Homotopy and Applications 18 (2016) (to appear)
arXiv:1305.6100
[83] H. Matsumura. “Commutative ring theory.” Cambridge Univ. Press. 1986.
[84] J.P.May “A concise course in algebraic topology.” University of Chicago Press, Chicago, IL, 1999.
x+243 pp. ISBN: 0-226-51182-0; 0-226-51183-9
[85] J.Rognes “Galois extensions of structured ring spectra. Stably dualizable groups.” Mem. Amer. Math.
Soc. 192 (2008), no. 898, viii+137 pp.
[86] D.J. Rusin “The cohomology of the groups of order 32.” Math. Comp. 53 (1989), no. 187, 359385.
[87] P. Schenzel “Proregular sequences, local cohomology, and completion.” Math. Scand. 92 (2003), no. 2,
161–180.
[88] S. Schwede, “Morita theory in abelian, derived and stable model categories.” Structured ring spectra,
33–86, London Math. Soc. Lecture Note Ser., 315, Cambridge Univ. Press, Cambridge, 2004.
[89] S. Schwede “Symmetric spectra” Book in preparation, available from author’s webpage.
[90] S. Schwede and B.E.Shipley “Stable model categories are categories of modules.” Topology 42 (2003),
no. 1, 103–153.
[91] S. Schwede and B.E.Shipley “Algebras and modules in monoidal model categories.” Proc. London
Math. Soc. (3) 80 (2000), no. 2, 491-511.
[92] B.E.Shipley “HZ-algebra spectra are differential graded algebras.” American J. Math 129 (2007) 351379
[93] E.H.Spanier and J.H.C.Whitehead “A first approximation to homotopy theory.” Proc. Nat.Acad. Sci.
U. S. A. 39, (1953). 655–660
The University of Sheffield, Sheffield, S3 7RH UK
E-mail address: [email protected]
63
| 0math.AC
|
The N-Tuple Bandit Evolutionary Algorithm
for Game Agent Optimisation
Simon M Lucas, Jialin Liu and Diego Perez-Liebana
arXiv:1802.05991v1 [cs.NE] 16 Feb 2018
School of Electrical Engineering and Computer Science
Queen Mary University of London
London, UK
Email: {simon.lucas, jialin.liu, diego.perez}@qmul.ac.uk
Abstract—This paper describes the N-Tuple Bandit Evolutionary Algorithm (NTBEA), an optimisation algorithm developed
for noisy and expensive discrete (combinatorial) optimisation
problems. The algorithm is applied to two game-based hyperparameter optimisation problems. The N-Tuple system directly
models the statistics, approximating the fitness and number of
evaluations of each modelled combination of parameters. The
model is simple, efficient and informative. Results show that the
NTBEA significantly outperforms grid search and an estimation
of distribution algorithm.
Index Terms—Keywords: Estimation of Distribution Algorithm, Evolutionary Algorithm, Hyper-Parameter Optimisation,
Rolling Horizon Evolution, Game Playing Agent, Noisy Optimisation.
I. I NTRODUCTION
This paper describes the N-Tuple Bandit Evolutionary Algorithm (NTBEA) and its application to optimising the parameters of a rolling horizon evolution game-playing agent.
The NTBEA combines evolutionary search with Multi-Armed
Bandit algorithms (MABs) in order to provide an algorithm
which is robust to noise, has an explicit way to balance the
trade-off between exploration and exploitation, and provides
a statistical model of the fitness landscape as an additional
output.
The applications of this type of algorithm are numerous. In
our research we have already applied it successfully to hyperparameter optimisation [1] and automated game tuning [2].
Furthermore, if the inherent fitness landscape is flat, then the
exploration term provides a means for performing novelty
search [3].
A. Estimation of Distribution Algorithms
Estimation of Distribution Algorithms (EDAs) [4], [5], [6],
[7], [8] are a powerful class of Evolutionary Algorithms (EAs).
Instead of using the mutation and crossover operators to
generate offspring from the fittest parents, the population is
sampled from an estimated probability distribution of selected
individuals, which is updated iteratively using the fitness
evaluations of individuals. In addition to potentially making
the search more efficient and robust to noise, EDAs also have
the benefit of learning a model which offers greater insight into
the nature of the problem. The model may provide information
regarding key parameter combinations that tend to lead to good
or bad solutions.
The NTBEA models not only the fitness of points in the
search space, but also estimates how much each point has been
visited. This enables explicit modelling of exploration, which
has two distinct benefits. First, it enables a principled way to
avoid becoming trapped in local optima, since the search can
be diverted to less well explored areas. Second, it provides
a natural way to perform a type of novelty search. This is
especially important when dealing with problems that involve
a flat reward landscape [9].
B. Combinatorial Multi-Armed Bandits
Rolet and Teytaud [10] modeled the sub-domains of the
search space as arms of a bandit and allocated the computational resource dynamically during the evaluation step of an
evolution strategy for continuous noisy optimization. When
searching in a discrete space, the NTBEA has significant similarity to the Combinatorial Multi-Armed Bandits (CMABs),
formalised by Gai et al. [11] and Chen et al. [12]. Gai et
al. applied CMABs to a channel allocation problem in mobile communications, and showed how considering pair-wise
interactions led to better performance than a naive univariate
model, but their model was particular to resource allocation
problems. The recent work of Ontañón [13] on CMABs is
the closest to this paper. Ontañón used CMABs for multiagent control in the microRTS game, a relatively simple but
highly challenging real-time strategy (RTS) game. Ontañón’s
CMABs model each dimension of the search space as an
individual local MAB, in which each arm is a legal value
of the corresponding dimension, together with a global MAB
with all the legal value combinations as arms. These value
combinations are sometimes called macro-arms.
C. N-Tuple Systems
In this paper we use an N-Tuple system to model the
fitness landscape, and combine it with an evolutionary search
algorithm. N-Tuple systems were first developed and applied
to pattern recognition in the late 1950s [14], and among many
other applications have also been applied to model value
functions in games [15], [16]. The standard way of using an
N-Tuple system is to model a high-dimensional feature space
by considering a set of lower-dimensional projections of that
space, defined by the set of N-Tuples. Each N-Tuple forms a
weak model of the space, but combining them provides a good
model. There are some novel aspects of the N-Tuple systems
used in this paper that will be described later in Section II.
The notion of successively improving an individual or
a population via variation operators and fitness evaluations
makes the NTBEA distinct from the standard CMAB methods.
A CMAB operates via a pure bandit-based sampling procedure, whereas the approach described in this paper uses an
evolutionary algorithm (EA) to perform the population generation and variation, and a bandit landscape fitness model (for
brevity called just the model or the model space) to perform
the selection. In this paper an N-Tuple system is used as the
model, but other choices of model would also be possible. NTuple systems are a good choice due to their speed, simplicity,
interpretability and reasonable accuracy. Essentially, the EA
searches the model space to improve the sample efficiency in
the problem domain.
In this paper, a variation of the (1, λ)-EA is used, in which
the current best individual is selected as the parent only to
generate the next population of size λ and is not included in
the new population.
The instance of the algorithm described in this paper is similar to the N-Tuple Bandit Evolutionary Algorithm introduced
by Kunanusont et al. [2], extended from the Bandit-based
Random Mutation Hill Climber [17]. Kunanusont et al. [2]
applied the NTBEA to automatic game parameter tuning of a
two-player adversarial video game, a very noisy problem, as
some stochastic AI agents were used to evaluate the evolved
stochastic games. The NTBEA significantly outperformed
two evolutionary algorithms, a simple Random Mutation Hill
Climber and a univariate Bandit EA [2]. The NTBEA was
also applied within the General Video Game AI (GVGAI [18])
framework for evolving game rules and parameters, and has
successfully evolved variations of the game Aliens to favour
either a short-term or a long-term planning agent [19].
In this paper, we formalise the NTBEA, and present its
applications for hyper-parameter optimisation. The NTBEA
makes use of the statistics of the previously evaluated solutions
and balances the trade-off between exploring a new (or less
evaluated) solution and exploiting the best one found so
far. The NTBEA is particularly useful when the evaluation
function is expensive or noisy. For example, in the context
of automatic game design, a game needs to be played several
times to approximate an accurate win rate if the game or the
agents are stochastic, while the execution time of a single
game may take a significant amount of time or it is difficult to
perform an evaluation. The notion of expensive in this context
is clearly subjective, but really means any problem where the
evaluation time is a significant issue in the opinion of the
algorithm user.
Using N-Tuples to approximate fitness statistics and visit
counts can be related to locally sensitive hashing (LSH), and
it is worth noting how the N-Tuple system described in this
paper could also be used for enhancing exploration in video
games, similar to the way LSH in the form of Context Tree
Switching (CTS) was applied to boost performance on the
challenging game of Montezuma’s Revenge [20].
D. Hyper-Parameter Optimisation
The task of tuning the parameters of a game-playing agent
can be viewed as a Hyper-Parameter Optimisation problem.
Bergstra and Bengio [21] state how the most popular methods
are Grid Search and manual search, but show how competitive
the performance of random search is for these problems.
More recently the Hyperband paper [22] shows how recent
approaches such as Sequential Model-based Algorithm Configuration (SMAC) [23] and Tree-structured Parzen Estimator
(TPE; [24]) in some cases only perform similarly to random
search. They outperformed random search in a direct rankbased comparison, but only by a small margin. When the
random search was repeated twice and the best solution was
picked, random search performed best [22] (albeit not quite a
fair comparison, but one that illustrates the small margin of
improvement).
The NTBEA approach has a similar overall algorithmic
framework to SMAC [23], in the way a model is iteratively
updated and used to decide which point in the search space to
sample next. Where SMAC uses random forests for the model,
the NTBEA uses N-Tuple systems. There are also significant
differences in the approaches; SMAC only estimates the value
of the objective function for each parameter configuration,
whereas NTBEA also estimates the visit count, which is
essential for the bandit-based exploration pressure. There are
also some conceptual differences, with SMAC running T trials
of each sample configuration (via its INTENSIFY procedure)
while by default NTBEA runs a single evaluation each time,
in order to make maximum use of the bandit model.
Other aspects of hyper-parameter optimisation in general
involve early stopping of unpromising solutions, and also
sharing of the system parameters between competing solutions: both of these are used to great effect by Jadeberg et
al [25]. Here we only consider when each solution is fully
evaluated at least once, as in a complete game is played,
though actually games could be abandoned very early when
one player is obviously weak. For now we note that there
are many approaches to hyper-parameter optimisation, but that
manual and grid search are in strong use, and that random
search performs surprisingly well.
The rest of this paper is structured as follows. Section II
introduces the overall structure of the NTBEA. The test
problem is described in Section III, then results are illustrated
and discussed in Section IV. Section V concludes and lists
some future work.
II. T HE N-T UPLE BANDIT E VOLUTIONARY A LGORITHM
In this section, we describe the system architecture of the
algorithm and then explain each part in more detail. Figure 1
illustrates the three key components of the NTBEA, a bandit
landscape model, an evolutionary algorithm and a noisy evaluator, i.e., a fitness function corrupted by noise. We assume that
the execution time of querying the bandit landscape model is
negligible compared to evaluating a candidate solution on the
target problem.
Fig. 1: Key components of the N-Tuple Bandit Evolutionary
Algorithm (NTBEA). A model of the fitness landscape is built
and updates iteratively using the evaluations of solutions. The
search for candidate solutions is performed in the model space,
i.e., the N-Tuple Bandit Fitness Landscape model, where
candidate solutions can be evaluated quickly (N-Tuple systems
are well known for their speed [26], and our experimental
tests support this). This is then sampled in the real (relatively
expensive) problem search space, where the fitness value is
expected to be noisy. The model is then updated with the
fitness value of this solution point. The process repeats until
some termination condition is met.
The algorithm works as follows. The search begins by
sampling a single solution point uniformly at random in the
search space. This search point is referred to as the current
point. The fitness of the current point is evaluated once, in
the problem domain, using the noisy evaluator. In theory, no
resampling is needed, even for noisy problems, since the UCB
module in the algorithm aims to take care of re-evaluating an
identical point if needed. Note that the algorithm also works
for noise-free problems, though we focus on noisy problems
in this paper.
The current point is then stored in the bandit fitness landscape model (referred to as the model from now on), together
with its fitness value. The model is then searched within
the neighborhood of the current point. The neighborhood is
defined using the number of neighbors and the proximity
distribution to the current point, which is controlled by a
mutation operator. The solution point in the neighborhood
with the highest estimated Upper Confidence Bounds value
(UCB; defined later in Section II-A) is then selected as the
new current point. The process iterates until the evaluation
budget is used up, or some other termination condition is met.
The algorithm is described in Algorithm 1.
Note that this description outlines the simplest approach
where only a single solution is held as the current focus of
the search, similar to using only one parent. Population-based
versions are also possible, and would be more appropriate for
parallel hardware.
A. Estimating UCB Values: an N -Tuple Approach
A key part of the algorithm is the value function used to
sample in a large search space. In this context, large means that
the size of the search space is larger than the number of fitness
evaluations allowed, thus it is impossible to properly evaluate
(proper evaluation may involve resampling due to noise) each
of the possible solutions points. Instead, we need to model the
relationship between points in the search space and sample
accordingly.
The algorithm UCB1, a simple multi-armed bandit algorithm, is introduced by Auer et al. [27]. The UCB value of
any arm i is defined as:
r
ln n
.
(1)
UCBi = X̂i + k
ni +
The empirical mean reward for playing arm i is X̂i : this is the
exploitation term. The right hand term controls the exploration,
where n is the total number of times this bandit has been
played, and ni is the number of times the arm i has been
played. The term k is called the exploration factor: higher
values of k lead to explorative search, low values lead to more
greedy or exploitative search. Each dimension of the search
space is modelled as an independent MAB, with an arm for
each possible value. The value is used to control whether
each arm should be pulled at least once. In the standard UCB
formula, is set to zero ensuring that each arm is pulled once
in turn, but for our purposes this would be impractical, as the
macro-arm consisting of the entire N -Tuple would force an
exhaustive exploration of the search space.
Additionally, we also model combinations of arms as superbandits. For example, in a d-dimensional search space where
each dimension has n possible values, the 10-wide superbandit would have nd arms.
We modify this to be an aggregate over all the N -Tuples in
the N -Tuple System model. Let N be the N -Tuple indexing
function such that Nj (x) indexes the j th bandit for search
space point x. Initially we compute the aggregate UCB value
for solution point x as an unweighted arithmetic average, as
defined in (2):
m
vU CB (x) =
1 X
UCBNj (x) ,
m j=1
(2)
where m denotes the total number of bandits in the system.
More sophisticated algorithms are possible, for example by
combining the individual outputs using a Bayesian method,
but for simplicity and proof of concept we begin by using the
arithmetic average.
The N-Tuple systems have ideal properties for use as fitness
landscape models, in that they offer an extremely fast one-shot
training and good accuracy. They are ideally suited to modelling discrete spaces, but can also be applied to continuous
spaces with some degree of compromise. In this paper we are
dealing with discrete search spaces, so they are already a good
fit.
The concept is as follows. Given a d-dimensional search
space, we sub-sample its dimensions with a number of N tuples. The value of N ranges from 1 up to d, though may
miss out values in between. If all the tuples were considered,
then the total number of bandits is 2d .
In the standard N-Tuple systems each entry in the look-uptable stores a single value for each class, normally related to
the probability of that index occurring given that class, or for
game position value function approximation, the value of that
index occurring.
Algorithm 1 The N-Tuple Bandit Evolutionary Algorithm.
Require: S: search space
Require: f itness: noisy solution evaluator
Require: n ∈ N + : number of neighbors
Require: p ∈ (0, 1): mutation probability
Require: f lipOnce ∈ {true, f alse}: flip at least once or not
Require: nbEvals: total number of evaluations allowed
1: t = 0
. Counter for fitness evaluations
2: M odel ← ∅
. Initialise the fitness landscape model
3: current ← a random point ∈ S
4: while t < nbEvals do
5:
value ← f itness(current)
6:
add < current, value > to LM odel
7:
8:
9:
P opulation ← N EIGHBORS(M odel, current, n, p, f lipOnce)
11:
12:
function N EIGHBORS(model, x, n, p, f lipOnce)
P opulation ← ∅
. Initialise empty set
d ← |x|
. Get the dimension
for k ∈ {1, . . . , n} do
neighbor ← x
i←0
if f lipOnce then
i ← randomly selected from {1, 2, . . . , d}
for j ∈ {1, . . . , d} do
if i == j then
Randomly mutate value of neighborj
else
if RAN D < p then
Randomly mutate value of neighborj
Add neighbor to P opulation
return P opulation
current ← arg maxx∈P opulation vU CB (x)
t←t+1
10: return LM odel
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
In our model, however, each N-Tuple has a look-up table
(LUT) that stores statistical summaries of the values it encounters; the basic numbers stored are the number of samples, the
sum of the fitness of samples, and the sum of the square of
the fitness of the samples. This enables the mean, the standard
deviation and the standard error to be calculated for each entry
in the table. This provides all we need for (1) and (2), and
beyond: calculating the mean, standard deviation or standard
error for each parameter combination being modelled provides
useful insight into the nature of the system under optimisation.
B. Algorithm
The NTBEA algorithm is outlined in Algorithm 1 and
operates as follows. It begins by choosing a random solution
point in the search space, which is called the current point.
Since we are dealing with discrete search spaces, each point
is represented as a vector of integers, where each element
of the vector is an index to the currently selected value in
that dimension. The actual value chosen may be of any type,
common types are integer, double and boolean.
The following steps are then repeated until the fitness
evaluation budget has been exhausted:
• It makes a (noisy) fitness evaluation of the current point
and stores it in the N-Tuple Fitness Landscape Model
as the value for that solution point (lines 5 and 6 in
Algorithm 1).
• Using a mutation operator to generate a set of unique
neighbors of the current solution, as the population of
current iteration (line 7).
• Using the fitness landscape model, the algorithm calculates the estimated UCB value of each solution by (1)
and (2). Then, it sets the current solution as the one in
the population (neighbors from the previous step) with
the highest estimated UCB value (line 8).
When the fitness evaluation budget has been exhausted, the
method searches and recommends a neighbor of all of the
evaluated solutions, in which each dimension is set to the value
with maximal approximate value defined in (2).
C. Illustrative Example
Consider a 5-dimensional space, modeled using five 1-tuples
and one 5-tuple. Suppose we now enter the four vectors (three
unique) in the search space together with their associated
fitness values as given in Table Ia.
The first 1-tuple will have a LUT that has 2 entries, with
LU T [0 ∗ ∗ ∗ ∗] having a mean of 1, and LU T [1 ∗ ∗ ∗ ∗]
having a mean of 32 . The single 5-tuple will have three nonempty entries, with LU T [12340] having a mean of 0.5, and
LU T [11111] and LU T [00110] both having means of 1. Note
that the object at each index is a statistical summary object as
described above that does not directly store the mean but can
calculate it; we describe the mean value to best illustrate the
operation, and because it feeds directly into (1) and (2). An
example of some of the statistics that are stored in the system
or can be calculated directly is shown in Table Ib.
In this example only the empirical average and the number
of evaluations is output for each index of each N -tuple, though
as mentioned previously the standard deviation and standard
error are also available. For each N -tuple, only the non-null
table entries are shown, i.e., ones in which the index (can also
be thought of as a pattern) occurs at least once.
III. N OISY O PTIMISATION OF G AME AGENT PARAMETERS
This section describes an application to optimising the
parameters of a rolling horizon evolutionary game-playing
agent. The agent is optimised to get as high average score on
the game as possible. This is a noisy optimisation problem
with two distinct sources of noise: (i) the game itself is
stochastic; (ii) the optimised agent follows a stochastic policy,
such that given the same game state, it may play differently
if simulating more than once.
The behavior of the agent is controlled by a number of
parameters, some of which may have a critical effect on
TABLE I: Illustrative example, in which only 1-tuples and
d-tuple are considered (d = 5).
•
(a) Evaluated solutions (entries) and corresponding fitness values.
Solution
[1, 2, 3, 4, 0]
[1, 1, 1, 1, 1]
[0, 0, 1, 1, 0]
[1, 2, 3, 4, 0]
fitness
1
1
1
0
•
(b) Some of the statistics that are stored in the system or can be
calculated directly.
N -tuple
1-tuple
5-tuple
Pattern
[0, ∗, ∗, ∗, ∗]
[1, ∗, ∗, ∗, ∗]
[∗, 0, ∗, ∗, ∗]
[∗, 1, ∗, ∗, ∗]
[∗, 2, ∗, ∗, ∗]
[∗, ∗, 1, ∗, ∗]
[∗, ∗, 3, ∗, ∗]
[∗, ∗, ∗, 1, ∗]
[∗, ∗, ∗, 4, ∗]
[∗, ∗, ∗, ∗, 0]
[∗, ∗, ∗, ∗, 1]
[0, 0, 1, 1, 0]
[1, 1, 1, 1, 1]
[1, 2, 3, 4, 0]
Mean
1
2
3
1
1
1
2
1
1
2
1
1
2
2
3
1
1
1
1
2
Nb. of eval.
1
3
1
1
2
2
2
2
2
3
1
1
1
2
the performance of the agent, and where the combination of
parameter values is important.
The problem addressed is similar to the hyper-parameter
optimisation problem that is topical in machine learning; it has
been shown many times that optimising a known architecture
can produce state of the art results (e.g., [24], [21]).
A. Rolling Horizon Evolutionary Agent
We aim to optimise the parameters of a rolling horizon
evolutionary agent. Rolling Horizon Evolutionary Algorithms
(RHEAs) are called rolling horizon as each of the individuals
in the population is an action sequence of a fixed planning
(time) horizon, hp , thus RHEAs plan ahead hp actions. On
each subsequent game step, the horizon rolls one step further
within the hp window. Every individual is evaluated by evaluating the state after simulating hp actions in its corresponding
sequence or earlier, if a termination state is reached before
performing all the hp actions. Then, only the first action of
the best individual is applied. This procedure repeats with the
updated state.
The rolling horizon technique can be integrated to different evolutionary algorithms such as Genetic Algorithms and
Coevolutionary Algorithms [28]. In this work, we combined
a simple (1 + 1)-EA with rolling horizon. This agent will be
referred to as RHEA in the rest of the paper.
The performance of RHEAs can be boosted in many ways.
For instance, recently, RHEAs have been applied to General
Video Game Playing, and proved that the population size and
sequence length (i.e., planning horizon) had significant effect
on the performance of the algorithms [29], [30], [31]. Some
of the key parameters are
• sequence length: planning horizon;
•
•
shift buffer enabled or not: if disabled, at any timestep
t + 1, the initial population is reset to random, otherwise,
each of the individuals from the population at timestep
t (previous optimisation procedure) shifts its action sequence forward and fills the last position by a random
action;
resampling number: in the stochastic case, how many
time an individual is re-evaluated (equals to 1 if no reevaluation is allowed).
mutation probability: how likely a mutation occurs at
every dimension;
flip at least one bit: indicates if at least one mutation
should occur at each time.
B. Test Problems
We optimise the rolling horizon evolutionary agent to play
two games: simplified Java implementations of Asteroids and
Planet Wars. In these test cases, the RHEA agent model
sequences of actions in next game ticks as its individuals, and
evolve the individuals over time aiming at maximising some
fitness value.
1) Asteroids: Asteroids, released to great acclaim in 1979
is one of the classic video games of all time, and Atari’s
most profitable1 . The original game was implemented in
special vector graphics hardware, featured memorable sound
effects and smooth physics-based movements. The challenge
for players is to avoid being hit by asteroids, while aiming at
them accurately and to also hit the flying saucers before they
hit the player’s spaceship. There are three sizes of asteroids:
large, medium and small. Each screen starts with a number
of large rocks: as each one is hit either by the player, or by
a missile fired by an enemy flying saucer, it splits in to a
smaller one: large rocks split into two medium ones, medium
into two small ones, small ones disappear. There is also a
score progression, with the score per rock increasing as the
size decreases.
Strong play requires good perceptual and motor skills for
avoiding collisions and shooting accurately at the targets.
There are also some interesting strategies which may easily
elude novice players. One of them is to control the number of
asteroids on the screen by shooting one large one at a time,
then picking off a single medium rock and each small one it
gives rise to, before breaking another large rock.
The largest score for any item is the 1, 000 points awarded
for shooting the small flying saucer. These fire at the player
with deadly accuracy, so it is important to shoot them quickly
as the missiles they fire are fast moving and tricky to avoid.
Many strong players will shoot all but one remaining asteroid
(a small one) and then lie in wait for flying saucers at the
edge of the screen, firing at them immediately as soon as they
appear. There is some devil in the detail here, regarding the
best place to wait and the best angle to fire at (the screen
wraps around both vertically and horizontally, so from a single
position it is possible to shoot at both sides).
1 https://en.wikipedia.org/wiki/Asteroids_(video_game)
For our experiments we used a Java implementation of the
game where we can simulate the game 10, 000 times faster
than the real game. Although using the ALE or MAME version
of the game would have been possible, there are some distinct
advantages to having our own implementation:
•
•
It runs much faster than the ALE or MAME versions.
Having access to the source code makes it straightforward
to vary details of the game in order to test various
hypotheses. For example, we can test the effects of penalising each missile fired, or giving the large and medium
rocks zero value to test the long-term planning ability of
the agents under test. An example is the use of the twoplayer version of this implementation in automatic game
parameter tuning [32].
In this simplified version, there are at each time twelve legal
actions at each game tick, being a combination of the three
steer actions (LEFT, CENTER, RIGHT), and whether the ship
is thrusting ( THRUST) and / or firing (FIRE). Unlike the
original game there is no HYPERSPACE action. The player
starts with 0 as game score and 3 lives, and an additional
life every 10, 000 points. The player gets 200, 100 or 50
points every time it hit an large, medium or small asteroid,
respectively, and loses 10 points for every missile fired. Losing
a life has a penalty of −200 points. The game stops when all
lives are lost, or 1, 000 game ticks have elapsed. Currently
there are no flying saucers, but we plan to add this feature for
the next set of experiments.
Figure 2a gives a screenshot of the game screen. Pink lines
illustrate the simulations of the RHEA agent which controls
the spaceship; these are shown for illustrative purposes only,
and ignore the fact that each rollout involves other changes to
the game state such as firing missiles and the rocks moving
and splitting.
2) Planet Wars: Planet Wars is a simple but challenging
Real Time Strategy game that was run as a highly successful
Google Game AI Challenge in 2011 by the University of
Waterloo in Canada. For the work in this paper we implemented a simpler version of the game to make it faster to allow
rapid running of experiments while retaining some challenging
aspects of the original game. The game runs at more than 10
million game ticks per second.
The aim game of Planet Wars is to occupy all the planets,
where each planet is occupied either by player 1, player 2, or
by a neutral. Each planet has a number of ships belonging to
the owner of the planet. To invade a planet a player sends a
number of ships to the planet to be invaded from the planet
it owns. For our experiments we used an implementation with
the following rules:
•
•
•
No neutral planets: the ships on each planet are either
owned by player 1 or player 2.
At each game tick, a player moves by shifting ships to
or from a player’s ship buffer, or by moving it’s current
planet of focus.
When a player transfers ships it is always between it’s
buffer and the current planet of focus.
At each game tick the score for each player is the sum
of all the ships on each planet it owns, plus the ships
stored in it’s buffer. We have two versions of the game:
the easiest for the planning agents, and the one used in
this paper also adds in the ships in a player’s buffer to
it’s score, a more deceptive version of the game does not
include this.
For research purposes, the advantage of the modified action
space is that it makes it compatible with the General Video
Game AI framework, giving direct access to a large number
of game-playing agents for comparison purposes.
•
IV. A PPLICATIONS AND R ESULTS
We aim to optimise the parameters of the RHEA agent
(described in Section III-A) with NTBEA and two baseline
algorithms, Grid Search and a multi-valued version of the
Sliding Window compact Genetic Algorithm (SWcGA), an
EDA proposed by Lucas et al. [33] recently.2 The NTBEA
used in the experiments takes into account all the d 1-tuples,
d(d−1)
2-tuples and the only d-tuple.
2
A. Experimental Setting
1) Fitness Function: In each case the fitness function is
based on the outcome of a single game, which leads to noisy
fitness functions.
a) Asteroids: The evaluation function is the game score
when the game terminates, the higher the better. The calculation of the game score is explained in Section III-B1. The
value of k should be set relative to the score distribution; for
Asteroids k was set to 5, 000.
b) Planet Wars: The evaluation function is based on
playing a single game, with a value of +1 if the agent wins
or −1 for a loss, and 0 for a draw (which is very unlikely to
occur). This gives an extremely noisy fitness function. Note
that an optimiser can choose to resample a particular point in
the search space, i.e., playing multiple games using the agent
with an identical parameter setting, in order to get an estimate
of win rate. However, with a fixed small evaluation budget,
the optimiser will run fewer iterations compared to sampling
exactly one.
The NTBEA optimises the parameters of a RHEA agent
controlled as player 1, versus a fixed-parameter version of the
RHEA with well-chosen parameters based on the authors’
experience playing as player 2. The game is symmetric and
the planning budget for both players was set to 2, 000 game
ticks. For Planet Wars, the value of k in the NTBEA was set
to 1.0.
2) Search Space: The RHEA agent is characterised by the
parameters shown in Table II, where each one is given a type
and legal values. The parameters have been detailed previously
in Section III-A. The mutation probability for a d-dimensional
problem is calculated by nbM utatedP oints/d.
2 The original Compact Genetic Algorithm (cGA) and the sliding window
version described by Lucas et al. [33] only handled binary strings. The version
used in this paper handles integer strings and can use absolute or relative
fitness measures; the version used here used absolute fitness measures.
TABLE II: Search space of the parameter settings.
Variable
Asteroids
sequenceLength
Planet Wars
nbM utatedP oints
f lipAtLeastOneBit
useShif tBuf f er
nbResamples
Type
Integer
Integer
Integer
boolean
boolean
Integer
Legal values
5, 10, 15, 20, 50, 100, 150
5, 10, 15, 20, 50
0.0, 1.0, 2.0, 3.0
f alse, true
f alse, true
1, 2, 3
The RHEA agent is very flexible and allows any compatible evolutionary algorithm to be plugged in to control the
evolutionary process. For these experiments we used a Random Mutation Hill Climber / (1 + 1)-Evolutionary Algorithm
(the distinction between the two depends on the parameters
chosen by the hyper-parameter optimiser).
The search space of the parameter settings is actually the
search space of RHEA instances, while every RHEA given a
parameter setting is considered as a distinct instance. As shown
in Table II, the size of the search space, in other words, total
number of possible RHEA instances, when playing Asteroids
and Planet Wars, is 336 and 240, respectively.
3) Budget: We allowed an optimisation budget of the same
value as the size of search space, thus 336 game evaluations
in the case of Asteroids and 240 in the case of Planet Wars.
Note that a “game evaluation” refers to a whole game playing,
which lasts at most 1, 000 game ticks. At each game tick, the
playing agent RHEA has a budget of 2, 000 forward model
calls (simulations) to find an optimal action to play. Given that
there are 336 and 240 points in the search space of the two
games, this means an attempted uniform Grid Search cannot
even sample each point twice. The NTBEA works by building
up statistics for each value in each individual dimension, as
well as tuples of values, and the full-width n-tuple. Hence the
optimiser is able to make good use of the gathered information.
B. Results
Each of the three optimisation algorithms, Grid Search,
SWcGA and NTBEA, are given the same budget to optimise a
RHEA agent on each of the games. The recommended agent
instance at the end of an optimisation is validated by playing
100 games. NTBEA was run with a neighborhood-size of 50,
and SWcGA with a sliding window of size 50.
Table III analyses the average fitness value obtained by
the RHEA instances recommended by three tested algorithms over 10 optimisation trials on Asteroids and Planet
War. RHEAGridSearch , RHEASW cGA and RHEAN T BEA
denote the RHEA agent instance recommended at the end of
optimisation by Grid Search, SWcGA and NTBEA, respectively. A random agent, which uniformly randomly selects
an action at every game tick, is also tested for comparison.
Unsurprisingly, the random agent performs poorly both in
games.
Given the same budget, the NTBEA significantly outperforms the baseline algorithms, Grid Search and SWcGA.
Of the three, Grid Search is by far the worst. The
Wilcoxon Signed-Rank Test was run for significance between
TABLE III: Average fitness value over 10 optimisation trials.
Standard error is given after ±. Note that in the case of Planet
War, an average of 0 refers to a win rate of 50%.
Agent
Random
RHEAGridSearch
RHEASW cGA
RHEAN T BEA
Planet Wars
-0.9400 ± 0.3400
-0.53 ± 0.12
0.39 ± 0.07
0.54 ± 0.02
Asteroids
1,091 ± 88
7,716 ± 330
8,439 ± 110
8,756 ± 41
RHEASW cGA and RHEAN T BEA , establishing a significance difference between the measures with p = 0.01928
for Planet Wars. The results for Asteroids follow a normal
distribution under the Shapiro-Wilk Normality Test, and the
difference between RHEASW cGA and RHEAN T BEA is
significant with a student t-test p-value of 0.040131.
In addition to the results presented in Table III the authors
have also made many other test runs with various evaluation
budgets, both higher and lower than the budgets used for
the table, and in all the tests made so far the NTBEA has
outperformed the SWcGA, and also outperformed grid search
where applicable (grid search was not applied to cases where
the budget was smaller than the number of points in the search
space). A more thorough set of experiments, including ones
with much larger search spaces is on-going work.
Figure 2 is an illustration of the plot of game score in
real-time by the program to show how score varies over the
planning horizon of the algorithm.
C. Solutions Found
The best solutions found showed common traits for both
games, but with some significant differences. The best solutions always used the shift buffer, and set nResamples to
1. Flipping a bit was normally preferred, and a large value
of nbM utatedP oints of 2.0 or 3.0 gave best performance.
The best sequence length for Planet Wars was either 10 or 20,
and for Asteroids a sequence length of 100 worked best; the
big difference in performance for various parameter settings
demonstrates the importance and impact of undertaking effective parameter tuning. The output of the NTBEA provides
clear statistics on the parameter combinations that were tested,
but these run in to pages of output. We are currently developing
graphical tools to convey this information in a more convenient
form.
V. C ONCLUSION AND F UTURE W ORK
In this paper, we propose an N-Tuple Bandit Evolutionary Algorithm (NTBEA), combining the strength of N-Tuple
systems, multi-armed bandit algorithms and evolutionary algorithms. The proposed approach is compared to two baseline
algorithm, Grid Search and Sliding Window compact Genetic Algorithm, on a hyper-parameter optimisation problem:
evolving a rolling horizon evolutionary game-playing agent.
NTBEA significantly outperformed the baseline algorithms,
and provides an explicit control over exploitation versus exploration.
(a) Screenshot of the game screen of Asteroids. Polygons represent
asteroids. Solid rectangle represents the spaceship, and pink lines
illustrate the simulations of the rolling horizon evolutionary agent
which controls the spaceship. Red solid circulars represent bullets
fired by the spaceship.
(b) Real-time plots showing how the score varies over the planning
horizon of the algorithm. The figure shows 10 plots, each of length
100 with a different color. Each of the plots refers to a simulation
in Figure 2a.
Fig. 2: Screenshot of the game screen of Asteroids and real-time plot of game scores in the simulations.
The next step is the application of the NTBEA to more
expensive problems, such as optimising some game-playing
agents for playing the 2-player games in the General Video
Game AI framework [34]. To ensure the stable performance,
the 2-player agent should be evaluated by simulating multiples games against multiple opponent models, which will be
computationally expensive. Additional, it will be challenging
to test the NTBEA on larger search space, not only in terms
of the dimension number but also the number of legal actions
on each of the dimensions.
One thing worth taking into account but not considered
in this paper explicitly is the prior knowledge on the dependencies between dimensions. As mentioned previously in
Section II, if all the tuples were considered, the number of
bandits increases exponentially. Though the use of tuples takes
care of the potential dependencies and it is not necessary to
consider all the tuples, we still need to decide in advance which
are the ones to be included in the model (e.g., all the possible
1-tuples, 2-tuples and the single d-tuple were included in our
test case). Some prior knowledge on the problem can facilitate
this selection step and reduces the amount of memory required
by NTBEA.
The comparison between the NTBEA and other EDAs will
be interesting, and more generally a comparison between
NTBEA and a wide range of existing hyper-parameter optimisation algorithms such as SMAC. It should be noted though
that many of the latter algorithms are not set up to handle
extremely noisy fitness functions such as the games used in
this paper.
In summary, the NTBEA is a significant new algorithm
that has been shown to work well both for automated game
tuning, and now for optimising a game playing agent on two
very different games. The algorithm is well suited to problems
where the fitness function is noisy and expensive, gives good
performance, and provides useful statistics on the contribution
of each combination of parameters as a consequence of the
underlying N-Tuple model.
R EFERENCES
[1] C. F. Sironi, J. Liu, D. Perez-Liebana, R. D. Gaina, I. Bravi, S. M.
Lucas, and M. H. Winands, “Self-adaptive mcts for general video game
playing,” in European Conference on the Applications of Evolutionary
Computation. Springer, 2018.
[2] K. Kunanusont, R. D. Gaina, J. Liu, D. Perez-Liebana, and S. M. Lucas,
“The n-tuple bandit evolutionary algorithm for automatic game improvement,” in Evolutionary Computation (CEC), 2017 IEEE Congress on.
IEEE, 2017.
[3] J. Lehman and K. O. Stanley, “Abandoning objectives: Evolution through
the search for novelty alone,” Evolutionary Computation, no. 2, pp. 198
– 223, 2011.
[4] H. Mühlenbein and G. Paass, “From recombination of genes to the estimation of distributions i. binary parameters,” in International conference
on parallel problem solving from nature. Springer, 1996, pp. 178–187.
[5] P. Larrañaga and J. A. Lozano, Estimation of distribution algorithms: A
new tool for evolutionary computation. Springer Science & Business
Media, 2001, vol. 2.
[6] C. González, J. A. Lozano, and P. Larranaga, “Mathematical modeling
of discrete estimation of distribution algorithms,” in Estimation of
Distribution Algorithms. Springer, 2002, pp. 147–163.
[7] J. A. Lozano, Towards a new evolutionary computation: advances on
estimation of distribution algorithms. Springer Science & Business
Media, 2006, vol. 192.
[8] M. Hauschild and M. Pelikan, “An introduction and survey of estimation of distribution algorithms,” Swarm and Evolutionary Computation,
vol. 1, no. 3, pp. 111–128, 2011.
[9] J. Decock and O. Teytaud, “Noisy optimization complexity under locality assumption,” in Proceedings of the twelfth workshop on Foundations
of genetic algorithms XII. ACM, 2013, pp. 183–190.
[10] P. Rolet and O. Teytaud, “Bandit-based estimation of distribution algorithms for noisy optimization: Rigorous runtime analysis.” in LION.
Springer, 2010, pp. 97–110.
[11] Y. Gai, B. Krishnamachari, and R. Jain, “Learning multiuser channel
allocations in cognitive radio networks: A combinatorial multi-armed
bandit formulation,” in New Frontiers in Dynamic Spectrum, 2010 IEEE
Symposium on. IEEE, 2010, pp. 1–9.
[12] W. Chen, Y. Wang, and Y. Yuan, “Combinatorial multi-armed bandit:
General framework and applications,” in International Conference on
Machine Learning, 2013, pp. 151–159.
[13] S. Ontanón, “Combinatorial multi-armed bandits for real-time strategy
games,” Journal of Artificial Intelligence Research, vol. 58, pp. 665–702,
2017.
[14] W. W. Bledsoe and I. Browning, “Pattern recognition and reading by
machine,” in Proceedings of the Eastern Joint Computer Conference,
1959, pp. 232 – 255.
[15] S. M. Lucas, “Learning to Play Othello with N-Tuple Systems,” Australian Journal of Intelligent Information Processing, vol. 4, pp. 1–20,
2008.
[16] T. P. Runarsson and S. M. Lucas, “Preference learning for move
prediction and evaluation function approximation in othello,” IEEE
Transactions on Computational Intelligence and AI in Games, vol. 6,
no. 3, pp. 300–313, 2014.
[17] J. Liu, D. Pérez-Liébana, and S. M. Lucas, “Bandit-based random
mutation hill-climbing,” in Evolutionary Computation (CEC), 2017
IEEE Congress on. IEEE, 2017, pp. 2145–2151.
[18] D. Perez-Liebana, S. Samothrakis, J. Togelius, S. M. Lucas, and
T. Schaul, “General Video Game AI: Competition, Challenges and
Opportunities,” in Thirtieth AAAI Conference on Artificial Intelligence,
2016, pp. 4335–4337.
[19] D. Perez-Liebana, J. Liu, and S. M. Lucas, “General video game AI as
a tool for game design,” Tutorial at IEEE Conference on Computational
Intelligence and Games (CIG), 2017.
[20] M. Bellemare, S. Srinivasan, G. Ostrovski, T. Schaul, D. Saxton,
and R. Munos, “Unifying count-based exploration and intrinsic
motivation,” in Advances in Neural Information Processing
Systems 29, D. D. Lee, M. Sugiyama, U. V. Luxburg,
I. Guyon, and R. Garnett, Eds.
Curran Associates, Inc.,
2016, pp. 1471–1479. [Online]. Available: http://papers.nips.cc/paper/
6383-unifying-count-based-exploration-and-intrinsic-motivation.pdf
[21] J. Bergstra and Y. Bengio, “Random search for hyper-parameter optimization,” Journal of Machine Learning Research, vol. 13, no. Feb, pp.
281–305, 2012.
[22] L. Li, K. Jamieson, G. DeSalvo, A. Rostamizadeh, and A. Talwalkar,
“Hyperband: A novel bandit-based approach to hyperparameter optimization,” arXiv preprint arXiv:1603.06560, 2016.
[23] F. Hutter, H. H. Hoos, and K. Leyton-Brown, “Sequential model-based
optimization for general algorithm configuration,” in Proceedings of
LION 5, 2011, pp. 507 – 523.
[24] J. S. Bergstra, R. Bardenet, Y. Bengio, and B. Kégl, “Algorithms
for hyper-parameter optimization,” in Advances in Neural Information
Processing Systems, 2011, pp. 2546–2554.
[25] M. Jaderberg, V. Dalibard, S. Osindero, W. M. Czarnecki, J. Donahue,
A. Razavi, O. Vinyals, T. Green, I. Dunning, K. Simonyan,
C. Fernando, and K. Kavukcuoglu, “Population based training of neural
networks,” CoRR, vol. abs/1711.09846, 2017. [Online]. Available:
http://arxiv.org/abs/1711.09846
[26] R. Rohwer and M. Morciniec, “A theoretical and experimental account
of n-tuple classifier performance,” Neural Computation, vol. 8, no. 3,
pp. 629–642, 1996.
[27] P. Auer, N. Cesa-Bianchi, and P. Fischer, “Finite-time analysis of the
multiarmed bandit problem,” Machine learning, vol. 47, no. 2-3, pp.
235–256, 2002.
[28] J. Liu, D. Pérez-Liébana, and S. M. Lucas, “Rolling horizon coevolutionary planning for two-player video games,” in Computer Science and
Electronic Engineering (CEEC), 2016 8th. IEEE, 2016, pp. 174–179.
[29] R. D. Gaina, J. Liu, S. M. Lucas, and D. Pérez-Liébana, “Analysis
of vanilla rolling horizon evolution parameters in general video game
playing,” in European Conference on the Applications of Evolutionary
Computation. Springer, 2017, pp. 418–434.
[30] R. D. Gaina, S. M. Lucas, and D. Pérez-Liébana, “Population seeding
techniques for rolling horizon evolution in general video game playing,”
in Evolutionary Computation (CEC), 2017 IEEE Congress on. IEEE,
2017, pp. 1956–1963.
[31] R. D. Gaina, S. M. Lucas, and D. Perez-Liebana, “Rolling horizon evolution enhancements in general video game playing,” in Computational
Intelligence and Games (CIG), 2017 IEEE Conference on. IEEE, 2017,
pp. 88–95.
[32] J. Liu, J. Togelius, D. Pérez-Liébana, and S. M. Lucas, “Evolving
game skill-depth using general video game ai agents,” in Evolutionary
Computation (CEC), 2017 IEEE Congress on. IEEE, 2017, pp. 2299–
2307.
[33] S. M. Lucas, J. Liu, and D. Pérez-Liébana, “Efficient noisy optimisation
with the multi-sample and sliding window compact genetic algorithms,”
in Computational Intelligence (SSCI), 2017 IEEE Symposium Series on.
IEEE, 2017.
[34] R. D. Gaina, A. Couetoux, D. J. N. J. Soemers, M. H. M. Winands,
T. Vodopivec, F. Kirchgeβner, J. Liu, S. M. Lucas, and D. Perez-Liebana,
“The 2016 two-player GVGAI competition,” IEEE Transactions on
Computational Intelligence and AI in Games, 2017.
| 2cs.AI
|
Quantum Information Processing manuscript No.
(will be inserted by the editor)
A quantum genetic algorithm with quantum
crossover and mutation operations
arXiv:1202.2026v5 [cs.NE] 22 Nov 2013
Akira SaiToh · Robabeh Rahimi · Mikio
Nakahara
Received: date / Accepted: date
Abstract In the context of evolutionary quantum computing in the literal
meaning, a quantum crossover operation has not been introduced so far.
Here, we introduce a novel quantum genetic algorithm which has a quantum
crossover procedure performing crossovers among all chromosomes in parallel
for each generation. A complexity analysis shows that a quadratic speedup is
achieved over its classical counterpart in the dominant factor of the run time
to handle each generation.
Keywords Genetic algorithm · Quantum computing · Computational
complexity
PACS 03.67.Ac · 87.23.-n · 89.70.Eg
A. SaiToh
Quantum Information Science Theory Group, National Institute of Informatics, 2-1-2 Hitotsubashi, Chiyoda, Tokyo 101-8430, Japan
R. Rahimi
Institute for Quantum Computing, University of Waterloo, 200 University Avenue West,
Waterloo, Ontario N2L 3G1, Canada
M. Nakahara
Department of Physics, Kinki University, 3-4-1 Kowakae, Higashi-Osaka, Osaka 577-8502,
Japan
A. SaiToh · M. Nakahara
Research Center for Quantum Computing, Interdisciplinary Graduate School of Science and
Engineering, Kinki University, 3-4-1 Kowakae, Higashi-Osaka, Osaka 577-8502, Japan
Present Address:
A. SaiToh
Department of Computer Science and Engineering, Toyohashi University of Technology, 1-1
Hibarigaoka, Tenpaku-cho, Toyohashi, Aichi 441-8580, Japan
E-mail: [email protected]
2
Akira SaiToh et al.
1 Introduction
Continuous development has been performed on genetic algorithms [19,11,32].
Along with the development of quantum computing [15,37], quantum-inspired
classical algorithms for evolutionary computing have been developed [36,16,
17,18,35,34,6,24,33] (see also a review [54] and references therein). In addition, classical genetic algorithms to evolve quantum circuits have also been
studied by several authors [52,44,2,45,53,38,25,22,43,23,29,26,30,7,46] (see
also review articles [10,9]). These algorithms are, however, designed to work
on classical computers. Quantum genetic algorithms (QGAs), in its literal
meaning, nonetheless, have gathered comparably little attention and a few
works [40,49,27,50,28] have been performed so far. Evolutionary computing
on quantum architectures will achieve more attention if there is a scenario
to establish significant improvement over classical counterparts. Indeed, Malossini et al. [27,28] claimed that the computational complexity of an evolutionary step between generations is exponentially fast in their quantum algorithm
in comparison to a classical one. It seems, however, that they overlooked the
complexity of index-to-string conversion circuits or, otherwise, overlooked the
circuit complexity of a variant of the inversion-about-average operation. Their
claim is thus misleading as we will discuss in Sect. 2. Recently, Johannsen et
al. [20] applied quantum search algorithms to several optimization problems
in a certain context of evolutionary computing. Nevertheless, it is obscure
how much cost is spent for the internal quantum circuit of a variant of the
inversion-about-average operation used for amplitude amplification (namely,
the operation denoted as ASφ0 A−1 in the convention of Ref. [4]) in their approach.
Let us briefly summarize the conventional approach of QGAs and its problem. Here, we omit discussions on incomplete or physically unfeasible works
on QGAs, which were summarized by Sofge in Ref. [41]. The aim of a genetic
algorithm is, in most cases, to find an individual (typically, an input string)
with a very high fitness value for a given problem. It starts with O(|A|w ) initial
individuals where |A| is the size of the alphabet A and w is the length of a
schema expected to be a building block for a given problem. Let us restrict the
problem by representing each individual as a chromosome encoded as a binary
string with the length n. We need ∼ N = 2n initial chromosomes for the worst
case. With a quantum register, one may use n qubits to make a superposition
of N chromosomes. We regard the probability to find a chromosome on the
computational basis as its (normalized) population.
First we briefly overview the selection strategies in conventional QGAs. In
short, a selection is an operation to enhance populations of individuals with
high fitness values and to decrease those with low fitness values. In quantum
computing, it is natural to utilize variants of Grover’s algorithm [12] for this
purpose. In fact, a variant of the Grover search for a priori unknown number r
of solutions (the Grover-BBHT search) [3] and that for finding the maximum
[8,1] are the essential parts
p of the QGAs developed in Refs. [49,27,50,28].
The query complexity O( N/r) for a variant of the Grover search dominates
A quantum genetic algorithm with quantum crossover and mutation operations
3
the total complexity in the QGA of Udrescu et al. [49,50]. For the QGA
of Malossini et al. [27,28], there is a different factor to consume time. As
mentioned, this will be discussed in detail in Sect. 2.
Second we overview the strategies for crossovers and mutations in conventional QGAs. In a short explanation of these terms, a crossover is an operation
to exchange substrings of two chromosomes and a mutation is an operation
to flip certain bits of a chromosome. These are effective operations to enlarge the search space. So far, quantum crossover operations have not been
developed. Malossini et al. [27,28] used classical crossover and mutation operations; Udrescu et al. [49,50] did not use crossover and mutation operations.
Johannsen et al. [20] introduced quantum mutation operations in an application of quantum amplitude amplification [4] to some optimization problems;
nevertheless, crossovers were not used. It is, in fact, in general difficult to manipulate populations of quantum states for handling crossovers if we imitate
the classical way in a straightforward manner (See also Ref. [10] which explained the difficulty in a slightly different manner). Picking-up two particular
individuals and making a crossover costs O(n) quantum gates if we use O(n)
ancillary qubits. Since there are many possible pairs for a crossover at each
generation, there is no speedup over a classical crossover. More specifically
speaking, this approach needs to look up classical data of chromosomes to
specify a pair of chromosomes. Thus there cannot be any speedup. In addition, this requires an exponentially large classical memory in comparison to
the size of a quantum register. We will show a different approach to handle
crossovers in our algorithm in Sect. 3.
As another direction for developing quantum evolutionary computing, which
we do not pursue in this contribution, one may use so-called quantum fixedpoint search algorithms [14,48]. They increase the population of a superposition of target chromosomes by iterative applications of unitary amplitudeamplification operations. Unlike the original Grover search, there is no drawback in an excessive iteration; the probability of finding a target grows monotonically as 1 − ǫ2t+1 where ǫ is the probability of finding a non-target chromosome and t is the number of queries. The query complexity of this approach
is, however, as large as that of the exhaustive search: When the number r of
targets is very small in comparison to N , we have ǫ2t+1 = (1 − r/N )2t+1 ≃
1 − (2t + 1)r/N , which implies that O(N/r) queries are required to achieve
a sufficiently small error probability, say 1/2. Thus for r ≪ N , Grover’s algorithm should be chosen instead of fixed-point quantum search algorithms.
In addition, the asymptotic optimality of ǫ2t+1 as an error reduction speed
was proved [5] for fixed-point quantum search methods. It is also known that
the query complexity of the Grover search is optimal [3] in general as a unitary process for unsorted search. Therefore, it is unlikely1 that a fixed-point
1 There is another drawback in the use of a quantum fixed-point search. It requires
es = 1 − (1 − eiπ/3 ) P |ςl ihςl | and U
et = 1 − (1 − eiπ/3 )|τm ihτm |
phase shift operations U
l
with source states |ςl i and target states |τm i (l and m are labels) in addition to another
appropriate unitary operation [14]. One may alternatively use Ûs = 1−(1−eiπ/3 )|SihS| and
Ût = 1 − (1 − eiπ/3 )|T ihT | where |Si is an equally-weighted superposition of |ςl i and |T i is
4
Akira SaiToh et al.
quantum search is effectively used in an evolutionary computing instead of the
standard Grover search and its variants.
In this contribution, we propose a quantum genetic algorithm that involves
a quantum crossover and a quantum mutation. It uses the quantum search for
finding the maximum [8,1] for the selection procedure. Although this selection
strategy looks similar to those of conventional QGAs [49,27,50,28], it is different in the point that we use tailored inversion operations for the quantum
search. Our algorithm achieves a quadratic speedup over its classical counterpart in the dominant factor of the run time to update each generation as an
algorithm involving all possible crossovers for a chosen crossing site.
This paper is organized in the following way. Section 2 discusses on the
algorithm of Malossini et al. [27,28]. Section 3 describes our algorithm. The
procedures of the crossover, the mutation and the selection are developed in the
standard quantum circuit model. Computational complexities of the algorithm
are evaluated in Sect. 4. The obtained results are discussed in Sect. 5 and
summarized in Sect. 6.
2 A conventional quantum genetic algorithm
Here we discuss the conventional quantum genetic algorithm proposed by Malossini et al. [27,28]. There is another conventional algorithm proposed by
Udrescu et al. [49,50], which is not introduced here. The main difference is
that Udrescu et al. did not employ crossovers and mutations while Malossini
et al. employed them as classical operations.
The problem we consider is given as follows.
Problem: Suppose there are N individuals with indices 0, . . . , N − 1. There is
a fitness function f : {0, . . . , N − 1} → [0, 1]. Find one of the individuals with
sufficiently large fitness values.
The algorithm of Malossini et al. is described in Fig. 1 (Algorithm 1).
They analyzed their algorithm and claimed [27,28] that the selection procedure
e log N
e)
internally requires a very small number of oracle calls in contrast to O(N
e is the number
oracle calls required by a classical selection procedure, where N
of individuals in a generation. In short, their claim was that the number of the
Grover iterations in step (b) (of Fig. 1) was O(1).
Their complexity analysis was, however, misleading. A more accurate description of the complexity is given in the following way. There are two possible
e is approximately equal to N ; (B) N
e ≨ N.
cases we should consider: (A) N
Let us consider the case (A). We can preset the index x to represent the xth
e ≃ N . Then the given fitness
individual in a good approximation thanks to N
function can be used as it is, without index conversion. This indicates that we
an equally-weighted superposition of |τm i. (This kind of alternative operation for a quantum
search was used in Ref. [47].) The problem is that ςl ’s are highly-random nonconsecutive
chromosomes in the context of evolutionary computing. There is no known way to construct
es or Ûs within poly(log N
e ) cost in such a case, where N
e is the number of ςl ’s.
U
A quantum genetic algorithm with quantum crossover and mutation operations
5
Algorithm 1 (Malossini et al. [27, 28])
Main Procedure:
e ≤ N individuals.
Start from a generation consisting of N
Repeat:
1. Use the Quantum Selection Procedure to get one index. Do the same procedure
to get another.
2. Classically make a crossover for the two individuals corresponding to the indices
obtained in 1. Classically make a proper mutation.
3. Classically replace randomly chosen two individuals with the two offsprings
obtained in 2.
Quantum Selection Procedure:
e possible ones in the present generation
(i) Choose randomly an index y among N
and compute its fitness f (y). Set the threshold Fy ←− f (y).
(ii) Perform τ times:
P
(a) Initialize the quantum registers to √1
x |xi|yi, where x’s are the indices
e
N
of the present generation.
(b) Apply the Grover-BBHT search [3], where we internally use an oracle
that inverts the signs of the marked states, namely states |xi such that
f (x) ≥ Fy . This enhances the amplitudes of marked states after its Grover
iteration.
(c) Measure the left register in the computational basis and get the new index
y ′ . If f (y ′ ) > Fy , then y ← y ′ and Fy ← f (y ′ ).
(iii) Return the index y.
Fig. 1 Description of Malossini et al.’s algorithm.
e ≃ N,
use nothing but the standard Grover-BBHT search for N indices. As N
the quantum selection procedure is called only once or a very small number of
times. Suppose there are r nearly-optimal individuals that are acceptable as
solutions for a given problem. Inside the quantum selection procedure, as the
selection goes on in (ii), the oracle used in step (b) has less marked states. The
number of marked states converges to r. Therefore,
the query complexity of a
q
p
e
single step in a very later selection stage is O( N /r) = O( N/r). This is the
accurate description of the query complexity for the present case and it is the
same as that Udrescu et al. gave for their algorithm [49,50]. On average, each
time of the repetition in (ii) virtually extinguishes a half of the individuals by
increasing
the threshold. Thus, by considering the sum of the geometric series
p
i r) (i = 0, 1, 2, . . .), the average total query complexity is found to be
N/(2
p
O( N/r).
Let us now consider the case (B). In this case, the quantum selection procedure is called several (or more) times. The important fact is that the chromosomes of individuals in a generation are not consecutive binary strings. The
e binary strings out of N possible ones. There are two
generation contains N
options to handle them: (B-1) consecutive integers are assigned as pointers to
nonconsecutive chromosomes; (B-2) nonconsecutive chromosomes are used as
they are.
(B-1): In this case, the index x is a pointer to the chromosome κx for which
the fitness function works appropriately. (In short, this is a workaround for
the fact that the original Grover-BBHT search should start from a uniform
superposition of consecutive indices while the generation consists of noncon-
6
Akira SaiToh et al.
secutive chromosomes.) Therefore, for the superposition |ψ0 i = √1
e
N
P
x |xi|yi,
the fitness function cannot be applied directly. We need a conversion circuit
to interpret x as κx . Then the fitness function is applied and it returns f (κx ).
Each conversion should involve O(log N ) CCNOT gates. Thus, the total circuit
e log N ). As a result, a single oracle call
depth for the index conversions is O(N
e log N ). Indeed, the query
accompanies the additional time complexity O(N
e is much smaller than N . However, the cost of index
complexity is small if N
conversions hinders speedup over a classical selection procedure.
P
(B-2): In this case, the state |xi in the superposition |ψ0 i = √1
x |xi|yi
e
N
keeps the actual chromosome x as an index. One should use the Grover-search
routine somehow without index conversions by choosing |ψ0 i as its initial state.
Indeed, this is possible [4] in the case where one can provide the operation
L = I − 2|ψ0 ihψ0 | as the inversion-about-average operation instead of the
standard one for the original Grover search. Nevertheless, L should be gen†
erated by sandwiching I − 2|0ih0| with U(a) and U(a)
in the present context,
where the unitary operation U(a) corresponds to the initialization step (a).
e log N ) (see, e.g., Ref. [51]). (There is
Thus the circuit complexity of L is O(N
a little confusing result by Soklakov and Schack [42]; they showed that a state
preparation, namely an initialization, is possibly performed within polynomial
cost in log N for some cases. Nevertheless, considering the internal cost of the
e poly(log N )) time for general
special oracle they use, their method spends O(N
cases including the present case where the parent set consists of random indices.) As a consequence, an expensive circuit for L should be used subsequent
to every query. Obviously, the time complexity in this case is as large as the
one in case (B-1).
One may also think of replacing the Grover-search routine with the generalized Grover search for nonconsecutive integer sets [13]. Nevertheless, it is
required to find an appropriate unitary transformation replacing the standard
Hadamard transformation H ⊗ · · · ⊗ H. There is no known way to find it efficiently when a random integer set is given and target integers are unknown
in advance.
There is, in fact, a way to reduce the circuit complexity of L introduced in
(B-2) if the algorithm is modified so that it uses an efficient pseudo-randomizer
(or, pseudo-scrambler) circuit instead of directly using a random number generator. Our algorithm introduced in the next section takes this approach.
In addition to the above discussions, we should mention that Algorithm 1
e log N ) space to keep a generation in a
has another problem: it takes O(N
classical memory. This is usually quite larger than the O(log N ) space that is
enough for quantum search (neglecting the space internally used by an oracle
circuit). Our algorithm is designed not to face this problem.
A quantum genetic algorithm with quantum crossover and mutation operations
7
3 Algorithm with quantum crossover and quantum mutation
As we have seen in the previous sections, conventional quantum genetic algorithms were not designed to achieve quantum speedup in their selection
procedures in case a generation consists of nonconsecutive chromosomes. In
addition, a quantum crossover operation has not been developed so far. Here,
we propose an algorithm using a quantum crossover and a quantum mutation.
We use a variant of the Grover search with tailored subroutines for the selection procedure, whose query complexity is the main factor of the total time
complexity. We achieve a quadratic speedup over a classical counterpart in
a dominant factor of time complexity as a genetic algorithm with crossovers
among all parents. It is novel in the sense that a simultaneous crossover using a superposition is achieved. The speedup partly relies on an efficient internal structure of a pseudo-randomizer circuit, which will be explained in
Appendix A.
3.1 Algorithm flow
We introduce our algorithm in Fig. 2 (Algorithm 2). In this algorithm, the
quantum register is accessible from its subroutines as a kind of global variables.
The procedures called inside the algorithm are described in corresponding
subsections 3.2-3.5.
3.2 Preparing the initial state of a quantum register
In this subsection, we define the procedure init_reg(R, z).
This procedure is intended to prepare a superposition corresponding to the
generation given as a set X = {R(q)}q ∪{z} with q = 00 · · · 0c−2 1c−1 , . . . , 10 · · · 1c−1
⊗n
e = 2c ). The procedure starts with the quantum state |0i⊗c
(thus ♯X = N
.
a |0i
P
1
√
The desired superposition is |ϕi =
x∈X |ax ia |xi with ax the address
e
N
pointing x. We opt to use consecutive addresses 0, . . . , 2c − 1. The procedure
is now defined in Fig. 3. The order of time complexity of this procedure equals
to the internal circuit complexity poly(cn) of the pseudo-randomizer R. An
explicit example to construct R as a quantum circuit is given in Appendix A.
3.3 Crossover
Here, we construct a quantum crossover procedure quantum_crossover(l). It
is a 1-point crossover acting on the chromosomes simultaneously.
Recall that the original generation is given as a set X of n-bit integers x
e . Each x has its left side xleft and its right side xright separated
with ♯X = N
by the crossing site. The crossing site is placed between the (l − 1)th and the
lth qubits as specified by the parameter. Hence, the bit length of xleft is l.
8
Akira SaiToh et al.
Algorithm 2
Consider a threshold fth for fitness values, which is considered to be sufficiently large.
t ←− 0.
REPEAT 1.-9.:
1. Construct a pseudo-randomizer R that maps a c-bit string to a pseudo-random
n-bit string. It should be implemented as a circuit whose input and output are
integer couples (a, 0) and (a, R(a)), respectively, where a is a c-bit integer. The
circuit should consist of poly(cn) elementary reversible logic gates. We require R
to regard 00 · · · 0c−1 as an exception and map it to 00 · · · 0n−1 . We also require
e = 2c ≨ N = 2n . Once constructed, R is fixed until next t. An explicit example
N
to construct R is given in Appendix A.
2. IF t = 0 THEN generate a random c-bit string γ and set z ←− R(γ) (otherwise,
z is the best chromosome found in the (t − 1)th trial) ENDIF
3. Generate a random c-bit string γ ′ and set u ←− R(γ ′ ).
4. CALL init_reg(R, z) defined in Sect. 3.2 twice to make two identical quantum
P
states that are both |ϕi = √1
x∈X |ax ia |xi where ax is the address pointing
e
N
5.
6.
7.
8.
9.
to x ∈ X; X = {R(q)}q ∪{z} with q = 00 · · · 0c−2 1c−1 , . . . , 10 · · · 1c−1 ; subscript
“a” stands for the address portion. We write the entire unitary operation of this
procedure as Uinit . That is, |ϕi⊗2 = Uinit (|0ia |0i)⊗2 .
CALL quantum_crossover(l) defined in Sect. 3.3 for the current quantum register, which is a 1-point crossover with the crossing site, chosen at one’s convenience, placed between the (l − 1)th and the lth bits of a chromosome. All
possible crossovers for this crossing site are performed simultaneously. This procedure is an identity map as a quantum operation acting on |ϕi⊗2 .
Apply the quantum mutation (Sect. 3.4) to the current quantum register. We
write the entire unitary operation of this procedure as Umut . Apply the same
mutation classically to u.
CALL quantum_selection(Uinit , Umut , u) defined in Sect. 3.5 for the current
quantum register and obtain the output chromosome z.
IF f (z) ≥ fth THEN RETURN z and EXIT ENDIF
t ←− t + 1. Refresh the quantum register.
Fig. 2 Description of our algorithm.
PROCEDURE init_reg(R, z):
P2c −1
⊗c ⊗ I ⊗n to
(i) Make a superposition √1♯X
j=0 |jia |0 · · · 0i by applying H
⊗n .
|0i⊗c
a |0i
(ii) Apply the pseudo-randomizer R implemented as a unitary operation mapping
|jia |0i to |jia |xj i with xj an n-bit pseudo-random number for j = 1, . . . , 2c − 1.
By assumption, |0ia |0i is mapped to |0ia |0i.
(iii) Apply a 00 · · · 0c−1 -controlled X z0 ⊗· · ·⊗X zn−1 to map |0ia |0i to |0ia |zi, where
zk is the kth bit of z (k = 0, . . . , n − 1).
(iv) RETURN the current state, namely |ϕi.
Fig. 3 Description of procedure init reg(R, z).
With the procedure, we generate a superposition of all the children that are
combinations of xleft ’s and xright ’s together with their parents with the same
weight as children.
The state of the quantum register in the beginning of this procedure is
⊗2
|ϕi
=
!
1 X
left
right
√
|ax ia |x i|x
i ⊗
♯X x∈X
!
1 X
′ left
′ right
√
i .
|ax′ ia |x i|x
♯X x′ ∈X
A quantum genetic algorithm with quantum crossover and mutation operations
9
This state has the components |xleft i|xright i|x′ left i|x′ right i besides the addresses.
left
We relabel the qubits so that the middle portion |xright i|x′ i is put aside from
our minds. Let us conceal them by denoting as |∗xx′ i. In addition, we denote
right
the main portion |xleft x′
i with the subscript “main”. The state |ϕi⊗2 with
the new qubit labels is written as
1 X X
right
imain |∗xx′ i.
|ax ax′ ia |xleft x′
♯X
′
x∈X x ∈X
We have at most (♯X)2 distinct chromosomes in this state. In this way, all
possible crossovers are performed at once by the relabelling. The resultant
state is a superposition of all of the children together with their parents. (The
parents are involved because the values of x and x′ may coincide.) This is
desirable as a crossover because sometimes some parents have higher fitness
values than any child.
The procedure described above is formally written as shown in Fig. 4.
PROCEDURE quantum_crossover(l):
(i) We have the quantum register in the state |ϕi⊗2 . The original labels of its
qubits are 0, . . . , 2c + 2n − 1. We relabel them as 0, . . . , c − 1, 2c, . . . , 2c + l − 1,
{z
} |
{z
}
|
c
2c + n, . . . , 2c + 2n − l − 1,
|
{z
}
n−l
2c + l, . . . , 2c + n − 1.
|
{z
}
c, . . . , 2c − 1,
|
{z
}
c
l
2c + 2n − l, . . . , 2c + 2n − 1,
{z
}
|
l
n−l
(ii) RETURN
Fig. 4 Description of procedure quantum crossover(l).
As is obvious, this procedure is an identity map as a quantum operation.
Once the crossover process is completed, one may use a mutation as an option.
This is going to be explained in the next subsection.
3.4 Mutation
In classical genetic algorithms, randomly selected chromosomes are affected
by a mutation, which is typically certain bit flips acting on randomly-chosen
places. Here, we consider the mutation procedure described in Fig. 5. Although
it is written as a classical routine, it can be trivially interpreted as a quantum
circuit. As a quantum circuit, this mutation procedure tmp_mut is realized
by a multiple-bit controlled multiple-bit NOT gate. (In the example mentioned in Fig. 5, the gate is “0-controlled 1-controlled 1-controlled
0-controlled 1-controlled NOT NOT” with control bits specified by the
first template and the target bits specified by the second template.) The gate
10
Akira SaiToh et al.
PROCEDURE tmp_mut:
(i) Let us randomly generate the first template like ***0*1***1**0****1*, which
specifies a schema to mutate. Using this template, we pick up chromosomes with
specified bits like 0,1,1,0,1 in the specified places.
(ii) We also use the second template like *X*************X*** in which X’s can be
placed only on the places where *’s (namely, “don’t care” symbols) are placed
in the first template. Using this template, we apply the bit flip X to the specified
places of the chromosomes picked up in (i).
(iii) RETURN
Note: Technically, we often wish to avoid a mutation for the best chromosome z found
so far by the present time step. This is realized by choosing the first template
so that this does not happen.
Fig. 5 Description of procedure tmp mut.
right
acts on the portion |xleft x′
imain ; the addresses ax and ax′ are untouched.
Therefore, the resultant state can be written as
|ϕi
e =
right
1 X X
xleft xe′
imain |∗xx′ i,
|ax ax′ ia |e
♯X
′
(1)
x∈X x ∈X
right
where (e
xleft xe′
)’s are the chromosomes after the mutation process.
The next step is to apply a natural selection to the chromosomes living in
the superposition |ϕi.
e
3.5 Selection
In this subsection, we introduce our selection procedure. It is intended to find
right
a chromosome having the maximum fitness among (e
xleft xe′
)’s. It utilizes
the quantum search for finding the maximum [8,1]. As we have mentioned,
conventional QGAs [49,27,50,28] have similar selection strategies. The difference from them is that we use tailored inversion operations for the quantum
search.
Our selection procedure is called with three arguments: Uinit , Umut and u.
We have the state (1) at the beginning of this procedure. The procedure is now
defined in Fig. 6. Here in the text, we do not repeat its description. It should be
e + (28/5)(log2 N
e )2 ⌉
noted that, in the procedure, we set kterm = η × ⌈(45/2)N
with integer constant η ≥ 1.
Let us give an explanation about kterm , namely, the number of iterations.
The defined procedure is the same as the well-known quantum search algorithm for finding the maximum [8,1] except for the definitions of the inversion operation for targets and the inversion-about-average operation. In other
words, we perform the quantum search for finding the maximum in the subright
xleft xe′
imain |∗xx′ i}. As proved by Dürr and Høyer [8],
space span{|ax ax′ ia |e
the probability for the output
to
be the maximum is at least 1/2 if the number
√
of iterations is ⌈(45/2) M + (7/5)(log2 M )2 ⌉ with M the number of indices.
e 2 since there are N
e 2 distinct addresses.
In the present context, we have M = N
A quantum genetic algorithm with quantum crossover and mutation operations
11
PROCEDURE quantum_selection(Uinit , Umut , u):
FOR k ←− 0 TO kterm − 1:
(i) Set
X
U1 = I ⊗ I ⊗ I − 2
I ⊗ |yihy| ⊗ I
f (y)≥f (u)
where the left and the right I’s act on the address states |ax ax′ i and the states
|∗xx′ i, respectively. This is the oracle function that can be implemented as
follows. First, we attach ancillary qubits as blocks (I) and (II) in the state
√
right
xleft xe′
) is set
|0 · · · 0i(I) |−i(II) with |−i = (|0i − |1i)/ 2. For each xx′ , f (e
as a value of the block (I) by using f implemented as a quantum circuit. Let
us write this operation as Uf . Then we compare the block (I) with f (u); we
right
flip the qubit (II) if f (e
xleft xe′
) ≥ f (u). We apply Uf † (this disentangles the
ancillary qubits from the main register) and remove the ancillary qubits.
(ii) Set
U2 = I ⊗ I ⊗ I − 2|ϕih
e ϕ|.
e
This operation is composed in the following way.
†
†
,
Umut
U2 = Umut Uinit [I − 2(|0i⊗c |0i⊗n h0|⊗c h0|⊗n )⊗2 ]Uinit
where we also relabel qubits according to quantum_crossover(l).
(iii) Apply the Grover-BBHT search [3] for which we use U1 and U2 instead of
the standard operations, namely, the inversion operation for targets and the
inversion-about-average operation, respectively. As for the starting state of the
search, we use |ϕi.
e
(iv) Measure the main register and obtain the chromosome u′ .
IF f (u′ ) > f (u) THEN u ←− u′ ENDIF
END FOR
RETURN u.
Fig. 6 Description of procedure quantum selection(Uinit , Umut , u).
Therefore, after kterm iterations, we find the output chromosome having the
right
maximum fitness among (e
xleft xe′
)’s, with the probability at least 1−(1/2)η .
By setting η sufficiently large, say, around 16 to 24, we have the desired output
with almost certainty.
4 Computational cost
We are going to evaluate the computational cost of each process of Algorithm
2 (Fig. 2) in Sect. 4.1. The total computational cost to handle a generation
will be derived in Sect. 4.2 and compared with that of classical counterpart in
Sect. 4.3.
4.1 Costs of each procedure
For a single call of each procedure, the costs are evaluated as follows.
Cost of the initialization The procedure to prepare the initial state, described
e log N ) elementary quantum gates.
in Sect. 3.2, uses poly(cn) = poly(log N
12
Akira SaiToh et al.
Cost of the crossover The crossover described in Sect. 3.3 makes use of two
identical (c + n)-qubit states. This procedure does not use any quantum operation but relabels qubits. This takes O(log N ) time.
Cost of the mutation The mutation described in Sect. 3.4 involves a single
gate that looks like, say,
∗ − NOT − ∗ − C0 − ∗ − C1 − ∗ − ∗ − ∗ − C1 − ∗ − ∗
−C0 − ∗ − ∗ − NOT − ∗ − C1 − ∗
with multiple control bits (0-control C0 ’s and 1-control C1 ’s) and multiple
NOT gates placed according to the corresponding templates, where symbol
∗ stands for an untouched qubit. Such a gate can be realized by O(log N )
elementary quantum gates with O(log N ) ancillary qubits.
e)
Cost of the selection The selection described in Sect. 3.5 consumes O(N
2
e chromosomes.
queries to find the best chromosome in the set of at most N
Each query accompanies the operation U2 that invokes Uinit and Umut , and
also their inverse operations. It is easy to find that the internal cost of U2 is
poly(cn) according to the costs for Uinit and Umut . In addition, the cost to
prepare the starting state is poly(cn). We may also mention that, the internal
cost of the fitness function is a certain small factor, typically poly(n), as conventionally assumed [11]. Therefore, the circuit complexity of the procedure
e poly(log N
e log N )). (We assume that the circuit depth is on the order
is O(N
of the circuit size, namely, the number of elementary quantum gates.) As for
space, we use poly(log N ) qubits in total, considering a typical fitness function
consuming poly(n) space. When the fitness function is designed to use O(n)
space, O(log N ) qubits are enough, although we do not assume this case for
evaluating the space complexity.
4.2 Total cost
Comparing the costs of the four procedures, the dominant cost is the circuit
complexity for the selection procedure. Therefore, we find that our algorithm
uses
e poly(log N
e log N ))
O(N
(2)
elementary quantum gates for each t. This is the time complexity of our algorithm for handling each generation. It is quadratically faster than classically
expected amount considering the fact that the number of combinations is
e 2 ) in our crossover procedure (see the next subsection for the details). As
O(N
for the space complexity, we spend poly(log N ) qubits and O(log N ) classical
bits as is easily evaluated from the description of Algorithm 2 (Fig. 2).
A quantum genetic algorithm with quantum crossover and mutation operations
13
4.3 Comparison with a classical counterpart
The classical counterpart of our algorithm is the one described in Fig. 7 (Algorithm 3). As is obvious from its structure, it should have the same output
and the same number of iterations as our algorithm with almost certainty as
long as the same random seed is used for step 1 to construct R for each value
of t. The computational costs of individual procedures in Algorithm 3 are as
Algorithm 3
Consider a threshold fth for fitness values, which is considered to be sufficiently
large.
t ←− 0.
REPEAT 1.-7.:
1. Construct the pseudo-randomizer R found in Algorithm 2 (Fig. 2) as a classical
function. Fix R until next t. Using R, we generate a generation X = {R(q)}
e = 2c ≨ N = 2n .
with q = 00 · · · 0c−2 1c−1 , . . . , 10 · · · 1c−1 . We require N
e − 1 n-bit-length chromosomes in X presently.)
(Thus, we have N
2. IF t 6= 0 THEN put z into X (z is the best chromosome found in the (t − 1)th
trial) ELSE generate a random c-bit string γ and put R(γ) into X ENDIF
e chromosomes in X.)
(Now we have N
3. Split all x ∈ X between the (l − 1)th bit and the lth bit. This makes l-bit strings
xleft ’s and (n − l)-bit strings xright ’s. Generate the set X̂ consisting of all x’s
and all of their children that are all the combinations of xleft ’s and xright ’s.
4. Apply the template-based mutation tmp_mut, introduced in Sect. 3.4, as a classical procedure to all chromosomes in X̂.
5. Find the chromosome z having the best fitness value among those in X̂.
6. IF f (z) ≥ fth THEN RETURN z and EXIT ENDIF
7. t ←− t + 1.
Fig. 7 Description of a classical counterpart of our algorithm.
follows.
e poly(cn)) basic operations.
– Generating X in the steps 1.-2. takes O(N
2
2
e
e
– We need to use O(nN ) space and O(nN ) basic operations to perform all
e parents.
the crossovers among N
e 2 ) basic operations.
– Mutations acting on the individuals of X̂ take O(nN
e 2 poly(n)) basic
– The selection to find the best individual from X̂ takes O(N
operations considering the cost poly(n) of calculating a fitness value.
e , the time and space complexities are O(N
e 2 poly(log N ))
As n = log2 N and c = log2 N
2
e log N ), respectively, for each t, i.e., for handling each generation.
and O(N
In contrast, as we have seen in Sect. 4.2, the time and space complexities of
e poly(log N
e log N ))
our quantum genetic algorithm (described in Fig. 2) are O(N
and poly(log N ), respectively, for each t. Therefore, neglecting the difference
e log N ) and poly(log N ), we have achieved a quadratic
between poly(log N
speedup over its classical counterpart together with an exponential reduction
in space.
The classical counterpart has been constructed by keeping the one-byone correspondence with the quantum algorithm. Thus there is a possibility
14
Akira SaiToh et al.
that a better classical algorithm with the same behavior exists. This is in
fact the case for Algorithm 3. For a fairer comparison, now we reform Algorithm 3 and reduce its space complexity. The algorithm described in Fig. 8
(Algorithm 4) has the same output and the same number of repetitions as
Algorithm 3 while its space complexity is exponentially reduced. We use the
same pseudo-randomizer construction and the same mutation procedure as
e = 2c and N = 2n for integers c and n satisfying
before. Of course, we set N
e 2 ) times for each value of t. We
1 ≤ c < n. In this algorithm, R is called O(N
Algorithm 4
Consider a threshold fth for fitness values, which is considered to be sufficiently
large.
t ←− 0.
REPEAT 1.-8.:
1. Construct the pseudo-randomizer R : {0, 1}c → {0, 1}n as a classical operation.
2. Construct the mutation process as a map M as a classical operation.
e − 1}, z ←− R(γ) (otherwise, z is
3. IF t = 0 THEN for a random γ ∈ {1, . . . , N
the best chromosome found in the (t − 1)th step) ENDIF
e − 1}, j ←− R(γ ′ ).
4. For a random γ ′ ∈ {1, . . . , N
e
5. FOR a ←− 0 TO N − 1:
IF a = 0 THEN x ←− z ELSE x ←− R(a) ENDIF
e − 1:
FOR b ←− 0 TO N
IF b = 0 THEN y ←− z ELSE y ←− R(b) ENDIF
Crossover x and y and obtain children v and w.
Find the best chromosome g among the chromosomes M (x),
M (y), M (v) and M (w).
IF f (g) > f (j) THEN j ←− g ENDIF
END FOR
END FOR
6. z ←− j.
7. IF f (z) ≥ fth THEN RETURN z and EXIT ENDIF
8. t ←− t + 1.
Fig. 8 Description of an improved classical counterpart of our algorithm.
know that R internally takes poly(cn) time. Therefore, this algorithm spends
e 2 poly(log N
e log N )) time for handling each generation. As for the space
O(N
complexity, it spends only poly(log N ) space, which is clear from the algorithm
structure.
In comparison to this enhanced classical algorithm, our quantum algorithm
still has a quadratically small time complexity as shown in Eq. (2).
5 Discussion
How to perform crossovers in a quantum manner was a pending problem in
conventional quantum genetic algorithms [49,27,50,28]. In fact, a selective
crossover for specific two chromosomes is expensive when they are component
states of a superposition. Even if we attach address states pointing to the
component states, we need to look up the classical data of the chromosomes
to construct a quantum circuit realizing the unitary operation for this purpose,
A quantum genetic algorithm with quantum crossover and mutation operations
15
or more specifically, for placing address-controlled bit-flip gates appropriately.
In our algorithm (Algorithm 2 shown in Fig. 2), we have avoided to mimic a
classical way and chosen a different approach. We use two identical copies of a
superposition corresponding to a generation and utilize relabelling of qubits so
as to handle all possible combinations of substrings simultaneously. Obviously,
the classical counterpart of our algorithm is the one that seeks for the best
chromosome (after a mutation) among all possible crossovers for a chosen
crossing site for each generation. Comparing our algorithm with the classical
counterpart, we concluded that we have achieved a considerable reduction in
the computational cost.
One may, however, claim that usually at most several crossovers are performed for a single generation in a classical genetic algorithm. Indeed, our
algorithm is not aimed to be a quantum counterpart of a common classical
genetic algorithm. As we discussed in Sect. 2, a straightforward conversion
of a common classical algorithm into quantum one by simply incorporating
a quantum search into the selection procedure has a problem: we need to either interpret nonconsecutive integers to consecutive ones or use an expensive
construction for the inversion-about-average operation in order to perform
the Grover-BBHT search, which causes a significant loss of performance. This
problem should be resolved so as to find a meaningful quantum counterpart for
the common case. Seemingly, the following workaround looks fine: (i) Use the
pseudo-randomizer R used in Algorithm 2 instead of a random number generator to generate initial chromosomes of a generation. (ii) Apply a small number
of crossovers. (iii) Use a selection procedure similar to that of Algorithm 2.
Nevertheless, as we have discussed, it is not known how to construct the procedure (ii) as a unitary operation without the expensive process of looking
up classical data of chromosomes. Therefore, it is the fact that a meaningful quantum counterpart is not easily found for a common classical genetic
algorithm.
In view of the search space covered by each generation, a simultaneous
crossover is, of course, desirable. Use of a superposition for this purpose was
discussed [49,50] but not developed previously. In this sense, we have made a
meaningful improvement by introducing Algorithm 2.
Besides the crossover, let us discuss on the selection procedure. Our algorithm uses a variant of the Grover-BBHT search to achieve a quadratic
speedup over its classical counterpart. The internal cost for each query is kept
polynomial in the length of a chromosome because of the polynomial cost of
our pseudo-randomizer, as described in Sect. 4. Apart from the complexity,
there is some room to find a different design for the selection. Our algorithm
is designed to carry over only the best chromosome z to the next generation,
among the chromosomes existing after the quantum crossover and mutation
procedures. Since projective measurements are used in the quantum selection,
it is inevitable to demolish other chromosomes. This can be a drawback because some of them may possess good fitness values albeit not the best. To
mitigate this severe selection, one may keep the values of z obtained in several elder generations as classical data. These values can be put into a later
16
Akira SaiToh et al.
generation by modifying the register-initialization procedure slightly: one can
modify the pseudo-randomizer so that it does not touch several input strings;
then one can map them to the kept values of z. In this way, one may maintain
a better diversity for high-fitness chromosomes. This is one possible extension
of our algorithm.
There have not been many studies on quantum genetic algorithms so far.
It is hoped that several or more different designs of genetic procedures will be
developed for quantum computers.
6 Summary
We have proposed a genetic algorithm whose crossover, mutation and selection procedures have been all constructed as quantum routines so that quantum parallelism is effectively used. Its crossover procedure performs crossovers
among all chromosomes of a generation. The run time of our algorithm to
update each generation is quadratically faster than that of its classical counterpart, apart from negligible factors.
Acknowledgements A.S. is thankful to Shigeru Yamashita for his comment. A.S. and
M.N. were supported by the “Open Research Center” Project for Private Universities:
matching fund subsidy from MEXT. R.R. is supported by Industry Canada and CIFAR.
A An example of constructing the pseudo-randomizer R
In this appendix, we show an example to construct the pseudo-randomizer R used in Algorithms 2, 3 and 4. It should map a c-bit string to a pseudo-random n-bit string except for
00 · · · 0c−1 that is mapped to 00 · · · 0n−1 . Its internal circuit complexity should be poly(cn).
As we use a quantum circuit to realize it in Algorithm 2, it is desirable to employ a circuit
structure that is originally unitary.
Consider inputs a ∈ {0, 1}c . We design a circuit that maps a0 · · · ac−1 00 · · · 0n−1 to
a0 · · · ac−1 κ0 · · · κn−1 with κ = R(a), an n-bit pseudo-random number (here, a0 · · · ac−1
and κ0 · · · κn−1 are the binary representations of a and κ). By the definition of R, the
circuit preserves 00 · · · 0c−1 00 · · · 0n−1 . This circuit is generated by function gen_r_circ()
described in Fig. 9. As is clear from the description, the circuit output from this function
FUNCTION gen_r_circ():
Comment: We use wires v0 , . . . , vc−1 , w0 , . . . , wn−1 .
FOR i ←− 0 TO c − 1:
(1) Use a random number generator to generate an n-bit integer γ. Write its binary
representation as γ0 · · · γn−1 .
(2) Using the wire vi as the control wire (namely, the control bit), output the gate
controlled-(X0 )γ0 ⊗· · ·⊗(Xn−1 )γn−1 with Xk the bit flip gate acting on the wire
wk (k = 0, . . . , n − 1). In this gate, the bit flips are active under the condition
that vi = 1.
END FOR
Fig. 9 Description of function gen r circ().
can be directly used as a quantum circuit. Using the circuit C = gen_r_circ(), we have
A quantum genetic algorithm with quantum crossover and mutation operations
17
C
|ai|0i 7→ |ai|R(a)i ∀a ∈ {0, 1}c . The circuit complexity of C is O(cn) because, for each i,
at most n CNOT gates are used to decompose the gate output from step (2). In addition,
gen_r_circ() spends O(c poly(n)) time when a common random number generator [21, 31]
is used in step (1).
Note that the function gen_r_circ() is called only once for each t, in the beginning of
step 1 in Algorithms 2, 3 and 4. We have only to reuse the circuit C for the use of the
pseudo-randomizer until t is incremented.
It is expected that outputs from the circuit C possess good uniformity if we use a good
random number generator in step (1) of gen_r_circ() for generating C. Let us write γ as
γ(i) to emphasize
its dependence on i. For a nonzero input a0 · · · ac−1 , the kth bit of the
Pc−1
output R(a) is
i=0 ai · γk (i) mod 2. This indicates that, for two different inputs a and
a′ , the kth bits of R(a) and R(a′ ) differ with probability 1/2 in the ideal case where γ(i)’s
are generated from a true random number generator. This is because a and a′ differ by at
least a single bit. It also indicates that two different bits, the kth and the k ′ th bits, of R(a)
for a nonzero input a differ with the probability 1/2 in the ideal case. This is because γk (i)
and γk′ (i) differ with the probability 1/2.
Now we show the result of our numerical test of C. We tried statistical tests of randomness [21, 39] to test pseudo-random numbers output from C, using NIST’s Statistical
Test Suite (STS) (version 2.1.1) [39]. We set c = 10 and n = 32. Mersenne Twister (MT)
(version mt19937ar) [31] was used to generate γ in step (1) of gen_r_circ(). We used the
seed value 121212 and did not reset MT during the circuit generation. The circuit C output
from gen_r_circ(), of course, consisted of 10 outputs from step (2). For this C, we used the
inputs a ∈ {0, 1}c \{00 · · · 0c−1 } from smaller to larger and obtained corresponding outputs
R(a) by numerical computation. We obtained 1023 × 32 bits in total in the outputs, since
210 − 1 = 1023. We regarded them as a serial bit string from left to right and used STS in
its default setting to test the string. In the execution of STS, we used 25 binary sequences
with length 1200 as samples from the string. The following tests were tried with the default
parameter values in STS: the Frequency Test, the Block Frequency Test, the Cumulative
Sums Test, the Runs Test, the Longest-Run-of-Ones Test, the Binary Matrix Rank Test, the
Spectral DFT Test and the Serial Test. The string passed the tests except for the Binary
Matrix Rank Test. It should be noted that the input length was too small for the binary
matrix rank test [39]. In addition, randomness is not very strictly required for the use in
evolutionary computing. Therefore, considering the tests that the string passed, we may
claim that gen_r_circ() generates a usable pseudo-randomizer circuit for our algorithm.
We conducted another test: We generated ten circuits by calling gen_r_circ() ten times
without resetting MT, using the seed value 676767. For each circuit, we performed the
same process as above to obtain the serial bit string. We obtained ten serial bit strings in
total and tested the concatenated string using STS. As samples input to STS, we used 25
binary sequences with length 12000. The concatenated string passed the tests except for the
Binary Matrix Rank Test and the Spectral DFT Test. It was unexpected that it did not
pass the spectral DFT test. It requires a further investigation to reveal the reason of this
phenomenon.
The results of the first and the second tests are summarized in Table 1. In summary for this appendix, we found that a pseudo-randomizer circuit whose outputs possess
enough randomness for the use in evolutionary computing can be generated by the function gen_r_circ(). It is hoped that the function will be improved so as to achieve better
randomness for the sake of general use.
References
1. Ahuja, A., Kapoor, S.: A quantum algorithm for finding the maximum (1999).
arXiv:quant-ph/9911082
2. Barnum, H., Bernstein, H.J., Spector, L.: A quantum circuit for OR (1999). arXiv:quantph/9907056
3. Boyer, M., Brassard, G., Høyer, P., Tapp, A.: Tight bounds on quantum searching.
Fortschr. Phys. 46, 493–505 (1998)
18
Akira SaiToh et al.
Table 1 List of test results for the first and the second tests we performed (see the text
for the details of the tests). Each value is a P -value (see Ref. [39] for its meaning for each
test). The asterisk ∗ indicates a failure to pass the test. We should mention that the values
for the Longest-Run-of-Ones Test accidentally coincided, while the internally-used results
for sample sequences were different.
Frequency Test
Block Frequency Test
Cumulative Sums Test (Forward)
Cumulative Sums Test (Reverse)
Runs Test
Longest-Run-of-Ones Test
Binary Matrix Rank Test
Spectral DFT Test
Serial Test
1st Test
0.021262
0.105618
0.105618
0.001691
0.141256
0.875539
0.000000 ∗
0.000533
0.041438
2nd Test
0.186566
0.001156
0.029796
0.057146
0.010606
0.875539
0.000000 ∗
0.000000 ∗
0.186566
4. Brassard, G., Høyer, P., Tapp, A.: Quantum counting. In: K.G. Larsen, S. Skyum,
G. Winskel (eds.) Proceedings of Automata, Languages and Programming, 25th International Colloquium (ICALP’98) (LNCS 1443), pp. 820–831. Aalborg, Denmark, 13-17
July 1998, Springer-Verlag, Berlin (1998). arXiv:quant-ph/9805082
5. Chakraborty, S., Radhakrishnan, J., Raghunathan, N.: Bounds for error reduction with
few quantum queries. In: C. Chekuri, K. Jansen, J. Rolim, L. Trevisan (eds.) Proceedings
of the 9th International Workshop on Randomization and Computation (RANDOM
2005) (LNCS 3624), pp. 245–256. Berkeley, CA, 22-24 August 2005, Springer-Verlag,
Berlin (2005)
6. Chen, M., Quan, H.: Quantum-inspired evolutionary algorithm based on estimation
of distribution. In: Proceedings of the 2nd International Conference on Bio-Inspired
Computing: Theories and Applications (BIC-TA 2007), pp. 17–19. Zhengzhou, China,
14-17 September 2007, IEEE Press, Piscataway, NJ (2007)
7. Ding, S., Jin, Z., Yang, Q.: Evolving quantum oracles with hybrid quantum-inspired
evolutionary algorithm (2006). arXiv:quant-ph/0610105
8. Dürr, C., Høyer, P.: A quantum algorithm for finding the minimum (1996). arXiv:quantph/9607014
9. Gepp, A., Stocks, P.: A review of procedures to evolve quantum algorithms (2007).
arXiv:0708.3278
10. Giraldi, G.A., Portugal, R., Thess, R.N.: Genetic algorithms and quantum computation
(2004). arXiv:cs/0403003
11. Goldberg, D.E.: Genetic Algorithms in Search, Optimization, and Machine Learning.
Addison-Wesley, Reading, MA (1989)
12. Grover, L.K.: A fast quantum mechanical algorithm for database search. In: Proceedings
of the 28th Annual ACM Symposium on Theory of Computing (STOC 1996), pp. 212–
219. Philadelphia, PA, 22-24 May 1996, ACM Press, New York, NY (1996)
13. Grover, L.K.: Quantum search on structured problems (1998). arXiv:quant-ph/9802035
14. Grover, L.K.: Fixed-point quantum search. Phys. Rev. Lett. 95, 150501–1–4 (2005)
15. Gruska, J.: Quantum Computing. McGraw-Hill, London (1999)
16. Han, K.H., Kim, J.H.: Genetic quantum algorithm and its application to combinatorial
optimization problem. In: Proceedings of the 2000 Congress on Evolutionary Computation (CEC2000), pp. 1354–1360. La Jolla, CA, 16-19 July 2000, IEEE Press, Piscataway,
NJ (2000)
17. Han, K.H., Kim, J.H.: Quantum-inspired evolutionary algorithm for a class of combinatorial optimization. IEEE Trans. Evol. Comput. 6(6), 580–593 (2002)
18. Han, K.H., Kim, J.H.: Quantum-inspired evolutionary algorithms with a new termination criterion, hǫ gate, and two-phase scheme. IEEE Trans. Evol. Comput. 8(2),
156–169 (2004)
A quantum genetic algorithm with quantum crossover and mutation operations
19
19. Holland, J.H.: Adaptation in Natural and Artificial Systems: An Introductory Analysis
with Applications to Biology, Control and Artificial Intelligence. The University of
Michigan Press, Ann Arbor, MI (1975)
20. Johannsen, D., Kuru, P.P., Lengler, J.: Can quantum search accelerate evolutionary
algorithms? In: M. Pelikan, J. Branke (eds.) Proceedings of the 12th Annual Genetic
and Evolutionary Computation Conference (GECCO-2010), pp. 1433–1440. Portland,
OR, 7-11 July 2010, ACM, New York, NY (2010)
21. Knuth, D.E.: The Art of Computer Programming, Vol. 2: Seminumerical Algorithms,
3rd Ed. Addison-Wesley, Reading, MA (1997). Chap. 3
22. Leier, A., Banzhaf, W.: Evolving Hogg’s quantum algorithm using linear-tree GP. In:
E. Cantú-Paz, J.A. Foster, K. Deb, L.D. Davis, R. Roy, U.M. O’Reilly, H.G. Beyer,
R. Standish, G. Kendall, S. Wilson, M. Harman, J. Wegener, D. Dasgupta, M.A. Potter,
A.C. Schultz, K.A. Dowsland, N. Jonoska, J. Miller (eds.) Proceedings of the Genetic
and Evolutionary Computation Conference 2003 (GECCO-2003), Part I (LNCS 2723),
pp. 390–400. Chicago, IL, 12-16 July 2003, Springer-Verlag, Berlin (2003)
23. Leier, A., Banzhaf, W.: Comparison of selection strategies for evolutionary quantum
circuit design. In: K. Deb (ed.) Proceedings of the Genetic and Evolutionary Computation Conference 2004 (GECCO-2004), Part II (LNCS 3103), pp. 557–568. Seattle, WA,
26-30 June 2004, Springer-Verlag, Berlin (2004)
24. Liao, R., Wang, X., Qin, Z.: A novel quantum-inspired genetic algorithm with expanded
solution space. In: Proceedings of the 2010 Second International Conference on Intelligent Human-Machine Systems and Cybernetics (IHMSC 2010), pp. 192–195. Nanjing,
China, 26-28 August 2010, IEEE Computer Society, Los Alamitos, CA (2010)
25. Lukac, M., Perkowski, M.: Evolving quantum circuits using genetic algorithm. In:
A. Stoica, D. Keymeulen, J. Lohn (eds.) Proceedings of the 2002 NASA/DoD Conference on Evolvable Hardware, pp. 177–181. Alexandria, VA, 15-18 July 2002, IEEE
Computer Society, Los Alamitos, CA (2002)
26. Lukac, M., Perkowski, M., Goi, H., Pivtoraiko, M., Yu, C.H., Chung, K., Jee, H., Kim,
B.G., Kim, Y.D.: Evolutionary approach to quantum and reversible circuits synthesis.
In: S.N. Yanushkevich (ed.) Artificial Intelligence in Logic Design, pp. 201–257. Kluwer
Academic Publisher, Dordrecht (2004)
27. Malossini, A., Blanzieri, E., Calarco, T.: QGA: quantum genetic algorithm (2004). Technical Report: #DIT-04-105, Dec. 2004, Univ. Trento, http://www.dit.unitn.it
28. Malossini, A., Blanzieri, E., Calarco, T.: Quantum genetic optimization. IEEE Trans.
Evol. Comput. 12(2), 231–241 (2008)
29. Massey, P., Clark, J.A., Stepney, S.: Evolving quantum circuits and programs through
genetic programming. In: K. Deb (ed.) Proceedings of the Genetic and Evolutionary Computation Conference 2004 (GECCO-2004), Part II (LNCS 3103), pp. 569–580.
Seattle, WA, 26-30 June 2004, Springer-Verlag, Berlin (2004)
30. Massey, P., Clark, J.A., Stepney, S.: Human-competitive evolution of quantum computing artefacts by genetic programming. Evol. Comput. 14(1), 21–40 (2006)
31. Matsumoto, M., Nishimura, T.: Mersenne Twister: a 623-dimensionally equidistributed
uniform pseudorandom number generator. ACM Trans. Model. Comput. Sim. 8, 3–30
(1998). Http://www.math.sci.hiroshima-u.ac.jp/˜m-mat/MT/mt.html
32. Mitchell, M.: An Introduction to Genetic Algorithms. MIT Press, Cambridge, MA
(1996)
33. Mohammed, A.M., Elhefnawy, N.A., El-Sherbiny, M.M., Hadhoud, M.M.: Quantum
crossover based quantum genetic algorithm for solving non-linear programming. In:
Proceedings of the 8th International Conference on INFOrmatics and Systems (INFOS2012), pp. BIO–145–153. Cairo, Egypt, 14-16 May 2012, IEEE, Piscataway, NJ
(2012)
34. Nakayama, S., Imabeppu, T., Ono, S.: Pair swap strategy in quantum-inspired evolutionary algorithm (2006). In the Late-breaking papers of the 2006 Genetic and Evolutionary
Computation Conference (GECCO-2006), Seattle, WA, 8-12 July 2006
35. Nakayama, S., Imabeppu, T., Ono, S., Iimura, I.: Consideration on pair swap strategy in
quantum-inspired evolutionary algorithm. IEICE Trans. Inf. Sys. J89-D(9), 2134–2139
(2006). In Japanese
20
Akira SaiToh et al.
36. Narayanan, A., Moore, M.: Quantum-inspired genetic algorithms. In: Proceedings of
the IEEE 3rd International Conference on Evolutionary Computation (ICEC96), pp.
61–66. Nagoya, Japan, 20-22 May 1996, IEEE Press, Piscataway, NJ (1996)
37. Nielsen, M.A., Chuang, I.L.: Quantum Computation and Quantum Information. Cambridge University Press, Cambridge (2000)
38. Rubinstein, B.I.P.: Evolving quantum circuits using genetic programming. In: Proceedings of the 2001 Congress on Evolutionary Computation (CEC2001), pp. 144–151. Seoul,
Korea, 27-30 May 2001, IEEE Press, Piscataway, NJ (2001)
39. Rukhin, A., Soto, J., Nechvatal, J., Smid, M., Barker, E., Leigh, S., Levenson, M., Vangel, M., Banks, D., Heckert, A., Dray, J., Vo, S.: A statistical test suite for random and pseudorandom number generators for cryptographic applications (2010).
NIST Special Publication 800-22, Revision 1a,
http://csrc.nist.gov/groups/ST/toolkit/rng/index.html
40. Rylander, B., Soule, T., Foster, J., Alves-Foss, J.: Quantum evolutionary programming.
In: L. Spector, E.D. Goodman, A. Wu, W.B. Langdon, H.M. Voigt, M. Gen, S. Sen,
M. Dorigo, S. Pezeshk, M.H. Garzon, E. Burke (eds.) Proceedings of the Genetic and
Evolutionary Computation Conference (GECCO-2001), pp. 1005–1011. San Francisco,
CA, 7-11 July 2001, Morgan Kaufmann, San Francisco (2001)
41. Sofge, D.A.: Prospective algorithms for quantum evolutionary computation. In: P.D.
Bruza, W. Lawless, K. van Rijsbergen, D.A. Sofge, B. Coecke, S. Clark (eds.) Proceedings of the 2nd Quantum Interaction Symposium (QI-2008), pp. 98–105. Oxford, UK,
26-28 March 2008, College Publications, London (2008). arXiv:0804.1133
42. Soklakov, A.N., Schack, R.: Efficient state preparation for a register of quantum bits.
Phys. Rev. A 73, 012307–1–13 (2006)
43. Spector, L.: Automatic Quantum Computer Programming: A Genetic Programming
Approach. Springer, New York (2004, Paperback Ed. 2007)
44. Spector, L., Barnum, H., Bernstein, H.: Genetic programming for quantum computers.
In: J.R. Koza (ed.) Genetic Programming 1998: Proceedings of the Third Annual Conference (GP-98), pp. 365–374. Madison, WI, 22-25 July 1998, Morgan Kaufmann, San
Francisco (1998)
45. Spector, L., Barnum, H., Bernstein, H., Swamy, N.: Finding a better-than-classical
quantum AND/OR algorithm using genetic programming. In: Proceedings of the 1999
Congress on Evolutionary Computation (CEC1999), pp. 2239–2246. Washington, D.C.,
6-9 July 1999, IEEE Press, Piscataway, NJ (1999)
46. Spector, L., Klein, J.: Machine invention of quantum computing circuits by means of
genetic programming. AI EDAM 22, 275–283 (2008)
47. Tanaka, Y., Ichikawa, T., Tada-Umezaki, M., Ota, Y., Nakahara, M.: Quantum oracles
in terms of universal gate set. Int. J. Quant. Inf. 9, 1363–1381 (2011)
48. Tulsi, T., Grover, L.K., Patel, A.: A new algorithm for fixed point quantum search.
Quant. Inf. Comput. 6, 483–494 (2006)
49. Udrescu, M., Prodan, L., Vlăduţiu, M.: Grover’s algorithm and the evolutionary approach of quantum computation (2004). ACSA Report, ”Politehnica” University of
Timisoara, 15 Oct. 2004, http://www.acsa.upt.ro/publications/index.htm
50. Udrescu, M., Prodan, L., Vlăduţiu, M.: Implementing quantum genetic algorithms:
A solution based on Grover’s algorithm. In: Proceedings of the 3rd Conference on
Computing Frontiers, pp. 71–81. Ischia, Italy, 3-5 May 2006, ACM Press, New York
(2006)
51. Ventura, D., Martinez, T.: Initializing the amplitude distribution of a quantum state.
Found. Phys. Lett. 12, 547–559 (1999)
52. Williams, C.P., Gray, A.G.: Automated design of quantum circuits. In: C.P. Williams
(ed.) Quantum Computing and Quantum Communications: First NASA International
Conference (LNCS 1509), pp. 113–125. Palm Springs, CA, 17-20 February 1998,
Springer-Verlag, Berlin (1999)
53. Yabuki, T., Iba, H.: Genetic algorithms for quantum circuit design–evolving a simpler
teleportation circuit. In: L.D. Whitley, D.E. Goldberg, E. Cantú-Paz, L. Spector, I.C.
Parmee, H.G. Beyer (eds.) Proceedings of the 2000 Genetic and Evolutionary Computation Conference (GECCO-2000), pp. 425–430. Las Vegas, NV, 8-12 July 2000, Morgan
Kaufmann, San Francisco (2000)
A quantum genetic algorithm with quantum crossover and mutation operations
21
54. Zhang, G.: Quantum-inspired evolutionary algorithms: a survey and empirical study. J.
Heuristics 17, 303–351 (2011)
| 9cs.NE
|
Journal of Nonlinear Systems and Applications ()
Copyright c 2009 Watam Press
http://www.watam.org/JNSA/
STATISTICS ON GRAPHS, EXPONENTIAL FORMULA AND
COMBINATORIAL PHYSICS
arXiv:0910.0695v2 [cs.DM] 11 Feb 2010
Laurent Poinsot, Gérard H. E. Duchamp, Silvia Goodenough and Karol A. Penson
2
Abstract. The concern of this paper is a famous combinatorial
formula known under the name “exponential formula”. It occurs
quite naturally in many contexts (physics, mathematics, computer
Partial semigroups
Let us call partial semigroup a semigroup with a partially
defined associative law (see for instance [6] for usual semigroups and [1, 14, 18] for more details on structures with
a partially defined binary operation). More precisely, a
partial semigroup is a pair (S, ∗) where S is a set and ∗
is a (partially defined) function S × S → S such that the
two (again partially defined) functions S × S × S → S
science). Roughly speaking, it expresses that the exponential generating function of a whole structure is equal to the exponential
of those of connected substructures. Keeping this descriptive statement as a guideline, we develop a general framework to handle many
different situations in which the exponential formula can be applied.
Keywords. Combinatorial physics, Exponential generating function, Partial semigroup, Experimental mathematics.
1
†‡
∗
(x, y, z) 7→ (x ∗ y) ∗ z and (x, y, z) 7→ x ∗ (y ∗ z)
(1)
coincide (same domain and values). Using this requirement one can see that the values of the (partially defined)
functions S n → S
Introduction
(x1 , · · · , xn ) 7→ ET (x1 , · · · , xn )
Applying the exponential paradigm one can feel sometimes incomfortable wondering whether “one has the
right” to do so (as for coloured structures, for example). The following paper is aimed at giving a rather
large framework where this formula holds.
Exponential formula can be traced back to works by
Touchard and Ridell & Uhlenbeck [20, 17]. For an other
exposition, see for example [4, 7, 9, 19].
We are interested to compute various examples of EGF
for combinatorial objects having (a finite set of) nodes
(i.e. their set-theoretical support) so we use as central
concept the mapping σ which associates to every structure, its set of (labels of its) nodes.
We need to draw what could be called “square-free decomposable objects” (SFD). This version is suited to our
needs for the “exponential formula” and it is sufficiently
general to contain, as a particular case, the case of multivariate series.
(2)
obtained by evaluating the expression formed by labelling
by xi (from left to right) the ith leaf of a binary tree T
with n nodes and by ∗ its internal nodes, is independant
of T . We will denote x1 ∗ · · · ∗ xn their common value.
In this paper we restrict our attention to commutative
semigroups. By this we mean that the value x1 ∗ · · · ∗
xn does not depend on the relative order of the xi . A
nonempty partial semigroup (S, ∗) has a (two-sided and
total) unit ǫ ∈ S if, and only if, for every ω ∈ S, ω ∗ ǫ =
ω = ǫ∗ω. Using associativity of ∗, it can be easily checked
that if S has a unit, then it is unique.
Example 2.1. Let F be a set of sets (resp. which contains ∅ as an element) and which is closed under the disjoint sum ⊔, i.e., if A, B ∈ F such that A ∩ B = ∅, then
A ∪ B(= A ⊔ B) ∈ F . Then (F, ⊔) is a partial semigroup
(resp. partial semigroup with unit).
3
∗ L.
Poinsot, G. H. E. Duchamp and S. Goodenough are affiliated to Laboratoire d’Informatique Paris Nord, Université ParisNord 13, CNRS UMR 7030, 99 av. J.-B. Clément, F 93430 Villetaneuse, France (emails: {ghed,laurent.poinsot}@lipn-univ.paris13.fr,
[email protected]).
† K. A. Penson is affiliated to Laboratoire de Physique Théorique
de la Matière Condensée, Université Pierre et Marie Curie, CNRS
UMR 7600, Tour 24 - 2e ét., 4 pl. Jussieu, F 75252 Paris cedex 05,
France (email: [email protected]).
‡ Manuscript received October 05, 2009. This work was supported by the French Ministry of Science and Higher Education
under Grant ANR PhysComb.
Square-free decomposable partial semigroups
+
Let 2(N ) be the set of all finite subsets of the positive
integers N+ and (S, ⊕) be a partial semigroup with unit
+
(here denoted ǫ) equipped with a mapping σ : S → 2(N ) ,
called the (set-theoretic) support mapping. Let D be the
domain of ⊕. The triple (S, ⊕, σ) is called square-free
decomposable (SFD) if, and only if, it fulfills the two following conditions.
1
2
L. Poinsot, G. H. E. Duchamp, S. Goodenough and K.A. Penson
• Direct sum (DS):
1. σ(ω) = ∅ iff ω = ǫ;
2. D = {(ω1 , ω2 ) ∈ S 2 : σ(ω1 ) ∩ σ(ω2 ) = ∅};
3. For all ω1 , ω2 ∈ S, if (ω1 , ω2 ) ∈ D then σ(ω1 ⊕
ω2 ) = σ(ω1 ) ∪ σ(ω2 ).
• Levi’s property (LP): For every ω1 , ω2 , ω 1 , ω 2 ∈ S
such that (ω1 , ω2 ), (ω 1 , ω 2 ) ∈ D and ω1 ⊕ ω2 = ω 1 ⊕
ω 2 , there are ωij ∈ S for i = 1, 2, j = 1, 2 such that
(ωi1 , ωi2 ), (ω1j , ω2j ) ∈ D, ωi = ωi1 ⊕ωi2 and ω j = ω1j ⊕ω2j
for i = 1, 2 and j = 1, 2.
Remark 3.1. The second and third conditions of (DS)
imply that σ(ω1 ⊕ω2 ) = σ(ω1 )⊔σ(ω2 ) whenever (ω1 , ω2 ) ∈
D (which means that σ(ω1 )∩σ(ω2 ) = ∅), where ⊔ denotes
the disjoint sum.
Example 3.1. As example of this setting we have:
is equal to ǫ, then ωi 6= ωj for every i, j ∈ {1, . . . , n} such
n
G
σ(ωi ).
that i 6= j. Moreover σ(⊕ni=1 ωi ) =
i=1
(ωi )ni=1
Lemma 3.2. Let
be a finite family of elements
of S with pairwise disjoint supports. Suppose that for
i
i
i = 1, · · · , n, ωi = ⊕nk=1
ωik , where (ωik )nk=1
is a finite
n
n
i
ωik .
family of elements of S. Then ⊕i=1 ωi = ⊕i=1 ⊕nk=1
These lemmas are useful to define the sum of two or
more elements of S using respective sum decompositions.
Now, an atom in a partial semigroup with unit S is any
object ω 6= ǫ which cannot be split, formally
ω = ω1 ⊕ ω2 =⇒ ǫ ∈ {ω1 , ω2 } .
(3)
The set of all atoms is denoted by atoms(S). Whenever
the square-free decomposable semigroup S is not trivial,
i.e., reduced to {ǫ}, atoms(S) is not empty.
1. The positive square-free integers, σ(n) being the set Example 3.2. The atoms obtained from examples 3.1:
of primes which divide n, the atoms being the prime
1. The atoms of 3.1.2 are the primes.
numbers.
2. The atoms of 3.1.3 are connected graphs.
2. All the positive integeres (S = N+ ), under the usual
integer multiplication, σ(n) being the set of primes
3. The atoms of 3.1.4 are the endofunctions for which
which divide n.
the domain is a singleton.
3. Graphs, hypergraphs, (finitely) coloured, weighted
4. The atoms of 3.1.5 are the monomials.
graphs, with nodes in N+ , σ(G) being the set of nodes
The prescriptions (DS,LP) imply that decomposition
and ⊕ the juxtaposition (direct sum) when the set
of
objects into atoms always exists and is unique.
of nodes are mutually disjoint.
Proposition 3.1. Let (S, ⊕, σ) be SFD. For each ω ∈
4. The set of endofunctions f : F → F where F is a
S
there is one and only one finite set of atoms A =
finite subset of N+ .
{ω1 , · · · , ωn } such that ω = ⊕ni=1 ωi . One has A = ∅
5. The (multivariate) polynomials in N[X], X = {xi : iff ω = ǫ.
i ∈ I}, with I ⊆ N+ , being a nonempty set of (commuting or not) variables, with σ(P ) = Alph(P ) the
Exponential formula
set of indices of variables that occur in a polynomial 4
P , and ⊕ = +.
In this section we consider (S, ⊕, σ) as a square-free
6. For a given finite or denumerable field, the set of irre- decomposable partial semigroup with unit.
ducible monic polynomials is denumerable. Arrange
In the set S, objects are conceived to be “measured” by
them in a sequence (Pn )n∈N+ , then the square-free
monic (for a given order on the variables) polynomi- different parameters (data in statistical language). So, to
als is SFD, σ(P ) := {n ∈ N+ : Pn divides P } and ⊕ get a general purpose tool, we suppose that the statistics
takes its values in a (unitary) ring R of characteristic zero
being the multiplication.
that is to say which contains Q (as, to write exponential
7. Rational complex algebraic curves; σ(V ) being the generating series it is convenient to have at hand the fracset of monic irreducible bivariate polynomials van- tions 1 ). Let then c : S → R be the given statistics. For
n!
ishing on V .
F a finite set and each X ⊆ S, we define
In what follows we write ⊕ni=1 ωi instead of ω1 ⊕· · ·⊕ωn
(if n = 0, then ⊕ni=1 ωi = ǫ) and we suppose that (S, ⊕, σ)
is SFD for the two following lemmas.
Lemma 3.1. Let ω1 , . . . , ωn ∈ S such that ⊕ni=1 ωi is
defined. Then for every i, j ∈ {1, . . . , n} such that i 6= j,
it holds that σ(ωi ) ∩ σ(ωj ) = ∅. In particular, if none ωk
XF := {ω ∈ X : σ(ω) = F } .
(4)
In order to write generating series, we need
X
1. that the sums c(XF ) :=
c(ω) exist for every
ω∈XF
finite set F of N+ and every X ⊆ S;
Statistics on Graphs, Exponential Formula and Combinatorial Physics
2. that F → c(XF ) would depend only of the cardinal- where
ity of the finite set F of N+ , for each fixed X ⊆ S;
an =
X Y
fcard(p)
3
(8)
π∈Πn p∈π
3. that c(ω1 ⊕ ω2 ) = c(ω1 ).c(ω2 ).
We formalize it in
(LF) Local finiteness. — For each finite set F of N+ , the
subset SF of S is a finite set.
(Eq) Equivariance. —
with Πn being the set of all partitions of [1..n] (in particX zn
ular for n = 0, a0 = 1) and ez =
∈ R[[z]].
n!
n≥0
In what follows [1..n] denotes the interval {j ∈ N+ : 1 ≤
j ≤ n}, reduced to ∅ when n = 0. Let (S, ⊕, σ) be a
locally finite SFD and c be a multiplicative equivariant
statistics.
For every subset X of S one sets the following
card(F1 ) = card(F2 ) =⇒ c(atoms(S)F1 ) = c(atoms(S)F2 ) .
exponential
generating series
(5)
∞
(Mu) Multiplicativity. —
X
zn
.
(9)
EGF(X; z) =
c(X[1..n] )
n!
c(ω1 ⊕ ω2 ) = c(ω1 ).c(ω2 ) .
(6)
n=0
Remark 4.1. a) In fact, (LF) is a property of the set thm 4.1 (exponential formula). Let S be a locally finite
S, while (Eq) is a property of the statistics. In practice, SFD and c be a multiplicative equivariant statistics. We
we choose S which is locally finite and choose equivariant have
statistics for instance
EGF(S; z) = c(ǫ) − 1 + eEGF(atoms(S);z) ..
(10)
(number of cycles) (number of fixed points)
c(ω) = x
y
In particular if c(ǫ) = 1 (for instance if c is proper and
R is an integral domain),
for some variables x, y.
EGF(S; z) = eEGF(atoms(S);z) .
(11)
b) More generally, it is typical to take integer-valued
partial (additive) statistics c1 , · · · ci , · · · , cr (for every
Proof — Let n = 0. Then the unique element of S∅
c (ω) c (ω)
c (ω)
ω ∈ S, ci (ω) ∈ N) and set c(ω) = x11 x22 · · · xrr .
is ǫ. Therefore c(S∅ ) = c(ǫ). Now suppose that n > 0
and let ω ∈ S[1..n] . According to proposition 3.1, there
c) The set of example 3.1.2 is not locally finite, but
is a unique finite set {α1 , . . . , αk } ⊆ atoms(S) such that
other examples satisfy (LF): for instance 3.1.3 if one asks
ω = ⊕ki=1 αi . By lemma 3.1, {σ(αi ) : 1 ≤ i ≤ k} is a parthat the number of arrows and weight is finite, 3.1.1.
tition of [1..n] into k blocks. Therefore ω ∈ atoms(S)P1 ⊕
A multiplicative statistics is called proper if c(ǫ) 6= 0. · · · ⊕ atoms(S)Pk where Pi = σ(αi ) for i = 1, . . . , k. We
It is called improper if c(ǫ) = 0. In this case, for every can remark that α1 ⊕ · · · ⊕ αk is well-defined for each
(α1 , . . . , αk ) ∈ atoms(S)P1 × · · · × atoms(S)Pk since the
ω ∈ S, c(ω) = 0 as c(ω) = c(ω ⊕ ǫ) = c(ω)c(ǫ) = 0.
If R is a integral domain and if c is proper, supports are disjoint. Now, one has, thanks to the partithen c(ǫ) = 1 because c(ǫ) = c(ǫ ⊕ ǫ) = c(ǫ)2 , tions of [1..n]
therefore 1 = c(ǫ). Note that for each X ⊆ S,
G M
X
S[1..n] =
atoms(S)p
(12)
c(ǫ) if ǫ ∈ X
c(X∅ ) =
c(ω) =
. For every
p∈π
π∈Π
n
0
if ǫ 6∈ X
X Y
ω∈X∅
X
c(S[1..n] ) =
c(atoms(S)p )
(13)
finite subset X of S, we also define c(X) :=
c(ω),
ω∈X
then we have in particular c(∅) = 0 (which is not the
same as c(S∅ ) = c({ǫ}) if c is proper). The requirement
(LF) implies that for every X ⊆ S and every finite set
F of N+ , c(XF ) is defined as a sum of a finite number of terms because XF ⊆ SF , and therefore XF is finite.
π∈Πn p∈π
as, for disjoint (finite) sets F and G of N+ , it is easy to
check that c(XF ⊕ XG ) = c(XF )c(XG ) for every X ⊆ S
and because the disjoint union as only a finite number of
factors. Therefore due to equivariance of c on sets of the
form atoms(S)F , one has
X Y
Now, we are in position to state the exponential formula
c(S[1..n] ) =
c(atoms(S)[1..card(p)] ) .
(14)
as it will be used throughout the paper. Let us recall the
π∈Πn p∈π
usual exponential formula for formal power series in R[[z]]
(see [13,X
19] for more details on formal power series). Let But c(atoms(S)[1..card(p)] ) is the card(p)th coefficient of
the series EGF(atoms(S); z). Therefore due to the
f (z) =
fn z n . Then we have
usual exponential formula, EGF(S; z) = c(ǫ) − 1 +
n≥1
eEGF(atoms(S);z) . Now if c(ǫ) = 1, then we obtain
X zn
EGF(atoms(S);z)
.
(7) EGF(S; z) = e
ef =
an
n!
n≥0
4
5
L. Poinsot, G. H. E. Duchamp, S. Goodenough and K.A. Penson
Two examples
(one-parameter) groups eλΩ where Ω =
X
α(ω)ω is
ω∈HWC
an element of HWC , with all - but a finite number of
them - the complex numbers α(ω) equal to 0, and ω a
word on the alphabet {a, a† } leads to the necessity of
solving the Normal Ordering Problem, i.e., the reduction
of the powers of Ω to the form
X
Ωn =
βi,j (a† )i aj .
(18)
(15)
The examples provided here pertain to the class of labelled graphs where the “classic” exponential formula applies, namely Burnside’s Classes1 Burn a,b , defined, for
0 ≤ a < b two integers, as the class of graphs of numeric
endofunctions f such that
fa = fb
where f n denotes the nth power with respect to functional composition. Despite of its simplicity, there are
still (enumerative combinatorial) open problems for this
class and only B1,ℓ+1 gives rise to an elegant formula
[8, 19] (see also [11], for the idempotent case: ℓ = 1 and
compare to exact but non-easily tractable formulas in [4]
for the general case in the symmetric semigroup, and in
[12] for their generalization to the wreath product of the
symmetric semigroup and a finite group).
In the sequel, Normal (Ωn ) denotes such a sum. This
problem can be performed with three indices in general
and two in the case of homogeneous operators that is
operators for which the “excess” e = i − j is constant
along the monomials (a† )i aj of the support (for which
βi,j 6= 0). Thus, for
X
βi,j (a† )i aj
(19)
Ω=
i−j=e
one has, for all n ∈ N,
The second example: the class of finite parti∞
tions which can be (and should here) identified as
X
n
† ne
Normal
(Ω
)
=
(a
)
SΩ (n, k)(a† )k ak
(20)
graphs of equivalence relations on finite subsets F ⊆
+
k=0
N .
Call this class “Stirling class” as the number of such graphs with support [1..n] and k conwhen e ≥ 0, and
nected components is exactly the Stirling number of
!
∞
the second kind S2 (n, k) and, using the statistics
X
n
†
k
k
(21)
Normal (Ω ) =
SΩ (n, k)(a ) a an|e|
x(number of points) y (number of connected components) , one obk=0
tains
X
x
xn
(16) otherwise. It turns out that, when there is only one anS2 (n, k) y k = ey(e −1) .
n!
n,k≥0
nihilation, one gets a formula of the type (x, y are formal
Examples of this kind bring us to the conclusion that commutative variables)
bivariate statistics like Burna,b (n, k), S2 (n, k) or S1 (n, k)
P
X
xn
xn
SΩ (n, k) y k = g(x)ey n≥1 SΩ (n,1) n!
(22)
(Stirling numbers of the second and first kind) are better
n!
n,k≥0
understood through the notion of one-parameter group,
conversely such groups naturally arinsing in Combinatowhich is a generalization of formula (16). A complete
rial Physics lead to such statistics and new ones some of
study of such a procedure and the details to perform the
which can be interpreted combinatorially.
solution of the normal ordering problem may be found in
[5].
6
Generalized Stirling numbers in
7
Combinatorial Physics
Conclusion
In this paper, we have broadened 2, 3 the domain of application of the exponential formula, a tool originated from
statistical physics. This broadening reveals us, together
with the essence of “why this formula works”, a possibility
of extension to denominators other than the factorial and,
on the other hand, provides a link with one-parameter
groups whose infinitesimal generators are (formal) vector
The complex algebra generated by these two symbols and fields on the line. The general combinatorial theory of
this unique relation, the Heisenberg-Weyl algebra, will be the correspondence (vector fields ↔ bivariate statistics)
here denoted by HWC . The consideration of evolution is still to be done despite the fact that we have already a
wealth of results in this direction.
1
In Quantum Mechanics, many tools boil down to the consideration of creation and annihilation operators which
will be here denoted respectively a† and a. These two
symbols do not commute and are subject to the unique
relation
[a, a† ] = 1 .
(17)
The name is related to the notion of free Burnside semigroups,
namely the quotient of the free semigroup A+ , where A is a finite
alphabet, by the the smallest congruence that contains the relators
ω n+m = ω n , ω ∈ A+ . For more details see [15].
2 A part of our setting can be reformulated in the categorical
context [2, 3]
3 Another direction is the q-exponential formula [10, 16].
Statistics on Graphs, Exponential Formula and Combinatorial Physics
Acknowledgements
We would like to thank Christian Krattenthaler (from
Wien) for fruitful discussions.
The research of this work was supported, in part, by the
Agence Nationale de la Recherche (Paris, France) under
Program No. ANR-08-BLAN-0243-2. We would like also
to acknowledge support from “Projet interne au LIPN
2009” “Polyzêta functions”.
References
[1] R. H. Bruck, A survey of binary systems, Ergebnisse der Mathematik und ihrer Grenzgebiete, new series, vol. 20, BerlinGttingen-Heidelberg, Springer, 1958.
[2] F. Bergeron, G. Labelle, and P. Leroux, Combinatorial Species
and Tree-Like Structures, Cambridge University Press, 1999.
[3] P. J. Cameron, C. Krattenthaler, and T. W. Müller, Decomposable functors and the exponential principle II, to appear.
[4] A. Dress and T. W. Müller, Decomposable functors and the exponential principle, Advances in Mathematics, vol. 129, pp. 188221, 1997.
[5] G. H. E. Duchamp, K. A. Penson, A. I. Solomon, A. Horzela and
P. Blasiak, One-parameter groups and combinatorial physics, in
Proc. of the Third International Workshop on Contemporary
Problems in Mathematic Physics (COPROMAPH3), PortoNovo (Benin), 2003. arXiv:quant-ph/0401126
[6] S. Eilenberg, Automata, Languages and Machines - volume A,
Academic Press, 1974.
[7] P. Flajolet, R. Sedgewick, Analytic Combinatorics, Cambridge
University Press, 2008.
[8] I. P. Goulden and D. M. Jackson, Combinatorial enumeration,
John Wiley & Sons, Inc., 1983.
[9] A. Joyal, Une théorie combinatoire des séries formelles, Advances in Mathematics, vol. 42, pp. 1-82, 1981.
[10] I. M. Gessel, A q-analog of the exponential formula, Discrete
Mathematics 306 (2006).
[11] B. Harris and L. Schoenfeld, The number of idempotent elements in symmetric semigroups, Journal of Combinatorial Theory, Series A, vol. 3, pp. 122-135, 1967.
[12] C. Krattenthaler and T. W. Müller, Equations in finite semigroups: explicit enumeration and asymptotics of solution numbers, Journal of Combinatorial Theory, Series A, vol. 105,
pp. 291-334, 2004.
[13] S. Lang, Complex analysis, Springer, 1999.
[14] E. S. Ljapin and A. E. Evseev, The Theory of Partial Algebraic
Operations, Kluwer Academic, 1997.
[15] A. Pereira do Lago and I. Simon, Free Burnside Semigroups,
Theoretical Informatics and Applications, vol. 35, pp. 579-595,
2001.
[16] C. Quesne, Disentangling q-Exponentials: A General Approach, International Journal of Theoretical Physics, Vol. 43,
No. 2, February 2004
[17] R. J. Ridell and G. E. Uhlenbeck, On the theory of the virial
development of the equation of state of monomatomic gases, J.
Chem. Phys., vol. 21, pp. 2056-2064, 1953.
[18] G. Segal, Configuration-spaces and iterated loop-spaces, Inventiones Mathematicae, vol. 21 (3), pp. 213-221, 1973.
[19] R. Stanley, Enumerative Combinatorics - Volume I, in Studies
in Advanced Mathematics, vol. 49, Cambridge University Press,
1997.
[20] J. Touchard, Sur les cycles des substitutions, Acta Mathematica, vol. 70, pp. 243-297, 1939.
5
| 5cs.CE
|
arXiv:1710.09875v1 [cs.NE] 26 Oct 2017
Phase Transitions in Image Denoising via Sparsely
Coding Convolutional Neural Networks
Jacob Carroll
Department of Physics & Center for Soft Matter and Biological Physics
Virginia Tech
Blacksburg, VA 24061
[email protected]
Nils Carlson
Department of Computer Science and Engineering
New Mexico Tech
Socorro NM, 87801
Garrett T. Kenyon
CCS-3, Information Sciences
Los Alamos National Laboratory
Los Alamos, NM, 87545
Abstract
Neural networks are analogous in many ways to spin glasses, systems which are
known for their rich set of dynamics and equally complex phase diagrams. We
apply well-known techniques in the study of spin glasses to a convolutional sparsely
encoding neural network and observe power law finite-size scaling behavior in
the sparsity and reconstruction error as the network denoises 32×32 RGB CIFAR10 images. This finite-size scaling indicates the presence of a continuous phase
transition at a critical value of this sparsity. By using the power law scaling relations
inherent to finite-size scaling, we can determine the optimal value of sparsity for
any network size by tuning the system to the critical point and operate the system
at the minimum denoising error.
1
Introduction
Spin glasses and neural networks are very analogous and draw many parallels in their dynamics.
Generally, a spin glass is a model of disordered magnetism. The simplest model of a spin glass, the
Ising model, is a network of N "spins" {σi } which take on the discrete values, connected by a weight
matrix Jij ∈ R that represents the strength of connection between the spins. The dynamics of these
systems is determined by the values of randomly chosen Jij , which are generally time independent
(3).
The similarity of these spin glass systems with neural networks is of interest to us because spin
glasses have been a focus of research in statistical physics for the last fifty years, and a large library
of machinery and techniques has been developed to deal with them. We would like to apply this
machinery to the field of neural networks.
For this paper we used PetaVision, a high performance neural simulation toolbox (1), to construct
sparsely coding convolutional neural networks and examine the relationship between the network’s
efficiency and sparsity. Interesting behavior in the efficiency of the networks as the sparsity was varied
led us to analyze the finite-size scaling of the network, a technique more commonly used in the study
of spin glasses, and discovered power law relationships that indicate a continuous (second-order)
phase transition is occurring in the networks as sparsity is varied.
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
2
Neural network
We used two networks in our simulation, both built using PetaVision. The first network was a
sparse auto-encoder network that trained the filter kernels of a convolutional layer using a Locally
Competitive Algorithm, as defined by Rozell et al. (7), as it attempted to iteratively converge on a
sparse representation of different input images. The second network (see Figure 1) used the same
sparsely encoding convolutional layer that was trained by the autoencoder to denoise images that had
very high Gaussian noise added to them.
The input for both networks were images from the CIFAR-10 image set (6).The image set was divided
into two parts. The first 50,000 images were used for training the filter kernels of the sparsely coding
convolutional layer for different levels of sparsity. Then, 10,000 additional images had very high
Gaussian noise added to them and were denoised by the denoising network for each level of sparsity
used in training.
We observed a distinct minimum in the percent reconstruction error of the noisy images as the sparsity
of the network was varied that displayed behavior analogous to continuous phase transitions seen in
spin glasses (see figure 2) (3; 9). With this as our motivation we investigated the presence of a phase
transition in our system.
Input Layer
Noise Layer
Input Error
Layer
Sparsely Coding Input Reconstruction
Convolutional Layer
Layer
Figure 1: A schematic of the denoising network. The Input Error Layer computes the difference
between the Noise Layer and the Input Reconstruction Layer, an alternative implementation of lateral
inhibition in LCA (8). Feature learning utilizes a local Hebbian rule to implement stochastic gradient
descent.
3
Phase transitions and finite-size scaling
A phase of a system is defined as a subspace of the microscopic system parameters where the system’s
dynamics obey the same macroscale laws and relations everywhere in that subspace. The space of
system parameters can have many phases, and the system can transition between them as system
control parameters change. The point of transition between two (or more) phases is known as the
critical point.
Phase transitions have been subject of significant study in Condensed Matter Physics, and it is well
established that the occurrence of a continuous phase transition1 is accompanied by a singularity
at the critical point in one or more system parameters when the system is of infinite size (9). It is
impossible to achieve infinite system sizes computationally, but this theory can be expanded to finite
systems where these singularities become truncated and rounded. These minima or maxima that the
singularities turn into at finite system sizes follow very specific relations with system size:
Location of Minima ∼ L−1/ν
−γ/ν
Height of Minima ∼ L
,
(1)
(2)
where L is the linear system size. This behavior is known as finite-size scaling (9; 2). The exponents
ν and γ two examples of "critical exponents". The critical exponents describe the behavior of the
system as it approaches the critical point (9; 2). Thus we can identify a phase transition in our network
by the existence and behavior of minima and maxima in the space of system parameters as we vary
the system size, which in our case will be the number of neurons in the convolutional layer. The
exponents we record, ν̄ and γ̄, will be proportional to γ and ν through some effective dimension of
our system.
1
A continuous (second-order) phase transition has a continuous change in the dynamics of the system as it
transitions between phases, while first-order phase transitions are discontinuous.
2
4
Results
The parameters of the system that we are interested in are the fraction of active neurons and the
average percent reconstruction error of our noisy images:
10,000
X ksi − ŝi k2
1
Perr =
,
(3)
10, 000 i=1
ksi k2
where Perr is the average percent reconstruction error, si is the ith original image before it has
Gaussian noise added to it, ŝi is the ith reconstruction of the noised image taken from the sparsely
coding convolutional layer (1; 7; 8).
213 neurons
214 neurons
215 neurons
216 neurons
217 neurons
0.26
0.28
●
0.24
Percent Reconstruction Error
The fraction of active neurons is controlled by a parameter λ, as described Rozell et al. (7), that
behaves monotonically with the sparsity of active neurons and inversely with the fraction of active
neurons. Through λ we can control the fraction of active neurons and observe how the average percent
reconstruction error behaves as the fraction of active neurons is varied. We observed a minimum in
Perr occur as we varied the fraction of active neurons for many different system sizes. These results
are summarized in Figure 2.
●
●
●
●
●
●
●
●
●
0.22
●● ●
0.00
0.02
0.04
0.06
0.08
0.10
0.12
Fraction Active Neurons
Figure 2: Average percent active error vs. fraction active neurons
0.03
0.04
0.05
101.91x−0.756
●
0.01
0.02
Fraction Active Neurons
0.2225
0.2215
0.2205
●
0.2195
Minimum Percent Error
10−0.621x−0.00745
214
215
216
217
214
System Size (Number of Neurons)
215
216
217
System Size (Number of Neurons)
(a) Height of minima vs. system size
(b) Location of minima vs system size
Figure 3: The power law behavior of the minimum average percent reconstruction error (a), and the
fraction of active neurons at that minimum (b). We report ν̄ = 1.32 ± 0.04 and γ̄ = 0.0099 ± 0.0095.
We measured the shift in height and location of the minima in Perr as the system size was varied,
and plot each on a log-log plot (see Figures 3 (a), and 3 (b)). We observe power law behavior in both
3
the location and height of the minima as the system size is varied. This satifies the finite-size scaling
requirements as defined in equations 1 and 2. This finite-size scaling behavior indicates a continuous
phase transition is occurring as the sparsity of the network is varied.
5
Discussion
The existence of phase transitions in neural networks is not unique to this sparsely coding convolutional system. The auto-associative network proposed by Hopfield (5) was shown by Hertz et al. (4)
to display a first-order phase transition in its memory capacity. If the number of patterns recorded by
the network exceeds a "critical fraction" of the network size, the output of the network is maximally
disordered (4).
We propose a similar mechanism is responsible for the observed continuous phase transition of our
sparsely coding convolution network, where the fraction of active neurons is analogous to the "critical
fraction" of learned patterns in the auto-associative network. If our network’s fraction of active
neurons is too far above the "critical fraction", the network will have the freedom to reconstruct the
noise in the image, while if the fraction of active neurons is too low, the network will only reconstruct
image components for which it has learned strong priors. These two different regions of dynamics
form our "phases". The existence of a phase transition in the average percent reconstruction error
of the network as the fraction of active neurons is varied guarantees the persistence of the power
law behavior seen in Figure 3 (b). This power law behavior allows us to predict the optimal fraction
of active neurons for any system size, which in turn can be tuned to through the parameter λ, as
described by Rozell et al. (7), to ensure that any sparsely coding convolutional network is operating
at the optimal level of sparsity.
The critical behavior of the network allows us to always achieve the minimum denoising error by
operating the network at this critical value of sparsity.
Acknowledgments
We gladly acknowledge helpful discussions with Uwe C. Täuber.
This work was supported by the Los Alamos National Laboratory under contract DE-AC5206NA25396.
Computations were performed using the Darwin Computational Cluster at Los Alamos National
Laboratory.
References
[1] Petavision. URL github.com/PetaVision/OpenPV.
[2] J. Cardy. Scaling and renormalization in statistical physics. Cambridge University Press, 1996. ISBN
0521499593.
[3] V. Dotsenko. An introduction to the theory of spin glasses and neural networks. World Scientific Lecture
Notes in Physics. World Scientific Publishing Company, 1995. ISBN 9810218737.
[4] J. A. Hertz, A. S. Krogh, and R. G. Palmer. Introduction to the theory of neural computation. Addison-Wesley
Publishing Company, 1991. ISBN 0201515601.
[5] J. J. Hopfield. Neural networks and physical systems with emergent collective computational abilities.
Proceedings of the National Academy of Sciences, 79(8):2554–2558, 1982. URL http://www.pnas.org/
content/79/8/2554.abstract.
[6] A. Krizhevsky. Learning multiple layers of features from tiny images. 2009.
[7] C. J. Rozell, D. H. Johnson, R. G. Baraniuk, and B. A. Olshausen. Sparse coding via thresholding and local
competition in neural circuits. Neural Computation, 20:2526–2563, 2008.
[8] P. F. Schultz, D. M. Paiton, W. Lu, and G. T. Kenyon. Replicating kernels with a short stride allows sparse
reconstructions with fewer independent kernels. arXiv preprint arXiv:1406.4205, 2014.
[9] U. C. Täuber. Critical dynamics: a field theory approach to equilibrium and non-equilibrium scaling
behavior. Cambridge University Press, 2014. ISBN 9780521842235.
4
| 9cs.NE
|
THE Q-TENSOR SQUARE OF FINITELY GENERATED
NILPOTENT GROUPS, q ≥ 0
arXiv:1603.05424v1 [math.GR] 17 Mar 2016
NORAÍ R. ROCCO AND EUNICE C. P. RODRIGUES
Abstract. In the present paper the authors extend to the q−tensor square G ⊗q G
of a group G, q a non-negative integer, some structural results due to R. D. Blyth,
F. Fumagalli and M. Morigi concerning the non-abelian tensor square G⊗G (q = 0).
The results are applied to the computation of G⊗q G for finitely generated nilpotent
groups G, specially for free nilpotent groups of finite rank. We also generalize to
all q ≥ 0 results of M. Bacon regarding an upper bound to the minimal number
of generators of the non-abelian tensor square G ⊗ G when G is a n−generator
nilpotent group of class 2. We end by computing the q−tensor squares of the free
n−generator nilpotent group of class 2, n ≥ 2, for all q ≥ 0. This shows that the
above mentioned upper bound is also achieved for these groups when q > 1.
1. Introduction
Let G and Gϕ be groups, isomorphic via ϕ : g 7→ g ϕ for all g ∈ G. Consider the
group ν(G), introduced in [23] as
(1)
ϕ
ν(G) = G ∪ Gϕ | [g, hϕ ]k = [g k , (hk )ϕ ] = [g, hϕ ]k , ∀g, h, k ∈ G .
It is a well known fact (see [23]) that the subgroup Υ(G) = [G, Gϕ ] of ν(G) is isomorphic to the non-abelian tensor square G ⊗ G, as defined by Brown and Loday
in their seminal paper [8]. A modular version of the operator ν was considered in
[10], where for any non-negative integer q the authors introduced and studied a group
ν q (G), which in turn is an extension of the so called q-tensor square of G, G⊗q G, first
defined by Conduché and Rodriguez-Fernandez in [11] (see also [14], [7]). In order
to describe the group ν q (G), if q ≥ 1 then let Gb = {b
k | k ∈ G} be a set of symbols,
b be
one for each element of G (for q = 0 we set Gb = ∅, the empty set). Let F (G)
b
b
b
the free group over G and ν(G) ? F (G) be the free product of ν(G) and F (G). As G
and Gϕ are embedded into ν(G) we shall identify the elements of G (respectively of
b Let J denote the normal closure in
Gϕ ) with their respective images in ν(G) ∗ F (G).
Date: March 19, 2018.
2010 Mathematics Subject Classification. 20F45, 20E26, 20F40.
Key words and phrases. Non-abelian tensor square; q-tensor square; nilpotent groups.
1
2
ROCCO AND RODRIGUES
b of the following elements, for all b
ν(G) ∗ F (G)
k, kb1 ∈ Gb and g, h ∈ G :
(2)
(3)
(4)
(5)
−1
d
g) ;
k g (k
g −1 b
−1
d
g) ;
k g ϕ (k
(g ϕ )−1 b
q
q
(b
k)−1 [g, hϕ ] b
k [g k , (hk )ϕ ]−1 ;
d1 (kb1 )−1 (
(b
k)−1 kk
q−1
Y
q−1−i −1
[k, (k1−i )ϕ ]k
) ;
i=1
(6)
[b
k, kb1 ] [k q , (k1q )ϕ ]−1 ;
(7)
[
[g,
h] [g, hϕ ]−q .
Definition 1.1. The group ν q (G) is defined to be the factor group
(8)
b
ν q (G) := (ν(G) ∗ F (G))/J.
Note that for q = 0 the sets of relations (2) to (7) are empty; in this case we have
∼
b
ν (G) = ν(G) ∗ F (G))/J
= ν(G).
Let R1 , . . . , R6 be the sets of
S relations corresponding to (2), . . . , (7), respectively,
and let R be their union, R = 6i=1 Ri . Therefore, ν q (G) has the presentation:
D
E
q
ϕ b
ϕ k
k
k ϕ −1
ϕ kϕ
k
k ϕ −1
ν (G) = G, G , G | R, [g, h ] [g , (h ) ] , [g, h ] [g , (h ) ] , ∀g, h, k ∈ G .
0
There is an epimorphism ρ : η q (G) G, g 7→ g, hϕ 7→ h, b
k 7→ k q . On the other
hand the inclusion of G into ν(G) induces a homomorphism ı : G → ν q (G). We
have g ıρ = g and thus ı is injective. Similarly the inclusion of Gϕ into ν(G) induces
a monomorphism : Gϕ → ν q (G). Thus we shall identify the elements g ∈ G and
g ϕ ∈ Gϕ with their respective images g ı and (g ϕ ) in ν q (G).
b By relations
Now let G denote the subgroup of ν q (G) generated by the images of G.
ϕ
q
q
(4), G normalizes the subgroup T = [G, G ] in ν (G) and hence Υ (G) = T G =
[G, Gϕ ]G is a normal subgroup of ν q (G). Thus we obtain ν q (G) = Gϕ · (G · Υq (G)),
where the dots mean internal semidirect products. It should be noted that the actions
of G and Gϕ on Υq (G) are those induced by the defining relations of ν q (G): for
any elements g, x ∈ G, hϕ , y ϕ ∈ Gϕ and b
k ∈ Gb we have [g, hϕ ]x = [g x , (hx )ϕ ] and
ϕ
ϕ
d
d
x ); similarly, [g, hϕ ]y = [g y , (hy )ϕ ] and (b
y ). In addition, for any
(b
k)x = (k
k)y = (k
q
yϕ
ϕ yϕ
q
τ ∈ Υ (G), (gτ ) = g[g, y ]τ ∈ GΥ (G).
By [10, Proposition 2.9] Υq (G) is isomorphic to the q-tensor square G ⊗q G, for all
q ≥ 0. We then get a result (see [10, Corollary 2.11]) analogous to one due to Ellis in
[14]: ν q (G) ∼
= G n (G n (G ⊗q G)); this generalizes a similar result found in [23] for
q = 0.
The commutator approach to G⊗G for the case q = 0, provided by the isomorphism
between G ⊗ G and the subgroup [G, Gϕ ] of ν(G) (see [23], and also [15]), has proven
suitable to treat of non-abelian tensor products of groups, Schur multipliers and many
other relevant invariants involving covering questions in groups; see for instance,
references [15], [24], [20], [4], [12], [21] and the GAP Package “POLYCYCLIC” in [13].
Q-TENSOR SQUARE OF NILPOTENT GROUPS
3
The extension of the existing theory from q = 0 to all non-negative integers q, as
addressed for instance in [10], broadens the scope of these connections, now in a hat
(“power”) and commutator approach to the q-tensor square, q ≥ 0.
In section 2 we extend to G ⊗q G, q ≥ 0, some structural results found in [5] and
[24] concerning G ⊗ G. In section 3 it is established an upper bound for the minimal
number of generators of G ⊗q G when G is a finitely generated nilpotent group of
class 2, thus generalizing a result of Bacon found in [2]. We end by computing the
q-tensor square of the free nilpotent group of rank n ≥ 2 and class 2, Nn,2 , q ≥ 0;
this will show, as in the case q = 0 (see [2, Theorem 3.2]), that the cited upper bound
is also attained for these groups when q > 1, although in this case Nn,2 ⊗q Nn,2 is a
non-abelian group.
Notation is fairly standard (see for instance [22]). If x and y are elements of a
group G then we write y x for the conjugate x−1 yx and [x, y] for the commutator
x−1 y −1 xy. Our commutators are left normed: [x, y, z] = [[x, y], z] for all x, y, z ∈ G,
and so on, recursively, for commutators of higher weights. The order of x (resp. of G)
is written o(x) (resp. |G|). As usual, γi (G) denotes the ith term of the lower central
series of G. For future reference we recall the well known Hall-Witt identity:
(9)
[x, y −1 , z]y [y, z −1 , x]z [z, x−1 , y]x = 1, ∀x, y, z ∈ G.
In view of the isomorphism given by [10, Proposition 2.9], from now on we identify
G ⊗q G with the subgroup Υq (G) = [G, Gϕ ]G ≤ ν q (G) and write [g, hϕ ] in place of
g ⊗ h, for all g, h ∈ G. Following [10] we write ∆q (G) for the subgroup h[g, g ϕ ]|g ∈
Gi ≤ Υq (G), which by Lemma 2.1 (vii) is a central subgroup of ν q (G). We write
τ q (G) for the factor group ν q (G)/∆q (G). The subgroup Υq (G)/∆q (G) of τ q (G) is
isomorphic to the q-exterior square G ∧q G. In order to avoid any confusion we
usually write [G, Gϕ ]τ ( G) to identify the q-exterior square G ∧q G with the image of
[G, Gϕ ] in τ q (G). We shall eventually write T to denote the subgroup [G, Gϕ ] of ν q (G)
in order to distinguish it from the nonabelian tensor square G ⊗ G ∼
= [G, Gϕ ] ≤ ν(G)
in the case q = 0.
The material presented here incorporates part of the doctoral thesis [25] of the
second named author, written under the supervision of the first.
2. Some Structural Results
In this section we extend results found in [5] and [24] related to the non-abelian
tensor square, from G ⊗ G to G ⊗q G, q ≥ 0. We begin by including some previous,
technical results for future references.
The following basic properties are consequences of the defining relations of ν q (G).
Lemma 2.1. [10, Lemma 2.4] Suppose that q ≥ 0. The following relations hold in
ν q (G), for all g, h, x, y ∈ G.
ϕ
(i) [g, hϕ ][x,y ] = [g, hϕ ][x,y] ;
(ii) [g, hϕ , xϕ ] = [g, h, xϕ ] = [g, hϕ , x] = [g ϕ , h, xϕ ] = [g ϕ , hϕ , x] = [g ϕ , h, x];
(iii) If h ∈ G0 (or if g ∈ G0 ) then [g, hϕ ][h, g ϕ ] = 1;
(iv) [b
x, [g, hϕ ]] = [b
x, [g, h]];
g
q
(v) (b
x) = x
b[x , g ϕ ];
4
ROCCO AND RODRIGUES
(vi) If [g, h] = 1 then [g, hϕ ] and [h, g ϕ ] are central elements of ν q (G), of the
same finite order dividing q. If in addition g, h are torsion elements of orders
o(g), o(h), respectively, then the order of [g, hϕ ] divides the gcd(q, o(g), o(h)).
(vii) [g, g ϕ ] is central in ν q (G), for all g ∈ G;
(viii) [g, hϕ ][h, g ϕ ] is central in ν q (G);
(ix) [g, g ϕ ] = 1, for all g ∈ G0 ;
(x) If [x, g] = 1 = [x, h], then [g, h, xϕ ] = 1 = [[g, h]ϕ , x].
Corollary 2.2. Let G be any group and let g, h be arbitrary elements in G. Then
(i) [G0 , Gϕ ] = [G, G0 ϕ ];
(ii) [G0 , Z(G)ϕ ] = 1;
(iii) If gG0 = hG0 then [g, g ϕ ] = [h, hϕ ];
(iv) If o0 (x) denotes the order of a coset xG0 ∈ G/G0 , then [g, hϕ ][h, g ϕ ] has order
dividing the gcd(q, o0 (g), o0 (h));
(v) The order of [h, hϕ ] divides the gcd(q, o0 (h)2 , 2o0 (h)).
Proof. Part (i) follows directly from Lemma 2.1 (iii) (see also [5, Corollary 1.2 (iii)]).
As for part (ii), see [23, Proposition 2.7 (i)].The remaining parts are appropriate
adaptations of [24, Lemma 3.1 (v)], using (7) and Lemma 2.1 (vi).
For our purposes we establish the following proposition, which may have its own
interest.
Proposition 2.3. Let G be a nilpotent group of class 2. Then the following hold
in ν q (G):
(i) G centralizes [G, Gϕ ];
(ii) [G0 , Gϕ ] (= [G, G0 ϕ ]) is a central subgroup of ν q (G);
(iii) Υq (G) (∼
= G ⊗q G ) is nilpotent of class at most 2.
Proof. (i) follows straightforward from 2.1 (iv) and relation (2), once G has nilpotency
class 2.
(ii). For all g, h ∈ G and c ∈ G0 we have:
[c, g ϕ ]h = [c, g ϕ ][c, g ϕ , h]
= [c, g ϕ ][c, g, hϕ ]
(by Lemma 2.1, (i))
(since [c, g] = 1, as G0 ≤ Z(G))
= [c, g ϕ ]
ϕ
= [c, g ϕ ]h
(by definition of ν q (G)).
ϕ
b by Lemma 2.1 (iv) and relations (2) we have that b
In addition, for all b
k ∈ G,
k [c,g ] =
\
[c,g] ) = b
b
k [c,g] = (k
k, since [c, g] = 1. This proves part (ii) (using the definition of
q
ν (G)), because [G0 , Gϕ ] is generated by all those [c, g ϕ ] above.
(iii). That Υq (G) is nilpotent and has nilpotency class at most 3 follows from [10,
Proposition 2.7, (i)]. Now, Υq (G) = [G, Gϕ ]G and thus, once G centralizes [G, Gϕ ],
we have
(Υq (G))0 = [[G, Gϕ ]G, [G, Gϕ ]G] ≤ [G, Gϕ ]0 [G, G].
Induction arguments can be used, together with Lemma 2.1 (i), (ii), (iv) and (v) and
defining relations (6) – (7) to get:
Q-TENSOR SQUARE OF NILPOTENT GROUPS
5
(a) [G, Gϕ ]0 = [G0 , (G0 )ϕ ] (see also [5, Proposition 1.3 (i)] or [23, Theorem 3.3]);
(b) [G, G] ≤ hGb0 i[G0 , Gϕ ] ≤ [G0 , Gϕ ].
Consequently, (Υq (G))0 ≤ [G0 , Gϕ ], which by part (ii) is central in ν q (G). This completes the proof.
For a finitely generated abelian group A, its q-tensor square Υq (A) can be computed by repeated applications of the following two results from [10].
Lemma 2.4. [10, Corollary 2.16] Let G = N × H be a direct product and set N =
N/N 0 N q , H = H/H 0 H q . Then
(i) Υq (G) = Υq (N ) × [N, H ϕ ][H, N ϕ ] × Υq (H);
(ii) [N, H ϕ ] ∼
= (N ⊗Zq H) ∼
= [H, N ϕ ].
Lemma 2.5. [10, Theorem 3.1] Let Cn (resp. C∞ ) be the cyclic group of order n
(resp. ∞), q a non- negative integer and d = gcd(n, q). Then
C ∞ ⊗q C ∞ ∼
= C∞ × Cq ;
if d is odd;
C n × C d ,
q
∼
Cn ⊗ Cn = Cn × Cd ,
if d is even and either 4|n or 4|q;
C × C , otherwise.
2n
d/2
Qr
Thus, if A = i=1 Ci is a direct product of the cyclic groups Ci , i = 1, . . . , r, where
Ci = hxi i, then
r
Y
Y
Υq (A) =
Υq (Ci ) ×
[Ci , Cjϕ ][Cj , Ciϕ ].
i=1
1≤i<j≤r
xbi i and [Ci , Cjϕ ][Cj , Ciϕ ] = h [xi , xϕj ][xj , xϕi ], [xi , xϕj ] i.
Here we have Υ (Ci ) =
Since ∆q (A) = h [a, aϕ ] | a ∈ Ai, we observe, like in [24, Proposition 3.3], that ∆q (A) =
h[xi , xϕi ], [xj , xϕk ][xk , xϕj ] | 1 ≤ i ≤ r, 1 ≤ j < k ≤ ri and thus it does not depend on
the particular set X = {x1 , . . . , xr } of generators of A. Consequently, we can write
q
h[xi , xϕi ],
q
Υq (A) = ∆q (A)EX
(A),
q
where EX
(A) = hxbi , [xj , xϕk ] | 1 ≤ i ≤ r, 1 ≤ j < k ≤ ri.
Remark 2.6. If x and y are commuting elements in any group G then by relations
(5)–(7) and Lemma 2.1 (vi) we get
q
q
x
cy = x
b yb [x, y ϕ ]−(2) = x
b yb [y, xϕ ]−(2) = yc
x,
q
q
and hence [x, y ϕ ]−(2) = [y, xϕ ]−(2) . In particular, if q = 2 then [x, y ϕ ] = [y, xϕ ].
This means for instance that in the decomposition of Υq (A) found above, the groups
d
n) =
[Ci , Cjϕ ] and [Cj , Ciϕ ] are not necessarily independent. Moreover, the identity (x
n
q
(b
x)n [x, xϕ ]−( 2 )(2) shows that the subgroups hxb i and h[x , xϕ ]i of Υq (C ) may have non
i
i
i
i
q
(A) is not
trivial intersection. Consequently, unlike the case q = 0, the subgroup EX
q
necessarily a complement of ∆ (A) (see also [5, Section 2]).
6
ROCCO AND RODRIGUES
Now let G be any group and write Gab = G/G0 . The natural projection G Gab
induces an epimorphism π : ν q (G) → ν q (Gab ). We denote by π0 the restriction of π to
Υq (G). By [10, Lemma 2.14 (iii)] we have that Ker(π0 ) = [G0 , Gϕ ][G, G0 ϕ ]hGb0 i, which
reduces to [G0 , Gϕ ]hGb0 i, by force of Corollary 2.2 (i). In addition, using relations (5)
and (7), an induction argument as in the proof of the Proposition2.3 (ii) shows that
d
0 ) ≤ [G0 , Gϕ ] and, consequently, Ker(π ) = [G0 , Gϕ ].
(G
0
The next Lemma extends [5, Lemma 2.1] to all q ≥ 0 (see also [24, Proposition
3.3]). We shall omit the proof.
Lemma 2.7. Let q be a non negative integer and G be a group such that Gab is finitely
generated. Assume that Gab is a direct product of the cyclic groups Ci = hxi G0 i, for
i = 1, . . . , r and set
E q (G) = hxbi , [xj , xϕk ] | 1 ≤ i ≤ r, 1 ≤ j < k ≤ r rangle [G0 , Gϕ ].
Then,
(i) ∆q (G) = h[xi , xϕi ], [xj , xϕk ][xk , xϕj ] | 1 ≤ i ≤ r, 1 ≤ j < k ≤ ri;
(ii) Υq (G) = ∆q (G)E q (G).
With the above notation, let π1 denote the restriction of π0 to ∆q (G), π1 : ∆q (G)
∆ (Gab ), and let N = Ker(π1 ). Therefore, N = ∆q (G)∩[G0 , Gϕ ] (= ∆q (G) ∩ E q (G)) ,
a central subgroup of Υq (G).
Our next theorem generalizes, to all q ≥ 0, Proposition 2.2 in [5], which in turn
improves Proposition 3.3 in [24].
q
Theorem 2.8. Let q ≥ 0 and assume that Gab is finitely generated. Then, with the
notation of Lemma 2.7, the following hold:
(i) Υq (G)/N ∼
= ∆q (Gab ) × (G ∧q G);
(ii) If q ≥ 1 and q is odd, then N = 1 and thus ∆q (G) ∼
= ∆q (Gab ) and Υq (G) ∼
=
q
ab
q
∆ (G ) × (G ∧ G);
(iii) For q = 0 or q ≥ 2 and q even, if Gab has no element of order two or if G0
has a complement in G, then also N = 1, ∆q (G) ∼
= ∆q (Gab ) and Υq (G) ∼
=
q
ab
q
∆ (G ) × (G ∧ G);
(iv) For q ≥ 2 and q even, if Gab has no element of order
2, then ∆q (G) is a
homocyclic abelian group of exponent q, of rank t+1
;
2
ab
(v) If G is free abelian of rank t, then the conclusion of the previous item holds
if q = 0.
for all q ≥ 1, while ∆q (G) is free abelian of rank t+1
2
Proof. (i): By Lemma 2.7 (ii) we have
Υq (G)
∆q (G)E q (G) ∼ ∆q (G) E q (G)
=
×
.
=
N
N
N
N
Now, ∆q (G)/N ∼
= ∆q (Gab ) and
E q (G)/N = E q (G)/(∆q (G) ∩ E q (G)) ∼
= Υq (G)/∆q (G) ∼
= G ∧q G.
This proves (i).
Q-TENSOR SQUARE OF NILPOTENT GROUPS
7
(ii), (iii), (iv) and (v): Suppose that the torsion subgroup of Gab is the direct product
of the cyclic groups hxi G0 i of order ni , 1 ≤ i ≤ s, and let the free part of Gab be
the direct product of the cyclic groups hyj G0 i, 1 ≤ j ≤ t. Thus, o0 (xi ) = ni and
o0 (yj ) = ∞. Set X := {xi | 1 ≤ i ≤ s} and Y := {yj | 1 ≤ j ≤ t}. Then G is generated
by X ∪ Y ∪ G0 . Using Lemma 2.1 and Corollary 2.2 (see also [24, Proposition 3.3 and
Remark 5]) we find that ∆q (G) is generated by the set ∆X ∪ ∆Y ∪ ∆XY , where
∆X = {[xi , xϕi ], [xj , xϕk ][xk , xϕj ] | 1 ≤ i ≤ s, 1 ≤ j < k ≤ s},
∆Y = {[yj , yjϕ ], [yk , ylϕ ][yl , ykϕ ] | 1 ≤ j ≤ t, 1 ≤ k < l ≤ t},
∆XY = {[xi , yjϕ ][yj , xϕi ] | 1 ≤ i ≤ s, 1 ≤ j ≤ t}.
Set nik = gcd(ni , nk ). Parts (iv) and (v) of Corollary 2.2 give ([xi , xϕk ][xk , xϕi ])nik = 1
and ([xi , yjϕ ][yj , xϕi ])ni = 1, while [xi , xϕi ]ni ∈ Ker(π0 ), ∀i, k = 1, . . . , s, i < k, ∀j =
1, . . . t. Actually, Ker(π0 ) is generated by the set {[xi , xϕi ]ni | 1 ≤ i ≤ s}, by [24,
Proposition 3.5]. Again by Corollary 2.2 (v), we get that if ni (= o0 (xi )) is odd,
then [xi , xϕi ]ni = 1, while [xi , xϕi ]2ni = 1 if ni is even. Consequently, N = Ker(π0 )
is an elementary abelian 2-group of rank at most r2 (Gab ), the 2 − rank of Gab (see
also [24, Corollary 3.6]). On the other hand, we should take into account that q is
involved in the upper bound found in Corollary 2.2 ( v). Thus, if q ≥ 1 and q is
odd, then gcd(q, 2ni ) = gcd(q, ni ) | ni and hence [xi , xϕi ]ni = 1, for all i = 1, . . . , s.
Therefore N = 1 in this case, proving part (ii). It should be also clear that N = 1
if r2 (Gab ) = 0. This proves (iii) in the case where Gab has no element of order 2.
Now, if G0 has a complement C in G, then every g ∈ G can be written as g = xh
with x ∈ C and h ∈ G0 . Corollary 2.2 (iii) says that [g, g ϕ ] = [x, xϕ ] and thus
∆q (G) = h[x, xϕ ] | x ∈ Ci = ∆q (Gab ). This completes the prof of part (iii) (see
also [5, Proposition 2.2]). Finally, we observe that [xi , xϕi ] = 1 = [xi , yjϕ ][yj , xϕi ]
in the case where r2 (Gab ) = 0 and q ≥ 2, q even. Here we have ∆q (G) = h∆Y i
and [yj , yjϕ ]q = 1 = ([yk , ylϕ ][yl , ykϕ ])q , ∀j, k, l = 1, . . . , t, k < l. That each of these
t+1
generators has order q follows immediately from Lemma 2.5, where we found
2
q
C∞ ⊗ C∞ ∼
= C∞ × Cq . Part (v) follows by an analogous argument , as in part (iv);
the last assertion can be also found in [24, Corollary 3.6]. The proof is complete.
We state the next Lemma for easy of reference, which in a certain sense extends
ideas found in [19] for the case q = 0. A proof is given in [8] for q = 0 (see also [5,
Proposition 3.2] for an alternative proof for this case) and in [16] for q ≥ 1.
Lemma 2.9. Let F/R be a free presentation of a group G. Then
G ∧q G ∼
= F 0 F q /[R, F ]Rq .
Notice that there is a map
ρ : ν q (G) −→ G, g 7−→ g, g ϕ 7−→ g and b
k 7−→ k q .
Let ρ0 = ρ|Υq (G) : Υq (G) −→ G, [g1 , g2 ϕ ] 7−→ [g1 , g2 ], b
k 7−→ k q .
q
q
Following [10] we write θ (G) = Ker(ρ) and µ (G) = Ker(ρ0 ) = Υq (G) ∩ θq (G). It
follows that Υq (G)/µq (G) ∼
= G0 Gq . If G = F/R is a free presentation of G, then
(10)
H 2 (G, Zq ) ∼
= R ∩ F 0 F q /Rq [R, F ] = (G ∧q G) ∩ M q (G),
8
ROCCO AND RODRIGUES
where M q (G) = R/Rq [R, F ] is the q-multiplier of G. From this we obtain (see for
instance [10, Theorem 2.12]):
(11)
µq (G)/∆q (G) ∼
= H 2 (G, Zq ),
for all q ≥ 0.
Corollary 2.10. Let Fn be the free group of rank n. Then
(i) For q ≥ 1,
(n+1)
Fn ⊗q Fn ∼
= Cq 2 × (Fn )0 (Fn )q .
(ii) ([9, Proposition 6]) For q = 0,
(n+1
)
Fn ⊗ Fn ∼
C
= ∞ 2 × (Fn )0 .
Proof. Since Fnab is free abelian of rank n, by Theorem 2.8 (ii), (iii) and (v), we have:
Υq (Fn ) ∼
= ∆q (F ab ) × (Fn ∧q Fn ).
n
n+1
2
(
(i): If q ≥ 1 then ∆q (Fnab ) ∼
= Cq
)
and, by Lemma 2.9 (i) with G = Fn and R = 1,
Fn ∧ Fn ∼
= (Fn )0 (Fn )q .
q
This proves (i).
(n+1)
(ii): If q = 0 then ∆q (Fnab ) ∼
= C∞ 2 and, again by the previous Lemma,
Fn ∧ Fn ∼
= (Fn )0 .
This completes the proof.
Corollary 2.11. Let Nn,c = Fn /γc+1 (Fn ) be the free nilpotent group of class c ≥ 1
and rank n > 1. Then
(i) For q ≥ 1,
(
Nn,c ⊗q Nn,c ∼
= Cq
n+1
2
)
×
(Fn )0 (Fn )q
.
γc+1 (Fn )q γc+2 (Fn )
(ii) ([6, Corollary 1.7]) For q = 0,
(n+1)
Nn,c ⊗ Nn,c ∼
= C∞ 2 × (Nn,c+1 )0 .
Proof. (i) and (ii) follow by similar arguments as above, taking into account that here
we have R = γc+1 (Fn ) and thus [R, F ], as in Lemma 2.9, is precisely γc+2 (Fn ).
3. q-Tensor Squares of Nilpotent Groups of Class 2
In this section we restrict our considerations to finitely generated nilpotent groups
G of class two. We begin with a general result concerning polycyclic groups found in
[10]; this generalizes to all q ≥ 0 a result due to Blyth and Morse in [4] for q = 0,
which in turn extends to all polycyclic groups a similar result for finite solvable groups
found in [24].
Lemma 3.1. ([10, Corolary 3.6]) Let G be a polycyclic group with a polycyclic generating sequence pgs(G) = (a1 , . . . , an ) . Then
Q-TENSOR SQUARE OF NILPOTENT GROUPS
9
(i) [G, Gϕ ], a subgroup of ν q (G), q ≥ 0, is generated by
ϕ
ϕ
ϕ
ϕ β
ϕ
α
[G, G ] = [ai , ai ], [ai , aj ][aj , ai ], [ai , (aj ) ], for 1 ≤ i < j ≤ n, 1 ≤ k ≤ n ,
(ii) Υq (G), a subgroup of ν q (G), q ≥ 1, is generated by
ϕ
ϕ
ϕ
ϕ β d
q
α
Υ (G) = [ai , ai ], [ai , aj ][aj , ai ], [ai , (aj ) ], (ak ), for 1 ≤ i < j ≤ n, 1 ≤ k ≤ n ,
(
1
where α =
±1
if o(ai ) < ∞
and β =
if o(ai ) = ∞
(
1
±1
if o(aj ) < ∞
if o(aj ) = ∞.
ϕ
ϕ
(iii) ∆q (G) is generated by the set {[ai , aϕ
i ], [ai , aj ][aj , ai ], for 1 ≤ i < j ≤ n}.
Now let G be a finitely generated nilpotent group of class two and assume that G
is generated by g1 , g2 , · · · , gn . Thus, any element g ∈ G can be written as
n
Y
Y
(12)
g=
gimi
[gj , gk ]ljk ,
i=1
1≤j<k≤n
where the exponents mi and ljk are integers. Consequently, G has the following
polycyclic generating set
(13)
{gi , 1 ≤ i ≤ n} ∪ {[gj , gk ], 1 ≤ j < k ≤ n}.
The following theorem extends a result of Bacon in [2, Theorem 3.1] (see also [3])
for all q ≥ 0. We provide a proof for the case q = 0 using the commutator approach;
the general case follows straightforward from this case and Lemma 3.1 (ii), but we
shall prove it in this case too, for the sake of completeness.
Theorem 3.2. Let G be a nilpotent group of class two with d(G) = n, then
2
(i) ([2, Theorem 3.1]) d([G, Gϕ ]) ≤ n(n +3n−1)
;
3
2 +3n+2)
n(n
(ii) d(G ⊗q G) ≤
, for all q ≥ 0;
3
(iii) In particular, if G has finite exponent e(G) and gcd(q, e(G)) = 1 , then
d(G ⊗q G) ≤ n2 .
Proof. On assuming that G is generated by g1 , . . . , gn then we obtain the polycyclic
generating set given by (13). Thus, by Lemma 3.1 (i) we have that Υ(G) = [G, Gϕ ]
is generated by the following set of elements:
{[giα , (gjϕ )β ] : 1 ≤ i, j ≤ n}∪
{[giα , ([gj , gk ]ϕ)β ] : 1 ≤ i ≤ n, 1 ≤ j < k ≤ n}∪
{[[gj , gk ]β , (giϕ )α ] : 1 ≤ i ≤ n, 1 ≤ j < k ≤ n}∪
{[[gr , gs ]α , ([gt , gu ]ϕ )β ] : 1 ≤ r < s ≤ n, 1 ≤ t < u ≤ n},
where α, β ∈ {−1, 1}.
Now by Lemma 2.1, parts (ii), (iii), (ix), (x), and the fact that G has class 2, we can
further reduce the above set to obtain
(14)
{[gi , gjϕ ] : 1 ≤ i, j ≤ n} ∪ {[gi , [gj , gk ]ϕ ] : 1 ≤ i ≤ n, 1 ≤ j < k ≤ n}.
10
ROCCO AND RODRIGUES
This new set has n2 generators of the form [gi , gjϕ ] and n(n − 1) generators of the form
[gi , [gj , gi ]ϕ ]. It remains to count the generators of the form [gi , [gj , gk ]ϕ ], when i, j, k
are all distincts and j < k. Now by [23, Corollary 3.2] ν(G) has nilpotency class at
most 3 and thus, again
[gjϕ , gkϕ , gi ][gkϕ , giϕ , gj ][gi , gj , gkϕ ] = 1.
It then follows that [gi , [gj , gk ]ϕ ] = [gj , [gi , gk ]ϕ ][gk , [gj , gi ]ϕ ]. Therefore,
1
d([G, Gϕ ]) ≤ n(n2 + 3n − 1).
3
This proves part (i).
(ii): Part (i) also proves (ii) in case q = 0, giving us the better bound for d(G ⊗ G).
Thus, we shall assume q ≥ 1. Since Υq (G) = [G, Gϕ ]G it suffices to control the number
of generators of both [G, Gϕ ] and G. We already know by part (i) that [G, Gϕ ] is
generated by the set {[gi , giϕ ] : 1 ≤ i, j ≤ n} ∪ {[gi , [gj , gk ]ϕ ] : 1 ≤ i ≤ n, 1 ≤ j <
k ≤ n}. Now, by definition the subgroup G is generated by Gb = {b
g , g ∈ G}. By
Qq−1
q
ϕ −i g q−1−i b
c
) h, for all
the defining relations (5) of ν (G) we have gh = gb ( i=1 [g, (h ) ]
ϕ
c
b
b
b
g, h ∈ G, and hence gh ≡ gb h (mod [)G, G ], for all gb, h ∈ G. Since every element
g ∈ G has a unique expression in the form (12) and given the fact that, for every
ϕ q
ϕ
\
commutator [gj , gk ] ∈ G0 , [g
j , gk ] = [gj , gk ] ∈ [G, G ] (by relations (7)), we see in later
ϕ
stage that G is generated, modulo [G, G ], by the n elements gb1 , . . . , gbn . Therefore,
we conclude that d(Υq (G)) ≤ 13 (n3 + 3n2 + 2n). This proves part (ii).
(iii): If in particular G has finite exponent e(G) and gcd(q, e(G)) = 1, then we see by
Lemma 2.1 (vi) that all generators of the forms [gi , giϕ ] and [gi , [gj , gk ]ϕ ] are trivial.
Consequently, d(G ⊗q G) ≤ n2 in this case. The proof is complete.
In [1] Aboughazi computed the nonabelian tensor square of the Heisenberg group
H = F2 /γ3 (F2 ), where F2 denotes the free group of rank 2. There, it is found that
H⊗H ∼
= Z6 , thus showing that the bound in Theorem 3.2 is sharp. Later, Bacon in
[2, Theorem 3.2] computed Nn,2 ⊗ Nn,2 for all n ≥ 2, to show that the bound is also
reached for the free n−generator nilpotent group of class 2, n > 1: Nn,2 ⊗ Nn,2 is a
free abelian group of rank 31 n(n2 + 3n − 1).
It is not difficult to extend these results to the q−tensor square Υq (Nn,2 ), q ≥ 1, to
show that the bound in Theorem 3.2 (ii) is also attained. In fact, the next proposition
is but a specialization of Corollary 2.11. We write M (G) to denote the Schur multiplier
H2 (G, Z) of G.
Proposition 3.3. Let Nn,2 be the free nilpotent group of rank n > 1 and class 2,
Nn,2 = Fn /γ3 (Fn ). Then,
(i) ([2, Theorem 3.2]) Nn,2 ⊗ Nn,2 is free abelian of rank 31 n(n2 + 3n − 1). More
precisely,
0
Nn,2 ⊗ Nn,2 ∼
.
= ∆(Fnab ) × M (Nn,2 ) × Nn,2
n+1
q
0
(ii) Nn,2 ⊗q Nn,2 ∼
Nn,2
, where Mn (3) = 31 (n3 − n) is
= (Cq )(( 2 )+Mn (3)) × Nn,2
the q−rank of γ3 (Nn,2 )/γ3 (Nn,2 )q γ4 (Nn,2 ), according to the Witt’s formula.
Q-TENSOR SQUARE OF NILPOTENT GROUPS
11
Consequently, for q > 1
1
d(Nn,2 ⊗q Nn,2 ) = (n3 + 3n2 + 2n).
3
Proof. As in the proof of Corollary 2.11, using Lemma 2.9 we can write
G ∧q G ∼
= F 0 F q /[R, Fn ]Rq .
n
n
In view of (10),
H 2 (G, Zq ) ∼
= R ∩ Fn0 Fnq /Rq [R, Fn ].
Taking into account that R = γ3 (Fn ) ≤ Fn0 Fnq we have the exact sequence
γ3 (Fn )
Fn0 Fnq
Fn0 Fnq
(15)
1→
→ 1.
→
→
γ4 (Fn )γ3 (Fn )q
γ4 (Fn )γ3 (Fn )q
γ3 (Fn )
M (3)
Here we find that H2 (Nn,2 , Zq ) ∼
= γ3 (Fn )/γ4 (Fn )γ3 (Fn )q ∼
= Zq n , for all q ≥ 0, where,
by the Witt’s formula for the rank of γr (Fn )/γr+1 (Fn ),
r
1X
Mn (r) =
µ(d)n d ,
r
d|r
where µ denotes the Möbius function.
(i): If q = 0 then the exact sequence (15) splits and thus we have, by Corollary 2.11,
1
(n3 −n)
(n+1)
(n+1)
(n)
3
Nn,c ⊗ Nn,c ∼
.
= C∞ 2 × γ3 (Fn )/γ4 (Fn ) × γ2 (Fn )/γ3 (Fn ) ∼
= C∞ 2 × C∞2 × C∞
This is the result of Bacon, that d(Nn,2 ⊗ Nn,2 ) = 31 n(n2 + n − 1).
(ii): If q > 1 then we see by the generators of Nn,2 ⊗q Nn,2 found in Theorem 3.2
that the image by ρ0 of the subgroup hb
gi , [gj , gkϕ ] | 1 ≤ i ≤ n, 1 ≤ k < j ≤ ni
q
0
Nn,2
of Nn,2 , while the subgroup h[gj, gi , gkϕ ]i, where [gj, gi , gk ]
is the subgroup Nn,2
is a basic commutator of γ3 (Fn )/γ4 (Fn ), is isomorphic to γ3 (Fn )/γ4 (Fn )γ3 (Fn )q ∼
=
H2 (Nn,2 , Zq ), a homocyclic abelian group of exponent q and q−rank Mn (3) = 31 (n3 −
n). Consequently, also in this case we get that the sequence (15) splits and we then
find that
(n+1)+Mn (3)
q
0
× Nn,2
Nn,2
.
Nn,2 ⊗q Nn,2 ∼
= Cq 2
Therefore, d(Nn,2 ⊗q Nn,2 ) = 31 n(n2 + 3n + 2), thus showing that the upper bound
given in Theorem 3.2 (ii) is also achieved when q > 1.
References
[1] Aboughazi, R., Produit Tensoriel du Groupe D’Heisenberg, Bull. Soc. math. France, 115 (1987),
95–106.
[2] Bacon, M., On the non-abelian Tensor Square of a Nilpotent Group of Class Two, Glasgow
Math. J., 3 (1994), 291–295.
[3] Bacon, M. and L.-C. Kappe The nonabelian tensor square of a 2-generator p-group of class 2.
Arch. Math. (Basel) 61 (1993), 508–516.
[4] Blyth, R. D., Morse, R. F., Computing the nonabelian tensor square of polycyclic groups, J.
Algebra 321 (2009), 2139–2148.
[5] Blyth, R. D., Fumagalli, F., Morigi, M., Some structural results on the non- abelian tensor
square of groups, J. Group Theory, 13, No.1 (2010), 83–94.
12
ROCCO AND RODRIGUES
[6] Blyth, R. D., Moravec, P., Morse, R. F. On the nonabelian tensor squares of free nilpotent
groups of finite rank. In Computational Group Theory and the Theory of Groups, Contemporary
Mathematics 470 (American Mathematical Society, 2008), pp. 27–44.
[7] Brown, R., q-perfect Groups and Universal q-central Extensions, Publ. Mat., 34 (1990), 291-297.
[8] Brown, R. and Loday, J.-L., Van Kampen Theorems for Diagrams of Spaces, Topology, 26
(1987), 311–335.
[9] Brown, R., Johnson, D. L., Robertson, E. F., Some computations of non-abelian tensor products
of groups, J. Algebra 111 (1987), 177– 202, .
[10] Bueno, T. P. and Rocco, N. R., On the q-tensor square of a group, J. Group Theory, 14 (2011),
785–805.
[11] Conduché D., Rodriguez-Fernandez, C., Non-abelian Tensor and Exterior Products modulo q
and Universal q-central Relative Extensions, J. Pure Appl. Algebra, 78, No.2 (1992), 139–160.
[12] Eick, B., Nickel, W., Computing the Schur multiplicator and the nonabelian tensor square of a
polycyclic group, J. Algebra 320 No.2 (2008), 927–944.
[13] Eick, B., Nickel, W., POLYCYCLIC - Computation with polycyclic groups, A GAP 4 package. In The Gap Group, GAP—Groups, Algorithms, and Programming, Version 4.5.7, 2014
(http://www.gap–system.org).
[14] Ellis, G., Tensor Product and q-crossed Modules, J. London Math. Soc., 2 (51) No.2 (1995),
243–258.
[15] Ellis, G., Leonard, F., Computing Schur multipliers and tensor products of finite groups, Proc.
Royal Irish Acad., 95A (1995), 137–147.
[16] Ellis, G., Rodríguez-Fernández, C., An exterior product for the homology of groups with integral
coefficients modulo p, Cah. Top. Géom. Diff. Cat. 30, 339–343, (1989).
[17] Gilbert, N. D., Higgins, P. J., The non-abelian tensor product of groups and related constructions, Glasgow Math. J. 31 (1989), 17–29.
[18] McDermott, A., The Nonabelian Tensor Product of Groups: Computations and Structural Results, PhD Thesis, Nat. Univ. Ireland, Gallway, 1998.
[19] Miller, C., The Second Homology Group of a Group; Relations Among Commutators, Proc.
Amer. Math. Soc. 3 (1952), 588–595.
[20] Nakaoka, I. N., Non abelian tensor products of solvable groups, J. Group Theory 3 (2000),
157–167.
[21] Nakaoka, I. N. and Rocco, N. R., A survey of non-abelian tensor products of groups and related
constructions, Bol. Soc. Paran. Mat. 30 1 ( 2012), 77–89.
[22] Robinson, Derek J. S., A Course in the Thoery of Groups, second edition, Graduate Texts in
Mathematics 80, Springer-Verlag New York (1996).
[23] Rocco, N. R.,On a Construction Related to the Non-Abelian Tensor Square of a Group, Bol.
Soc. Bras. Mat., 22, No.1 (1991), 63–79.
[24] Rocco, N. R., A Presentation for a Crossed Embedding of Finite Solvable Groups, Comm. in
Algebra,22(6) (1994), 1975–1998.
[25] Rodrigues, Eunice C. P., Cotas Superiores para o Expoente e o número mínimo de geradores do
Quadrado q-Tensorial de Grupos Nilpotentes, Doctoral Thesis (in Portuguese), Universidade de
Brasília, Brasilia, DF, Brazil, 2011. Available at http://repositorio.unb.br/handle/10482/8717.
Departamento de Matematática, Universidade de Brasília, Brasilia-DF, 70910-900
Brazil
E-mail address: [email protected]
Departamento de Matemática, Universidade Federal de Mato Grosso, Rondonópolis-MT, 85735-001 Brazil
E-mail address: [email protected]
| 4math.GR
|
1
Curvature of Hypergraphs via Multi-Marginal Optimal Transport
Shahab Asoodeh1, Tingran Gao2 , and James Evans3
arXiv:1803.08584v1 [cs.IT] 22 Mar 2018
Abstract
We introduce a novel definition of curvature for hypergraphs, a natural generalization of graphs, by introducing a multimarginal optimal transport problem for a naturally defined random walk on the hypergraph. This curvature, termed coarse scalar
curvature, generalizes a recent definition of Ricci curvature for Markov chains on metric spaces by Ollivier [Journal of Functional
Analysis 256 (2009) 810-864], and is related to the scalar curvature when the hypergraph arises naturally from a Riemannian
manifold. We investigate basic properties of the coarse scalar curvature and obtain several bounds. Empirical experiments indicate
that coarse scalar curvatures are capable of detecting “bridges” across connected components in hypergraphs, suggesting it is an
appropriate generalization of curvature on simple graphs.
I. I NTRODUCTION
Complex systems or datasets are often modeled as weighted simple graphs in network science problems. While edges in these
simple graphs qualitatively characterize similarity or adjacency relations among entities represented by the graph vertices, the
edge weights are frequently used to quantify the nature of the interactions between pairs of nodes. Simple yet powerful as these
simple graph models are, many recent work reported the importance of understanding higher-order interactions among more
than a pair of nodes, rendering simple graphs insufficient as a natural model for capturing the network structure information in
these practices. Applications of this type include spatial network [1], image tagging [2], image retrieval [3], cellular networks
[4], and co-authorship network [5], to name just a few. Hypergraphs have been proposed as a replacement to tackle this
difficulty.
Roughly speaking, a hypergraph H = (V, E) consists of a finite set V of vertices and a set of hyperedges E ⊆ 2V –
just as edges in a simple graph that can be identified with vertex pairs, hyperedge E ∈ E are subsets of V . The ubiquitous
influence in modeling complex networks fostered numerous recent developments in the theory and algorithms of hypergraphs,
including extensive studies of the spectral and algebraic properties such as hypergraph Laplacian [6], hypergraph partitioning
[7], Cheeger’s inequality for hypergraph [8], and spectrum of hypergraphs [9].
Among many tools developed for better understanding the geometry of graphs, the graph Ricci curvature [10]–[14] has
attracted an increasing amount of interest in the past years. In his original work [10], Ollivier defined coarse Ricci curvature
for metric measure spaces, including simple graphs and Markov chains as special cases. In a nutshell, the coarse Ricci curvature
summarizes the behavior of shortest paths with close-by starting points and parallel initial directions: two such paths tend to
get closer to each other in a metric space of positive Ricci curvature, and further if the space is negatively curved. On simple
graphs, this notion of Ricci curvature has found applications ranging from bounding the chromatic number [15] and analyzing
the Internet topology [16] to measuring the stability in financial networks [17], brain structural connectivity [18], and similarity
of networks [19].
This paper proposes a novel definition of curvature for hypergraphs by generalizing Ollivier’s coarse Ricci curvature through
a multi-marginal optimal transport framework (see e.g. [20]–[22]). Analogous to coarse Ricci curvatures, our definition of
hypergraph curvature is grounded upon differential geometric intuitions, and reduces to the graph Ricci curvature when the
hypergraph is a simple graph. While the coarse Ricci curvature is defined for pairs of points, which are naturally identified with
edges in simple graphs, we need to adjust the construction to account for ≥ 3 vertices joined by a hyperedge simultaneously;
consequently, the geometric information captured in our definition is a summary of a small neighborhood enclosing all the
end nodes of a hyperedge, as opposed to the directional information revealed by the Ricci curvature. In fact, our construction
corresponds to the scalar curvature of Riemannian manifolds under appropriate manifold assumptions analogous to the manifold
Ricci curvature example in [10].
The rest of the paper is structured as follows. We list a few useful notation in Section II-A. After a brief review of coarse
Ricci curvature in Section II-B, we define coarse scalar curvature in Section II-C. Based on this definition, we then propose
our notion of hypergraph curvature and investigate its properties in Section II-D. We derive a closed form for the curvature
of complete uniform hypergraphs in Section III and a general lower bound for hyperpaths in Section IV. In Section V, we
provide a detailed consistency result for the definition of coarse scalar curvature in a Riemannian manifold setting. Finally, we
conclude the paper in Section VI.
1 Computation
Institute and Institute of Genomics and System Biology, The University of Chicago, Chicago, IL 60637 [email protected]
of Statistics, The University of Chicago, Chicago, IL 60637 [email protected]
3 Computation Institute and Department of Sociology, The University of Chicago, Chicago, IL 60637 [email protected]
2 Department
2
II. H YPERGRAPH C URVATURE VIA M ULTI -M ARGINAL O PTIMAL T RANSPORT
A. Notations
P
PFor each vertex i of a hypergraph H = (V, E), we use di := E∈E 1{i∈E} to denote the degree of vertex i and d(E) :=
r∈V 1{r∈E} to denote the cardinality of hyperedge E ∈ E. Similar to graph, we use N (i) for the neighbors of vertex i, i.e.,
N (i) = {j ∈ V : ∃E ∈ E, (i, j) ∈ E}. For a pair of vertices i and j of a hypergraph, d(i, j) denotes the shortest distance, i.e.,
d(i, j) = r if there exist r interesting hyperedges E1 , . . . , Er , such that i ∈ E1 , j ∈ Er , and Ek ∩ Ek+1 6= ∅ for 1 ≤ k ≤ r − 1.
We will always denote M for a d-dimensional Riemannian manifold. We use expx (·) : Tx M → M to denote the exponential
map. For any x ∈ M and v ∈ Tx M with kvk = 1, we denote Ricx (v, v) for the Ricci curvature at x ∈ M in the direction of
v ∈ Tx M , defined as
d
1 X
hR (v, ei ) v, ei i
(1)
Ricx (v, v) =
d − 1 i=2
where v, e2 , · · · , ed constitutes an orthonormal basis for Tx M , and R is the Riemannian curvature tensor. We will often write
Ric (v, v) for Ricx (v, v) when the point x is fixed throughout the discussion. Averaging out Ricx (v, v) for v ranging in an
orthonormal basis of Tx M gives rise to the scalar curvature at x ∈ M :
d
Scal (x) =
1X
Ricx (zi , zi )
d i=1
(2)
where z1 , · · · , zd constitutes an orthonormal basis for Tx M . Equivalently, the scalar curvature can be obtained from averaging
out the Ricci curvature over the unit sphere in the tangent plane, i.e. (c.f. [23, Exercise 4.9])
Z
1
Ricx (v, v) dSd−1 (v)
(3)
Scal (x) =
ωd−1 Sd−1 (0)
where Sd−1 (0) is the unit sphere in Tx M and ωd−1 is the volume of the standard (d − 1)-dimensional sphere in Rd .
For any set A ∈ Rn , we let P (A) denote the set of all probability measures defined on A.
B. Graph Ricci Curvature
Given a graph G = (V, E) and a pair of vertices x, y ∈ V , Ollivier [10] defined the curvature of edge (x, y) ∈ E as
κ(x, y) := 1 −
W (mx , my )
,
d(x, y)
(4)
where d(x, y) is the shortest distance from x to y, mi is the uniform random walk starting at i ∈ V , and W (mx , my ) is the
Wasserstein distance between mx and my given by
X
W (mx , my ) :=
inf
d(r, s)π(r, s),
(5)
π∈Π(mx ,my )
(r,s)∈V 2
where Π(mx , my ) is the set of all joint distributions having mx and my as marginals (i.e., the set of all couplings of mx and
my ). He then showed that positive curvature is equivalent to the contraction of the random walk under Wasserstein’s distance
which in turn leads to the existence of a unique stationary distribution.
To justify that (4) is a valid discrete version of Ricci curvature, Ollivier argued as follows. In a d-dimensional Riemannian
manifold (M, dM ), consider the random walk (c.f. Definition 1) dmεx for each x ∈ M and ε > 0 given by
dmεx (s) :=
1B(x,ε) (s)
dvol(s),
vol(B(x, ε))
(6)
where B (x, ε) = {z ∈ M | dM (x, z) < ε} is the metric ball with radius ε centered at x and 1{x∈A} is the indicator function
of set A. It is then shown [10, Example 7] that for sufficiently small δ = dM (x, y) > 0
W1 mεx , mεy
ε2 Ric(v, v)
1−
=
+ O(ε3 + ε2 δ),
(7)
dM (x, y)
2(n + 2)
where v ∈ Tx M is a unit tangent vector at x such that expx (δv) = y and where W1 mεx , mεy is the L1 -Wasserstein distance
between mεx and mεy
Z
ε
ε
inf
dM (x, y) dπ (x, y) .
(8)
W1 mx , my :=
π∈Π(mεx ,mεy ) M×M
In Riemannian geometry, Ricci curvature Ricx (v, v) is, up to a scaling factor, the average of the sectional curvatures of all
two-dimensional subspaces of Tx M passing through v [23, §4] and hence, it measures the coupling of the random walks. In
this context, if the curvature of a point in a manifold is zero, it is locally on a Euclidean space, positive if it is locally on an
sphere and negative if it locally on a hyperbolic space.
3
C. Coarse Scalar Curvature for Metric Spaces
We begin our construction of coarse scalar curvature by defining random walks on a metric space, which was used in [10]
to define coarse Ricci curvatures.
Definition 1 ( [10] Definition 1). Let (M, dM ) be a Polish metric space equipped with its Borel σ-algebra. A random walk
m on M is a family of probability measures {mx | x ∈ M } satisfying
• The map x 7→ mx is measurable;
• Each mx has finite first moment.
Definition 2 (Coarse Scalar Curvature). For a collection of n points Xn := {x1 , · · · , xn } in a metric space (M, dM ) with
random walk m := {mx | x ∈ M }, define the coarse scalar curvature for Xn by
κ ({Xn }) := 1 −
W1 (Xn )
,
c(x1 , . . . , xn )
(9)
where W1 (Xn ) is the minimum of the multi-marginal optimal transport problem
Z
W1 (Xn ) := inf
c (ξ1 , · · · , ξn ) dπ (ξ1 , · · · , ξn ) ,
π∈Π(Xn )
with
M×···×M
n
o
Π (Xn ) := Π (m1 , · · · , mn ) = π ∈ P (M × · · · × M ) | (Projk )# π = mxk , ∀1 ≤ k ≤ n ,
with (A)# π being the push-forward of measure π under mapping A and
c (ξ1 , · · · , ξn ) := inf
z∈M
n
X
dM (ξi , z) .
(10)
i=1
It must be mentioned that the multi-marginal optimal transport problem is first studied in Gangbo and Świȩch [22] where
they showed the necessary conditions for the existence and uniqueness of the optimizer when c(ξ1 , · · · , ξn ) is the sum of
pairwise ℓ2 distances. The coarse scalar curvature is closely tied to the multi-marginal optimal transport problem among n ≥ 2
probability measures mx1 , · · · , mxn , which is a direct generalization of the pairwise Wasserstein distance between measures.
We will make frequent use of the following well-known facts in the theory of multi-marginal optimal transport problems (see
e.g. [20]–[22], [24], [25] and the references therein):
Proposition 1 (Multi-marginal and barycenter [20]). The minimum of the multi-marginal optimal transport problem is equal
to the minimum of the Wasserstein barycenter problem, i.e.
W1 (Xn ) =
inf
ν∈P(M)
n
X
W1 (mxi , ν)
(11)
i=1
where the L1 -Wasserstein distance W1 (m, ν) for any m, ν ∈ P (M ) is defined in (8).
The minimization problem in (11) is called Wasserstein barycenter of Xn .
Proposition 2 (Duality [25]).
W1 (Xn ) =
sup
n Z
X
(f1 ,··· ,fn )∈F i=1
fi (ξ) dmxi (ξ)
(12)
M
where the supremum is taken over
n
o
n
X
fi (ξi ) ≤ c (ξ1 , · · · , ξn ) .
F := (f1 , · · · , fn ) fi ∈ L1 (M, mxi ) ∀1 ≤ i ≤ |ex | ,
i=1
We postpone the justification of our nomenclature of “scalar curvature” in Section V, under an appropriate manifold setting.
Briefly speaking, at least when the hypergraph arises from a Riemannian manifold in a natural geometric construction, the
coarse scalar curvature (9) is asymptotically lower bounded by the scalar curvature of the Riemannian manifold.
4
D. Hypergraph Curvature
Let a hypergraph H = (V, E) with V = {1, 2, . . . , N } and a hyperedge E = {1, 2, . . . , n} be given. Inspired by coarse
scalar curvature, we wish to define the curvature for each hyperedge E of H. To this goal, we first need to define random
walk over hypergraphs. It is natural to define the (uniform) random walk started at vertex i ∈ V on H as the following: for
each j ∈ V
X
1
1
,
(13)
mi (j) =
di d(E) − 1
E∈E:(i,j)∈E
and hence we associate E with n probability measures
mi ∈ P(V ), 1 ≤ i ≤ n. Replacing Xn with {1, 2, . . . , n} in
P
Definition 2 and defining c(x1 , . . . , xn ) = minz∈V ni=1 d(xi , z), we can define a multi-marginal optimal transport problem
associated with E ∈ E as
X
W (E) :=
min
c(xn )π(xn ),
(14)
π∈Π(m1 ,m2 ,...,mn )
xn ∈V n
n
where x = (x1 , x2 , . . . , xn ). It is worth mentioning that c(1, 2, . . . , n) = n − 1 as {1, 2, . . . , n} is a hyperedge. In general,
c(x1 , . . . , xn ) ≥ n − 1 for distinct x1 , . . . , xn .
Definition 3 (Hypergraph Curvature). The curvature of a hyperedge E = {1, 2, . . . , n} ∈ E is defined as
W (E)
.
n−1
Note that if xi ∈
/ N (i), then π(x1 , . . . , xi , . . . , xn ) = 0 for all π ∈ Π(m1 , . . . , mn ). Hence, we have either c(x1 , . . . , xn ) ≤
3(n − 1) or π(x1 , . . . , xn ) = 0 for all π ∈ Π(m1 , . . . , mn ). This then demonstrates that, similar to graph curvature, −2 ≤
κ(E) ≤ 1.
Specializing Propositions 1 and 2 to the hypergraph setting, we now provide two equivalent formulations for hypergraph
curvature.
Barycenter: Although the minimization problem in (14) is a linear program, its complexity is exponential in N . However, it
turns out [26] that Wasserstein barycenter problem (11) can be solved quite efficiently. The equivalence between barycenter
problem and multi-marginal optimal transport problem justifies to define the Wasserstein barycenter of hyperedge E as
κ(E) := 1 −
bar(E) :=
inf
ν∈P(V )
Following mutatis mutandis Proposition 1, we have
n
X
W (mi , ν).
(15)
i=1
bar(E)
.
(16)
n−1
The term P
barycenter makes sense by recalling that for the Euclidean space the barycenter of points {xi }ni=1 is given by
n
arg minx i=1 kxi − xk2 . Consequently, bar(E) is the barycenter of points {mi }ni=1 in the Wasserstein space (i.e., a
metric space with Wasserstein distance).
Duality Following mutatis mutandis Proposition 2, we can write the following dual formula for W (E)
κ(E) = 1 −
κ(E) = 1 −
n
X
1
Emi [fi (X)],
sup
n − 1 (f1 ,...,fn )∈K i=1
(17)
where Eν [·] is the expectation operator with respect to measure ν and K is the set of all integrable real-valued functions
on V such that for any vector (x1 , . . . , xn ) ∈ V n
n
X
fk (xk ) ≤ c(x1 , . . . , xn ).
(18)
k=1
After defining hypergraph curvature, a natural question is whether or not this definition reduces to graph Ricci curvature (4) if
the hypergraph is indeed a simple graph, i.e., the cardinality of each hyperedge is two. We answer this question in affirmative by
invoking the barycenter interpretation. If H is in fact a graph and E is a hyperedge (i.e., n = 2 and thus E = {1, 2}), then bar(E)
equals either m1 or m2 , because for any ν ∈ P(V ) the triangle inequality implies W (m1 , ν) + W (m2 , ν) ≥ W (m1 , m2 ).
Thus, hypergraph curvature coincides with (4).
It is a well-known fact that Ollivier-Ricci curvature of edge (x, y) depends heavily on the number of common neighbors of
x and y. Specifically, if N (x) ∩ N (y) = ∅, then κ(x, y) ≤ 0 (see e.g., [14]). We now prove a similar result for hypergraphs.
Theorem 1. For any hyperedge E with cardinality m, we have
X |N (υ) ∩ N (ϑ) ∩ F |
.
W (E) ≥ 1 − min
υ,ϑ∈E
(d(F ) − 1)dυ
F ∈E
5
Proof. Let again E be denoted by {1, 2, . . . , n}. The proof relies on the dual formula of W (E). In order to make use of (17),
we need to find a set of functions {fj }nj=1 that satisfy the constraint (18). Fix two vertices υ and ϑ in E. Set fi ≡ 0 for
i ∈ E\{ϑ, υ} and suppose
fυ (r) + fϑ (s) ≤ d(r, s),
(19)
for all pairs of vertices (r, s). Then we have for any vector (x1 , . . . , xn ) ∈ V n
n
X
fi (xi ) =
fυ (xυ ) + fϑ (xϑ )
i=1
≤
d(xυ , xϑ ) = min [d(xυ , z) + d(xϑ , z)]
≤
c(x1 , . . . , xn ).
z∈V
Hence, the constraints fi ≡ 0 for i ∈ E\{ϑ, υ} and (19) are sufficient to satisfy (18).
Letting Cυ,ϑ denote the set of real-valued functions satisfying (19), we can write
X
X
fυ (r)mυ (r) +
W (E) ≥
sup
fϑ (r)mϑ (r)
fυ ,fϑ ∈Cυ,ϑ r∈V
=
sup
X
f ∈A r∈V
r∈V
f (r) [mυ (r) − mϑ (r)] ,
where the last equality is due to [27, Theorem 1.14] and A is the set of all real-valued 1-Lipschitz functions f on V , that is
A := {f : V → R : |f (r) − f (s)| ≤ d(r, s)}.
Now consider the following function
f (r) =
(
2, if r ∈ N (υ)\N (ϑ),
1, otherwise.
Clearly, f ∈ A and therefore,
W (E)
≥
X
f (r) [mυ (r) − mϑ (r)]
r∈V
=
X
2
mυ (r) +
=
mυ (r) − 1
r∈N (υ)∩N (ϑ)
r∈N (υ)\N (ϑ)
=
X
1 − mυ (N (υ) ∩ N (ϑ))
X |N (υ) ∩ N (ϑ) ∩ F |
.
1−
(d(F ) − 1)dυ
F ∈E
In light of this theorem, we have
X |N (υ) ∩ N (ϑ) ∩ F |
.
υ,ϑ∈E
(d(F ) − 1)dυ
κ(E) ≤ min
(20)
F ∈E
If H happens to be a simple graph, then for every edge E = (υ, ϑ), we have mυ (N (υ) ∩ N (ϑ)) =
of triangles supported on edge E. Hence, the bound (20) implies
κ(υ, ϑ) ≤
∆
,
max{dυ , dϑ }
∆
dυ ,
where ∆ is the number
(21)
which appears in [14, Theorem 4].
Remark 1 (Curvature as a projection of Boltzmann). Cuturi [26] introduced an entropic regularized version of Wasserstein
distance between two measures as
X
π(r, s)d(r, s) − εH(π)
Wε (µ, ν) :=
inf
π∈Π(µ,ν)
=
ε
inf
(r,s)∈V ×V
π∈Π(µ,ν)
D(πkKε ) − ε log Z(ε),
(22)
where ε > 0 is the regularization parameter, H(·) denotes the Shannon entropy function, Kε is the Boltzmann distribution
P
−d(x,y)/ε
and Z(ε) := (x,y)∈V ×V e−d(x,y)/ε . It follows that W (µ, ν) = limε↓0 Wε (µ, ν) [?]. As
defined as Kε (x, y) := e Z(ε)
mentioned in [26], despite the theoretically-guaranteed convergence, the procedure cannot work beyond a graph-dependent
value ε0 beyond which some entries of Kε are represented as zeroes in memory. Since H(·) is a strictly concave function, the
6
optimization problem in (22) has a unique solution π ∗ which corresponds to the projection of Kε onto Π(µ, ν). Following this
spirit, we can define the entropic regularized Wasserstein barycenter problem as
:=
Wε (E)
=
n
inf
ν∈P(V )
n
X
Wε (mi , ν)
i=1
ε inf D PkKε⊗n − nε log Z(ε)
P∈C
where C := {π1 ⊗ π2 ⊗ · · · ⊗ πn ∈ P (V × V ) : πi 1 = mi , and ∃η ∈ P(V ) s.t. πiT 1 = η}. Hence, Wε (E) is the
unique projection of Kε⊗n onto C. Assuming ε is sufficiently small, it follows that W (m1 , . . . , mn ) is an approximation of
this projection.
III. C OMPLETE UNIFORM
HYPERGRAPHS
Graph Ricci curvature turns out to have a simple formula for complete graphs KN . In particular, it is shown [14] that
−2
κ(x, y) = N
N −1 for any edges (x, y) in KN . In this section, we show that complete uniform hypergraphs with N vertices have
the same curvature as KN .
n
Definition 4. A hypergraph HN
= (V, E) is called complete n-uniform for n ≤ |V | = N if E = [V ]n , where [V ]n is the
collection of all n-subsets of V .
−1
2
n
Notice that HN
= KN . Recall that V = {1, 2, . . . , N }. It follows that in HN
, we have d(j) = N
n−1 , j ∈ V and d(E) = n
−2
for each E ∈ E. Furthermore, any pair of (i, j) ∈ E is contained in N
n−2 many hyperedges. Therefore, according to (13),
n
the random walk started at j ∈ V of HN is
mj (r) =
1{r6=j}
N −1
∀r ∈ V.
(23)
n
Lemma 1. For every hyperedge E of HN
, we have
W (E) =
n−1
.
N −1
In particular,
N −2
.
N −1
In light of this lemma, the curvature of complete n-uniform hypergraphs is independent of n. This is, in fact, a result of the
normalization in Definition 3.
κ(E) =
n
Proof. Again let E = {1, 2, . . . , n}. Notice that for HM
, we have d(r, s) = 1 for any distinct pair of vertices r, s ∈ V . Thus,
we can write for a measure ν ∈ P(V )
n
X
W (m1 , ν) =
i=1
=
min
π (1) ∈Π(m1 ,ν)
min
π (1) ∈Π(m1 ,ν)
n
=
N X
N
X
d(r, s)π (1) (r, s) + · · · +
r=1 s=1
N
X
"
1−
π
(1)
#
(r, r) + · · · +
r=1
1X
kmi − νk,
2 i=1
min
π (n) ∈Π(mn ,ν)
min
π (n) ∈Π(mn ,ν)
"
N X
N
X
r=1 s=1
N
X
1−
π
d(r, s)π (n) (r, s)
(n)
1
n
Pn
i=1
mi . Recalling that mi (r) =
1{r6=i}
N −1 ,
we obtain
W (E) =
n−1
.
N −1
(r, r)
r=1
where the last equality follows from [27, Exercise 1.17] and for two vectors a and b, ka − bk :=
obtain
n
n
1X
1X
bar(E) = inf
kmi − νk =
kmi − m̄k,
ν∈P(V ) 2
2 i=1
i=1
where m̄ =
#
P
i
|a(i) − b(i)|. Thus, we
7
Fig. 1. A hyperpath with 42 vertices and 7 hyperedges. For the green hyperdge, we have m = 10 and β = 7. Note that, unlike path graph,
hyperpath might have cycles.
IV. E XAMPLES
In this section, we focus on the computation of hypergraph curvature in two examples to illustrate the differences and
similarities with the graph curvature. The first example is a natural generalization of infinite path Pn ; a simple graph that has
n vertices with n − 2 vertices of degree 2 and the other two of degree one.
Example 1. It is shown [28] that Pn is one of the few graphs (among graphs with girth at least 5) with constant zero curvature.
We now demonstrate in the following theorem that a similar statement does not hold for the hyperpath; a hypergraph whose
vertices have degree at most 2. For the ease of presentation, we assume that any two intersecting hyperedges have exactly one
common vertex; see Fig. 2.
Theorem 2. For any hyperedge E of a hyperpath described above, we have
W (E) ≤ β
3n − 4
2n − 3
+ (n − β − 1)
,
2(n − 1)
2(n − 1)
where n := d(E) ≥ 3 and β := |{v ∈ E : ∄R ∈ E, v ∈ R}| < n. In particular, the curvature of hyperedge E is bounded
from below by
−(n − β − 2)
κ(E) ≥
.
2(n − 1)
Proof Sketch. The proof relies on the simple observation that for the hyperedge E = {1, 2, . . . , n}, we have W (E) ≤
P
n
i=1 W (mi , ν) for any probability measure ν ∈ P(V ). In particular,
W (E) ≤
n
X
W (mi , m1 ).
(24)
i=2
Assuming that {1, 2, . . . , n − β} are vertices in E that are shared with other hyperedges and {n − β + 1, . . . , n} are isolated
vertices inside E, we can employ a tedious (yet standard) linear-programming argument to show that
W (m1 , mj ) =
and
W (m1 , mk ) =
3n − 4
,
2(n − 1)
2n − 3
,
2(n − 1)
∀j ∈ {2, . . . , n − β},
(25)
∀k ∈ {n − β + 1, . . . , n}.
(26)
Plugging (25) and (26) into (24), we obtain the result.
In light of this result, we have κ(E) > 0 if only one vertex of E is shared, i.e., β > n − 2. In other words, the leaves of a
hyperpath have positive curvature which is different from graph curvature in that in simple graphs (with girth at least 5) each
edge connecting to a leaf has zero curvature, see [28, Theorem 3.3].
Example 2. As a toy example, consider the hypergraph H = (V, E) with V = {1, 2, . . . , 13} and E = {E1 , E2 , E3 , E4 },
where E1 = {1, 2, 3}, E2 = {2, 4, . . . , 7}, E3 = {6, . . . , 11}, and E4 = {7, 11, 12, 13}, as illustrated in Fig. 2. Using (13),
we can compute the probability measures associated to each hyperedge. For instance, the random walk started at vertex 2 is
µ2 = [0.25, 0, 0.25, 0.125, 0.125, 0.125, 0.125, 0, . . ., 0]. Since there are only 13 vertices, we can solve the optimization problem
(14) (as a linear program) for each hyperedge. Solving this optimization problem, we obtain W (E1 ) = 1, W (E2 ) = 2.38,
W (E3 ) = 2.08, and W (E4 ) = 1.45. Consequently, it follows that κ(E1 ) = 0.5, κ(E2 ) = 0.4, κ(E3 ) = 0.58, and κ(E4 ) = 0.52.
Informally speaking, the hyperedge with the lowest curvature is the bridge connecting two components of hypergraphs. This
is similar to the intuition of graph Ricci curvature, as experimentally observed in [29], that edges with negative curvature are
locally shortcuts of two component of graph.
8
4
E2 5
2
3
E1
8
6
7
11
13 12
1
9 E3
10
E4
Fig. 2. The hypergraph studied in Example 2.
V. C OARSE S CALAR C URVATURE
IN A
R IEMANNIAN M ANIFOLD S ETTING
In this section, we establish a consistency result for the definition of coarse scalar curvature (9) under a Riemannian manifold
setting. We will assume the hypergraphs of interest stem from a natural geometric model encoding adjacency relations among
discrete points randomly sampled from a Riemannian manifold.
Definition 5 (ε-Neighborhood Hypergraphs). Let (M, dM ) be a complete metric space. For any ε > 0, an ε-neighborhood
hypergraph H = (V, E) supported on M consists of the following data:
• A finite set of vertices V = {vi ∈ M, 1 ≤ i ≤ N };
• A set of hyperedges E, where each E ∈ E is a subset of V satisfying
diam (E) := sup dM (u, v) < 2ε;
u,v∈E
Every subset U of vertices with diam (U ) < 2ε corresponds to a hyperedge EU ∈ E.
We will denote |E| for the number of vertices joined by hyperedge E; |V | and |E| will be used to denote the number of vertices
and hyperedges in H, respectively.
•
The definition of ε-neighborhood hypergraphs assumes that the spatial proximity of the vertices is faithfully reflected in the
connectivity: the diameter of every hyperedge — understood as a discrete subset E ⊂ V — is bounded from above by 2ε.
In the rest of this section, we will assume M is an orientable d-dimensional complete Riemannian manifold with finite
volume, and the distance function dM : M × M → [0, ∞) is induced from the Riemannian length structure. As often
encountered in the manifold learning setup (see e.g. [30] and [31]), we assume the vertices of H are sampled i.i.d. uniformly
on M with respect to the standard volume measure dvol/vol (M ). Our goal in this section is to establish conditions under
which the coarse scalar curvature (9) converges to the scalar curvature on M in the limit |V | → ∞ and ε → 0. We begin
our discussion by summarizing some geometric and statistical properties of the Riemannian medians that we will need in the
following two lemmas.
Lemma 2. Let M be an orientable d-dimensional complete Riemannian manifold with
• Finite volume vol (M ) < ∞;
• Finite, positive injectivity radius 0 < Inj (M ) < ∞;
• Uniformly bounded sectional curvature Kx (u, v) < D for D ∈ R and all x ∈ M , u, v ∈ Tx M .
Assume U = {x1 , · · · , xn } is a finite discrete subset of M contained in a geodesic ball B (x, ε) of sufficiently small radius
ε > 0 satisfying
n
o
(
min D−1/2 π/4, Inj (M ) /2
if D > 0
2ε <
(27)
Inj (M ) /2
otherwise
and U is not totally contained in any geodesic on M . Then there exists a unique x̄ ∈ M minimizing the moment function
M ∋ z 7→
n
X
dM (z, xi ) ∈ [0, ∞) ,
i=1
and x̄ lies in the smallest closed, geodesically convex subset of B (x, ε) containing U .
Proof. The bound on ε and the assumption that U is not totally contained in any geodesic on M together ensures the uniqueness
of x̄, according to [32, Theorem 3.1]; the existence can be found e.g. in [33] or [34]. The last assertion about the location of
x̄ is a consequence of [32, Proposition 2.4].
The minimizer x̄ is known as the Riemannian median. Note that the diameter bound 2ε only depends on the injectivity
radius when the Riemannian manifold has negative sectional curvatures; this is of particular interest to us since many real
world networks demonstrate “negatively curved” behaviors, see e.g., [29], [35], [36].
9
∞
Lemma 3. Assume M , ε > 0 satisfy the same assumptions as in Lemma 2. Let {Xn }n=1 be a sequence of i.i.d. random points
uniformly distributed in a geodesic ball B (x, ε) of radius ε > 0 centering at x ∈ M , and denote µn for the Riemannian
median of X1 , · · · , Xn for all n ∈ N. Then µn → x̄ a.s. as n → ∞, where x̄ is the unique Riemannian median of the uniform
distribution on B (x, ε).
Proof. The existence and uniqueness of µn and x̄ follows from Lemma 2; note here that x̄ need not coincide with x in
general, by the characterization of Riemannian medians established in [32, Proposition 2.1 and Theorem 2.2]. The almost sure
convergence has been established in [34, Corollary 4.1].
We now turn to investigating the coarse scalar curvature of hypergraphs generated from a geometric probabilistic model:
saturated ε-neighborhood hypergraphs supported on a Riemannian manifold M , with vertices uniformly distributed on M . We
consider the same random walk on the Riemannian manifold M as in [10], namely, the one given in (6): for any x ∈ M and
ε > 0,
1B(x,ε)
dmεx :=
dvol,
(28)
vol(B(x, ε))
which is essentially the standard volume measure on M restricted and renormalized on B (x, ε). For simplicity of statement,
let us denote x̂E for the Riemannian median — when it exists and is unique — of the vertices connected by a hyperedge E
in a saturated hypergraph G = (V, E).
∞
Theorem 3. Let M , ε > 0 be as assumed in Lemma 2. Let {vi }i=1 be a sequence of i.i.d. random points sampled uniformly
on M with respect to the standard Riemannian volume measure. For any N ∈ N, let VN := {vi }N
i=1 and HN = (VN , EN ) be
an ε-neighborhood hypergraph supported on M . If there exists a hyperedge EN ∈ EN for each Hn such that
• EN $ EN +1 for all sufficiently large N ∈ N
• there exists x̄ ∈ M such that x̂EN → x̄ as n → ∞
• EN ⊂ B (x̄, ε) for all sufficiently large N ∈ N
then the coarse scalar curvature of hyperedge EN satisfies, for all sufficiently small ε > 0,
lim sup κ (EN ) ≥
N →∞
ε2
Scal (x̄) + O ε3 .
2 (d + 2)
(29)
Proof. Since curvature is a local quantity, we may assume without loss of generality that M is connected and even compact. Let
ε > 0 be sufficiently small such that the geodesic ball B (x̄, ε) is geodesically convex neighborhood of x̄ (c.f. [23, Proposition
4.2]). To ease notations, write ℓN := |EN | and denote x1 , · · · , xℓN ∈ VN for the vertices of HN connected by the hyperedge
EN ∈ E. By Proposition 1 and the definition of Riemannian median, we have
Z
dvol (y)
W1 (my , mx̄ )
P ℓN
vol (B (x̄, ε))
W1 (mxi , mx̄ ) ℓN →∞ B(x̄,ε)
W1 (EN )
(30)
−−−−→ Z
≤ Pi=1
ℓ
N
dvol (y)
c (x1 , · · · , xℓn )
i=1 dM (xi , x̂EN )
dM (y, x̄)
vol (B (x̄, ε))
B(x̄,ε)
where the limit follows from the law of large number and the Lipschitz continuity of the function (see e.g. [32, §2])
Z
dM (x, y) dvol (y) .
x 7→
B(x,ε)
We know from [10, Example 7] that, for any y ∈ B (x̄, ε),
ε2
Ric (vx̄,y , vx̄,y ) + O ε3 dM (x̄, y)
W1 (my , mx̄ ) = 1 −
2 (d + 2)
where vx̄,y is a unit tangent vector in Tx̄ M such that
expx̄ (dM (x̄, y) vx̄,y ) = y.
It follows that
Z
ε2
W1 (my , mx̄ ) dvol (y) =
dM (x̄, y) dvol (y) −
2 (d + 2)
B(x̄,ε)
B(x̄,ε)
!
Z
Z
+ O ε3
dM (x̄, y) dvol (y)
B(x̄,ε)
Z
Ric (vx̄,y , vx̄,y ) dM (x̄, y) dvol (y)
B(x̄,ε)
10
which, when plugged back into (30), gives
ε2
W1 (EN )
≤ 1 + O ε3 −
lim sup
2 (d + 2)
N →∞ c (x1 , · · · , xℓN )
dvol (y)
Ric (vx̄,y , vx̄,y ) dM (x̄, y)
vol
(B (x̄, ε))
B(x̄,ε)
Z
.
dvol (y)
dM (x̄, y)
vol (B (x̄, ε))
B(x̄,ε)
Z
A straightforward calculation using geodesic normal coordinates reveals
Z
Scal (x)
d
ε2
dvolM (y)
=ε
−
+ O ε3
dM (x, y)
vol (B (x, ε))
d+1
3 (d + 1) (d + 2) (d + 3)
B(x,ε)
(31)
(32)
and
Z
#
!
"
2
dvolM (y)
|Scal (x)|
ε2
Scal (x)
2
3
Ric (vx,y , vx,y ) dM (x, y)
(33)
=ε
+
− kR (x)k +O ε
vol (B (x, ε))
d+1
3 (d + 2) (d + 3)
d+1
B(x,ε)
Pd
2
where kR (x)k = i=1 |Ric (ei , ei )| for an arbitrary orthonormal basis e1 , · · · , ed of Tx̄ M . Plugging (32) and (33) back into
(31) to conclude that
W1 (EN )
ε2
lim sup
(34)
≤1−
Scal (x̄) + O ε3 .
2 (d + 2)
ℓN →∞ c (x1 , · · · , xℓN )
Theorem 3 indicates that coarse scalar curvature asymptotically upper bounds the scalar curvature of the Riemannian manifold,
when the hypergraph is constructed from uniformly sampling the manifold in a natural way. This justifies the nomenclature
of "scalar curvature" in our definition. We conjecture that coarse scalar curvature also asymptotically lower bounds the scalar
curvature in the same setting, but will have to leave that for future work.
VI. C ONCLUSION
AND FUTURE WORK
In this paper, we propose a novel definition of curvature for hypergraphs by generalizing coarse Ricci curvature to coarse
scalar curvature through multi-marginal optimal transport. Our definition is shown to be consistent with graph curvature in
that (i) it reduces to graph curvature if the hypergraph of interest is indeed a simple graph, and (ii) it shares several properties
with graph curvature. In particular, it is experimentally observed that, analogous to graph curvature, hypergraph curvature
can be used to determine the bridge between components in the network. We are currently computing hypergraph curvature
in real-world hypergraph networks (in particular co-authorship or cellular networks) to observe this centrality property of
hypergraph curvature. We are also applying hypergraph curvature to characterize dynamic effects in large dynamic network
(in particular, financial network). Intuitively, hypergraph curvature provides a computational method for detecting changes in
dynamic networks, characterizing fast evolving network components, as well as identifying stable network region. On the
theoretical side, we are interested in gaining better understandings of our hypergraph curvature with deeper insights from
differential geometry.
R EFERENCES
[1] E. Demir, C. Aykanat, and B. B. Cambazoglu, “Clustering spatial networks for aggregate query processing: A hypergraph approach,” Information
Systems, vol. 33, no. 1, pp. 1–17, 2008. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S030643790700021X
[2] A. Chasapi, C. Kotropoulos, and K. Pliakos, “Adaptive algorithms for hypergraph learning,” in IEEE Int. Conf. Acoustics, Speech and Signal Processing
(ICASSP), March 2016, pp. 1179–1183.
[3] Y. Fang and Y. Zheng, “Metric learning based on attribute hypergraph,” in IEEE Int. Conf. Image Processing (ICIP), Sept 2017, pp. 3440–3444.
[4] S. Klamt, H. U-U, and F. Theis, “Hypergraphs and Cellular Networks,” PLoS Comput Biol, vol. 5, no. 5, p. e1000385, May 2009.
[5] P. Ji and J. Jin, “Coauthorship and citation networks for statisticians,” The Annals of Applied Statistics, vol. 10, no. 4, pp. 1779–1812, 12 2016.
[6] T.-H. H. Chan, Z. G. Tang, and C. Zhang, “Spectral Properties of Laplacian and Stochastic Diffusion Process for Edge Expansion in Hypergraphs,”
[Preprint] arXiv:1510.01520v1, 2017.
[7] R. Shaydulin, J. Chen, and I. Safro, “Relaxation-Based Coarsening for Multilevel Hypergraph Partitioning,” [Preprint] arXiv:1710.06552v2, 2017.
[8] A. Louis, “Hypergraph Markov Operators, Eigenvalues and Approximation Algorithms,” [Preprint] arXiv:1408.2425v2, 2016.
[9] A. Banerjee, “On the spectrum of hypergraphs,” [Preprint] arXiv:1711.09356v2, 2017.
[10] Y. Ollivier, “Ricci curvature of Markov chains on metric spaces,” Journal of Functional Analysis, vol. 256, no. 3, pp. 810–864, 2009.
[11] Y. Lin, L. Lu, and S.-T. Yau, “Ricci Curvature of Graphs,” Tohoku Mathematical Journal, vol. 63, no. 4, pp. 605–627, 2011.
[12] E. Saucan, A. Samal, M. Weber, and J. Jost, “Discrete curvatures and network analysis,” Max Planck Institute for Mathematics in the Sciences,
Preprint, 2017. [Online]. Available: http://www.mis.mpg.de/de/publications/preprints/2017/prepr2017-34.html
[13] F. Bauer, J. Jost, and S. Liu, “Ollivier-Ricci curvature and the spectrum of the normalized graph Laplace operator,” Mathematical Research Letters,
vol. 19, pp. 1185–1205, 11 2012.
[14] J. Jost and S. Liu, “Ollivier’s Ricci curvature, local clustering and curvature-dimension inequalities on graphs,” Discrete and Computational Geometry,
vol. 51, no. 2, pp. 300–322, Mar 2014.
[15] H. J. Cho and S. H. Paeng, “Ollivier’s Ricci curvature and the coloring of graphs,” European Journal of Combinatorics, vol. 34, no. 5, pp. 916–922,
2013.
11
[16] C. C. Ni, Y. Y. Lin, J. Gao, X. D. Gu, and E. Saucan, “Ricci curvature of the Internet topology,” in IEEE Conference on Computer Communications
(INFOCOM), April 2015, pp. 2758–2766.
[17] R. Sandhu, T. Georgiou, and A. Tannenbaum, “Ricci curvature: An economic indicator for market fragility and systemic risk,” Science Advances, vol. 2,
no. 5, 2016.
[18] H. Farooq, Y. Chen, T. Georgiou, A. Tannenbaum, and C. Lenglet, “Network Curvature as a Hallmark of Brain Structural Connectivity,” bioRxiv, 2017.
[19] R. Sandhu, T. Georgiou, E. Reznik, L. Zhu, I. Kolesov, Y. Senbabaoglu, and A. Tannenbaum, “Graph curvature for differentiating cancer networks,”
Scientific Reports, vol. 5, 7 2015.
[20] G. Carlier and I. Ekeland, “Matching for Teams,” Economic Theory, vol. 42, no. 2, pp. 397–418, 2010.
[21] Y. H. Kim and Pass, B., “Multi-marginal Optimal Transport on Riemannian Manifolds,” American Journal of Mathematics, vol. 137, no. 4, pp. 1045–1060,
2017.
[22] W. Gangbo and A. Świȩch, “Optimal maps for the multidimensional Monge-Kantorovich problem,” Communications on Pure and Applied Mathematics,
vol. 51, no. 1, pp. 23–45, 1998.
[23] M. P. Do Carmo, Riemannian Geometry. Springer, 1992.
[24] M. Agueh and G. Carlier, “Barycenters in the Wasserstein Space,” SIAM Journal on Mathematical Analysis, vol. 43, no. 2, pp. 904–924, 2011.
[25] H. G. Kellerer, “Duality Theorems for Marginal Problems,” Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete, vol. 67, no. 4, pp. 399–432,
1984.
[26] M. Cuturi, “Sinkhorn Distances: Lightspeed Computation of Optimal Transport,” in Advances in Neural Information Processing Systems, 2013, pp.
2292–2300.
[27] C. Villani, Topics in Optimal Transportation. American Mathematical Society, 2003.
[28] B. B. Bhattacharya and S. Mukherjee, “Exact and asymptotic results on coarse ricci curvature of graphs,” Discrete Mathematics, vol. 338, no. 1, pp. 23
– 42, 2015.
[29] C.-C. Ni, Y.-Y. Lin, J. Gao, X. D. Gu, and E. Saucan, “Ricci Curvature of the Internet Topology,” ArXiv e-prints, Jan 2015. [Online]. Available:
arXiv:1501.04138v1
[30] M. Belkin and P. Niyogi, “Laplacian Eigenmaps for Dimensionality Reduction and Data Representation,” Neural Comput., vol. 15, no. 6, pp.
1373–1396, Jun. 2003. [Online]. Available: http://dx.doi.org/10.1162/089976603321780317
[31] R. R. Coifman and S. Lafon, “Diffusion Maps,” Applied and Computational Harmonic Analysis, vol. 21, no. 1, pp. 5–30, 2006, special Issue: Diffusion
Maps and Wavelets. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S1063520306000546
[32] L. Yang, “Riemannian Median and its Estimation,” LMS Journal of Computation and Mathematics, vol. 13, pp. 461–479, 2010.
[33] P. T. Fletcher, S. Venkatasubramanian, and S. Joshi, “The Geometric Median on Riemannian Manifolds with Application to Robust Atlas Estimation,”
NeuroImage, vol. 45, no. 1, pp. S143–S152, 2009.
[34] M. Arnaudon, F. Barbaresco, and L. Yang, “Medians and Means in Riemannian Geometry: Existence, Uniqueness and Computation,” in Matrix Information
Geometry. Springer, 2013, pp. 169–197.
[35] O. Narayan and I. Saniee, “Large-scale curvature of networks,” Phys. Rev. E, vol. 84, p. 066108, Dec. 2011.
[36] R. Albert, B. DasGupta, and N. Mobasheri, “Topological implications of negative curvature for biological and social networks,” Phys. Rev. E, vol. 89,
p. 032811, Mar 2014.
| 7cs.IT
|
SINGULARITIES OF GENERIC LINKAGE OF ALGEBRAIC VARIETIES
arXiv:1207.1082v2 [math.AG] 18 Jun 2013
WENBO NIU
Abstract. Let Y be a generic link of a subvariety X of a nonsingular variety A. We give
a description of the Grauert-Riemenschneider canonical sheaf of Y in terms of the multiplier
ideal sheaves associated to X and use it to study the singularities of Y . As the first application,
we give a criterion when Y has rational singularities and show that log canonical threshold
increases and log canonical pairs are preserved in generic linkage. As another application
we give a quick and simple liaison method to generalize the results of de Fernex-Ein and
Chardin-Ulrich on the Castelnuovo-Mumford regularity bound for a projective variety.
1. Introduction
Let A be either a nonsingular affine variety or a projective space Pn over the field of complex
number C. Two closed subvarieties of A are said to be geometrically linked if the union of
them is a complete intersection in A. Let us fix a reduced equidimensional subscheme X of
A of codimension c for now. By choosing c equations carefully among the defining equations
of X we define a complete intersection V and obtain a subscheme Y as the closure of the
complement of X in V . Then Y is geometrically linked to X via V , i.e., Y ∪ X = V and Y
has no common irreducible components with X. If such complete intersection V is chosen to
be as general as possible, then Y is called a generic link of X.
The study of linkage, or the theory of liaison, of algebraic varieties can be traced back to
hundreds of years ago. The recent work in this area was initiated by Peskine and Szpiro [PS74].
After that, linkage has attracted considerable attention and has been developed widely and
deeply from both geometric and algebraic point of views.
Let us state the construction of a generic link more concretely to make our introduction
clear. Assume further that X is defined by an ideal IX generated by equations f1 , · · · , ft . The
essential point is to create a general complete intersection V by using the equations fi ’s. There
are three situations in current research to do this, which we list as follows.
Situation A. Suppose A = Spec R is affine. A complete intersection V is defined by the
equations αi = Ui,1 f1 + · · · + Ui,t ft , for i = 1, · · · , c, where Ui,j ’s are variables. V is actually
defined in an extended space Spec R[Ui,j ]. By abuse of notation, we still write A = Spec R[Ui,j ]
and X is defined by IX [Uij ]. This only occurs in the introduction. New notation which avoids
any ambiguity will be introduced in Definition 2.3 and adopted henceforth.
Situation B. Suppose A = Spec R is affine. A complete intersection V is defined by the
equations αi = ai,1 f1 + · · · + ai,t ft , for i = 1, · · · , c, where ai,j ’s are general scalars in C.
Situation C. Suppose A = Pn and the homogeneous equations fi ’s of X have degrees d1 ≥
d2 ≥ · · · ≥ dt . A complete intersection V is defined by choosing general equations αi from
H 0 (Pn , IX (di )), for i = 1, · · · , c.
2010 Mathematics Subject Classification. 13C40, 14M06.
Key words and phrases. Generic linkage, singularity, log canonical pair, multiplier ideal sheaf, CastelnuovoMumford regularity.
1
In each of the three situations, once having a complete intersection V in hand, a generic link
Y of X can be obtained by algebraic construction, namely, by an ideal IY := (IV : IX ). (We
sometimes use sheaf notation, e.g. IX , to view an ideal as a sheaf.)
The terminology of generic linkage usually refers to Situation A. The theory based on this
setting has been founded and developed deeply by Huneke and Ulrich in the last three decades,
mainly from algebraic side of the story (e.g. [HU88], [HU87]). Situation C was studied from the
geometric side. The book of Miglior [Mig98] gives a outline of the theory along this direction.
It is worth mentioning that the way to construct a generic link in Situation C is quite classical
and a typical application can be found in the work of Betram, Ein and Lazarsfeld [BEL91] on
the Castelnuovo-Mumford regularity bound for a smooth projective variety. Situation B can
be considered as a specialization of Situation A as well as a local version of Situation C. As
an important technique it has been used in recent studies of singularities, for instance, in the
work of de Fernex and Docampo [dFD12] and the work of Ein and Mustata [EM09].
However, in contrast to the quick and deep development of singularity theories in the past
decades, much less has been known about the singularities in generic linkage. A few special
examples have drawn attention recently. One important case, which serves as a guideline of
this paper, is a result of Chardin and Ulrich [CU02] that if X is a local complete intersection
and has rational singularities then a generic link of X has rational singularities. But the machinery behind the result seems quite mysterious from the geometric point of view. Meanwhile,
experience gained from research in the past gives the intuition that the singularities of a generic
link seem to be worse than X itself.
The main theorem of this paper in the context of Situations A, B and C is the following.
Recall that A is either a nonsingular affine variety or a projective space and X is a reduced
equidimensional subscheme of codimension c. We simply write them as a pair (A, cX).
Theorem 1.1. Let Y be a generic link (if it is nonempty) to a pair (A, cX) via V . Then Y
is reduced equidimensional of codimension c, its Grauert-Riemenschneider canonical sheaf is
ωYGR ≃ I (A, cX) · OY ⊗ ωV ,
where I (A, cX) is the multiplier ideal sheaf associated to the pair (A, cX) and ωV is the
dualizing sheaf of V , and it fits into the following commutative diagram
(1.1.1)
ωYGR
≃
_
// I (A, cX) · OY ⊗ ωV
_
tr
i
ωY
≃
// IX · OY ⊗ ωV
where the bottom isomorphism is canonical, tr is the trace map, and the inclusion i is induced
by I (A, cX) ֒→ IX (cf. Lemma 2.5). Furthermore, we have inequalities for log canonical
thresholds of pairs as
lct(A, Y ) ≥ lct(A, X).
It should be noticed that in the application of the theorem the dualizing sheaf ωV can be
replaced by ωA |V in Situations A and B and by ωPn (d1 + · · · + dc )|V in Situation C. Also the
generic link Y could be reducible in Situations B and C. There are some special case in which
Y is irreducible discussed in Section 4.
The aforementioned result of Chardin and Ulrich is an immediate consequence of the theorem. Also the problem that when a generic link has rational singularities now becomes clear.
Furthermore the theorem also implies that log canonical pairs are preserved in generic linkage.
It should be mentioned that, as pointed out by the referee, the inequality of log canonical
thresholds in the theorem is rather straightforward in Situations B and C.
2
As another application of the theorem in Situation C of projective spaces, we can generalize
and extend the results of de Fernex and Ein [dFE10] and Chardin and Ulrich [CU02] on the
Castelnuovo-Mumford regularity bound for a projective variety. The proof is based on liaison
theory and provides a natural, geometric approach to study regularity. The original proofs in
[CU02] and [dFE10] were very different, and the new proof here finds its way somewhere in the
middle, leading to a statement which recovers both results. This was kindly suggested by the
referee. The following is a simple version and we will give its full general version in Corollary
4.3.
Corollary 1.2. Let X ⊂ Pn be a reduced equidimensional subscheme of codimension c defined
by the equations of degree d1 ≥ d2 ≥ · · · ≥ dt and assume that the pair (Pn , cX) is log canonical
except possibly at finitely many points. Then
reg X ≤
c
X
di − c + 1
i=1
and equality holds if and only if X is a complete intersection in Pn .
We conclude this introduction by briefly stating the organization of this paper. We take Situation A as our framework, following the footprints of Huneke and Ulrich. We prove Theorem
1.1 in Situations A and B in Section 3. Section 4 is devoted to the case of projective varieties
and the Castelnuovo-Mumford regularity bound.
Acknowledgement. Special thanks are due to professor Bernd Ulrich who introduced the
author to this subject and spent his valuable time on discussion. This paper would not be
possible without his generous help and encouragement. The author also would like to thank
professor Lawrence Ein for his insightful knowledge and inspiring suggestions which enrich the
paper. The author’s thanks also goes to professor Joseph Lipman for his patient reading and
kind suggestions, and goes to the referee for his nice comments and suggestions which improve
the quality of the paper.
2. Generic linkage, singularities and multiplier ideal sheaves
Throughout this paper, we work over the field k := C. By a variety we mean a reduced
irreducible scheme of finite type over k. A subscheme is always assumed to be closed. We
shall briefly review basic facts about linkage, singularities and multiplier ideal sheaves in this
section.
Definition 2.1. Let A be either a nonsingular affine variety over k or a projective space Pnk .
Let X and Y be two subschemes of A. We say that X and Y are geometrically linked if X ∪ Y
is a complete intersection V in A and X and Y are equidimensional, no embedded components
and have no common irreducible components. We also say Y is linked to X via V , or vice
versa.
Suppose Y is geometrically linked to X via V in A as defined above. Let ωY be a dualizing
sheaf of Y (in this paper, we use dualizing sheaf and canonical sheaf interchangeably). One
important fact is an isomorphism ωY ≃ IX · OV ⊗ ωV , where IX is the defining ideal sheaf
of X in A and ωV is a dualizing sheaf of V . If A is an affine nonsingular variety, then one can
deduce that
(2.1.1)
ωY ≃ IX · OY ⊗ ωA ,
3
where ωA is a dualizing sheaf of A. If A is a projective space Pnk and V is cut out by homogeneous equations of degrees d1 ≥ d2 ≥ · · · ≥ dc , where c = codim V , then one has
(2.1.2)
ωY ≃ IX · OY ⊗ ωA (d1 + · · · + dc ).
It is also well-know that Y is Cohen-Macaulay if and only if X is Cohen-Macaulay. The modern
approach to study linkage goes back to Peskine and Szpiro [PS74], and we refer to their work
for more general theory of linkage.
Definition 2.1 suggests that the study of linkage involves an ambient space containing the
varieties concerned. Thus it is natural to consider a variety and its ambient space as a pair.
Besides, the concept of pairs is a successful approach to study singularities of varieties, which
is the main motivation of this paper.
Definition 2.2. A pair (A, cX) contains a nonsingular variety A over k, a subscheme X and
a nonnegative real number c. If A = Spec R is affine and X is defined by an ideal IX , then we
c ).
say the pair (A, cX) is an affine pair and also use an alternative notation (R, IX
Now we give the definition of generic linkage mentioned in Situation A and take it as our
framework in Section 3.
c ) be an affine pair such that X is reduced equidiDefinition 2.3. Let (Ak , cXk ) = (Rk , IX
k
k
mensional and c = codimAk Xk . We construct a generic link of Xk as follows. Fix a generating
set (f1 , · · · , ft ) of IXk . Let (Uij ), 1 ≤ i ≤ c, 1 ≤ j ≤ t, be a c × t matrix of variables. Set
R := Rk [Uij ] and IX := IXk · Rk [Uij ] and define A = Spec R and X = Spec R/IX . Notice that
IX is still generated by (f1 , · · · , ft ) in R. We define a complete intersection V inside A by an
ideal
IV := (α1 , · · · , αc ) = (Ui,j ) · (f1 , · · · , ft )T ,
that is
αi := Ui,1 f1 + Ui,2 f2 + · · · + Ui,t ft ,
for 1 ≤ i ≤ c.
Then a generic link Y to Xk via V is defined by an ideal IY := (IV : IX ).
Remark 2.4. In the definition above, the subscheme V is a complete intersection in A [Hoc73].
The subschemes Y and X of A are geometrically linked so that IY = (IV : IX ), IX = (IV : IY )
and IV = IX ∩ IY and therefore Y is equidimensional without embedded components and has
no common components with X [HU85, 2.1, 2.5]. Furthermore Y is actually integral [HU85,
2.6]. In the sequel we only need the fact that Y is reduced equidimensional. It is from (2.1.1)
that a dualizing sheaf ωY of Y is ωY ≃ IX · OY ⊗ ωA , where ωA is a dualizing sheaf of A.
Let A be a nonsingular variety over k and {Xi }, i = 1, · · · , m, are m subschemes of A. By
Hironaka’s resolution of singularities, there is a birational
projective morphism f : A′ −→ A
P
′
−1
such that A is a nonsingular variety, f (Xi ) =
ai,j Ej for i = 1, · · · , m, where ai,j ’s are
nonnegative integers and Ej ’s are prime divisors of A′ such that the union of Ej ’s with the
exceptional locusP
Exc(f ) is a simple normal crossing divisor. The morphism f is called a log
resolution of (A, i Xi ).
′
A of (A, X) such that f −1 (X) =
PsLet (A, cX) be a pair. Take a log resolution f : AP−→
s
i=1 ai Ei and the relative canonical divisor KA′ /A =
i=1 ki Ei . We say that the pair (A, cX)
is log canonical if ki − c · ai ≥ −1 for all i. The log canonical threshold of (A, X) is defined to
be
ki + 1
}.
lct(A, X) := min{
i
ai
The multiplier ideal sheaf I (A, cX) associated to the pair (A, cX) is defined to be
X
ai Ei ⌋),
I (A, cX) := f∗ OA′ (KA′ /A − ⌊c
4
where ⌊c
P
ai Ei ⌋ is the round down of the Q-divisor c
P
ai Ei .
Lemma 2.5. Let A be a nonsingular variety over k and X ⊂ A be a reduced equidimensional
subscheme of codimension c defined by an ideal sheaf IX . Let I (A, cX) be the multiplier ideal
sheaf associated to the pair (A, cX). Then I (A, cX) ⊆ IX .
Proof. Since at each generic point p of X one has I (A, cX)p = IX,p and IX is radical the
result is then clear.
The following lemma is due to Lawrence Ein, which gives a criterion to compare multiplier
ideal sheaves with ideal sheaves.
Ein’s Lemma. Let A be a nonsingular variety and X ⊂ A be a reduced equidimensional
subscheme of codimension c defined by an ideal sheaf IX . Then I (A, cX) = IX if and only
if I (A, (c−1)X) = OA . In particular if the pair (A, cX) is log canonical then I (A, cX) = IX .
Proof. The inclusion I (A, cX) ⊆ IX is from Lemma 2.5. It then suffices to show that IX ⊆
I (A, cX) if and only if I (A, (c − 1)X) is trivial. If c = 1, then there is nothing to prove. So
in the sequel we assume c > 1. We shall follow notation and terminologies in [EM09, Section
7].
The inclusion IX ⊆ I (A, cX) is true if and only if for any prime divisor E over X we
have an inequality ordE IX + ordE K /A − c · ordE IX ≥ 0. This is equivalent to the inequality
ordE K /A − (c − 1) · ordE IX ≥ 0, which is equivalent to I (A, (c − 1)X) is trivial.
Now suppose (A, cX) is log canonical. This means that for any prime divisor E over X
we have ordE K /A − c · ordE IX ≥ −1. If the center of E is outside X then ordE IX = 0
and ordE K /A ≥ 0 since A is nonsingular. And thus ordE K /A − c · ordE IX ≥ −ordE IX .
On the other hand, if the center of E is inside X, then ordE IX ≥ 1 and again we have
ordE K /A −c·ordE IX ≥ −1 ≥ −ordE IX . Hence in any case we have ordE K /A −c·ordE IX ≥
−ordE IX , which implies, as showed above, that I (A, (c − 1)X) is trivial.
Definition 2.6. Let A be a nonsingular variety over k and X a reduced subscheme of A. A
morphism ϕA : A −→ A is a factorizing resolution of X inside A if the following hold:
(1) ϕA is an isomorphism at the generic point of every irreducible component of X. In
particular, the strict transform X of X is defined.
(2) The morphism ϕA and ϕX := ϕA |X are resolution of singularities of A and X, respectively, and the union of X with the exceptional locus Exc(ϕA ) has simple normal
crossings.
(3) If IX and IX are the defining ideals of X and X in A and A, respectively, then there
exists an effective divisor G on A such that
IX · OA = IX · OA (−G).
The divisor G is supported on Exc(ϕA ) and hence has normal crossing with X.
Remark 2.7. The above definition is borrowed from [EIM11, Definition 2.10]. The existence
of a factorizing resolution is proved in [BVU03, Theorem 1.2]. In addition, we can assume that
the morphism ϕA is isomorphic over the open set A \ X.
Let us conclude this section by briefly reviewing the definition of rational singularities and
Grauert-Riemenschneider canonical sheaves. Let X be a reduced equidimensional scheme of
finite type over k. Let f : X ′ −→ X be a resolution of singularities of X. Then the GrauertGR of X is defined to be ω GR := f ω ′ , where ω ′ is the
Riemenschneider canonical sheaf ωX
∗ X
X
X
GR is independent on the choice of the
canonical sheaf of X ′ . It turns out that the sheaf ωX
5
GR is canonically a subsheaf of ω , a
resolution of singularities f (cf. [Laz04a]). Furthermore ωX
X
GR
dualizing sheaf of X, via a trace map tr : ωX ֒→ ωX . Recall that X has rational singularities
if f∗ OX ′ = OX and Ri f∗ OX ′ = 0 for i > 0. It is well-know that X has rational singularities if
GR = ω (cf. [Kol97]).
and only if X is Cohen-Macaulay and ωX
X
3. Generic linkage of affine varieties
In this section, we first study the singularities of generic linkages under the framework of
Huneke and Ulrich as Definition 2.3 in Situation A. In this case, Theorem 1.1 will be proved
by using Proposition 3.1, Proposition 3.2 and Proposition 3.7. Some direct consequences will
also be given which describe singularities of generic linkage. Then we specialize our results to
a Zariski open set of a scalar matrices space, which takes care of Situation B.
Proposition 3.1. With notation as in Definition 2.3 let Y be a generic link to a pair (Ak , cXk ).
Then
ωYGR ≃ I (A, cX) · OY ⊗ ωA ,
where I (A, cX) is the multiplier ideal sheaf associated to the pair (A, cX).
Proof. Let ϕk : Ak −→ Ak be a factorizing resolution of singularities of Xk inside Ak , so that
IXk · OAk = IX k · OAk (−Gk ) where X k is the strict transform of Xk , Gk is an effective divisor
supported on the exceptional locus of ϕk , and furthermore X k and the exceptional locus of ϕk
are simple normal crossings. The morphism ϕk can be assumed to be an isomorphism over the
open set Ak \Xk (cf. Definition 2.6 and Remark 2.7).
By tensoring k[Uij ] to the factorizing resolution ϕk , we obtain a factorizing resolution of
singularities of X inside A as
ϕ : A −→ A,
such that IX · OA = IX · OA (−G), where X is the strict transform of X, G is an effective
divisor supported on the exceptional locus of ϕ, and X and exceptional locus of ϕ are simple
normal crossings. Notice that by the construction, we actually have A = Ak ⊗k Spec k[Uij ],
X = X k ⊗k Spec k[Uij ] and G = Gk ⊗k Spec k[Uij ].
Claim 3.1.1.
The ideal sheaf IV · OA has a decomposition as
IV · OA = IV · OA (−G)
where IV is an ideal sheaf on A and is a local complete intersection.
Proof of Claim 3.1.1. The question is local. Recall that ϕk : Ak −→ Ak is the factorizing
resolution of singularities of Xk inside Ak . Let U k = Spec Rk be an affine open set of Ak such
that the effective divisor Gk is defined by an equation g ∈ Rk and we have a decomposition
IXk · Rk = IX k · (g) on U k . Now since IXk · Rk = (f1 , · · · , ft ) · Rk we can write fi = f i g where
f i ∈ Rk for i = 1, · · · , t so that IX k = (f 1 , · · · , f t ).
Recall that the factoring resolution ϕ : A −→ A is obtained by tensoring Spec k[Uij ] to the
factoring resolution ϕk . Write R = Rk ⊗ k[Uij ] which is a faithfully flat ring extension of Rk
and then U = U k ⊗ Spec k[Uij ] = Spec R is an affine open set of A. Notice that on U the
ideal IX = IX k · R and the effective divisor G is still generated by the equation g. Recall
that the ideal IV = (α1 , · · · , αc ), where αi = Ui,1 f1 + Ui,2 f2 + · · · + Ui,t ft . Thus if write
αi = Ui,1 f 1 + Ui,2 f 2 + · · · + Ui,t f t and set IV = (α1 , · · · , αc ), then IV is a complete intersection
on U and we have a decomposition IV ·R = IV ·(g) on U , which finishes the proof of Claim 3.1.1.
6
e −→ A be the blowing-up of A along X such that I · O e = O e(−T ), where
Now let µ : A
X
A
A
e −→ A. Notice that the supports of
T is an exceptional divisor of µ. Denote by ψ = (φ ◦ µ) : A
divisors T , µ∗ (G) and the exceptional locus of ψ are simple normal crossings. We write KA/A
e
to be the relative canonical divisor of the morphism ψ.
Claim 3.1.2.
We have the following statements.
(1) The ideal sheaf IV · OAe can be decomposed as
IV · OAe = IVe · OAe(−T ).
e and defines a local complete intersection Ve of codiwhere IVe is an ideal sheaf on A
mension c.
(2) The scheme Ve is nonsingular and irreducible and its dualizing sheaf is
− c(T + µ∗ G)) ⊗ ψ ∗ ωA ,
ωVe ≃ OVe (KA/A
e
where ωA is a dualizing sheaf of A.
(3) The scheme Ve is the strict transform of Y via ψ.
Proof of Claim 3.1.2. We work locally on affine open sets as in the proof of Claim 3.1.1.
Assume that Ak = Spec Rk and A = Spec R, where R = Rk ⊗k k[Uij ]. Recall that, as we
showed in the proof of Claim 3.1.1, the ideal IX = (f 1 , · · · , f t ) · R where for each i the
generator f i is in the ring Rk , and that the complete intersection IV = (α1 , · · · , αc ), where
e say
αi = Ui,1 f 1 + Ui,2 f 2 + · · · + Ui,t f t for i = 1, · · · , c. Now take a canonical affine cover of A,
U = Spec R[f 2 /f 1 , · · · , f t /f 1 ] such that the exceptional divisor T is given by the element f 1
α1 , · · · , α
ec ).
on U . For i = 1, · · · , c write α
ei = Ui,1 + Ui,2 f 2 /f 1 + · · · + Ui,t f t /f 1 and set IVe = (e
Then on the open set U we have IV · OU = IVe · (f1 ) and IVe defines an irreducible nonsingular
variety of Ve on U , which prove the statement (1) and the first part of the statement (2) in the
claim.
Next we compute the dualizing sheaf of Ve . Notice that we have IV ·OAe = IVe ·OAe(−T −µ∗ G).
Since IV is a complete intersection in A of codimension c we have a surjective morphism
e a surjective morphism
⊕c OA −→ IV −→ 0, which induces on A
c
M
OAe(T + µ∗ G) −→ IVe −→ 0.
e is
Thus it is clear that the determinant of the normal bundle of Ve inside A
det NVe /Ae = OVe (−c(T + µ∗ G)).
Then by the adjunction formula, we have
− c(T + µ∗ G)) ⊗ ψ ∗ ωA ,
ωVe ≃ ωAe ⊗ det NVe /Ae = OVe (KA/A
e
which proves the second part of the statement (2) in the claim. For the statement (3), just
notice that the morphism ψ is an isomorphism at the generic point of Y . Thus we finish the
proof of Claim 3.1.2.
Now twisting the short exact sequence 0 −→ IVe −→ OAe −→ OVe −→ 0 by the divisor
− c(T + µ∗ G)) ⊗ ψ ∗ ωA , we obtain an exact sequence
OAe(KA/A
e
∗
∗
∗
∗
(3.1.3) 0 → IVe ·OAe(KA/A
e −c(T +µ G))⊗ψ ωA → ωVe → 0.
e(KA/A
e −c(T +µ G))⊗ψ ωA → OA
7
Push down this sequence via ψ. Notice that by the definition of multiplier ideal sheaves, we
obtain
− c(T + µ∗ G)) = I (A, cX).
ψ∗ OAe(KA/A
e
Now we make the following claim.
Claim 3.1.4.
We have the following statements for the sequence 3.1.3.
(1) Let I (A, cV ) be the multiplier ideal sheaf associated to the pair (A, cV ), then
− c(T + µ∗ G))) = I (A, cV ).
ψ∗ (IVe · OAe(KA/A
e
(2) We have the vanishing
− c(T + µ∗ G)) = 0,
Ri ψ∗ (IVe · OAe(KA/A
e
for i > 0.
e be the blowing-up of A
e along Ve such that I e ·
Proof of Claim 3.1.4. Let ν : A′ −→ A
V
OA′ = OA′ (−S), where S is an exceptional divisor of ν. Notice that KA′ /Ae = (c − 1)S and
KA′ /A = KA′ /Ae + ν ∗ KA/A
e . Thus we have
− cν ∗ (T + µ∗ G) = KA′ /A − c(S + ν ∗ T + (µ ◦ ν)∗ G).
−S + ν ∗ KA/A
e
Write the divisor D := (S + ν ∗ T + (µ ◦ ν)∗ G). We notice that IV · OA′ = OA′ (−D). Since
ν∗ OA′ (−S) = IVe and Ri ν∗ OA′ (−S) = 0 for i > 0, we obtain that
− c(T + µ ∗ G))
ν∗ OA′ (KA′ /A − cD) = IVe · OAe(KA/A
e
and
Ri ν∗ OA′ (KA′ /A − cD) = 0,
for i > 0.
Write f := ψ ◦ ν : A′ −→ A. The divisor −D is f -nef and then by the Kawamata-Viehweg
vanishing theorem we have
Ri f∗ OA′ (KA′ /A − cD) = 0,
for i > 0.
Also by the definition of multiplier ideal sheaves we have f∗ OA′ (KA′ /A − cD) = I (A, cV ).
Now by using the spectral sequence
E2p,q = Rp ψ∗ (Rq ν∗ OA′ (KA′ /A − cD)) ⇒ Rp+q f∗ OA′ (KA′ /A − cD),
we immediately have that
− c(T + µ∗ G)) = 0,
Ri ψ∗ (IVe · OAe(KA/A
e
for i > 0,
and
− c(T + µ∗ G))) = I (A, cV ),
ψ∗ (IVe · OAe(KA/A
e
which finish the proof of Claim 3.1.4.
Finally, we push down the sequence (3.1.3) via ψ to obtain an exact sequence
0 −→ I (A, cV ) ⊗ ωA −→ I (A, cX) ⊗ ωA −→ ωYGR −→ 0.
Thus by restricting to Y we see I (A, cX) · OY ⊗ ωA ≃ ωYGR , which proves the proposition.
8
Next we show that the isomorphism proved in Proposition 3.1 is canonical in the sense
that it fits into the commutative diagram 1.1.1. To this end, we need to track carefully the
isomorphisms constructed in the proposition. Since canonical sheaves are only unique up to
isomorphism we need to fix our canonical sheaves uniformly in arguments. There is a canonical
canonical sheaf, namely sheaf of regular differential forms, developed by Kunz [KW88], offering
a concrete way to do so. We follow the notation of Lipman [Lip84] to denote by ω
e the sheaf of
regular differential forms.
Let Z be a reduced of pure dimension d scheme of finite type over k. Denote by KZ the locally
constant sheaf of total quotient ring of OZ . Its ring of global sections is K(Z) := KZ (Z) which
is a product of residue fields of the generic points of Z. Let ΩdZ/k := ∧d Ω1Z/k be the d-th exterior
power of the sheaf of Kählar differential one-form. Let Ω̄dK(Z)/k be the locally constant sheaf
of meromorphic d-forms on Z so that its module of global sections is Ω̄dK(Z)/k (Z) = ΩdK(Z)/k .
The sheaf ω
eZ of regular differential forms of degree d of Z is defined in [KW88, Section 3] and
it is a subsheaf of Ω̄dZ/k . Now let f : Z ′ −→ Z be a resolution of singularities of Z so that f is
isomorphic at the generic points of Z. Then pushdown the inclusion ω
eZ ′ ֒→ Ω̄dK(Z ′ )/k via f we
have f∗ ω
eZ ′ ֒→ f∗ Ω̄dK(Z ′ )/k . But since f is generically isomorphism we see that K(Z ′ ) = K(Z)
and f∗ Ω̄dK(Z ′ )/k = Ω̄dK(Z)/k . Thus f∗ ω
eZ ′ is naturally included in Ω̄dK(Z)/k as a subsheaf. The
eZ is then the natural inclusion as subsheaves of Ω̄dK(Z)/k .
trace map tr : f∗ ω
eZ ′ ֒→ ω
Proposition 3.2. With notation as in Proposition 3.1, the isomorphism ωYGR ≃ I (A, cX) ·
OY ⊗ ωA is canonical in the sense that it fits into the following commutative diagram
ωYGR
(3.2.1)
≃
// I (A, cX) · OY ⊗ ωA
_
_
tr
i
≃
ωY
// IX · OY ⊗ ωA
where the bottom isomorphism is given by 2.1.1, tr is the trace map, and the inclusion i is
induced by I (A, cX) ֒→ IX (cf. Lemma 2.5).
Proof. Keep notation and construction as in the proof of Proposition 3.1. Assume that d :=
dim X. We first make the adjunction isomorphism ωV ≃ ωA ⊗OA OV precisely by using
regular differential forms mentioned above. Recall that V is a complete intersection defined
by IV = (α1 , · · · , αc ). Then it is clear that IV /IV2 = ⊕OV ᾱi . Following notation of [Lip84,
Section 13] we set a sheaf
eA /IV ω
eA ),
HV,A := H omOV (∧c IV /IV2 , ω
which is torsion free since ω
eA is locally free. Notice that HV,A = det NV /A ⊗OA ω
eA where NV /A
is the normal sheaf of V in A. We have the following commutative diagram
(3.2.2)
ω
eV _
Ω̄dK(V )/k
≃
// HV,A
≃
// ω
eA ⊗OA OV
iV
≃
aV
// HV,A ⊗O KV
V
jV
≃
bV
// ω
eA ⊗OA KV
where the left commutative square follows from [HS97, Theorem 2.3] (see also [Lip84, Corollary
13.7]) and the right commutative square is a consequence of that the sheaves inside H om are
all locally free and det NV /A ≃ OV . The morphism iV and jV are injective because HV,A
and ω
eA ⊗OA OV are torsion free. Furthermore the vertical morphisms in the diagram can
9
be thought of induced by tensoring with the natural inclusion OV ֒→ KV . Thus HV,A and
ω
eA ⊗OA OV are naturally as subsheaves of those locally constant sheaves at the bottom of the
−1
diagram. The adjunction isomorphism ω
eV ≃ ω
eA ⊗OA OV is then induced by a−1
V ◦ bV , i.e.
−1
ωA ⊗OA OV ). The isomorphisms aV and bV can be described precisely and
ω
eV = (a−1
V ◦ bV )(e
determined completely at each generic point of V . Assume v is one generic point of V with the
residue field k(v). The local ring OA,v has a regular system of parameters α1 , · · · , αc , x1 , · · · , xd .
Then locally at an open set of V containing only v the sheaf Ω̄dK(V )/k = k(v)dx̄1 ∧ · · · ∧ dx̄d ,
the sheaf HV,A ⊗ KV = k(v)ξ where ξ maps ᾱ1 ∧ · · · ∧ ᾱc to dα1 ∧ · · · dαc ∧ dx1 ∧ · · · ∧ dxd and
the sheaf ω
eA ⊗OA KV = k(v)dα1 ∧ · · · dαc ∧ dx1 ∧ · · · ∧ dxd . Then on this open neighborhood
of v the bottom line of the diagram 3.2.2 can be written as
k(v)dx̄1 ∧ · · · ∧ dx̄d −→ k(v)ξ −→ k(v)dα1 ∧ · · · ∧ dαc ∧ dx1 · · · ∧ dxd .
The isomorphisms aV and bV are defined in the way that aV (dx̄1 ∧ · · · ∧ dx̄d ) = ξ and bV (ξ) =
dα1 ∧ · · · ∧ dαc ∧ dx1 · · · ∧ dxd .
Next we need to make the isomorphism ωY ≃ IX · OY ⊗ ωA clearly. Write the sheaf
HY,V := H omOY (OY , ω
eV /IY /V ω
eV )
where IY /V = IY · OV . It is a torsion free since ω
eV is locally free. There is a fundamental local
homomorphism (cf. [Lip84, 13.1])
h : H omOV (OY , ω
eV ) −→ HY,V
which in our case is induced by the natural quotient morphism ω
eV −→ ω
eV /IY /V ω
eV . We have
the following commutative diagram
(3.2.3)
ω
eY _
≃
// H omO (OY , ω
eV )
V
// ω
eV ⊗OV IX · OV
uY
h
≃
HY,V
// ω
eV ⊗OV OY
jY
iY
Ω̄dK(Y )/k
≃
aY
// HY,V ⊗O KY
Y
≃
bY
// ω
eV ⊗OV KY .
The left hand side big square is commutative checked by definition directly. The right two
small squares are commutative because ω
eV is locally free and H omOV (OY , OV ) = IX · OV .
The morphisms iY and iY ◦ h are injective since sheaves involved are all torsion free. Now let v
be a generic point of Y which is also a generic point of V . Suppose that the local ring OA,v has a
regular system of parameters α1 , · · · , αc , x1 , · · · , xd . Then locally at an open set of Y containing
only v we see that the sheaves Ω̄dK(Y )/k = k(v)dx̄1 ∧ · · · ∧ dx̄d , HY,V ⊗OY KY = k(v)ξY where ξY
maps 1 to dx̄1 ∧ · · · ∧ dx̄d , and ω
eV ⊗OV KY = k(v)dx̄1 ∧ · · · ∧ dx̄d . Thus it is easy to check that
the morphism bY ◦aY is the identity. Hence in the locally constant sheaf Ω̄dK(Y )/k = ω
eV ⊗OY KY
the canonical sheaf ω
eY is exactly the sheaf ω
eV ⊗OV IX · OV . Furthermore since Y is generic
linked to X via V we see that IX · OV = IX · OY .
10
Now tensoring the diagram (3.2.2) with OY over OV . Notice that KV ⊗OV OY = KY since
IY /V · KV = KX . And combining the diagram (3.2.3) together we have
(3.2.4)
=
ω
eY _
// ω
eV ⊗OV IX · OV
// ω
eA ⊗OA IX · OV
// ω
eA ⊗OA OY
uY
ω
eV ⊗OV OY
jY
=
Ω̄dK(Y )/k
bY ◦aY
≃
// ω
eV ⊗OV KY
(bV ◦aV )⊗1Y
// ω
eA ⊗OA KV ⊗OV OY .
The two horizontal morphisms on the top and right of the diagram can be thought of as the
restriction of the right bottom morphism of locally constant sheaves on their subsheaves. Thus
the sheaf ω
eY is the image of IX ⊗ ω
eA under the following morphisms
(3.2.5)
(bV ◦aV )⊗1−1
Y
IX ⊗ ω
eA ֒→ ω
eA −→ ω
eA ⊗ OY −→ ω
eA ⊗OA KY −−−−−−−−
→ Ω̄dK(Y )/k
where except (bV ◦ aV ) ⊗ 1−1
Y the rest morphisms are all natural ones.
Now we look at the canonical sheaf ω
eVe . Denote by E := T + µ∗ G an effective divisor on
e Recall that Ve is locally a complete intersection on A
e and ∧c I e /I 2 = O e(cE). The normal
A.
A
V
Ve
sheaf NVe /Ae = OVe (−cE). We define the sheaf
eAe)
eAe/IVe ω
HVe ,Ae = H omOVe (∧c IVe /IV2e , ω
which is torsion free since ω
eAe is free. Exactly as in the situation of V on A, we have the
following commutative diagram
(3.2.6)
ω
eVe
≃
_
Ω̄d
K(Ve )/k
≃
aVe
≃
// H e e
V ,A
// ω
eAe ⊗O e OVe (−cE)
A
// H e e ⊗O K e
e
V ,A
V
V
≃
bVe
// ω
eA ⊗OA KV
eAe(−cE) under the morphisms
Thus the sheaf ω
eVe is the image of the sheaf ω
(3.2.7)
(a e ◦b e )−1
V
V
eAe ⊗OAe KVe −−−
eAe ⊗ OVe −→ ω
eAe −→ ω
ω
eAe(−cE) ֒→ ω
−−
−−→ Ω̄dK(Ve )/k
where except of (aVe ◦ bVe )−1 all morphisms are natural ones. Push down (3.2.7) via ψ. Notice
eA . Also since the birational morphism ψ
eA and ψ∗ ω
eAe = ω
that ψ∗ (e
ωAe(−cE)) = I (A, cX) ⊗ ω
is an isomorphism over A\X so it is an isomorphism around generic points of Y and Ve and
eA ⊗OA KY and ψ∗ Ω̄d e
therefore ψ∗ (e
ωAe ⊗OAe KVe ) = ω
= Ω̄dK(Y )/k as locally constant sheaves.
K(V )/k
Furthermore at one generic point v of Y which is identical to a generic point of Ve since ψ is
an isomorphism around v, we can choose the same local equation of the local ring OA,v and
−1
OA,v
e , for instance, α1 , · · · , αc , x1 , · · · , xd . Then we can check that the morphism (aVe ◦ bVe )
eVe is the image
is the same as the morphism (bV ◦ aV ) ⊗ 1−1
Y . Thus we see that the sheaf ψ∗ ω
of I (A, cX) ⊗ ωA under the morphisms
(3.2.8)
(bV ◦aV )⊗1−1
Y
eA ⊗OA KY −−−−−−−−
→ Ω̄dK(Y )/k.
I (A, cX) ⊗ ω
eA ֒→ ω
eA −→ ψ∗ (e
ωAe ⊗ OVe ) −→ ω
11
The fact that I (A, cX) ⊗ ω
eA is mapped surjectively to ψ∗ ω
eVe is guaranteed by Claim 3.1.4 in
the proof of Proposition 3.1.
e
Finally we compare (3.2.5) with (3.2.8). We have the following commutative diagram on A
(3.2.9)
ψ∗ ω
eA
// ψ ∗ ω
eA ⊗ OVe
// ψ ∗ ω
eA ⊗OAe KVe
// ω
eAe ⊗ OVe
// ω
eAe ⊗OAe KVe
ω
eAe
where the vertical morphisms are induced by the morphism ψ ∗ Ω1A/k −→ Ω1e . Push down the
A/k
eA then we have the commutative diagram
diagram and notice that ψ∗ ω
eAe = ω
(3.2.10)
ω
eA
ω
eA ⊗ ψ∗ OVe
// ω
eA ⊗OA KY
V
// ω
eA ⊗OA KY
99
g ssss
s
sss
sss
// ψ∗ (e
ω e ⊗ Oe )
A
eA −→
Since ω
eA ⊗ ψ∗ OVe is naturally a OY -module the morphism g then factors through ω
ω
eA ⊗ OY −→ ω
eA ⊗ ψ∗ OVe . Now it is clear that the proposition follows from (3.2.5) and
(3.2.8).
Corollary 3.3. With notation as in Definition 2.3, let Y be a generic link to an affine pair
(Ak , cXk ). Then ωYGR = ωY if and only if IXk = I (Ak , cXk ), where I (Ak , cXk ) is the
multiplier ideal sheaf associated to the pair (Ak , cXk ).
Proof. Let I (A, cX) be the multiplier ideal sheaf associated to the pair (A, cX). By the
commutative diagram (3.2.1) we have ωYGR = ωY if and only if I (A, cX) · OY = IX · OY if and
only if IX + IY = I (A, cX) + IY . Intersecting with IX and noticing that IX ∩ IY = IV and
I (A, cX) ⊆ IX , we see that IX + IY = I (A, cX) + IY if and only if IX = I (A, cX) + IV .
Now since the morphism A −→ Ak is smooth, we then have I (A, cX) = I (Ak , cXk ) · OA
by [Laz04b, 9.5.45]. Also notice that IX = IXk · OA and the ring extension OAk −→ OA is
faithfully flat. Thus intersecting with OAk , we conclude that IX = I (A, cX) + IV if and only
if IXk = I (Ak , cXk ).
Now we can easily deduce a criterion when a generic link has rational singularities. It turns
out that multiplier ideal sheaves determine rational singularities of a generic link.
Corollary 3.4. With notation as in Definition 2.3, let Y be a generic link to an affine pair
(Ak , cXk ). Then Y has rational singularities if and only if Xk is Cohen-Macaulay and IXk =
I (Ak , cXk ), where I (Ak , cXk ) is the multiplier ideal sheaf associated to the pair (Ak , cXk ).
Proof. Y has rational singularities if and only if Y is Cohen-Macaulay and ωYGR = ωY . Then
the result is clear from above.
Corollary 3.5. With notation as in Definition 2.3, let Y be a generic link to an affine pair
(Ak , cXk ). Suppose that the pair (Ak , cXk ) is log canonical and Xk is Cohen-Macaulay. Then
Y has rational singularities.
Proof. By Ein’s Lemma, that (Ak , cXk ) is log canonical implies IXk = I (Ak , cXk ). Then the
result follows from above.
Remark 3.6. Let us go back to the result of Chardin and Ulrich mentioned in Introduction.
Still with notation as in Definition 2.3 suppose that Xk is a local complete intersection with
12
rational singularities. Then by the inversion of adjunction [EM09] the pair (Ak , cXk ) is log
canonical. Thus Corollary 3.4 says that a generic link Y of Xk has rational singularities.
Proposition 3.7. With notation as in Definition 2.3 let (Ak , cXk ) be an affine pair and let
Y be a generic link to X via a complete intersection V . Then
lct(A, Y ) ≥ lct(A, V ) = lct(A, X) = lct(Ak , Xk ).
Proof. Keep notation and construction as in the proof of Proposition 3.1. First of all it is clear
that lct(A, X) = lct(Ak , Xk ) and since IV ⊆ IY we have lct(A, Y ) ≥ lct(A, V ). Thus it suffices
to show lct(A, V ) = lct(A, X).
Recall that ϕk : Ak −→ Ak is a factorization resolution of singularities of Xk inside Ak and
X k is the strict transform of Xk . Denote by Exc(ϕk ) = ∪si=1 Eik the exceptional locus of ϕk
with
where Eik are prime divisors with normal crossing support. Then X k has normal
Pcrossing
s
k
k
k
E1 , · · · , Es (cf. Definition 2.6). We then can write the effective divisor Gk = i=1 ai Ei and
P
the relative canonical divisor KAk /Ak = si=1 ki Eik
Recall also that the factorizing resolution of singularities ϕ : A −→ A of X inside A is
then obtained by tensoring Spec k[Uij ] with the resolution ϕk . Write Ei := Eik ⊗k Spec k[Uij ]
for i = 1, · · · , s. Then it is clear that the exceptional locus of ϕ is Exc(ϕ) = ∪si=1 Ei with
X of X has normal crossing with
simple normal crossing support, and the strict transform
Ps
E1 , · · · , E
.
Furthermore
the
effective
divisor
G
=
a
E
s
i=1 i i and the relative canonical divisor
P
s
KA/A = i=1 ki Ei .
e −→ A is the composition of ϕ with the blowing-up µ : A
e −→ A
Now the morphism ψ : A
∗
ei = µ (Ei ) for i = 1, · · · , s. Since X has
along X with the exceptional divisor T . We set E
e
ei has simple
normal crossing with Ei we see that Ei is a prime divisor and Exc(ψ) = T ∪si=1 E
e −→ A is a log resolution of X inside A. Notice that
normal crossing support. Thus ψ : A
= (c − 1)T . We then can write
KA/A
e
(3.7.1)
= (c − 1)T +
KA/A
e
s
X
i=1
ei ,
ki E
ψ −1 (X) = T +
s
X
i=1
ei .
ai E
e (cf. Claim
Claim 3.7.2. Recall that Ve is nonsingular locally complete intersection on A
e1 , · · · , E
es .
3.1.2). Then Ve has normal crossing with T, E
Proof of Claim 3.7.2. The question is local so we just need to look at local equations. Let
U k = Spec Rk be an affine open set of Ak such that IX k = (f 1 , · · · , f t ) ⊂ Rk and Eik has a
local equation hi ∈ Rk for i = 1, · · · , s (cf. proof of Claim 3.1.1). Let U = U k ⊗ Spec k[Uij ] be
the corresponding affine open set in A. Write R = Rk ⊗ k[Uij ]. Then IX = (f 1 , · · · , f t ) · R and
each Ei is still defined by the equation hi in the ring R. Now let U1 = Spec R[f 2 /f 1 , · · · , f t /f 1 ]
e over U . Then on U1 the divisor T is defined by the equation f 1 .
be one canonical cover of A
e
Notice that each Ei is still defined by the local equation hi ∈ R[f 2 /f 1 , · · · , f t /f 1 ]. On U1 the
α1 , · · · , α
ec ), where
variety Ve is defined by IVe = (e
α
ei = Ui,1 + Ui,2 f 2 /f 1 + · · · + Ui,t f 2 /f 1 ,
for i = 1, · · · , c.
Now we just need to show on U1 , IVe , f 1 , h1 , · · · , hc are normal crossings. Notice that
f 1 , h1 , · · · , hc are already normal crossings by the construction and they all sit in the ring
ei ’s over Rk . Thus a local calculation shows
Rk . But IVe is essentially defined by variables α
that IVe meets f 1 , h1 , · · · , hc as normal crossings. This finishes the proof of Claim 3.7.2.
13
e along Ve with the exceptional divisor S and
Now recall ν : A′ −→ A is the blowing-up of A
′
′
∗
ei for i = 1, · · · , s and T ′ = ν ∗ T . By Claim 3.7.2 above we
f = ν ◦ ψ : A −→ A. Write Ei = ν E
see that T ′ , E1′ , · · · , Es′ are all prime divisors and Exc(f ) = S ∪ T ′ ∪si=1 Ei′ are simple normal
crossings. Thus f : A′ −→ A is a log resolution of (A, V + X), which we use to compute log
canonical thresholds. Notice that KA′ /Ae = (c − 1)S. From (3.7.1) we can write
KA′ /A = (c − 1)S + (c − 1)T ′ +
s
X
ki E ′ ,
f −1 (X) = T ′ +
s
X
ai Ei′ .
i=1
i=1
Since IV · OAe = IVe · OAe(−ψ −1 (X)) (cf. Claim 3.1.1 and 3.1.2) we then have
f −1 (V ) = S + T ′ +
s
X
ai Ei′ .
i=1
Finally by the definition of log canonical threshold we see that
lct(A, X) = min{
ki + 1 (c − 1) + 1 (c − 1) + 1
,
,
}
ai
1
0
and
ki + 1 (c − 1) + 1 (c − 1) + 1
,
,
}.
ai
1
1
Therefore lct(A, X) = lct(A, V ) as required.
lct(A, V ) = min{
Corollary 3.8. With notation as in Definition 2.3, if IX = I (A, cX), where I (A, cX) is the
multiplier ideal sheaf associated to the pair (A, cX), then
IV = I (A, cV ) and IY = I (A, cY ),
where I (A, cV ) and I (A, cY ) are multiplier ideal sheaves associated to the pairs (A, cV ) and
(A, cY ), respectively.
Proof. By Ein’s Lemma, IX = I (A, cX) if and only if I (A, (c−1)X) = OA . Thus lct(A, X) >
(c − 1) and therefore by Theorem 3.7 lct(A, Y ) ≥ lct(A, V ) = lct(A, X) > (c − 1). Hence the
multiplier ideal sheaves I (A, (c − 1)Y ) and I (A, (c − 1)V ) are all trivial. The result then
follows by using Ein’s Lemma again.
Remark 3.9. In the above corollary, the equality IX = I (A, cX) is equivalent to the equality
IXk = I (Ak , cXk ), where I (Ak , cXk ) is the multiplier ideal sheaf associated to the pair
(Ak , cXk ). This is because the morphism A −→ Ak is smooth and the ring extension Rk −→ R
is faithfully flat.
Corollary 3.10. With notation as in Definition 2.3, let Y be a generic link to an affine pair
(Ak , cXk ). Suppose that the pair (Ak , cXk ) is log canonical. Then the pair (A, cY ) is also log
canonical.
Proof. Since, by assumption, (Ak , cXk ) is log canonical and, by Lemma 2.5, I (Ak , cXk ) ⊆ IXk
we see that lct(Ak , Xk ) = c. Thus by Theorem 3.7, we have lct(A, Y ) ≥ c. But by Lemma 2.5,
we have lct(A, Y ) ≤ c. Therefore lct(A, Y ) = c and thus (A, cY ) is log canonical.
Proof of Theorem 1.1 in Situation A. First of all in this case, Y cannot be empty. The
fact that Y is reduced is standard by [HU85, 2.6]. The rest of the theorem follows from
Proposition 3.1, Proposition 3.2 and Proposition 3.7.
14
Remark 3.11. Using results established above, we then look at a sequence of generic linkages.
Precisely, let (Ak , cXk ) be an affine pair as in Definition 2.3 and set Y0 := Xk and A0 := Ak .
We denote a generic link of Y0 to be Y1 , which is in a nonsingular ambient space A1 . We
can continue to construct a generic link of Y1 as Y2 in an nonsingular ambient space A2 .
Consequently, we get a sequence Y0 , Y1 , · · · , such that each Yi is a generic link of Yi−1 and is
in a nonsingular variety Ai . Now we list some interesting consequences from the above results
as follows.
(1) If IY0 = I (A0 , cY0 ), then IYi = I (Ai , cYi ), i.e. equality of multiplier ideal with ideal
is preserved by generic linkages.
(2) If (A0 , cY0 ) is log canonical then (Ai , cYi ) is log canonical, i.e. log canonical pair is
preserved by generic linkage.
(3) If (A0 , cY0 ) is log canonical and Y0 is rational then (Ai , cYi ) is log canonical and Yi is
rational, i.e. log canonical plus rational is preserved by generic linkages.
(4) lct(A0 , Y0 ) ≤ lct(A1 , Y1 ) ≤ · · · ≤ c, i.e. log canonical thresholds increase in generic
linkages but bounded above by c.
Notice that in (4) we get an increasing but bounded above sequence. Thus it must have a
limit, which we denoted by lct∞ (A0 , Y0 ). It would be very interesting to know if lct∞ (A0 , Y0 )
is independent on the choice of generic linage sequence Y1 , Y2 , · · · . It has been conjectured by
the author that after finitely many steps of generic link sequences we will have lct∞ (A0 , Y0 ) = c.
However, it was pointed out by Bernd Ulrich that the conjecture is not true because otherwise
if we start with a Cohen-Macaulay Y0 we will end up with a variety Y which has rational
singularities, but it is not the case.
Remark 3.12. There is a conjecture made by the author in [Niu11, Conjecture 1.4] which
asserts that if X is a local complete intersection with log canonical singularities then a generic
link Y of X is also a local complete intersection with log canonical singularities. Now it is clear
that this conjecture is false. One main reason is that Y cannot be a local complete intersection.
However, Corollary 3.10 says that the log canonical pair is preserved by generic linkages and
in the conjecture the pair (A, cX) is actually log canonical.
In the last of this section, we consider specialization problem stated in Situation B. It is
a direct consequence of Situation A by restricting to the general fiber over Spec k[Ui,j ]. This
quick proof of the theorem was suggested by the referee.
Proof of Theorem 1.1 in Situation B. Keep the notation and construction in the proof of
Proposition 3.1. Let A = Spec k[Ui,j ] be the affine space parameterizing the scalar matrices
(ai,j )c×t . Recall that we have the following diagram
(3.12.1)
Ve
e
// A
ψ
Y
// A
where Ve is the embedded resolution of Y (cf. Claim 3.1.2). It is a diagram over A. Thus it
is enough to prove the theorem for a general fiber of this diagram over A. Since we assume a
general fiber of Y over A is nonempty the morphism Y −→ A must be dominant. (In fact the
only case that a general fiber of Y is empty is that Xk is a complete intersection. In this case
we cannot make generic link under Situation B.) Thus the image of Y must contain a open
set U of A. By replacing A by this open set we may assume that the morphism Y −→ A is
15
surjective. Now over a general point p ∈ A, the diagram is
Vep
ep
// A
ψp
Yp
// Ap
ep is the blowing-up of Ak along X k and the morphism ψp is
Notice that by the construction A
an isomorphism over Ak /Xk . Also by the generic smoothness Vep is the embedded resolution of
Yp . Thus Yp is generically reduced. But Yp does not have embedded components and therefore
is reduced. Now the rest of the argument is exact the same as the proof of the theorem in
Situation A.
Remark 3.13. Corollary 3.5, 3.8 and 3.10 are still true in Situation B. However, in Corollary
3.3 and 3.4 the “if” parts are still true but “only if” parts are not in general.
Remark 3.14. As in Remark 3.11 (4), we can look at an increasing sequence of log canonical
thresholds lct(A, Y0 ) ≤ lct(A, Y1 ) ≤ · · · ≤ c, in which Yi is a generic link of Yi−1 in Situation B.
It was pointed out by Lawrence Ein that by using ACC for log canonical thresholds [dFEM10],
after finitely many steps the number lct becomes stable in the sequence, which we denoted by
lct∞ (A, Y0 ). At the moment, it is not clear to us if this number depends on the specific linkage
sequence or not. It would be very interesting to have a further investigation on this number.
We hope that this new invariant will have some application in linkage classes of varieties.
4. Generic linkage of projective varieties and Castelnuovo-Mumford
regularity
In this section, we study a generic link of a subvariety of Pn in Situation C. The main idea is
inspired by the work of Betram, Ein and Lazarsfeld [BEL91]. Thus we shall be brief in proofs.
Throughout this section, we assume that A is a projective nonsingular variety over k and L
is a line bundle on A generated by its global sections. We shall prove the following theorem
which can be applied to a slightly more general case than Situation C.
Theorem 4.1. Suppose that X ⊂ A is a reduced equidimensional subscheme of codimension
c scheme-theoretically defined by the t sections si ∈ H 0 (A, Ldi ) with d1 ≥ d2 ≥ · · · ≥ dt . Take
c general sections αi ∈ H 0 (A, IX ⊗ Ldi ) for i = 1, · · · , c and let V be a complete intersection
defined by the vanishing of α1 , · · · , αc . Let Y be a subscheme of A defined by IY := (IV : IX ).
Then either Y is empty or else:
(1) Y is reduced equidimensional of codimension c (possibly reducible) geometrically linked
to X via V .
(2) ωYGR ≃ I (A, cX) · OY ⊗ ωA ⊗ Ld1 +···+dc , where I (A, cX) is the multiplier ideal sheaf
associated to the pair (A, cX), and it fits into the following commutative diagram
ωYGR
≃
_
// I (A, cX) · OY ⊗ ωA ⊗ Ld1 +···+dc
_
tr
i
ωY
≃
// IX · OY ⊗ ωA ⊗ Ld1 +···+dc
where the bottom one is given by 2.1.2, tr is the trace map, and the inclusion i is
induced by I (A, cX) ֒→ IX (cf. Lemma 2.5).
(3) lct(A, Y ) ≥ lct(A, V ) = lct(A, X).
16
Proof. Take a log resolution of singularities for the pair (A, X) as f : A −→ A such that
IX · OA = OA (−E) where E is an effective divisor and Exc(f ) ∪ E is a divisor with simple
normal crossing support. We may also assume that the morphism f is an isomorphism over
the open set A\X.
For i = 1, · · · , t we denote by bi the sub-linear system of |Ldi | determined by the vector
space H 0 (A, IX ⊗ Ldi ). We use notation (si )0 to be the effective divisor in the linear system
bi defined by the zero locus of the section si . Since X is defined by the vanishing of sections
(s1 ,··· ,st )
si we have a surjective morphism ⊕t L−di −−−−−−→ IX −→ 0. Pulling back this surjective
morphism via f , we then obtain a surjective morphism
(4.1.1)
t
M
(f ∗ s1 ,··· ,f ∗ st )
f ∗ L−di −−−−−−−−−→ OA (−E) −→ 0.
(Since f is dominant the induced morphism on the linear system f ∗ : bi −→ |f ∗ Ldi | is actually
injective. So we think of bi naturally as a sub-linear system of |f ∗ Ldi | under f ∗ .) Denote
by Bi = f ∗ bi − E the sub-linear system of |f ∗ Ldi (−E)| obtained from f ∗ bi by removing the
base locus divisor E. Then the section f ∗ si naturally gives rise to a section σi of f ∗ Ldi (−E)
defining an effective divisor Fi in the linear system Bi . Thus from the surjectivity of (4.1.1),
we deduce a surjection
t
M
(σ1 ,··· ,σt )
f ∗ L−di (E) −−−−−−→ OA −→ 0,
and therefore we have
(4.1.2)
F1 ∩ F2 ∩ · · · ∩ Ft = φ.
We make the following observation.
Claim 4.1.3. For each i = 1, · · · , t, one has
(a) The system B1 is base point free.
(b) For each i ≥ 1, the base locus Bs(Bi ) of the system Bi is inside the support of the
divisor Fj for j ≥ i.
Proof of Claim 4.1.3. For the statement (a), by the definition of X, we see that the sheaf
IX ⊗ Ld1 is generated by its global sections. Let W1 = H 0 (A, IX ⊗ Ld1 ) so that b1 = |W |. So
we have a surjective morphism W1 ⊗ Ld1 −→ IX −→ 0. Thus we have a surjective morphism
W1 ⊗ f ∗ L−d1 −→ OA (−E) −→ 0, i.e., W1 ⊗ f ∗ L−d1 (E) −→ OA −→ 0. Thus we see that B1 is
base point free.
For the statement (b), notice that when i = 1 the result is trivial from (a). We prove the
first nontrivial case when i = 2. It is from the definition of base loci that Bs(B2 ) ⊂ F2 . Now we
show Bs(B2 ) ⊂ F3 . Denote by δ32 the linear system |Ld2 −d3 | which is base point free. Notice
that we have an inclusion δ32 + (s3 )0 ⊂ b2 . Thus pulling back via f , we have the inclusion
f ∗ δ32 + f ∗ (s3 )0 ⊂ f ∗ b2 and therefore by subtracting the divisor E we see f ∗ δ32 + f ∗ (s3 )0 − E ⊂
f ∗ b2 − E. Recall that F3 = f ∗ (s3 )0 − E and B2 = f ∗ b2 − E. Thus we have an inclusion
f ∗ δ32 + F3 ⊂ B2 .
From this, the linear system B2 − F3 contains the system f ∗ δ32 , which is base point free. Thus
the base locus Bs(B2 ) is contained in F3 . Similar argument works for all j ≥ i, which proves
the Claim 4.1.3.
Now since B1 is base point free, by Bertini’s theorem (Cf. [Har77, Corollary III.10.9]) we
can take a general element D1 ∈ B1 such that (i) D1 is nonsingular and equidimensional; (ii)
17
no components of D1 are contained in the support of E ∪ Exc(f ); (iii) D1 has normal crossing
with Exc(f ); (iv) D1 ∩ F2 ∩ · · · ∩ Ft = φ. Here the reason for (iv) is that since the section σ1
is nowhere vanishing on F2 ∩ · · · ∩ Ft , the general section of B1 is then nowhere vanishing on
F2 ∩ · · · ∩ Ft .
Now by Claim 4.1.3 the base locus Bs(B2 ) is inside Fj for j ≥ 2. Thus we have Bs(B2 ) ⊂
F2 ∩ F3 ∩ · · · ∩ Ft . By the fact (4.1.2) and the choice of D1 , the linear system B2 is base
point free on D1 . Thus by Bertini’s theorem again we can choose a general element D2 ∈ B2
such that (i) D1 ∩ D2 is nonsingular and equidimensional; (ii) no components of D1 ∩ D2 are
contained in the support of E ∪ Exc(f ); (iii) D1 ∩ D2 has normal crossing with Exc(f ); (iv)
D1 ∩ D2 ∩ F3 ∩ · · · ∩ Ft = φ.
We then can iterate such argument by c times to end up with a subscheme
Y := D1 ∩ D2 ∩ · · · ∩ Dc
of A such that Y is either empty or else (i) Y is nonsingular and equidimensional; (ii) no
component of Y is contained in the support of E ∪ Exc(f ); (iii) Y has normal crossing with
Exc(f ); (iv) Y ∩ Fc+1 ∩ · · · ∩ Ft = φ. Notice that each effective divisor Di is a general element
in the linear system Bi .
Now each Di naturally corresponds to a general section αi ∈ H 0 (A, IX ⊗ Ldi ). (Recall that
the divisor Di + E is in f ∗ bi .) Those α1 , · · · , αc cut out a complete intersection V on A. Let Y
be the subscheme of A linked to X via V , i.e., Y is defined by an ideal sheaf IY := (IV : IX ).
It is well-know that Y is equidimensional of codimension c without embedded components and
with no common components with X. Notice that at least set-theoretically Y = f (Y ). Recall
that the morphism f is an isomorphism over the open set A\X. Thus by the construction
of Y we see that Y ∩ f −1 (U ) is isomorphic to Y ∩ U . Therefore f is an isomorphism at the
generic points of Y and then Y is the strict transform of Y . Thus Y is generically smooth and
therefore is reduced. The rest of the statement (1) are all standard result, so we would not
repeat here.
By the construction of Y , we have a surjective morphism ⊕c f ∗ L−di (E) −→ IY −→ 0. Thus
the normal bundle of Y inside A can be easily calculated and the rest argument for (2) is
exactly the same as the proof of Propositions 3.1 and 3.2. The reducedness of Y in (1) is also
clear since Y is generic reduced and has no any embedded components. Finally, the statement
(3) has the same proof as in Proposition 3.7 so we would not repeat here.
Remark 4.2. If A = Pn , we can actually reduce Situation C to Situation B, at least in the case
that all equations have the same degree. This was kindly suggested by the referee. Here let us
explain this reduction briefly. We choose general homogeneous coordinates for Pn so that all
the equations fi ’s restricted to the same degree polynomials, say gi ’s, on the standard charts.
Then we choose a general matrix B = (bij )c×t and we can find a invertible lower triangular
c × c matrix A so that A · B = (B1 |B2 ), where B1 is a c × c upper triangular matrix. Now the
complete intersection (β1 , · · · , βc ) = A · B · (g1 , · · · , gt )T will give back general sections αi in
H 0 (Pn , IX (di )).
As an application of above theorem, using an idea of [CU02] we can generalize results
of de Fernex and Ein [dFE10, Corollary 1.4] and Chardin and Ulrich [CU02, Theorem 0.1]
on the Castelnuovo-Mumford regularity bound. The bound was first established in [BEL91]
for nonsingular case and then generalized in [CU02] for rational singular case. [BEL91] and
[CU02] allow variety to have some not too bad very singular loci but [dFE10] cannot allow this
situation, which we are able to handle now (cf. [dFE10, Remark 5.2]). Recall that a coherent
sheaf F on the projecive space Pn is said to be m-regular if H i (Pn , F (m − i)) = 0 for all
18
i > 0. The minimal such number m is called the Castelnuovo-Mumford regularity of F and is
denoted by reg F . If X is a subscheme of Pn defined by an ideal sheaf IX then the regularity
of X is defined to be reg X = reg IX .
Corollary 4.3. Let X ⊂ Pn be a reduced equidimensional subscheme of codimension c defined
by the equations of degree d1 ≥ d2 ≥ · · · ≥ dt and let Z = Supp(IX /I (Pn , cX)) where
I (Pn , cX) is the multiplier ideal sheaf. Assume that dim Z ≤ 1 and each one dimensional
irreducible component of Z has at least one point at which X is a local complete intersection.
Then
c
X
di − c + 1
reg X ≤
i=1
and equality holds if and only if X is a complete intersection in Pn .
Proof. If X is a complete intersection then the result is clear. So in the sequel we assume X
is not a complete intersection. Then take a generic link Y of X via a complete intersection V
cut out by the general equations of X of degree d1 ≥ d2 ≥ · · · ≥ dc . By the assumption on
Z we can choose V general such that Y ∩ Z has dimension ≤ 0, i.e. no any dimension one
irreducible component of Z can be contained in Y . This means that from Theorem 4.1 we
have the following exact sequence
(4.3.1)
0 −→ ωYGR −→ ωY −→ Q −→ 0
where dim Supp Q ≤P0.
c
Denote by d :=
j=1 dj and r := dim V = dim X = dim Y . It is clear that from the
Kawamata-Viehweg vanishing theorem we can deduce reg ωYGR = r + 1 and therefore reg ωY =
r + 1 by the sequence (4.3.1). Now recall ωY = IX · OV ⊗ ωV . Since V is a complete
intersection we see reg IV = d − c + 1 and ωV = OV (d − n − 1) by Koszul resolution and thus
ωY = IX · OV (d − n − 1). Now from the exact sequence
0 −→ IV −→ IX −→ IX · OV −→ 0,
it is immediately that reg IX ≤ d − c.
Remark 4.4. In the corollary, by Ein’s Lemma the set Z can be equivalently defined as
Z := {x ∈ Pn | I (Pn , (c − 1)X) is non trivial at x}.
The condition on Z in the corollary already includes conditions discussed in de Fernex and Ein
[dFE10] and Chardin and Ulrich [CU02] and is certainly more general. It is worth mentioning
that the above liaison method to bound the regularity also can be used directly to study so
called multiregularity once the correct corresponding form of generic linkage can be built.
In Theorem 4.1, a generic link Y is usually not necessarily irreducible. One important
case in applications is when X is cut out by sections of the same degree, i.e., d1 = · · · = dt .
If the defining ideal sheaf IX has enough sections then it is possible that a generic link Y
is irreducible. One way to see this is by using the s-invariant of IX with respect to L,
which measures the positivity of IX . We recall its definition and refer to [Laz04a] for further
reference.
Definition 4.5. Given an ideal sheaf I on A let µ : W = BlI A −→ A be the blowing-up of A
along the ideal I with an exceptional Cartier divisor E on W , such that I · OW = OW (−E).
Let L be an ample line bundle on A. We define the s-invariant of I with respect to L to be
the positive real number
sL (I ) := min{ s | sµ∗ L − E is nef }.
19
Here sµ∗ L − E is considered as an R-divisor on W .
Remark 4.6. For example, in the projective space case the line bundle is usually taken to be
the hyperplane divisor and simply write the s-invariant as s(I ). In this case suppose that d is
an integer such that I (d) is generated by global sections, then it is easy to see that s(I ) ≤ d,
i.e., the s-invariant of I is always bounded by a generating degree of I . There is an example
in [CEL01] showing that the s-invariant could be an irrational number.
Corollary 4.7. With notation and assumption as in Theorem 4.1 assume further that L is
ample and X is cut out by the sections of the same degree, i.e., d1 = · · · = dt = d. If
the s-invariant of IX with respect to L is strictly smaller than d, then Y is nonempty and
irreducible.
Proof. We keep notation as in the proof of Theorem 4.1 but set d1 = · · · = dt = d. Now this
time the linear systems B1 , · · · , Bt are all the same as the linear system B := f ∗ b − E, where
b is the sub-linear system of |Ld | determined by the vector space H 0 (A, IX ⊗ Ld ). Notice that
B is a sub-linear system of |f ∗ Ld (−E)| and is base point free. Also notice that dim B = dim b
since f is dominant and E is the basic locus of f ∗ b.
All we need is to show the subscheme Y , which is the intersection of c general divisors
Di ∈ B, is irreducible. Recall that B is base point free, it then gives a morphism to a
projective space
φB : A −→ Pr ,
where r = dim B such that f ∗ Ld (−E) = φ∗B OPr (1). We claim that the morphism φB is
generically finite. To see this, let µ : A′ = BlX A −→ A be the blowing up of A along X
with an exceptional divisor F such that IX · OA′ = OA′ (−F ). By the universal property of
blowing-ups we have g : A −→ A′ such that f = g ◦µ and g ∗ F = E. Notice that g is generically
finite. Now the system B′ := µ∗ b − F is base point free and then gives a morphism φB′ to Pr ,
which commutes with φB , i.e. φB = φB′ ◦ g
A❄
❄❄
❄❄
❄
g ❄❄
φB
A′
// Pr
>>
⑥
⑥⑥
⑥
⑥⑥
⑥⑥ φB′
Since the s-invariant of IX with respect to L is strictly smaller than d the line bundle µ∗ Ld (−F )
is ample. But µ∗ Ld (−F ) = φ∗B′ OPr (1) so the morphism φB′ is finite. Thus φB is generically
finite and we have dim φB (A) = dim A. Now by the theorem of [Laz04a, 3.3.1], the subscheme
Y is nonempty and irreducible since dim φB (A) > c. Therefore the generic link Y to X is also
nonempty and irreducible.
Remark 4.8. Having Theorem 4.1 in hand, it is then easy to get those similar corollaries
mentioned in the previous sections (cf. Remark 3.11, 3.13 and 3.14). So we leave them to the
reader.
References
[BEL91]
[BVU03]
[CEL01]
Aaron Bertram, Lawrence Ein, and Robert Lazarsfeld. Vanishing theorems, a theorem of Severi, and
the equations defining projective varieties. J. Amer. Math. Soc., 4(3):587–602, 1991.
A. Bravo and O. Villamayor U. A strengthening of resolution of singularities in characteristic zero.
Proc. London Math. Soc. (3), 86(2):327–357, 2003.
Steven Dale Cutkosky, Lawrence Ein, and Robert Lazarsfeld. Positivity and complexity of ideal
sheaves. Math. Ann., 321(2):213–234, 2001.
20
[CU02]
Marc Chardin and Bernd Ulrich. Liaison and Castelnuovo-Mumford regularity. Amer. J. Math.,
124(6):1103–1124, 2002.
[dFD12] Tommaso de Fernex and Roi Docampo. Jacobian discrepancies and rational singularities. 2012.
[dFE10] Tommaso de Fernex and Lawrence Ein. A vanishing theorem for log canonical pairs. Amer. J. Math.,
132(5):1205–1221, 2010.
[dFEM10] Tommaso de Fernex, Lawrence Ein, and Mircea Mustaţă. Shokurov’s ACC conjecture for log canonical thresholds on smooth varieties. Duke Math. J., 152(1):93–114, 2010.
[EIM11] Lawrence Ein, Shihoko Ishii, and Mircea Mustata. Multiplier ideals via Mather discrepancy. Arxiv,
2011.
[EM09]
Lawrence Ein and Mircea Mustaţă. Jet schemes and singularities. In Algebraic geometry—Seattle
2005. Part 2, volume 80 of Proc. Sympos. Pure Math., pages 505–546. Amer. Math. Soc., Providence,
RI, 2009.
[Har77]
Robin Hartshorne. Algebraic geometry. Springer-Verlag, New York, 1977. Graduate Texts in Mathematics, No. 52.
[Hoc73]
M. Hochster. Properties of Noetherian rings stable under general grade reduction. Arch. Math.
(Basel), 24:393–396, 1973.
[HS97]
Reinhold Hübl and Gerhard Seibert. The adjunction morphism for regular differential forms and
relative duality. Compositio Math., 106(1):87–123, 1997.
[HU85]
Craig Huneke and Bernd Ulrich. Divisor class groups and deformations. Amer. J. Math., 107(6):1265–
1303 (1986), 1985.
[HU87]
Craig Huneke and Bernd Ulrich. The structure of linkage. Ann. of Math. (2), 126(2):277–334, 1987.
[HU88]
Craig Huneke and Bernd Ulrich. Algebraic linkage. Duke Math. J., 56(3):415–429, 1988.
[Kol97]
János Kollár. Singularities of pairs. In Algebraic geometry—Santa Cruz 1995, volume 62 of Proc.
Sympos. Pure Math., pages 221–287. Amer. Math. Soc., Providence, RI, 1997.
[KW88]
Ernst Kunz and Rolf Waldi. Regular differential forms, volume 79 of Contemporary Mathematics.
American Mathematical Society, Providence, RI, 1988.
[Laz04a] Robert Lazarsfeld. Positivity in algebraic geometry. I, volume 48 of Ergebnisse der Mathematik und
ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics
and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics]. Springer-Verlag, Berlin,
2004. Classical setting: line bundles and linear series.
[Laz04b] Robert Lazarsfeld. Positivity in algebraic geometry. II, volume 49 of Ergebnisse der Mathematik und
ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics
and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics]. Springer-Verlag, Berlin,
2004. Positivity for vector bundles, and multiplier ideals.
[Lip84]
Joseph Lipman. Dualizing sheaves, differentials and residues on algebraic varieties. Astérisque,
(117):ii+138, 1984.
[Mig98]
Juan C. Migliore. Introduction to liaison theory and deficiency modules, volume 165 of Progress in
Mathematics. Birkhäuser Boston Inc., Boston, MA, 1998.
[Niu11]
Wenbo Niu. A bound for the Castelnuovo-Mumford regularity of log canonical varieties. J. Pure
Appl. Algebra, 215(9):2180–2189, 2011.
[PS74]
C. Peskine and L. Szpiro. Liaison des variétés algébriques. I. Invent. Math., 26:271–302, 1974.
Department of Mathematics, Purdue University, West Lafayette, IN 47907-2067, USA
E-mail address: [email protected]
21
| 0math.AC
|
Curve Reconstruction via the Global Statistics of Natural Curves
Ehud Barnea and Ohad Ben-Shahar
Dept. of Computer Science, Ben-Gurion University
Beer-Sheva, Israel
arXiv:1711.03172v2 [cs.CV] 18 Nov 2017
{barneaeh, ben-shahar}@cs.bgu.ac.il
20
Abstract
10
0
-10
Reconstructing the missing parts of a curve has been
the subject of much computational research, with applications in image inpainting, object synthesis, etc. Different
approaches for solving that problem are typically based on
processes that seek visually pleasing or perceptually plausible completions. In this work we focus on reconstructing the
underlying physically likely shape by utilizing the global
statistics of natural curves. More specifically, we develop
a reconstruction model that seeks the mean physical curve
for a given inducer configuration. This simple model is both
straightforward to compute and it is receptive to diverse additional information, but it requires enough samples for all
curve configurations, a practical requirement that limits its
effective utilization. To address this practical issue we explore and exploit statistical geometrical properties of natural curves, and in particular, we show that in many cases
the mean curve is scale invariant and oftentimes it is extensible. This, in turn, allows to boost the number of examples
and thus the robustness of the statistics and its applicability. The reconstruction results are not only more physically
plausible but they also lead to important insights on the reconstruction problem, including an elegant explanation why
certain inducer configurations are more likely to yield consistent perceptual completions than others.
-20
-30
-40
-50
-60
-70
-80
-20
0
20
(a)
40
60
80
100
60
80
100
(b)
20
20
10
10
0
0
-10
-10
-20
-20
-30
-30
-40
-40
-50
-50
-60
-60
-70
-70
-80
-80
-20
0
20
40
(c)
60
80
100
-20
0
20
40
(d)
1. Introduction
Figure 1. Curve reconstruction via global statistics of natural
curves. (a) To measure natural curve prior for a given relative
inducer configuration, curve fragments having the desired relative inducers (end points + tangent orientation) are collected from
ground truth curves labeled by humans [8]. (b) The curves are then
normalized to the same frame of reference and sampled along their
arc length with n equally spaced points. In this example n = 5
and to avoid clutter only 4 curves are shown. (c) Each point i
along the arc length of the reconstructed curve is calculated as the
mean of the corresponding points in the dataset, or more abstractly,
as the expected value of the distribution of all ith points of curves
that match the same inducers. (d) While some variance is observed
among the centers of all 245 curves, the distribution appears rather
tight, somewhat anisotropic, and approximately normal, suggesting that the expected value nicely represents the likely outcome.
The reconstruction of visual curves is the process of filling in curve fragments that are completely unobservable due
to occlusion or adversarial acquisition conditions. When attempting this task one could pursue one of two different
goals, either the original physical shape or the perceived
one. (Note that oftentimes these two curves are quite different.). Different models for the generation of missing parts
of curves were indeed suggested [25, 13, 18, 11, 2, 10, 19,
33] and employed in various applications such as image inpainting [26, 23, 24, 29], and computer graphics [10, 9].
With much of the theoretical work done in the context
of curve completion that supports perceptual plausibility,
most studies have focused on defining different shape criteria for the generation of visually pleasing curves (for example those with minimal total curvature [25, 18, 11] or minimal change of curvature [13]) rather than relying on any
measurable prior on natural curves. Additionally, the computation of these curves entails a minimization procedure
that usually requires an iterative scheme that is susceptive
to local minima and cannot be easily expanded to include
1
additional information (such as the curvature at the visible
end points or the shape of the occluder, among others). Critically, no consensus exists in the literature regarding which
completion criteria are preferred, and evaluation is hardly
performed vis-a-vis physical objects (whose shape one tries
to reconstruct) or perceptual findings.
Somewhat differently, and avoiding arbitrary shape constraints as above, Ben-Yosef and Ben-Shahar suggested a
biologically plausible theory based on an abstraction of the
primary visual cortex and a least action criterion [2]. This
was recently elaborated as a constrained Elastica model in
the unit tangent bundle with emergent properties like sensitivity to curvature [1]. Other works employed probabilistic schemes such as first-order and multi-scale higher-order
Markov models [31, 19]. These models too, however, suffer
from some of the issues discussed above.
In this work we suggest to incorporate a stronger and
more realistic prior for reconstruction by following the
global shape statistics of natural curves. Like others, we
will assume that the reconstruction is performed between
two inducers – a pair of end points, their tangent orientation,
and possibly additional information about the observed part
of the curve as it penetrates the occluder. Simply put, we
first sample from the distribution of natural curves that corresponds to each relative inducer configuration. This operation can be done empirically from annotated datasets [8].
Representing each curve as a discrete set of n points distributed uniformly along the arc length, we then extract
the point-wise mean of the collected curves, from which
a “mean curve” can be compiled. In contrast to other methods employing only local statistics of adjacent point pairs,
this computation examines curves in their entirety, thus employing global statistics.
Even when lacking any local consideration, this process
generates surprisingly smooth and visually pleasing curves,
as shown in Fig. 1. And yet, for many inducer configurations, especially for longer curves, the number of observed
samples may be small, prohibiting robust statistics and thus
proper reconstruction. To alleviate this problem we show
that the mean curve is scale invariant in many cases and further investigate its extensibility, allowing to collect and generate many more samples for each configuration, providing
a significantly more robust reconstruction process.
In addition to their pleasing appearance, the generated
curves and our proposed completion model enjoy several favorable properties. By design, the resultant curves strongly
represent the distribution of physical image curves and thus
the structure of objects in natural images. By definition of
the expected value, these curves are also “closest” to all
dataset curves with the same inducer configuration. The
proposed method can also easily extend to include the curvature at the end points, the shape of the occluder, or any
other information, by simply conditioning the measured
probability distribution (and thus the reconstruction computation) on the desired properties or conditions. Furthermore, the proposed reconstruction approach provides novel
insights into the reconstruction problem. For example, we
show that for “convenient” configurations (when the inducers are relatable [12] or even just “facing” each other)
the distribution of curves in natural images is quite narrow
and the mean curve is similar to most ground-truth curves,
providing a reconstruction that closely matches the ground
truth physical curves in most cases. However, curves with
“abnormal” or “difficult” inducer configurations (e.g., when
the inducers face away from each other) exhibit much
greater variance, which implies that any reconstruction, regardless of the underlying principle employed, is likely to
generate a curve that does not reproduce the original shape.
Such an analysis suggests that in these cases additional information is strongly recommended, as well as a model that
can exploit it (like the one suggested here).
2. Prior art
Different methods have been suggested for the generation of curves, with either the goal of completing curves
when no information is known apart for the inducer configuration (usually the result of occlusions), or when the
curve part to complete is visible but difficult to discern
due to noisy or faint edge appearance. To the best of our
knowledge no work has focused on reconstruction of missing curve parts to match the physical reality, and so our review of the prior art focuses on methods obtaining visually
pleasing or perceptually plausible completions over large
gaps.
Much of the previous work on curve completion is based
on defining a set of axioms or properties that the completed
curve should satisfy, in what has been dubbed the axiomatic
approach [2, 1]. Given such axioms, a model and computational approach that generate such curves is then developed. Among the first is the Biarc model suggested by
Ullman [25] that seeks curves of least curvature that are
also smooth, isotropic, and extensible. Loosening some of
these constraints, the Biarc model constructs the completed
curve using two circular arcs (that meet at a curvature discontinuity). Taking least curvature strictly later gave rise
to the Elastica model [18, 11, 3, 22, 30, 16]. A different
property, suggested by Kimia et al. [13], is the minimization of change in curvature, resulting in a family of visually pleasing curves known as Euler Spirals. This completion model has been improved later in various follow up
papers [32, 33, 27].
Employing a different approach, in what has been
dubbed the “mechanistic approach” [2, 1], some works define probabilistic models for the generation of maximum
probability curves or for the calculation of point-wise probabilities of belonging to a curve. In Williams and Jacobs
2
3. Global Curve Statistics - Analysis and Reconstruction
[31], for example, the completion is described as the most
probable random walk between the curve inducers. Being
a first order Markov model, this work can be seen as minimizing some local probabilistic measure along the curve
while ignoring long range interactions. A similar model
was also employed together with local detection responses
for grouping of visible edges [6]. In a later work, it was
shown that this Markov assumption does not comply with
curvature statistics of natural curves, and so a higher order
model was suggested incorporating multiple scales for the
task of edge boundary detection [19].
Given an input inducer configuration C = (I1 , I2 ) for
inducers I1 = (Ix1 , Iy1 , Iθ1 ) and I2 = (Ix2 , Iy2 , Iθ2 ) at locations I1xy = (Ix1 , Iy1 ), I2xy = (Ix2 , Iy2 ) and orientations Iθ1 ,
Iθ2 , we seek to generate a curve represented as a set of n
points α(C) = {x1 , ..., xn } that closely matches the way
physical visual (i.e., image) curves behave between such
two inducers. That latter behavior will be measured from
a large collection of observable image curves.
For this data driven reconstruction we suggest to employ
the distribution of natural curves for a given configuration:
Seeking to benefit the advantages of both approach types
while avoiding arbitrary or unfounded assumptions about
the shape of the completed curves, Ben-Yosef and BenShahar [2] moved the computation framework from the image plane to a space that abstracts the primary visual cortex where the visual system allegedly performs the curve
completion task. With a proper abstraction the problem was
then considered in the unit tangent bundle and the completion criteria become ones that are inspired by how biological neural circuits may behave. The simple models that
emerged, first one that seeks the shortest path in the tangent bundle [2] and later the tangent bundle Elastica [1],
are very simple to describe (though not necessarily to solve)
and provide visually pleasing curves. Like in previous models, no perceptual validation or evaluation against ground
truth were yet performed.
P (x1 , ..., xn |I1 , I2 ) .
(1)
Previous works that employed the statistics of curves sought
the most likely completion, which they were able to calculate by making assumptions that enable to express this distribution with smaller functions that capture local statistics
of shape. Here, we suggest a global approach of estimating
the mean curve instead, and as observed in Fig. 1d, it may
also provide a good estimation of the most likely curve.
In the following we describe the data collection and completion process, as well as its practical difficulties to generate proper reconstructions when the inducers are too far,
among other conditions. To make the process more generally applicable, we show that the mean curve possesses
certain properties that facilitate more stable and visually
appealing completion even under those challenging conditions.
While most of the prior art in curve completion sought
the perceptual completed curve, it is also reasonable to seek
the likely physical curve between inducers, a task that lends
itself to properties of real word image curves and natural
image statistics. Indeed, some prior art did employ various
statistics for the estimation of different aspects of natural
images, showing that they follow properties such as scale
invariance [20], or that the lengths of curve segments between two locally maximal curvature points follow a power
low [19]. Focusing on the co-occurrence of oriented edges,
a strong preference of curve elements to co-linearity was
shown by employing the statistics of edge pairs [7], and
further high-order structure was shown by examining the
statistics of edge triplets [14]. In this paper we combine
both schools, i.e., we explore natural image curve statistics
for the task of long range curve reconstruction that best fits
the physical reality (rather than, for example, the perceptual
process or its outcomes).
3.1. Collection and Representation of the Prior
As exemplified in Fig. 1, the basic reconstruction of a
curve between a given inducer configuration C follows several steps. Note that most of the computations can and are
done just once as preprocessing and need not repeat themselves for each reconstruction query, which essentially can
be answered by a lookup operation.
As a first step, we collect a set of natural image curve
fragments with configurations that are similar to C. To that
end, we employ the existing Curve Fragment Ground-Truth
Dataset (CFGD) [8] that contains ground-truth annotations
of perceived image curves collected from three different annotators on highly varied scenes from the Berkeley Segmentation Dataset (BSDS) [15], leading to ~40K perceived
curves represented as a set of ordered points. Via proper
processing we consider any sub-curve of each CFGG curve
as a fragment in its own right, an operation implemented by
choosing any possible pair of points along a given CFGD
curve as starting and ending points of a fragment. This process provides a total of ~19M fragments.
Next, to make these fragments useful for our statistical
analysis, we represent them by the relative configuration of
their inducers:
Our investigation of the statistics of curves follows the
framework of active shape models [4], employing the mean
and variance of point correspondences and allowing a simpler investigation of shape variability relative to more complex methods for averaging curves [21].
3
Cr = p = (px , py , pθ ) = (Ix1 − Ix2 , Iy1 − Iy2 , Iθ1 − Iθ2 )
and rotated back to configuration C, providing α(C). The
reconstructed curve as a whole is obtained by interpolating
these points, either linearly (as in Fig. 1c) or using more
elaborate methods.
As exemplified in Fig. 1, this process often provides intuitive, visually pleasing, and veridical reconstructions. Such
results come as no surprise once we examine the nearly
normal distribution of the curve sample points, for example the one for the central curve point depicted in Fig. 1d.
Note that this happens despite the observation that the center point, being “furthest” away from both inducers along
the arc length, is the one where curves are most likely to
show the greatest variation.
That said, and as Fig. 2 shows, the quality of results
highly depends on the number of ground truth curve fragments that are observed in the prior (denoted as darker blue
curves), or more formally, on the power of the statistics
for any given inducer configuration. Empirically, reconstructions appear noisy when sample size is small, a situation more common for example when inducers are oriented
away from each other, especially for larger gaps (i.e. long
curves). Clearly, one way to address this difficulty is collecting more data and annotating many more curves, two
operations that require much human labor. However, one
can do much better even with existing data (in our case,
the CFGD collection [8]) and indeed, to alleviate the problem we now examine and exploit certain properties of the
mean curve that allow “sharing” fragments between configurations, boosting the number of examples and the power
of the statistics, and thus greatly improve the results in the
“problematic” cases also.
(2)
where pxy = (px , py ) and pθ are the location and orientation of the second inducer relative to the first one. In
particular, we translate and rotate the curve fragment with
configuration C such that the location of its first inducer
I1xy = (Ix1 , Iy1 ) overlaps the origin of the coordinate system,
and its orientation Iθ1 coincides with the X axis. This operation effectively assumes invariance of shape to translations
and rotations and allows a representation of the transformed
fragments by the second inducer only. Assuming also invariance to reflection, we finally reflect all fragments with
positive relative elevation (i.e., py > 0) so all curves end
in the bottom half of the coordinate system. Finally, for a
quick lookup of fragments later on, they are stored according to their relative configuration p. All this pre-processing
is exemplified in Fig. 1a,b.
To extract the probability distribution of natural curves
that match C, we now collect all curve fragments that fit
the corresponding relative inducer configuration, allowing
some tolerance in position and orientation. Formally, we include all curve fragments with relative configuration Cer =
kp −qxy k
< t1 and dπ (pθ , qθ ) < t2 , where dπ
q, such that xy
kpxy k
is the angular distance between two angles. That is, we include fragments if their inducer configurations deviate from
pxy only marginally in normalized distance and in relative
orientation, where the margin is determined by two predefined small thresholds t1 and t2 . Note that since the collected fragments have similar but not identical configurations, they are further finely transformed to allow meaningful pooling in the reconstruction step. In practice, they are
just slightly scaled so their end point qxy exactly matches
pxy . In the interest of space, the few trivial technical details
of this step are listed in the supplementary materials. Note
that small discrepancies in orientations qθ and pθ may still
persist but do not affect the reconstruction process.
3.2. Basic Curve Reconstruction
m(Cr )
X
1
αi (Cr ) ,
m(Cr ) i=1
>300
200
200
100
100
0
(a)
As the reconstruction of missing curve fragments is
based on the “mean curve”, the latter must be represented
in a convenient way. For configuration C with relative configuration Cr , we represent transformed fragments in the
collected data as a set of n points αi (Cr ) = {x1 , ..., xn }
sampled from the original points such that the arc length
between each two adjacent points in αi (Cr ) is equal (blue
points in Fig. 1b). The mean curve α(Cr ) is estimated as:
α(Cr ) =
>300
0
(b)
Figure 2. Basic reconstructions of different inducer configurations
based on the mean curve. All configurations share one inducer
(red) but have different second inducers (green), in this case at the
same distance (80 pixels). Curve colors represent the number of
corresponding fragments that were found in the dataset and thus
used during reconstruction. Note how larger sample sets allow
more intuitive and visually pleasing reconstructions.
3.3. Boosting via Invariance to Scale
(3)
If curve fragments (or curves in general) were scale
invariant, curves (either physical or reconstructed) would
scale similarly to their inducer configuration. Hence, if we
where m is the number of available fragments αi for relative configuration Cr (Fig. 1c). Finally, α(Cr ) is translated
4
could obtain evidence to that effect, the distribution of curve
fragments at one scale could help determine the reconstructions at other scales also and the scarcity of samples for certain configurations, especially for longer curves, would be
alleviated. Clearly, with such a capacity, one would even be
able to generate reconstructions of arbitrarily long curves at
arbitrary scales even if such curves were never observed in
the prior. In this section we show that the mean curve is
indeed scale invariant in most cases, facilitating data boosting and better reconstructions in general. We note that these
findings are congruent with earlier natural image statistics
about edge elements and image curves [20, 7], though here
we show them for the shape of whole curves. For the interest of space we focus the presentation on the curve center
points.
Loosely speaking, the reconstructed curve for some inducer configuration is invariant to scale if it is similar to the
reconstruction at any other scale after proper scaling of the
configuration. Since scaling is analogous to changing viewing distance, and since such an operation does not change
the observed orientation, scaling essentially affects only the
distance between inducers, and the “scale” s of a curve fragment is determined by that distance alone.
With this in mind, to examine the scale invariance hypothesis we explore the effect of scale on the reconstruction
at a base relative configuration Cr = p with nominal scale 1
(i.e., with kpxy k = 1). At each scale s, fragments for relative configuration Cer = (spxy , pθ ) are collected and transformed as described in Sec. 3.1, and then brought to a common scale by transforming them once more such that the
distance between their inducers is 1. Focusing on the center
points of the transformed fragments from the original scale
s, let µs and σs be their mean and standard deviation. We
summarize the information across scales as the standard deviation of the means µs (see Fig 3b ) and the expected value
of the standard deviations σs (Fig 3c).
Since the STD of µs is the mean difference between the
reconstructed central curve points across scales, one would
expect it to be smaller the more invariant the curves to scale.
In Fig. 3 we examine this difference in the reference scale of
1. As shown, for “normal” inducer configurations the STD
of µs is very small compared to the scale, suggesting that
the mean curve is indeed invariant to scale in these cases.
For other, more “abnormal” configurations we see greater
STD, suggesting that the calculated mean curve is different
across scales. Note that this in itself does not necessarily indicate that the mean curve is scale variant in such difficult
configurations, for the average σs in these cases is large as
well (Fig. 3c), requiring more samples to guarantee an accurate estimation of the mean according to Chebyshev’s inequality applied for the empirical mean [17].
Following this analysis, we now leverage the invariance
of the mean curve to scaling (assuming scale invariance also
when it could not be asserted) by extracting the distribution of natural curves that match Cr = p from fragments
Cer = q that are close to any of its scaled versions. Formally,
we include all curve fragments with relative configuration
Cer = q, such that the angular distance between vectors
pxy ,qxy and orientation distance dπ (pθ , qθ ) are smaller
than predefined thresholds t1 ,t2 . To quickly collect such
curves, fragments are stored ahead of time according to the
angle to qxy from the X axis, and their orientation qθ .
The process of utilizing curves at all scales greatly improves the results and now provides intuitive, visually pleasing, and more veridical reconstructions (see Fig. 4b) for
many more inducer configurations (Fig. 2). Another favorable property is the ability to generate curves regardless
of the distance between given inducers. That said, it can
be seen that the number of examples, for some “difficult”
configurations remains small and thus the reconstruction in
such cases is still noisy, a problem that would possibly be
alleviated with the inclusion of additional samples. To this
end, we now turn to explore yet another possible property
of curves known as extensibility.
3.4. Boosting via Midway Extensibility
As discussed above, scale invariance alone does not suffice to increase the number of samples and qualify as a robust prior in all cases. Typically, the cases that remain problematic have their inducers face away from each other and
so the mean curve bends rapidly near the inducers and keeps
a straighter shape half way from them (Fig. 4b). Due to
this behavior, another inducer placed in the center of such
curve (and oriented along its tangent) would represent much
smaller orientation relative to either of the inducers, effectively better “facing” both of them and thus forming less difficult configurations with either. As observed in Figs. 2,4b,
such simpler configuration tend to have many more dataset
samples and smoother reconstructions. Therefore, generating a curve by reconstructing two sub-curves from the original inducers to the curve’s center is likely to provide a more
robust, visually pleasing and veridical reconstruction.
Such a scheme, which we refer to as midway extensibility, is a special case of general extensibility, a property
sought after in axiomatic approaches to curve completion
(e.g., [25], cf. [2, 1]). A reconstruction approach is considered extensible if for any pair of oriented inducers I3 ,I4
extracted from a reconstructed curve α between inducers
I1 ,I2 , it generates a curve β that is identical to α between
I3 ,I4 . Here, we explore the extensibility of the mean curve
by its center point, and define a reconstruction scheme as
midway extensible if for any generated curve α between
inducers I1 ,I2 , α is identical to the curve β = β1 ⊕ β2 ,
calculated as the concatenation of curves β1 and β2 that are
generated between inducers I1 ,I3 and I3 ,I2 , where I3 is the
inducer extracted from the center of α.
5
Horizontal inducers
STD of µs
Average of σs
(b)
(c)
𝜃$
𝜃"
𝑦$
𝑦"
(a)
Figure 3. Empirical analysis of scale invariance for a subset of inducer configurations. (a) Considered are all pairs of horizontal inducers
with two arbitrary orientations. (b) A map of the STD of µs of the curve center point across different scales shows very stable behavior
when the inducers are “facing” each other with gradually decaying stability as the inducers “face” away from each other. (c) The average
of σs across scales similarly shows very stable behavior when the inducers roughly “face” each other, gradually turning unstable as they
“face” away from each other. The two maps in conjunction imply scale invariance for facing inducers, and no conclusive insights otherwise.
To investigate the property of midway extensibility in
the context of our mean curve we employ the scale invariant procedure described in Sec.3.3 to generate a curve β
(Fig. 4c) for each curve α in Fig. 4b by generating two
scale invariant curves β1 , β2 from the original inducers to
the center of α. While this was done for all inducer configurations, to determine whether the mean curve exhibits
midway extensibility we compare the corresponding reconstructions α and β only in cases where α was generated
with many observations, as these are the only cases where
the reconstruction is likely to be precise in the first place.
A careful inspection of their shapes reveals that while some
differences between such corresponding curves can be observed, they are typically insignificant relative to the scale
of the curve. To measure the deviation from extensibility we
calculate the Fréchet distance [5] relative to the distance between inducers for corresponding curves in Fig. 4 and in the
complementary figure in the supplementary material, covering the range of relative configurations. Considering only
curves for which more than 400 fragments were observed,
the maximal distance between such curves is 4.9% of the
distance between inducers, and the mean is 2.5%. Thus,
curves generated by employing midway extensibility do not
vary much from their counterparts that were accurately reconstructed without extensibility.
Establishing such empirical midway extensibility, we reiterate that its sought-after effect is for cases where relative
configurations enjoy only few samples in the prior. Checking the reconstruction facilitated by the two sub curves now
shows significant improvement in such cases (compare the
relevant cases in Fig. 4b,c).
ity when the number of samples is lacking), with Libcornu’s
implementation of Euler Spiral [32, 28], the latter being a
prominent method for the completion of curves.
Reconstruction for all selected curve configurations can
be seen in Fig. 4 and in the supplementary material. Results
are shown at a single scale and for the lower half visual field,
a consequence of the scale and mirror reflection invariance
properties of both models. Comparing the results visually
suggests that the Euler spiral, being an analytic regularized
model, is generally smoother, but it often provides reconstructions that appear unnatural and/or far from veridical (in
a sense of corresponding to the actual physical curve as represented by the prior). When it comes to reconstruction of
missing physical curve fragments our data driven approach
thus provides an intrinsic advantage. A demonstration to
that effect can be shown with the real objects in Fig. 5.
4.1. Reconstruction Evaluation
In addition to the qualitative results discussed above, we
examined the accuracy of the mean curve as a reconstruction scheme for natural occluded curves. In our representation, the mean curve is, by construction, the one “closest”
to all observable n-point natural planar curves (abstractly
represented as vectors in 2n dimensional space). However,
uniquely shaped curves do occur in nature and they may be
far from the mean. The extent to which such curves can
be reconstructed, and what is the information or the prior
needed to facilitate such reconstructions, remain to be studied. That said, we still seek to obtain a quantitative measure
of a method’s ability to properly reconstruct fragments of
natural curves, and to do so we first split the original dataset
into training and test images, the latter comprises 10% of
the dataset. No curve fragment from the test set was used
when extracting the prior and computing the mean curve for
reconstruction while no curve from the train set was used for
4. Experimental Results
We compare the reconstructions based on the mean
curve (assuming scale invariance and employing extensibil6
Euler Spiral
Scale invariant (SI) Mean
1
SI Extensible Mean
>300
>300
200
200
100
100
0
>300
0
>300
200
200
100
100
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0
(a)
(b)
(c)
Figure 4. Reconstruction results based on (a) Euler spiral, (b) scale invariant mean curve, and (c) extensible and scale invariant mean curve,
where darker curve colors represent a larger number of samples from which the curve was calculated. Red circles in (c) mark the points
in which two reconstructions are connected to a single curve. As exemplified, the final mean curve is not only data driven and thus more
veridical, it is also more compact and it exhibits much smoothness even though no regularization or smoothing was performed.
Figure 5. Example results of reconstruction over large gaps of possibly occluded curves based on our approach (green) and Euler spiral
(magenta). Image parts below the curves are assumed to be completely occluded and are displayed for comparison with the ground-truth.
queries or as ground truth for performance evaluation.
More specifically, a set of 5000 curve fragments and their
inducers were randomly selected from the test images such
that fragments are uniformly distributed by their scale (as
the distance between their inducers) to better represent a
scenario such as the reconstruction of curves over randomly
sized gaps. For each ground-truth curve α∗ from this test
set, we then generated the mean curve α = {y1 , ..., yn } that
matches its inducer configuration (I1 , I2 ) while using the
fragments and the prior collected only from train images.
We then evaluate the relative reconstruction error (RRE) of
the generated curve α as the Fréchet distance d [5] between
both curves relative to the distance between their inducers:
RRE =
d(α∗ , α)
I1xy − I2xy
.
Reconstruction results are shown in Fig. 6, where each
point in the X axis represents an RRE threshold and its
corresponding value in the Y axis represents the ratio of
accurately reconstructed curves (ARC) for which the error is less than that threshold. Overall results are summarized by the area under the curve (AUC), calculated as the
average ARC at 101 RRE points from 0 to 1 (where the
limit of 1 RRE was chosen arbitrarily). As can be seen,
the mean curve provides more accurate reconstructions than
those based on Euler Spiral, with AUC scores of 0.887 and
0.876, respectively. The only cases where the Euler Spiral
is marginally more accurate are those with extremely low
threshold RRE < 0.017 which, as we verified, comprises
of straight lines only.
(4)
While one may argue that the differences in performance
7
Ratio of accurately reconstruced curves
significantly more accurate than Euler Spiral, with an AUC
of 0.492. Clearly, the mean curve reconstruction is far more
veridical in such cases.
Correct reconstructions per permitted error
1
Euler spiral
Mean curve
0.8
5. Conclusions and Future Work
0.6
We have suggested a method for the reconstruction of
missing or occluded curves by employing the global statistics of natural contours. At the foundation of our approach
is the generation of the mean curve, estimated as the mean
of the distribution of natural image curves along their arc
length. As we have shown, the estimated mean curve provides an intuitive, visually pleasing, and most importantly, a
more veridical reconstruction. A main disadvantage of this
procedure, namely the required number of curve fragments
sampled to represent the statistics of natural curves loyally,
was overcome by exploring and employing certain properties of curves. We have shown that the mean curve is invariant to scale in many cases and that employing this property
for sharing fragment samples across scales provides better
reconstructions. We also explored midpoint extensibility of
the mean curve and have shown that incorporating this property improves the reconstructed curves when the statistical
power of the prior is particularly low.
0.4
0.2
0
0
0.2
0.4
0.6
0.8
1
Relative reconstruction error
Figure 6. Reconstruction evaluation showing the ratio of accurate
reconstructions for each accuracy threshold defined according to
the relative reconstruction error. As shown, the graph of the mean
curve dominates over the graph of Euler Spiral completions.
Correct reconstructions per permitted error
Ratio of accurately reconstruced curves
1
0.8
0.6
The data in Fig. 1d shows that for the inducer configuration exemplified, the observed curve centers lie in a small
region relative to the distance between the two inducers, exhibiting relatively small variance. In such cases, the center of the mean curve is therefore close to most data points
(i.e., real natural curves), and so it could be considered a
correct reconstruction (i.e. with low enough error). Unfortunately, however, this is not always the case, as the mean
variance could grow larger for certain inducer configurations, as already shown in Fig. 3b (e.g., for inducers facing
away from each other). In such cases, where the variance of
natural curves is considerable, no single curve can be close
to most real curves, regardless of the reconstruction method
that generated it, and thus cannot be considered correct in
most cases.
0.4
0.2
Euler spiral
Mean curve
0
0
0.2
0.4
0.6
0.8
1
Relative reconstruction error
Figure 7. Evaluation of reconstruction results for curves with difficult inducer configurations, showing the ratio of accurately reconstructed curves for each accuracy threshold defined according
to the relative reconstruction error. The mean curve provides many
more accurate reconstructions than the Euler Spiral.
in Fig. 6 are not big, it is also suspected that many of the
details are washed by the frequent cases of “convenient”
inducer configurations where both approaches yield similar
results. Indeed, observing the different reconstructions such
as in Fig. 4, it can be seen that the mean curve and Euler
Spiral differ more for the “difficult” configurations that are
underrepresented in our dataset and in the prior in general.
Therefore, we conducted a focused evaluation with an additional random set of 1000 curves, all with “difficult” configurations such that θ1 > π2 and θ2 < π2 according to the representation in Fig. 3a. Evaluation of reconstruction results
over this difficult set are presented in Fig. 7. This time, the
mean curve method obtains an AUC score of 0.621 and is
An immediate consequence of the above is that to obtain correct reconstruction one must reduce the variance (or
variability) of observed data that fit the conditions for which
reconstruction is attempted. Since the prior in the natural
world is fixed, the only way of achieving reduced variance
is by introducing additional constraints on the curve fragments that are relevant for a reconstruction query, i.e., by
requiring more information about the inducers or the reconstruction scenario as a whole. Such typical information may
include the curvature of the inducers, their polarity, or the
shape of the occluder, etc, and incorporating such considerations into the computational scheme is part of our ongoing
and future research.
8
References
[20] D. Ruderman and W. Bialek. Statistics of natural images:
Scaling in the woods. Phys. Rev. Lett., 73(6):814–818, 1994.
3, 5
[21] T. B. Sebastian, J. J. Crisco, P. N. Klein, and B. B. Kimia.
Constructing 2d curve atlases. In Mathematical Methods in
Biomedical Image Analysis, 2000. Proceedings. IEEE Workshop on, pages 70–77. IEEE, 2000. 3
[22] E. Sharon, A. Brandt, and R. Basri. Completion energies and
scale. IEEE Trans. Pattern Anal. Mach. Intell., 22(10):1117–
1131, 2000. 2
[23] T. Shibata, A. Iketani, and S. Senda. Image inpainting based
on probabilistic structure estimation. Proc. ACCV, pages
109–120, 2011. 1
[24] J. Sun, L. Yuan, J. Jia, and H.-Y. Shum. Image completion
with structure propagation. ACM Transactions on Graphics
(ToG), 24(3):861–868, 2005. 1
[25] S. Ullman. Filling-in the gaps: The shape of subjective contours and a model for their creation. Biol. Cybern., 25(1):1–
6, 1976. 1, 2, 5
[26] V. Voronin, V. Marchuk, A. Sherstobitov, and K. Egiazarian. Image inpainting using cubic spline-based edge reconstruction. In IS&T/SPIE Electronic Imaging, pages
82950I–82950I. International Society for Optics and Photonics, 2012. 1
[27] D. J. Walton and D. S. Meek. An improved euler spiral algorithm for shape completion. In Computer and Robot Vision, 2008. CRV’08. Canadian Conference on, pages 237–
244. IEEE, 2008. 2
[28] D. J. Walton and D. S. Meek. G1 interpolation with a single
cornu spiral segment. Journal of Computational and Applied
Mathematics, 223(1):86–96, 2009. 6
[29] Y. Wei and S. Liu. Domain-based structure-aware image inpainting. Signal, Image and Video Processing, 10(5):911–
919, 2016. 1
[30] I. Weiss. 3d shape representation by contours. CVGIP,
41:80–100, 1988. 2
[31] L. Williams and D. Jacobs. Stochastic completion fields: A
neural model of illusory contour shape and salience. Neural
Comp., 9(4):837–858, 1997. 2
[32] J. Xu, M. D. Collins, and V. Singh. Incorporating topological constraints within interactive segmentation and contour
completion via discrete calculus. In CVPR, 2013. 2, 6
[33] H. Zhou, J. Zheng, and X. Yang. Euler arc splines for curve
completion. Computers & Graphics, 36(6):642–650, 2012.
1, 2
[1] O. Ben-Shahar and G. Ben-Yosef. Tangent bundle elastica
and computer vision. IEEE Trans. Pattern Anal. Mach. Intell., 37:161–174, 2015. 2, 3, 5
[2] G. Ben-Yosef and O. Ben-Shahar. A tangent bundle theory for visual curve completion. IEEE Trans. Pattern Anal.
Mach. Intell., 34(7):1263–1280, 2012. 1, 2, 3, 5
[3] A. Bruckstein and A. Netravali. On minimal energy trajectories. CVGIP, 49(3):283–296, 1990. 2
[4] T. F. Cootes, C. J. Taylor, D. H. Cooper, and J. Graham.
Active shape models-their training and application. CVIU,
61:38–59, 1995. 3
[5] T. Eiter and H. Mannila. Computing discrete fréchet distance. Technical report, Tech. Report CD-TR 94/64, Information Systems Department, Technical University of Vienna, 1994. 6, 7
[6] P. Felzenszwalb and D. McAllester. A min-cover approach
for finding salient curves. In Computer Vision and Pattern
Recognition Workshop, 2006. CVPRW’06. Conference on,
pages 185–185. IEEE, 2006. 3
[7] W. Geisler, J. Perry, B. Super, and D. Gallogly. Edge cooccurrence in natural images predicts contour grouping performance. Vision Res., 41(6):711–724, 2001. 3, 5
[8] Y. Guo and B. Kimia. On evaluating methods for recovering image curve fragments. In Computer Vision and Pattern
Recognition Workshops (CVPRW), 2012 IEEE Computer Society Conference on, pages 9–16. IEEE, 2012. 1, 2, 3, 4
[9] G. Harary and A. Tal. The natural 3d spiral. In Computer
Graphics Forum, volume 30, pages 237–246. Wiley Online
Library, 2011. 1
[10] G. Harary and A. Tal. 3d euler spirals for 3d curve completion. Computational Geometry, 45(3):115–126, 2012. 1
[11] B. Horn. The curve of least energy. ACM Trans. Math. Software, 9(4):441–460, 1983. 1, 2
[12] P. Kellman and T. Shipley. A theory of visual interpolation in
object perception. Cognitive Psychology, 23:141–221, 1991.
2
[13] B. Kimia, L. Frankel, and A. Popescu. Euler spiral for shape
completion. In Proc. POCV, 1999. 1, 2
[14] M. Lawlor and S. W. Zucker. Third-order edge statistics:
Contour continuation, curvature, and cortical connections. In
NIPS, 2013. 3
[15] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database
of human segmented natural images and its application to
evaluating segmentation algorithms and measuring ecological statistics. In ICCV, pages 416–423, 2001. 3
[16] W. Mio, A. Srivastava, and X. Liu. Contour inferences for
image understanding. Int. J. Comput. Vision, 69(1):137–144,
2006. 2
[17] M. Mohri, A. Rostamizadeh, and A. Talwalkar. Foundations
of machine learning. MIT press, 2012. 5
[18] D. Mumford. Elastica and computer vision. In B. Chandrajit, editor, Algebric Geometry and its applications. SpringerVerlag, 1994. 1, 2
[19] X. Ren and J. Malik. A probabilistic multi-scale model for
contour completion based on image statistics. ECCV, pages
312–327, 2002. 1, 2, 3
9
| 1cs.CV
|
Graph-Guided Banding of the Covariance Matrix
arXiv:1606.00451v2 [stat.ME] 15 Feb 2018
Jacob Bien∗
Data Sciences and Operations
Marshall School of Business
University of Southern California
Abstract
Regularization has become a primary tool for developing reliable estimators of the covariance
matrix in high-dimensional settings. To curb the curse of dimensionality, numerous methods
assume that the population covariance (or inverse covariance) matrix is sparse, while making
no particular structural assumptions on the desired pattern of sparsity. A highly-related, yet
complementary, literature studies the specific setting in which the measured variables have a
known ordering, in which case a banded population matrix is often assumed. While the banded
approach is conceptually and computationally easier than asking for “patternless sparsity,” it
is only applicable in very specific situations (such as when data are measured over time or
one-dimensional space). This work proposes a generalization of the notion of bandedness that
greatly expands the range of problems in which banded estimators apply.
We develop convex regularizers occupying the broad middle ground between the former
approach of “patternless sparsity” and the latter reliance on having a known ordering. Our
framework defines bandedness with respect to a known graph on the measured variables. Such
a graph is available in diverse situations, and we provide a theoretical, computational, and
applied treatment of two new estimators. An R package, called ggb, implements these new
methods.
Keywords: covariance estimation, high-dimensional, hierarchical group lasso, network
1
Introduction
Understanding the relationships among large numbers of variables is a goal shared across many scientific areas. Estimating the covariance matrix is perhaps the most basic step toward understanding
these relationships, and yet even this task is far from simple when sample sizes are small relative to
the number of parameters to be estimated. High-dimensional covariance estimation is therefore an
active area of research. Papers in this literature generally describe a set of structural assumptions
that effectively reduces the dimension of the parameter space to make reliable estimation tractable
given the available sample size. Some papers focus on eigenvalue-related structures and shrinkage
(e.g., Johnstone 2001, Fan et al. 2008, Won et al. 2013, Donoho et al. 2013, Ledoit & Wolf 2004,
among many others); others assume that correlations are constant or blockwise constant (such as
in random effects models); and a very large number of papers introduce sparsity assumptions on
∗
The author thanks Christian Müller for a useful conversation and acknowledges the support of the NSF grant
DMS-1405746 and NSF CAREER Award DMS-1748166.
1
the covariance matrix (or its inverse). Still other papers suggest combinations of these structures
(Chandrasekaran et al. 2010, Bien & Tibshirani 2011, Luo 2011, Rothman 2012, Fan et al. 2013,
Liu et al. 2013, Xue et al. 2012).
Given the large number of papers focused on the sparsity assumption, it is perhaps surprising
to note that nearly all such papers fall into one of just two categories:
1. Methods that place no assumption on the pattern of sparsity but merely require that
the number of nonzeros be small. This assumption is placed on either the inverse covariance
matrix (Dempster 1972, Meinshausen & Bühlmann 2006, Yuan & Lin 2007, Rothman et al.
2008, Banerjee et al. 2008, Friedman et al. 2008, Peng et al. 2009, Yuan 2010, Cai et al. 2011,
Khare et al. 2014) or on the covariance matrix itself (Bickel & Levina 2008b, Rothman et al.
2009, Lam & Fan 2009, Bien & Tibshirani 2011, Rothman 2012, Xue et al. 2012, Liu et al.
2013).
2. Methods that assume the variables have a known ordering. These methods generally
assume a banded (or tapered) structure for the covariance matrix, its inverse, or the Cholesky
factor (Wu & Pourahmadi 2003, Huang et al. 2006, Levina et al. 2008, Bickel & Levina 2008a,
Wu & Pourahmadi 2009, Rothman et al. 2010, Cai et al. 2010, Cai & Yuan 2012, Bien et al.
2015a, Yu & Bien 2017).
We can view these two categories of sparsity as extremes on a spectrum. Methods in the first
category tackle a highly challenging problem that is widely applicable. Being completely neutral
topthe
pattern of zeros leads to an enormous number of sparsity patterns to choose from: there are
(2) for an s-sparse model. By constrast, methods in the second category address a more modest
s
problem that is much less widely applicable. Using a simple banding estimator, there are only p
sparsity patterns to choose from (one for each bandwidth) and there are few examples other than
data collected over time in which such estimators apply.
This paper introduces a new category of sparsity patterns that aims to bridge the wide divide between these two categories of sparsity pattern. The goal is to produce an estimator that
combines the problem-specific, targeted perspective of banded covariance estimation with the wide
applicability of the patternless sparsity approaches. The key idea is to generalize the notion of
banding beyond situations in which the variables have a known ordering. In particular, we replace
the assumption that the variables “have a known ordering” with the assumption that they “lie on
a known graph.” The latter is a generalization because variables having a known ordering can be
expressed as variables lying on a path graph.
To be precise, we will assume that we observe n independent copies of a random vector X ∈ Rp ,
where the p variables are associated with a known graph G = ([p], E). We will refer to G as the seed
graph. The right panel of Figure 1 shows a traditional B-banded matrix. The term “banded” refers
to the diagonal strip of nonzero elements that are within some distance B (called the “bandwidth”)
of the main diagonal. The middle panel of Figure 1 shows the graph with this adjacency matrix.
Observing that this graph, denoted GB , is the B-th power of the path graph shown in the left
panel of Figure 1 suggests the following generalization of the term “banded.” We will use dG (j, k)
to denote the length of the shortest path between nodes j and k on G.
Definition 1. We say that a matrix Σ is B-banded with respect to a graph G if supp(Σ) ⊆
E(GB ), that is Σjk = 0 if dG (j, k) > B.
2
●
●
●
●
●
●
●
●
●
Seed graph
●
●
●
●
●
●
●
●
●
●
●
●
5
●
B=4
●
●
●
10
●
15
●
●
●
●
●
●
●
●
20
●
●
●
●
●
●
5
10
15
20
Figure 1: A matrix with bandwidth B = 4 (right panel) can be thought of as the adjacency matrix of the
B-th power (center panel) of a path graph G (left panel). To form the B-th power of a graph G, one connects
any pair of nodes that are within B hops of each other on G.
●
●
●
●
●
●
●
●
●
●
●
●
5
●
Seed graph
B=4
●
10
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
15
●
●
●
●
●
20
●
●
●
5
10
15
20
Figure 2: Given a seed graph G (left panel), one forms GB (B = 4 case shown in center panel) by connecting
any pair of nodes within B hops of each other in G. Definition 1 refers to the resulting adjacency matrix
(right panel) as B-banded with respect to the graph G.
Bickel & Lindner (2012) have a similar notion, for a metric ρ, which they call ρ-generalized
banded operators.
Figure 2 shows the sparsity pattern of a 4-banded matrix with respect to a seed graph. The
bandwidth B can vary from B = 0, which corresponds to Σ being a diagonal matrix, to B being
the diameter of G, which corresponds to Σ being a dense matrix when G is a connected graph
(when G is not connected, the densest Σ attainable is, up to permutation of the variables, a blockdiagonal matrix with completely dense blocks). Despite the simple interpretation of taking powers
of a seed graph, the resulting sparsity patterns can be quite complex and varied depending on the
seed graph. There are many applications in which measured variables lie on a known graph that
would be relevant to incorporate when estimating the covariance matrix Σ:
• Image data: Consider a p1 × p2 pixeled image. The p = p1 p2 pixels lie on a two-dimensional
lattice graph. Nearby pixels would be expected to be correlated.
• fMRI data: One observes three-dimensional images of the brain over time. The voxels (threedimensional pixels) lie on a three-dimensional lattice.
3
●
●
●
●
●
●
●
●
●
B5 = 1
B8 = 1
B6 = 4
B4 = 2
B3 = 2
●
●
●
●
Seed graph
●
5
10
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
15
●
●
●
●
●
●
20
●
●
●
5
10
15
20
Figure 3: Given a seed graph G (left panel) and a set of node-specific bandwidths, one forms a
new graph (shown in center panel) by connecting any pair of nodes within Bj hops of node j in
G (five of the 20 nodes have nonzero bandwidths with values shown in figure; the remaining 15
nodes have zero bandwidths). Definition 2 refers to the resulting adjacency matrix (right panel) as
(B1 , . . . , Bp )-banded with respect to the graph G.
• Social network data: A social media company wishes to estimate correlations between users’
purchasing behaviors (e.g., do they tend to buy the same items?). The social network of users
is the known seed graph.
• Gene expression data: Genes that encode proteins that are known to interact may potentially
be more likely to be co-expressed than genes whose proteins do not interact. Biologists have
constructed protein interaction networks (Das & Yu 2012), which may be taken as seed
graphs, leveraging protein interaction data for better estimation of the covariance matrix of
gene expression data.
In each example described above, the naive approach would be to vectorize the data and ignore
the graph structure when estimating the covariance matrix. However, doing so when the truth is
banded with respect to the known graph would be unnecessarily working in the patternless sparsity
category when in fact a much simpler approach would work.
In some situations, a known seed graph is available, but the assumption of a single bandwidth B
applying globally is unrealistic. We can therefore extend the definition of bandedness with respect
to a graph as follows:
Definition 2. We say that a matrix Σ is (B1 , . . . , Bp )-banded with respect to a graph G, if
Σjk = 0 for all (j, k) satisfying dG (j, k) > max{Bj , Bk }.
In words, this definition says that each variable has its own bandwidth, which defines a neighborhood size, and it cannot be correlated with any variables beyond this neighborhood. For example,
on a social network, the bandwidth can be thought informally as having to do with the influence of
a person. A very influential friend might be correlated with not just his friends but with his friends’
friends (this would correspond to Bj = 2). By contrast, a “lone wolf” might be uncorrelated even
with his own friends (Bj = 0). Figure 3 shows an example of a sparsity pattern emerging from the
same seed graph as in Figure 2 but with variable bandwidths.
We propose two estimators in this paper that make use of a latent overlapping group lasso
penalty (Jacob et al. 2009, Obozinski et al. 2011), which is used in situations in which one desires
4
a sparsity pattern that is a union of a set of predefined groups. Both Definitions 1 and 2 can
be described as such, so the difference between the two methods lies in the choice of groups. In
Section 2, we introduce and study the graph-guided banding estimator, which produces estimates of
the covariance matrix that are banded in the sense of Definition 1. The limited number of sparsity
patterns attainable by this estimator makes it potentially undersirable in certain situations, and
we therefore introduce and study (in Section 3) the graph-guided banding with local bandwidths
estimator, which attains bandedness in the more flexible sense of Definition 2. Section 4 compares
these methods empirically both in simulation and in naturally occurring data. Section 2.3 concludes
with a discussion of related work, insights learned, and future directions.
To our knowledge, there is little work on using known graphs in high-dimensional covariance
estimation. Cai & Yuan (2015) study minimax lower bounds and adaptive estimation of covariance matrices when variables lie on a d-dimensional grid. Deng & Yuan (2009) consider inverse
covariance estimation in the setting where variables lie in a metric space. Their approach uses a
non-negative Garrotte semidefinite program formulation to perform inverse covariance estimation
in which inverse covariance decays with increasing distance between variables. The method requires
an initial estimate of the inverse covariance matrix (when n > p, one takes S −1 ) and then shrinks
so that variables farther apart have smaller partial covariances.
Notation: Given a p × p matrix A and a set of indices g ∈ {1, . . . , p}2 , we write Ag to denote the
p × p matrix that agrees with A on g and is zero on g c , and we write |g| for the cardinality of the
set g. The graph GB is the Bth power of G, meaning the graph in which any nodes that are within
B of each other in G are connected. The constraint Σ δIp means that the smallest eigenvalue
of Σ is at least δ. Given a square matrix Σ, the matrix Σ− = Σ − diag(Σ) agrees with Σ on
off-diagonals but is zero on the diagonal. We write an . bn to mean that there exists a constant
C > 0 for which an ≤ Cbn for n sufficiently large.
2
Graph-guided banding with global bandwidth
Definition 1 introduces the simplest notion of bandedness with respect to a graph. In particular,
a single bandwidth B is postulated to apply globally across the entire graph. This means that the
support of Σ is given by one of the following groups of parameters:
gb = {jk : 1 ≤ dG (j, k) ≤ b}
for b = 1, . . . , diam(G),
(1)
where diam(G) is the diameter of the seed graph G. The group lasso (Yuan & Lin 2006) is a
well-known extension of the lasso (Tibshirani 1996) that is useful for designing estimators in which
groups of variables are simultaneously set to zero. Given a sample
of n independent random vectors
P
X1 , . . . , Xn ∈ Rp , form the sample covariance matrix, S = ni=1 (Xi − X̄n )(Xi − X̄n )T . We define
the graph-guided banding estimator with global bandwidth (GGB-global) as
1
2
− c
Σ̂λ = arg min
kS − ΣkF + λP (Σ; G) s.t. [Σ ]gM = 0 ,
(2)
ΣδIp 2
where
P (Σ; G) =
min
{V(b) ∈Rp×p }
(M
X
wb kV(b) kF
b=1
s.t. ΣgM =
M
X
b=1
5
)
Vg(b)
b
is the latent overlapping group lasso penalty (Jacob et al. 2009) with group structure given by (1).
(b)
Each Vgb matrix has the “graph-guided” sparsity pattern gb , so heuristically, this penalty favors
Σ̂λ that can be written as the sum of a small number of such sparse matrices. While the penalty
is itself defined by an optimization problem, in practice we can solve (2) directly without needing
to evaluate P (Σ; G), as described in the next section.
The parameter M will typically be taken to be diam(G); however, in certain applications in
which diam(G) is very large and a reasonable upper bound on B is available, it may be compuc = 0 only has an effect
tationally expedient to choose a much smaller M . The constraint [Σ− ]gM
when M < diam(G). Of course, one must be cautious not to inadvertently choose M < B since
then the resulting estimates will always be sparser than the true covariance matrix. Thus, if one
does not have a reliable upper bound for B, one should simply take M = diam(G). The tuning
parameter λ ≥ 0 controls the sparsity level of Σ̂λ (with large values of λ corresponding to greater
sparsity). The wg ≥ 0 are weights that control the relative strength of the individual `2 norms
in the penalty. In studying the theoretical properties of the estimator (Section 2.2), it becomes
clear that wb = |gb |1/2 is a good choice. The parameter δ is a user-specified lower bound on the
eigenvalues of Σ̂λ . While our algorithm is valid for any value of δ, the most common choice is
δ = 0, which ensures that Σ̂λ is a positive semidefinite matrix. In certain applications, a positive
definite matrix may be required, in which case the user may specify a small positive value for δ.
In the special case that G is a path graph (and dropping the eigenvalue contraint), the GGB-global estimator reduces to the estimator introduced in Yan & Bien (2017). In the next section, we show
that (2) can be efficiently computed.
2.1
Computation
The challenge of solving (2) lies in the combination of a nondifferentiable penalty and a positive
semidefinite constraint. Without the penalty term, Σ̂λ could be computed with a single thresholding
of the eigenvalues; without the eigenvalue constraint, Σ̂λ is nothing more than the proximal operator
of λP (·; G) evaluated at S. The following lemma shows that we can solve (2) by alternate application
of the proximal operator of λP (·; G) and a thresholding of the eigenvalues. We present this result
for general convex penalties since the same result will be used for the estimator introduced in
Section 3.
Lemma 1. Consider the problem
Σ̂ = arg min
ΣδIp
1
2
kS − ΣkF + λΩ(Σ) ,
2
(3)
where Ω is a closed, proper convex function, and let
1
2
ProxλΩ (M) = arg min
kM − ΣkF + λΩ(Σ)
Σ
2
denote the proximal operator of λΩ(·) evaluated at a matrix M. Algorithm 1 converges to Σ̂.
Proof. See Appendix A.
In Appendix A, we show that Algorithm 1 corresponds to blockwise coordinate ascent on the
dual of (3). While the alternating direction method of multipliers (ADMM) would be an alternative
6
Algorithm 1 Algorithm for solving (3)
• Initialize C.
• Repeat until convergence:
1. B ← (S + C) − ProxλΩ (S + C)
2. C ← U · diag([Λii + δ]+ ) · UT where B − S = UΛUT is the eigendecomposition.
• Return Σ̂ ← S − B + C.
approach to this problem, we prefer Algorithm 1 because, unlike ADMM, its use does not require
choosing a suitable algorithm-specific parameter. Invoking the lemma, all that remains for solving
(2) is to specify an algorithm for evaluating ProxλP (·;G) . The group structure (1) used in this
estimator is special in that
g1 ⊂ g2 ⊂ · · · ⊂ gM .
Yan & Bien (2017) study the use of the latent overlapping group lasso with such hierarchically
nested group structures and show that, for group structures of this form, the proximal operator
can be evaluated essentially in closed form (which is not true more generally). Algorithm 3 of Yan
& Bien (2017) shows that evaluating the proximal operator amounts to identifying a sequence of
“breakpoints” k0 < k1 < . . . < km and then applying groupwise soft-thresholding to the elements
in gki+1 − gki for i = 0, . . . , m − 1.
In practice, we solve (3) for a decreasing sequence of values of λ. We initialize C in Algorithm
1 as 0 for the first value of λ and then for each subsequent value of λ we use the value of C from
the previous λ as a warm start.
2.2
Theory
In this section, we study the statistical properties of the GGB estimator (2). For simplicity, we will
focus on the special case in which δ = −∞ (or, equivalently, we drop the eigenvalue constraint);
however, in Section 2.2.3 we provide conditions under which our results apply to the δ = 0 case.
We introduce two conditions under which we prove these results.
Assumption 1 (distribution of data). We draw n i.i.d. copies of X ∈ Rp , having mean 0 and
covariance Σ∗ . The random vector has bounded variances, maxj |Σ∗jj | ≤ κ̄, and sub-Gaussian
√ ∗
2
marginals: E[etXj / Σjj ] ≤ eCt for all t ≥ 0. Both C and κ̄ are constants.
Assumption 2 (relation between n and p). We require that p ≤ eγ1 n for some γ1 > 0.
Assumptions 1 and 2 are sufficient for establishing that the sample covariance matrix S concentrates
around Σ∗ with high probability for large n. These assumptions are made in Bien et al. (2015a),
and similar assumptions appear throughout the literature.
2.2.1
Frobenius norm error
We begin by producing the rate of convergence of our estimator in Frobenius norm.
7
Theorem 1. Suppose Σ∗ is B ∗ -banded with q
respect to the graph G (in the sense of Definition 1)
and Assumptions 1–2 hold. Then with λ = x log(max{p,n})
(for x sufficiently large), wb = |gb |1/2 ,
n
δ = −∞, and M = diam(G), there exists a constant c > 0 such that
kΣ̂λ − Σ∗ k2F .
(p + |gB ∗ |) log(max{p, n})
n
with probability at least 1 − c/ max{p, n}.
Proof. See Appendix B.
The term p + |gB ∗ | is the sparsity level of Σ∗ (with the initial p representing the diagonal
elements). Theorem 1 shows that when p < n and moreover p + |gB ∗ | is o(n/ log n), the GGB
estimator is consistent in Frobenius norm. By contrast, consistency of the sample covariance
matrix S in this setting requires that p2 be o(n/ log n). Thus, when |gB ∗ | p2 , GGB is expected
to substantially outperform the sample covariance matrix. When p > n, the theorem does not
establish consistency of GGB. This fact is not, however, a shortcoming of the GGB method; rather,
it is a reflection of the difficulty of the problem and the fact that the assumed sparsity level is
in this case high relative to n. Even when B ∗ = 0 (i.e., Σ∗ is a diagonal matrix), we still must
estimate p free parameters on the basis of only n observations. More generally, we know that the
rate of convergence in Theorem 1 cannot, up to log factors, be improved by any other method. To
see this, note that in the special case that G is a path graph, Bien et al. (2015a) prove (in Theorem
7 of their paper) that the optimal rate is B ∗ p/n over the class of B ∗ -banded matrices (assuming
√
B ∗ ≤ n). Since |gB ∗ | ∼ B ∗ p for a path graph, we see that the GGB estimator is within a log
factor of the optimal rate in Frobenius norm.
If we are interested in estimating the correlation matrix R∗ = diag(Σ∗ )−1/2 Σ∗ diag(Σ∗ )−1/2
instead of the covariance matrix, then we can apply our estimator to the sample correlation matrix
R̂ = diag(S)−1/2 · S · diag(S)−1/2 . Because the diagonal entries of the correlation matrix do not
need to be estimated, we can show (under one additional assumption) that the rate of convergence
has |gB ∗ | in place of p + |gB ∗ |.
Assumption 3 (bounded non-zero variances). Suppose that there exists a constant κ > 0 (i.e.,
not depending on n or p) such that minj Σ∗jj ≥ κ.
Theorem 2. Suppose Σ∗ is B ∗ -banded with respect to the graph G (in the sense of Definition 1)
and Assumptions
1, 2, and 3 hold. Then there exist positive constants x and c for which taking
q
λ=x
log(max{p,n})
,
n
wb = |gb |1/2 , δ = −∞, and M = diam(G),
kΣ̂λ (R̂) − R∗ k2F .
|gB ∗ | log(max{p, n})
n
with probability at least 1 − c/ max{p, n}.
When G has a small number of edges (e.g., with all but a small number of nodes having 0
degree), then we can have |gB ∗ | n < p, in which case the above result establishes the consistency
of the graph-guided estimator even when p > n.
8
2.2.2
Bandwidth recovery
The remaining properties of this section can be established under a condition on the minimum
signal strength (analogous to the “beta-min” conditions arising in the regression context).
Assumption 4 (minimum signal strength). The nonzero elements of Σ∗ must be sufficiently large:
min |Σ∗jk | > 2λ.
jk∈gB ∗
q
log(max{p,n})
One can think of λ as on the order of
and measuring the “noise level” of the
n
data, in particular kS − Σ∗ k∞ . Thus, this assumption ensures that the nonzeros stand out from the
noise. Under this additional assumption, our estimator’s bandwidth, B(Σ̂λ ), matches the unknown
bandwidth B ∗ with high probability.
Theorem 3. Adding Assumption 4 to the conditions of Theorem 1, there exists a constant c > 0
such that B(Σ̂λ ) = B ∗ with probability at least 1 − c/ max{p, n}.
Proof. See Appendix C.
2.2.3
Positive semidefinite constraint
The results presented above are for δ = −∞. The next result provides an additional condition
under which the previous results hold for the δ = 0 case (i.e., the situation in which the estimator
is forced to be positive semidefinite).
Suppose all nonzero elements of Σ∗ are within a factor of τ of each other:
maxjk∈gB∗ |Σ∗jk |
≤ τ.
minjk∈gB∗ |Σ∗jk |
Assumption 5 (minimum eigenvalue). The minimum eigenvalue of Σ∗ must be sufficiently large:
p
λmin (Σ∗ ) > 4xτ log(p)/n max |gjB ∗ |.
j
The quantity maxj |gjB ∗ | is the maximal degree of the true covariance graph (defined by the
sparsity pattern of Σ∗ ).
Theorem 4. Adding Assumptions 4 and 5 to the conditions of Theorem 1, Theorems 1 and 3 hold
for δ = 0.
2.3
Discussion & the small world phenomenon
The theorems above assume that M = diam(G), but the results in fact apply more generally so
long as M ≥ B ∗ . While in general we may not know B ∗ , there are some applications in which
one may know enough about the mechanism of the graph-based dependence to provide an upper
bound on B ∗ . In such situations, choosing an M that is much smaller than diam(G) may lead to
substantial computational improvements (both in terms of timing and memory).
9
Examining the Frobenius-norm bound above, we see that the structure of the seed graph G plays
an important role in the rates of convergence. While it is tempting to think of B ∗ as controlling
the sparsity level (and therefore the difficulty of the problem), it is important to observe that this
quantity is in fact a function of both B ∗ and the graph G. For example, in the Frobenius-norm
bound, the quantity |gB ∗ | is the number of pairs of variables that are within B ∗ hops of each other in
the graph G. In conventional banded estimation, where G is a path seed graph, the dependence on
B ∗ is linear: |gB ∗ | ∼ B ∗ p. More generally, when G is a d-dimensional lattice graph |gB ∗ | ∼ (B ∗ )d p
(when B ∗ is small enough to avoid edge effects). However, intuition provided by lattice graphs does
not extend well to more
general graphs. For a particularly extreme example, when G is a complete
p
graph, then |g1 | = 2 . Many large graphs in nature have small diameter (Watts & Strogatz 1998),
which means that |gB ∗ | may be quite large even when B ∗ is small. The idea of “six degrees of
separation” (Travers & Milgram 1969) means that when using a social network as a seed graph,
|g6 | ∼ p2 (in fact, recent work suggests six might even be an overestimate, Backstrom et al. 2011).
When diam(G) is very small, Definition 1 is restricted to a very limited class of sparsity
patterns. In the next section, we consider a much larger class of sparsity patterns, using the notion
of local bandwidths introduced in Definition 2.
3
Graph-guided banding with local bandwidths
Definition 2 describes a different kind of graph-guided banding, in which each variable j has an
associated bandwidth Bj∗ , which can be thought of as a neighborhood size. Consider sets of the
form
gjb = {k : 1 ≤ dG (j, k) ≤ b}, for 1 ≤ b ≤ Mj , 1 ≤ j ≤ p,
which is the set of variables within b hops of j in the seed graph G. We can describe a sparsity
pattern of the type in Definition 2 as
i h
i
[ h
S∗ =
{j} × gjBj∗ ∪ gjBj∗ × {j} .
1≤j≤p
To see this, note that jk ∈ S ∗ is equivalent to dG (j, k) ≤ max{Bj∗ , Bk∗ }. We desire an estimator
whose sparsity patterns are unions of the {gjb }, so we use the latent overlapping group lasso
penalty (Jacob et al. 2009), which is designed for such situations. We define the graph-guided
banding estimator with local bandwidths (GGB-local) as
1
2
Σ̃λ = arg min
kS − ΣkF + λP̃ (Σ; G) ,
(4)
ΣδIp 2
where
P̃ (Σ; G) =
min
{V(jb) ∈Rp×p }
Mj
p X
X
wjb kV(jb) kF
j=1 b=1
s.t.
ΣgjMj =
Mj
X
b=1
Vg(jb)
jb
.
The only difference between this estimator and the GGB-global estimator of (2) is in the choice
of groups. For a fixed j, the groups are nested, gj1 ⊆ · · · ⊆ gjMj . When Mj is the diameter of
the connected component that contains j, then gjMj represents all variables that are connected to
variable j. For j 6= k, we can have gjb ∩gkb 6= ∅ without one group contained completely in the other.
The fact that this group structure is not simply hierarchically nested, as in the global bandwidth
estimator, complicates both the computation and the theoretical analysis of the estimator.
10
3.1
Computation
By Lemma 1, we can solve (4) by Algorithm 1. The challenge of evaluating the GGB-local estimator
lies therefore in efficiently evaluating the proximal operator ProxλP̃ . Viewing this problem as an
optimization problem over the latent variables {V(jb) } suggests a simple blockwise coordinate
descent approach, which we adopt here.
3.2
Theory
In Theorem 1, we saw that the Frobenius-norm convergence depends on the number of nonzero
off-diagonal elements, |gB∗ |, a quantity that we observed in Section 2.3 could be quite large—
as large as O(p2 )—even when B ∗ is very small. In other words, for such “small-world” seed
graphs G, the assumption of graph-guided bandedness with a small global bandwidth does not
necessarily correspond to Σ∗ being a sparse matrix. The notion of graph-guided bandedness with
local bandwidths, as in Definition 2, provides finer control, allowing us to describe sparse Σ∗ even
with “small-world” seed graphs G. The purpose of the next theorem is to show that, under such
a notion of sparsity, the GGB-local estimator can attain Frobenius-norm consistency when Σ∗ is
sparse.
Theorem 5. Suppose Σ∗ is (B1∗ , . . . , Bp∗ )-banded with respect to a graph G (in the sense of Defq
inition 2) and Assumptions 1-2 hold. Then with λ = x log(max{p,n})
(for x sufficiently large),
n
wjb = |gjb |1/2 , δ = −∞, and Mj = maxk dG (j, k),
r
log(max{p, n}) p log(max{p, n})
∗ 2
∗
kΣ̃λ − Σ kF . |S |
+
n
n
with probability at least 1 − c/ max{p, n}.
Proof. See Appendix D.
The two terms on the right-hand side of Theorem 5 represent estimation of the
q off-diagonal and
∗
∗
∗
, then
diagonal elements of Σ . If Σ is sparse enough—in particular that |S | ≤ p log(max{p,n})
n
∗
estimating all of Σ is as difficult as estimating its diagonal elements alone (i.e., just the p variances).
A simple example of this situation is when Σ∗ is (B1∗ , . . . , Bp∗ )-banded with respect to a graph G
q
and Bj∗ > 0 for at most log(max{p,n})
variables j. Another example is when all Bj∗ ≤ 1 and the
n
q
sum of degrees of all j with Bj∗ = 1 is bounded by p log(max{p,n})
.
n
Theorem 5 suggests that consistency in Frobenius norm might not be achieved in the p > n
regime even in the sparsest situation of a diagonal matrix. As noted earlier, this is not a shortcoming
of our method but rather a reflection of the difficulty of the task, which requires estimating p
independent parameters even in the sparsest situation of a diagonal matrix.
As noted previously, if instead we were interested in estimating the correlation matrix, we can
attain a rate as in Theorem 5 but without the second term by applying our estimator to the sample
correlation matrix.
Theorem 6. Suppose Σ∗ is (B1∗ , . . . , Bp∗ )-banded with respect to a graph G (in the sense of Definition 2) and Assumptions 1, 2, and 3 hold. Then there exists a constant x for which taking
11
λ=x
q
log(max{p,n})
,
n
wjb = |gjb |1/2 , δ = −∞, and Mj = maxk dG (j, k),
r
kΣ̃λ (R̂) −
R∗ k2F
∗
. |S |
log(max{p, n})
n
with probability at least 1 − c/ max{p, n}.
Consistency follows even when p n as long as the right-hand side approaches 0.
4
4.1
Empirical study
Simulations
All simulations presented in this paper were performed using the simulator package (Bien 2016b,a)
in R (R Core Team 2017). We investigate the advantages of exploiting a known graph structure
(both in the global and local bandwidths settings), and we also study our two methods’ robustness
to graph misspecification. We compare the two GGB methods to each other and to two other
methods, which represent the “patternless sparsity” and traditional banding approaches.
Perhaps the simplest and most commonly used sparse covariance estimator is to simply softthreshold each off-diagonal element of the sample covariance matrix:
soft
Σ̂jk = sign(Sjk ) [|Sjk | − λ]+ .
This method (Rothman et al. 2009), along with its close relatives (Rothman 2012, Liu et al. 2013,
Xue et al. 2012), do not make use of the graph G and corresponds to (3) with Ω being the `1 norm
and δ = −∞. Thus, comparisons to this method allow us to understand what is gained and what
is lost by using the graph G versus seeking “patternless sparsity.”
We also wish to compare our new estimators to a more traditional (i.e., non-graph-guided)
banding approach. Graph-guided banding is not the only way one might incorporate the graph
G. A different approach would be to find an ordering of the nodes based on G and then apply a
traditional banding estimator with respect to this ordering. To this end, we seek an ordering of the
variables that travels along G as much as it can, making jumps to another part of G as sparingly
as possible. We express this as a traveling salesman problem (TSP), which we then solve using an
off-the-shelf TSP solver (see Appendix F for details and for images of the learned paths). We then
apply the convex banding method using this ordering (Bien et al. 2015a,b).
While each of these methods can be solved with a positive semidefinite constraint, we compare
their performances without such a constraint. Doing so allows us to observe the fraction of the
time each unconstrained estimator is positive semidefinite: Table 5 in Appendix G shows that the
banding estimators are observed in the study to always be positive semidefinite, whereas this is not
the case for the soft-thresholding estimator. It follows that the reported results for the banding
methods in fact also pertain to these estimators with δ = 0.
To generate a covariance matrix that is (B1∗ , . . . , Bp∗ ) banded with respect to a graph G, we take
(
Σ∗jk =
1
dG (j,k) 1
n
o
dG (j, k) ≤ max(Bj∗ , Bk∗ )
a
if j 6= k
if j = k,
12
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Figure 4: Two-dimensional lattice graph; seed graph (left panel), global bandwidth (center panel), local
bandwidths (right panel)
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
● ●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
● ●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Figure 5: Scale-free graph; seed graph (left panel), local bandwidths (right panel)
where a is chosen to ensure that the minimum eigenvalue of Σ is at least σ 2 (we take σ = 0.01
throughout).
Initially, we compare the methods under four GGB structures: two-dimensional lattice with
global bandwidth, two-dimensional lattice with local bandwidths, scale-free graph with global bandwidth, and scale-free graph with local bandwidths. In the first two scenarios, we take G to be a
two-dimensional square lattice having 20 variables per side (so p = 400). The global bandwidth is
taken to be B ∗ = 4, and the local bandwidths are generated according to
0 with probability 0.90
∗
Bj = 1 with probability 0.06
4 with probability 0.04
(see Figure 4). The resulting Σ∗ has roughly 10% of elements nonzero in the global case and about
1% in the local case. In the third and fourth scenarios, we take G to be a scale free network (as
generated by the function sample pa in the R package igraph, Csardi & Nepusz 2006) with p = 400
nodes. The same two choices of global and local bandwidths lead to Σ∗ having, respectively, about
75% and 10% nonzeros (see Figure 5). The marked difference in sparsity levels is a manifestation
of the differing neighborhood structures between these two graph types.
13
Table 1: A comparison of mean Frobenius error at CV-best (averaged over 200 replicates).
2-d lattice (20 by 20), b = 4
2-d lattice (20 by 20) w/ var-bw
Scale free (p = 400), b = 4
Scale free (p = 400) w/ var-bw
GGB-global
7.605 (0.015)
4.244 (0.002)
6.740 (0.009)
3.273 (0.002)
GGB-local
8.381 (0.031)
3.170 (0.008)
6.543 (0.015)
2.867 (0.007)
Soft thresholding
11.634 (0.014)
4.347 (0.007)
7.655 (0.004)
3.552 (0.005)
TSP + hierband
11.562 (0.031)
4.779 (0.003)
7.727 (0.009)
3.661 (0.002)
Table 2: A comparison of mean operator error at CV-best (averaged over 200 replicates).
2-d lattice (20 by 20), b = 4
2-d lattice (20 by 20) w/ var-bw
Scale free (p = 400), b = 4
Scale free (p = 400) w/ var-bw
4.1.1
GGB-global
2.185 (0.007)
0.989 (0.003)
4.350 (0.010)
1.160 (0.001)
GGB-local
2.428 (0.015)
0.756 (0.005)
4.110 (0.016)
1.008 (0.005)
Soft thresholding
3.224 (0.010)
0.997 (0.004)
4.695 (0.004)
1.212 (0.002)
TSP + hierband
3.149 (0.016)
1.024 (0.003)
4.709 (0.009)
1.238 (0.001)
Lattice versus scale-free graphs and global versus local bandwidths
We generate samples of size n = 300 under the four scenarios (and repeat 200 times). Tables 1 and
2 present Frobenius-norm error and operator-norm errors, kΣ̂−Σ∗ kF and kΣ̂−Σ∗ kop , respectively;
table 3 presents the entropy error, − log det(Σ̂Σ∗−1 ) + trace(Σ̂Σ∗−1 ) − p. Each method has a single
tuning parameter, and we report the performance of the method with the value of the tuning
parameter chosen by 5-fold cross validation. In particular, we randomly
partition the training data
P
into 5 folds and, for an estimator Σ̂λ , choose λ̂CV-best = arg minλ 51 5k=1 kS(k) −Σ̂λ (S(−k) )kF , where
S(k) is the sample covariance matrix in which only the kth fold is included and S(−k) is the sample
covariance matrix in which the kth fold is omitted. In the 2-d lattice model, GGB-global does best
when the true model has global bandwidth, as we would expect it to. Interestingly, GGB-local does
best not just when the true model has local bandwidths but also in the scale-free model with a global
bandwidth. This is a reminder of the important role that the seed graph’s structure plays in the
statistical properties of the estimators. Finally, Figure 6 shows ROC curves for the four methods.
The results are consistent with the results from the tables. In particular, GGB-global does best in
the first scenario, and GGB-local does best in the other scenarios. Interestingly, TSP + hiernet
does not appear to be an effective approach. Unlike the graph-guided approaches, TSP + hiernet
does worse than soft-thresholding in nearly every scenario. For example, in Figure 6, we see that
TSP + hiernet’s ability to identify nonzero elements of Σ∗ is no better than random guessing. This
poor performance is to be expected in this scenario since such a network cannot be well captured
by an ordering of the nodes. For the two-dimensional lattice graph scenarios, we find that such
an approach is at least partially successful, which makes sense since the ordering captures some
aspects of the underlying seed graph.
Table 3: A comparison of mean entropy error at CV-best (averaged over 200 replicates).
2-d lattice (20 by 20), b = 4
2-d lattice (20 by 20) w/ var-bw
Scale free (p = 400), b = 4
Scale free (p = 400) w/ var-bw
GGB-global
7719 ( 24)
25823 ( 55)
72606 (146)
52047 ( 70)
GGB-local
7819 ( 69)
18088 (119)
68087 (191)
44188 (186)
14
Soft thresholding
10234 ( 34)
25358 ( 89)
68654 ( 62)
53332 ( 86)
TSP + hierband
11870 ( 73)
27188 ( 62)
68518 (119)
54304 ( 76)
2−d lattice (20 by 20), b = 4
2−d lattice (20 by 20) w/ var−bw
1.00
0.75
0.50
0.25
Specificity
Method
0.00
GGB − global
Scale free (p = 400), b = 4
GGB − local
Scale free (p = 400) w/ var−bw
Soft thresholding
1.00
TSP + hierband
0.75
0.50
0.25
0.00
0.00
0.25
0.50
0.75
1.00 0.00
0.25
0.50
0.75
1.00
Sensitivity
Figure 6: ROC curves (formed by varying the tuning parameter for each method), show the sensitivity and specificity of identifying the nonzero elements of Σ∗ . For visibility, only five of the 200
replicates are shown. (Plot made using R package ggplot2, Wickham 2009.)
15
4.1.2
Effect of graph misspecification
The previous simulation studies establish the rather obvious fact that when one knows the true
graph G with respect to which Σ∗ is banded, one can better estimate Σ∗ with an estimator that
uses G. In practice, however, one might typically use a graph G0 that is (hopefully) close to G but
not exactly identical. This section seeks to understand the effect of graph misspecification on the
effectiveness of graph-guided banding. We simulate data as above but imagine that we only have
access to a corrupted version of the underlying graph G.
Let π ∈ [0, 1] control the degree of corruption to G. Of course there are many ways to corrupt a
graph, but we choose a scheme that will preserve the properties of the graph and the sparsity level
of the covariance matrix (so as not to confound graph distortion with other factors such as sparsity
level, number of triangles, etc.). In particular, we take G to be an Erdos-Renyi graph with p = 400
vertices and 760 edges (this is chosen to match the number of edges in the previous examples). We
take Σ∗ to have global graph-guided bandwidth B ∗ = 4 with respect to this graph G. For each
edge in G, with probability π we remove it and then connect two currently unconnected nodes
chosen at random. We choose G to be an Erdos-Renyi graph so that the distorted graph G0 is
itself Erdos-Renyi with the same number of edges. By keeping all graph properties fixed, we can
be confident that as we vary π, any differences in performance of the methods will be attributable
to G and G0 having different edges.
Figure 7 shows the performance of the three methods as the probability of reassigning an edge
is varied from 0 (no distortion of G) to 1 (all edges are distorted). As expected, we see both
GGB methods deteriorating as the distortion level of the seed graph G0 is increased. When π = 0
(i.e., G = G0 ) GGB-global is best (which makes sense since Σ∗ has a global bandwidth). When
approximately π = 0.25 of the edges are distorted, the best obtainable operator-norm error by
the GGB methods becomes worse than that of soft-thresholding. Soft-thresholding is, of course,
unaffected by graph distortion since it does not use G0 at all. Surprisingly, in Frobenius-norm error
the two GGB methods remain better than soft-thresholding even at very high levels of distortion.
A deeper examination of this simulation reveals that the GGB methods are more biased but less
variable than the soft-thresholding estimates, leading to a smaller Frobenius-norm error. We do
expect this behavior to be problem-specific and in general would expect GGB methods with highly
distorted graphs to perform worse than soft-thresholding.
4.2
Application to images of handwritten digits
In this section, we apply our estimators to a dataset of (16 × 16)-pixel images of handwritten digits
(Le Cun et al. 1990, Hastie et al. 2009)1 . There are between (roughly) 700 and 1500 images per
digit type. For each digit type, we take a random n = 90 images to form the “training set” sample
covariance matrix Strain and use the remaining data to form the “test set” sample covariance matrix
Stest (our results are based on 20 replications of this process). A natural seed graph in this situation
is the two-dimensional lattice graph formed by connecting adjacent pixels. We compare our two
GGB estimators to the soft-thresholding estimator. We fit each covariance estimator to the training
data and compute its best attainable (over choice of tuning parameter) distance in Frobenius norm
to Stest ,
min kΣ̂λ (Strain ) − Stest k2F .
λ≥0
1
available at http://statweb.stanford.edu/~tibs/ElemStatLearn/datasets/zip.digits/
16
6
●
●
●
15
●
●
●
●
●
●
●
●
●
●
●
●
●
●
5
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
3
4
●
●
●
2
Best operator error
10
●
●
●
●
1
5
Best Frobenius error
●
●
0
●
●
0.0
GGB − global
GGB − local
Soft thresholding
0.2
0
●
0.4
0.6
0.8
1.0
0.0
Probability of distortion
0.2
0.4
0.6
0.8
1.0
Probability of distortion
Figure 7: The effect of graph misspecification on the performance of the global and local graph-guided banding
methods in terms of Frobenius and operator norms.
Table 4: A comparison of best attainable Frobenius-norm test error (averaged over 20 replicates). Standard
errors are reported in parentheses. For all but two of the digits, a GGB method has lower test error than
soft-thresholding in all 20 out of 20 replications.
Digit
Digit
Digit
Digit
Digit
Digit
Digit
Digit
Digit
Digit
0
1
2
3
4
5
6
7
8
9
GGB-global
10.8 (0.1)
2.6 (0.1)
10.9 (0.2)
8.9 (0.1)
9.4 (0.1)
10.7 (0.2)
9.2 (0.2)
7.3 (0.1)
9.3 (0.1)
7.8 (0.2)
GGB-local
10.8 (0.1)
2.6 (0.1)
10.9 (0.2)
8.9 (0.1)
9.5 (0.1)
10.8 (0.2)
9.1 (0.2)
7.4 (0.1)
9.4 (0.1)
7.9 (0.2)
Soft thresholding
11.0 (0.1)
2.6 (0.1)
11.3 (0.1)
9.3 (0.1)
9.9 (0.1)
11.2 (0.2)
9.2 (0.2)
7.5 (0.1)
9.8 (0.1)
8.0 (0.2)
Table 4 reports this quantity, averaged over the 20 random train-test splits of the dataset, for
each digit. GGB-local has a lower test error than soft-thresholding in all 20 replicates for the
digits 2-9; GGB-global has a lower test error than soft-thresholding in all 20 replicates for all digits
other than 0, 1, and 6. In fact, there is no digit for which soft-thresholding wins in a majority
of replicates. This indicates that exploiting the neighborhood structure inherent to image data is
useful for covariance estimates in this context.
5
Discussion
This paper generalizes the notion of banded covariance estimation to situations in which there is
a known underlying graph structure to the variables being measured. The traditional notion of a
banded matrix corresponds to the special case where this graph is a path encoding the ordering
of the variables. We propose two new “graph-guided” estimators, based on convex optimization
17
problems: GGB-global enforces a single bandwidth across all variables while GGB-local allows each
variable to have its own bandwidth. There are many interesting avenues of future research. In this
paper, we assume the seed graph is given, and the challenge is in determining the bandwidth(s).
One could imagine situations in which the seed graph itself is unavailable (but in which we do
believe that such a graph exists). In such a setting, can one identify the smallest seed graph G for
which the covariance matrix’s sparsity pattern is banded with respect to G? This generalizes the
problem of discovering an ordering for which one can apply a traditional known-ordering estimator
considered by Wagaman & Levina (2009), Zhu & Liu (2009). A simple heuristic strategy might
involve forming a minimum spanning tree based on the complete graph with edge weights given
by the sample covariance matrix. The graph-guided banding estimators would then have sparsity
patterns with this minimum spanning tree seed graph serving as backbone. Another future avenue
of research is in applying the graph-guided structure to the inverse covariance matrix. Finally,
we have seen (e.g., in Theorem 1 and in the empirical study) that the properties of the seed
graph G influence the success of the GGB estimators. More generally understanding how the
graph properties of G affect the difficulty of the estimation problem itself would be a particularly
interesting direction of future study.
A
Proof of Lemma 1 of main paper
Lemma 2. Consider the problem
Σ̂ = arg min
ΣδIp
1
2
kS − ΣkF + λΩ(Σ) ,
2
where Ω is a closed, proper convex function, and let
1
2
ProxλΩ (M) = arg min
kM − ΣkF + λΩ(Σ)
Σ
2
denote the proximal operator of λΩ(·) evaluated at a matrix M. Algorithm 1 converges to Σ̂.
Proof. Problem (3) is equivalent to
1
min
kS − Σk2F + λΩ(Σ̃)
2
Σ,Σ̃
s.t. Σ = Σ̃,
Σ δIp ,
which has Lagrangian
1
L(Σ, Σ̃; B, C) = kS − Σk2F + λΩ(Σ̃) + hΣ − Σ̃, Bi + hδIp − Σ, Ci.
2
The dual function is given by minimizing over Σ and Σ̃:
n
o
1
inf L(Σ, Σ̃; B, C) = δhIp , Ci + inf
kS − Σk2F + hΣ, B − Ci + inf λΩ(Σ̃) − hΣ̃, Bi
Σ
2
Σ,Σ̃
Σ̃
1
1
= δhIp , Ci − kB − C − Sk2F + kSk2F − [λΩ]∗ (B)
2
2
18
where [λΩ]∗ is the convex conjugate of λΩ. Also,
Σ̂ = S − B̂ + Ĉ.
The dual optimal (B̂, Ĉ) is given by solving
1
min kB − C − Sk2F + [λΩ]∗ (B) − δhIp , Ci s.t.
B,C 2
C 0,
which is a convex problem. Furthermore, by Tseng (2001), it can be solved by alternately optimizing
over B (holding C fixed) and over C (holding B fixed). The solution to the optimization over B is
the proximal operator of [λΩ]∗ evaluated at C + S. The Moreau decomposition (see, e.g., Parikh
& Boyd 2014) for the closed, proper convex function λΩ gives us
B̂(C) = Prox[λΩ]∗ (S + C) = S + C − ProxλΩ (S + C).
Optimizing over C while holding B fixed requires solving
1
Ĉ(B) = min kB − C − Sk2F − δhIp , Ci
C 2
s.t. C 0,
which in terms of the eigenvalue decomposition of B − S = UΛUT becomes
1
Ĉ(B) = min kC − UΛUT k2F − δhIp , Ci s.t. C 0
C 2
or Ĉ(B) = UD̂(B)UT where
1
D̂(B) = min kD − Λk2F − δhIp , Di
D 2
s.t. D 0,
which is solved by
D̂(B) = diag([Λii + δ]+ ).
B
Proof of Theorem 1 of main paper (Frobenius error)
Proof. Since δ = −∞ and M = diam(G), the estimator is more simply
1
2
Σ̂ := Σ̂λ = arg min
kS − ΣkF + λP (Σ; G)
Σ
2
where
P (Σ; G) =
min
{V(b) ∈Rp×p }
(M
X
wb kV(b) kF
b=1
s.t.
Σ− =
M
X
(5)
)
Vg(b)
b
.
b=1
In what follows, whenever we write a sum with subscript b, it is understood to run along b =
1, . . . , M ; this also holds for the {V(b) } notation, which is short for {V(b) }M
b=1 .
19
We begin by observing that (5) is a strongly convex problem in Σ, and thus there is a unique
minimizer Σ̂, although the minimizing set of V̂(b) ’s may not be unique. Also, since diag(Σ) does
not appear in the penalty, we see that diag(Σ̂) = diag(S). Thus, we can write (5) as
1 −
−
2
kS − ΣkF + λP (Σ; G) .
(6)
Σ̂ = arg min
Σ
2
We can write (6) in terms of the V(b) ’s:
(
X
X
1
{V̂(b) } ∈ arg min
kS −
V(b) k2F + λ
wb kV(b) kF
2
V(b)
b
)
s.t.
(b)
Vg c
b
= 0 ∀b .
(7)
b
Note that we have replaced S− by S (for notational convenience) since gb does not include diagonal
elements. The lemma below provides necessary and sufficient conditions for {V̂(b) } to be a solution
to (7).
P
(b)
Lemma 3. Writing Σ̂ = b V̂(b) , {V̂(b) } is a solution to (7) iff. for each b ∈ {1, . . . , M }, V̂gc = 0
b
and
(b)
λwb V̂gb
[S − Σ̂]gb =
if V̂(b) 6= 0
kV̂(b) kF
and k[S − Σ̂]gb kF ≤ λwb otherwise.
Proof. Problem (7) is equivalent to a non-overlapping group lasso problem, and this lemma provides
the KKT conditions of this problem, which are well-known.
Corollary 1. Let {V̂(b) } be a solution to (7). Then,
k[S − Σ̂]gb kF < λwb =⇒ V̂(b) = 0.
Suppose supp(Σ∗ ) = gB ∗ . If we can construct a solution {V̂(b) } to (7) that has
max wb−1 k[S − Σ̂]gb kF < λ,
b>B ∗
then, by Corollary 1, we know that V̂(b) = 0 for all b > B ∗ or
{b : V̂(b) 6= 0} ⊆ {1, . . . , B ∗ }.
In fact, the corollary implies that this holds for all solutions, not just for our constructed {V̂(b) }.
Consider the restricted problem
(
)
B∗
B∗
X
X
1
(b)
(b) 2
(b)
(b) B ∗
∗
kS −
V kF + λ
wb kV kF s.t. Vgc = 0 for 1 ≤ b ≤ B ,
{V̄ }b=1 ∈ arg min
b
2
V(b)
b=1
b=1
PB ∗
(b)
and write Σ̄ =
b=1 V̄ . This restricted problem is of an identical form to (7) but with M
replaced by B ∗ . Thus, Lemma 3 applies if we replace b = 1, . . . , M with b = 1, . . . , B ∗ : For each
b ≤ B∗,
(b)
λwb V̄gb
if V̄(b) 6= 0
[S − Σ̄]gb =
kV̄(b) kF
20
and k[S − Σ̄]gb kF ≤ λwb otherwise.
P
(b) =
Now, for b > B ∗ , let V̄(b) = 0. We show that {V̄(b) }M
is a solution to (7). Clearly, M
b=1 V̄
b=1
PB ∗ (b)
= Σ̄, so by Lemma 3 it simply remains to show that
b=1 V̄
k[S − Σ̄]gb kF ≤ λwb for b > B ∗ .
For b > B ∗ :
k(S − Σ̄)gb k2F = k(S − Σ̄)gB∗ k2F + kSs(B∗ +1):b k2F
2
2
∗
≤ λ2 wB
∗ + k(S − Σ )s(B ∗ +1):b kF
2
∗ 2
≤ λ2 wB
∗ + |s(B ∗ +1):b | · kS − Σ k∞ .
Now, on the event
Aλ = {kS − Σ∗ k∞ < λ},
(8)
we have
2
2 2
k(S − Σ̄)gb k2F < λ2 (wB
∗ + |s(B ∗ +1):b |) = λ wb ,
using that wb2 = |gb | = |gB ∗ | + |s(B ∗ +1):b |. This establishes, by Lemma 3, that {V̄(b) } is a solution
to (7) and, by Corollary 1 that {b : V̂(b) 6= 0} ⊆ {1, . . . , B ∗ } for all solutions {V̂(b) } to (7). Given
the support constraints of each V̂(b) , it follows that the overall Σ̂ has bandwidth no larger than
B∗.
Proposition 1. If wb = |gb |1/2 , δ = −∞, and M = diam(G), then B(Σ̂λ ) ≤ B ∗ on Aλ .
Having established the above proposition about the bandwidth, it is straightforward to prove a
result about the estimation error. On Aλ ,
k(Σ̂ − Σ∗ )gB∗ kF ≤ k(S − Σ∗ )gB∗ kF + k(S − Σ̂)gB∗ kF ≤ kS − Σ∗ k∞ · |gB ∗ |1/2 + λwB ∗ < 2λwB ∗
and Σ̂gB∗ = 0. Thus,
2
kΣ̂ − Σ∗ k2F = kdiag(S) − diag(Σ∗ )k2F + k(Σ̂ − Σ∗ )gB∗ k2F < λ2 (p + 4wB
∗ ).
(9)
Proposition 2. If wb = |gb |1/2 , δ = −∞, and M = diam(G), then
kΣ̂ − Σ∗ k2F ≤ λ2 (p + 4|gB(Σ∗ ) |)
on Aλ .
Under Assumptions 1 and 2 of the main paper, there exists a constant c > 0 for which
P (Aλ ) ≥ 1 − c/ max{p, n}
(10)
p
if we take λ = x log max{p, n}/n with x sufficiently large. This follows directly from bound
(A.4.1) of Theorem A.1 of Bien et al. (2015c). Theorem 1 of the main paper follows.
21
C
Proof of Theorem 3 of main paper (bandwidth recovery)
We prove a result that is a bit more general than Theorem 3 of the main paper.
Theorem 1. Under the minimum signal condition
min kΣ∗gb kF /|gb |1/2 > 2λ,
b≤B ∗
and the conditions of Theorem 1 of the main paper, B(Σ̂) = B(Σ∗ ) with probability at least 1 −
c/ max{p, n}.
Proof. On the event Aλ defined in (8), for b ≤ B ∗ ,
k(Σ̂ − Σ∗ )gb kF ≤ k(S − Σ∗ )gb kF + k(S − Σ̂)gb kF
< λ|gb |1/2 + k(S − Σ̂)gb kF
≤ λ|gb |1/2 + λwb = 2λwb .
Thus,
max wb−1 k(Σ̂ − Σ∗ )gb kF < 2λ.
b≤B ∗
We already know that B(Σ̂) ≤ B ∗ , so it suffices to show that
min kΣ̂gb kF /wb > 0,
b≤B ∗
which follows since
kΣ̂gb kF /wb ≥ kΣ∗gb kF /wb − k(Σ̂ − Σ∗ )gb kF /wb > 2λ − 2λ = 0.
The result then follows from (10).
Corollary 2. Under the minimum signal condition
min |Σ∗jk | > 2λ,
jk∈gB ∗
and the conditions of Theorem 1 of the main paper, B(Σ̂) = B(Σ∗ ) with probability at least 1 −
c/ max{p, n}.
Proof. For any b ≤ B ∗ ,
kΣ∗gb k2F /|gb | =
X
[Σ∗jk ]2 /|gb | >
jk∈gb
X
jk∈gb
Thus, the conditions of Theorem 1 hold.
22
4λ2 /|gb | = 4λ2 .
D
Proof of Theorem 5 (Frobenius error of local bandwidths estimator)
Proof. Let Σ̃ := Σ̃λ be the minimizer of Equation (4) of the main paper. By definition,
1
1
kS − Σ̃k2F + λP̃ (Σ̃; G) ≤ kS − Σ∗ k2F + λP̃ (Σ∗ ; G).
2
2
Since
kS − Σ̃k2F = kS − Σ∗ k2F + kΣ̃ − Σ∗ k2F − 2hS − Σ∗ , Σ̃ − Σ∗ i,
we have
1
kΣ̃ − Σ∗ k2F + λP̃ (Σ̃; G) ≤ hS − Σ∗ , Σ̃ − Σ∗ i + λP̃ (Σ∗ ; G).
2
(11)
−1
By Lemma 3 of Obozinski et al. (2011), the dual norm is P̃ ∗ (B; G) = maxjb wjb
kBgjb kF , and so
hS − Σ∗ , Σ̃ − Σ∗ i ≤ λ2 p + P̃ ∗ (S − Σ∗ ; G)P̃ (Σ̃ − Σ∗ ; G).
(12)
Define the index sets I = {jb : 1 ≤ b ≤ Mj } and I ∗ = {jb : 1 ≤ b ≤ Bj∗ }.
Observe that with wjb = |gjb |1/2
−1
kBgjb kF ≤ kBk∞ .
max wjb
jb∈I
It follows that
hS − Σ∗ , Σ̃ − Σ∗ i < λ2 p + λP̃ (Σ̃ − Σ∗ ; G),
as long as kS − Σ∗ k∞ < λ, that is on the event Aλ , as defined in (8). Thus, combining this with
(11),
h
i
1
kΣ̃ − Σ∗ k2F + λP̃ (Σ̃; G) ≤ λ2 p + λP̃ (Σ̃ − Σ∗ ; G) + λP̃ (Σ∗ ; G).
2
holds on Aλ . Using the triangle inequality (P̃ is a norm),
h
i
1
kΣ̃ − Σ∗ k2F ≤ λ2 p + λ P̃ (Σ̃ − Σ∗ ; G) − P̃ (Σ̃; G) + P̃ (Σ∗ ; G) ≤ λ2 p + 2λP̃ (Σ∗ ; G).
2
By definition of P̃ ,
P̃ (Σ∗ ; G) ≤
X
wjb kV(jb) kF
(13)
jb∈I
(jb)
for any choice of {V(jb) } such that Vgc = 0 and Σ∗ =
(j 0 b0 )
Consider taking Vjk
P
jb V
jb
(jb) .
= 0, for all jk ∈ {1, . . . , p}2 and j 0 b0 ∈ I, unless
(jBj∗ )
(i) dG (j, k) ≤ min{Bj∗ , Bk∗ }, in which case we take Vjk
(kBk∗ )
(ii) Bj∗ < dG (j, k) ≤ Bk∗ , in which case we take Vjk
23
(jBj∗ )
= Vkj
(kBk∗ )
= Vkj
(kBk∗ )
= Vjk
= Σ∗jk .
(kBk∗ )
= Vkj
= 12 Σ∗jk .
(jBj∗ )
(iii) Bk∗ < dG (j, k) ≤ Bj∗ , in which case we take Vjk
(jBj∗ )
= Vkj
= Σ∗jk .
We verify that this choice of {V(jb) } satisfies the necessary constraints. Clearly, V(jb) = 0 for
b 6= Bj∗ . If jk 6∈ gjBj∗ , then dG (j, k) > Bj∗ , which excludes cases (i) and (ii), the only two cases in
(jBj∗ )
(jBj∗ )
(jB ∗ )
) could be nonzero. Thus, Vgc ∗j = 0.
jB
j
P
0 b0 )
∗
(j
We next verify that j 0 b0 ∈I V
= Σ . In case (i),
which Vjk
(and Vkj
(jBj∗ )
0 0
X
[V(j b ) ]jk = Vjk
(kBk∗ )
= Σ∗jk .
+ Vjk
j 0 b0 ∈I
In case (ii),
X
0 0
(kBk∗ )
0 0
(jBj∗ )
[V(j b ) ]jk = Vjk
= Σ∗jk .
j 0 b0 ∈I
In case (iii),
X
[V(j b ) ]jk = Vjk
= Σ∗jk .
j 0 b0 ∈I
(jBj∗ )
Now, we have that |Vjk
X
∗
| ≤ |Σ∗jk |, which implies that kV(jBj ) kF ≤ kΣ∗gjB∗ kF . Thus,
j
wjb kV(jb) kF =
X
wjBj∗ kV
(jBj∗ )
kF ≤
j
jb∈I
X
wjBj∗ kΣ∗gjB∗ kF ≤ kΣ∗ k∞
j
j
X
|gjBj∗ |.
j
Now, each jk ∈ S ∗ appears only in gjBj∗ or gkBk∗ (or potentially both). Thus, by (13),
2|S ∗ |.
∗
∗
It follows that P̃ (Σ ; G) ≤ 2kΣ k∞
|S ∗ |,
P
j
|gjBj∗ | ≤
and so (on Aλ )
kΣ̃ − Σ∗ k2F ≤ λ2 p + 4kΣ∗ k∞ |S ∗ |λ.
Assumptions 1 and 2 of the main paper are used (as in Appendix B) to show that Aλ holds with
the desired probability.
E
Proofs about GGB on sample correlation matrix
We begin by establishing a concentration inequality for the correlation matrix that will be used
in the proofs of both of these theorems. This result is similar to Lemma A.2 of Liu et al. (2017)
except that we assume sub-Gaussian rather than Gaussian data. It is also similar to Lemma D.5
of Sun et al. (2017) except that we assume that S has been centered by the sample mean.
Lemma 4. Suppose that Assumptions 1, 2, and 3 hold. There exists a constant x (i.e., a value
not depending on n or p) such that for all n and p for which log{p, n}/n < 1/x2 ,
p
P(kR̂ − R∗ k∞ > x log(max{p, n})/n) ≤ 16/ max{p, n}.
24
Proof. The following proof is based on the proof of Lemma A.2 of Liu et al. (2017). Certain
details have been fleshed out and some modifications have been made to accommodate the differing
assumptions.
Let Aj = {|Sjj − Σ∗jj | ≤ Σ∗jj /2}. Observing that Aj = {1/2 ≤ Sjj /Σ∗jj ≤ 3/2}, it follows that
Sjj Skk
≤ (3/2)2
Σ∗jj Σ∗kk
Sjj Skk 1/2
=⇒ 1/2 ≤
≤ 3/2
Σ∗jj Σ∗kk
Aj ∩ Ak =⇒ (1/2)2 ≤
1/2
1/2
1/2
=⇒ (1/2)(Σ∗jj Σ∗kk )jk ≤ (Sjj Skk )jk ≤ (3/2)(Σ∗jj Σ∗kk )jk
1/2
1/2
1/2
=⇒ |(Sjj Skk )jk − (Σ∗jj Σ∗kk )jk | ≤ (1/2)(Σ∗jj Σ∗kk )jk .
(14)
Now,
{|R̂jk − R∗jk | > t} = {|Sjk − Σ∗jk | > t(Sjj Skk )1/2 }
⊆ {|Sjk − Σ∗jk | > t(Sjj Skk )1/2 } ∩ (Aj ∩ Ak )
[
(Aj ∩ Ak )c .
By (14) and the triangle inequality,
{|Sjk − Σ∗jk | > t(Sjj Skk )1/2 } ∩ (Ai ∩ Aj ) ⊆ {|Sjk − Σ∗jk | > (t/2)(Σ∗jj Σ∗kk )1/2 },
so
{|R̂jk − R∗jk | > t} ⊆ {|Sjk − Σ∗jk | > (t/2)(Σ∗jj Σ∗kk )1/2 }
[
(Aj ∩ Ak )c .
Now,
{max |R̂jk − R∗jk | > t} ⊆
j6=k
i
[h
{|Sjk − Σ∗jk | > (t/2)(Σ∗jj Σ∗kk )1/2 } ∪ Acj ∪ Ack
j6=k
=
[
{|Sjk − Σ∗jk | > (t/2)(Σ∗jj Σ∗kk )1/2 } ∪
⊆
Acj
j
j6=k
[
[
{|Sjk − Σ∗jk | > (t/2)κ} ∪
[
Acj
j
j6=k
⊆ {max |Sjk −
j6=k
Σ∗jk |
> (t/2)κ} ∪ {max |Sjj − Σ∗jj | > (1/2)κ}
j
∗
⊆ {kS − Σ k∞ > (κ/2) min{t, 1}}.
Since R̂jj = 1 = R∗jj , maxj6=k |R̂jk − R∗jk | = kR̂ − R∗ k∞ and thus we have established that
P(kR̂ − R∗ k∞ > t) ≤ P(kS − Σ∗ k∞ > (κ/2) min{t, 1}).
By Lemma A.1 of Bien et al. (2015c) (with a typo corrected), there exist constants c1 , c2 such
that
2
P(kS − Σ∗ k∞ > ) ≤ 2p2 e−c2 n(/κ̄) + 8pe−c1 n/κ̄
for any 0 < < 2κ̄. This constraint on implies that (/κ̄)2 < 2(/κ̄) and so
2
2
2p2 e−c2 n(/κ̄) + 8pe−c1 n/κ̄ < 2p2 e−c2 n(/κ̄) + 8pe−c1 n(/κ̄)
25
2 /2
.
Taking c3 = min{c2 , c1 /2},
2
2
P(kS − Σ∗ k∞ > ) ≤ (2p2 + 8p)e−c3 n(/κ̄) ≤ 16p2 e−c3 n(/κ̄) .
Observing that = (κ/2) min{t, 1} < κ̄ < 2κ̄, we have that
2 /(2κ̄)2
P(kR̂ − R∗ k∞ > t) ≤ 16p2 e−c3 nκ
min{t2 ,1}
.
p
We take t = x log(max{p, n})/n for x = 12(κ̄/κ)2 /c3 . By Assumption 3, κ and κ̄ are treated as
constants. If x2 log(max{p, n})/n then min{t2 , 1} = t2 and
2 /(2κ̄)2 x2
P(kR̂ − R∗ k∞ > t) ≤ 16p2 e−c3 κ
log(max{p,n})
2 /(2κ̄)2 x2
≤ 16 max{p, n}2−c3 κ
.
Thus, we just need that 2 − c3 κ2 /(2κ̄)2 x2 ≤ −1, which can be ensured by choosing
x = 12(κ̄/κ)2 /c3 .
We define the event
Bλ = {kR̂ − R∗ k∞ < λ}.
Lemma 4 establishes that there exist constants x and c for which choosing λ = x
ensures that P(Bλ ) ≥ 1 − c/ max{n, p}.
E.1
(15)
q
log(max{p,n})
n
Proof of Theorem 2 (Frobenius error of global GGB on correlation matrix)
Let Σ̂λ (R̂) be the minimizer of Equation (2) of the main paper where S is replaced by R̂.
Proof of Theorem 2. The proof is nearly identical to that of Theorem 1 given in Section C. We
take Σ̂ to be Σ̂λ (R̂), we replace S with R̂, and we replace Aλ with Bλ . Instead of (9), we get
2
kR̂ − R∗ k2F = k(Σ̂ − Σ∗ )gB∗ k2F < 4λ2 wB
∗,
which leads to the stated result.
E.2
Proof of Theorem 6 (Frobenius error of local GGB on correlation matrix)
Let Σ̃λ (R̂) be the minimizer of Equation (4) of the main paper where S is replaced by R̂.
Proof of Theorem 6. The proof is nearly identical to that of Theorem 5 given in Section D. We
take Σ̃ to be Σ̃λ (R̂) and replace S with R̂. Equation 12 becomes instead
hR̂ − R∗ , Σ̃ − R∗ i ≤ P̃ ∗ (R̂ − R∗ ; G)P̃ (Σ̃ − R∗ ; G).
Following the same steps as in that proof, we note that on the set Bλ , defined in (15),
kΣ̃ − Σ∗ k2F ≤ 4kΣ∗ k∞ |S ∗ |λ.
26
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
● ●
●
●
●
●
● ●
● ●
● ●
●
●
● ●
● ●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
● ● ● ●
●
● ●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●●
●
●
●
● ●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
● ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
● ●
●
●
●
● ●
● ● ●
● ●
●
●
●
Figure 8: TSP solutions used for the two graphs used in the simulation studies
F
Traveling salesman problem
Let C be a p × p matrix of costs, with Cjk being the cost of traveling from node j to node k. The
traveling salesman problem (TSP) seeks a loop v1 → v2 → . . . vp → v1 through all the nodes such
P
that p−1
`=1 Cv` v`+1 + Cvp v1 is minimal.
In our context, we have a graph G and wish to find an ordering of the nodes that is strongly
informed by G. In particular, we would like this path to travel along edges of G as much as possible
and when this is not possible and the path must “fly” between two nodes that are not connected
by a node, that it take the shortest flight possible. We therefore take
Cjk = dG (j, k) − 1.
We use the R package TSP (Hahsler & Hornik 2017), which provides an interface to the Concorde
solver (Applegate et al. 2017), and Figure 8 shows the resulting paths.
G
An empirical study of positive semidefiniteness
For each of the models considered in Section 4.1.1 of the main paper, we record the proportion
of the time each method is positive semidefinite. In particular, this average is over both the 200
simulation replicates and the values of the tuning parameter of each method. Table 5 shows that
throughout this study the banding methods (GGB and hierband) are always observed to give
positive semidefinite estimates. This is not the case for the soft-thresholding estimator.
H
A higher dimensional example
In Section 4.1.1, we study four models in which p = 400 and n = 300. The covariance matrix has
p(p + 1)/2 ≈ 80, 000 parameters. Thus, the number of parameters to be estimated greatly exceeds
27
Table 5: A comparison of propor PSD (averaged over a grid of tuning parameters and 200 replicates).
2-d lattice (20 by 20), b = 4
2-d lattice (20 by 20) w/ var-bw
Scale free (p = 400), b = 4
Scale free (p = 400) w/ var-bw
GGB
1.000
1.000
1.000
1.000
- global
(0.000)
(0.000)
(0.000)
(0.000)
GGB
1.000
1.000
1.000
1.000
- local
(0.000)
(0.000)
(0.000)
(0.000)
1.0
0.4
0.6
0.8
0.6
0.8
0.2
0.2
0.0
0.0
0.0
0.4
Specificity
0.8
0.6
0.4
GGB − global bandwidth
GGB − local bandwidth
Soft thresholding
0.2
Specificity
TSP + hierband
1.000 (0.000)
1.000 (0.000)
1.000 (0.000)
1.000 (0.000)
2−d lattice (20 by 20)
1.0
2−d lattice (50 by 50)
Soft thresholding
0.349 (0.008)
0.625 (0.008)
0.998 (0.001)
1.000 (0.000)
1.0
0.0
Sensitivity
0.2
0.4
0.6
0.8
1.0
Sensitivity
Figure 9: ROC curves for three methods when p = 400 and p = 2, 500.
the number of samples. Nonetheless, we can ask what happens if p is even larger than this. We
revisit the example of the 2-d lattice, increasing the side length from 20 to 50 so that p = 2, 500. In
this case the covariance matrix has p(p + 1)/2 ≈ 3, 000, 000 parameters. We focus in particular on
the GGB methods and soft thresholding. Figure 9 provides a side-by-side comparison of the ROC
curves (five replicates are shown per method) when one increases from p = 400 to p = 2, 500. The
curves in the two plots are different although this is difficult to discern by eye. This suggests that
we are effectively in the same regime in these two simulated scenarios.
References
Applegate, D., Bixby, R., Chvatal, V. & Cook, W. (2017), Concorde TSP Solver.
URL: http://www.math.uwaterloo.ca/tsp/concorde/
Backstrom, L., Boldi, P., Rosa, M., Ugander, J. & Vigna, S. (2011), ‘Four degrees of separation.
doi: arxiv1111. 4570v1’.
Banerjee, O., El Ghaoui, L. E. & d’Aspremont, A. (2008), ‘Model selection through sparse maximum
likelihood estimation for multivariate gaussian or binary data’, Journal of Machine Learning
Research 9, 485–516.
Bickel, P. J. & Levina, E. (2008a), ‘Regularized estimation of large covariance matrices’, Annals of
Statistics pp. 199–227.
28
Bickel, P. & Levina, E. (2008b), ‘Covariance regularization by thresholding’, Annals of Statistics
36, 2577–2604.
Bickel, P. & Lindner, M. (2012), ‘Approximating the inverse of banded matrices by banded matrices
with applications to probability and statistics’, Theory of Probability & Its Applications 56(1), 1–
20.
Bien, J. (2016a), simulator: An Engine for Running Simulations. R package version 0.2.0.
URL: https://CRAN.R-project.org/package=simulator
Bien, J. (2016b), ‘The simulator:
arXiv:1607.00021 .
An engine to streamline simulations’, arXiv preprint
Bien, J., Bunea, F. & Xiao, L. (2015a), ‘Convex banding of the covariance matrix’, Journal of the
American Statistical Association DOI: 10.1080/01621459.2015.1058265 .
Bien, J., Bunea, F. & Xiao, L. (2015b), hierband: Convex Banding of the Covariance Matrix. R
package version 1.0.
URL: https://CRAN.R-project.org/package=hierband
Bien, J., Bunea, F. & Xiao, L. (2015c), ‘Supplementary materials to “convex banding of the covariance matrix”’, Journal of the American Statistical Association .
Bien, J. & Tibshirani, R. (2011), ‘Sparse estimation of a covariance matrix’, Biometrika 98(4), 807–
820.
Cai, T., Liu, W. & Luo, X. (2011), ‘A constrained l1 minimization approach to sparse precision
matrix estimation’, Journal of the American Statistical Association 106(494), 594–607.
Cai, T. T. & Yuan, M. (2012), ‘Adaptive covariance matrix estimation through block thresholding’,
The Annals of Statistics 40(4), 2014–2042.
Cai, T. T. & Yuan, M. (2015), ‘Minimax and adaptive estimation of covariance operator for random
variables observed on a lattice graph’, Accepted to Journal of the American Statistical Association
.
Cai, T. T., Zhang, C.-H. & Zhou, H. H. (2010), ‘Optimal rates of convergence for covariance matrix
estimation’, The Annals of Statistics 38(4), 2118–2144.
Chandrasekaran, V., Parrilo, P. A. & Willsky, A. S. (2010), Latent variable graphical model selection
via convex optimization, in ‘Communication, Control, and Computing (Allerton), 2010 48th
Annual Allerton Conference on’, IEEE, pp. 1610–1613.
Csardi, G. & Nepusz, T. (2006), ‘The igraph software package for complex network research’,
InterJournal Complex Systems, 1695.
URL: http://igraph.org
Das, J. & Yu, H. (2012), ‘Hint: High-quality protein interactomes and their applications in understanding human disease’, BMC systems biology 6(1), 92.
Dempster, A. P. (1972), ‘Covariance selection’, Biometrics 28(1), 157–175.
29
Deng, X. & Yuan, M. (2009), ‘Large gaussian covariance matrix estimation with markov structures’,
Journal of Computational and Graphical Statistics 18(3).
Donoho, D. L., Gavish, M. & Johnstone, I. M. (2013), ‘Optimal shrinkage of eigenvalues in the
spiked covariance model’, arXiv preprint arXiv:1311.0851 .
Fan, J., Fan, Y. & Lv, J. (2008), ‘High dimensional covariance matrix estimation using a factor
model’, Journal of Econometrics 147(1), 186–197.
Fan, J., Liao, Y. & Mincheva, M. (2013), ‘Large covariance estimation by thresholding principal
orthogonal complements’, Journal of the Royal Statistical Society: Series B (Statistical Methodology) 75(4), 603–680.
Friedman, J., Hastie, T. & Tibshirani, R. (2008), ‘Sparse inverse covariance estimation with the
graphical lasso’, Biostatistics 9, 432–441.
Hahsler, M. & Hornik, K. (2017), TSP: Traveling Salesperson Problem (TSP). R package version
1.1-5.
URL: https://CRAN.R-project.org/package=TSP
Hastie, T., Tibshirani, R. & Friedman, J. (2009), The Elements of Statistical Learning; Data
Mining, Inference and Prediction, Second Edition, Springer Verlag, New York.
Huang, J., Liu, N., Pourahmadi, M. & Liu, L. (2006), ‘Covariance matrix selection and estimation
via penalised normal likelihood’, Biometrika 93(1), 85.
Jacob, L., Obozinski, G. & Vert, J.-P. (2009), Group lasso with overlap and graph lasso, in ‘Proceedings of the 26th annual international conference on machine learning’, ACM, pp. 433–440.
Johnstone, I. M. (2001), ‘On the distribution of the largest eigenvalue in principal components
analysis’, Annals of statistics pp. 295–327.
Khare, K., Oh, S.-Y. & Rajaratnam, B. (2014), ‘A convex pseudolikelihood framework for high
dimensional partial correlation estimation with convergence guarantees’, Journal of the Royal
Statistical Society: Series B (Statistical Methodology) .
Lam, C. & Fan, J. (2009), ‘Sparsistency and rates of convergence in large covariance matrix estimation’, Annals of Statistics 37(6B), 4254–4278.
Le Cun, B. B., Denker, J. S., Henderson, D., Howard, R. E., Hubbard, W. & Jackel, L. D. (1990),
Handwritten digit recognition with a back-propagation network, in ‘Advances in neural information processing systems’, Citeseer.
Ledoit, O. & Wolf, M. (2004), ‘A well-conditioned estimator for large-dimensional covariance matrices’, Journal of multivariate analysis 88(2), 365–411.
Levina, E., Rothman, A. & Zhu, J. (2008), ‘Sparse estimation of large covariance matrices via a
nested lasso penalty’, The Annals of Applied Statistics pp. 245–263.
Liu, H., Wang, L. & Zhao, T. (2013), ‘Sparse covariance matrix estimation with eigenvalue constraints’, Journal of Computational and Graphical Statistics .
30
Liu, H., Wang, L. et al. (2017), ‘Tiger: A tuning-insensitive approach for optimally estimating
gaussian graphical models’, Electronic Journal of Statistics 11(1), 241–294.
Luo, X. (2011), ‘Recovering model structures from large low rank and sparse covariance matrix
estimation’, arXiv preprint arXiv:1111.1133 .
Meinshausen, N. & Bühlmann, P. (2006), ‘High dimensional graphs and variable selection with the
lasso’, Annals of Statistics 34, 1436–1462.
Obozinski, G., Jacob, L. & Vert, J. (2011), ‘Group lasso with overlaps: The latent group lasso
approach’, arXiv preprint arXiv:1110.0413 .
Parikh, N. & Boyd, S. P. (2014), ‘Proximal algorithms.’, Foundations and Trends in optimization
1(3), 127–239.
Peng, J., Wang, P., Zhou, N. & Zhu, J. (2009), ‘Partial correlation estimation by joint sparse
regression models’, Journal of the American Statistical Association 104(486).
R Core Team (2017), R: A Language and Environment for Statistical Computing, R Foundation
for Statistical Computing, Vienna, Austria.
URL: https://www.R-project.org/
Rothman, A. (2012), ‘Positive definite estimators of large covariance matrices’, Biometrika
99(3), 733–740.
Rothman, A. J., Levina, E. & Zhu, J. (2009), ‘Generalized thresholding of large covariance matrices’,
Journal of the American Statistical Association 104(485), 177–186.
Rothman, A., Levina, E. & Zhu, J. (2008), ‘Sparse permutation invariant covariance estimation’,
Electronic Journal of Statistics 2, 494–515.
Rothman, A., Levina, E. & Zhu, J. (2010), ‘A new approach to Cholesky-based covariance regularization in high dimensions’, Biometrika 97(3), 539.
Sun, Q., Tan, K. M., Liu, H. & Zhang, T. (2017), ‘Graphical nonconvex optimization for optimal
estimation in gaussian graphical models’, arXiv preprint arXiv:1706.01158 .
Tibshirani, R. (1996), ‘Regression shrinkage and selection via the lasso’, J. Royal. Stat. Soc. B.
58, 267–288.
Travers, J. & Milgram, S. (1969), ‘An experimental study of the small world problem’, Sociometry
pp. 425–443.
Tseng, P. (2001), ‘Convergence of a block coordinate descent method for nondifferentiable minimization’, Journal of optimization theory and applications 109(3), 475–494.
Wagaman, A. & Levina, E. (2009), ‘Discovering sparse covariance structures with the isomap’,
Journal of Computational and Graphical Statistics 18(3), 551–572.
Watts, D. J. & Strogatz, S. H. (1998), ‘Collective dynamics of ‘small-world’ networks’, Nature
393(6684), 440–442.
31
Wickham, H. (2009), ggplot2: elegant graphics for data analysis, Springer New York.
URL: http://had.co.nz/ggplot2/book
Won, J.-H., Lim, J., Kim, S.-J. & Rajaratnam, B. (2013), ‘Condition-number-regularized covariance estimation’, Journal of the Royal Statistical Society: Series B (Statistical Methodology)
75(3), 427–450.
Wu, W. B. & Pourahmadi, M. (2003), ‘Nonparametric estimation of large covariance matrices of
longitudinal data’, Biometrika 90(4), 831–844.
Wu, W. B. & Pourahmadi, M. (2009), ‘Banding sample autocovariance matrices of stationary
processes’, Statistica Sinica 19(4), 1755.
Xue, L., Ma, S. & Zou, H. (2012), ‘Positive-definite l1-penalized estimation of large covariance
matrices’, Journal of the American Statistical Association 107(500), 1480–1491.
Yan, X. & Bien, J. (2017), ‘Hierarchical sparse modeling: A choice of two group lasso formulations’,
Statistical Science 32(4), 531–560.
Yu, G. & Bien, J. (2017), ‘Learning local dependence in ordered data’, Journal of Machine Learning
Research 18(42), 1–60.
Yuan, M. (2010), ‘High dimensional inverse covariance matrix estimation via linear programming’,
The Journal of Machine Learning Research 11, 2261–2286.
Yuan, M. & Lin, Y. (2006), ‘Model selection and estimation in regression with grouped variables’,
Journal of the Royal Statistical Society, Series B 68, 49–67.
Yuan, M. & Lin, Y. (2007), ‘Model selection and estimation in the Gaussian graphical model’,
Biometrika 94(1), 19–35.
Zhu, Z. & Liu, Y. (2009), ‘Estimating spatial covariance using penalised likelihood with weighted
l1 penalty’, Journal of Nonparametric Statistics 21(7), 925–942.
32
| 10math.ST
|
Exploiting ConvNet Diversity for Flooding Identification
Keiller Nogueira1 , Samuel G. Fadel2 , Ícaro C. Dourado2 , Rafael de O. Werneck2 , Javier A. V.
Muñoz2 , Otávio A. B. Penatti3 , Rodrigo T. Calumby4 , Lin Tzy Li2,3 , Jefersson A. dos
Santos1 , Ricardo da S. Torres2
arXiv:1711.03564v1 [cs.CV] 9 Nov 2017
1
Departamento de Ciência da Computação, Universidade Federal de Minas Gerais (UFMG)
{keiller.nogueira,jefersson}@dcc.ufmg.br
2
Institute of Computing, University of Campinas (Unicamp)
{samuel.fadel,icaro.dourado,rafael.werneck,lintzyli,rtorres}@ic.unicamp.br
3
Advanced Technologies Group, SAMSUNG Research Institute
[email protected]
4
Department of Exact Sciences, University of Feira de Santana (UEFS)
[email protected]
November 13, 2017
Abstract
Flooding is the world’s most costly type of natural disaster in terms of both economic losses and
human causalities. A first and essential procedure towards flood monitoring is based on identifying
the area most vulnerable to flooding, which gives authorities relevant regions to focus. In this
work, we propose several methods to perform flooding identification in high-resolution remote
sensing images using deep learning. Specifically, some proposed techniques are based upon unique
networks, such as dilated and deconvolutional ones, while other was conceived to exploit diversity of
distinct networks in order to extract the maximum performance of each classifier. Evaluation of the
proposed algorithms were conducted in a high-resolution remote sensing dataset. Results show that
the proposed algorithms outperformed several state-of-the-art baselines, providing improvements
ranging from 1 to 4% in terms of the Jaccard Index.
Keywords: Flooding identification, Natural disaster, Remote Sensing, Inundation, MediaEval,
Satellites.
1
Introduction
Natural disaster monitoring is a fundamental task to create prevention strategies, as well as to help
authorities to act in the control of damages, coordinate rescues, and help victims. Among all kinds
of natural hazards, flooding is possibly the most extensive and devastating one, destroying buildings,
roads, bridges; tearing out trees; devastating agriculture; causing mudslides; and threatening human
lives [1]. All these consequences make such events to be considered as the world’s most costly type of
natural disaster in terms of both economic losses and human causalities as pointed out by the disaster
statistical review released every year by the Centre for Research on the Epidemiology of Disasters [2].
Some of these disasters happen annually [3], for example in humid tropics and subtropical climates,
where river flooding is a recurrent natural phenomenon due to excessive rain within a short period
of time. However, other events may occur atypically [4], which is the case of flooding caused by
hurricanes, such as the recent triple hurricanes (e.g., Harvey, Maria, and Irma) that brought massive
flooding in several countries [5].
Although extremely important, floods are difficult to monitor, because they are highly dependent
on several local conditions, such as precipitation, slope of terrain, drainage network, protective structures, and land cover [6]. A first and essential step towards such monitoring is based on identifying
1
areas most vulnerable to flooding, helping authorities to focus on such regions while monitoring inundations. Remotely sensed data play a crucial role in identifying such areas, since it allows the capture
of whole inundated regions during a flooding event, allowing a better understanding of what and how
it is flooded. Although there are lots of works [7,8] performing flooding detection using remote sensing
data integrated with elevation maps in order to augment the amount and type of information available
for an efficient flood management, as far as we know, there are no works that focus on identifying
flooding areas using only remote sensing data. Because of the importance of such task and the lack of
works dealing with it, a subtask (called Flood-Detection in Satellite Images) of the 2017 Multimedia
Satellite Task [9], which was part of the traditional MediaEval Benchmark, was proposed to leverage
the development of methods for identifying flooding events in high-resolution remote sensing images.
In this paper, we present our proposed methods, which won the aforementioned task, to automatic
identify flooding areas in high-resolution remote sensing images using deep learning paradigm. Deep
learning [10], commonly represented as multi-layered neural networks, can learn simultaneously the
features and the classifiers. In other words, during the training process, a deep neural network is able
to learn both the features and the classifier in a unified manner, adjusting itself to better represent
the characteristics of the data and their labels. Among all deep learning-based networks, a specific
type, called Convolutional (Neural) Networks, ConvNets or CNNs [10], is the most popular for learning
visual features in computer vision applications, including remote sensing [11,12]. This type of network
relies on the natural stationary property of an image, i.e., the statistics of one part of the image are
assumed to be the same as those of any other part. Furthermore, deep ConvNets can be considered
as an inherently multiscale approach since they usually obtain different levels of abstraction for the
data, ranging from local low-level information in the initial layers (e.g., corners and edges), to more
semantic descriptors, mid-level information (e.g., object parts) in intermediate layers, and high-level
information (e.g., whole objects) in the final layers.
We introduce several approaches to identify flooding areas of remote sensing images exploiting
the advantages of ConvNets. Some methods are based uniquely on networks with distinct properties,
including: (i) dilated convolutions [13], which, unlike standard ConvNets, process the input without
downsampling it, and (ii) deconvolution layers, such as SegNet [14], in which a coarse feature map is
upsampled outputting a dense map with the same resolution of the original image. Another approach
exploits the diversity of distinct networks in order to extract the maximum performance of each
classifier. In summary, the contributions of the paper are: (i) novel ConvNet architectures specialized
in identifying flooding areas; and (ii) a new strategy to exploit network diversity for inundation
identification. Obtained results of the proposed methods represent the state of the art, in terms of the
Jaccard Index, in a remote-sensing-based flooding detection task. These results made us the winner
of the Flood-Detection in Satellite Images, a subtask of 2017 Multimedia Satellite Task [9].
2
Proposed Methods
In this section, we present the proposed techniques to perform flooding identification. The first
approaches (Section 2.1) are based upon new and specific architectures to handle such important
task. The other method (Section 2.2) was conceived to better combine the diversity of distinct, but
complementary networks.
2.1
Network Architectures
All networks conceived specifically for flooding identification are based on ConvNets and have architectures illustrated in Figure 1. Some use dilated convolutions while others are based on deconvolutional
networks.
Specifically, two architectures are based on the concept of dilated convolutions [13]. In these layers,
the convolution filter is expanded by dilation rate. Given this rate, the weights are placed far away at
given intervals and the kernel size increases by allowing gaps (or “holes”) inside their filters. Therefore,
networks composed of these layers allow the receptive field to expand but preserving the resolution,
i.e., without downsampling the input data. This procedure represents a great advantage in terms
2
of computational processing, as well as in terms of learning, since internal feature maps do not lose
resolution (and maybe useful information).
The first architecture, presented in Figure 1a, is composed of seven layers: six dilated convolutions
(that are responsible for learning the patterns of the input images) and a final 1 × 1 convolution
layer, which is responsible for identifying the flooding areas. There are no pooling or normalization
operations inside this network. Specifically, the first two convolutions have 5 × 5 filters with dilation
rate 1. The following two convolutions have 4 × 4 filters and rate 2 while the last two convolutions
have smaller filters (3 × 3) with 4 dilation rate.
The second proposed network, presented in Figure 1b, is also based on dilated convolution layers.
This network shares the same architecture of the first one, except for the fact that there are additional
max-pooling layers between each dilated convolutions layer. Each pooling has 3 × 3 kernel and stride
of 1. Although pooling usually reduces the resolution of the input data, in this case, it also preserves
the resolution because of zero padding used in the input. The core idea of this network is to conserve
the benefits of dilated convolutional layers but adding known advantages of pooling layers, such as
invariance to small rotations and translations, as well as the preservation of the most important
features.
inputDimage
overlapping
patches
DilatedDConvolutionD1
DilatedDConvolutionD2
KernelD=D5x5
ReLUs =D64
Dilatation RateD=D1
Kernel =D5x5
ReLUsD=D64
Dilatation RateD=D1
Kernel =D4x4
ReLUsD=D128
Dilatation RateD=D2
2Dclasses:
flooding
non-flooding
KernelD=D3x3
ReLUsD=D256
Dilatation RateD=D4
KernelD=D3x3
ReLUs =D256
Dilatation Rate =D4
Kernel =D4x4
ReLUs =D128
Dilatation Rate =D2
ClassificationDLayer
DilatedDConvolutionD6
...
DilatedDConvolutionD5
DilatedDConvolutionD3
DilatedDConvolutionD4
predictedDimage
(a) Dilated ConvNet #1
inputDimage
overlapping
patches
...
Dilated
ConvolutionD1
KernelD=D5x5
ReLUs =D64
Dilatation
RateD=D1
2Dclasses:
flooding
non-flooding
KernelD=D3x3
StrideD=D1x1
Classification
Layer
Max-PoolingD6
Max-PoolingD1
KernelD=D3x3
StrideD=D1x1
KernelD=D3x3
ReLUs =D256
Dilatation
Rate =D4
Dilated
ConvolutionD6
Dilated
ConvolutionD2
KernelD=D5x5
ReLUs =D64
Dilatation
RateD=D1
KernelD=D3x3
StrideD=D1x1
Max-PoolingD5
Max-PoolingD2
KernelD=D3x3
StrideD=D1x1
KernelD=D3x3
ReLUs =D256
Dilatation
Rate =D4
Dilated
ConvolutionD5
Dilated
ConvolutionD3
Kernel =D4x4
ReLUsD=D128
Dilatation
RateD=D2
KernelD=D3x3
StrideD=D1x1
Max-PoolingD4
Max-PoolingD3
KernelD=D3x3
StrideD=D1x1
Kernel =D4x4
ReLUsD=D128
Dilatation
RateD=D2
Dilated
ConvolutionD4
predictedDimage
(b) Dilated ConvNet #2
inputyimage
overlapping
patches
Convolutiony1
Max-Poolingy1
Convolutiony2
Max-Poolingy2
Convolutiony3
Max-Poolingy3
...
Kernely=y5x5
ReLUs =y64
Kernely=y2x2
Stridey=y2x2
Kernely=y3x3
ReLUs =y128
Kernely=y2x2
Stridey=y2x2
Kernely=y3x3
ReLUs =y256
Kernely=y2x2
Stridey=y2x2
2yclasses:
flooding
non-flooding
Kernely=y5x5
ReLUs =y64
Kernely=y2x2
ReLUsy=y64
Stridey=y2x2
Kernely=y3x3
ReLUs =y64
Kernely=y2x2
ReLUsy=y64
Stridey=y2x2
Kernely=y3x3
ReLUs =y128
Kernely=y2x2
ReLUsy=y128
Stridey=y2x2
Classification
Layer
Convolutiony6
Deconvolutiony3
Convolutiony5
Deconvolutiony2
Convolutiony4
Deconvolutiony1
predictedyimage
(c) Deconvolution ConvNet #1
inputMimage
overlapping
patches
ConvolutionM1
Max-PoolingM1
ConvolutionM2
Max-PoolingM2
ConvolutionM3
Max-PoolingM3
...
KernelM=M4x4
ReLUs =M64
KernelM=M2x2
StrideM=M2x2
KernelM=M4x4
ReLUs =M128
KernelM=M2x2
StrideM=M2x2
KernelM=M3x3
ReLUs =M256
KernelM=M2x2
StrideM=M2x2
2Mclasses:
flooding
non-flooding
KernelM=M3x3
ReLUs =M64
KernelM=M3x3
ReLUsM=M64
StrideM=M2x2
KernelM=M4x4
ReLUs =M128
KernelM=M3x3
ReLUsM=M128
StrideM=M2x2
KernelM=M3x3
ReLUs =M128
KernelM=M3x3
ReLUsM=M256
StrideM=M1x1
Classification
Layer
ConvolutionM6
DeconvolutionM3
ConvolutionM5
DeconvolutionM2
ConvolutionM4
DeconvolutionM1
predictedMimage
(d) Deconvolution ConvNet #2
Figure 1: ConvNet architectures proposed in this work.
The two remaining networks are based on deconvolutional networks [14]. This type of network has
two modules: the first receives input images, learns the visual features by using standard convolution
and pooling layers, and outputs a coarse feature map while the second receives this map as input,
3
learns to upsample these features by using several deconvolution layers, and outputs a dense prediction
map with the same resolution of the original image. Both modules work together without distinction
and can be trained end-to-end by using standard feedforward and backpropagation algorithms.
The two architectures, presented in Figures 1c and 1d, are based on the solution discussed in [14].
The encoder of both is based on three standard convolution layers intercalated with max-pooling
operations. The only difference is the kernel of the convolutions, i.e., while the first layer uses a larger
kernel in the first layer (in order to learn more information about flooding regions directly from the
input image), the other one bets on similar filter sizes, which should extract the same amount of
information in all layers resulting in a balanced learning of the flooding patterns. The deconvolutional
part has similar architecture with the same number of layers (three deconvolutional ones intercalated
with standard convolutions) differing only in the size of deconvolutional and convolutional kernels.
The premise was conserved, i.e., while the first deconvolutional network bets on larger kernel, the
second tries to use similar filter sizes.
2.2
Combination
As previously explained, we also proposed another strategy to solve the flooding detection task, which
aims to exploit the diversity of distinct ConvNets. The main premise of the proposed method is that the
previous presented ConvNets learn and produce distinct outcomes, which are dense prediction maps.
This difference should make ConvNets complementary to each other. Therefore, a clever combination
of such outcomes should improve the final prediction map if compared with the ConvNets individual
results.We propose a combination method based upon Support Vector Machines (SVM).
DistinctFConvNets
TrainedFwithFDifferentFData
TestFImage
DilatedNetF#1
ConcatenatedFPredictions
FinalFPrediction
DilatedNetF#2
...
LinearFSVM
DeconvNetF#1
DeconvNetF#2
Figure 2: Pipeline for the prediction phase of the proposed combination approach.
The proposed method is divided into three main steps: (i) extraction: In this phase, an image
is processed by all proposed network, which produce distinct outcomes (i.e., different probability or
prediction maps). All these maps (that have the same resolution of the original input image) are then
concatenated creating a feature vector that, in fact, represents the input image. (ii) learning: In this
step, the SVM receives the aforementioned feature vector as well as the ground-truth flooding map
for all training data. Then, it independently process each pixel of these images, learning which and
when each classifier is better; and (iii) prediction: This final step receives feature vectors of testing
images and, using the trained SVM, outputs the improved prediction map for each test image. This
final step is illustrated in Figure 2.
3
Experiments
In this section, we present the experimental setup. Also, we present and discuss the obtained results
of the proposed methods comparing them with the best performing teams of the Flood-Detection in
Satellite Images subtask of the 2017 Multimedia Satellite Task.
4
3.1
Dataset
(a) Location 1
(b) Location 2
(c) Location 3
(d) Location 4
(e) Location 5
(f) Location 6
Figure 3: Examples of patches for all locations of the training set with its respective ground-truth, in
which white regions refer to flooded area and black ones correspond to background.
The dataset consists of satellite image patches collected from eight different flooding events June 1st,
2016 to May 1st, 2017. Each image patch is composed of four bands (red, green, blue, and near
infrared bands) and has resolution of 320 × 320 pixels, with a ground-sample distance (GSD) of 3.7
meters and an orthorectified pixel size of 3 meters [9].
The training set is composed of 462 image patches unevenly extracted from six locations. Among
these images, 92 (20%) were employed as internal validation set to evaluate the proposed algorithms
while the remaining 370 images were used to train the proposed methods. Some examples of image
patches for each of these locations are presented in Figure 3. Two test sets were released in this
dataset: the Same Locations test set contains 216 unseen patches unevenly extracted from the same
region presented in the training set, while the New Locations test set contains 58 unseen patches
extracted from a region not present in the training set. Figure 4 presents some examples of both test
sets. It is important to highlight that, until the submission of this current paper, ground-truth of the
test sets were not released by the organization of the competition.
5
(a) Location 1
(b) Location 2
(e) Location 5
(c) Location 3
(f) Location 6
(d) Location 4
(g) New Location
Figure 4: Examples of patches for both test sets.
3.2
Experimental Evaluation
In order to assess the performance of generated segmentation masks for flooded areas in the satellite
image patches, the intersection-over-union metric (also known as Jaccard Index), was used. The
metric measures the accuracy for the pixelwise classification, and as defined as IoU = (T P +FT PP +F N ) ,
where T P , F P , and F N are the numbers of true positive, false positive, and false negative pixels,
respectively, determined over the whole test set.
3.3
Experimental Protocol
First, we trained all ConvNets (presented in Section 2.1) using overlapping patches of size 25 × 25
extracted from all training images (independent of the location). In the prediction phase, we also
extracted overlapping patches with the same resolution from the testing images and averaged the
probabilities outputted by the network. Among all networks, the best one (in our internal validation
set) is reported as ConvNet 25 × 25.
Another proposed method relied on training the aforementioned ConvNets using larger overlapping
patches, with 50 × 50 pixels, also extracted from all training images. The motivation behind this
strategy is based on the entire context that could be extract from the input patches and improve
the learning process. The prediction phase is similar to the previous strategy. Considering this
configuration, the best network (in our internal validation set) is referred to in the next section as
ConvNet 50 × 50.
The Location ConvNets strategy is based on the idea of creating specialized ConvNets for each
flooding event. Since the dataset has six distinct flooding event locations, we propose to train a
specific Dilated ConvNet #1 (using patches of 25 × 25) for each location. The prediction is similar
to the other proposed protocols, except for the fact that, in this case, each ConvNet was used in its
respective location. For the New Locations test set, we combined the outcomes extracted from each
ConvNet (trained specifically for each location) using a linear SVM, as proposed in Section 2.2.
The Fusion-SVM strategy expands above idea. Differently from previous scheme, in this procedure, an SVM is used to create prediction maps for both test sets (and not only for the New
Locations test set). Based on the premise that distinct ConvNets (trained using different input data)
produce distinct (and possible complementary) outcomes, we propose to combine the predictions extracted from all above ConvNets using a linear SVM, as presented in Section 2.2. In this way, the
SVM should be able to learn when and how these networks complement each other in order to improve
6
the final performance. Specifically, the SVM receives as input concatenated probabilities extracted
from all previously trained ConvNets, which include: (i) all four ConvNets presented in Section 2.1
trained with overlapping 25 × 25 patches, (ii) all four networks introduced in Section 2.1 trained with
overlapping 50 × 50 patches, and (iii) six Dilated ConvNet #1 trained specifically for each location of
the training set.
Another strategy relied on exploiting the diversity of distinct ConvNets by combining all outcomes
of previous methods using a majority voting scheme, which is referred to in following section as
Fusion-MV.
It is important to emphasize that all proposed methods were created using TensorFlow framework
and will have code released upon the acceptance of this paper. When training, all the aforementioned
protocol used the same hyper-parameters, i.e., learning rate, weight decay, momentum and number of
epochs as 0.01, 0.0005, 0.9 and 20, respectively.
3.4
Baselines
The baselines evaluated in this work were, in fact, the best performing approaches proposed for the
Flood-Detection in Satellite Images subtask of the 2017 Multimedia Satellite Task. An overview of
the such methods (which includes state-of-the-art methods, such as Generative Adversarial, Deconvolutional and Fully Convolutional Networks) are presented in Table 1.
3.5
Results and Discussion
All results for the test sets are presented in Table 1. These are the official results released by the
Mediaeval since no ground-truth for the test set was released yet. Also, it is worth mentioning that,
for the two strategies of learning ConvNets using all available training data (but with distinct patch
size), the Dilated ConvNet #1 yielded the best results among all experimented networks. Therefore,
these were the ones submitted to the competition as well as reported in this work as ConvNet 25 × 25
and ConvNet 50 × 50.
For both test sets, the best solution was obtained by combining the probabilities of all trained
ConvNets using a Linear SVM. This technique yielded state-of-the-art results in both test set, outperforming all baselines by, at least, 4% in the Same Locations test set and 1% in the New Locations
test set (in terms of Jaccard Index). Some samples of this obtained results are presented in Figure 5.
For the Same Locations test set, training a network for each location (Location ConvNets) or
training a ConvNet with all available data (ConvNet 25 × 25) achieved the same result. This may
indicate that the proposed architecture can, in fact, extract and interpret all feasible information from
the whole data, which is a great advantage given that it reduces the number of networks to train and,
consequently, the processing time. This conclusion does not hold for the New Locations test set. In
this set, training a specific network for each location (Location ConvNets) achieved higher performance
(aside the Fusion-SVM strategy) when compared to unique networks trained with the whole training
set (such as ConvNet 25 × 25 and ConvNet 50 × 50). This indicates that specific Location Network
can learn details that may not be useful for classification of a known image, but that is important for
unseen data, which is the case.
Another relevant outcome is that increasing the size of the input patch (ConvNet 50×50) decreases
the final result, a conclusion that holds for both datasets. We believe that this is because of the amount
of training patches generated in each case. More specifically, a large amount of data may be used for
training with smaller patch sizes while large patches means less data to train. This corroborates with
the fact that deep learning really needs a large amount of labeled data to train [10].
Finally, for both sets, the worst result was obtained using the majority voting scheme. This may
be justified by the fact that Majority Voting is not so robust to aggregate information from multiple
networks, when they disagree in the classification. This fact can be overcome by using a machine
learning technique to capture about the opinions of the ConvNets.
7
Table 1: IoU (%) results of the proposed method and baselines for both test sets. Higher values of
IoU indicates better performance.
Methods
Test Set
Overview
Same Locations
WISC [15]
NDVI plus SVM-RBF
K-Means to cluster and classify
80
81
83
77
CERTH-ITI [16]
Mahalanobis dist. with stratified cov.
75
56
BMC [17]
ResNet-152 and random forest
37
40
UTAOS [18]
Gen.
Gen.
Gen.
Gen.
Gen.
82
80
83
83
81
73
70
74
74
73
DFKI [19]
VGG13-FCN with RGB data
VGG13-FCN with RGB and NIR data
VGG13 adapted to be a DeconvNet
73
84
84
69
70
74
Dilated 25 × 25
Dilated 50 × 50
Location ConvNets
Fusion-SVM
Fusion-MV
Dilated ConvNet #1 (25 × 25 patches)
Dilated ConvNet #1 (50 × 50 patches)
Dilated ConvNet #1 trained per location
SVM over concatenated predictions
MV over concatenated predictions
87
86
87
88
78
82
80
84
84
49
Baselines
Proposed
4
New Locations
Adv.
Adv.
Adv.
Adv.
Adv.
Net.
Net.
Net.
Net.
Net.
with
with
with
with
with
0.78
0.94
0.50
0.35
0.12
threshold
threshold
threshold
threshold
threshold
Conclusion
In this paper, we propose new approaches based on Convolutional Neural Networks to perform detection of flooding areas in remote sensing images. Specifically, we proposed four distinct architectures
based on dilated convolutions [13] and deconvolution layers [14]. Furthermore, different strategies to
combine such networks were proposed and evaluated.
Experimental results have showed that the methods are effective and robust. We have achieved
state-of-the-art performance, in terms of Jaccard Index, in a specific dataset proposed for the FloodDetection in Satellite Images subtask of the 2017 Multimedia Satellite Task. The proposed methods
outperformed all baselines, winning that subtask challenge. Such results show that our proposed
approaches are effective and robust to identify flooding areas (independent if it is for a recurrent or
atypical event). This identification process performed by our proposed algorithms may help authorities
to keep focus on most vulnerable regions while monitoring forecast inundations, which may aid in
coordinate rescues, and help victims.
As future work, we intend to use different post-processing methods, such as Conditional Random
Fields, in order to exploit the contextual information.
Acknowledgments
The authors thank FAPESP (grants #2013/50169-1, #2013/50155-0, #2014/50715-9, #2014/122361, and #2016/18429-1), FAPEMIG, CNPq, and CAPES.
References
[1] NOAA/NWS (National Weather Service), Floods: The awesome power (2005).
[2] CRED, Centre for research on the epidemiology of disasters (cred) (2017).
URL http://www.emdat.be/
8
[3] T. De Groeve, Flood monitoring and mapping using passive microwave remote sensing in namibia,
Geomatics, Natural Hazards and Risk 1 (1) (2010) 19–35.
[4] B. P. Harman, S. Heyenga, B. M. Taylor, C. S. Fletcher, Global lessons for adapting coastal
communities to protect against storm surge inundation, Journal of Coastal Research 31 (4) (2013)
790–801.
[5] C. Eric Levenson, 3 storms, 3 responses: Comparing harvey, irma and maria (2017).
URL http://edition.cnn.com/2017/09/26/us/response-harvey-irma-maria/index.html
[6] V. Klemas, Remote sensing of floods and flood-prone areas: an overview, Journal of Coastal
Research 31 (4) (2014) 1005–1013.
[7] L. Pulvirenti, N. Pierdicca, G. Boni, M. Fiorini, R. Rudari, Flood damage assessment through
multitemporal cosmo-skymed data and hydrodynamic models: The albania 2010 case study,
JSTARS 7 (7) (2014) 2848–2855.
[8] A. D’Addabbo, A. Refice, G. Pasquariello, F. P. Lovergine, D. Capolongo, S. Manfreda, A
bayesian network for flood detection combining sar imagery and ancillary data, TGRS 54 (6)
(2016) 3612–3625.
[9] B. Bischke, P. Helber, C. Schulze, S. Venkat, A. Dengel, D. Borth, The multimedia satellite task
at mediaeval 2017: Emergence response for flooding events, in: Proc. of the MediaEval 2017
Workshop, Dublin, Ireland.
[10] I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT Press, 2016.
[11] K. Nogueira, W. O. Miranda, J. A. Dos Santos, Improving spatial feature representation from
aerial scenes by using convolutional networks, in: Conference on Graphics, Patterns and Images
(SIBGRAPI), IEEE, 2015, pp. 289–296.
[12] K. Nogueira, O. A. Penatti, J. A. dos Santos, Towards better exploiting convolutional neural
networks for remote sensing scene classification, Pattern Recognition 61 (2017) 539–556.
[13] F. Yu, V. Koltun, Multi-scale context aggregation by dilated convolutions, in: ICLR, 2016.
[14] V. Badrinarayanan, A. Kendall, R. Cipolla, Segnet: A deep convolutional encoder-decoder architecture for image segmentation, arXiv preprint arXiv:1511.00561.
[15] N. Tkachenko, A. Zubiaga, R. Procter, Wisc at mediaeval 2017: Multimedia satellite task, in:
Working Notes Proc. MediaEval Workshop, 2017, p. 2.
URL http://slim-sig.irisa.fr/me17/Mediaeval_2017_paper_12.pdf
[16] K. Avgerinakis, A. Moumtzidou, S. Andreadis, E. Michail, I. Gialampoukidis, S. Vrochidis,
I. Kompatsiaris, Visual and textual analysis of social media and satellite images for flood detection @ multimedia satellite task mediaeval 2017, in: Working Notes Proc. MediaEval Workshop,
2017, p. 2.
URL http://slim-sig.irisa.fr/me17/Mediaeval_2017_paper_31.pdf
[17] X. Fu, Y. Bin, L. Peng, J. Z. Y. Yang, H. T. Shen, Bmc@mediaeval 2017 multimedia satellite
task via regression random forest, in: Working Notes Proc. MediaEval Workshop, 2017, p. 2.
URL http://slim-sig.irisa.fr/me17/Mediaeval_2017_paper_46.pdf
[18] K. Ahmad, P. Konstantin, M. Riegler, N. Conci, P. Holversen, Cnn and gan based satellite and
social media data fusion for disaster detection, in: Working Notes Proc. MediaEval Workshop,
2017, p. 2.
URL http://slim-sig.irisa.fr/me17/Mediaeval_2017_paper_15.pdf
9
[19] B. Bischke, P. Bhardwaj, A. Gautam, P. Helber, D. Borth, A. Dengel, Detection of flooding
events in social multimedia and satellite imagery using deep neural networks, in: Working Notes
Proc. MediaEval Workshop, 2017, p. 2.
URL http://slim-sig.irisa.fr/me17/Mediaeval_2017_paper_51.pdf
10
(a) Location 1
(b) Location 2
(c) Location 3
(d) Location 4
(e) Location 5
(f) Location 6
(g) New Location
Figure 5: Examples of some test images and the obtained results achieved by using SVM with the
aggregated probabilities. White areas refer to flooded regions and black areas correspond to background.
11
| 1cs.CV
|
1
Z2Z4-Additive Cyclic Codes: Kernel and Rank
Joaquim Borges
Steven T. Dougherty
Cristina Fernández-Córdoba
arXiv:1711.03784v2 [cs.IT] 16 Jan 2018
Roger Ten-Valls
Abstract
β
A Z2 Z4 -additive code C ⊆ Zα
2 × Z4 is called cyclic if the set of coordinates can be partitioned
into two subsets, the set of Z2 coordinates and the set of Z4 coordinates, such that any cyclic shift of
the coordinates of both subsets leaves the code invariant. Let Φ(C) be the binary Gray map image of
C. We study the rank and the dimension of the kernel of a Z2 Z4 -additive cyclic code C, that is, the
dimensions of the binary linear codes hΦ(C)i and ker(Φ(C)). We give upper and lower bounds for these
parameters. It is known that the codes hΦ(C)i and ker(Φ(C)) are binary images of Z2 Z4 -additive codes
R(C) and K(C), respectively. Moreover, we show that R(C) and K(C) are also cyclic and we determine
the generator polynomials of these codes in terms of the generator polynomials of the code C.
Keywords Z2 Z4 -additive cyclic codes, Gray map, kernel, rank.
I. I NTRODUCTION
Denote by Z2 and Z4 the rings of integers modulo 2 and modulo 4, respectively. We denote
the space of n-tuples over these rings as Zn2 and Zn4 . A binary code is any non-empty subset C
J. Borges is with the Department of Information and Communications Engineering, Universitat Autònoma de Barcelona,
08193-Bellaterra, Spain (e-mail: [email protected]).
S. T. Dougherty is with the Department of Mathematics, University of Scranton, Scranton, PA 18510, USA (email:[email protected]).
C. Fernández-Córdoba is with the Department of Information and Communications Engineering, Universitat Autònoma de
Barcelona, 08193-Bellaterra, Spain (e-mail: [email protected]).
R. Ten-Valls is with the Department of Information and Communications Engineering, Universitat Autònoma de Barcelona,
08193-Bellaterra, Spain (e-mail: [email protected]).
This work has been partially supported by the Spanish MINECO grants TIN2016-77918-P and MTM2015-69138-REDT, and
by the Catalan AGAUR grant 2014SGR-691.
January 20, 2018
DRAFT
2
of Zn2 , and if that subset is a vector space then we say that it is a linear code. Any non-empty
subset C of Zn4 is a quaternary code and a submodule of Zn4 is called a linear code over Z4 .
In 1994, Hammons et al. discovered that some good non-linear binary codes can be seen as
the Gray map images of linear codes over Z4 , [9]. From then on, the study of codes over Z4
and other finite rings has been developing and the construction of Gray maps has been a topic
of study.
In Delsarte’s 1973 paper (see [5]), he defined additive codes as subgroups of the underlying
abelian group in a translation association scheme. For the binary Hamming scheme, namely
when the underlying abelian group is of order 2n , the only structures for the abelian group are
those of the form Zα2 × Zβ4 , with α + 2β = n. This means that the subgroups C of Zα2 × Zβ4 are
the only additive codes in a binary Hamming scheme. Hence, the study of codes in Zα2 × Zβ4
became important. These codes are called Z2 Z4 -additive codes. In recent times, the structure
and properties of Z2 Z4 -additive codes have been intensely studied (see [2]).
In [1], Z2 Z4 -additive cyclic codes were introduced and in [3], the duality was studied. In this
paper, we study the rank and the kernel of Z2 Z4 -additive cyclic codes, taking into account the
known results for general codes over Z4 and Z2 Z4 -additive codes [6], [7], [8].
The paper is organized as follows. In Section II, we recall the necessary concepts and properties
on Z2 Z4 -additive codes and Z2 Z4 -additive cyclic codes. In Section III, we give the main results
of the paper about the rank and the kernel of Z2 Z4 -additive cyclic codes. We prove that both
the binary span and the kernel are binary images of Z2 Z4 -additive cyclic codes, we determine
the possible values of the dimensions of the corresponding binary images, and we compute the
generator polynomials of these codes.
II. P RELIMINARIES
A. Z2 Z4 -additive codes
A Z2 Z4 -additive code C is a subgroup of Zα2 × Zβ4 (see [2]). Since C is a subgroup of Zα2 × Zβ4 ,
it is also isomorphic to a commutative structure of the form Zγ2 × Zδ4 and it has |C| = 2γ+2δ
codewords.
Let X (respectively Y ) be the set of Z2 (respectively Z4 ) coordinate positions, so |X| = α
and |Y | = β. Unless otherwise stated, the set X corresponds to the first α coordinates and Y
corresponds to the last β coordinates. Let CX be the binary punctured code of C formed by
deleting the coordinates outside X. Define similarly the quaternary code CY .
January 20, 2018
DRAFT
3
Let Cb be the subcode of C generated by all order two codewords and let κ be the dimension
of (Cb )X , which is a binary linear code. For the case α = 0, we write κ = 0. With all these
parameters, we say that a code C is of type (α, β; γ, δ; κ).
For a vector u ∈ Zα2 × Zβ4 we write u = (u | u′ ), where u = (u0 , . . . , uα−1 ) ∈ Zα2 and
u′ = (u′0 , . . . , u′β−1) ∈ Zβ4 .
In [2], it is shown that a Z2 Z4 -additive code is permutation equivalent to a Z2 Z4 -additive code
with standard generator matrix of the form:
0
0
Iκ Tb 2T2
GS = 0 0 2T1 2Iγ−κ 0
,
0 Sb Sq
R
Iδ
(1)
where Ik is the identity matrix of size k × k; Tb , Sb are matrices over Z2 ; T1 , T2 , R are matrices
over Z4 with all entries in {0, 1} ⊂ Z4 ; and Sq is a matrix over Z4 .
A Z2 Z4 -additive code C is said to be separable if C = CX × CY . Otherwise the code is said
to be non-separable.
Let u′ = (u′0 , . . . , u′n−1) be an element of Zn4 such that u′i = ũ′i + 2û′i , for i = 0, . . . , n − 1
and with ũ′i , û′i ∈ {0, 1}. As in [9], the Gray map φ from Zn4 to Z2n
2 is defined by
φ(u′) = (û′0 , . . . , û′n−1, ũ′0 + û′0 , . . . , ũ′n−1 + û′n−1).
The extended Gray map Φ is the map from Zα2 × Zβ4 to Zα+2β
given by
2
Φ(u | u′) = (u | φ(u′)).
B. Z2 Z4 -additive cyclic codes
Cyclic codes have been a primary area of study for coding theory, [10]. Recently, the class
of Z2 Z4 -additive cyclic codes has been defined in [1].
Let u = (u | u′) ∈ Zα2 × Zβ4 , then the cyclic shift π is given by π(u) = (π(u) | π(u′)) where
π(u) = π(u0 , u1, . . . , uα−1 ) = (uα−1 , u0 , u1, . . . , uα−2 ) and π(u′ ) = (u′β−1 , u′0, u′1 , . . . , u′β−2) .
We say that a Z2 Z4 -additive code C is cyclic if π(C) = C.
There exists a bijection between Zα2 × Zβ4 and Rα,β = Z2 [x]/(xα − 1) × Z4 [x]/(xβ − 1) given
by:
(u0 , u1, . . . , uα−1 | u′0 , . . . , u′β−1) 7→
(u0 + u1 x + · · · + uα−1 xα−1 | u′0 + · · · + u′β−1xβ−1 ).
January 20, 2018
DRAFT
4
Therefore, as usual in the study of cyclic codes, any codeword is identified as a vector or as a
polynomial.
From now on, the binary reduction of a polynomial p(x) ∈ Z4 [x] will be denoted by p̃(x).
Let p(x) ∈ Z4 [x] and (b(x) | a(x)) ∈ Rα,β and consider the following multiplication p(x) ⋆
(b(x) | a(x)) = (p̃(x)b(x) | p(x)a(x)). From [1], Rα,β is a Z4 [x]-module with respect to this
multiplication.
Let u′ (x) = ũ′ (x) + 2û′ (x) be the polynomial representation of u′ ∈ Zn4 . Then, the polynomial
version of the Gray map is φ(u′ (x)) = (û′(x), ũ′ (x) + û′(x)). In the following, a polynomial
p(x) ∈ Z2 [x] or Z4 [x] will be denoted simply by p.
Using the polynomial representation, an equivalent definition of Z2 Z4 -additive cyclic codes is
the following.
Definition 1 ([1]). A subset C ⊆ Rα,β is called a Z2 Z4 -additive cyclic code if C is a Z4 [x]submodule of Rα,β .
From [1], if β is odd, we know that if C is a Z2 Z4 -additive cyclic code then it is of the form
h(b | 0), (ℓ | f h + 2f )i,
(2)
where f hg = xβ −1 in Z4 [x], b divides xα −1 in Z2 [x], and we can assume that deg(ℓ) < deg(b).
The polynomials satisfying these conditions are said to be in standard form. In this case, we
have that |C| = 2α−deg(b) 4deg(g) 2deg(h) . From now on, we assume that β is odd. Then f , g and
h are pairwise coprime polynomials. Since h and g are coprime, there exist polynomials λ and
µ, that will be used later along the paper, such that
λh + µg = 1.
(3)
Lemma 2 ([3, Corollary 2]). Let C be a Z2 Z4 -additive cyclic code of type (α, β; γ, δ; κ) with
C = h(b | 0), (ℓ | f h + 2f )i. Then, b divides
We can put the generator matrix (1)
Iκ 1 T
0 Iκ 2
0 0
0 0
January 20, 2018
xβ −1
f˜
gcd(b, ℓ) and b divides h̃ gcd(b, ℓg̃).
in the following form, called the standard form:
0
0
Tb1 0
Tb2 2T2
0
0
.
0 2T1 2Iγ−(κ1 +κ2 ) 0
R
Iδ
S′ S
(4)
DRAFT
5
The next theorem relates the parameters of the type of a Z2 Z4 -additive code to its generator
polynomials.
Theorem 3 ([3, Theorem 5 and Proposition 6]). Let C be a Z2 Z4 -additive cyclic code of type
(α, β; γ, δ; κ) with C = h(b | 0), (ℓ | f h + 2f )i, where f hg = xβ − 1. Then
γ = α − deg(b) + deg(h),
δ = deg(g),
κ = α − deg(gcd(ℓg̃, b))
and
κ1 = α − deg(b),
κ2 = deg(b) − deg(gcd(b, ℓg̃)),
δ1 = deg(gcd(b, ℓg̃)) − deg(gcd(b, ℓ)) and δ2 = deg(g) − δ1 .
It is well known that if C is a Z2 Z4 -additive code, then the Z2 Z4 -linear code C = Φ(C) is
not linear in general. The linearity of these codes was studied in [8]. The key to establish this
linearity was the fact that
Φ(v + w) = Φ(v) + Φ(w) + Φ(2v ∗ w),
(5)
where ∗ denotes the component-wise product. It follows immediately that Φ(C) is linear if and
only if 2u ∗ v ∈ C, for all u, v ∈ C.
It is shown in [4] that, for a Z2 Z4 -additive cyclic code C, CX is a binary cyclic code and CY
is a linear cyclic code over Z4 . Moreover, if Φ(C) is linear, then φ(CY ) is also linear but the
converse is not true in general. The characterization of linear cyclic codes over Z4 of odd length
whose Gray map images are linear binary codes was given in [11]. Let p be a divisor of xn − 1
in Z2 [x] with n odd and let ξ be a primitive nth root of unity over Z2 . The polynomial (p ⊗ p)
is defined as the divisor of xn − 1 in Z2 [x] whose roots are the products ξ i ξ j such that ξ i and
ξ j are roots of p.
Theorem 4 ([11, Theorem 20]). Let D = hf h + 2f i be a Z4 -additive cyclic code of odd length
n and where f hg = xn − 1. The following properties are equivalent.
1) gcd(f˜, (g̃ ⊗ g̃)) = 1 in Z2 [x];
2) φ(D) is a binary linear code of length 2n.
January 20, 2018
DRAFT
6
This result was generalized for Z2 Z4 -additive cyclic codes of type (α, β; γ, δ; κ) with β odd.
Theorem 5 ([4]). Let C = h(b | 0), (ℓ | f h + 2f )i be a Z2 Z4 -additive cyclic code of length
α + β, β odd, and where f hg = xβ − 1. The following properties are equivalent.
˜
fb
, (g̃ ⊗ g̃)) = 1 in Z2 [x];
1) gcd( gcd(b,ℓg̃)
2) Φ(C) is a binary linear code of length α + 2β.
As a result, it is completely characterized when a Z2 Z4 -additive code C has binary linear
image under the Gray map, just considering its generator polynomials. The next step is study
the rank and the dimension of the kernel for those codes whose image Φ(C) is not linear.
III. K ERNEL
AND
R ANK
OF
Z2 Z4 -A DDITIVE C YCLIC C ODES
For an additive code C ⊆ Zα2 × Zβ4 , the kernel of Φ(C) is defined as ker(Φ(C)) = {v ∈
Zα+2β
2
| v + Φ(C) = Φ(C)}. Define K(C) = {v ∈ Zα2 × Zβ4 | Φ(v) ∈ ker(Φ(C))}. Let
rank(Φ(C)) = dim(hΦ(C)i) and R(C) = {v | v ∈ Zα2 × Zβ4 , φ(v) ∈ hΦ(C)i}. It is clear that
K(C) ⊆ C ⊆ R(C).
It is known that if C is a Z2 Z4 -additive code, then hΦ(C)i and ker(Φ(C)) are both Z2 Z4 -linear
codes ([8]). Therefore, R(C) and K(C) are both Z2 Z4 -additive codes. In the next sections we will
see that if the code C is a Z2 Z4 -additive cyclic code, then R(C) and K(C) are also Z2 Z4 -additive
cyclic. Therefore, the following proposition will be useful to relate the generator polynomials
of C to the generator polynomials of R(C) and K(C).
Proposition 6. Let C0 = h(b | 0), (ℓ | f h + 2f )i and C1 = h(b′ | 0), (ℓ′ | f ′ h′ + 2f ′ )i be
Z2 Z4 -additive cyclic codes with C0 ⊆ C1 . Then
1) f ′ divides f ;
2) gcd(b′ , ℓ′ ) divides gcd(b, ℓ).
Proof. Since C0 ⊆ C1 , we have that (C0 )Y = hf h + 2f i ⊆ (C1 )Y = hf ′ h′ + 2f ′ i. Therefore, by
[6, Theorem 3], f ′ divides f .
Finally, since C0 and C1 are cyclic Z2 Z4 -additive codes, then clearly (C0 )X = hgcd(b, ℓ)i and
(C1 )X = hgcd(b′ , ℓ′ )i. Then, since (C0 )X ⊆ (C1 )X , gcd(b′ , ℓ′) divides gcd(b, ℓ).
In order to study the rank and the kernel of a Z2 Z4 -additive code C, it is necessary to consider
the code Cb .
January 20, 2018
DRAFT
7
Proposition 7 ([4]). Let C be a Z2 Z4 -additive cyclic code with C = h(b | 0), (ℓ | f h + 2f )i.
Then Cb = h(b | 0), (µ̄ℓg̃ | 2f )i.
Note that if C = Cb , then (5) is satisfied and the code Φ(C) is linear. In this case, δ = 0 and,
by Theorem 3, g = 1. Therefore Cb = h(b | 0), (ℓ | 2f )i.
A. Kernel of Z2 Z4 -Additive Cyclic Codes
In this section, we will study the kernel of Z2 Z4 -additive cyclic codes. We will prove that, for
a Z2 Z4 -additive cyclic code C, the code K(C) is also cyclic and we will establish some properties
of its generator polynomials. We will show that there does not exist a Z2 Z4 -additive cyclic code
for all the possible values of the dimension of the kernel as for general Z2 Z4 -additive codes.
Let C be a Z2 Z4 -additive code. By (5), we can give the following definition of K(C) (see [8]):
K(C) = {v ∈ C | 2v ∗ w ∈ C, ∀w ∈ C}.
Lemma 8. Let C be a Z2 Z4 -additive code. Then, K(C)Y ⊆ K(CY ).
Proof. Let C be a Z2 Z4 -additive cyclic. Let v = (v | v ′ ) ∈ C. We have that v ∈ K(C) if and
only if 2v ∗ w ∈ C, ∀w = (w | w ′) ∈ C. Since 2v ∗ w = (0 | 2v ′ ∗ w ′ ), we have that if v ∈ K(C)
then v ′ ∈ K(CY ) and the statement follows.
Example 1. Let C be the Z2 Z4 -additive cyclic code in
Z2 [x]
hx−1i
[x]
× hxZ34−1i
generated by h(1 | x + 1)i,
where f = 1 and h = x − 1. Note that C is of type (1, 3; 1, 2; 1) and the generator matrix of C
in standard form, (1), is
1 2 0 0
0 3 1 0 .
0 3 0 1
Since f = 1, by Theorem 4, we know that K(CY ) = CY = h(x+1)i. We have that the generator
matrix of K(C) in standard form is
1 2 0 0
0 2 2 0
0 2 0 2
and therefore K(C) = h(1 | 2)i. Hence, K(C)Y
January 20, 2018
K(CY ).
DRAFT
8
The following theorems determine an upper and a lower bound for the kernel of a Z2 Z4 additive code and that there exists a Z2 Z4 -additive code of type (α, β, γ, δ, κ) for all possible
values of the kernel.
Theorem 9 ([8]). Let C be a Z2 Z4 -additive code with parameters (α, β, γ, δ, κ). Then γ + δ ≤
dim(ker(Φ(C)) ≤ γ + 2δ.
Theorem 10 ([8]). Let α, β, γ, δ, κ be integers satisfying
α, β, γ, δ, κ ≥ 0,
0<δ+γ ≤β+κ
and
α + β > 0,
κ ≤ min(α, γ).
(6)
Then, there exists a Z2 Z4 -linear code C of type (α, β; γ, δ; κ) with dim(ker(C)) = γ + 2δ − k̄
if and only if
k̄ = 0,
k̄ ∈ {0} ∪ {2, . . . , δ} and k̄ even,
k̄ ∈ {0} ∪ {2, . . . , δ},
if s = 0,
if s = 1,
if s ≥ 2,
where s = β − (γ − κ) − δ.
We will see that not all possible values for the kernel of a Z2 Z4 -additive code C are possible
if C is cyclic. First, we will determine some properties of the kernel of a Z2 Z4 -additive cyclic
code.
Proposition 11. Let C = h(b | 0), (ℓ | f h + 2f )i be a Z2 Z4 -additive cyclic code. Then
α − deg(b) + deg(h) + deg(g) ≤ |K(C)| ≤ α − deg(b) + deg(h) + 2 deg(g).
Proof. Straightforward from Theorems 9 and 3.
Note that the upper bound is sharp when the code has binary linear image, i.e., C = K(C).
Moreover, the lower bound is tight when K(C) only has the all-zero vector and all order two
codewords; that is, C = Cb .
Proposition 12. Let C be a Z2 Z4 -additive cyclic code, then K(C) ⊆ CX × K(CY ).
Proof. Let v = (v, v ′) ∈ K(C). For all w = (w, w ′) ∈ C, 2v ∗ w ∈ C since v is in the kernel.
Then v ∈ CX and 2v ′ ∗ w ′ ∈ CY , for all w ′ ∈ CY which gives v ′ ∈ K(CY ) and v ∈ CX × K(CY )
and the result follows.
January 20, 2018
DRAFT
9
Proposition 13. If C is a separable Z2 Z4 -additive code then K(C) = CX × K(CY ).
Proof. Let v ∈ CX × K(CY ), then 2v ∗ w = (0 | 2v ′ ∗ w ′) for all w ∈ C. Since v ′ ∈ K(CY ),
then 2v ′ ∗ w ′ ∈ CY . Moreover, since C is a separable Z2 Z4 -additive cyclic code, 2v ∗ w = (0 |
2v ′ ∗ w ′ ) ∈ C. Therefore v ∈ K(C) and CX × K(CY ) ⊆ K(C).
Now, by Proposition 12, we have K(C) ⊆ CX × K(CY ). w ∈ C. Since C is a separable
Z2 Z4 -additive cyclic code, 2v ′ ∗ w ′ ∈ CY for all w ′ ∈ CY and so v ′ ∈ K(CY ). Therefore
The following example shows that if the Z2 Z4 -additive code C is non-separable, then the
kernel is not necessarily CX × K(CY ).
Example 2. Let C = h(1 | x + 1)i, the Z2 Z4 -additive code of Example 1, with CX = h1i and
CY = hx + 1i. We have seen that K(CX ) = CX , K(CY ) = CY and K(C) = h(1 | 2)i. Therefore
K(C) ( CX × K(CY ).
Therefore, if the code is non-separable, the equality is not satisfied in general.
From Proposition 12, we obtain that dim(ker(Φ(C))) ≤ κ + dim(ker(φ(CY ))). However, we
can give a bound that is more accurate.
Proposition 14. Let C be a Z2 Z4 -additive code, then dim(ker(Φ(C))) ≤ κ1 + dim(ker(φ(CY ))).
Proof. Define C0 = {v = (v | v ′ ) ∈ C | v ′ = 0}. We have that C0 ⊆ K(C), and dim(ker(C0 )) =
κ1 .
Let v = (v | v ′ ) ∈ K(C). If v′ = 0, then v ∈ C0 . Otherwise, v ′ ∈ K(C)Y ⊆ K(CY ) by
Lemma 8 and, therefore, dim(ker(Φ(C))) ≤ κ1 + dim(ker(φ(CY ))).
From the generator matrix G of C given in (4), we have that the code CY has a generator
matrix of the form
2T2
0
0
2T1 2Iγ−(κ +κ ) 0 .
1
2
S
R
Iδ
(7)
By [3, Proposition 1], we know that the code CY has type 4δ 2γ−κ1 . The minimum value for
the dimension of ker(φ(CY )) is δ + γ − κ1 .
Theorem 15. Let C be a Z2 Z4 -additive code. If K(CY ) is a minimum then K(C) is a minimum.
January 20, 2018
DRAFT
10
Proof. If K(CY ) is a minimum, by Proposition 14, then dim(ker(Φ(C))) ≤ κ1 + δ + γ − κ1 =
γ + δ.
In the previous statements, we compute an upper bound of the kernel of a Z2 Z4 -additive code
C by considering the kernel of a code over Z4 . Now we shall give the exact value of the kernel
of a Z2 Z4 -additive code C in terms of CX and the kernel of a linear code over Z4 , C ′ . As we have
seen, in the case of a separable code C, the code C ′ is exactly CY and the value dim(ker(Φ(C)))
is dim(ker(Φ(CX ))) + dim(ker(Φ(CY ))), where CY is a cyclic code over Z4 . If the code C is not
separable, then C ′ is not necessarily cyclic.
Let C be a Z2 Z4 -additive code with generator matrix in the form of (4) and let C ′ be the
subcode generated by
0 0 0
0 0 S′
2T1 2Iγ−(κ1 +κ2 ) 0
S
R
Iδ
!
.
(8)
Theorem 16. Let C be a Z2 Z4 -additive with generator matrix in the form of (4) and let C ′ be
the subcode generated by the matrix in (8). Then dim(ker(Φ(C)) = κ1 + κ2 + dim(ker(φ(CY′ ))).
Proof. Let C be a Z2 Z4 -additive cyclic code with generator matrix G in the form of (4). Let
{ui = (ui | u′i )}γi=1 be the first γ rows and {vj = (vj | vj′ )}δj=1 the last δ rows of G. Define the
1 +κ2
codes C¯ = h{ui }κi=1
i, C ′ = h{ui }γi=κ1 +κ2 +1 , {vj }δj=1i.
By [8], v ∈ K(C) if and only if 2v ∗ w ∈ C for all w ∈ C. We have that v ∈ C̄ is of
¯ = κ1 + κ2 . Let
order 2 and hence 2v ∗ w = 0, ∀w ∈ C. Then, C¯ ⊆ K(C) and dim(ker(C))
¯ we have that v ∈ K(C) if and only
v = (v | v ′ ) ∈ C ′ . Since 2v ∗ w = 0 for all w ∈ C,
if 2v ∗ w ∈ C ′ for all w ∈ C ′ ; that is, v ∈ K(C ′ ). Finally, 2v ∗ w = (0 | 2v ′ ∗ w ′ ) ∈ C ′
if and only if 2v ′ ∗ w ′ ∈ (C ′ )Y , and hence dim(ker(Φ(C ′ ))) = dim(ker(Φ((C ′ )Y ))). Therefore,
dim(ker(Φ(C))) = κ1 + κ2 + dim(ker(φ(CY′ ))).
Now we will establish the kernel of a Z2 Z4 -additive cyclic codes taking into account its
generator polynomials. In the following theorem we shall prove that if C is Z2 Z4 -additive cyclic
code, then K(C) is also cyclic. This result is a generalization of the case when C is a linear
cyclic code over Z4 that is given in [6].
Theorem 17. Let C be a Z2 Z4 -additive cyclic code. Then K(C) is a Z2 Z4 -additive cyclic code.
Proof. We know that K(C) is a Z2 Z4 -additive code so we just have to show that if u = (u | u′ ) ∈
K(C) then π(u) ∈ K(C). That is, we want to show that 2π(u) ∗ w ∈ C, for all w ∈ C.
January 20, 2018
DRAFT
11
Let u ∈ K(C), w ∈ C. Then 2π(u) ∗ w = π(2u ∗ π −1 (w)). We have that u ∈ K(C) and
π −1 (w) ∈ C, therefore 2u ∗ π −1 (w) ∈ C by (5). Since the code C is cyclic, π(2u ∗ π −1 (w)) ∈ C,
which gives that 2π(u) ∗ w ∈ C, and π(u) ∈ K(C).
Corollary 18. Let C = h(b | 0), (ℓ | f h + 2f )i be a Z2 Z4 -additive cyclic code, where f hg =
xβ − 1. Then, K(C) = h(bk | 0), (ℓk | fk hk + 2fk )i, where fk hk gk = xβ − 1 and
1) f divides fk ;
2) gcd(b, ℓ) divides gcd(bk , ℓk ).
Proof. By Theorem 17, K(C) is cyclic and therefore K(C) = h(bk | 0), (ℓk | fk hk + 2fk )i, where
fk hk gk = xβ − 1. Since K(C) ⊆ C, the result follows from Proposition 6.
Let C = h(b | 0), (ℓ | f h + 2f )i be a Z2 Z4 -additive cyclic code, where f hg = xβ − 1. In [4]
it is proved that
C = h(b | 0), (ℓ′ | f h), (µ̃ℓg̃ | 2f )i,
(9)
where ℓ′ = ℓ − µ̃ℓg̃.
Lemma 19. Let C = h(b | 0), (ℓ | f h+2f )i be a Z2 Z4 -additive cyclic code, where f hg = xβ −1.
Let h(b | 0), (ℓk | f hk + 2f )i ⊂ C, for k|g. Then ℓk = k̃ℓ + (1 − k̃)µ̃ℓg̃ (mod b), for µ in (3).
Proof. Let C = h(b | 0), (ℓ′ | f h), (µ̃ℓg̃ | 2f )i, where ℓ′ = ℓ − µ̃ℓg̃, as in (9). Since (ℓk |
f hk + 2f ) ∈ C and (ℓk | f hk + 2f ) = c1 (b | 0) + c2 (ℓ′ | f h) + c3 (µ̃ℓg̃ | 2f ), we obtain c2 = k,
c3 = 1 and ℓk = k̃ℓ′ + µ̃ℓg̃ (mod b) = k̃ℓ + (1 − k̃)µ̃ℓg̃ (mod b).
Theorem 20. Let C = h(b | 0), (ℓ | f h+2f )i be a Z2 Z4 -additive cyclic code, where f hg = xβ −1.
Then, K(C) = h(b | 0), (ℓk | f hk + 2f )i, where k divides g and ℓk = k̃ℓ + (1 − k̃)µ̃ℓg̃ (mod b),
for µ in (3).
Proof. By Theorem 17, K(C) is cyclic and then K(C) = h(bk | 0), (ℓk | fk hk + 2fk )i. Clearly,
bk = b. Since Cb ⊆ K(C) ⊆ C, by Proposition 7 and Proposition 6, we conclude that fk = f .
Since K(C)Y = hf hk + 2f i ⊆ CY , with an argument analogous to that of [6, Theorem 9] we
obtain that hk = hk with k a divisor of g.
Let ℓ′ = ℓ − µ̃ℓg̃. By (9), (ℓ′ | f h), (µ̃ℓg̃ | 2f ) ∈ C. Therefore, ℓk = k̃ℓ′ + µ̃ℓg̃ (mod b) =
k̃ℓ + (1 − k̃)µ̃ℓg̃ (mod b).
January 20, 2018
DRAFT
12
Theorem 10 shows that there exists a Z2 Z4 -additive code for all possible values of the kernel
for a given type (α, β; γ, δ; κ). Considering the last theorem, the next example illustrates that this
result is not true for Z2 Z4 -additive cyclic codes; i.e., for a given type (α, β; γ, δ; κ) there does
not always exist a Z2 Z4 -additive cyclic code for all possible values of the kernel. Furthermore, it
shows that there does not always exist a Z2 Z4 -additive cyclic code for a given type (α, β; γ, δ; κ).
Example 3. By Theorem 10, there exists a Z2 Z4 -additive code C of type (2, 7; 2, 3; κ) with
dim(ker(Φ(C))) = kd , for all kd ∈ {5, 6, 8}. We will see that there does not exist any cyclic
Z2 Z4 -additive code of type (2, 7; 2, 3; κ), with dimension of the kernel in {6, 8}.
Let α = 2 and β = 7. We have that x7 − 1 = (x − 1)(x3 + 2x2 + x + 3)(x3 + 3x2 + 2x + 3).
Let C = h(b | 0), (ℓ | f h + 2f )i be a Z2 Z4 -additive cyclic code of type (2, 7; 2, 3; κ), where
f hg = x7 − 1.
By Theorem 3, deg(g) = 3 and deg(b) = deg(h) ≤ 2. Let {p3 , q3 } = {(x3 + 2x2 + x+ 3), (x3 +
3x2 +2x+3)}. Assume without loss of generality that g = p3 and, since deg(h) ≤ 2, we have that
˜
fb
, (g̃ ⊗ g̃)) 6= 1.
q3 divides f . It is easy to see that gcd(q3 , (p˜3 ⊗ p˜3 )) 6= 1 and therefore gcd( gcd(b,ℓg̃)
Hence, by Theorem 5, there does not exist a Z2 Z4 -additive code of type (2, 7; 2, 3; κ) with linear
Gray image. Thus, dim(ker(Φ(C))) 6= 8.
By Theorem 20, K(C) = h(b | 0), (ℓk | f hk+2f )i where k divides g. By the previous argument,
k 6= 1 and then we have that k = g = p3 and K(C) = h(b | 0), (ℓk | 2f )i. Therefore K(C) does
not contain codewords of order 4, thus dim(ker(Φ(C))) = γ + δ = 5.
Finally, we will give Z2 Z4 -additive cyclic codes of type (2, 7; 2, 3; κ), for different values of
κ. Recall that κ ≤ min{α, γ} = 2 and κ = α − deg(gcd(b, ℓg̃)), then
•
κ = 2: C = h(x − 1 | 0), (1 | (x3 + 2x2 + x + 3)(x − 1) + 2(x3 + 2x2 + x + 3))i, or
C = h(x − 1 | 0), (1 | (x3 + 3x2 + 2x + 3)(x − 1) + 2(x3 + 3x2 + 2x + 3))i.
•
κ = 1: C = h(x − 1 | 0), (0 | (x3 + 3x2 + 2x + 3)(x − 1) + 2(x3 + 3x2 + 2x + 3))i, or
C = h(x − 1 | 0), (0 | (x3 + 2x2 + x + 3)(x − 1) + 2(x3 + 2x2 + x + 3))i.
•
κ = 0: In this case, deg(gcd(b, ℓg̃)) = 2 and therefore, gcd(b, ℓg̃) = x2 − 1. Note that p̃3
and q̃3 are not divisors of x2 − 1 over Z2 , thus there does not exist ℓ with deg(ℓ) < 2 such
that ℓg̃ = x2 − 1. There does not exist a Z2 Z4 -additive cyclic code of type (2, 7; 2, 3; 0).
The statement in Theorem 20 is also true for any maximal Z2 Z4 -additive cyclic subcode of a
Z2 Z4 -additive cyclic code C whose Gray image is a linear subcode of Φ(C).
January 20, 2018
DRAFT
13
Corollary 21. Let C = h(b | 0), (ℓ | f h + 2f )i be a cyclic code, where f hg = xβ − 1. Then, if C1
is a maximal Z2 Z4 -additive cyclic subcode with Φ(C1 ) linear, then C1 = h(b | 0), (ℓk | f hk+2f )i,
where k divides g and ℓk = k̃ℓ + (1 − k̃)µ̃ℓg̃ (mod b), for µ in (3).
The kernel of a binary code is the intersection of all its maximal linear subspaces. Therefore,
if C1 , C2 , . . . Cs are all the maximal subcodes of a Z2 Z4 -additive code C such that φ(Ci ) is a
linear subcode of φ(C), for 1 ≤ i ≤ s, then
K(C) =
s
\
Ci .
(10)
i=1
In [6] it is proved that if C1 = hf h1 + 2f i and C2 = hf h2 + 2f i are quaternary cyclic codes of
odd length n, then C1 ∩C2 = hf lcm(h1 , h2 )+2f i. We will give a similar result for Z2 Z4 -additive
cyclic codes.
Proposition 22. Let C = h(b | 0), (l | f h + 2f )i be a Z2 Z4 -additive cyclic code. Let C1 = h(b |
0), (lk1 | f hk1 + 2f )i and C2 = h(b | 0), (lk2 | f hk2 + 2f )i be Z2 Z4 -additive maximal subcodes
of C whose images under the Gray map are linear subcodes of φ(C). Then
C1 ∩ C2 = h(b | 0), (lk′ | f hk ′ + 2f )i,
where k ′ = lcm(k1 , k2 ) and ℓk′ = k̃ ′ ℓ + (1 − k̃ ′ )µ̃ℓg̃ (mod b), for µ in (3).
Proof. Let C = h(b | 0), (l | f h + 2f )i be a Z2 Z4 -additive cyclic code and let Ci = h(b | 0), (lki |
f hki + 2f )i be a Z2 Z4 -additive maximal subcode of C whose image under the Gray map is a
linear subcode of φ(C).
We first consider (C1 )Y ∩(C2 )Y . By [6], (C1 )Y ∩(C2 )Y = hf lcm(hk1 , hk2 )+2f i = hf hk ′ +2f i,
where k ′ = lcm(k1 , k2 ). Since h(b|0)i ∈ C1 ∩ C2 , then C1 ∩ C2 = h(b | 0), (lk′ | f hk ′ + 2f )i, where
ℓk′ = k̃ ′ ℓ + (1 − k̃ ′ )µ̃ℓg̃ (mod b), for µ in (3) by Lemma 19.
Lemma 23. Let C be a Z2 Z4 -additive cyclic code and let D be a maximal cyclic subcode with
linear binary image. Then, K(C) ⊆ D.
Proof. If K(C) 6⊆ D, then consider the Z2 Z4 -additive code D ′ generated by K(C) ∪ D ∪ {2u ∗ v |
u, v ∈ K(C) ∪ D}. Since the binary image of K(C) ∪ D is cyclic, D ′ is a cyclic subcode of C.
Moreover, since 2u ∗ v ∈ D ′ , for all u, v ∈ D ′ , we have that D ′ has linear binary image, leading
to a contradiction since we are assuming that D is maximal.
January 20, 2018
DRAFT
14
Theorem 24. Let C = h(b | 0), (ℓ | f h+2f )i be a Z2 Z4 -additive cyclic code, where f hg = xβ −1.
Assume that k1 , . . . , ks are the divisors of g of minimum degree such that
!
g̃
g̃
f˜b
,
⊗
= 1,
gcd
gcd(b, ℓ kg̃˜ ) k̃i k̃i
i
for i = 1, . . . , s. Then,
K(C) = h(b | 0), (ℓ′k | f hk ′ + 2f )i,
where k ′ = lcm(k1 , . . . , ks ) and ℓk′ = k̃ ′ ℓ + (1 − k̃ ′ )µ̃ℓg̃ (mod b), for µ in (3).
Proof. Assume that k1 , . . . , ks are the divisors of g of minimum degree such that
!
g̃
f˜b
g̃
⊗
= 1,
gcd
,
gcd(b, ℓ kg̃˜ ) k̃i k̃i
i
for i = 1, . . . , s. Let Di be a cyclic subcode of C, where Di = h(b | 0), (ℓki | f hki + 2f )i, for
some ℓki . Note that Φ(Di ) is linear by Theorem 5. Since ki is a polynomial of minimum degree
dividing g, then Di is a maximal cyclic subcode of C with linear binary image. Then each Di
extends to Ci which is a maximal subcode of C, not necessarily cyclic, with linear binary image.
Note that every maximal code with linear image must contain a cyclic code with linear image,
e.g., every maximal code contains K(C) that is cyclic with linear image. By Lemma 23, we know
K(C) ⊆ Di and, therefore K(C) ⊆ ∩i Di . But ∩i Ci = K(C) ⊆ ∩i Di ⊆ ∩i Ci , so K(C) = ∩i Di . By
Corollary 21 and Proposition 22, the result follows.
Example 4. Let x7 − 1 = (x − 1)p3 q3 over Z4 . Let C = h(1 | 0), (0 | f )i of type (1, 7; 1, 6; 1)
with f = (x − 1), h = 1 and g = p3 q3 .
˜
fb
Note that gcd( gcd(b,ℓg̃)
, (g̃⊗g̃)) = x−1 6= 1. We have that all maximal cyclic subcodes of C with
binary linear image are C1 = h(1 | 0), (0 | f p3 + 2f )i, and C2 = h(1 | 0), (0 | f q3 + 2f )i. Clearly,
k ′ = lcm(p3 , q3 ) = p3 q3 and then K(C) = h(1 | 0), (0 | f p3 q3 + 2f )i = h(1 | 0), (0 | 2f )i.
B. Rank of Z2 Z4 -Additive Cyclic Codes
In this section, we will study the rank of a Z2 Z4 -additive cyclic code C. We will prove that
R(C) is also cyclic and we will establish some properties of its generator polynomials. However,
we will show that there does not exist a Z2 Z4 -additive cyclic code for all possible values of the
rank, in contrast to what is exhibited in the following results for a Z2 Z4 -additive code.
January 20, 2018
DRAFT
15
Proposition 25 ([8]). Let C be a Z2 Z4 -linear code of binary length n = α + 2β and type
(α, β; γ, δ; κ). Then, rank(Φ(C)) ∈ {γ + 2δ, . . . , min(β + δ + κ, γ + 2δ + δ2 )}.
Theorem 26 ([8]). Let α, β, γ, δ, κ be integers satisfying (6). Then, there exists a Z2 Z4 -linear
code C of type (α, β; γ, δ; κ) with rank(Φ(C)) = r if and only if
δ
)}.
r ∈ {γ + 2δ, . . . , min(β + δ + κ, γ + 2δ +
2
Proposition 27 ([8]). Let C be a Z2 Z4 -additive code of type (α, β; γ, δ; κ) and let C = Φ(C).
Let G be a generator matrix of C as in (1) and let {ui }γi=1 be the rows of order two and {vj }δj=1
the rows of order four in G. Then, hCi is generated by {Φ(ui )}γi=1 , {Φ(vj ), Φ(2vj )}δj=1 and
{Φ(2vj ∗ vk )}1≤j<k≤δ . Moreover, hCi is both binary linear and Z2 Z4 -linear.
Corollary 28 ([6]). Let C be a Z2 Z4 -additive code of type (α, β; γ, δ; κ) and let C = Φ(C). Let
G be a generator matrix of C as in (1) and let {ui }γi=1 be the rows of order two and {vj }δj=1
the rows of order four in G. Then,
R(C) = C ∪ h{2vj ∗ vk }1≤j<k≤δ i.
Lemma 29. Let C be a Z2 Z4 -additive code. Then, R(C)Y = R(CY ).
Proof. Let {ui = (ui | u′i )}γi=1 be the first γ rows and {vj = (vj | vj′ )}δj=1 the last δ rows
of G. By Corollary 28, R(C) = C ∪ h{2vj ∗ vk }1≤j<k≤δ i. By the same argument, R(CY ) =
CY ∪ h{2vj′ ∗ vk′ }1≤j<k≤δ i. Since 2vj ∗ vk = (0 | 2vj′ ∗ vk′ ) for all 1 ≤ j < k ≤ δ, the statement
follows.
The following theorem shows that if C is Z2 Z4 -additive cyclic code, then R(C) is also cyclic.
As in the case of K(C), this theorem is a generalization of the case when C is a linear cyclic
code over Z4 [6].
Theorem 30. Let C be a Z2 Z4 -additive cyclic code. Then R(C) is Z2 Z4 -additive cyclic code.
Proof. Let x ∈ R(C). By Corrollary 28, R(C) is generated by C and {2v∗w | v, w ∈ C}, then x =
u + 2v ∗ w, for some u, v, w ∈ C. As C is a Z2 Z4 -additive cyclic code then π(u), π(v), π(w) ∈ C
and 2π(v) ∗ π(w) ∈ R(C). Thus, π(x) = π(u) + 2π(v) ∗ π(w) ∈ R(C) and R(C) is Z2 Z4 -additive
cyclic code.
The next proposition is straightforward from Theorems 26 and 3.
January 20, 2018
DRAFT
16
Proposition 31. Let C be a Z2 Z4 -additive cyclic code. Then
α − deg(b) + deg(h) + 2 deg(g) ≤ rank(Φ(C)) ≤
deg(g)
).
min(α + β + deg(g) − deg(gcd(b, ℓg̃)), α − deg(b) + deg(h) + 2 deg(g) +
2
For a Z2 Z4 -additive cyclic code C, define C1 = h(b | 0)i and C2 = h(ℓ | f h + 2f )i. Since
C = C1 ∪ C2 and C1 ∩ C2 = {0}, we have that
rank(Φ(C)) = rank(Φ(C1 )) + rank(Φ(C2 )).
(11)
If the code C is separable, then ℓ = 0 and rank(Φ(C1 )) = rank(Φ(CX )). Moreover, C2 = h(0 |
f h + 2f )i, and therefore rank(Φ(C2 )) = rank(Φ(CY )). We obtain the following result.
Proposition 32. If C is a separable Z2 Z4 -additive cyclic code, then R(C) = R(CX ) × R(CY )
and rank(Φ(C)) = κ1 + rank(Φ(CY )).
Note that, if C is not separable, rank(Φ(C)) is not necessarily equal to κ1 + rank(Φ(CY )) as
it is shown in the following example.
Example 5. Consider the Z2 Z4 -additive
1
0
0
0
0
code generated by the following matrix.
0 0 0 0 0
1 0 0 0 0
0 1 2 0 0
0 0 1 1 0
0 0 1 0 1
(12)
In [4, Example 2] it was proved that φ(CY ) is binary linear whereas Φ(C) is not binary linear.
In this example we have that κ1 = 2. Since Φ(CY ) is linear, rank(Φ(CY )) = 5. Nevertheless,
Φ(C) is not binary linear and rank(Φ(C)) = 8 > 5 + 2.
In fact, rank(Φ(C)) is always greater or equal to κ1 + rank(Φ(CY )).
Proposition 33. Let C be a Z2 Z4 -additive cyclic code, then rank(Φ(C)) ≥ κ1 + rank(Φ(CY )).
Proof. Let C be a Z2 Z4 -additive cyclic code. By (11), rank(Φ(C)) = rank(Φ(C1 ))+rank(Φ(C2 )) =
κ1 + rank(Φ(C2 )) . By Proposition 27, in order to determine the rank, we have to consider the
set of vectors 2v ∗ w, for v = (v | v ′ ), w = (w | w ′ ) ∈ C. Since for all v ∈ C if w ∈ C1
we obtain 2v ∗ w = 0, we just have to consider v, w ∈ C2 . We have that if 2v ′ ∗ w ′ 6∈ CY
January 20, 2018
DRAFT
17
then 2v ∗ w 6∈ C. Therefore, rank(Φ(C2 )) ≥ rank(Φ((C2 )Y )) = rank(Φ(CY )) and, therefore,
rank(Φ(C1 )) ≥ κ1 + rank(Φ(CY )).
Now we can determine the rank of a Z2 Z4 -additive code C as the rank of CX and the rank of
a linear code over Z4 , C ′ . As in the case of the kernel, when C is separable we have seen that
C ′ = CY , but if C is not separable such a code C ′ may not be cyclic over Z4 .
Theorem 34. Let C be a Z2 Z4 -additive cyclic code with generator matrix in the form of (4) and
let C ′ be the subcode generated by the matrix in (8). Then,
rank(Φ(C)) = κ1 + κ2 + rank(Φ(CY′ )).
Proof. Let C be a Z2 Z4 -additive cyclic code with generator matrix G in the form of (4). Let
{ui = (ui | u′i )}γi=1 be the first γ rows and {vj = (vj | vj′ )}δj=1 the last δ rows of G. Define the
1 +κ2
codes C¯ = h{ui }κi=1
i and C ′ = h{ui }γi=κ1 +κ2 +1 , {vj }δj=1 i.
By Corollary 28 we have
R(C) = h{ui }γi=1 , {vj }δj=1, {2vj ∗ vk }1≤j<k≤δ i,
¯ = h{ui }κ1 +κ2 i,
R(C)
i=1
R(C ′ ) = h{ui }γi=κ1 +κ2 +1 , {vj , 2vj }δj=1, {2vj ∗ vk }1≤j<k≤δ i.
¯ ∪ R(C ′ ). Moreover, R(C)
¯ ∩ R(C ′ ) = {0} due to the fact that for
Note that R(C) = R(C)
¯ Therefore, rank(Φ(C)) = rank(Φ(C))
¯ +
all 1 ≤ j < k ≤ δ, 2vj ∗ vk = (0 | 2vj′ ∗ vk′ ) 6∈ C.
rank(Φ(C ′ )) = κ1 + κ2 + rank(Φ(C ′ )). Finally, by Lemma 29, rank(Φ(C ′ )) = rank(Φ(CY′ )) and
the statement follows.
Theorem 35. Let C = h(b | 0), (ℓ | f h+2f )i be a Z2 Z4 -additive cyclic code, where f hg = xβ −1.
Then, R(C) = h(br | 0), (ℓr | f h + 2 fr )i, where r is a divisor of f and br divides b.
Proof. By Theorem 30, R(C) is a Z2 Z4 -additive cyclic code, therefore R(C) = h(br | 0), (ℓr |
fr hr + 2fr )i. Since (b | 0) ∈ R(C), it is clear that br divides b. By [8, Lemma 3], C and R(C)
have the same number of order four codewords and since C ⊆ R(C) we have that fr hr = f h.
Then gr = g. By Proposition 6, we know that fr divides f and hence there exists r ∈ Z4 [x]
such that fr =
January 20, 2018
f
r
and hr = hr. Therefore, R(C) = h(br | 0), (ℓr | f h + 2 fr )i.
DRAFT
18
Let Ch(b | 0), (ℓ | f h + 2f )i be a Z2 Z4 -additive cyclic code. In the following example, we
will see that R(C) = h(br | 0), (ℓr | fr hr + 2fr )i, where br 6= b, if we consider the generators of
R(C) in standard form.
Example 6. Let x7 − 1 = (x − 1)p3 q3 over Z4 . Let C = h((x − 1) | 0), (1 | (x − 1) + 2)i, with
f = 1, h = x − 1 and g = p3 q3 . If we compute R(C) we obtain that R(C) = h(1 | 0), (0 |
(x − 1) + 2)i.
Z2 Z4 -additive
As it is shown in Theorem 26, there exists a Z2 Z4 -additive code for any possible value of
the rank. Nevertheless, the following example gives a particular type (α, β; γ, δ; κ) such that it
is not possible to construct a Z2 Z4 -additive cyclic code with a specific, and valid, value of the
rank.
Example 7. Let C = h(b | 0), (ℓ | f h + 2f )i be a Z2 Z4 -additive cyclic code of type (2, 7; 2, 3; κ).
By Theorem 26, rank(Φ(C)) ∈ {8, 9, 10, 11}. We will see that there does not exist any cyclic
Z2 Z4 -additive code of type (2, 7; 2, 3; κ), C, with rank(Φ(C)) ∈ {8, 9, 10}.
Let x7 − 1 = (x − 1)p3 q3 over Z4 , with p3 and q3 as in Example 3. By Theorem 3, deg(g) = 3
and deg(b) = deg(h) ≤ 2. Assume without loss of generality that g = p3 and, since deg(h) ≤ 2,
we have that q3 divides f . We have already proved, in Example 3, that there does not exist a
Z2 Z4 -additive code of type (2, 7; 2, 3; κ) with linear Gray image. Thus, rank(Φ(C)) 6= 8.
By Theorem 35, R(C) = h(b | 0), (ℓr | fr hr + 2fr )i where r divides f and hr = hr. Since
rank(Φ(C)) ∈ {8, 9, 10, 11}, we have that deg(r) ≤ 3 as |R(C)| = 43 22+deg(r) ≤ 211 . Since
gcd(q̃3 , p̃3 ⊗ p̃3 ) 6= 1 we have that q3 must divide r. Therefore deg(r) ≥ 3, and by the previous
argument, we know that r = q3 . So, rank(Φ(C)) ∈
/ {9, 10}.
Finally, we will give Z2 Z4 -additive cyclic codes of type (2, 7; 2, 3; κ), for different values of
κ, such that rank(Φ(C)) = 11. Recall that κ ≤ min{α, γ} = 2 and κ = α − deg(gcd(b, ℓg̃)),
then
•
κ = 2: C = h(x − 1 | 0), (1 | (x3 + 2x2 + x + 3)(x − 1) + 2(x3 + 2x2 + x + 3))i, or
C = h(x − 1 | 0), (1 | (x3 + 3x2 + 2x + 3)(x − 1) + 2(x3 + 3x2 + 2x + 3))i.
•
κ = 1: C = h(x − 1 | 0), (0 | (x3 + 3x2 + 2x + 3)(x − 1) + 2(x3 + 3x2 + 2x + 3))i, or
C = h(x − 1 | 0), (0 | (x3 + 2x2 + x + 3)(x − 1) + 2(x3 + 2x2 + x + 3))i.
•
κ = 0: As in Example 3, there does not exist a Z2 Z4 -additive cyclic code of type (2, 7; 2, 3; 0).
January 20, 2018
DRAFT
19
Proposition 36. Let C = hf h + 2f i be a cyclic code over Z4 of length n, with f hg = xn − 1.
Then,
R(C) = f h + 2
f
,
r
where r is the Hensel lift of gcd(f˜, g̃ ⊗ g̃).
Proof. From [6], we have that R(C) = hf h + 2 fr i, where r divides f .
Since R(C) is the minimum cyclic code over Z4 containing C whose image under the Gray
map is linear, then r is the polynomial of minimum degree dividing f satisfying that hf h + 2 fr i
˜
has linear image. This is equivalent, by [11], to the condition gcd fr̃ , g̃ ⊗ g̃ = 1. Therefore,
the polynomial r of minimum degree dividing f satisfying this condition is the Hensel lift of
gcd(f˜, g̃ ⊗ g̃).
Proposition 37. Let C = h(b | 0), (ℓ | f h + 2f )i be a Z2 Z4 -additive cyclic code, where f hg =
xβ − 1, such that Φ(C) is not linear and Φ(CY ) is linear. Then,
R(C) = h(br | 0), (ℓr | f h + 2f )i,
where br = gcd(b, µ̃g̃ℓ), µ is as in (3), and ℓr = ℓ − µ̃g̃ℓ (mod br ).
Proof. Let C = h(b | 0), (ℓ | f h + 2f )i be a Z2 Z4 -additive cyclic code, CY = hf h + 2f i. Let G
be a generator matrix of C as in (1) and let {ui = (ui | u′i )}γi=1 be the rows of order two and
{vj = (vj | vj′ )}δj=1 the rows of order four in G. By Corollary 28, R(C) = C∪h{2vj ∗vk }1≤j<k≤δ i.
We have that Φ(C) is not linear, therefore there exist i, k ∈ {1, . . . , δ} such that 2vj ∗ vk =
(0 | 2vj′ ∗ vk′ ) 6∈ C. Since Φ(CY ) is linear, 2vj′ ∗ vk′ ∈ h2f h, 2f i. Moreover, h(0 | 2f h)i ∈ C and
therefore R(C) = C ∪ h(0 | 2f )i. Considering the generators polynomials of R(C) and µ as in
(3), we have
R(C) = h(b | 0), (ℓ − µ̃g̃ℓ | f h), (µ̃g̃ℓ | 2f ), (0 | 2f )i
= h(b | 0), (µ̃g̃ℓ | 0), (ℓ − µ̃g̃ℓ | f h + 2f )i
= h(gcd(b, µ̃g̃ℓ) | 0), (ℓ − µ̃g̃ℓ | f h + 2f )i.
Therefore, considering the polynomial generators of R(C) in standard form, we have that
br = gcd(b, µ̃g̃ℓ) and ℓr = ℓ − µ̃g̃ℓ (mod br ).
January 20, 2018
DRAFT
20
Theorem 38. Let C = h(b | 0), (ℓ | f h+2f )i be a Z2 Z4 -additive cyclic code, where f hg = xβ −1.
Then,
f
R(C) = h(br | 0), (ℓr | f h + 2 )i,
r
where r is the Hensel lift of gcd(f˜, g̃ ⊗ g̃), br = gcd(b, µ̃g̃ℓ), µ is as in (3), and ℓr = ℓ − µ̃g̃ℓ.
Proof. Let C = h(b | 0), (ℓ | f h + 2f )i be a Z2 Z4 -additive cyclic code. By Theorem 35,
R(C) = h(br | 0), (ℓr | f h + 2 fr )i, where r is a divisor of f and br divides b.
Consider the quaternary code CY = hf h+2f i. By Lemma 29, we have that (R(C))Y = R(CY ).
Therefore, by Proposition 36, R(CY ) = f h + 2 f , where r is the Hensel lift of gcd(f˜, g̃ ⊗ g̃).
r
Note that R(CY ) = CY ∪
2 fr
.
From Corollary 28, C = C ∪ h{2vj ∗ vk }1≤j<k≤δ i, where {ui = (ui | u′i )}γi=1 are the rows
of order four of the generator matrix of C as in (1). Note that, for all uj , uk , 1 ≤ j ≤ k ≤ δ,
2uj ∗ uk = (0 | 2u′j ∗ u′k ), where 2u′j ∗ u′k ∈ R(CY ) = CY ∪ h2 fr and, therefore, (0 | 2u′j ∗ u′k ) ∈
C ∪ h(0 | 2 fr ) . Hence, we have that R(C) = C ∪ h(0 | 2 fr ) .
Therefore, for µ as in (3),
R(C) = h(b | 0), (ℓ − µ̃g̃ℓ | f h), (µ̃g̃ℓ | 2f ), (0 | 2 fr )i
= h(b | 0), (µ̃g̃ℓ | 0), (ℓ − µ̃g̃ℓ | f h + 2 fr )i
= h(gcd(b, µ̃g̃ℓ) | 0), (ℓ − µ̃g̃ℓ | f h + 2 fr )i.
From the last equation, and considering the polynomial generators of R(C) in standard form,
we have that br = gcd(b, µ̃g̃ℓ) and ℓr = ℓ − µ̃g̃ℓ (mod br ).
Example 8. Let α = 3 and β = 7. Consider, as in Example 3, C = h(x − 1 | 0), (0 | x − 1)i
where f = x − 1 and h = 1. As we have seen, C does not have linear binary image. Then, by
Theorem 38, we have that R(C) = h(1 | 0), (0 | (x − 1) + 2)i where r = f = x − 1.
Example 9. Let α = 3 and β = 15. Consider C = h(x − 1 | 0), (1 | f h + 2f )i where
f = x4 + 2x2 + 3x + 1 and h = (x − 1)(x4 + x3 + x2 + x + 1). Then, by Theorem 38, we have
that R(C) = h(1 | 0), (0 | f h + 2 fr )i where r = f = x4 + 2x2 + 3x + 1.
IV. C ONCLUSIONS
Given a Z2 Z4 -additive cyclic code C, we have shown that the codes K(C) and R(C) are also
Z2 Z4 -additive cyclic. We have computed the generator polynomials of these codes in terms of
January 20, 2018
DRAFT
21
the generator polynomials of C. Using these results, we have concluded that the dimensions
of the binary images of K(C) and R(C), i.e. the dimension of the kernel and the rank of C,
cannot take all the possible values as for a general Z2 Z4 -additive code. In other words, if a
Z2 Z4 -additive code is cyclic, then the set of possible values for the rank and the dimension of
the kernel becomes more restrictive.
R EFERENCES
[1] T. Abualrub, I. Siap and N. Aydin, “Z2 Z4 -additive cyclic codes,” IEEE Transactions on Information Theory, vol. 60, pp.
1508-1514, 2014.
[2] J. Borges, C. Fernández-Córdoba, J. Pujol, J. Rifà and M. Villanueva, “Z2 Z4 -linear codes: generator matrices and duality,”
Designs, Codes and Cryptography, vol. 54, pp. 167-179, 2010.
[3] J. Borges, C. Fernández-Córdoba, R. Ten-Valls, “Z2 Z4 -additive cyclic codes, generator polynomials and dual codes,” IEEE
Transactions on Information Theory, vol. 62, No. 10, pp. 6348-6354, 2016.
[4] J. Borges, S.T. Dougherty, C. Fernández-Córdoba, R. Ten-Valls,
“Binary images of Z2 Z4 -additive cyclic codes,”
arXiv:1707.03214v1, 2017.
[5] P. Delsarte, “An algebraic approach to the association schemes of coding theory,” Philips Res. Rep. Suppl., vol. 10, 1973.
[6] S.T. Dougherty, C. Fernández-Córdoba. “Kernels and ranks of cyclic and negacyclic quaternary codes,” Designs, Codes
and Cryptography, vol. 81, no. 2, pp. 347-364, 2016.
[7] C. Fernández-Córdoba, J. Pujol and M. Villanueva, “On rank and kernel of Z4 -linear codes,” Lecture Notes in Computer
Science, no. 5228, pp. 46-55, 2008.
[8] C. Fernández-Córdoba, J. Pujol and M. Villanueva, “Z2 Z4 -linear codes: rank and kernel,” Designs, Codes and Cryptography, vol. 56, pp. 43-59, 2010.
[9] A.R. Hammons, P.V. Kumar, A.R. Calderbank, N.J.A. Sloane, P. Solé, “The Z4 -linearity of kerdock, preparata, goethals
and related codes,” IEEE Transactions on Information Theory, vol. 40, pp. 301-319, 1994.
[10] F.J. MacWilliams, N.J.A. Sloane. The Theory of Error-Correcting Codes. North-Holland Publishing Company, Amsterdam,
New York, Oxford, 1975.
[11] J. Wolfmann. “Binary Images of Cyclic Codes over Z4 ,” IEEE Transactions on Information Theory, vol. 47, pp. 1773-1779,
2001.
January 20, 2018
DRAFT
| 7cs.IT
|
L ONG
SHORT- TERM MEMORY AND L EARNING - TO LEARN IN NETWORKS OF SPIKING NEURONS
arXiv:1803.09574v1 [cs.NE] 26 Mar 2018
Guillaume Bellec*, Darjan Salaj*, Anand Subramoney*, Robert Legenstein & Wolfgang Maass
Institute for Theoretical Computer Science
Graz University of Technology
Austria
{bellec,salaj,subramoney,legenstein,maass}@igi.tugraz.at
* = first authors
A BSTRACT
Networks of spiking neurons (SNNs) are frequently studied as models for networks of neurons in the brain, but also as paradigm for novel energy efficient
computing hardware. In principle they are especially suitable for computations
in the temporal domain, such as speech processing, because their computations
are carried out via events in time and space. But so far they have been lacking
the capability to preserve information for longer time spans during a computation,
until it is updated or needed - like a register of a digital computer. This function is
provided to artificial neural networks through Long Short-Term Memory (LSTM)
units. We show here that SNNs attain similar capabilities if one includes adapting
neurons in the network. Adaptation denotes an increase of the firing threshold of
a neuron after preceding firing. A substantial fraction of neurons in the neocortex
of rodents and humans has been found to be adapting. It turns out that if adapting
neurons are integrated in a suitable manner into the architecture of SNNs, the performance of these enhanced SNNs, which we call LSNNs, for computation in the
temporal domain approaches that of artificial neural networks with LSTM-units.
In addition, the computing and learning capabilities of LSNNs can be substantially enhanced through learning-to-learn (L2L) methods from machine learning,
that have so far been applied primarily to LSTM networks and apparently never
to SSNs.
This preliminary report on arXiv will be replaced by a more detailed version in
about a month.
1
I NTRODUCTION
Deep learning in artificial neural networks (ANNs) usually employs feedforward networks. Feedforward networks tend to support computing and learning mechanisms that differ strongly from those
of recurrently connected networks of neurons, which one finds in the brain. Recurrent ANNs have
been applied very successfully for demanding tasks such as speech processing and video recognition,
which require integration of information over time. But virtually all of these applications involve recurrent ANNs that employ specialized modules for preserving information over time, such as Long
Short-Term memory (LSTM) units (Hochreiter & Schmidhuber, 1997). A simpler type of memory
module had been introduced for artificial neural networks by Mikolov et al. (2014). Modules of this
type have been missing in spiking neural network (SNN) models, and as a result, their performance
for tasks that require integration over longer timespans has remained inferior. A first step for alleviating this problem was made through the inclusion of short-term synaptic plasticity in randomly
connected SNNs (Maass et al., 2002). This improved their capabilities for tasks that require integration of information over hundreds of ms, but did not solve the problem for tasks that involve the
behaviourally relevant time scale of seconds and longer.
Neurons in the brain are subject to intracellular processes on many different times scales, from ms
to s, minutes, hours, and days. We show that by including just one of their longer lasting processes,
neuronal adaptation, and by simultaneously considering structured rather than randomly connected
1
network architectures, one arrives at recurrent SNNs with substantially enhanced computing and
learning capabilities. We will refer to these SNNs with longer short-term memory as LSNNs.
2
T HE LSNN
MODEL
We use simple leaky integrate-and-fire neuron models (see Supplementary Information). One population R of regularly firing neurons in the LSNN consists of standard neurons of this type, whereas
a simple adaptation mechanism is included for neurons in a second subpopulation A. Both populations receive spike trains from a population X of external input neurons. Results of computations in
the LSNN are read out by a population Y of external linear readout neurons, see Fig. 1A.
In neuroscience the term neural adaptation refers to a broad range of phenomena and mechanisms. It
was already discovered in the 19th century that the intensity of a constant stimulus tends to cause a
diminishing sensation, see e.g. the Wikipedia article “Neural adaptation”. A closer look shows that
these phenomena take place simultaneously on several time scales, from hundreds of ms to s, hours,
or even days. In the context of neuron models, adaptation refers to changes in the firing threshold or
subthreshold currents of a neuron that reduce its excitability in response to its own firing (Gerstner
et al., 2014). Experimental data on that are documented for example in the form of a distribution of
measured adaptation indices of neurons for large sets of neurons, both from mouse and humans, in
the online Allen Brain Atlas (Allen Institute, 2018). This adaptation index measures the increase of
interspike intervals of a neuron that receives a constant input current. Common ways for modeling
such data are described in (Gerstner et al., 2014; Pozzorini et al., 2015; Gouwens et al., 2018; Teeter
et al., 2018). We are using here the arguably simplest form of integrating adaptation into a neuron
model: We assume that the firing threshold Bj (t) of neuron j increases by some fixed amount β for
each spike of this neuron j, and then decays exponentially back to a baseline value b0j with a time
constant τa,j . The neural dynamics for a discrete time step of δt = 1 reads as follows
Bj (t)
bj (t + 1)
= b0j + βbj (t),
=
ρj bj (t) + (1 − ρj )zj (t),
(1)
(2)
where ρj = exp(− τδt
) and zj (t) is the spike train of neuron j. Note that this dynamics of thresha,j
olds of adaptive spiking neurons is similar to the dynamics of the state of context neuron in (Mikolov
et al., 2014).
We optimize the weights, and in some cases also the connections of an LSNN through standard
optimization algorithms for specific ranges of tasks. This optimization algorithm is not claimed to be
biologically realistic. But it might correspond functionally to evolutionary processes that optimized
LSNNs in the brain for tasks that were relevant for survival. For the three experiments that we
report here we used backprop through time (BPTT). BP has already been applied by (Courbariaux
et al., 2016) and (Esser et al., 2016) to feedforward networks of spiking neurons, and by (Huh &
Sejnowski, 2017) to recurrent networks of spiking neurons. The gradient is backpropagated through
spikes by replacing in these approaches the non-existent derivative of the membrane potential at
the time of a spike by a pseudo-derivative that smoothly increases from 0 to 1, and then decays
back to 0. For our applications we needed to modify the pseudo-derivative in order to achieve
stable performance of BPTT for larger time spans, that required backpropagation through several
1000 layers of an unrolled feedforward network of spiking neurons: we reduced (“dampened”) the
amplitude of the pseudo-derivative by a factor < 1.
3
L ONG SHORT- TERM MEMORY IN LSNN S
Store-Recall task:
We first test the performance of LSNNs on a very simple working memory task (see Fig. 1D): The
network has to store a bit – encoded by spikes from input neurons – in working memory if it receives
simultaneously a “store” instruction, which is given in the form of spikes from other input neurons.
The stored bit has to be recalled by the network after a subsequent “recall” instruction, that is also
given in the form of spikes from dedicated input neurons. At the same time other spike inputs should
be ignored by the network.
2
Figure 1: Store-recall task. A An LSNN consists of a regular spiking and an adaptive neural
population. B A standard LSNN architecture that corresponds to the ANN architecture of Mikolov
et al. (2014) C A sample LSNN architecture that resulted from using Deep R, a biologically inspired
technique that prunes and rewires gradually the connectivity matrices during training (Bellec et al.,
2018). D The raster plot shows the activity of a network after convergence. From top to bottom,
the sub-panels show spike rasters from input neurons (X), regular spiking neurons (R), and adaptive
neurons (A) followed by the output of the readout and the dynamics of the adaptive thresholds of
adaptive neurons (this network is based on the standard architecture).
3
More precisely, the input neurons marked in blue and red at the top of Fig. 1D are Poisson neurons
that produce random spike sequences. An input bit to the network is encoded by spiking activity at
50 Hz in the corresponding input channels X0 or X1 for a pattern duration D = 200 ms. “Store”
and “recall” instructions are analogously encoded through firing of populations XS and XR of other
Poisson input neurons at 50 Hz for D = 200 ms. Otherwise input neurons are assumed to be silent
in this simple experiment.
When a “recall” instruction is received, the network should output the bit value received concurrently
with the most recent “store” instruction before that. After a “store”, a “recall” instruction is given
during a subsequent time period of length D with probability π. This results in an expected delay
of D
π between a “store” and “recall” instruction. The next “store” appears similarly in a subsequent
period of length D with probability π.
During a recall cue, the binary output is supposed to be encoded by the sign of the time-averaged
membrane potential of a linear readout neuron Y . The LSNN consisted of 60 regular spiking and
20 adaptive neurons. An example trial after training is shown in Fig. 1D.
A summary of results on performance of trained networks for several versions of the store-recall
tasks is given in Tables 1 and 2 of the Supplement. A SNN without adapting neurons can solve the
store-recall task only if working memory for a few 100 ms suffices. LSNNs work best if the time
constants of their adapting neurons are in the same range as the required durations of working memory in the network. If the adapting neurons of an LSNN have a spread of different time constants,
such LSNN generally performs well for a variety of tasks that require working memory durations
with a similar spread of memory durations. Table 2 shows that the architecture of an LSNN also
influences its performance for these tasks. In general, the standard architecture (Fig. 1 B) works
quite well. But first tests where the network architecture was optimized together with its parameters
suggest that more specific architectures, such as the one shown in Fig. 1 C may perform substantially better for specific tasks (see lowest row of Table 2). We applied there the rewiring method
Deep R (Bellec et al., 2018). Deep R converges theoretically to an optimal network configuration by
continuously updating the set of active connections in a biologically inspired manner (Kappel et al.,
2015; 2017; Bellec et al., 2018).
Sequential MNIST: We next tested the performance of LSNNs on a task that requires continuous
updates of working memory over a longer time span: sequential MNIST. The task is to classify the
handwritten digits of the MNIST datasets when the pixels are presented one after another.
We compare the performance of LSNNs with that of LSTM networks. For fair comparison, the size
of the LSNN is chosen to match the number of parameters of the LSTM network. This leads to
140 regular spiking and 100 adaptive neurons with adaptation time constant 700 ms in the LSNN,
in comparison to 128 LSTM units.
The pixels are presented sequentially in 784 time steps, after which the network is required to output
the class of the presented digit. For LSNNs, the gray level of each pixel is encoded into the spikes
of 80 input neurons. Several different ways of encoding pixel values by spikes can be used. Fig. 2C
shows a version where the gray level of each pixel is encoded by population coding through the firing
probability of the 80 input neurons, 1ms for each pixel. Somewhat better performance is achieved
when each of the 80 input neurons is associated with a particular threshold for the gray value, and
this input neuron fires whenever the gray value of the currently presented pixel crosses its threshold
(two input neurons were used per threshold, one spiked at threshold crossings from below, and one
at the crossings from above; this input convention is chosen for the LSNN results of Fig. 2B). In
either case, an additional input neuron becomes active when the presentation of the 784 pixel values
is finished, in order to prompt an output from the network. The firing of this additional input neuron
is shown at the top right of the top panel of Fig. 2C. The softmax of 10 linear output neurons Y
is trained to produce, during that time segment, the label of the sequentially presented handwritten
digit (see yellow shading around 800 ms of the output neuron for label 6 in the plot of the dynamics
of the output neurons Y in Fig. 2C; this output was correct). In contrast to the store-recall task, we
see that the firing thresholds of adapting neurons undergo complex temporal modifications.
The output of the network is determined by averaging the readout output over the 56 ms following
the presentation of the digit. The network is trained by minimizing the cross entropy between the
softmax of the averaged readout and the label distributions.
4
Figure 2: Sequential MNIST. A The goal of this task is to classify images of handwritten digits
when the pixels are shown sequentially row by row. B The performance of spiking neural networks
is compared for three different setups: without adapting neurons (LIF), an LSNN in the standard
architecture, but with connections from the regular to the adaptive spiking pool (LSNN), and an
LSNN with randomly initialized connectivity that was rewired during training (LSNN rewired). For
comparison, the performance of an artificial fully connected RNN and an LSTM with standard gray
level inputs are shown. C Dynamics of a network classifying a digit as in Figure 1.
5
A performance comparison is given in Figure 2B. LSNNs achieve a 93% classification accuracy on
the test set using the standard architecture with added connections from the regular to the adaptive
spiking pool. A higher performance of 96% is reached if the network is rewired during training
with DEEP R (Bellec et al., 2018). The same performance is achieved by an LSTM receiving
continuous gray levels instead of spikes. The first and fourth bars in Fig. 2B show that this range
of accuracy is out of reach for spiking or nonspiking neural networks without enhanced working
memory capabilities.
Each image presentation with this setup takes 784 ms. To test if the model can also integrate information over a longer time span we expanded this duration to 3.36 s by presenting each pixel for 4
milliseconds instead of 1. An LSNN trained on this expanded presentation duration achieved in a
first experiment 83% classification accuracy (all adaptive neurons had an adaptation time constant
of 3500 ms).
4
L EARNING - TO -L EARN FOR LSNN S
Some of the most intriguing applications of LSTM networks were demonstrations of Learning-toLearn (L2L) or Meta-learning capabilities of this type of neural networks. This research direction
was initiated by (Hochreiter et al., 2001), and has lead in recent years to a host of remarkable
new methods and results, see e.g. (Wang et al., 2016; Gupta et al., 2018). The reason why LSTM
networks have been used for most work on L2L is that they can naturally accommodate two levels of
learning and representation of learned insight: network parameters such as synaptic weights can be
used to encode “innate” or transfer knowledge from preceding learning of more-or-less related tasks.
At the same time, a sufficiently long short-term memory in the neural network can accumulate and
apply knowledge gained during a single learning iteration (generally consisting of many episodes)
for a specific learning task. Of course, many other schemes for encoding, fusing, and applying
results of long-term meta-learning and results from the current learning task are possible and should
be explored.
The results of (Hochreiter et al., 2001) show that besides backprop there are likely to be many other
methods which recurrent neural networks can use to learn nonlinear functions from a teacher. This
supervised learning scenario is also of interest from the perspective of understanding learning in
the brain, since learning to predict future sensory inputs, the consequences of an action, or sensory
input from one modality in terms of sensory inputs from other modalities, are all learning scenarios
where the target output values are given to networks of neurons in the brain – with often some delay
– from the environment or internal processing modules. Another aspect of this L2L approach is
that it produces a host of new ideas and algorithms for implementing learning in recurrent neural
networks. This is of interest from the perspective of neuroscience, because on the basis of the
currently available experimental data it is not clear to what extent synaptic weights are able to
support fast learning by storing knowledge from previous trials for the current learning task. It is
also not clear whether spike-timing-dependent plasticity (STDP) is the primary local learning rule in
neural networks of the brain, and which other local learning rules are there implemented (Lisman &
Spruston, 2010; Titley et al., 2017). In addition, experimental data for STDP question its usefulness
for fast and reliable learning, since STDP is subject to high synapse-to-synapse variability, and
requires dozens of repetitions of pairing of pre- and postsynaptic firing (Froemke et al., 2010).
We show here that the L2L paradigm of Hochreiter et al. (2001) for LSTM networks can be ported to
the biologically more realistic LSNNs. L2L results for SNNs have the advantage that many features
and fingerprints of resulting new learning algorithms for SNNs, such as neural activity (spike trains),
changes in the excitability of neurons and synaptic weights, and changes in the network dynamics
through learning can be compared with corresponding measurements from networks of neurons
in the brain. At the same time, the resulting new methods for learning nonlinear functions are
also of interest for spike-based neuromorphic chips such as Brainscales (Schemmel et al., 2010),
SpiNNaker (Furber et al., 2013), True North (Esser et al., 2016), chips from ETH Zürich such as
(Qiao et al., 2015), and Loihi (Davies et al., 2018). Nonlocal learning rules such as backprop are
real challenges for these neuromorphic devices, and alternative local rules for learning nonlinear
functions are therefore of particular interest.
The standard setup of L2L involves a large, possibly infinitely large, family F of learning tasks
C. Some parameters P of a neural network N , in our case an LSNN, are optimized in an outer
6
optimization loop to support fast learning of a randomly drawn task C from F. Like in (Hochreiter
et al., 2001) we let all synaptic weights W of N belong to the set of parameters P that are optimized
through the outer loop, i.e., through meta-learning. Hence the network is forced to encode all results
from learning the current task C in its internal state, in particular in its short-term memory. Thus the
synaptic weights W of the neural network N encode an efficient algorithm for learning arbitrary
tasks C from F, rather than results of learning a particular task C. The teacher input for learning a
particular function C from F is given in (Hochreiter et al., 2001) in a delayed manner: The target
output value of C is given after N has provided its guessed output value for the preceding input
values. This avoids that N learns to cheat by simply using the teacher input as its guessed output
value. We are using the same convention in our experiment.
The learning methods that can possibly be encoded by the synaptic weights W range over a substantially larger range of local update rules than commonly considered for network learning. Note
for example that most commonly considered local update rules are formulated as functions where
a parameter update (e.g., a weight update) depends in a particular but smooth manner on previous
parameter values and specific aspects of the recent network activity, such as spike times or rates of
particular neurons. As long as this function is smooth, it can be approximated by a neural network.
But importantly, ANY smooth update function – even a function that was previously never considered for neural network learning can be approximated by a neural network. The L2L approach of
(Hochreiter et al., 2001) allows in principle in the outer optimization loop a search over the best
performing ones among all these potential local update rules. They are only restricted to those that
can be implemented by the neural network N . In particular, they are constrained to local learning
rules if the neural network N only supports local rules for computing and changing the content of
short-term memories, as is the case for the LSNNs N that we are considering here. For example,
the network N is free to use the delayed teacher input for computing the error that it made on the
preceding step, and to transmit this error through its internal spiking activity to other neurons in N ,
in particular also to adapting neurons.
One of the more challenging family F of tasks considered by Hochreiter et al. (2001) for LSTM
networks was the class of all quadratic functions over two real-valued variables. We found that
L2L for this family F works also very well for LSNNs. We present here instead a potentially even
more interesting variation of this L2L task, where F consists of all functions of two real-valued
variables that can be realized by a minimal class of artificial neural networks: small two-layer target
neural networks TN, i.e., feedforward networks with one hidden layer, consisting of sigmoidal (nonspiking) neurons. We found that LSNNs can learn to learn quickly to reproduce the input-output
behavior of any such ANN TN, by just using their network dynamics and their internal states for
learning and emulating this ANN. Note that BPTT was only used for the outer learning loop, in
order to determine the synaptic weights W of the LSNN. But these could not be optimized for a
particular target network TN. Constraints of compute time did not yet allow us to emulate learning
for larger target networks TN, and we also did not try to optimize the architecture of the LSNN for
this task. Hence the results below provide just a proof of concept.
The target network TN represents a function C(x1 , x2 ), with 2 inputs x1 and x2 and 2 hidden
neurons, with a sigmoid activation at both the hidden and output layers (shown in Fig. 3A). The
outputs of this function, which serve as the targets for the output Y of the LSNN, are scaled to be
between [0, 1].
7
Figure 3: (Caption next page.)
8
Figure 3: Learning to Learn for SNNs. A Illustration of the two-layer feed-forward target network
(TN) used to generate the targets for the LSNN to learn. B The architecture of the LSNN used for
the learning to learn task. It received as inputs at each step x1 , x2 and the target from the previous
step C(x01 , x02 ). Its prediction Ĉ(x1 , x2 ) was output through a linear readout. C Performance during
training and testing. A new function (target network TN) was randomly chosen for each episode.
The LSNN is enabled by its weights W to learn approximations to the input-output behavior of
target networks TN that was significantly better than the best linear approximation to its inputoutput behavior. D The first 100 steps in a sample test episode where the network initially had high
error, but after few steps learned to predict the new targets successfully. E The correlation of the
output of the LSNN vs target outputs at the start, middle and end of training. F Sample spike raster
during testing, of the input neurons (every third neuron, top panel), and the regular (middle panel)
and adaptive neuron populations (bottom panel) for the same episode as in (D). The point where the
error goes to 0 is marked by a red cross corresponding to the one in (D).
The training procedure was as follows: Network training was divided into training episodes. At
the start of each training episode, a new set of input pairs x1 , x2 ∈ [−1, 1] and new weights
(between [−1, 1]) for the target network TN were randomly chosen and used to generate targets
C(x1 , x2 ) ∈ [0, 1]. These input pairs and targets were used as training data and presented to the
LSNN during the episode. The LSNN parameters were updated using BPTT to minimize the mean
squared error between the LSNN output and the target in the training set, using gradients computed
over batches of 10 such episodes. In other words, each weight update included gradients calculated
on the input/target pairs from 10 different TNs, which prevented the LSNN from specializing on
predicting the output of one single TN.
In each episode, for the given TN, 500 different input/target pairs were presented, one per step.
Each step lasted for 20 timesteps, which corresponds to 20 ms. The LSNN received at each step
three input values as input: x1 , x2 and C(x01 , x02 ), where x01 , x02 were the inputs for the preceding
step (i.e., C(x01 , x02 ) was the target of the previous step; it was set to 0 in the first step) during both
the training and testing process, as illustrated in Fig. 3B. While testing, the weights of the LSNN
remained fixed, and it was required to predict targets from functions Ctest ∈ F that it had never
seen before.
The input population consisted of three subpopulations of 100 neurons each, one population for each
of the inputs x1 , x2 and C(x01 , x02 ). These input populations produced Poisson spike trains based on
a Gaussian population-rate encoding of the input values, between 0 and 200 Hz. The range of each
input was split evenly among all the neurons responsible for that input, and for each given analog
input value, the firing rate of the neurons had a Gaussian activity profile with mean at that analog
value, and a variance of 0.1.
We used an LSNN architecture that contained, in addition to the connections of the standard architecture, connections from the regular-spiking to the adaptive population (see Fig. 3B), since performance of the standard architecture was much worse. The network contained 100 in each of the
regular spiking and adaptive populations. A sample spike raster is shown in Fig. 3F (which is for
the same episode as shown in Fig. 3D). The top panel shows the input values.
During training, there was an initial steep drop in error (see Fig. 3C) when the network learned to
output values in the same range as the target functions, see Fig. 3E left and middle. The performance
of the network was compared to a linear baseline: for each TN, an optimal linear regressor was
computed using the same input/target pairs presented to the LSNN. The mean squared error for the
predictions of this linear regressor on these input/target pairs was the linear baseline.
During testing, even though the network was shown target functions Ctest (x1 , x2 ) it had not seen
before, the test error remained low. The temporal evolution of test errors for the first 100 steps in a
sample test episode for a new target function C is shown in Fig. 3D. It can be seen that the network
had a high error in the first few steps of the episode, but it quickly learned to output the correct
targets, even though its weights W remained fixed. Thus the LSNN had learnt to learn in other
ways than through changes of its synaptic weights.
9
5
D ISCUSSION
We have shown that the inclusion of adapting neurons significantly enhances computing and learning
capabilities of recurrent networks of spiking neurons. In particular we have shown that the resulting
LSNN is able to integrate and maintain information over several seconds in a working memory.
In fact, it is reasonable to assume that any duration for working memory can be achieved with
corresponding large time constants of adapting neurons. Furthermore, our tests so far suggest that
LSNNs are able to approach the computational power and learning capability of the generally best
performing ANN model for temporal processing tasks: LSTM networks. BPTT tends to work well
for LSNNs because the error gradient can be propagated through the slowly changing firing threshold
of an adapting neuron without causing vanishing or exploding gradients. Hence adapting neurons
appear to provide “highways” for propagating errors backwards even over 100s and 1000s of layers
of an unrolled LSNN.
A remarkable feature of a realization of working memory in SNNs through increased firing thresholds is that this method is very energy-efficient: When a memory is stored through an increased firing
threshold, this tends to reduce firing activity in the SNN. In contrast, working memory was previously modelled in SNNs via persistent firing of a subset of neurons, i.e., through increased firing
activity. In neuroscience there is an ongoing debate to what extent working memory is implemented
in the brain through persistent firing or through activity-silent mechanisms (Stokes, 2015). Interestingly, experimental data show that also one form of long-term memory, familiarity, is expressed in
some brain regions (perirhinal and entorhinal cortex) through reduced activity during the presentation of a familiar stimulus (Eichenbaum et al., 2007). These results are consistent with a memory
model based on longer lasting increases of the firing threshold of specific subsets of neurons, see
also (Titley et al., 2017).
The implementation of short-term memory through neural adaptation is somewhat analogous to an
interesting variant of long short-term memory in artificial neural networks (Mikolov et al., 2014).
They introduced a particular population of neurons, termed context neurons, whose state changes on
a slower time scale than that of the other neurons. If one interprets the firing threshold of adapting
neurons as their “state”, then the dynamics of this state is analogous to the dynamics of context
neurons in their approach. However, in the non-spiking setting of (Mikolov et al., 2014) one cannot
see an advantage of this model with regard to energy-efficiency of a physical realization: the state
of context neurons has to be read out at every discrete time step, like for all other neurons. In other
words, a clocked network cannot make use of the possibility to convey information by not sending
a message, as one can do in SNNs. This option, to maintain and use information by not firing, is
extensively used by LSNNs.
In (Costa et al., 2017) an LSTM module was modelled through a recurrent network of non-spiking
neurons, i.e., of sigmoidal neurons whose analog outputs are interpreted as firing rates. Their model
is consistent with models for working memory based on persistent firing. It achieved remarkable
results for several datasets, including the same sequential MNIST data that we have considered for
the case of spike-based network inputs and spike-based network computations.
We have shown in Fig. 3 that the inclusion of adapting neurons in SNNs opens the door to porting
methods and results on L2L from LSTM networks to SNNs. Our result suggests new ways of storing
the “program” of a computation in SNNs: In the states of adapting neurons, i.e., in the working
memory of the LSNN. Similarly as previously shown for artificial neural networks in (Hochreiter
et al., 2001), this approach simultaneously opens the door to the exploration of new classes of
learning algorithms for SNNs that could previously not be considered. In particular, our result
shows that LSNNs can learn the input/output behaviour of feedforward artificial neural network
with a hidden layer by just using readily available local mechanisms. This shows that LSNNs are
able to emulate the function of backprop learning in quite different ways than considered so far.
We will show in (Bellec et al., 2018) that LSNNs also pave the way for porting new ideas for learning
to learn from rewards (Wang et al., 2016) to networks of spiking neurons. In particular, one can see
that LSNNs are able to learn nontrivial sequential behaviors, and that they can make use of control
transfer learning, as shown there for LSTM networks.
10
ACKNOWLEDGMENTS
Research leading to these results has in parts been carried out on the Human Brain Project PCP
Pilot Systems at the Juelich Supercomputing Centre, which received co-funding from the European
Union (Grant Agreement #604102). We gratefully acknowledge Sandra Diaz, Alexander Payser
and Wouter Klijn from the Simulation Laboratory Neuroscience of the Jülich Supercomputing Centre for their support. This research was carried out under partial support by the Human Brain Project
of the European Union #720270 and #785907.
R EFERENCES
Allen Institute. c 2018 Allen Institute for Brain Science. Allen Cell Types Database, cell feature search.
Available from: celltypes.brain-map.org/data. 2018.
Guillaume Bellec, Anand Subramoney, Robert Legenstein, and Wolfgang Maass. Networks of spiking neurons
learn to learn. In Preparation, 2018.
Rui Costa, Ioannis Alexandros Assael, Brendan Shillingford, Nando de Freitas, and Tim Vogels. Cortical
microcircuits as gated-recurrent neural networks. In Advances in Neural Information Processing Systems,
pp. 272–283, 2017.
Matthieu Courbariaux, Itay Hubara, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized neural networks: Training deep neural networks with weights and activations constrained to+ 1 or-1. arXiv preprint
arXiv:1602.02830, 2016.
Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with
on-chip learning. IEEE Micro, 38(1):82–99, 2018.
Howard Eichenbaum, Andrew P Yonelinas, and Charan Ranganath. The medial temporal lobe and recognition
memory. Annu. Rev. Neurosci., 30:123–152, 2007.
Steven K. Esser, Paul A. Merolla, John V. Arthur, Andrew S. Cassidy, Rathinakumar Appuswamy, Alexander
Andreopoulos, David J. Berg, Jeffrey L. McKinstry, Timothy Melano, Davis R. Barch, Carmelo di Nolfo,
Pallab Datta, Arnon Amir, Brian Taba, Myron D. Flickner, and Dharmendra S. Modha. Convolutional networks for fast, energy-efficient neuromorphic computing. Proceedings of the National Academy of Sciences,
113(41):11441–11446, November 2016. ISSN 0027-8424, 1091-6490. doi: 10.1073/pnas.1604850113.
Robert C Froemke, Dominique Debanne, and Guo-Qiang Bi. Temporal modulation of spike-timing-dependent
plasticity. Frontiers in synaptic neuroscience, 2:19, 2010.
Steve B Furber, David R Lester, Luis A Plana, Jim D Garside, Eustace Painkras, Steve Temple, and Andrew D
Brown. Overview of the spinnaker system architecture. IEEE Transactions on Computers, 62(12):2454–
2467, 2013.
Wulfram Gerstner, Werner M. Kistler, Richard Naud, and Liam Paninski. Neuronal dynamics: From single
neurons to networks and models of cognition. Cambridge University Press, 2014.
Nathan W Gouwens, Jim Berg, David Feng, Staci A Sorensen, Hongkui Zeng, Michael J Hawrylycz, Christof
Koch, and Anton Arkhipov. Systematic generation of biophysically detailed models for diverse cortical
neuron types. Nature communications, 9(1), 2018.
Abhishek Gupta, Russell Mendonca, YuXuan Liu, Pieter Abbeel, and Sergey Levine. Meta-reinforcement
learning of structured exploration strategies. arXiv preprint arXiv:1802.07245, 2018.
Geoffrey Hinton, Nitish Srivastava, and Swersky Kevin. Lecture 6d - a separate, adaptive learning rate for each
connection. Slides of Lecture Neural Networks for Machine Learning, 2012.
Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780,
1997.
Sepp Hochreiter, A Steven Younger, and Peter R Conwell. Learning to learn using gradient descent. In
International Conference on Artificial Neural Networks, pp. 87–94. Springer, 2001.
Dongsung Huh and Terrence J Sejnowski. Gradient descent for spiking neural networks. arXiv preprint
arXiv:1706.04698, 2017.
11
David Kappel, Stefan Habenschuss, Robert Legenstein, and Wolfgang Maass. Network Plasticity as Bayesian
Inference. PLOS Computational Biology, 11(11):e1004485, 2015.
David Kappel, Robert Legenstein, Stefan Habenschuss, Michael Hsieh, and Wolfgang Maass. Reward-based
stochastic self-configuration of neural circuits. arXiv preprint arXiv:1704.04238, 2017.
Diederik P Kingma and Jimmy Ba.
arXiv:1412.6980, 2014.
Adam: A method for stochastic optimization.
arXiv preprint
John Lisman and Nelson Spruston. Questions about STDP as a general model of synaptic plasticity. Frontiers
in synaptic neuroscience, 2:140, 2010.
Wolfgang Maass, Thomas Natschläger, and Henry Markram. Real-time computing without stable states: A new
framework for neural computation based on perturbations. Neural computation, 14(11):2531–2560, 2002.
Tomas Mikolov, Armand Joulin, Sumit Chopra, Michael Mathieu, and Marc’Aurelio Ranzato. Learning longer
memory in recurrent neural networks. arXiv preprint arXiv:1412.7753, 2014.
Christian Pozzorini, Skander Mensi, Olivier Hagens, Richard Naud, Christof Koch, and Wulfram Gerstner.
Automated high-throughput characterization of single neurons by means of simplified spiking models. PLoS
computational biology, 11(6):e1004275, 2015.
Ning Qiao, Hesham Mostafa, Federico Corradi, Marc Osswald, Fabio Stefanini, Dora Sumislawska, and Giacomo Indiveri. A reconfigurable on-line learning spiking neuromorphic processor comprising 256 neurons
and 128k synapses. Frontiers in neuroscience, 9:141, 2015.
Johannes Schemmel, Daniel Brüderle, Andreas Grübl, Matthias Hock, Karlheinz Meier, and Sebastian Millner.
A wafer-scale neuromorphic hardware system for large-scale neural modeling. In Circuits and systems
(ISCAS), proceedings of 2010 IEEE international symposium on, pp. 1947–1950. IEEE, 2010.
Mark G. Stokes. ‘Activity-silent’ working memory in prefrontal cortex: a dynamic coding framework. Trends
in Cognitive Sciences, 19(7):394–405, 2015.
Corinne Teeter, Ramakrishnan Iyer, Vilas Menon, Nathan Gouwens, David Feng, Jim Berg, Aaron Szafer,
Nicholas Cain, Hongkui Zeng, Michael Hawrylycz, et al. Generalized leaky integrate-and-fire models classify multiple neuron types. Nature communications, 1(1):1–15, 2018.
Heather K Titley, Nicolas Brunel, and Christian Hansel. Toward a neurocentric view of learning. Neuron, 95
(1):19–32, 2017.
Jane X Wang, Zeb Kurth-Nelson, Dhruva Tirumala, Hubert Soyer, Joel Z Leibo, Remi Munos, Charles
Blundell, Dharshan Kumaran, and Matt Botvinick. Learning to reinforcement learn. arXiv preprint
arXiv:1611.05763, 2016.
12
S UPPLEMENTARY I NFORMATION
Neuron model: In continuous time the spike trains xi (t) and zj (t) are formalized as sums of
Dirac pulses. Neurons are modeled according to a standard adaptive leaky integrate-and-fire model.
A neuron j spikes as soon at its membrane potential Vj (t) is above its threshold Bj (t). At each
spike time t, the membrane potential Vj (t) is reset by subtracting the current threshold value Bj (t).
Importantly at each spike the threshold Bj (t) of an adaptive neuron is increased by a constant β.
Then the threshold decays back to a baseline value b0j . Between spikes the membrane voltage Vj (t)
and the threshold Bj (t) are following the dynamics
τm V̇j (t)
τa,j Ḃj (t)
= −Vj (t) + Rm Ij (t)
=
b0j
− Bj (t),
(3)
(4)
where τm is the membrane time constant, τa,j is the adaptation time constant and Rm is the membrane resistance. The input current Ij (t) is defined as the weighted sum of spikes from external
inputs and other neurons in the network:
X
X
in
rec
Ij (t) =
Wji
xi (t − din
Wji
zi (t − drec
(5)
ji ) +
ji ),
i
i
in
rec
where Wji
and Wji
denote respectively the input and the recurrent synaptic weights and din
ji
rec
and dji the corresponding synaptic delays. All network neurons are connected to a population of
out
. When network neuron j spikes, the output synaptic strength
readout neurons with weights Wkj
out
Wkj is added to the membrane voltage yk (t) of all readout neurons k. yk (t) also follows the
dynamics of a leaky integrator τm ẏk (t) = −yk (t). The physical units of voltages, currents, delays,
Dirac pulses and synaptic weights are respectively Volt, Ampere, Second, Hertz and Coulomb.
Implementation in discrete time: Our simulations were performed in discrete time with a time
step δt = 1 ms. In discrete time, the spike trains are modeled as binary sequences xi (t), zj (t) ∈
{0, 1}. Neuron j emits a spike at time t if it is currently not in a refractory period, and its membrane
potential Vj (t) is above its threshold Bj (t). During the refractory period following a spike, zj (t) is
fixed to 0. The dynamics of the threshold is defined by Bj (t) = b0j + βbj (t) where β is a constant
which scales the deviation bj (t) from the baseline b0j . The neural dynamics in discrete time reads as
follows
Vj (t + 1) =
bj (t + 1) =
αVj (t) + (1 − α)Rm Ij (t) − Bj (t)zj (t)
ρj bj (t) + (1 − ρj )zj (t),
(6)
(7)
where α = exp(− τδtm ) and ρj = exp(− τδt
). The term Bj (t)zj (t) implements the reset of the
a,j
membrane voltage after each spike. The current Ij (t) is the weighted sum of the incoming spikes.
The definition of the input current in equation (5) holds also for discrete time, with the difference
that spike trains now assume values in {0, 1}.
Parameter values: For adaptive neurons, we used βj = 1.7, and for regular spiking neurons we
used βj = 0 (i.e. Bj is constant). The baseline threshold voltage was b0j = 0.01 and the membrane
time constant τm = 20 ms. Initial network weights were drawn from a Gaussian distribution Wji ∼
√w0 N (0, 1), where nin is the number of afferent neurons in the considered weight matrix (i.e.,
nin
the number of columns of the matrix), N (0, 1) is the zero-mean unit-variance Gaussian distribution,
and w0 is a weight scaling factor chosen to be w0 = R1Vm δt . For this choice, a spike transmitted
with synaptic weight w0 resulted in an increase of post-synaptic potential of 1 − α. In this way,
the standard initialization of synaptic weights initialized the networks in a stable regime producing
realistic firing rates.
In store-recall the test performance is computed over sequences of 32 consecutive store-recall pairs
and averaged over 32 random input sequences. During training with BPTT the depth of the unrolled
network is adapted to the distribution of store-recall delays. For expected delays of 2, 4, and 6 s
it resulted in 3200, 5200 and 8000 unrolled time steps respectively. All networks were trained for
500 iterations. We used the Adam optimizer (Kingma & Ba, 2014) with default parameters and a
13
learning rate of 0.01, with sequences presented in batches of sizes indicated in Table 1 and 2 (as
“batch size”). To avoid unrealistically high firing rates, the loss function contains a regularization
term that minimizes the distance of average firing rate between individual neurons and a target firing
rate of 10 Hz.
In sequential MNIST all networks were trained for 36, 000 iterations with approximately 66, 000
parameters, with a batch size of 256, using the default Adam optimizer, and a learning rate initialized
at 0.01 and decayed by a factor 0.8 every 2500 iterations. The baseline artificial RNN contains 128
hidden units and uses hyperbolic tangent as the activation function. The LIF network is formed by a
fully connected population of 220 regular spiking neurons.
In the learning to learn task, the network was trained for 5, 000 iterations with a batch size of 10
episodes, using the RMSProp (Hinton et al., 2012) optimizer, and a learning rate of 0.001.
Propagation of gradients in recurrent networks of LIF neurons: In artificial recurrent neural
networks such as LSTMs, gradients can be computed with back propagation through time (BPTT).
For BPTT in spiking neural networks, complications arise from the non-differentiability of the output of spiking neurons, and from the fact that gradients need to be propagated either through continuous time or through many time steps if time is discretized. Therefore, in (Courbariaux et al., 2016;
Esser et al., 2016) it was proposed to use a pseudo-derivative.
dzj (t)
:= max{0, 1 − |vj (t)|},
dvj (t)
V (t)−B (t)
(8)
where vj (t) denotes the normalized membrane potential vj (t) = j Bj (t)j . This made it possible
to train deep feed-forward networks of deterministic binary neurons (Courbariaux et al., 2016; Esser
et al., 2016). We observed that this convention tends to be unstable for very deep (unrolled) recurrent
networks of spiking neurons (see Fig. 1E). To achieve stable performance we dampened the increase
of back propagated errors through spikes by using a pseudo-derivative of amplitude γ < 1 (typically
γ = 0.3):
dzj (t)
:= γ max{0, 1 − |vj (t)|}.
(9)
dvj (t)
Note that in adaptive neurons, gradients can propagate through many time steps in the dynamic
threshold. This propagation is not affected by the dampening.
14
expected delay between “store”
and “recall” instructions
D (ms)
batch size
LSNN τa = 200ms
LSNN τa = 2s
LSNN τa = 4s
LSNN τa = 6s
LSNN τa = 8s
LSNN τa spread over [0, 8]s
LIF
200ms
2s
4s
6s
uniform in [0-8]s
50
24
0.11
0.24
0.30
0.58
0.40
0.41
0.08
200
24
51
1.59
7.62
7.24
16.3
3.08
49
200
24
48
1.97
3.77
9.17
45.1
3.25
49
200
24
48
15.7
8.27
5.06
34.0
8.12
51
200
256
49
16.9
6.96
3.44
8.33
4.36
49
Table 1: Test error (%) on store-recall tasks with different delays, and different time constants
of adapting neurons in LSNNs with the standard architecture. Mean test errors (over 3 random
network initializations on test batch of size 512) reached by different networks after 500 training
iterations. Best and second best results per task (column) are marked with boldface (D: pattern
presentation time). For comparison, the performance of a fully connected recurrent SNN without
adapting neurons is shown in the last row (LIF).
expected delay
D (ms)
batch size
Standard
Fully connected
Deep R
4s
200
24
3.25
6.39
0.41
uniform in [0-8]s
200
256
4.36
31.8
5.17
Table 2: Test error (%) of different LSNN architectures on two store-recall tasks. Mean test
errors (over 3 random network initializations on test batch of size 512) reached by different networks
after 500 training iterations. Rows show results for SSNs with different architectures. The adaptation
time constants were spread for all these networks uniformly over [0, 8] s, like in the second to last
row of Table 1.
15
| 9cs.NE
|
Comparison between two common collocation approaches based on
radial basis functions for the case of heat transfer equations arising
in porous medium
K. Parand1,a , S. Abbasbandyb,∗, S. Kazemb , A.R. Rezaeia
a
arXiv:1008.2348v2 [math-ph] 16 Aug 2010
b
Department of Computer Sciences, Shahid Beheshti University, G.C., Tehran, Iran
Department of Mathematics, Imam Khomeini International University, Ghazvin 34149-16818, Iran
Abstract
In this paper two common collocation approaches based on radial basis functions have been
considered; one be computed through the integration process (IRBF) and one be computed
through the differentiation process (DRBF). We investigated the two approaches on natural
convection heat transfer equations embedded in porous medium which are of great importance in the design of canisters for nuclear wastes disposal. Numerical results show that the
IRBF be performed much better than the common DRBF, and show good accuracy and
high rate of convergence of IRBF process.
Keywords: Collocation method; Nonlinear ODE; Radial Basis Functions; Direct Inverse
Multiquadric; Indirect Multiquadric; Porous media.
PACS: 47.56.+r, 02.70.Hm
1. Introduction
Natural convective heat transfer in porous media has received considerable attention
during the past few decades. This interest can be attributed due to its wide range of
applications in ceramic processing, nuclear reactor cooling system, crude oil drilling, chemical
reactor design, ground water pollution and filtration processes. External natural convection
in a porous medium adjacent to heated bodies was analyzed by Nield and Bejan [1], Merkin
[2, 3], Minkowycz and Cheng [4, 5], Pop and Cheng [6, 7], Ingham and Pop [8]. In all of
these analysis, it was assumed that boundary layer approximations are applicable and the
coupled set of governing equations were solved by numerical methods.
In this paper, the same approximations are applied to the problem of natural convection
about an inverted heated cone embedded in a porous medium of infinite extent. No similarity
∗
Corresponding author. Tel:(+98912) 1305326 Fax:(+98281) 3780040
Email addresses: [email protected] (K. Parand), [email protected] (S. Abbasbandy),
[email protected] (S. Kazem), [email protected] (A.R. Rezaei)
1
Member of research group of Scientific Computing.
August 17, 2010
solution exists for the truncated cone, but for the case of full cone, if the prescribed wall
temperature or surface heat flux is a power function of distance from the vertex of the
inverted cone similarity solutions exist [1, 6], a great deal of information is available on heat
and fluid flow about such cones as reviewed by Refs. [9, 10].
Bejan and Khair [11] used Darcy’s law to study the vertical natural convective flows
driven by temperature and concentration gradients. Nakayama and Hossain [12] applied
the integral method to obtain the heat and mass transfer by free convection from a vertical
surface with constant wall temperature and concentration. Yih [13] examined the coupled
heat and mass transfer by free convection over a truncated cone in porous media for variable
wall temperature and concentration or variable heat and mass fluxes and [14] applied the
uniform transpiration effect on coupled heat and mass transfer in mixed convection about
inclined surfaces in porous media for the entire regime. Cheng [15] used an integral approach
to study the heat and mass transfer by natural convection from truncated cones in porous
media with variable wall temperature and [16] studied the Soret and Dufour effects on the
boundary layer flow due to natural convection heat and mass transfer over a vertical cone in
a porous medium saturated with Newtonian fluids with constant wall temperature. Natural
convective mass transfer from upward-pointing vertical cones, embedded in saturated porous
media, has been studied using the limiting diffusion [17]. The natural convection along
an isothermal wavy cone embedded in a fluid-saturated porous medium are presented in
[18, 19]. Lai and Kulacki [20] studied the natural convection boundary layer flow along a
vertical surface with constant heat and mass flux including the effect of wall injection. In
[21] fluid flow and heat transfer of vertical full cone embedded in porous media have been
solved by Homotopy analysis method [22, 23].
Mathematical modeling of many problems in science and engineering leads to ordinary
differential equations (ODEs) [24–28]. The methods based on radial basis functions (RBF)
which are part of an emerging field of mathematics are famous ways to solve these kinds
of problems . First studied by Roland Hardy, an Iowa State geodesist, in 1968, these
methods allow for scattered data to be easily used in computations [29]. The concept of
solving DEs by using RBFs was first introduced by Kansa [30]. Since then, it has received
a great deal of attention from researchers. And consequently, many further interesting
developments and applications have been reported (e.g. Zerroukat et al.[31], Mai-Duy and
Tran-Cong[32, 33]). Essentially, in a typical RBF collocation method, each variable and its
derivatives are all expressed as weighted linear combinations of basis functions, where the
sets of network weights are identical. These closed forms of representations are substituted
with the governing equations as well as boundary conditions, and the point collocation
technique is then employed to discretize the system [27]. If all basis functions in networks
are available in analytic forms, the RBF collocation methods can be regarded as truly
meshless methods [34]. There are two basic approaches for obtaining new basis functions
from RBFs, namely direct approach (DRBF) based on a differential process (Kansa [30])
and indirect approach (IRBF) based on an integration process (Mai-Duy and Tran-Cong
[29, 32, 35]). Both approaches were tested on the solution of second order DEs and the
indirect approach was found to be superior to the direct approach (Mai-Duy and Tran-Cong
[32]).
2
In this paper we apply the DRBF and IRBF for solving natural convection of Darcian
fluid about a vertical full cone embedded in porous media prescribed surface heat flux which
is third order nonlinear ODE.
2. Problem formulation
Consider an inverted cone with semi-angle γ and take axes in the manner indicated in
Fig. 1(a). The boundary layer develops over the heated frustum x = x0 .
The boundary layer equations for natural convection of Darcian fluid about a cone are
[6]:
∂
∂
(ru) +
(rv) = 0,
∂x
∂y
ρ∞ βKg cos γ(T − T∞ )
u=
,
µ
∂T
∂2T
∂T
+v
=α 2.
u
∂x
∂y
∂y
(2.1)
(2.2)
For a thin boundary layer, r is obtained approximately xsin(γ). Suppose that a power law
of heat flux is prescribed on the frustum. Accordingly, the boundary conditions at infinity
are:
u = 0,
T = T∞ , if y → ∞
(2.3)
(2.4)
and at the wall are
v = 0 if y = 0.
If the surface heat flux qw [6] is prescribed, qw is obtained as
qw = −k(
∂T
)y=0 = A(x − x0 )λ ,
∂y
x0 ≤ x ≤ ∞.
For the case of a full cone (x0 = 0, F ig.1(b)) a similarity solution exists [6].
In the case of prescribed surface heat flux the similarity solution for the stream function
ψ and T where
u=
1 ∂ψ
,
r ∂y
v=−
1 ∂ψ
r ∂x
(2.5)
is of the form [6]:
ψ = αr(Rax )1/3 f (η),
1
qw x
T − T∞ =
(Rax )− 3 θ(η),
k
y
1/3
η = (Rax ) ,
x
(2.6)
3
where
Rax =
ρ∞ βgK cos(γ)qw x2
µαk
(2.7)
is the local Rayleigh number for the case of prescribed surface heat flux. The governing
equations become
f ′ = θ,
λ + 5 ′ 2λ + 1 ′
fθ −
f θ = 0,
θ′′ +
2
3
subjected to boundary conditions as:
f (0) = 0,
θ′ (0) = −1,
θ(∞) = 0.
Finally from Eqs. (2.8) and (2.9) we have:
(
′′
′ 2
f f − 2λ+1
(f ) = 0,
ODE. f ′′′ + λ+5
2
3
′′
B.C. f (0) = 0, f (0) = −1, f ′ (∞) = 0.
(2.8)
(2.9)
(2.10)
It is of interest to obtain the value of the local Nusselt number which is defined as [6]:
qw x
Nux =
.
(2.11)
k(Tw − T∞ )
From Eqs. (2.11), (2.6) and (2.7) it follows that the local Nusselt number which is interest
to obtain given by:
Nux = Ra1/3
x [−θ(0)].
(2.12)
3. RBF Functions
Let R+ = {x ∈ R, x ≥ 0} be the non-negative half-line and let φ : R+ → R be a
continuous function with φ(0) ≥ 0. A radial basis function on Rd is a function of the form
φ(kX − Xi k)
(3.1)
where X, Xi ∈ Rd and k.k denotes the Euclidean distance between X, Xi . If one chooses
N points {Xi }N
i=1 in R then by custom
s(X) =
N
X
i=1
λi φ(kX − Xi k);
λi ∈ R,
(3.2)
is called a radial basis function as well [36].
In order to explain RBF methods briefly, suppose that the one-dimensional input data point
set or the center set xi in the given domain Ω ⊆ R is given. The center point is not necessarily
structured, that is, it can have an arbitrary distribution. The arbitrary grid structure is one
of the major differences between the RBF method and other global methods. Such a meshfree grid structure yields high flexibility especially when the domain is irregular. In this
work the uniform grid is used for RBF approximation.
4
3.1. Properties of RBF
With a radial function φ(r) and with data values ui given at the locations xi , for i =
1, 2, ..., N the function
s(x) =
N
X
λi φi (x)
(3.3)
i=1
where r = ri = kx − xi k and φi (x) = φ(kx − xi k), interpolates the data if we choose the
expansion coefficients λi in such a way that s(xj ) = uj , for j = 1, 2, ..., N [37, 38]. The
expansion coefficients λi can therefore be obtained by solving the linear system AΛ = U,
where:
Aij = φ(kxj − xi k),
Λ = [λ1 , λ2 , ..., λN ]T ,
U = [u1 , u2 , . . . , uN ]T .
(3.4)
(3.5)
(3.6)
All the infinitely smooth RBF choices listed in Table (1) will give coefficient matrices
A in (3.4) which are symmetric and nonsingular [39], i.e. there is a unique interpolant of
the form (3.3) no matter how the distinct data points are scattered in any number of space
dimensions. In the cases of inverse quadratic, inverse multiquadric and GA the matrix
A is positive definite and, for multiquadric (MQ), it has one positive eigenvalue and the
remaining ones are all negative [39].
Interpolation using Conical splines and thin-plate splines (TPSs) can become singular in
multidimensions [34]. However, low-degree polynomials can be added to the RBF interpolant
to guarantee that the interpolation matrix is positive definite (a stronger condition than
nonsingularity). For example, for the Conical RBF and
Pm the TPS inPdNdimensions this
becomes the case if we use as an interpolant s(x) =
i=1 λi φ(kx − xi k)
i=1 ai pi (x) +
PN
together with the constraints j=1 λj pi (xj ) = 0, for i = 1, 2, ..., m. Here pi (x) denotes a
basis for polynomials of Pqd in Rd (Pqd denotes the space of d-variate polynomials of order
not exceeding q) and m = (q − 1 + d)!/(d!(q − 1)!) [40].
3.2. RBF Interpolation
One dimensional function u(x) to be interpolated or approximated can be represented
by an RBF as:
N
X
λi φi (x) = ΦT (x)Λ
(3.7)
ΦT (x) = [φ1 (x), φ2 (x), ..., φN (x)],
Λ = [λ1 , λ2 , ..., λN ]T ,
(3.8)
(3.9)
u(x) ≈ s(x) =
i=1
where
5
x is the input and λi s are the set of coefficients to be determined. By choosing N interpolate
nodes (xj , j = 1, 2, ..., N) in Ω ∪ ∂Ω, the function u(x) can be approximated in Ω ∪ ∂Ω.
uj =
N
X
λi φi (xj ),
(j = 1, 2, ..., N).
(3.10)
i=1
To brief discussion on coefficient matrix we define:
AΛ = U
(3.11)
where
U =[u1 , u2 , ..., uN ]T ,
(3.12)
A =[ΦT (x1 ), ΦT (x2 ), ..., ΦT (xN )]T
φ1 (x1 ) φ2 (x1 ) . . . φN (x1 )
φ1 (x2 ) φ2 (x2 ) . . . φN (x2 )
=
..
..
..
..
.
.
.
.
φ1 (xN ) φ2 (xN ) . . . φN (xN )
.
(3.13)
Note that φi (xj ) = φ(kxi − xj k) therefore φi (xj ) = φj (xi ) consequently A = AT .
The shape parameter c which is shown in Table (1) affects both the accuracy of the approximation and the conditioning of the interpolation matrix [44]. In general, for a fixed number
of N, smaller shape parameters produce the more accurate approximations, but also are
associated with a poorly conditioned A. The condition number also grows with N for fixed
values of the shape parameter c. Many researchers (e.g.[41, 42]) have attempted to develop
algorithms for selecting optimal values of the shape parameter. The optimal choice of the
shape parameter is still an open question. In practice it is most often selected by brute force.
Recently, Fornberg et. al.[43] have developed a Contour-Padé algorithm which is capable of
stably computing the RBF approximation for all c > 0 [44].
The following theorem about the convergence of RBF interpolation is discussed [45, 46].
Theorem 3.1. assume xi , (i = 1, 2, ..., N) are N nodes in Ω which is convex, let
h = max min kx − xi k2
x∈Ω 1≤i≤N
R
when φ̂(η) < c(1 + |η|)−(2l+d) for any u(x) satisfies (û(η))2 /φ̂(η)dη < ∞ we have
kuN (α) − u(α) k∞ ≤ chl−α
where φ(x) is RBF and the constant c depends on the RBF, d is space dimension, l and
α are nonnegative integer. It can be seen that not only RBF itself but also its any order
derivative has a good convergence.
6
3.3. Direct RBF for ODEs (DRBF)
In the direct method, the closed form DRBF approximating function (3.7) is first obtained from a set of training points, and its derivative of any order, e.g. pth order, can then
be calculated in a straightforward manner by differentiating such a closed form DRBF as
follows:
N
N
N
X
dk s(x)
dk X
dk φi (x) X
[k]
= k(
λi φi (x)) =
λi
=
λi Gi (x)
k
k
dx
dx i=1
dx
i=1
i=1
(3.14)
where
[k]
Gi (x) = dk φi (x)/dxk ,
k = 0, 1, ..., p.
Now we aim to apply the DRBF method for solving the ODEs in general form :
(
F (x, u, u′, ..., u(p−1) , u(p) ) = 0, a ≤ x ≤ b,
u(i) (ei ) = αi+1 ,
i = 0, 1, ..., p − 1,
(3.15)
where ei ∈ {a, b} and u(i) (x) = di u(x)/dxi , F is known function and {αi }pi=1 are known
constants. By substituting Eq. (3.7) in (3.15) and using Eq. (3.14) we have:
F (x, s, s′ , ..., s(p) ) = F (x,
N
X
i=1
[0]
λi Gi (x),
N
X
i=1
[1]
λiGi (x), ...,
N
X
[p]
λi Gi (x)).
i=1
Now, to obtain λi s (i = 1, 2, ..., N) we define the residual function:
Res(x) = F (x, s, s′ , ..., s(p) ).
(3.16)
The set of equations for obtaining the coefficients {λi }N
i=1 come from equalizing Eq. (3.16)
N −p
to zero at N − p interpolate nodes {xj }j=1 plus p boundary conditions:
(
Res(xj ) = 0,
j = 1, 2, ..., N − p,
(3.17)
PN
[k]
i=1 λi Gi (ei ) = αi+1 , k = 0, 1, ..., p − 1.
Since the direct approach is based on a differentiation process, all derivatives obtained here
are very sensitive to noise arising from the interpolation of DRBFs from a set of discrete
data points. Any noise here, even at the small level, will be badly magnified with an increase
in the order of derivative [29].
7
3.4. Indirect RBF for ODEs (IRBF)
In the indirect method, the formulation of the problem starts with the decomposition
of the highest order derivative under consideration into RBFs. The obtained derivative
expression is then integrated to yield expressions for lower order derivatives and finally
for the original function itself. In contrast, the integration process, where each integral
represents the area under the corresponding curve, is much less sensitive to noise. Based
on this observation, it is expected that through the integration process, the approximating
functions are much smoother and therefore have higher approximation power. Also To
numerically explore tile IRBF methods with shape parameters for which the interpolation
matrix is too poorly conditioned to use standard methods [47]. Let p be the highest order
of the derivative under consideration the boundary value ODEs in general form Eq. (3.15)
when (∃k s.t. ek = b) then we can define:
N
dp ŝ(x) X
=
λi φi (x),
dxp
i=1
Z X
Z
N
N
N
p−1
X
X
d ŝ(x)
[p−1]
=
λ
φ
(x)dx
=
λ
φ
(x)dx
=
λi hi (x) + d1 ,
i i
i
i
dxp−1
i=1
i=1
i=1
(3.18)
..
.
dŝ(x)
=
dx
=
Z X
N
[2]
λi hi (x)dx + d1
i=1
N
X
[1]
λi hi (x)
i=1
xp−3
xp−2
+ d2
+ ... + dp−1
(p − 2)!
(p − 3)!
xp−2
xp−3
+ d1
+ d2
+ ... + dp−1 ,
(p − 2)!
(p − 3)!
Z X
N
xp−3
xp−2
[1]
+ d2
+ ... + dp−1)dx
λi hi (x) + d1
ŝ(x) = (
(p − 2)!
(p − 3)!
i=1
=
N
X
i=1
[0]
λi hi (x) + d1
xp−2
xp−1
+ d2
+ ... + dp−1 x + dp ,
(p − 1)!
(p − 2)!
where
(R
φi (x)dx,
k = p − 1,
[k]
hi (x) = R [k+1]
hi (x)dx, k = 0, 1, ..., p − 2.
(3.19)
Substituting Eqs. (3.18) in (3.16) at N interpolate nodes {xj }N
j=1 plus p boundary conditions
p
the set of coefficients {λi }N
and
{d
}
is
obtained
as
follow
:
j j=1
i=1
(
Res(xj ) = 0,
ŝ(i) (ei ) = αi+1 ,
j = 1, 2, ..., N,
i = 1, 2, ..., p.
8
4. Solving the model
Consider governing equation of fluid flow and heat transfer of full cone embedded in
porous medium that is expressed by Eq. (2.10) for prescribed surface heat flux.
In the first step of our analysis, we approximate f (η) for solving the model by DRBF:
f (η) ≃ s(η) =
N
X
λi φi (η),
(4.1)
i=1
and f ′′′ (η) for solving the model by IRBF:
N
X
d3
f (η) ≃ 3 ŝ(η) =
λi φi (η).
dη
i=1
′′′
The general form of problem appear to:
(
Fλ (η, f, f ′ , f ′′ , f ′′′ ) = 0,
f (0) = 0, f ′′ (0) = −1, f ′ (∞) = 0.
(4.2)
(4.3)
To solve this problem we define residual function:
Res(η) = Fλ (η, s, s′, s′′ , s′′′ );
f or DRBF,
(4.4)
Res(η) = Fλ (η, ŝ, ŝ′ , ŝ′′ , ŝ′′′ );
f or IRBF.
(4.5)
The unknown coefficients {λi }N
i=1 come from equalizing Res(η) to zero at N interpolate
nodes ηi from Uniform distribution between 0 and η∞ which we set 9/2 for this problem.
4.1. Solving the model by DRBF
In the first step of solving, φi (η) is set by inverse multiquadric function which is shown in
Table (1). Now, the residual function is constructed by substituting Eq. (4.1) in Eq. (4.4):
Res(η) =
N
X
i=1
N
N
N
X
2λ + 1 X
λ+5 X
′′
′
)(
λi φi (η))(
λi φi (η)) − (
)(
λi φi (η))2 .
λi φi (η) + (
2
3
i=1
i=1
i=1
′′′
N −2
By using N − 2 interpolate nodes {ηj }j=1
plus two boundary conditions of Eq. (4.3)
′′
(f (0) = 0, f (0) = −1), the set of equations can be solved, consequently the coefficients
{λi }N
i=1 will be obtained:
Res(ηj ) = 0,
j = 1, 2, ..., N − 2,
P
N
i=1 λi φi (0) = 0,
PN
′′
i=1 λi φi (0) = −1.
9
Take into account φ′i (∞) = 0, for i = 1, 2, ..., N the infinity boundary condition (f ′ (∞) = 0)
is already satisfied.
Table (2) show the f ′ (η) for some λ in comparison with solutions of [21]. Also f ′ (η) for two
selected λ = 1/4 and 3/4 are showed in Table (3) in comparison with Runge-Kutta solution
is obtained by the MATLAB software command ODE45 which is used and applied by the
authors in ref. [21]. Absolute errors show that DRBF give us approximate solution with
a high degree of accuracy with a small N. The resulting graph of Eq. (2.10) is shown in
Figure (2).
4.2. Solving the model by IRBF
In the first of solving φi (η) is set by multiquadric function which is shown in Table (1).
Now, the residual function is constructed by substituting Eq. (4.2) in Eq. (4.5) and using
Eq. (3.18):
Z Z Z
Z
N
N
N
X
X
λ+5 X
)
λi
φi (η)dη
λi φi (η)dη
Res(η) =
λi φi (η) + (
2
i=1
i=1
i=1
Z
Z
N
2λ + 1 X
−(
)(
λi
φi (η)dη)2
3
i=1
=
N
X
i=1
N
N
X
d1 η 2
λ+5 X
[0]
[2]
)(
λi hi (η) +
+ d2 η + d3 )(
λi hi (η) + d1 )
λi φi (η) + (
2
2
i=1
i=1
N
2λ + 1 X
[1]
)(
λi hi (η) + d1 η + d2 )2 .
−(
3
i=1
By using N interpolate nodes {ηj }N
j=1 plus three boundary conditions of Eq. (4.3) the set of
3
equations can be solved, consequently the coefficients {λi }N
i=1 and {di }i=1 will be obtained.
In this method we put η∞ instead of infinity condition:
Res(ηj ) = 0,
j = 1, 2, ..., N,
PN λ h[0] (0) + d = 0,
i i
3
Pi=1
[2]
N
i=1 λi hi (0) + d1 = −1,
PN
[1]
i=1 λi hi (η∞ )d1 η∞ + d2 = 0.
Table (4) shows the f ′ (η) for some λ in comparison with solutions of [21]. Also f ′ (η)
for two selected λ = 1/4 and 3/4 are showed in Table (5) in comparison with Runge-Kutta
solution is obtained by the MATLAB software command ODE45 which is used and applied
by the authors in ref. [21]. Absolute errors show that IRBF give us approximate solution
with a high degree of accuracy with a small N. The resulting graph of Eq. (2.10) is shown
in Figure (3). A graph in figures (4) for λ = 2/3 show kResk2 for some N. Table (6) shows
kResk2 for some N and λ.
Comparison between DRBF solution in Table (3) with N = 12 and IRBF solution in Table
(5) with N = 10 for f ′ (η) show that the convergence of the IRBF method is faster, because
of using less numbers of collocation points.
10
5. Conclusion
In this paper we made a comparison between the two common collocation approaches
based on radial basis functions namely DRBF and IRBF methods on natural convection
equation about an inverted heated cone embedded in a porous medium of infinite extent
which are of great importance in the design of canisters for nuclear wastes disposal. These
functions are proposed to provide an effective but simple way to improve the convergence
of the solution by collocation method. The direct approach (DRBF) is based on a differentiation process, all derivatives are very sensitive to noise arising from the interpolation of
DRBFs from a set of discrete data points. Any noise, even at the small level, will be bad
magnified with an increase in the order of derivative. The indirect technique (IRBF) which
is based on integration process, each integral represents the area under the corresponding
curve, is much less sensitive to noise. Based on this observation, it is expected that through
the integration process, the approximating functions are much smoother and therefore have
higher approximation power. Additionally, through the comparison with other methods
such as HAM we show that the RBFs methods have good reliability and efficiency. Also
high convergence rates and good accuracy are obtained with the proposed method using
relatively low numbers of data points.
References
[1] D.A. Nield, A. Bejan, Convection in Porous Media, third ed., Springer-Verlag, New York, (2006).
[2] J.H. Merkin, Free convection boundary layers in a saturated porous medium with lateral mass flux, Int.
J. Heat Mass Transfer, 21 (1978) 1499-1504.
[3] J.H. Merkin, Free convection boundary layers on axisymmetric and two dimensional bodies of arbitrary
shape in a saturated porous medium, Int. J. Heat Mass Transfer, 22 (1979) 1461-1462.
[4] W.J. Minkowycz, P. Cheng, Local non-similar solutions for free convective flow with uniform lateral
mass flux in a porous medium, Lett. Heat Mass Transfer, 9 (1982) 159-168.
[5] W.J. Minkowycz, P. Cheng, F. Moalem, The effect of surface mass transfer on buoyancy induced Darcian
flow adjacent to a horizontal heated surface, Int. Commun. Heat Mass Transfer, 12 (1985) 55-65.
[6] P. Cheng, T.T. Le, I. Pop, Natural convection of a Darcian fluid about a cone, Int. Commun. Heat Mass
Transfer, 12 (1985) 705-717.
[7] I. Pop, P. Cheng, An integral solution for free convection of a Darcian fluid about a cone with curvature
effects, Int. Commun. Heat Mass Transfer, 13 (1986) 433-438.
[8] D.B. Ingham, I. Pop, Natural convection about a heated horizontal cylinder in a porous medium, J.
Fluid Mech. 184 (1987) 157-181.
[9] I. Pop, D.B. Ingham Convective heat transfer: mathematical and computational modeling of viscous
fluids and porous media, Pergamon Press, Oxford, 2001.
[10] K. Vafai, Handbook of porous media. Marcel Dekker, New York, 2000.
[11] A. Bejan, K.R. Khair, Heat and mass transfer by natural convection in a porous medium, Int. J. Heat
Mass Transfer, 28 (1985) 909-918.
[12] A. Nakayama, M.A. Hossain, An integral treatment for combined heat and mass transfer by natural
convection in a porous medium, Int. J. Heat Mass Transfer, 38 (1995) 761-765.
[13] K.A. Yih, Coupled heat and mass transfer by free convection over a truncated cone in porous media:
VWT/VWC or VHF/VMF, Acta Mech. 137 (1999) 83-97.
[14] K.A. Yih, Uniform transpiration effect on coupled heat and mass transfer in mixed convection about
inclined surfaces in porous media : the entire regime, Acta Mech. 132 (1999) 229-240.
11
[15] C.Y. Cheng, An integral approach for heat and mass transfer by natural convection from truncated cones
in porous media with variable wall temperature and concentration, Int. Commun. Heat Mass Transfer,
27 (2000) 437-548.
[16] C.Y. Cheng, Soret and Dufour effects on natural convection heat and mass transfer from a vertical cone
in a porous medium, Int. Commun. Heat Mass Transfer, 36 (2009) 1020-1024.
[17] S.U. Rahman, K. Mahgoub, A. Nafees, Natural Convective Mass Transfer from Upward Pointing Conical Surfaces in Porous Media, Chem. Eng. Commun. 194 (2007) 280-290.
[18] I. Pop, T.Y. Na, Naturnal convection of a Darcian fluid about a wavy cone, Int. Commun. Heat Mass
Transfer, 21 (1994) 891-899.
[19] I. Pop, T.Y. Na, Natural convection over a frustum of a wavy cone in a porous medium, Mech. Res.
Commun. 22 (1995) 181-190.
[20] F. C. Lai and F. A. Kulacki, Coupled heat and mass transfer by natural convection from vertical
surfaces in porous media, Int. J. Heat Mass Transfer, 34,4-5 (1991) 1189-1194.
[21] A.R. Sohouli, M. Famouri, A. Kimiaeifar, G. Domairry, Application of homotopy analysis method for
natural convection of Darcian fluid about a vertical full cone embedded in pours media prescribed surface
heat flux, Commun. Nonlinear Sci. Numer. Simul. 15 (7) (2010) 1691-1699.
[22] S. Abbasbandy, T. Hayat, Solution of the MHD Falkner-Skan flow by homotopy analysis method,
Commun. Nonlinear Sci. Numer. Simul. 14 (9-10) (2009) 3591-3598.
[23] S. Abbasbandy, Approximate solution for the nonlinear model of diffusion and reaction in porous
catalysts by means of the homotopy analysis method, Chem. Eng. J. 136 (2008) 144-150.
[24] K. Parand, M. Razzaghi, Rational Legendre approximation for solving some physical problems on
semi-infinite intervals, Phys. Scr. 69 (2004) 353-357.
[25] K. Parand, M. Dehghan, A. Pirkhedri, Sinc-collocation method for solving the Blasius equation, Phys.
Lett. A, 373 (44) (2009) 4060-4065.
[26] K. Parand, M. Dehghan, A.R. Rezaei, S.M. Ghaderi.An approximational algorithm for the solution
of the nonlinear Lane-Emden type equations arising in astrophysics using Hermite functions collocation
method, Comput. Phys. Commun. 2010; DOI:10.1016/j.cpc.2010.02.018.
[27] K. Parand, A.R. Rezaei, S.M. Ghaderi, An approximate solution of the MHD Falkner-Skan
flow by Hermite functions pseudospectral method, Commun. Nonlinear Sci. Numer. Simul. 2010;
Doi:10.1016/j.cnsns.2010.03.022.
[28] S. Abbasbandy, E. Shivanian, Exact analytical solution of a nonlinear equation arising in heat transfer,
Phys. Lett. A, 374 (4) (2010) 567-574.
[29] N. Mai-Duy, Solving high order ordinary differential equations with radial basis function networks, Int.
J. Numer. Meth. Engng. 62 (6) (2005) 824-852.
[30] E.J. Kansa, Multiquadrics-A scattered data approximation scheme with applications to computational
fluiddynamics II. Solutions to parabolic, hyperbolic and elliptic partial differential equations, Comput.
Math. Appl. 19 (8,9) (1990) 147-161.
[31] M. Zerroukat, H. Power, C.S. Chen, A numerical method for heat transfer problems using collocation
and radial basis functions, Int. J. Numer. Meth. Engng. 42 (1998) 1263-1278.
[32] N. Mai-Duy, T. Tran-Cong, Numerical solution of differential equations using multiquadric radial basis
function networks, Neural Netw. 14(2)(2001) 185-199.
[33] N. Mai-Duy, T. Tran-Cong, Numerical solution of NavierStokes equations using multiquadric radial
basis function networks, Int. J. Numer. Meth Fluids, 37 (2001) 65-86.
[34] B. Fornberg, Comparisons between pseudospectral and radial basis function derivative approximations,
IMA J. Numer. Anal. 30 (1) (2008) 149-172.
[35] N. Mai-Duy, T. Tran-Cong, Approximation of function and its derivatives using radial basis function
network methods, Appl. Math. Modelling, 27 (2003) 197-220.
[36] M.A. Golberg, C.S. Chen, H. Bowman, Some recent results and proposals for the use of radial basis
functions in the BEM, Eng. Anal. Bound. Elem. 23 (4) (1999) 285-296.
[37] M.D. Buhmann, Radial basis functions, Cambridge University Press, Cambridge, 2003.
[38] G.E. Fasshauer, Meshfree Approximation Methods with Matlab, World Scientific Publishing, Singapore,
12
(2007).
[39] M.J.D. Powell, The theory of radial basis function approximation in 1990, Advances in Numerical
Analysis, Clarendon, Oxford, (1992).
[40] M. Dehghan, A. Shokri, A meshless method for numerical solution of the one-dimensional wave equation
with an integral condition using radial basis functions, Numer. Algo. 52(3) (2009) 461-477.
[41] R.E. Carlson, T.A. Foley, The parameter r2 in multiquadric interpolation, Comput. Math. Appl. 21
(9) (1991) 29-42.
[42] S. Rippa, An algorithm for selecting a good parameter c in radial basis function interpolation, Adv.
Comput. Math. 11 (1999) 193-210.
[43] B. Fornberg, T. Dirscol, G. Wright, R. Charles, Observations on the behavior of radial basis function
approximations near boundaries, Comput. Math. Appl. 43 (2002) 473-490.
[44] S.A. Sarra, Adaptive radial basis function method for time dependent partial differential equations,
Appl. Numer. Math. 54 (2005) 79-94.
[45] Z.M. Wu, Radial basis function scattered data interpolation and the meshless method of numerical
solution of PDEs, J. Eng. Math. 19 (2) (2002), pp. 112 [In Chinese].
[46] Z.M. Wu, R. Schaback, Local error estimates for radial basis function interpolation of scattered data,
IMA J. Numer. Anal. 13 (1993) 13-27.
[47] B. Fornberg, G. Wright, Stable computation of multiquadric interpolants for all values of the shape
parameter, Computers Math. Applic. 48 (5/6) (2004) 853-867.
13
List of Figures
1
2
3
4
(a) Coordinate system for the boundary layer on a heated frustum of a cone,
(b) full cone, x0 = 0. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
DRBF approximation of f ′ (η) for different values λ = 0, 1/4, 1/3, 1/2, 3/4
and 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
IRBF approximation of f ′ (η) for different values λ = 0, 1/4, 1/3, 1/2, 3/4
and 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
kResk2 for λ = 2/3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
16
17
18
List of Tables
1
2
3
4
5
6
Some well–known functions that generate RBFs (r = kx − xi k = ri ), c > 0 .
A comparison between solutions obtained by [21] and the DRBF method for
f ′ (0) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison between DRBF solution and Runge-Kutta solution for f ′ (η) for
λ = 1/4 and λ = 3/4 with N = 12 . . . . . . . . . . . . . . . . . . . . . . . .
A comparison between solutions obtained by [21] and the IRBF method for
f ′ (0) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison between IRBF solution and Runge-Kutta solution for f ′ (η) for
λ = 1/4 and λ = 3/4 with N = 10 . . . . . . . . . . . . . . . . . . . . . . . .
kResk2 for different N and λ by IRBF . . . . . . . . . . . . . . . . . . . . .
14
20
21
21
22
22
23
xxxxxxxx
xxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxx r(x)xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
g
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxx
u
xxxxx
xxxxx
X0
x
v
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxx
xxxxxxxxx
xxxxxxxxx
xxxxxxxxx
y
(a)
(b)
Figure 1: (a) Coordinate system for the boundary layer on a heated frustum of a cone, (b) full cone, x0 = 0.
15
Figure 2: DRBF approximation of f ′ (η) for different values λ = 0, 1/4, 1/3, 1/2, 3/4 and 1
16
Figure 3: IRBF approximation of f ′ (η) for different values λ = 0, 1/4, 1/3, 1/2, 3/4 and 1
17
Figure 4: kResk2 for λ = 2/3
18
Nomenclature
A
prescribed constant
f
similarity function for stream function temperature
g
acceleration due to gravity parameter
K
permeability of the fluid-saturated porous medium
N ux local Nusselt number
qw
surface heat flux fluid-saturated porous medium
r
local radius of the cone fluid
Rax local Raleigh number
T
temperature
T∞
ambient temperature
u, v
velocity vector along x,y axis
x, y
Cartesian coordinate system
x0
distance of start point of cone from the vertex
Greek symbols
α
thermal diffusivity the fluid-saturated porous medium
β
expansion coefficient of the fluid
η
independent dimensionless
θ
similarity function for
λ
prescribed constants
µ
viscosity of the fluid
ρ∞
density of the fluid at infinity
ψ
stream function
19
Table 1: Some well–known functions that generate RBFs (r = kx − xi k = ri ), c > 0
Name of functions
Multiquadrics (MQ)
Inverse multiquadrics (IMQ)
Thin plate (polyharmonic)Splines (TPS)
Conical splines
Gaussian (GA)
Exponential spline
Definition
√
r 2√+ c2
1/( r 2 + c2 )
(−1)k+1 r 2k log(r)
r 2k+1
exp(−cr 2 )
exp(−cr)
20
Table 2: A comparison between solutions obtained by [21] and the DRBF method for f ′ (0)
λ
0
1/4
1/3
1/2
3/4
1
Runge-Kutta
Solution[21]
0.94760
0.91130
0.90030
0.87980
0.85220
0.82760
N
10
12
10
10
12
10
DRBF method
c
DRBF
Error with RK
3.46543
0.94750
0.0001
3.943
0.91086
0.00044
4.9665
0.90038
0.0008
5.36
0.87981
0.00001
5.23
0.85227
0.00007
5.89
0.82737
0.00023
Other methods
HAM[21]
0.94783
0.91119
0.90103
0.87964
0.85242
0.82726
Table 3: Comparison between DRBF solution and Runge-Kutta solution for f ′ (η) for λ = 1/4 and λ = 3/4
with N = 12
η
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
λ = 1/4, f ′(η)
DRBF
Runge-Kutta
Solution
Solution[21]
0.910886
0.911295
0.813122
0.813604
0.720720
0.721351
0.634656
0.635531
0.555601
0.556661
0.483877
0.484997
0.419467
0.420587
0.362196
0.363276
0.311655
0.312677
0.267358
0.268264
0.228756
0.229508
0.195268
0.195878
0.166342
0.166847
0.141450
0.141837
0.120090
0.120362
0.101811
0.102025
Error
with RK
0.000409
0.000482
0.000631
0.000875
0.001061
0.001120
0.001120
0.001080
0.001022
0.000906
0.000752
0.000610
0.000505
0.000387
0.000272
0.000214
21
λ = 3/4, f ′(η)
DRBF
Runge-Kutta
Solution
Solution[21]
0.852268
0.852193
0.755678
0.755377
0.666229
0.665448
0.584176
0.582985
0.509628
0.508141
0.442519
0.440849
0.382620
0.380907
0.329539
0.327973
0.282878
0.281536
0.242076
0.241013
0.206652
0.205832
0.176015
0.175434
0.149653
0.149275
0.127026
0.126821
0.107681
0.107596
0.091167
0.091196
Error
with RK
0.000075
0.000301
0.000781
0.001191
0.001487
0.001670
0.001713
0.001566
0.001342
0.001043
0.000820
0.000581
0.000378
0.000205
0.000085
0.000029
Table 4: A comparison between solutions obtained by [21] and the IRBF method for f ′ (0)
λ
0
1/4
1/3
1/2
3/4
1
Runge-Kutta
Solution[21]
0.94760
0.91130
0.90030
0.87980
0.85220
0.82760
N
10
10
10
10
10
10
c
1.860
2.005
2.050
2.150
2.418
2.380
IRBF method
IRBF
Error with RK
0.94758
0.00002
0.91128
0.00002
0.90030
0.00000
0.87979
0.00001
0.85206
0.00014
0.82762
0.00002
Other methods
HAM[21]
0.94783
0.91119
0.90103
0.87964
0.85242
0.82726
Table 5: Comparison between IRBF solution and Runge-Kutta solution for f ′ (η) for λ = 1/4 and λ = 3/4
with N = 10
η
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
2
2.5
3
3.5
4
4.5
λ = 1/4, f ′ (η)
IRBF
Runge-Kutta
Solution
Solution[21]
0.911278
0.911295
0.813594
0.813604
0.721302
0.721351
0.635394
0.635531
0.556440
0.556661
0.484753
0.484997
0.420334
0.420587
0.362989
0.363276
0.312347
0.312677
0.267934
0.268264
0.229197
0.229508
0.195584
0.195878
0.166546
0.166847
0.141548
0.141837
0.120107
0.120362
0.101771
0.102025
0.043769
0.043951
0.018310
0.018546
0.007391
0.007610
0.002716
0.002953
0.000719
0.000962
0.000010
0.000123
Error
with RK
0.000017
0.000010
0.000049
0.000137
0.000221
0.000244
0.000253
0.000287
0.000330
0.000330
0.000311
0.000294
0.000301
0.000289
0.000255
0.000254
0.000182
0.000236
0.000219
0.000237
0.000243
0.000113
22
λ = 3/4, f ′(η)
IRBF
Runge-Kutta
Solution
Solution[21]
0.852059
0.852193
0.755254
0.755377
0.665278
0.665448
0.582736
0.582985
0.507827
0.508141
0.440510
0.440849
0.380599
0.380907
0.327647
0.327973
0.281184
0.281536
0.240696
0.241013
0.205482
0.205832
0.175053
0.175434
0.148891
0.149275
0.126408
0.126821
0.107158
0.107596
0.090757
0.091196
0.038887
0.039223
0.016225
0.016574
0.006433
0.006832
0.002255
0.002668
0.000445
0.000913
−0.00001
0.000237
Error
with RK
0.000134
0.000123
0.000170
0.000249
0.000314
0.000339
0.000308
0.000326
0.000352
0.000317
0.000350
0.000381
0.000384
0.000413
0.000438
0.000439
0.000336
0.000349
0.000399
0.000413
0.000468
0.000247
λ
0
1/4
1/3
1/2
3/4
1
N=5
1.161037
0.851976
0.770207
0.637068
0.469112
0.355474
Table 6: kResk2 for different N and λ by IRBF
N=6
0.306257
0.278068
0.269714
0.216095
0.230425
0.221336
N=8
0.016229
0.018364
0.018842
0.016813
0.020280
0.020463
23
N=10
0.000108
0.000363
0.000494
0.000587
0.001051
0.001077
N=12
0.77e − 5
0.52e − 4
0.24e − 4
0.83e − 5
0.15e − 4
0.31e − 4
N=15
0.11e − 7
0.27e − 6
0.13e − 5
0.47e − 6
0.83e − 6
0.39e − 6
| 5cs.CE
|
AngularJS in the Wild: A Survey with 460 Developers
Miguel Ramos
Marco Tulio Valente
arXiv:1608.02012v3 [cs.SE] 27 Sep 2016
UFMG, Brazil
{miguel,mtov}@dcc.ufmg.br
Ricardo Terra
Gustavo Santos
UFLA, Brazil
[email protected]
RMoD Team, INRIA, France
[email protected]
Abstract
To implement modern web applications, a new family of
JavaScript frameworks has emerged, using the MVC pattern.
Among these frameworks, the most popular one is A NGU LAR JS, which is supported by Google. In spite of its popularity, there is not a clear knowledge on how A NGULAR JS
design and features affect the development experience of
Web applications. Therefore, this paper reports the results of
a survey about A NGULAR JS, including answers from 460
developers. Our contributions include the identification of
the most appreciated features of A NGULAR JS (e.g., custom
interface components, dependency injection, and two-way
data binding) and the most problematic aspects of the framework (e.g., performance and implementation of directives).
Categories and Subject Descriptors D.3.3 [Frameworks]
Keywords JavaScript, AngularJS, MVC frameworks.
1.
Introduction
JavaScript is a fundamental piece of modern Web applications. It was initially designed as a scripting language to extend web pages with small executable code. However, the
language is used nowadays to construct a variety of complex systems (Kienle 2010; Silva et al. 2015). As a result, we are observing the birth of new technologies and
tools—including JavaScript libraries and frameworks—to
solve common problems faced in the development of such
applications. For example, frameworks following the ModelView-Controller (MVC) architecture pattern (or variations
of it) are widely used nowadays, including systems such as
A NGULAR JS, BACKBONE . JS, and E MBER . JS. Among these
frameworks, A NGULAR JS is probably the most popular one.
This fact is evidenced by comparing the number of Google
searches (the most queried framework since 2013), the num-
ber of contributors in GitHub and the increasing number of
questions and answers in Stack Overflow (the framework
with more Q&A since mid-2013).
However, despite the increasing practical interest on A N GULAR JS, there is no clear knowledge on how the design
and features proposed by this framework affect the development experience of JavaScript software. More specifically, it
is not clear what are the most appreciated features of A N GULAR JS, what are the main problems faced by developers when using the framework, and which aspects of A N GULAR JS can be improved. Answers to these questions are
important to different developers. First, developers who use
A NGULAR JS can learn how to improve this usage and also
how to avoid bad A NGULAR JS programming practices. Second, developers who do not use JavaScript MVC frameworks
can understand the benefits and problems related to these
frameworks, by reviewing the case of A NGULAR JS. Third,
MVC framework builders can use our results to design more
powerful and usable frameworks.
This paper reports the results of a survey with 460 developers, when we collected their perceptions about A NGU LAR JS. We reveal the relevant features of the framework,
e.g., custom components, dependency injection, and twoway data binding. We also shed light on the most frequent
problems faced by A NGULAR JS developers, e.g., due to the
complexity of the API to declare directives.
The remainder of the paper is organized as follows. Section 2 introduces A NGULAR JS. Section 3 documents the
survey design and Section 4 presents the survey results.
Threats to validity are presented in Section 5. Section 6 discusses related work and Section 7 concludes.
2.
AngularJS in a Nutshell
In this section, we briefly describe the key components of
A NGULAR JS. A basic understanding of these components
is important to interpret our survey results.
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored.
Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists, contact
the Owner/Author(s). Request permissions from [email protected] or Publications Dept., ACM, Inc., fax +1 (212)
869-0481.
Modules: An A NGULAR JS application is a set of modules,
which act as containers to the different parts of the application. Modules can also depend on other modules.
PLATEAU ’16, November 01 2016, Amsterdam, Netherlands
Copyright © 2016 held by owner/author(s). Publication rights licensed to ACM.
ACM 978-1-4503-4638-2/16/11. . . $15.00
DOI: http://dx.doi.org/10.1145/3001878.3001881
Services: A NGULAR JS services are objects that encapsulate
code related with a specific concern. They are instantiated
1
2
3
4
5
6
7
8
9
10
< html lang = " en " ng-app = " todomvc " >
...
< header id = " header " >
<h1 > todos </ h1 >
< form ng-submit = " addTodo () " >
< input placeholder = " What needs to be done ? "
ng-model = " newTodo " / >
</ form >
</ header >
...
</ html >
Listing 1. Template sample
only once by factories or constructor functions. The created
singleton object is shared by the components that depend
on it (e.g., controllers, directives, filters, and other services).
Typically, A NGULAR JS services are stateless objects with
a set of methods that deal with specific concerns, such
as server requests, manipulation of arrays, asynchronous
operations, etc. A NGULAR JS also provides built-in services
to deal with common concerns in Web applications, such as
$http, $filter, and $timeout.
Templates: In Web applications, HTML documents are
parsed to generate the DOM (Document Object Model),
which is the data structure that models the final document
presented to users. A NGULAR JS supports DOM-based templates, which are written in HTML and contain proprietary
elements and attributes to render the dynamic interface of
web applications. Listing 1 shows a sample template. It
includes the definition of the document (html element in
line 1) and the markup for the main input of the application,
which is represented by the form element (line 5).
Directives: Directives are specific HTML markers used
in templates to define the UI behavior. In Listing 1, the
ng-app attribute (line 1) is an A NGULAR JS directive that
specifies the root node of the application. When directives
are executed by A NGULAR JS they can modify the DOM
structure and also register event handlers. During the
compilation process, A NGULAR JS traverses the DOM and
searches for all directives. The directives are executed in
order of priority generating the final DOM presented to
users. Directives can (1) use the same scope of the parent
element; (2) create a scope that inherits from the scope of
the parent element; or (3) create a completely new scope. It
is also possible to create custom directives.
Expressions: A NGULAR JS expressions are delimited by
double curly brackets ({{expression}}) or are the values
of some directive attributes. Literals (e.g., arrays ([]),
objects ({})), operators, and variables are examples of
elements that can be used in expressions. Expressions are
evaluated using a context represented by an object, called
scope. Variables and functions used in expressions must be
defined in the scope object. During the template compilation, when the ng-app directive is parsed, A NGULAR JS
creates an object representing the main application scope,
which is referenced as the $rootScope. Since directives
can define different scopes from $rootScope, expressions
from different parts of the template may be evaluated under
different scopes. When the value of an expression changes,
A NGULAR JS updates the view accordingly.
Controllers: A NGULAR JS controllers are used to initialize
the state of an application and provide an interface to update
it. Controllers are used with the ngController directive.
When this directive is used in a template, it receives the
name of a controller and the scope created by the directive
is passed as a parameter to the specified controller. The
controller must populate the scope object with properties
and methods that are used when evaluating expressions.
Digest Cycle: A NGULAR JS constantly maintains in sync the
state of the application with the view presented to the final
user. The framework provides this synchronization by comparing the current value of all variables in the scope referenced by the template expressions with their previous values. When a change is detected, the framework adequately
updates the DOM, in a process known as digest cycle.
3.
Survey Design
First, we performed a Mapping Study (Section 3.1) to gather
information about A NGULAR JS. We relied on the results
of this first study to construct the survey. The survey participants were selected among Stack Overflow users (Section 3.2).
3.1
Mapping Study
We use a mapping study to gather information about the use
of A NGULAR JS. A mapping study is more flexible than a
systematic review and it is recommended for studying emergent fields or technologies (Wohlin et al. 2012). For example, information about A NGULAR JS is primarily found in
blogs, forums, and Q&A sites. Therefore, using the Google
search engine, we initially focused on finding documents reporting the benefits and disadvantages of A NGULAR JS. To
this purpose, we used search queries such as “the best features of AngularJS” or “the bad parts of AngularJS”. We also
used two other strategies to reach more sources of information. First, we recursively accessed the references in the sites
already reviewed (a practice called “snowballing”). Second,
we performed additional search queries for frequently mentioned topics. As an example, we searched for “transclusion directives”, due to the frequency of references to this
topic as a complex A NGULAR JS concept. During the revision of blogs, we only considered well-written posts, by
authors with experience in software development. A complete list of the posts we consider is available in a companion
website.1 Finally, these documents were analyzed and classified by identifying topic trends, which were used for the
survey design.
1 https:// github.com/ aserg-ufmg/ angularjs-survey
Figure 1. Selecting the survey participants
3.2
Survey Construction and Participants
As a result of the mapping study, we built a 15-minute survey with 25 questions, organized in seven sections: (1) background of the participants, (2) key characteristics of A NGU LAR JS, (3) problems with A NGULAR JS templates, (4) debugging and testing, (5) development practices, (6) complex
concepts and features, and (7) A NGULAR JS 2.0. To avoid
unreliable responses, we asked developers to skip the questions they did not feel confident to answer. All the scales
used in the survey have an even number of points, to force
participants to make a choice. In multiple choice questions,
when the respondents could provide a response not included
in the set of answers, we included an Other option.
To find participants, we used the Stack Exchange API2
to search for A NGULAR JS developers in the Stack Overflow
community. As illustrated in Figure 1, we retrieved all existing tags with the substring “angularjs”. We retrieved 120,696
questions and 153,689 answers containing these tags. Next,
we extracted data about three types of users: users who own
a question (QO); users who own an answer (AO); or users
who edited a question or answer (E). In this way, we collected a total of 74,005 users. To find their e-mail address,
we matched each user nickname at StackOverflow with an
equivalent nickname at GitHub. In this way, we obtained a
total of 20,695 matched users (28%). However, only 8,084
users (10.9%) had public contact information at GitHub. In
a last step, 7,157 users were marked as potential participants in the survey, because they have valid e-mail address
and at least one operation with the collected answers that is
not a simple edition. These users were ranked considering
the number of A NGULAR JS-related questions and answers
at Stack Overflow. Each user received the following score:
S = QO +3∗AO, where QO is number of questions owned
by the user and AO is the number of answers he owns. We
gave an additional weight to answers since users who pro2 https:// api.stackexchange.com
vide answers tend to have more experience than those who
are looking for them.
We randomly selected 30 users from the middle of the
rank, with scores between 9 and 24, to run a pilot survey.
Their feedback helped us to correct typographical errors and
ambiguities in some questions. However, the most important
change was in the structure of some questions. In the initial survey version, we used ranking questions, when survey
respondents have to rank a list of items in order of importance. After this pilot study, we decided to change to rating
questions (when respondents just have to rank each item in
a scale ranging from 1 to 4) because some participants complained about the time to answer ranking questions. Due to
these changes in the survey, the responses obtained during
the pilot phase were discarded.
The final version of the survey—which is also available
in our companion website—was first sent to a group of 60
users. This time, we received complete responses and an improved response rate. Therefore, we extended the invitation
to the rest of the users by daily emailing groups of nearly
700 users from the top of the ranking. At one point, we decided to stop due to replies from developers saying that they
had a limited experience with A NGULAR JS. In total, we sent
the survey to 3,060 users with score between 3 and 831. We
received 460 complete responses, representing a response
rate of 15%. The survey was open during approximately one
month (from early November to early December 2015).
4.
Survey Results
4.1
Background
Figure 2 reveals the participants background. The majority
of the respondents (97.6%) have at least one year of experience in JavaScript (Figure 2(a)), and 74.8% have one to
three years of experience with A NGULAR JS (Figure 2(b)).
For 37.4% of the developers, the largest application implemented with A NGULAR JS has more than 10 KLOC (Fig-
(a) JavaScript experience
(b) AngularJS experience
(c) Largest application
Custom interface components
452 (98%)
Dependency injection
452 (98%)
Two−way data binding
451 (98%)
Behavior specified in HTML
440 (96%)
Predefined components
439 (95%)
Templates directly in HTML
447 (97%)
One solution to many problems
434 (94%)
Modeling with POJOs
443 (96%)
Ease of tests
428 (93%)
Google support
445 (97%)
40
20
0
1 (not valuable)
20
40
Percentage
2
3
60
80
Number of answers
Figure 2. Respondents’ background
100
4 (very valuable)
Figure 3. Key features and characteristics of AngularJS
ure 2(c)). Therefore, we can conclude that at least the participants are not novice A NGULAR JS developers.
6. Ease of writing tests: A NGULAR JS provides the ngMock
module to simulate logging operations, HTTP reqs, etc.
4.2
7. Two-way data binding: A NGULAR JS provides synchronization between data in the view and in the model.
Key Characteristics and Features of AngularJS
We asked developers about the the following features:
1. Pre-defined components for code organization: A NGU LAR JS has different components to modularize code,
which may help in separation of concerns.
2. Dependency injection: This design pattern is used by
A NGULAR JS to manage dependencies between components, to reduce coupling and increase testability.
3. Use of POJOs in model components: In A NGULAR JS,
models are implemented using Plain Old JavaScript Objects (POJOs). There is no need to extend proprietary
classes, for example, to provide accessor methods.
4. Templates in HTML: A NGULAR JS uses DOM-based
templates to simplify data binding operations, event mapping, and updating of large interface components.
5. Support to custom components: Custom directives can be
used as a DSL to define reusable UI components.
8. Use of HTML to declare UI behavior: The UI, including
its behavior, is defined in standard HTML documents.
9. One solution to manage many problems: A NGULAR JS
is an “opinionated framework”, meaning that its design
handles common decisions related with Web apps.
10. Supported by Google: This support may guarantee the
evolution and constant maintenance of the project.
Figure 3 reports the value given by developers to these
characteristics and features, ranging from not valuable
(score 1) to very valuable (score 4). At the right of the charts,
we include the number of answers of each item. Each individual item was rated by at least 93% of the participants.
The top-3 features with more positive scores are the support to custom components, the use of dependency injection,
and the support for two-way data binding. The characteristic
439 (95%)
Dump/Debug problems
449 (98%)
Modularity and SOC loss
434 (94%)
Low readability of HTML
448 (97%)
80
70
60
50
40
1 (never)
30 20 10
Percentage
2
3
0
10
20
30
Number of answers
Silent failures
40
4 (very often)
Figure 4. How often these items represent a real problem caused by using code in A NGULAR JS templates?
with the lowest number of positive answers is Google support, with 45.2% of the respondents seeing it as having no
value (score 1) or limited value (score 2).
4.3
Code in HTML Templates
In the mapping study, we detected four possible problems
related to placing code in A NGULAR JS templates:
1. Silent failures: In A NGULAR JS, when undefined functions or objects are used in templates no exceptions are
raised. As a consequence, applications might fail silently.
2. Code hard to debug: Since the code used in A NGULAR JS
templates is not pure JavaScript, it is not possible, for
example, to define break points.
3. Low readability: A NGULAR JS code might be spread all
over the HTML document, hindering readability.
4. Modularity and Separation of Concerns: The use of large
amount of JavaScript code in HTML is often seen as bad
smell (Nguyen et al. 2012), which might hinder separation of concerns.
We asked developers whether these issues are real problems in their daily development. In this case, the score 1
means the issue was never a problem and a score 4 means
it is a very frequent problem. As shown in Figure 4, none
of the issues have a major detrimental impact, according to
the respondents; they have at least 60% of the answers in the
low part of the scale (scores 1 or 2).
In a separate question, we asked if the developers had at
least once used large amounts of logic in HTML templates
and 26.3% of them answered positively. Since this is not a
recommended practice, we asked them to indicate possible
reasons for their decision The two most voted reasons are
A NGULAR JS design (54.8%) and the lack of experience in
A NGULAR JS (43.5%). Lack of experience in Web architecture and in JavaScript were also voted (20.9% and 4.3%, respectively). We also gave the respondents the possibility to
indicate other reasons, which were provided by 31 developers. These reasons include, for example, tight deadlines, special cases, easiness or laziness, prototyping purposes, etc.
4.4
Testing
First, we asked the participants to rate the frequency they
make use of mocking (provided by the ngMock module)
when testing their systems. From 441 answers, 72.8% indicated they never or rarely use this module (scores 1 or
2). Possible reasons for this result include limited usefulness of ngMock features, unfamiliarity with the module, and
few developers putting testing into practice. We also asked
the participants to rank how complex is testing A NGULAR JS
components, from very easy to very difficult. As presented in
Figure 5, services, filters, controllers, and providers received
the higher number of answers with scores 1 and 2. The reason is that these components are very common and usually
do not include complex code or code that deals with complex APIs. For example, most code in filters only make string
transformations. By contrast, the two components more difficult to test are transclusion directives and directives with
external templates. Probably, developers find these directives more difficult to test because they demand a deeper understanding of A NGULAR JS concepts. For example, when
creating directives, there are different types of transclusion,
different types of scopes, and different ways to interact with
the DOM API.
4.5
Complex Concepts and Features
We asked the participants to evaluate several characteristics of A NGULAR JS, which were originally identified in the
337 (73%)
Directives with external templates
350 (76%)
Simple HTML directives
371 (81%)
Providers
361 (78%)
Controllers
386 (84%)
Filters
369 (80%)
Services
381 (83%)
80
60
40
1 (very easy)
20
0
Percentage
2
3
20
40
Number of answers
Transclusion directives
60
4 (very difficult)
Performance degradation details
413 (90%)
Transclusion directives API
397 (86%)
Directive scopes
424 (92%)
Compile, link and controller functions
424 (92%)
Correct use of $scope.$apply
425 (92%)
Dependency registration methods
419 (91%)
Scopes prototypal inheritance
414 (90%)
External components integration
416 (90%)
Dependency injection methods
419 (91%)
80
60
40
1 (very simple)
20
0
20
Percentage
2
3
40
60
Number of answers
Figure 5. How difficult is it to test these A NGULAR JS components?
80
4 (very complex)
Figure 6. How complex are the the following A NGULAR JS aspects and features?
mapping study as complex concepts. The following characteristics were proposed: (1) use of different scopes in directives; (2) use of prototypes to simulate scope inheritance; (3)
different types of entities to register dependencies; (4) compile, link, and controller functions (necessary to implement
DOM-related directives); (5) transclusion directives; (6) correct use of $scope.$apply to manually trigger the digest
cycle; (7) tackling of performance degradation details; (8)
integration with external components and plug-ins; and (9)
correct usage of the syntax to inject dependencies (in order
to avoid unexpected results when minifying the code).
Figure 6 summarizes the developers’ classification from
very simple concept (score 1) to very complex one (score 4).
Tackling all the details that can lead to performance degradation was rated by 76% of the participants as a complex task.
The next three items in terms of complexity are transclusion
directives, the different scopes that can be used when building directives, and the correct use of the functions compile,
link, and controller. All these items are somehow re-
Integration with ES6 modules
New API for UI components definition
Component router
New mechanism for change detection
New mobile features
Server rendering
New template syntax
New API for dependency injection
No more $scope.$apply
Execution time measurement service
Other
0
10
20
30
40
50
60
70
Percent of developers
Figure 7. Most expected A NGULAR JS 2.0 features
lated with the implementation of directives. The remaining
items were rated as more simple than complex, mainly the
integration with external components and the use of the correct syntax to inject dependencies.
In a separate question, we asked the developers about the
frequency they create directives. As a result, 83.5% of the
participants answered they often or very often create their
own directives. Despite this fact, many tasks and features
related to the implementation of directives are usually considered complex by the survey participants.
4.6
• The two problems that arise more frequently regarding
the use of code in A NGULAR JS templates are the emergence of silent failures and the difficulty to dump/debug
the variables referenced in the HTML. New debug tools
and techniques can then be developed to alleviate these
problems faced by developers.
AngularJS 2.0
To reveal the most expected features or improvements in
A NGULAR JS 2.0, we selected the following features that
appeared in the mapping study: (1) the new API to define the main UI building blocks; (2) Zone.js (no more
$scope.$apply()); (3) server rendering; (4) the new logging service, called diary.js; (5) new mobile features (e.g.
support for touch event gestures); (6) the new template syntax; (7) the new change detection mechanism; (8) the new
API for injecting dependencies; (9) integration with ECMAScript 6 (ES6) modules; (10) component router, which
allows asynchronous loading.
Figure 7 indicates that the most expected feature is the
integration with ES6 modules (68.3%). The second most
expected feature is the new API to create UI components,
which is expected because the current API is perceived by
most developers as difficult to use and understand (see Section 4.5). The third most expected feature is the component
router, which allows the dynamic loading of UI components,
followed by the new change detection mechanism.
4.7
ponents by means of custom directives, the use of dependency injection, and the ease to set up two-way data binding. Therefore, JavaScript MVC framework builders can
embrace these characteristics and improve them by offering a more simple interface to build reusable UI components (without exposing final users to internal concepts
and decisions) and by using better mechanisms to detect
changes in the model (i.e., mechanisms that reduce the
number of details to be considered).
Key Findings and Implications
The main findings of our survey are as follows:
• Three characteristics of A NGULAR JS excel by the value
that developers give to them: the ability to create UI com-
• The two main reasons for placing large amounts of logic
in templates are the lack of experience in A NGULAR JS
and the design of the framework. This shows the importance of correctly training developers before they start to
use A NGULAR JS on complex applications. It also reveals
an opportunity for framework builders to investigate new
framework designs.
• The components that are more difficult to test in A NGU LAR JS are directives, mainly the ones using transclusion;
the remaining components (i.e., controllers, services, filters, etc.) are mostly considered easy to test.
5.
Threats to Validity
The first threat to validity relates to the execution of the mapping study. Due to the large amount of information on the
Internet about A NGULAR JS, it is possible that literature addressing more specific topics about the framework or presenting different points of view was not included.
Regarding the construction of the questionnaire, the main
threat is the insertion of ambiguous and leading questions.
We made our best to avoid this problem by constantly reviewing and improving the proposed questions. Additionally, we ran a pilot survey to identify and correct this type of
questions. In some questions, we also gave the participants
the opportunity to respond with an answer different from the
proposed ones by adding an “Other” option. Furthermore,
with the exception of the background questions, no question
was mandatory. Therefore, participants were not forced to
provide answers when they did not want or when they were
not sure about their answers.
There are also two threats related to the method used
to retrieve the participants’ e-mails. The first one is related
with the match between the Stack Overflow profile and the
GitHub profile of the participants. It is possible that a Stack
Overflow user has been matched with a homonym user in
GitHub (i.e., users who have exactly the same login name,
but who are different people). Additionally, it is possible
that the heuristic used to assess and rank the expertise of
the selected developers does not reflect the reality.
There are also some aspects that limit the generalization
of our results (external validity). First, the sample for the survey was selected only from the Stack Overflow community.
Therefore, it is possible that the findings in this study do not
apply to a different population. Moreover, constant and rapid
changes in Web development environments, including new
technologies and new versions of A NGULAR JS, can lead to
different results if the study is repeated in the future.
Finally, we have to mention threats related to human
behavior. For instance, we can mention the ordering of the
questions since one may provide context for the next one.
Another threat is the attitude of the participants towards the
topic of research. Their responses can introduce bias to make
A NGULAR JS appear in a positive or negative light. We can
mention the case of one participant who declared that he is a
contributor of the A NGULAR JS project.
Ocariza et al. (2015) proposed two types of inconsistencies that can be found in A NGULAR JS applications: (i) when
identifiers used in one layer are undefined in the lower layer;
and (ii) when values assigned to a variable, or returned by
a function, do not match their type in the view. According to the authors, both inconsistencies are not easily caught
during development and might cause bugs. However, in our
survey, only 39% of the respondents considered that silent
failures, corresponding to identifier inconsistency, are real
problems in their daily development. Moreover, 84.5% of
the respondents considered two-way data binding, which relates to type consistency, as a valuable feature. Both results
shed light over real problems developers face when they
use A NGULAR JS.
7.
Conclusion
This paper reported an empirical study about different aspects of A NGULAR JS based on opinions and experiences of
developers. Our main contributions include the identification
of the most appreciated features of A NGULAR JS (e.g., custom interface components, dependency injection, and twoway data binding) and the most problematic aspects of the
framework (e.g., performance degradation and implementation of directives). Future work includes an analysis of the
results using statistical tests. Interviews with A NGULAR JS
developers can also contribute to strengthen our findings.
Acknowledgments
Our research is supported by FAPEMIG and CNPq. We also
deeply thank the 460 developers who answered the survey.
References
B. Dobing and J. Parsons. How UML is used. Commununications
of the ACM, 49(5):109–113, 2006.
H. M. Kienle. It’s about time to take JavaScript (more) seriously.
IEEE Software, 27(3):60–62, 2010.
6.
Related Work
Some works have been focusing on practitioners’ use of
known technologies. Dobing and Parsons (2006) conducted
the first survey on how UML diagrams are used by practitioners and their clients. The authors gathered 182 responses
from analysts with average experience of 4.7 years in UML.
Class diagrams are being used regularly by the majority of
the respondents, followed by Sequence and Use Case diagrams. Some of the reasons why a UML diagram is not used
vary from “not well understood by analysts” to “insufficient
value to justify the cost”. In another study, Petre (2013) reported two years of interviews with practitioners. Most of
them (35 out of 50) currently do not use UML at all, due to
notation overhead, lack of context, etc. Both work highlight
the complexity of UML. They also suggest that more tooling
is needed for both newcomers and professionals in order to
use the language more effectively.
H. V. Nguyen, H. A. Nguyen, T. T. Nguyen, A. T. Nguyen, and T. N.
Nguyen. Detection of embedded code smells in dynamic web
applications. In 27th International Conference on Automated
Software Engineering (ASE), pages 282–285, 2012.
F. Ocariza, K. Pattabiraman, and A. Mesbah. Detecting inconsistencies in JavaScript MVC applications. In 37th International
Conference on Software Engineering (ICSE), pages 325–335,
2015.
M. Petre. UML in practice. In 35th International Conference on
Software Engineering (ICSE), pages 722–731, 2013.
L. Silva, M. Ramos, M. T. Valente, N. Anquetil, and A. Bergel.
Does Javascript software embrace classes? In 22nd International
Conference on Software Analysis, Evolution and Reengineering
(SANER), pages 73–82, 2015.
C. Wohlin, P. Runeson, M. Höst, M. C. Ohlsson, B. Regnell, and
A. Wesslén. Experimentation in software engineering. Springer,
2012.
| 2cs.AI
|
ActionFlowNet: Learning Motion Representation for Action Recognition
Joe Yue-Hei Ng1
Jonghyun Choi2
Jan Neumann3
Larry S. Davis1
1
2
3
University of Maryland, College Park
Allen Institute for AI
Comcast Labs, DC
arXiv:1612.03052v3 [cs.CV] 16 Feb 2018
{yhng,lsd}@umiacs.umd.edu
[email protected]
Abstract
We present a data-efficient representation learning approach to learn video representation with small amount of
labeled data. We propose a multitask learning model ActionFlowNet to train a single stream network directly from
raw pixels to jointly estimate optical flow while recognizing
actions with convolutional neural networks, capturing both
appearance and motion in a single model. Our model effectively learns video representation from motion information
on unlabeled videos. Our model significantly improves action recognition accuracy by a large margin (23.6%) compared to state-of-the-art CNN-based unsupervised representation learning methods trained without external large
scale data and additional optical flow input. Without pretraining on large external labeled datasets, our model, by
well exploiting the motion information, achieves competitive recognition accuracy to the models trained with large
labeled datasets such as ImageNet and Sport-1M.
1. Introduction
Convolutional Neural Networks have demonstrated great
success to multiple visual recognition tasks. With the help
of large amount of annotated data like ImageNet, the network learns multiple layers of complex visual features directly from raw pixels in an end-to-end manner without relying on hand-crafted features. Unlike image labeling, manual video annotation often involves frame-by-frame inspection and temporal trimming of videos that are expensive and
time consuming. This prohibits the technique to be applied
to other problem domains like medical imaging where data
collection is difficult.
We focus on effectively learning video motion representation for action recognition without large amount of external annotated video data. Following previous work [17,
28, 6] that leverages spatio-temporal structure in videos for
unsupervised or self-supervised representation learning, we
are interested in learning video representation from motion
information encoded in videos in addition to semantic la-
bels.
jan [email protected]
CxTxHxW
3x16x224x224
128x16x56x56
512x4x14x14
…
64x16x112x112
Optical Flow
256x8x28x28
64x16x56x56
128x8x28x28
256x4x14x14
512x2x7x7
…
concat
3D-conv
Action Label
3D-resnet
fc
Figure 1: ActionFlowNet for jointly estaimting optical flow
and recognizing actions. Orange and blue blocks represent
ResNet modules, where blue blocks represents strided convolution. Channel dimension is not shown in the figure.
Learning motion representation on videos from raw pixels is challenging. With large scale datasets such as Sports1M [10] and Kinetics [11], one could train a high capacity classifier to learn complex motion signatures for action recognition by extending image based CNN architectures with 3D convolutions for video action recognition [10, 26, 2]. However, while classification loss is an excellent generic appearance learner for image classification,
it is not necessarily the most effective supervision for learning motion features for action recognition. As shown in [2],
even with large amount of labeled video data, the model still
benefits from additional optical flow input stream. This suggests that the model is ineffective in learning motion representation for action recognition from video frames, and thus
alternative approach should be explored for learning video
representation.
Two-stream convolutional neural networks, which separately learn appearance and motion by two convolutional
networks on static images and optical flow respectively,
show impressive results on action recognition [22]. The
separation, however, fails to learn the interaction between
the motion and the appearance of objects, and introduces
additional complexity of computing the flow to the classification pipeline. In addition, human visual system does
not take optical flow as front end input signals but infer the
motion from raw intensities internally. Therefore, we focus
to learn both motion features and appearance directly from
raw pixels without hand-crafted flow input.
Encouraged by the success on estimating optical flow
with convolutional neural networks [7], we train a single stream feed-forward convolutional neural network ActionFlowNet - for jointly recognizing actions and estimating optical flow. Specifically, we formulate the learning
problem as multitask learning, which enables the network to
learn both appearance and motion in a single network from
raw pixels. The proposed architecture is illustrated in Figure 1. With the auxiliary task of optical flow learning, the
network effectively learns useful representations from motion modeling without a large amount of human annotation.
Based on the already learned motion modeling, the model
then only requires action annotations as supervision to learn
action class specific details, which results in requiring less
annotation to perform well for action recognition.
Our experiments and analyses show that our model successfully learns motion features for action recognition and
provide insights on how the learned optical flow quality affects action classification. We demonstrate the effectiveness
of our learned motion representation on two standard action
recognition benchmarks - UCF101 and HMDB51. Without
providing external training data or fine-tuning from already
well-trained models with millions of samples, we show
that jointly learning action and optical flow significantly
boosts action recognition accuracy compared to state-ofthe-art representation learning methods trained without external labeled data. Remarkably, our model outperforms
the models trained with large datasets Sports-1M pretrained
C3D by 1.6% on UCF101 dataset, showing the importance
of feature learning algorithms.
2. Related Work
Over the past few years, action recognition accuracy
has been greatly improved by learned features and various
learning models utilizing deep networks. Two-stream network architecture was proposed to recognize action using
both appearance and motions separately [22]. A number
of follow up methods have been proposed based on twostream networks that further improved action recognition
accuracies [5, 31, 30, 4, 18]. Our work is motivated by their
success in incorporating optical flow for action recognition,
but we focus on learning from raw pixels instead of relying
on hand-crafted representations.
Optical flow encodes motion between frames and is
highly related to action recognition. Our model is motivated by the success of FlowNet [7] and 3D convolutions
for optical flow estimation in videos [27], but emphasizes
on improving action recognition.
Pre-training the network with a large dataset helps to
learn appearance signatures for action recognition. Karpathy et al. proposed a “Slow Fusion” network for large scale
video classification [10]. Tran et al. trained a 3D convo-
lutional neural network (C3D) with a large amount of data
and showed the learned features are generic for different
tasks [26]. Recently, Carreira and Zisserman trained I3D
models [2] on the Kinetics dataset [11] and achieved strong
action recognition performance. In contrast, since training
networks on such large scale datasets is extremely computationally expensive, we focus on learning from small
amounts of labeled data. With only small amount of labeled data, we show that our model performs competitive
to models trained with large datasets.
Leveraging videos as a source for unsupervised learning has been suggested to learn video representations without large labeled data. Different surrogate tasks have been
proposed to learn visual representations from videos without any labels. Wang et al. trained a network to learn visual similarity for patches obtained from visual tracking in
videos [32]. Misra et al. trained a network to differentiate
the temporal order of different frames from a video [17].
Jacob et al. learned apperance features by predicting the future trajectories in videos [29]. Fernando et al. proposed
Odd-One-Out networks (O3N) to identify video sequences
that are out of order for self-supervised learning [6]. Our
work, similarly, uses video as an additional source for learning visual representation. However, in contrast to previous
work which focused on learning visual representations for
a single image, we learn motion representations for videos
which models more than a single frame. Vondrick et al.
used a Generatie Adversarial Network to learn a generative
model for video [28]. We focus on learning motion representations but not video generation.
Independent to our work, Diba et al. trained a two stream
network with flow estimation [3]. They based their network
on C3D with a two-stream architecture. Our work employs
a single stream network to learn both appearance and motion. While we both estimate motion and recognize actions
in the same model, we focus on learning motion representations without pretraining on large labeled datasets and provide more analysis to learn flow representations for action
recognition.
3. Approach
We propose a single end-to-end model to learn both motions and action classes simultaneously. Our primary goal
is to improve action classification accuracy with the help of
motion information; we use optical flow as a motion signature. Unlike previous methods that utilize externally computed optical flow as the input to their models, we only use
the video frames for input and simultaneously learn the flow
and class labels.
3.1. Multi-frame Optical Flow with 3D-ResNet
Fischer et al. proposed FlowNet [7] that is based on convolutional neural networks to estimate high quality optical
flow. Tran et al. proposed to use 3D convolution and deconvolution layers to learn multi-frame optical flow from
videos [27]. In addition, He et al. introduced residual networks (ResNet) to train a deeper convolutional neural network model by adding shortcut connections [8].
In addition to the benefit of easy training, ResNet is fully
convolutional, so is easily applied to pixel-wise prediction
of optical flow, unlike many architectures with fully connected layers including AlexNet [13] and VGG-16 [23]. In
contrast to other classification architectures like AlexNet
and VGG-16, which contains multiple max pooling layers
that may harm optical flow estimation, the ResNet architecture only contains one pooling layer right after conv1. We
believe the reduced number of pooling layers makes ResNet
more suitable for optical flow estimation where spatial details need to be preserved. Specifically, we use an 18 layers
ResNet, which is computationally efficient with good classification performance [8].
Taking advantage of ResNet for flow estimation, we extend ResNet-18 to 3D-ResNet-18 for multi-frame optical
flow estimation by replacing all k × k 2D convolutional
kernels with extra temporal dimension k × k × 3, inspired
by [27]. The deconvolution layers in the decoder are extended similarly. Skip connections from encoder to decoder
are retained as in [7] to obtain higher resolution information in the decoder. Unlike [7], we only use the loss on
the highest resolution to avoid downsampling in the temporal dimension. We do not apply temporal max pooling
suggested in [26, 27], but use only strided convolutions to
preserve temporal details. After the third residual block, the
temporal resolution is reduced by half when the spatial resolution is reduced.
Future Prediction. In addition to computing the optical flow between the T input frames, we train the model to
predict the optical flow on the last frame, which is the optical flow between the T th and (T + 1)st frames. There are
two benefits of training the model to predict the optical flow
of the last frame: 1) It is practically easier to implement a
model with the same input and output sizes, since the output sizes of deconvolution layers are usually multiples of
the inputs; and 2) Semantic reasoning is required for the
model to extrapolate the future optical flow given the previous frames. This possibly trains the model to learn better
motion features for action recognition, as also suggested by
previous work [29], which learned appearance feature by
predicting the future.
Following [7], the network is optimized over the endpoint error (EPE), which is the sum of L2 distance between
the ground truth optical flow and the obtained flow over all
pixels. The total loss for the multiple frame optical flow
model is the EPE of T output optical flow frames:
T X
X
koj,t,p − o[
(1)
j,t,p k2 ,
t=1
p
where oj,t,p is 2-dimensional optical flow vector of the tth
and the (t + 1)st frame in the j th video at pixel p.
Note that the T th optical flow frame oj,t is the future
optical flow for the T th and (T + 1)st input frames, where
the (T + 1)st frame is not given to the model.
3.2. ActionFlowNet
Knowledge Transfer by Finetuning. Finetuning a pretrained network is a common practice to transfer knowledge from different datasets and tasks. Unlike previous
work, where knowledge transfer has been accomplished between very similar tasks (image classification and detection or semantic segmentation), knowledge transfer in our
model is challenging since the goals of pixel-wise optical
flow and action classification are not obviously compatible.
We transfer the learned motion by initializing the classification network using a network trained for optical flow estimation. Since the network was trained to predict optical
flow, it should encode motion information in intermediate
levels which support action classification. However, finetuning a pretrained network is known to have the problem of
catastrophic forgetting. Specifically, when training the network for action recognition, the originally initialized flow
information could be destroyed when the network adapts
the appearance information. We prevent catastrophic forgetting by using the multitask learning framework.
ActionFlowNet. To force the model to learn motion features while training for action recognition, we propose
a multitask model ActionFlowNet, which simultaneously
learns to estimate optical flow, together with predicting the
future optical flow of the last frame, and action classification to avoid catastrophic forgetting. With optical flow as
supervision, the model can effectively learn motion features
while not relying on explicit optical flow computation.
In our implementation, we take 16 consecutive frames
as input to our model. In the last layer of the encoder,
global average pooling across the spatial-temporal feature
map, with size 512 × 2 × 7 × 7, is employed to obtain a
single 512 dimensional feature vector, followed by a linear
softmax classifier for action recognition. The architecture
is illustrated in Figure 1. The multitask loss is given as follows:
MT-Lossj = −1(yj = ybj ) log p(ybj ) +
{z
}
|
Classification Loss
λ
T X
X
t=1
|
koj,t,p − o[
j,t,p k2 , (2)
p
{z
Flow Loss
}
where 1(·) is a indicator function, yj and ybj are the
groundtruth and predicted action labels respectively of the
j th video. λ is a hyper-parameter balancing the classification loss and the flow loss, where optical flow estimation
can be seen as a regularizer for the model to learn motion
feature for classification.
Although previous work on multitask learning [16] suggests that sharing parameters of two different tasks may hurt
performance, this architecture performs well since optical
flow is known empirically to improve video action recognition significantly. In addition, our architecture contains
multiple skip connections from lower convolutional layers
to decoder. This allows higher layers in the encoder to focus on learning more abstract and high level features, without constraining them to remembering all spatial details for
predicting optical flow, which is beneficial for action recognition. This idea is central to Ladder Networks [20] which
introduced lateral connections to learn denoising functions
and significantly improved classification performance.
It is worth noting that this is a very general architecture
and requires minimal architectural engineering. Thus, it can
be trivially extended to learn more tasks jointly to adapt
knowledge from different domains.
ActionFlowNet Inference. During inference for action
classification, optical flow estimation is not required since
the motion information is already learned in the encoder.
Therefore, the decoder can be removed and only the forward pass of the encoder and the classifier are computed.
If the same backbone architecture is used, our model runs
at the same speed as a single-stream RGB network without extra computational overhead. Since the optical flow
estimation and flow-stream CNN are not needed, it is more
efficient than two-stream counterparts.
3.3. Two-Frame Based Models
In this section, we propose various models that take two
consecutive input frames. Experimenting with two-frame
models has three benefits. First, when there are multiple frames in the input, it is difficult to determine whether
the performance improvement comes from motion modeling or aggregating long term appearance information. Thus
for better analysis, it is desirable to use the two frame input. Second, training two-frame models is computationally
much more efficient than multi-frame models which take N
video frames and output N − 1 optical flow images. Third,
we can measure the effectiveness of external large scale optical flow datasets, such as the FlyingChairs dataset [7],
which provide ground-truth flow on only two consecutive
frames, for action recognition.
Learning Optical Flow with ResNet. Similarly, we use
ResNet-18 as our backbone architecture and learn optical
flow. Like FlowNet-S [7], we concatenate two consecutive
frames to produce a 6(ch) × 224(w) × 224(h) input for our
two frames model. At the decoder, there are four outputs
with different resolutions. The total optical flow loss is the
weighted sum of end-point error at multiple resolutions per
the following equation:
4
X
αr
X
r=1
[
(r)
(r)
koj,t,p − oj,t,p k2 ,
(3)
p
(r)
where oj,t,p is the optical flow vector of the rth layer output
and αr is the weighting coefficient of the rth optical flow
output. We refer to this pre-trained optical flow estimation
network as FlowNet.
We first propose an architecture to classify actions on
top of the optical flow estimation network, which we call
the Stacked Model. Then, we present the two-frame version
of ActionFlowNet to classify the actions and estimate the
optical flow, which we call the ActionFlowNet-2F.
3.3.1
Stacked Model
A straightforward way to use the trained parameters from
FlowNet is to take the output of FlowNet and learn a CNN
on top of the output, as shown in Figure 2. This is reminiscence of the temporal stream in [22] which learns a CNN
on precomputed optical flow. If the learned optical flow has
high quality, it should give similar performance to learning
a network on optical flow.
6x224x224
64x56x56
128x28x28
256x14x14
512x7x7
Optical
Flow 64x56x56
128x28x28
256x14x14
Action
Labels
512x7x7
FC
Softmax
ResNet
Figure 2: Network structure of the ‘Stacked Model’.
Since the output of FlowNet has 4 times lower resolution
than the original image, we remove the first two layers of
the CNN (conv1 and pool1) and stack the network on top
of it. We also tried to upsample the flow to the original
resolution and use the original architecture including conv1
and pool1, but this produces slightly worse results and is
computationally more expensive.
The stacked model introduces about 2x number of parameters compared to the original ResNet, and is also 2x
more expensive for inference. It learns motion features by
explicitly including optical flow as an intermediate representation, but cannot model appearance and motion simultaneously, similar to learning a CNN on precomputed optical
flow.
3.3.2
ActionFlowNet-2F
The multitask ActionFlowNet-2F architecture, as illustrated
in Figure 3, is based on the two-frame FlowNet with additional classifier. Similar to ActionFlowNet, classification is
performed by average pooling the last convolutional layer
in the encoder followed by a linear classsifier.
56x56x64
128x28x28
6x224x224
64x56x56
Optical
Flow
256x14x14
128x28x28
256x14x14
512x7x7 R2
R3
R4
R1
Action
Label
ResNet
FC
Softmax
Figure 3: Network structure of the ActionFlowNet-2F
Just as with the stacked model, the loss function is defined for each frame. For the tth frame in the j th video the
loss is defined as a weighted sum of classification loss and
optical flow loss:
MT-Lossj,t = −1(yj = ybj ) log p(ybj ) +
|
{z
}
Classification Loss
λ
4
X
r=1
|
4. Experiments
αr
X
[
(r)
(r)
koj,t,p − oj,t,p k2 , (4)
p
{z
Flow Loss
}
4.1. Datasets
We use two publicly available datasets, UCF101 and
HMDB51, to evaluate action classification accuracy. The
UCF101 dataset contains 13,320 videos with 101 action
classes [24]. The HMDB51 contains 6,766 videos with 51
action categories [14]. As the number of training videos
in HMDB51 is small, we initialized our models trained on
UCF101 and fine-tuned for HMDB51 similar to [22]
The UCF101 and HMDB51 do not have groundtruth optical flow annotation. Similar to [27], we use EpicFlow [21]
as a psuedo-groundtruth optical flow to train the motion part
of the network.
To experiment models with better learned the motion
signature, we also use FlyingChairs dataset [7] as it has
groundtruth optical flow since it is a synthetic dataset.
The FlyingChairs dataset contains 22,872 image pairs and
ground truth flow from synthetically generated chairs on
real images. We use the Sintel dataset [1], which provides
dense groundtruth optical flow, to validate the quality of optical flow models.
4.2. Experimental Setup
Overfitting Prevention. We use different data augmentations on different datasets and tasks. On the FlyingChairs
dataset for optical flow estimation, we augment the data
using multi-scale cropping, horizontal flipping, translation
and rotation following [7]. On the UCF101 dataset for optical flow estimation, we use multi-scale cropping and horizontal flipping, but do not use translation and rotation in
order to maintain the original optical flow distribution in
the data. On UCF101 dataset for action recognition, we use
color jittering [25], multi-scale cropping and horizontal flipping. Dropout is applied to the output of the average pooling
layer before the linear classifier with probability 0.5.
Optimization and Evaluation. The models are trained using Adam [12] for 40,000 iterations with batch size 128 and
learning rate 1 × 10−4 . For evaluation, we sample 25 random video segments from a video and run a forward pass to
the network on the 10-crops (4 corners + center with their
horizontal reflections) and average the prediction scores.
4.3. Improving Action Recognition
We first evaluate the action recognition accuracy by
the various proposed two-frame models described in Section 3.3, and then the multi-frame models in Section 3.2,
on both UCF101 and HMDB51 datasets. All models take
RGB inputs only without external optical flow inputs. The
recognition accuracies are summarized in Table 1.
Method
UCF101
HMDB51
Two-frame Models
Scratch
FlowNet fine-tune
Stacked
ActionFlowNet-2F (UCF101)
ActionFlowNet-2F (FlCh+UCF101)
ImageNet pretrained ResNet-18
51.3
66.0
69.6
70.0
71.0
80.7
23.9
29.1
42.4
42.4
42.6
47.1
Multi-frame Models
Multi-frame FlowNet fine-tune
ActionFlowNet (UCF101)
Sports-1M pretrained C3D [26]
Kinetics pretrained I3D [2]
80.8
83.9
82.3
95.6
50.6
56.4
53.5
74.8
Table 1: Action recognition accuracies of our models on
UCF101 and HMDB51 datasets (split 1). FlCh denotes
FlyingChairs dataset. “ActionFlowNet-2F (UCF101)” denotes its FlowNet part is pretrained on UCF101, and
“ActionFlowNet-2F (FlCh+UCF101)” denotes its FlowNet
part is pretrained on FlyingChairs dataset. All ActionFlowNets are then learned on UCF101 dataset for action
and flow. For reference, we additionally show the results
trained with large scale datasets [27, 2], but it is not directly
comparable since our models are trained with significantly
less annotation.
Two-frame Models. ‘Scratch’ is a ResNet-18 model
that is trained from scratch (random initialization) using UCF101 without any extra supervision, which represents the baseline performance without motion modeling. ‘FlowNet fine-tune’ is a model that is pretrained from
UCF101 for optical flow only, and then fine-tuned with action classification, which captures motion information by
initialized FlowNet. ‘Stacked’ is a stacked classification
model on top of optical flow output depicted in Figure 2.
Its underlying FlowNet is trained with UCF101 and is fixed
to predict optical flow, so only the CNN classifier on top
is learned. ‘ActionFlowNet-2F’ is the multitask model depicted in Figure 3, which is trained for action recognition
and optical flow estimation to learn both motion and appearance. We trained two versions of ActionFlowNet-2F:
one with FlowNet pretrained on UCF101 and one on FlyingChairs dataset.
As shown in the table, all proposed models - ‘FlowNet
fine-tune’, ‘Stacked’ and ‘ActionFlowNet-2F’ significantly
outperform ‘Scratch’ . This implies that our models can
take advantage of the learned motion for action recognition,
which is difficult to learn implicitly from action labels.
Both the Stacked model and two ActionFlowNet-2Fs
outperform the finetuning models by a large margin (up to
5.0% in UCF101 and up to 13.5% in HMDB51). As all
models are pretrained from the high quality optical flow
model, the results show that the knowledge learned from
previous task is prone to be forgotten when learning new
task without multitask learning. With extra supervision
from optical flow estimation, multitask models regularize
the action recognition with the effort of learning the motion
features.
While the Stacked model performs similarly to
ActionFlowNet-2F when trained only on UCF101,
ActionFlowNet-2F is much more compact than the Stacked
model, containing only approximately half the number of
parameters of the Stacked model. When ActionFlowNet2F is first pretrained with FlyingChairs, which predicts
better quality optical flow in EPE, and finetuned with the
UCF101 dataset, it further improves accuracy by 1%. This
implies that our multitask model is capable of transferring
general motion information from other datasets to improve
recognition accuracy further.
Our ActionFlowNet-2F still performs inferior compared to ResNet pretrained on ImageNet, especially in
UCF101 (71.0% vs 80.7%) because of the rich background
context appearance in the dataset. When evaluated on
HMDB51, where the backgrounds are less discriminative,
our ActionFlowNet-2F is only slightly behind the ImageNet
pretrained model (42.6% vs 47.1%), indicating that our
model learns strong motion features for action recognition.
Multi-frame Models. We train 16-frame ActionFlowNet on UCF101. The results are shown in the lower
part of Table 1. By taking more frames per model, our
multi-frame models significantly improve two-frame models (83.9% vs 70.0%). This confirms previous work [10, 19]
that taking more input frames in the model is important.
Remarkably, without pretraining on large amounts of labeled data, our ActionFlowNet outperforms the ImageNet
pretrained single frame model and Sports-1M pretrained
C3D. Our ActionFlowNet gives 1.6% and 2.9% improvements over C3D on UCF101 and HMDB51 repsectively.
The recently published I3D models [2] achieved strong
performance by training on the newly released Kinetics
dataset [11] with large amount of clean and trimmed labeled video data and performing 3D convolutions on 64 input frames instead of 16 frames. Although the I3D model
achieved better results compared to previous work, their
RGB model could still benefit from optical flow inputs,
which indicates that even with large amount of labeled data
the I3D model does not learn motion features effectively.
It should be noted that there is prior work that gives better results with the use of large scale datasets like ImageNet
and Kinetics dataset [2], or with the help of external optical
flow input [22]. Those results are not directly comparable
to us because we are using a significantly smaller amount
of labeled data - only UCF101 and HMDB51. Nevertheless, our method shows promising results for learning motion representations from videos. Even with only a small
amount of labeled data, our action recognition network outperforms methods trained with a large amount of labeled
data with the exception of the recently trained I3D models [2] which used ImageNet and Kinetics dataset [11]. We
envision the performance of ActionFlowNet would further
improve when trained on larger datasets like Kinetics and
taking more input frames in the model.
Method
ResNet-18 Scratch
VGG-M-2048 Scratch [22]
Sequential Verification [17]
VGAN [28]
O3N [6]
OPN [15]
FlowNet fine-tuned (ours)
ActionFlowNet-2F (ours)
ActionFlowNet (ours)
UCF101 Accuracy
51.3
52.9
50.9
52.1
60.3
59.8
66.0
70.0
83.9
Table 2: Results on UCF101 (split 1) from single stream
networks with raw pixel input and without pretraining on
large labeled dataset.
Comparison to state-of-the-arts. We compare our approach to previous work that does not perform pretraining
with external large labeled datasets in Table 2 on UCF101.
All models are trained only with UCF101 labels with different unsupervised learning methods. Our models significantly outperform previous work that use videos for unsupervised feature learning [17, 28, 6, 15]. Specifically,
even with only our two-frame fine-tuned model on UCF101,
the model obtain more than 5.9% improvement compared
to Sequential Verification, VGAN and O3N, indicating the
importance of motion in learning video representations.
When combined with multitask learning, the performance
improves to 70.0%. Finally, when extending our model to
16 frames by 3D convolutions, the performance of ActionFlowNet further boost to 83.9%, giving a 23.6% improve-
(a) Image
(b) Output flow from (c) ImageNet Model: (d) ActionFlowNet-2F:
ActionFlowNet-2F
Appearance Only
Motion and Appearance
Figure 4: Visualization of important regions for action recognition. Our ActionFlowNet-2F discovers the regions where the
motions are happening to be important while ‘Appearance Only’ captures discriminative regions based on the appearance.
tion, while the ImageNet pretrained network relies more on
background appearance, which may not directly relate to
the action itself. However, when appearance is discriminative - for example the writing on the board in the last example - our model can also focus on appearance, which is not
possible for models that learn from optical flow only.
4.3.2
4.3.1 Learning Motions for Discriminative Regions
We visualize what is learned from the multitask network by
using the method from [33] by using a black square to occlude the frames at different spatial locations and compute
the relative difference between classification confidence before and after occlusion. We visualize the two-frame based
ActionFlowNet-2F for more straightforward visualization.
We compare the discriminative regions discovered by
our multitask network with ones by the ImageNet pretrained
ResNet-18, which only models the discriminative appearances without motion. Figure 4 shows example results. The
visualization reveals that our model focuses more on mo-
0.5
0.0
WallPushups
ApplyEyeMakeup
BodyWeightSquats
ShavingBeard
PlayingDaf
Haircut
FrontCrawl
PlayingTabla
BlowDryHair
Knitting
YoYo
PlayingDhol
Shotput
ThrowDiscus
HandstandPushups
JavelinThrow
LongJump
PoleVault
Lunges
HeadMassage
SalsaSpin
JumpingJack
BreastStroke
TennisSwing
MilitaryParade
CleanAndJerk
BoxingSpeedBag
PlayingFlute
CricketShot
PullUps
Hammering
CuttingInKitchen
HighJump
BenchPress
PizzaTossing
Drumming
Mixing
Diving
Typing
PlayingViolin
HulaHoop
HorseRace
PommelHorse
VolleyballSpiking
Billiards
RockClimbingIndoor
RopeClimbing
Biking
BasketballDunk
BandMarching
Skijet
TrampolineJumping
HorseRiding
IceDancing
SoccerPenalty
JugglingBalls
TableTennisShot
GolfSwing
ParallelBars
CricketBowling
FloorGymnastics
Rowing
BrushingTeeth
HandstandWalking
SumoWrestling
Surfing
BlowingCandles
ApplyLipstick
SkyDiving
PushUps
PlayingPiano
Rafting
WritingOnBoard
Swing
SoccerJuggling
Punch
BalanceBeam
HammerThrow
PlayingGuitar
Bowling
UnevenBars
Kayaking
PlayingSitar
Archery
Skiing
FrisbeeCatch
PlayingCello
BabyCrawling
FieldHockeyPenalty
StillRings
CliffDiving
WalkingWithDog
Fencing
TaiChi
BoxingPunchingBag
JumpRope
SkateBoarding
Nunchucks
BaseballPitch
Basketball
MoppingFloor
ment over the best previous work. This shows that explicitly
learning motion information is important for learning video
representations.
Figure 5 shows the optical flow estimation and prediction
results from our multi-frame model. Although the model
does not have accurate optical flow groundtruth for training, the optical flow quality is fairly good. The model predicts reasonable future optical flow, which shows semantic
understanding from the model to the frames in addition to
simply performing matching between input frames.
(a) ActionFlowNet vs C3D
0.5
0.0
SoccerJuggling
JumpRope
CricketShot
PullUps
JumpingJack
BodyWeightSquats
Lunges
Nunchucks
WallPushups
FrontCrawl
PommelHorse
FloorGymnastics
TrampolineJumping
SalsaSpin
Haircut
CleanAndJerk
Archery
BoxingSpeedBag
BenchPress
HeadMassage
Swing
HandstandWalking
Drumming
PushUps
PoleVault
Knitting
HighJump
GolfSwing
RockClimbingIndoor
Shotput
JavelinThrow
RopeClimbing
YoYo
CricketBowling
HandstandPushups
Skijet
SkyDiving
Surfing
HulaHoop
BabyCrawling
VolleyballSpiking
WalkingWithDog
Kayaking
ThrowDiscus
BlowDryHair
Biking
LongJump
JugglingBalls
ApplyEyeMakeup
Typing
BreastStroke
SumoWrestling
TableTennisShot
Billiards
Diving
BasketballDunk
Basketball
Punch
PlayingTabla
BandMarching
HammerThrow
HorseRace
ParallelBars
TennisSwing
HorseRiding
IceDancing
Bowling
SoccerPenalty
PizzaTossing
MilitaryParade
UnevenBars
PlayingViolin
PlayingPiano
WritingOnBoard
PlayingDaf
FieldHockeyPenalty
Skiing
FrisbeeCatch
Rowing
Hammering
PlayingDhol
PlayingFlute
ApplyLipstick
StillRings
ShavingBeard
Mixing
CuttingInKitchen
BalanceBeam
TaiChi
PlayingGuitar
Fencing
PlayingSitar
Rafting
CliffDiving
SkateBoarding
PlayingCello
BoxingPunchingBag
BaseballPitch
BlowingCandles
MoppingFloor
BrushingTeeth
Figure 5: Optical flow and future prediction outputs from
our multi-frame model. The 1st and 3rd row shows an
example of input videos, and the 2nd and 4th row shows
the corresponding optical flow outputs. The last optical
flow output frames (in red border) are extrapolated rather
than computed within input frames. Only last 8 frames are
shown per sample due to space limit.
Optical Flow and Future Prediction
(b) ActionFlowNet vs ImageNet pretrained ResNet-18
Figure 6: Classwise accuracy improvement by ActionFlowNet over pretrained models. The blue bars show positive improvements and the red ones show otherwise.
4.3.3 Classes Improved By Learning Motions
We compare the per class accuracy for ActionFlowNet, ImageNet pretrained model and C3D. Not all action classes are
motion-centric - objects and their contextual (background)
appearances provide more discriminative information for
some classes [9], which can greatly benefit from large
amounts of labeled data. As shown in Figure 6, our model
better recognizes action classes with simple and discriminative motion like WallPushups and ApplyEyeMakeup, while
C3D and ImageNet models perform better on classes with
complex appearance like MoppingFloor and BaseballPitch.
(a) Frame with movements
highlighted with red
(b) EpicFlow
(c) FlowNet on FlyingChairs
(d) FlowNet on UCF101
4.4. Recognition and Optical Flow Quality
In this section, we study the effects of different optical flow models for action recognition based on the twoframe models. We train our optical flow models on FlyingChairs or UCF101 and evaluate their accuracies on the
Sintel dataset (similar to [7] that trains the model on FlyingChairs but tests on other datasets).
We investigate how the quality of the learned optical flow
affects action recognition. Since optical flow in the multitask model is collaboratively learned with the recognition
task, the quality of optical flow in the multitask model does
not directly affect recognition accuracy. Thus, we use our
Stacked model learned with different datasets, fix the optical flow part and train the classification part in the network
shown in Figure 2. We compare the end-point-error of different optical flow learners and the corresponding classification accuracy in Table 3.
Method
Stacked on FlyingChairs
Stacked on UCF101
ResNet on EpicFlow
EPE on Sintel
9.12
11.84
6.29
Classification
Accuracy (%)
51.7
69.6
77.7
Table 3: Comparison between End-Point-Error (EPE, lower
is better) and the classification accuracy. Interestingly, better optical flow does not always result in better action recognition accuracy. Refer to the text for discussion.
Action Recognition with Learned Flow. Surprisingly,
even with lower end-point-error the Stacked model pretrained on FlyingChairs performs significantly worse than
the one pretrained on UCF101 dataset (51.7% vs 69.6%),
as shown in Table 3. Compared to the model directly taking high quality optical flow as input (77.7%), our models
are still not as good as training directly on optical flow. We
believe this is because the quality of learned optical flow is
not high enough.
To understand how the learned optical flow affects action recognition, we qualitatively observe the optical flow
performance in Figure 7. Even though the end-point error on Sintel of the FlowNet pretrained on FlyingChairs
is low, the estimated optical flow has lots of artifacts in
Figure 7: Qualitative comparison of flow outputs. It shows
an example of small motion, where the maximum magnitude of displacement estimated from EpicFlow is only about
1.6px. FlowNet trained on FlyingChairs dataset fails to estimate small motion, since the FlyingChairs dataset consists
of large displacement flow.
the background and the recognition accuracy on top of that
is correspondingly low. We believe the reason is that the
FlyingChairs dataset mostly consists of large displacement
flow, and therefore the model performs badly on estimating
small optical flow, which contributes less in the EPE metric
when averaged over the whole dataset. This is in contrast to
traditional optimization based optical flow algorithms that
can predict small displacements well but have difficulties
for large displacements.
In addition, traditional optical flow algorithms such as
TV-L1 and EpicFlow explicitly enforce smoothness and
constancy. They are able to preserve object shape information when the flow displacements are small, which is important for action recognition. While our models perform
comparably to traditional optical flow algorithms in terms
of endpoint error, our model is not optimized for preserving
flow smoothness. This shows that end-point-error of optical
flow in public dataset may not be a good indicator of action
classification performance, since shape preservation is not
accounted for in the metric.
5. Conclusion
We presented a multitask framework for learning action
with motion flow, named ActionFlowNet. By using optical
flow as supervision for classification, our model captures
motion information while not requiring explicit optical flow
computation as input. Our model significantly outperforms
previous feature learning methods trained without external
large scale data and additional optical flow input.
Acknowledgments
This research was supported in part by funds provided from
the Office of Naval Research under grant N000141612713
entitled “Visual Common Sense Reasoning for Multi-agent
Activity Prediction and Recognition”.
References
[1] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A
naturalistic open source movie for optical flow evaluation.
In ECCV, 2012. 5
[2] J. Carreira and A. Zisserman. Quo vadis, action recognition?
a new model and the kinetics dataset. In CVPR, 2017. 1, 2,
5, 6
[3] A. Diba, A. M. Pazandeh, and L. V. Gool. Efficient twostream motion and appearance 3d cnns for video classification. arXiv preprint arXiv:1608.08851, 2016. 2
[4] C. Feichtenhofer, A. Pinz, and R. Wildes. Spatiotemporal
residual networks for video action recognition. In NIPS,
2016. 2
[5] C. Feichtenhofer, A. Pinz, and A. Zisserman. Convolutional
two-stream network fusion for video action recognition. In
CVPR, 2016. 2
[6] B. Fernando, H. Bilen, E. Gavves, and S. Gould. Selfsupervised video representation learning with odd-one-out
networks. In CVPR, 2017. 1, 2, 6
[7] P. Fischer, A. Dosovitskiy, E. Ilg, P. Häusser, C. Hazırbaş,
V. Golkov, P. van der Smagt, D. Cremers, and T. Brox.
FlowNet: Learning Optical Flow with Convolutional Networks. In ICCV, 2015. 2, 3, 4, 5, 8
[8] K. He, X. Zhang, S. Ren, and J. Sun. Deep Residual Learning
for Image Recognition. CVPR, 2016. 3
[9] M. Jain, J. van Gemert, and C. Snoek. What do 15,000 object
categories tell us about classifying and localizing actions? In
CVPR, 2015. 8
[10] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar,
and L. Fei-Fei. Large-scale Video Classification with Convolutional Neural Networks. In CVPR, 2014. 1, 2, 6
[11] W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijayanarasimhan, F. Viola, T. Green, T. Back, P. Natsev, et al.
The kinetics human action video dataset. arXiv preprint
arXiv:1705.06950, 2017. 1, 2, 6
[12] D. Kingma and J. Ba. Adam: A Method for Stochastic Optimization. In ICLR, 2015. 5
[13] A. Krizhevsky, I. Sutskever, and G. E. Hinton. ImageNet
Classification with Deep Convolutional Neural Networks. In
NIPS, 2012. 3
[14] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre.
HMDB: A large video database for human motion recognition. In ICCV, 2011. 5
[15] H.-Y. Lee, J.-B. Huang, M. Singh, and M.-H. Yang. Unsupervised representation learning by sorting sequences. In
ICCV, 2017. 6
[16] I. Misra, A. Shrivastava, A. Gupta, and M. Hebert. Crossstitch Networks for Multi-task Learning. In CVPR, 2016. 4
[17] I. Misra, L. C. Zitnick, and M. Hebert. Shuffle and learn:
unsupervised learning using temporal order verification. In
ECCV, 2016. 1, 2, 6
[18] J. Y.-H. Ng and L. S. Davis. Temporal difference networks
for video action recognition. In WACV, 2018. 2
[19] J. Y.-H. Ng, M. Hausknecht, S. Vijayanarasimhan,
O. Vinyals, R. Monga, and G. Toderici. Beyond Short Snippets: Deep Networks for Video Classification. In CVPR,
2015. 6
[20] A. Rasmus, H. Valpola, M. Honkala, M. Berglund, and
T. Raiko. Semi-Supervised Learning with Ladder Networks.
In NIPS, 2015. 4
[21] E. Revaud, P. Weinzaepfel, Z. Harchaoui, and C. Schmid.
EpicFlow: Edge-Preserving Interpolation of Correspondences for Optical Flow. In CVPR, 2015. 5
[22] K. Simonyan and A. Zisserman. Two-Stream Convolutional
Networks for Action Recognition in Videos. In NIPS, 2014.
1, 2, 4, 5, 6
[23] K. Simonyan and A. Zisserman. Very Deep Convolutional
Networks for Large-Scale Image Recognition. In ICLR,
2014. 3
[24] K. Soomro, A. R. Zamir, and M. Shah. UCF101: A Dataset
of 101 Human Action Classes From Videos in The Wild. In
CRCV-TR-12-01, 2012. 5
[25] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going Deeper with Convolutions. In CVPR, 2015. 5
[26] D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri.
Learning Spatiotemporal Features with 3D Convolutional
Networks. In ICCV, 2015. 1, 2, 3, 5
[27] D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri.
Deep End2End Voxel2Voxel Prediction. In CVPRW DeepVision Workshop, 2016. 2, 3, 5
[28] C. Vondrick, H. Pirsiavash, and A. Torralba. Generating
videos with scene dynamics. In NIPS, 2016. 1, 2, 6
[29] J. Walker, C. Doersch, A. Gupta, and M. Hebert. An uncertain future: Forecasting from static images using variational
autoencoders. In ECCV, pages 835–851. Springer, 2016. 2,
3
[30] L. Wang, Y. Xiong, Z. Wang, Y. Qiao, D. Lin, X. Tang, and
L. Van Gool. Temporal segment networks: Towards good
practices for deep action recognition. In ECCV, 2016. 2
[31] X. Wang, A. Farhadi, and A. Gupta. Actions ˜ transformations. In CVPR, 2016. 2
[32] X. Wang and A. Gupta. Unsupervised learning of visual representations using videos. In ICCV, 2015. 2
[33] M. D. Zeiler and R. Fergus. Visualizing and Understanding
Convolutional Networks. In ECCV, 2014. 7
| 1cs.CV
|
Kernel entropy estimation for linear processes
Hailin Sanga , Yongli Sangb and Fangjun Xuc
arXiv:1712.00196v1 [math.ST] 1 Dec 2017
a
Department of Mathematics, University of Mississippi, University, MS 38677, USA
[email protected]
b
Department of Mathematics, University of Louisiana at Lafayette, Lafayette, LA 70504, USA
[email protected]
c
School of Statistics, East China Normal University, Shanghai 200241, China and NYU-ECNU
Institute of Mathematical Sciences at NYU Shanghai, Shanghai 200062, China
[email protected]; [email protected]
Abstract
Let {Xn : n ∈ N} be a linear process with bounded
probability density function f (x). We
R
study the estimation of the quadratic functional R f 2 (x) dx. With a Fourier transform on the
kernel function and the projection method, it is shown that, under certain mild conditions, the
estimator
X
2
Xi − Xj
K
n(n − 1)hn
hn
1≤i<j≤n
has similar asymptotical properties as the i.i.d. case studied in Giné and Nickl (2008) if the linear
process {Xn : n ∈ N} has the defined short range dependence. We also provide an application
to L22 divergence and the extension to multivariate linear processes. The simulation study for
linear processes with Gaussian and α-stable innovations confirms our theoretical results. As
an illustration, we estimate the L22 divergences among the density functions of average annual
river flows for four rivers and obtain promising results.
Keywords: linear process, kernel entropy estimation, quadratic functional, projection operator
2010 Mathematics Subject Classification: Primary 60F05; Secondary 62M10, 60G10,
62G05.
1
Introduction
Let f (x) be the probability density function ofRa sequence of identically distributed observations
2
{Xn }∞
n=1 . The quadratic functional Q(f ) R= R f (x) dx plays an important role in the study
2
of quadratic
R Rényi entropy R(f ) = − ln( f (x) dx), Rényi (1970), and the Shannon entropy
S(f ) = − f (x) ln f (x) dx, Shannon (1948). Entropy is widely applied in the fields of information
theory, statistical classification, pattern recognition and so on since it is a measure of uncertainty
in a probability distribution. In the literature, different estimators for the quadratic functional and
entropies with independent data have been well studied. For example, the nearest-neighbor estimator [Leonenko, Pronzato and Savani (2008); Penrose and Yukich (2013)], the kernel estimator
[Hall and Marron (1987); Bickel and Ritov (1988); Giné and Nickl (2008)], the orthogonal projection estimator [Laurent (1996, 1997)] and the U -statistics estimator (a special kernel estimator)
[Leonenko and Seleznjev (2010); Källberg, Leonenko and Seleznjev (2012)] under the independence
assumption. It is a challenging problem to study the estimation of the quadratic functional and the
corresponding entropies for dependent case. In Källberg, Leonenko and Seleznjev (2014) the authors successfully extended the U -statistics method to m-dependence sequence. They showed the
rate optimality and asymptotic normality of the U -statistics estimator for multivariate sequence.
1
Ahmad (1979) obtained the strong consistency of the quadratic functional by orthogonal series
method for stationary time series with strong mixing condition. Nevertheless, to the best of our
knowledge, general results for quadratic functional estimations of regular time series data under
mild conditions are still unknown.
R
In this paper, we study the quadratic functional R f 2 (x) dx for the following linear process
Xn =
∞
X
ai εn−i ,
(1)
i=0
where the innovations εi are independent and identically distributed (i.i.d.) real-valued random
∞
P
ai εn−i convariables in some probability space (Ω, F, P) and ai are real coefficients such that
i=0
verges in distribution. We would apply the kernel method which was first introduced by Rosenblatt
(1956) and Parzen (1962). This method was proved to be successful in estimating the probability density functions and their derivatives, the regression functions and so on in the independent
setting; see the books [Devroye and Györfi (1985); Silverman (1986); Nadaraya (1989); Wand and
Jones (1995); Schimek (2000); Scott (2015)] and references therein. Kernel method was also proved
to be successful in estimating the density functions and regression functions for time series data
[see Tran (1992); Honda (2000); Wu and Mielniczuk (2002); Wu, Huang and Huang (2010)].
In the classical kernel estimation of the quadratic functional Q(f ) with i.i.d. observations
n
{X
R i }2i=1 , Giné and Nickl (2008) applied a convolution method to obtain the bias E Tn (hn ) −
R f (x) dx of the kernel estimator
X
Xi − Xj
2
Tn (hn ) =
.
(2)
K
n(n − 1)hn
hn
1≤i<j≤n
They then used the Hoeffding’s decomposition for U -statistics to study the stochastic part Tn (hn )−
E Tn (hn ). In this way, they showed the rate optimality and efficiency of the kernel estimator for
the quadratic functional. Krishnamurthy et. al. (2015) applied a similar method to study the
L22 divergence between two distributions.
R Nevertheless, this method does not work well in the
estimation of the quadratic functional R f 2 (x)dx for linear process Xn given in (1) due to the
dependence structure.
Instead we utilize the Fourier transform and projection methods to derive the asymptotic
properties of the kernel estimator for the quadratic functional on time series data. With the
help of the Fourier transform, one can easily separate the random part Xi − Xj and the kernel
function K in the estimator Tn (hn ) given in (2) and also obtain the Hoeffding’s decomposition for
Tn (hn ) − E Tn (hn ) when {Xn } is a linear process; see (33). In fact, the Hoeffding’s decomposition
works for general stationary processes. Then we use projection and some kind of chain rule to
determine which terms contribute to the limit and which terms are negligible. Our method is
different from the martingale approximation method applied in Wu and Mielniczuk (2002) when
studying the kernel density estimation for linear processes. Moreover, this method allows us to get
rid of the 2nd moment assumption on the innovations of linear processes, that is, E |ε1 |2 < ∞. This
allows innovations to take heavy tail distributions, which are often followed by data in finance,
see Mittnik, Rachev and Paolella (1998). We only need E |eιλε1 − φε (λ)|2 ≤ cγ,2 (|λ|2γ ∧ 1) or
E |eιλε1 − φε (λ)|4 ≤ cγ,4 (|λ|4γ ∧ 1) for some γ ∈ (0, 1], where φε (λ) is the characteristic function
of the innovations. The range γ ∈ (0, 1] here is optimal when ε1 has non-degenerate distribution.
In this paper, the innovations have non-degenerate distribution because of the existence of the
probability density function f (x) of the linear process. See Lemma 7.4. On the other hand,
our method gives a clear picture on the relationship between the linear process {Xn } and the
efficiency of the estimator Tn (hn ). In addition, this methodology would have applications in kernel
2
estimations. For example, we can use it to extend results in Krishnamurthy et. al. (2015) to
short memory linear processes, and simplify existing proofs and assumptions in kernel density
estimations. Furthermore, we could suggest a more general definition for short memory and long
memory linear processes (see Definition 2.1 below) and extend our results to multivariate linear
processes.
Throughout this paper, if not mentioned otherwise, the letter c, with or without a subscript,
denotes a generic positive finite constant √
whose exact value is independent of n and may change
from line to line. We use ι to denote −1. For a complex number z, we use z and |z| to
denote its conjugate and modulus, respectively. The notation k · k2 means [E | · |2 ]1/2 . For two
functions a(x) and b(x) of real numbers, a(x)∼b(x) means a(x)/b(x)R → 1 as x → ∞. For any
integrable function g(x), its Fourier transform is defined as gb(u) = R eιxu g(x) dx. Let φ(λ) be
∞
P
ai εn−i , φn (λ) the empirical characteristic
the characteristic function of linear process Xn =
i=0
function from observations {Xi }ni=1 and φε (λ) the characteristic function of the innovations. That
n
P
is, φ(λ) = E [eιλXi ], φn (λ) = n1
eιλXi and φε (λ) = E [eιλε1 ]. For each i ∈ N, define H(Xi ) :=
i=1
H(Xi )(λ) := eιλXi − φ(λ), λ ∈ R.
The paper has the following structure. The main results are given in Section 2. We then give
a few examples in Section 3, an application of the results to L22 divergence in Section 4, simulation
study and real data analysis in Section 5, and extensions to multivariate linear processes in Section
6. Section 7 is devoted to the proofs of Theorems 2.1 and 2.2, based on the Fourier transform.
2
Main results
If {εi : i ∈ Z} is a sequence of i.i.d. random variables in Lp (R) for some p > 0, E εi = 0 when p ≥ 1,
∞
P
and {ai }∞
|ai |2∧p < ∞, then the linear process Xn
i=0 is a sequence of real coefficients such that
i=0
given in (1) exists and is well-defined. For p = 2, the process has short memory (short range
∞
P
dependence) if
|ai | < ∞ and long memory (long range dependence) otherwise. As for linear
i=0
processes with symmetric α-stable innovations (0 < α ≤ 2), i.e., the law of innovations having
characteristic function E [eιλε1 ] = exp(−cα |λ|α ) for some positive constant cα only depending on
∞
∞
P
P
α, {Xn } has short memory if
|ai |α/2 < ∞ and it has long memory if
|ai |α/2 = ∞ but
∞
P
i=0
i=0
|ai |α < ∞, respectively. See, e.g., Hsing (1999).
i=0
Let f (x) be the probability density function of the linear process Xn =
∞
P
ai εn−i , n ∈ N defined
i=0 R
in (1). We study the estimation of the quadratic functional of f (x), that is, R f 2 (x)dx, when the
∞
P
linear process Xn =
ai εn−i has short memory in the sense of the following definition.
i=0
Definition 2.1 A linear process Xn =
∞
P
ai εn−i defined in (1) has short memory if
i=0
∞ q
X
Var (eιλai ε1 ) < ∞
i=1
3
for all λ close enough to 0, and long memory if
∞
∞ q
X
X
Var (eιλai ε1 ) = ∞, but
Var (eιλai ε1 ) < ∞
i=1
i=1
for all λ close enough to 0.
In fact this definition generalizes the definitions of short memory or long memory linear processes
stated at the beginning of this section. Note that Var (eιλε1 ) can also be written as E |eιλε1 −φε (λ)|2
or 1 − |φε (λ)|2 . By the first part of Lemma 7.3, this definition contains the original one for linear
processes with finite second moment innovations. Using the characteristic function of α-stable
distribution, E |eιλε1 − φε (λ)|2 ≤ c α2 ,2 (|λ|α ∧ 1). Hence it also contains the one introduced in Hsing
(1999) for symmetric α-stable innovations. See the details in Example 3.2.
Moreover, Definition 2.1 can be used to define short and long memory linear processes with
general infinite variance innovations. Therefore, it can be applied to define all short and long
memory linear processes with finite variance innovations or infinite variance innovations. The
advantage of Definition 2.1 is that we only need to know the behavior of Var (eιλai ε1 ) = 1 −
|φε (ai λ)|2 without any moment assumption. This definition relates the coefficients {ai } with the
distribution of the innovations which is fully represented by the characteristic function. By contrast,
the traditional definition only relates the coefficients {ai } with the moment information of the
innovations. The definition in Hsing (1999) is only for linear processes with symmetric α-stable
innovations and it only relates the coefficients {ai } with the parameter α of the α-stable innovations.
Throughout the paper, we use Definition 2.1 to classify linear processes with short or long memory.
R
To estimate the quadratic functional R f 2 (x)dx, we shall apply the kernel method
X
Xi − Xj
2
Tn (hn ) =
K
,
n(n − 1)hn
hn
1≤i<j≤n
R
R
where the kernel K is a symmetric and bounded function with R K(u) du = 1 and R u2 |K(u)| du <
∞. The bandwidth sequence hn satisfies 0 < hn → 0 as n → ∞. The following are the main results
of this paper.
Theorem 2.1 Assume that
∞
P
|ai |γ < ∞,
i=0
2γ
2
R |λ| |φε (λ)| dλ
R
< ∞ and
E |eιλε1 − φε (λ)|4 ≤ cγ,4 |λ|4γ ∧ 1
(3)
for some γ ∈ (0, 1]. We further assume that f is bounded. Then there exist positive constants c1
and c2 such that
Z
1
2γ
2
+ hn ,
(4)
E Tn (hn ) −
f (x) dx ≤ c1
n
R
n
1
ηn,γ
1 X 2
h2γ
n
E Tn (hn ) − E Tn (hn ) −
Yi ≤ c2 2 + 2 +
n
n hn
n
n
(5)
i=1
and, if additionally nhn → ∞ as n → ∞,
i
√ h
L
n Tn (hn ) − E Tn (hn ) −→ N (0, 4σ 2 ),
where Yi = 2 f (Xi ) −
R
n P
∞
P
2 (x) dx , η
f
=
|ai |γ and σ 2 = lim n−1 Var (Sn ) with
n,γ
R
n→∞
`=0 i=`
Z
n
X
2
Sn =
f (Xi ) −
f (x) dx .
R
i=1
4
(6)
Theorem 2.2 Assume that
∞
P
i=0
|ai |γ < ∞,
2γ
2
R |λ| |φε (λ)| dλ
R
< ∞ and
E |eιλε1 − φε (λ)|2 ≤ cγ,2 |λ|2γ ∧ 1
(7)
R
b
for some γ ∈ (0, 1]. We further assume that f is bounded and R |K(λ)|
dλ < ∞. Then there exist
positive constants c3 and c4 such that
Z
1
2
2γ
f (x) dx ≤ c3
E Tn (hn ) −
+ hn ,
(8)
n
R
n
1
hγn
1X
+√
Yi ≤ c4
E Tn (hn ) − E Tn (hn ) −
n
nhn
n
(9)
i=1
and, if additionally
√
nhn → ∞ as n → ∞,
i
√ h
L
n Tn (hn ) − E Tn (hn ) −→ N (0, 4σ 2 )
(10)
for some σ 2 ∈ (0, ∞).
Remark 2.1 By Definition 2.1, Theorems 2.1 and 2.2 are results for short memory linear processes. On the other hand, we work on the well-defined linear process (1) in Theorem 2.1, Theorem
∞
P
2.2 and the rest of the paper. If we assume ε1 ∈ Lγ (R), together with condition
|ai |γ < ∞ in the
i=0
theorems, then the linear process (1) is well-defined. However, the moment condition ε1 ∈ Lγ (R)
is not needed in the proofs of our results.
Remark 2.2 When {Xn } is an i.i.d. sequence, the kernel estimator (2) was studied in Giné
and Nickl R(2008) with assumptions that the density function f (x) is
and the smoothness
R bounded
1
2γ
2
2
condition R |λ| |φ(λ)|R dλ < ∞ of f (x) for γ restricted to (0, 2 ]. If R u |K(u)| du < ∞ is replaced
by a weaker condition R |u|β |K(u)| du < ∞ for some 0 < β ≤ 2, then the upper bounds in (4) and
(8) above should be replaced by a constant multiple of n1 + hβ∧2γ
.
n
Remark 2.3 The assumption (3) in Theorem 2.1 is required to obtain (5) and (6). To just get the
inequality (4), we only need to assume (7). Using Lemma 7.3, we could easily see that E |ε1 |4γ < ∞
and E |ε1 |2γ < ∞ are sufficient for these two assumptions, respectively. However, these two moment conditions are not necessary. For example, when the innovation ε1 takes symmetric α-stable
distributions with α ∈ (0, 2), E |eιλε1 − φε (λ)|4 ≤ c α4 ,4 (|λ|α ∧ 1), E |eιλε1 − φε (λ)|2 ≤ c α2 ,2 (|λ|α ∧ 1),
but E |ε1 |α = ∞. On the other hand, assumption (3) is stronger than assumption (7). This explains
why the results in Theorem 2.1 are better than the results in Theorem 2.2 if both assumptions (3)
and (7) hold for a same γ in (0, 1].
Remark 2.4 If the underlying linear
process {Xn } is i.i.d. or m-dependent, then the assumption
E |eιλε1 − φε (λ)|4 ≤ cγ,4 |λ|4γ ∧ 1 for some γ ∈ (0, 1] in Theorem 2.1 can be removed by making
some minor modifications of its proof, and ηn,γ can be replaced by 0. Furthermore, for the i.i.d.
case, the right hand side of (4) can be replaced by c1 h2γ
n .
R
Remark 2.5 In Theorems 2.1 and 2.2, if we further have R |λ|4γ |φε (λ)|2 dλ < ∞, then, using
2β for all β ∈ [0, 1], the inequality (5) can be improved to
b
b
the fact |K(λh
n ) − K(0)| ≤ cβ |λhn |
n
1X
E Tn (hn ) − E Tn (hn ) −
Yi
n
i=1
5
2
≤ c2
1
ηn,γ
h4γ
n
+
+
n2 hn
n2
n
while (9) to
n
E Tn (hn ) − E Tn (hn ) −
1
h2γ
1X
n
+√ .
Yi ≤ c4
n
nhn
n
i=1
As consequences of Theorem 2.1 and Theorem 2.2, we can easily obtain the following results.
Corollary 2.1 Under the assumptions in Theorem 2.1. Let hn have order n
then
Z
− 4γ
f 2 (x) dx = OP (n 4γ+1 );
Tn (hn ) −
2
− 4γ+1
. If 0 < γ ≤ 1/4,
(11)
R
if 1/4 < γ ≤ 1, then
Z
i
√ h
L
n Tn (hn ) −
f 2 (x) dx −→ N (0, 4σ 2 ),
(12)
R
and, if we additional requireR K(x) ≥ 0 for all x, the estimator − ln( n1 + Tn (hn )) of the quadratic
Rényi entropy R(f ) = − ln( R f 2 (x) dx) satisfies
i
√ h
1
4σ 2
L
,
n − ln( + Tn (hn )) − R(f ) −→ N 0, R 2
n
( R f (x) dx)2
where σ 2 = lim n−1 Var (Sn ) with Sn =
n→∞
n
P
f (Xi ) −
i=1
R
Rf
2 (x) dx
(13)
.
Corollary 2.2 Under the assumptions in Theorem 2.2. Let hn have order n
then
Z
− 2γ
Tn (hn ) −
f 2 (x) dx = OP (n 2γ+1 );
1
− 2γ+1
. If 0 < γ ≤ 1/2,
(14)
R
if 1/2 < γ ≤ 1, then
Z
i
√ h
L
f 2 (x) dx −→ N (0, 4σ 2 ),
n Tn (hn ) −
(15)
R
and, if we additional requireR K(x) ≥ 0 for all x, the estimator − ln( n1 + Tn (hn )) of the quadratic
Rényi entropy R(f ) = − ln( R f 2 (x) dx) satisfies
i
√ h
1
4σ 2
L
n − ln( + Tn (hn )) − R(f ) −→ N 0, R 2
n
( R f (x) dx)2
for some σ 2 ∈ (0, ∞).
3
Examples
In this section, we shall give a few examples of linear processes to illustrate our results.
6
(16)
Example 3.1 (Linear process with Gaussian white noise innovations) Here the distribuλ2
tion
Gaussian with mean 0 and variance σ 2 . Note that E [eιλε1 ] = exp(− 2σ
So
2 ).
R 2 of ε1 is
2 dλ < ∞, E |eιλε1 − φ (λ)|4 ≤ c
4 ∧ 1 and E |eιλε1 − φ (λ)|2 ≤ c
2∧1 .
λ
|φ
(λ)|
λ
λ
ε
ε
1,4
ε
1,2
R
Therefore, Theorems 2.1, 2.2 and Corollaries 2.1 and 2.2 hold for all short memory linear processes with Gaussian white noise innovations.
Example 3.2 (Linear process with symmetric α-stable innovations) Here the distribution of ε1 is the symmetric α-stableR (SαS) with 0 < α < 2. Note that E [eιλε1 ] = exp(−cα |λ|α )
for some positive constant cα . So R λ2 |φε (λ)|2 dλ < ∞, E |eιλε1 − φε (λ)|4 ≤ c α4 ,4 (|λ|α ∧ 1) and
E |eιλε1 − φε (λ)|2 ≤ c α2 ,2 (|λ|α ∧ 1). Therefore, Theorems 2.1 and 2.2 hold for all short memory
linear processes having symmetric α-stable innovations (0 < α < 2) with γ = α4 and γ = α2 , respectively. The central limit theorem (12) in Corollary 2.1 and the central limit theorem (15) in
Corollary 2.2 hold if 1 < α < 2. We should apply Theorem 2.2 and Corollary 2.2 instead of Theorem 2.1 and Corollary 2.1 P
in order to study all short-memory processes with α-stable innovations
and coefficients satisfying i |ai |α/2 < ∞. See Step 2 of Section 5 for the detailed discussion.
Example 3.3 (Causal ARM A(p, q) process) An ARM A(p, q) model has the form
φ(B)Xt = θ(B)εt ,
where εt are i.i.d random variables with E εt = 0 and E ε2t = σε2 , φ and θ are two polynomials with
no common zeros,
φ(z) = 1 − φ1 z − φ2 z 2 − · · · − φp z p , φp 6= 0,
θ(z) = 1 + θ1 z + θ2 z 2 + · · · − θq z q , θq 6= 0,
and B is the back shift operator with B k Xt = Xt−k . An ARM A(p, q) process is causal if it can be
∞
∞
P
P
written as a one-sided linear process Xt =
ai εt−i with
|ai | < ∞. It is causal if and only if
i=0
i=0
φ(z) 6= 0 for |z| ≤ 1. In this case, the coefficients of the linear process can be determined by solving
a(z) =
∞
X
ai z i =
i=0
φ(z)
,
θ(z)
where |z| ≤ 1. It is easy to see that the coefficients ai decay exponentially fast to zero. So
∞ P
∞
P
|ai |γ < ∞ for any γ > 0. This means that we can choose γ as small as possible, which will
`=0 i=`
enlarge the choices of the innovations ε. Moreover, the upper bound in (5) of Theorem 2.1 can be
2γ
improved to a constant multiple of n21hn + hnn .
4
An application to L22 divergence
Let f (x) and g(x) be the probability densities for independent linear processes Xn =
and Yn =
∞
P
j=0
∞
P
ai εn−i
i=0
bj ξn−j defined as in (1), respectively. Then the L22 divergence between f and g is
defined as
Z
D(f, g) :=
R
(f (x) − g(x))2 dx =
Z
f 2 (x) dx +
R
Z
R
7
g 2 (x) dx − 2
Z
f (x)g(x) dx.
R
The estimator for D(f, g) is
b g) = Tn (f, hn ) + Tn (g, hn ) − 2Tn (f, g, hn ),
D(f,
where
Xi − Xj
K
,
hn
1≤i<j≤n
X
Yi − Yj
2
K
Tn (g, hn ) =
n(n − 1)hn
hn
2
Tn (f, hn ) =
n(n − 1)hn
X
1≤i<j≤n
and
1
Tn (f, g, hn ) = 2
n hn
X
K
1≤i,j≤n
Xi − Yj
hn
.
Krishnamurthy et. al. (2015) studied the estimation of L22 divergence for multivariate independent
2n
i.i.d. samples {Xi }2n
i=1 ∼ f (x) and {Yi }i=1 ∼ g(x). They used data splitting in the estimation
which is necessary to obtain the asymptotic normality [see Gretton et. al. (2012); Krishnamurthy
et. al. (2015)]. However, in our case, the data splitting is not needed.
Using similar arguments as in the proof of Theorems 2.1 and 2.2, one can obtain the same
results for Tn (f, g, hn ). Together with Theorems 2.1 and 2.2, we then have the following results.
Theorem 4.1 Assume that the independent linear processes {Xn } and {Yn } satisfy the same
conditions as in Theorem 2.1. Then there exist positive constants c1 and c2 such that
1
2γ
b
E D(f, g) − D(f, g) ≤ c1
+ hn ,
(17)
n
n
X
b g) − E D(f,
b g) − 1
E D(f,
Wi
n
2
i=1
1
ηn,γ
h2γ
n
≤ c2 2 + 2 +
n hn
n
n
and, if additionally nhn → ∞ as n → ∞,
i
√ h
L
b g) − E D(f,
b g) −→
n D(f,
N (0, 4σ 2 ),
(18)
(19)
where
Z
Z
Z
f 2 (x) dx + g(Yi ) −
g 2 (x) dx + g(Xi ) + f (Yi ) − 2 f (x)g(x) dx ,
Wi = 2 f (Xi ) −
R
ηn,γ =
n P
∞
P
`=0 i=`
|ai |γ ∨
n P
∞
P
R
|bi |γ , and σ 2 = lim n−1 Var (Sn ) with Sn =
n→∞
`=0 i=`
Let hn have order n
R
2
− 4γ+1
1
2
n
P
Wi .
i=1
b g).
. Then similar results as in Corollary 2.1 can be obtained for D(f,
Theorem 4.2 With the same notations as in Theorem 4.1, assume that the independent linear
processes {Xn } and {Yn } satisfy the same conditions as in Theorem 2.2. Then there exists a
positive constant c such that
n
1
X
hγn
b g) − E D(f,
b g) − 1
E D(f,
Wi ≤ c
+√
n
nhn
n
(20)
i=1
and (17), (19) hold for some
Let hn have order n
1
− 2γ+1
σ2
∈ (0, ∞).
b g).
. Then similar results as in Corollary 2.2 can be obtained for D(f,
8
5
Simulation study and real data analysis
In this section we first conduct a simulation study to verify the central limit theorems in Section
2 for the kernel entropy estimator of linear processes with different coefficient sequences and innovations with stable or Gaussian laws. For this purpose, we shall draw normal quantile plots and
histograms, perform hypothesis tests and provide confidence intervals on the means. In the second
part of this section, we apply the kernel method to study the divergences between each pair of the
annual average flow distributions of four rivers on the earth.
The simulation study has two steps.
R
Step 1: Calculate or approximate the true value of the quadratic functional R f 2 (x)dx.
By the inverse Fourier transform, the probability density function f (x) of the linear process in
(1) and its characteristic function φ(t) = E[eιtX1 ] have the following relationship:
Z
1
f (x) =
φ(t)e−ιtx dt.
(21)
2π R
Also, we have
φ(t) = E[eιtXn ] =
∞
Y
E[eιtai εn−i ].
i=0
Γ(i+d)
Now let ai = Γ(d)Γ(i+1)
, i ≥ 0, for any d < 21 , d 6= 0, −1, −2, · · · . Applying Stirling’s formula,
√
Γ(x) ∼ 2πe−x+1 (x − 1)x−1/2 as x → ∞, ai ∼ id−1 /Γ(d) as i → ∞. In particular, under the
condition −1 < d < 21 , d 6= 0, the FARIMA(0, d, 0) processes Xn = (1 − B)−d εn , where B is the
Γ(i+d)
backshift operator, are causal and they have the form of linear process with ai = Γ(d)Γ(i+1)
. See
Bondon and Palma (2007) for the extension of causality of FARIMA(p, d, q) processes to the range
of −1 < d < 1/2. We shall perform simulation study for linear processes with Gaussian or Cauchy
innovations and some selected values of d.
t2
Case 1. Suppose the innovations {εi } are i.i.d. N (0, 1) random variables, then φε (t) = e− 2
t2
P∞
2
and hence φ(t) = e− 2 i=0 ai . By the Gauss’s theorem [Gauss (1866)] for hypergeometric series,
P∞ 2 Γ(1−2d)
1
i=0 ai = Γ2 (1−d) for any d < 2 , d 6= 0, −1, −2, · · · . See also Bailey (1935) or its direct calculation
in Sang and Sang (2017). By the formula (21) and the application of MATLAB,
√
Z
2 − πx2
1 ∞ − t22 ΓΓ(1−2d)
2 (1−d)
f (x) =
e
cos(tx)dt =
e 8
π 0
4
if we take d = −0.5. As a result,
Z
f 2 (x)dx = 0.2500 if d = −0.5.
R
In this way, we also calculate or approximate the value of
in Table 1.
Table 1: The approximate value of
d
R
2
R f (x)dx
−0.1
0.2801
−0.3
0.2678
R
Rf
2 (x)dx
−0.5
0.2500
R
Rf
2 (x)dx
for other d values as listed
for linear processes with Gaussian innovations
−0.7
0.2300
−0.9
0.2095
−1.5
0.1531
−2.5
0.0856
−3.5
0.0462
Case 2. If the innovations {εi } P
have i.i.d. symmetric α-stable distribution with 0 < α < 2,
∞
α
α
α
then φε (t) = e−|t| and φ(t) = e−|t| i=0 |ai | . In particular, if α = 1, we have the standard Cauchy
9
P∞
−|t| i=0 |ai |
distribution with φε (t) = e−|t| and
P∞φ(t) = e P100,000 . In this case, with the help of the software
|ai | and then calculate f (x) and the value of
we approximate i=0 |ai | by i=0
RMathematica,
2 (x)dx. The result is in Table 2.
f
R
Table 2: The approximate value of
d
R
2
R f (x)dx
R
Rf
−0.1
0.0934
2 (x)dx
−0.3
0.0806
for linear processes with Cauchy innovations
−0.5
0.0796
−0.7
0.0796
−0.9
0.0796
Step 2: For each case, linear processes with Gaussian or Cauchy innovations, and each memory
parameter d, we produce m = 1, 000 linear processes with n = 4, 096 observations {Xi }ni=1 in each
of them by applying the MATLAB code in Faÿ et. al. (2009) with some modification. For each
generated linear
R 2 process, we apply the estimator Tn (hn ) as in (2) to estimate the true quadratic
functional R f (x)dx. The estimation is realized in MATLAB. Here we choose the normal kernel,
x2
that is, K(x) = √12π e− 2 , in the estimator Tn (hn ). The other common kernel functions are also
applicable since the main theorems and corollaries only assume some basic conditions on the kernel
functions.
Γ(i+d)
1
For the chosen ai = Γ(d)Γ(i+1)
since
, i ≥ 0, and for all results in Section 2, we require γ > 1−d
P∞
γ
d−1
/Γ(d).
i=0 |ai | < ∞ and ai ∼ i
Case 1. Suppose the innovations εi , i ∈ Z, are i.i.d. N (0, 1) random variables. Then E [eιλε1 ] =
R 2
λ2
2
ιλε1 − φ (λ)|4 ≤ c
4
ιλε1 − φ (λ)|2 ≤
exp(− 2σ
ε
1,4 λ ∧ 1 and E |e
ε
2 ). So
R λ |φε (λ)| dλ < ∞, E |e
c1,2 λ2 ∧ 1 . In this case, we should apply Theorem 2.1 and Corollary 2.1 instead of Theorem
2.2 and Corollary 2.2. The convergence rate (11) in Corollary 2.1 is faster than the convergence
rate (14) in Corollary 2.2. The central limit theorem (12) in Corollary 2.1 holds for 1/4 < γ ≤ 1
while the central limit theorem (15) in Corollary 2.2 holds only for 1/2 < γ ≤ 1. See also Remark
1
2.3. To have the central limit theorem (12), we require 1−d
∨ 41 < γ ≤ 1 and the bandwidth hn
−
2
to have the order of n 4γ+1 . In this Gaussian case, we recommend to take γ = 1 and hn to have
2
the order of n− 5 to make (12) hold for all d < 0. To apply kernel method in estimation, one
should select an optimal bandwidth based on some criteria, for example, to minimize the mean
squared error.
R For the i.i.d. case, Giné and Nickl (2008) construct a kernel-based rate adaptive
in the first step. It is interesting
estimator of R f 2 (x)dx by selecting the bandwidth hn adaptively
R 2
to investigate the bandwidth adaptive kernel estimator of R f (x)dx for linear processes from both
theoretical and application viewpoints. However, it seems that the study in this direction is very
difficult. We leave it as an open question for future study. For the current simulation study, we
take hn = n−2/5 in the Gaussian case.
R
Case 2. In the α-stable case, R λ2 |φε (λ)|2 dλ < ∞, E |eιλε1 − φε (λ)|4 ≤ c α4 ,4 (|λ|α ∧ 1) and
E |eιλε1 − φε (λ)|2 ≤ c α2 ,2 (|λ|α ∧ 1). Therefore, Theorem 2.1 (Corollary 2.1) and Theorem 2.2
(Corollary 2.2) hold for all short memory linear processes having symmetric
α-stable innovations
P
α/4 < ∞ in Theorem
(0 < α < 2) with γ = α4 and γ = α2 , respectively. Hence we require ∞
|a
|
i=0 i
P
α/2 < ∞ in Theorem 2.2 and Corollary 2.2. For
2.1 and Corollary 2.1 and we only require ∞
|a
|
i
i=0
this reason, we should apply Theorem 2.2 and Corollary 2.2 instead of Theorem 2.1 and Corollary
2.1. Since γ = α2 , Theorem 2.2 holds for all α ∈ (0, 2) while in Corollary 2.2, we have (14) for
α ∈ (0, 1] and the central limit theorem (15) only for α ∈ (1, 2). Hence if 1 < α < 2, we should take
1
hn to have the order of n− α+1 to confirm the central limit theorem (15). If 0 < α ≤ 1, we should
√
not expect to have (15). But in this case we can take hn with nhn → ∞ to confirm the central
limit theorem (10) in Theorem 2.2. In the simulation study, we take hn = n−2/5 . In Theorem 2.2
and Corollary 2.2, the coefficients ai should have parameter d with d < 1 − γ1 = 1 − α2 .
10
Figure 1 and Figure 2 show
the histograms with kernel density fits and normal Q-Q plots of rn =
i
R 2
√ h
n Tn (hn ) − R f (x) dx from m = 1, 000 simulated linear processes with Gaussian innovations
and d = −0.1, −0.9, −3.5. It is clear that the realizations of rn are distributed as some normal
distribution in each case. We also have performed simulation study for d = −0.3, −0.5, −0.7, −1.5
and −2.5 and obtained similar plots and results.
We then perform hypothesis test H0 : µ = 0; H1 : µ 6= 0 for the mean of the asymptotic
distribution of rn and provide the 95% confidence interval for µ for each value of d. The result is
listed in Table 3. It is clear that rn is asymptotically centered in each case. Over all, the simulation
study here confirms the analysis for the Gaussian case based on the main results in Section 2.
2
Histogram: Gaussian case, d=-0.1
2.5
Histogram: Gaussian case, d=-0.9
8
1.8
Histogram: Gaussian case, d=-3.5
7
1.6
2
6
1.4
1.2
5
1.5
1
4
0.8
1
3
0.6
2
0.4
0.5
1
0.2
0
-1
-0.5
0
0.5
1
0
-1
1.5
-0.5
0
0.5
0
-0.3
1
-0.2
-0.1
0
0.1
0.2
0.3
Figure 1: The histograms with kernel density fits of rn from m = 1, 000 simulated processes with
Gaussian innovations and d = −0.1, −0.9, −3.5.
QQ Plot: Gaussian case, d=-0.1
0.8
Quantiles of Input Sample
Quantiles of Input Sample
0.2
0.6
0.6
0.4
0.2
0
-0.2
-0.4
-2
0
2
Standard Normal Quantiles
4
0.2
0
-0.2
-0.4
-0.8
-4
QQ Plot: Gaussian case, d=-3.5
0.15
0.4
-0.6
-0.6
-0.8
-4
QQ Plot: Gaussian case, d=-0.9
Quantiles of Input Sample
1
0.8
0.1
0.05
0
-0.05
-0.1
-0.15
-2
0
2
Standard Normal Quantiles
4
-0.2
-4
-2
0
2
4
Standard Normal Quantiles
Figure 2: The normal Q-Q plots of rn from m = 1, 000 simulated processes with Gaussian innovations and d = −0.1, −0.9, −3.5.
11
Table 3: The p-values on testing H0 : µ = 0; H1 : µ 6= 0 for the asymptotic distribution of rn and
the 95% confidence intervals for µ.
d
p-value
CI
d
p-value
CI
−0.1
0.3996
(-0.0191,0.0076)
−0.9
0.6705
(-0.0148, 0.0095)
−0.3
0.7810
(-0.0159,0.0120)
−1.5
0.8297
(-0.0064,0.0080)
−0.5
0.5058
(-0.0086,0.0174)
−2.5
0.2055
(-0.0021, 0.0098)
−0.7
0.6676
(-0.0099, 0.0154)
−3.5
0.2211
(-0.0058, 0.0013)
Figureh 3 and Figure 4 show
i the histograms with kernel density fits and normal Q-Q plots of
R 2
√
rn = n Tn (hn ) − R f (x) dx from m = 1, 000 simulated processes with Cauchy innovations and
d = −0.1, −0.5, −0.9. It is clear that the realizations of rn do not follow normal distribution in
the case d = −0.1 or d = −0.5. We also have performed simulation study for d = −0.3, −0.7 and
obtained similar plots and results. For d = −0.9, the histogram and normal Q-Q plot indicate that
the distribution of rn is very close to some normal distribution with mean 0. This is because the
central limit theorem (15) in Corollary 2.2 holds for α = 1 + with an arbitrary small positive
number, d < 1 − α2 , which is close to the case α = 1, d = −0.9 here. In the case 0 < α ≤ 1, we
have the central limit theorem (10) although (15) does not hold any more.
0.6
Histogram: Cauchy case, d=-0.1
1.2
0.5
1
0.4
0.8
0.3
0.6
0.2
0.4
0.1
0.2
Histogram: Cauchy case, d=-0.5
2.5
Histogram: Cauchy case, d=-0.9
2
1.5
1
0
-8
-6
-4
-2
0
2
4
0
-4
0.5
-3
-2
-1
0
1
2
0
-1.5
-1
-0.5
0
0.5
1
Figure 3: The histograms with kernel density fits of rn from m = 1, 000 simulated processes with
Cauchy innovations and d = −0.1, −0.5, −0.9.
12
QQ Plot: Cauchy case, d=-0.1
4
1.5
QQ Plot: Cauchy case, d=-0.5
0.6
1
2
1
0
-1
-2
-3
-4
0.5
0
-0.5
-1
-1.5
-2
-6
-4
-2
0
2
4
-3
-4
Standard Normal Quantiles
0.2
0
-0.2
-0.4
-0.6
-0.8
-2.5
-5
QQ Plot: Cauchy case, d=-0.9
0.4
Quantiles of Input Sample
Quantiles of Input Sample
Quantiles of Input Sample
3
-2
0
2
4
-1
-4
-2
Standard Normal Quantiles
0
2
4
Standard Normal Quantiles
Figure 4: The normal Q-Q plots of rn from m = 1, 000 simulated processes with Cauchy innovations
and d = −0.1, −0.5, −0.9.
Figure 5 and Figure 6 show the histograms with kernel density fits and normal Q-Q plots
√
of nTn (hn ) from m = 1, 000 simulated processes with stable innovations (α = 0.5) and d =
√
−1.8, −2.5, −3.5. It is clear that the values of nTn (hn ) do not follow normal distribution in the
case d = −1.8 while they do follow normal distribution in the case d = −3.5. For d = −2.5, the
√
histogram and normal Q-Q plot indicate that the distribution of nTn (hn ) is not Gaussian but
√
is close to some normal distribution. This confirms the analysis that nTn (hn ) follows a normal
distribution when d < 1 − α2 even if 0 < α ≤ 1.
Histogram: alpha=0.5, d=-1.8
6
12
Histogram: alpha=0.5, d=-2.5
5
10
25
4
8
20
3
6
15
2
4
10
1
2
5
0
0
0.1
0
0.2
0.4
0.6
0.8
1
0.2
0.3
0.4
0.5
Histogram: alpha=0.5, d=-3.5
30
0.6
0
0.14
0.16
0.18
0.2
0.22
0.24
0.26
0.28
√
Figure 5: The histograms with kernel density fits of nTn (hn ) from m = 1, 000 simulated processes
with stable innovations (α = 0.5) and d = −1.8, −2.5, −3.5.
13
QQ Plot: alpha=0.5, d=-1.8
0.6
Quantiles of Input Sample
Quantiles of Input Sample
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
-4
QQ Plot: alpha=0.5, d=-2.5
0.28
QQ Plot: alpha=0.5, d=-3.5
0.55
Quantiles of Input Sample
1
0.9
0.5
0.45
0.4
0.35
0.3
0.25
0.26
0.24
0.22
0.2
0.18
0.2
-2
0
2
Standard Normal Quantiles
4
0.15
-4
-2
0
2
Standard Normal Quantiles
4
0.16
-4
-2
0
2
4
Standard Normal Quantiles
√
Figure 6: The normal Q-Q plots of nTn (hn ) from m = 1, 000 simulated processes with stable
innovations (α = 0.5) and d = −1.8, −2.5, −3.5.
b g) to study
For the purpose of illustration, we apply the L22 kernel divergence estimator D(f,
the difference of annual average river flows among four rivers on the earth. The data sets consist of
annual average river flows (1,000 cubic meters per second) from Danube River at Orshava, Romania,
Gota River near Sjotop-Vannersburg, Sweden, Mississippi River near St. Louis, Missouri, USA and
Rhine River near Basle, Switzerland for 96 years between 1861 and 1956. The river flow data of
these rivers are well studied in ecology and statistics, e.g., Hipel and McLeod (1994). The original
data sets cover different time periods. We choose part of each data set from year 1861 to year 1956
in order to compare the flows in the same time period. The Kwiatkowski, Phillips, Schmidt, and
Shin (KPSS) test, the time series plots in Figure 7 and the autocorrelation plots in Figure 8 show
the stationarity and short memory property of the river flow level of each river.
We apply the estimator to calculate the divergence between the flow distributions of every two
rivers. As in the simulation study, we take the standard normal density function as the kernel
function. The right part of Figure 7 gives the kernel density estimations of river flows for each
river. Since they approximately follow some normal distributions, we let the bandwidth to be
hn = 96−0.4 ≈ 0.161. All the divergences are presented in Table 4. According to the density
curves in Figure 7, it is not a surprise that the biggest L22 divergence is between the Gota River
and the Rhine River. The density functions of these two rivers do not have much overlap although
they are very close in location. On the other hand, the density functions of Gota River and
Rhine River, in particular the density function of the Gota River, have much bigger values in
their supports than those of the other two rivers. The squared difference (f (x) − g(x))2 , not the
relative location between the supports of f (x) and g(x), makes contribution to the value of the
L22 divergence D(f, g). Mississippi River and Danube River have very similar density functions in
terms of supports and values. Mississippi River has the smallest maximum density and it has a
little bit overlap with Gota River or Rhine River. There is almost no overlap between the Danube
River and Gota River or Rhine River. This explains the order of the divergences in Table 4.
14
Danube River
Yearly flow
8
3.5
4
1870
1880
1890
1900
1910
1920
1930
1940
1950
1960
3
Gota River
0.8
Yearly flow
Danube
Gota
Mississippi
Rhine
6
2
1860
0.6
2.5
0.4
0.2
1860
1870
1880
1890
Yearly flow
1900
1910
1920
1930
1940
1950
1960
2
Mississippi River
10
1.5
5
0
1860
1870
1880
1890
1900
1910
1920
1930
1940
1950
1
1960
Rhine River
1.4
Yearly flow
Probability density curves of four river flows
4
1.2
0.5
1
0.8
0.6
1860
1870
1880
1890
1900
1910
1920
1930
1940
1950
0
-2
1960
Time (years)
0
2
4
6
8
10
12
Average annual river flow (1,000 cubic meters per second)
Figure 7: The time series plots and probability density functions of average annual river flow.
Sample ACF: Danube River
0.8
0.6
0.4
0.2
0
-0.2
-0.4
0
5
10
15
Sample ACF: Gota River
1
Sample Autocorrelation
Sample Autocorrelation
1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
20
0
5
Lag
Sample ACF: Mississippi River
0.8
0.6
0.4
0.2
0
-0.2
-0.4
0
5
10
15
20
15
Sample ACF: Rhine River
1
Sample Autocorrelation
Sample Autocorrelation
1
10
Lag
0.8
0.6
0.4
0.2
0
-0.2
-0.4
20
Lag
0
5
10
15
20
Lag
Figure 8: The sample autocorrelation functions of average annual river flow.
15
Table 4: River flow divergences
The rivers
Gota River, Rhine River
Gota River, Danube River
Gota River, Mississippi River
Rhine River, Danube River
Rhine River, Mississippi River
Danube River, Mississippi River
6
Divergence
2.7797
2.1421
2.0725
1.6129
1.5433
0.0348
Extensions to multivariate linear processes
√
In this section, we denote norm functions in various linear spaces by | · |. For x ∈ Rd , |x| := x xT
where xT is the transposition of x. For a d × d deterministic real matrix a, |a| := max |a xT |.
x∈Rd ,|x|=1
In the sequel, we give extensions to the following multivariate linear process
Xn =
∞
X
ai εn−i ,
(22)
i=0
where the innovations εi are i.i.d. d × 1 random vectors in some probability space (Ω, F, P) and ai
∞
P
are d × d deterministic real matrices such that
ai εn−i converges in distribution. It is easy to
i=0
see that the multivariate linear process Xn in (22) exists and is well defined when εi are i.i.d. d × 1
random vectors in Lp (Rd ) for some p > 0 and ai are d × d deterministic real matrices satisfying
∞
P
|ai |2∧p < ∞. Let f be the probability density function of Xn . To estimate the quadratic
i=0
functional of f , we would use the estimator
Tn (hn ) =
1
n(n − 1) det(hn )
X
K h−1
n (Xi − Xj ) ,
1≤i6=j≤n
R
R
where the kernel K is a symmetric and bounded function with Rd K(u) du = 1 and Rd |u|2 |K(u)| du <
∞. The bandwidth sequence hn are diagonal matrices diag[hn1 , hn2 , · · · , hnd ] satisfying 0 < hni →
0 for all i = 1, 2, · · · , d.
We require the following assumptions on ai and εi .
Assumption
R 1 det(a0 ) 6= 0, det(am ) 6= 0 for some m ≥ 1 and ai = d × d zero matrix for all
i > m, and Rd |λ|2γ |φε (λ)|2 dλ < ∞ for some γ ∈ (0, 1].
Assumption 2 det(a0 ) 6= 0, det(ap ) 6= 0 and det(aq ) 6= 0 with p = min{i ∈ N : det(ai ) 6= 0} and
∞
P
q = min{i ∈ N : det(ai ) 6= 0 and i > p} < ∞,
|ai |γ < ∞, E |eιλε1 − φε (λ)|4 ≤ cγ,4 |λ|4γ ∧ 1
i=0
R
and Rd |λ|2γ |φε (λ)|2 dλ < ∞ for some γ ∈ (0, 1], ∃ two indices ir s.t. det(air +r − air ) 6= 0 for each
r ∈ {1, 2, · · · , q}.
Assumption 3 det(a0 ) 6= 0, det(ap ) 6= 0 with p = min{i ∈ N : det(ai ) 6= 0} < ∞, ∃ indice ir such
∞
P
that det(air − air +r ) 6= 0 for each r ∈ {1, · · · , p},
|ai |γ < ∞, E |eιλε1 − φε (λ)|2 ≤ cγ,2 |λ|2γ ∧ 1
i=0
R
and Rd |λ|2γ |φε (λ)|2 dλ < ∞ for some γ ∈ (0, 1].
16
Using similar arguments as in the proof of Theorems 2.1 and 2.2 with some proper modifications,
we can get the following results.
Theorem 6.1 Under the assumption 1 or 2, we further assume that f is bounded. Then there
exist positive constants c1 and c2 such that
Z
1
2
2γ
f (x) dx ≤ c1
+ det(hn )
,
(23)
E Tn (hn ) −
n
Rd
n
E Tn (hn ) − E Tn (hn ) −
1X
Yi
n
2
≤ c2
i=1
ηn,γ
1
det(hn )2γ
+
+
n2 det(hn )
n2
n
(24)
and, if additionally n det(hn ) → ∞ as n → ∞,
i
√ h
L
n Tn (hn ) − E Tn (hn ) −→ N (0, 4σ 2 ),
where Yi = 2 f (Xi ) −
R
Rd
(25)
n P
∞
P
f 2 (x) dx , ηn,γ =
|ai |γ and σ 2 = lim n−1 Var (Sn ) with
n→∞
`=0 i=`
Sn =
n
X
Z
f (Xi ) −
2
f (x) dx .
Rd
i=1
Theorem 6.2 Under the assumption 3, we further assume that f is bounded and
∞. Then there exist positive constants c3 and c4 such that
Z
1
2
2γ
+ det(hn )
,
E Tn (hn ) −
f (x) dx ≤ c3
n
Rd
R
Rd
b
|K(λ)|
dλ <
(26)
n
1X
1
det(hn )γ
E Tn (hn ) − E Tn (hn ) −
Yi ≤ c4
+ √
n
n det(hn )
n
(27)
i=1
and, if additionally
√
n det(hn ) → ∞ as n → ∞,
i
√ h
L
n Tn (hn ) − E Tn (hn ) −→ N (0, 4σ 2 )
(28)
for some σ 2 ∈ (0, ∞).
Remark
6.1 In the i.i.d case, that is, det(a0 ) 6= 0 and ai = d × d zero matrix for all i ≥ 1, and
R
2γ |φ (λ)|2 dλ < ∞ for some γ ∈ (0, 1], then results in Theorem 6.1 still hold with the right
|λ|
ε
Rd
hand side of (23) replaced by a constant multiple of det(hn )2γ and ηn,γ in (24) by 0, respectively.
17
7
Proofs
In this section, the proofs of our main results, Theorems 2.1 and 2.2, are given. First we provide
two useful lemmas based on the projection method.
For each i ∈ Z, let Fi be the σ-field generated by {εk : k ≤ i}. Given an integrable complexvalued random variable Z, we define the following projection operator Pi as
Pi Z = E [Z|Fi ] − E [Z|Fi−1 ]
(29)
for each i ∈ Z. It is easy to see that
E [Pi Z Pj W ] = 0
(30)
for any two integrable complex-valued random variables Z and W if i 6= j.
The estimation in the following lemma is useful in deriving the order of the bias of the kernel
estimator (2).
Lemma 7.1 Suppose
∞
P
|ai |γ < ∞ and E |eιλε1 − φε (λ)|2 ≤ cγ,2 |λ|2γ ∧ 1 for some γ ∈ (0, 1].
i=0
Then there exists a positive constant c such that
i
h
X
E H(Xi )(λ)H(Xj )(λ)
1≤i6=j≤n
p Y
∞
h
i
X
≤ c n |φε (λa0 )| |λ|2γ |φε (λa0 )| + |λ|γ |φε (λap )| +
|φε (λ(a` − a`+r ))| ,
r=1 `=1
where p = min{i ∈ N : ai 6= 0} < ∞. Moreover, the right hand side of the above inequality is equal
to 0 if ai = 0 for all i ∈ N.
Proof. It suffices to show the case when ai 6= 0 for some i ∈ N. Recall the definition of projection
operator Pk in (29). Applying telescoping, (30) and the triangle inequality, we have
X
h
i
E H(Xi )(λ)H(Xj )(λ) ≤ 2
X
i
X
i
h
E Pk H(Xi )(λ)Pk H(Xj )(λ)
1≤i<j≤n k=−∞
1≤i6=j≤n
=2
X
i
X
i
h
E P0 H(Xi−k )(λ)P0 H(Xj−k )(λ) .
1≤i<j≤n k=−∞
Note that
∞
j−k−1
i−k−1
i
h ιP
i
Y
λ(a`+i−k −a`+j−k )ε−` ih Y
E P0 H(Xi−k )(λ)P0 H(Xj−k )(λ) = E e `=1
|φε (λa` )|2
φε (−λa` )
h
`=0
`=i−k
h
i
× E (eιλai−k ε0 − φε (λai−k ))(e−ιλaj−k ε0 − φε (−λaj−k )) .
Pn−1 Pn
Pi
Next, we decompose the triple sum i=1
j=i+1
k=−∞ into the sum with k < i, the sum with
k = i and i + 2 ≤ j ≤ n and the sum with k = i and j = i + 1. By Cauchy-Schwartz inequality, we
18
have
h
i
E H(Xi )(λ)H(Xj )(λ)
X
1≤i6=j≤n
≤2
n−1
X
n
X
i−1
X
|φε (λa0 )|2
p
1 − |φε (λai−k )|2
q
1 − |φε (λaj−k )|2
i=1 j=i+1 k=−∞
+2
n−1
X
j−i−1
Y
n
X
|φε (λa` )|
i=1 j=i+2 `=0
+2
n−1
X
∞
Y
q
p
|φε (λ(a` − a`+j−i ))| 1 − |φε (λa0 )|2 1 − |φε (λaj−i )|2
`=1
|φε (λa0 )|
i=1
∞
Y
|φε (λ(a` − a`+1 ))| .
`=1
Note that the second term on the right hand side of the above inequality is less than
2
n−1
X
n
X
|φε (λa0 )||φε (λap )|
q
1 − |φε (λaj−i )|2
i=1 j=i+1+p
+2
i+p
n−1
X X
|φε (λa0 )|
i=1 j=i+2
∞
Y
q
|φε (λ(a` − a`+j−i ))| 1 − |φε (λaj−i )|2 .
`=1
Therefore, using the conditions of the lemma,
X
n−1
n
h
i
X X
2γ
E H(Xi )(λ)H(Xj )(λ) ≤ c1 |λ|
i−1
X
|φε (λa0 )|2 |ai−k |γ |aj−k |γ
i=1 j=i+1 k=−∞
1≤i6=j≤n
γ
+ c1 |λ|
n−1
X
n
X
|φε (λa0 )||φε (λap )||aj−i |γ
i=1 j=i+1+p
+ c1
n−1
X
|φε (λa0 )|
p Y
∞
X
|φε (λ(a` − a`+r ))|.
r=1 `=1
i=1
This gives the desired result.
Remark 7.1 Note that in Lemma 7.1, if the linear process {Xn } is m-dependent, then am 6=
p Q
∞
P
0, m ≥ p and
|φε (λ(a` − a`+r ))| ≤ p|φε (λam )|. If lim ai = 0, a0 6= 0 and ai 6= 0 for
i→∞
r=1 `=1
infinite many i ∈ N, then for each r ∈ {1, 2, · · · , p}, qr = min{i ∈ N : ai − ai+r 6= 0} < ∞ and
p Q
p
∞
P
P
|φε (λ(a` − a`+r ))| ≤
|φε (λ(aqr − aqr +r ))|.
r=1 `=1
r=1
The next lemma is needed in the proof of Theorem 2.2.
Lemma 7.2 For any m ∈ N and γ ∈ (0, 1], if E |eιλε1 − φε (λ)|2 ≤ cγ,2 |λ|2γ ∧ 1 , then there exists
a positive constant c such that
h
E φn (λ) − φ(λ)
2
i
∞
i2
X
Y
γ m−1
ch
γ
≤
m+2
|ai | |λ|
|φε (λaj )| .
n
i=m
j=0
Proof. Obviously, E H(X1 ) = 0, E |H(X1 )|2 < ∞ and
n(φn (λ) − φ(λ)) =
n h
X
n
i X
eιλXi − E eιλXi =
H(Xi ).
i=1
i=1
19
Note that
P0 H(Xi ) = E [H(Xi )|F0 ] − E [H(Xi )|F−1 ]
for each i ∈ N. Then,
P0 H(Xi ) = E [eιλXi |F0 ] − E [eιλXi |F−1 ]
ι
=e
∞
i−1
λaj εi−j
j=i
ι
=e
∞
P
∞
P
i
P
P
h ι P λaj εi−j i
ι
λaj εi−j h ι
λaj εi−j i
E e j=0
− e j=i+1
E e j=0
i−1
λaj εi−j
j=i+1
h ι P λaj εi−j ih
i
E e j=0
eιλai ε0 − E eιλai ε0
:= I ∗ II ∗ III,
ι
where I = e
∞
P
i−1
λaj εi−j
j=i+1
h ι P λaj εi−j i i−1
Q
, II = E e j=0
=
φε (λaj ) and
j=0
III = eιλai ε0 − E [eιλai ε0 ] = eιλai ε0 − φε (λai ).
Since I and III are independent,
i
h
kP0 H(Xi )k22 = E P0 H(Xi )P0 H(Xi )
= [II ∗ II]E [I ∗ Ī]E [III ∗ III]
=
i−1
Y
|φε (λaj )|2 [1 − |φε (λai )|2 ]
j=0
for i ≥ 1 and kP0 H(X0 )k22 = 1 − |φε (λa0 )|2 . Consequently,
∞
X
kP0 H(Xi )k2 =
∞ Y
i−1
X
p
p
1 − |φε (λa0 )|2 +
|φε (λaj )| 1 − |φε (λai )|2
i=0
i=1 j=0
≤m+
∞ Y
i−1
X
p
|φε (λaj )| 1 − |φε (λai )|2
i=m j=0
≤m+
m−1
Y
∞ p
X
|φε (λaj )|
1 − |φε (λai )|2
i=m
j=0
≤ m + cγ
∞
X
γ
|ai |
|λ|
i=m
γ
m−1
Y
|φε (λaj )|.
j=0
This completes the proof.
Now we give the proof of Theorem 2.1.
Proof of Theorem 2.1. The proof will be done in several steps.
Step 1. We give the estimation for
Z
E [Tn (hn )] −
R
20
f 2 (x) dx .
By Fourier inverse transform, Tn (hn ) can be written as
X Z
1
−ιλ(Xi −Xj )
b
K(λh
dλ.
Tn (hn ) =
n )e
πn(n − 1)
R
(31)
1≤i<j≤n
Together with the Plancherel theorem, we have
Z
E [Tn (hn )] −
f 2 (x) dx
R
h
i
X Z
1
ιλXi
b
=
K(λh
− φ(λ))(e−ιλXj − φ(−λ)) dλ
n ) E (e
2πn(n − 1)
1≤i6=j≤n R
Z
1
2
b
b
K(λh
+
n ) − K(0) |φ(λ)| dλ.
2π R
(32)
2β for any β ∈ [0, 1],
b
b
Using Lemma 7.1 and the inequality |K(λh
n ) − K(0)| ≤ cβ |λhn |
Z
Z
Z
c1
2
2γ
2
2γ
E [Tn (hn )] −
f (x) dx ≤
(1 + |λ| )|φε (λ)| dλ + c1 hn
|λ|2γ |φ(λ)|2 dλ
n R
R
R
1
2γ
≤ c2
+ hn .
n
In the case that {Xi }ni=1 are i.i.d. random variables, the term (32) equals zero and hence E [Tn (hn )]−
R 2
2γ
R f (x) dx ≤ c2 hn . This is consistent with the result using convolution method as in Giné and
Nickl (2008).
Step 2. We give the decomposition for
Tn (hn ) − E Tn (hn ).
Using (31), we can obtain the following decomposition
Tn (hn ) − E Tn (hn ) = 2Nn + Dn − E [Dn ],
(33)
where
n
1 X
Nn =
2πn
Z
ιλXi
b
− φ(λ) φ(−λ) dλ
K(λh
n) e
R
i=1
and
1
Dn =
2πn(n − 1)
X
Z
ιλXi
b
K(λh
− φ(λ) e−ιλXj − φ(−λ) dλ.
n) e
1≤i6=j≤n R
Step 3. Here we estimate E |Dn |2 .
We first assume that {Xn } is m-dependent, that is, a0 6= 0, am 6= 0 and ai = 0 for all i > m.
In this case,
X Z
2
−ιλX
1
2
ιλXi
j
b
E |Dn | ≤ 2
E
K(λh
)
e
−
φ(λ)
e
−
φ(−λ)
dλ
n
n (n − 1)2
j−i>4m R
X Z
2
1
ιλXi
b
+ 2
E
K(λh
− φ(λ) e−ιλXj − φ(−λ) dλ .
(34)
n) e
2
n (n − 1)
R
0<j−i≤4m
21
For the first expectation on the right hand side of (34),
X Z
2
ιλXi
b
E
K(λh
− φ(λ) e−ιλXj − φ(−λ) dλ
n) e
j−i>4m R
Z
≤
R2
X
b 1 hn )||K(λ
b 2 hn )|
|K(λ
E (eιλ1 Xi1 − φ(λ1 ))(e−ιλ2 Xi2 − φ(−λ2 ))
|i1 −i2 |≤m
X
×
E (e−ιλ1 Xj1 − φ(−λ1 ))(eιλ2 Xj2 − φ(λ2 )) dλ1 dλ2
|j1 −j2 |≤m
Z
≤
R2
b 1 hn )||K(λ
b 2 hn )|
|K(λ
h
X
i2
|E eιλ1 Xi1 −ιλ2 Xi2 | + |φ(λ1 )φ(−λ2 )| dλ1 dλ2
|i1 −i2 |≤m
Z
h
i
b 1 hn )||K(λ
b 2 hn )| |φ(λ1 − λ2 )|2 + |φε (λ1 a0 )φε (λ2 am )|2 + |φ(λ1 )φ(λ2 )|2 dλ1 dλ2
|K(λ
2
Z
ZR
2
2
2
2
b 2 hn )|2 |φ(λ1 − λ2 )|2 dλ1 dλ2
b
|K(λ
|K(λ1 hn )| |φ(λ1 − λ2 )| dλ1 dλ2 + c4 n
≤ c4 n
2
2
R Z
ZR
2
2
2
+ c4 n
|φε (λ1 a0 )φε (λ2 am )| dλ1 dλ2 + c4 n
|φ(λ1 )φ(λ2 )|2 dλ1 dλ2
≤ c3 n2
R2
≤ c5
n2
hn
R2
,
(35)
where in the last inequality we used the Plancherel theorem.
Set Kn (x) = h1n K( hxn ). Then
X Z
ιλXi
b
E
− φ(λ) e−ιλXj − φ(−λ) dλ
K(λh
n) e
2
0<j−i≤4m R
≤ c6 n2
≤ c7 n
2
4m
X
k=1
4m
X
Z
ιλX1
b
K(λh
− φ(λ) e−ιλX1+k − φ(−λ) dλ
n) e
E
R
E Kn2 (X1
2
Z
2
ιλX1
b
K(λh
φ(−λ) dλ
n) e
− X1+k ) + c7 n E
R
k=1
+ c7 n2
2
4m
X
k=1
Z
Z
2
−ιλX1+k
b
K(λh
φ(λ) dλ + c7 n2
n) e
E
R
2
2
b
K(λh
n )|φ(λ)| dλ .
R
Hence
E
Z
X
ιλXi
b
K(λh
− φ(λ) e−ιλXj − φ(−λ) dλ
n) e
2
0<j−i≤4m R
≤ c8 n
2
Z
c2 (λ)||φε (λa0 )||φε (λam )| dλ + c8 n2
|K
n
R
Z
2
2
+ c8 n
|φ(λ)|2 dλ
Z
2
b
|K(λh
n )| dλ
Z
R
|φ(λ)|2 dλ
R
R
n2
≤ c9
,
hn
(36)
c2 (λ)| ≤
where in the last inequality we used the fact that |K
n
Combining (34), (35) and (36) gives
E |Dn |2 ≤
22
c10
.
n 2 hn
1
hn
R
RK
2 (x) dx
for all λ ∈ R.
In the sequel, we can assume that a0 6= 0 and ai 6= 0 for infinite many i ∈ N. Let p1 = min{i ∈
N : ai 6= 0} and p2 = min{i ∈ N : ai 6= 0, i > p1 }. Note that
kDn k2 ≤ I0 + I1 + · · · + Ip1 + Ip1 +1 ,
(37)
where
Z
1
Ip =
2πn(n − 1)
X
X
1≤i6=j≤n
k≤i, `≤j
|k−i|+|`−j|=p
b
K(λh
n)
R
Pk H(Xi )(λ)P` H(Xj )(−λ) dλ ,
2
for 0 ≤ p ≤ p1 and
Ip1 +1
Z
1
=
2πn(n − 1)
X
X
1≤i6=j≤n
k≤i, `≤j
|k−i|+|`−j|≥p1 +1
b
K(λh
n)
R
We first estimate I0 . Note that
Z
1
b
I0 =
K(λh
n)
2πn(n − 1) R
X
Pk H(Xi )(λ)P` H(Xj )(−λ) dλ .
2
Pi H(Xi )(λ)Pj H(Xj )(−λ) dλ ,
2
1≤i6=j≤n
and
I02
Z
c11
≤ 4
n
b
K(λh
n)
R
+
X
j−i=1
Z
c11
n4
Pi H(Xi )(λ)Pj H(Xj )(−λ) dλ
b
K(λh
n)
R
X
2
2
Pi H(Xi )(λ)Pj H(Xj )(−λ) dλ
j−i=2
2
2
Z
+ ··· +
+
c11
n4
X
c11
b
K(λh
Pi H(Xi )(λ)Pj H(Xj )(−λ) dλ
n)
4
n
R
j−i=p2
Z
X
2
b
K(λh
Pi H(Xi )(λ)Pj H(Xj )(−λ) dλ
n)
R
2
2
2
j−i≥p2 +1
:= I0,1 + I0,2 + · · · + I0,p2 + I0,p2 +1 .
Recall properties of the projection operator Pk defined in (29). It is easy to see that, for
1 ≤ q ≤ p2 ,
I0,q ≤
=
c12
n3
Z
c12
n3
Z
R2
R2
b 1 hn )||K(λ
b 2 hn )|
|K(λ
∞
Y
E [eι(λ2 −λ1 )(ak+q −ak )εj−k−q ] dλ1 dλ2
k=1
b 1 hn )||K(λ
b 2 hn )|
|K(λ
∞
Y
|φε ((λ2 − λ1 )(ak+q − ak ))| dλ1 dλ2 .
k=1
Since lim ai = 0 and ai 6= 0 for infinite many i ∈ N, there exist infinite many indices k such that
i→∞
ak+q 6= ak . Doing some simple calculations then gives
I0,q ≤
c13
,
n3 hn
23
for 1 ≤ q ≤ p2 . To estimate the last term in the upper bound of I02 , observe that
Z
c11
b 1 hn )K(−λ
b
I0,p2 +1 = 4
K(λ
2 hn )
n R2
X
E [Pi1 H(Xi1 )(λ1 )Pi2 H(Xi2 )(−λ2 )Pj H(Xj )(−λ1 )Pj H(Xj )(λ2 )] dλ1 dλ2
j−i1 ≥p2 +1, j−i2 ≥p2 +1
c11
= 4
n
Z
Z
c11
b
b
b 1 hn )K(λ
b 2 hn )I2 (λ1 , λ2 ) dλ1 dλ2
K(λ1 hn )K(−λ2 hn )I1 (λ1 , λ2 ) dλ1 dλ2 + 4
K(λ
n R2
R2
Z
c11
b 1 hn )K(−λ
b
+ 4
K(λ
2 hn )I3 (λ1 , λ2 ) dλ1 dλ2 ,
n R2
where
X
I1 (λ1 , λ2 ) =
E [Pi1 H(Xi1 )(λ1 )Pi2 H(Xi2 )(−λ2 )Pj H(Xj )(−λ1 )Pj H(Xj )(λ2 )] ,
j−i1 ≥p2 +1, j−i2 ≥p2 +1, i2 >i1
X
I2 (λ1 , λ2 ) =
E [Pi1 H(Xi1 )(λ1 )Pi2 H(Xi2 )(−λ2 )Pj H(Xj )(−λ1 )Pj H(Xj )(λ2 )] ,
j−i1 ≥p2 +1, j−i2 ≥p2 +1, i1 <i2
X
I3 (λ1 , λ2 ) =
E [Pi1 H(Xi1 )(λ1 )Pi2 H(Xi2 )(−λ2 )Pj H(Xj )(−λ1 )Pj H(Xj )(λ2 )] .
j−i1 ≥p2 +1, j−i2 ≥p2 +1, i2 =i1
Note that
X
|I1 (λ1 , λ2 )| ≤
|φε (ap1 (λ1 − λ2 ))| |φε (ap2 (λ1 − λ2 ))|
j−i1 ≥p2 +1, j−i2 ≥p2 +1, i2 >i1
i
h
× E (e−ιλ2 a0 εi2 − φε (−λ2 a0 ))(e−ι(λ1 −λ2 )aj−i2 εi2 − φε (−(λ1 − λ2 )aj−i2 )
X
|φε ((λ1 − λ2 )ap1 )| |φε ((λ1 − λ2 )ap2 )|
≤
j−i1 ≥p2 +1, j−i2 ≥p2 +1, i2 >i1
× e−ι(λ1 −λ2 )aj−i2 εi2 − φε (−(λ1 − λ2 )aj−i2 )
2
X
|φε ((λ1 − λ2 )ap1 )| |φε ((λ1 − λ2 )ap2 )| |λ1 − λ2 |γ |aj−i2 |γ
≤ c16
≤
j−i1 ≥p2 +1, j−i2 ≥p2 +1, i2 >i1
c17 n2 |λ1 − λ2 |γ |φε ((λ1 − λ2 )ap1 )| |φε ((λ1
− λ2 )ap2 )| .
Similarly,
|I2 (λ1 , λ2 )| ≤ c18 n2 |λ1 − λ2 |γ |φε ((λ1 − λ2 )ap1 )||φε ((λ1 − λ2 )ap2 )|.
Finally,
|I3 (λ1 , λ2 )| ≤
≤
X
E [Pi H(Xi )(λ1 )Pi H(Xi )(−λ2 )Pj H(Xj )(−λ1 )Pj H(Xj )(λ2 )]
j−i≥p2 +1
c19 n2 |φε ((λ1
− λ2 )ap1 )||φε ((λ1 − λ2 )ap2 )|.
So
I0,p2 +1
Z
c20
b 1 hn )||K(λ
b 2 hn )|(1 + |λ1 − λ2 |γ )|φε ((λ1 − λ2 )ap )||φε ((λ1 − λ2 )ap )| dλ1 dλ2
≤ 2
|K(λ
1
2
n R2
c21
≤ 2 .
n hn
24
Therefore,
I02 ≤ I0,1 + I0,2 + · · · + I0,p2 + I0,p2 +1 ≤
c22
.
n2 hn
(38)
Using similar arguments, for 1 ≤ p ≤ p1 , we can obtain
Ip2 ≤
c23
.
n 2 hn
(39)
We now estimate Ip1 +1 . Note that
1
(2π)2 n2 (n − 1)2
Ip21 +1 =
Z
R2
b 1 hn )K(−λ
b
K(λ
2 hn ) Jn (λ1 , λ2 ) dλ1 dλ2 ,
where
Jn (λ1 , λ2 ) =
i
h
E Pk1 H(Xi1 )(λ1 )Pk2 H(Xj1 )(−λ1 )Pk3 H(Xi2 )(−λ2 )Pk4 H(Xj2 )(λ2 ) .
X
i1 6=j1 , i2 6=j2
k1 ≤i1 , k2 ≤j1 , k3 ≤i2 , k4 ≤j2
|k1 −i1 |+|k2 −j1 |≥p1 +1
|k3 −i2 |+|k4 −j2 |≥p1 +1
Recall properties of the projection operator Pk defined in (29). By Cauchy-Schartz inequality and
the condition (3), we can obtain
|Jn (λ1 , λ2 )| ≤ c24 |λ1 |2γ |λ2 |2γ |φε (λ1 a0 )|2 + |φε (λ1 ap1 )|2 |φε (λ2 a0 )|2 + |φε (λ2 ap1 )|2
X
|ai1 −k1 |γ |aj1 −k2 |γ |ai2 −k3 |γ |aj2 −k4 |γ . (40)
×
i1 6=j1 , i2 6=j2
k1 ≤i1 , k2 ≤j1 , k3 ≤i2 k4 ≤j2
|k1 −i1 |+|k2 −j1 |≥p1 +1, |k3 −i2 |+|k4 −j2 |≥p1 +1
at least two k indices are identical,
identical k indices are greater than the others
There are only four possibilities for the k indices in the above summation: 1) all the k indices
are identical; 2) three k indices are identical and strictly larger than the last one; 3) two pairs of
different identical k indices; 4) one pair of identical k indices and strictly larger than the remaining
two different indices. For the first three possibilities, the summation in (40) is less than a constant
∞
P
multiple of n2 since
|ai |γ < ∞. For the last possibility, the summation in (40) is less than a
i=0
constant multiple of n2 ηn,γ where ηn,γ =
n P
∞
P
|ai |γ . Therefore,
`=0 i=`
Ip21 +1 ≤
c25 ηn,γ
n2
2
2γ
2
2
b
|K(λh
)|
|λ|
|φ
(λa
)|
+
|φ
(λa
)|
dλ .
n
ε
0
ε
p1
Z
R
Putting inequalities (37), (38), (39) and (41) together gives
ηn,γ
1
2
+ 2 .
kDn k2 ≤ c26
n2 hn
n
Step 4. Here we estimate
E |Nn − N n |2 ,
where
Nn =
1
2π
Z
b
K(0)
φn (λ) − φ(λ) φ(−λ) dλ.
R
25
(41)
Note that
E |Nn − N n |2 = E
1
2π
Z
b
b
K(λh
n ) − K(0) φn (λ) − φ(λ) φ(−λ) dλ
R
Z
2
ιλX1
b
b
K(λhn ) − K(0) e
φ(−λ) dλ
E
1
(2π)2 n
R
i2
1 h
= E Kn ∗ f (X1 ) − f (X1 ) ,
n
≤
where Kn (x) = h1n K( hxn ).
Since f is bounded,
2
E |Nn − N n |
Z
2
c27
Kn ∗ f (x) − f (x) dx
≤
n R
Z
c27
2
2
b
b
=
|K(λh
n ) − K(0)| |φ(λ)| dλ
(2π)2 n R
Z
c27
2
2
b
b
|K(λh
≤
n ) − K(0)| |φε (a0 λ)| dλ
(2π)2 n R
c28 h2γ
n
≤
,
n
where in the first equality we used the Plancherel theorem.
It is easy to see that
Z
n
1X
2
Nn =
f (Xi ) −
f (x) dx .
n
R
i=1
Combining all the above results gives
n
1X
E Tn (hn ) − E Tn (hn ) −
Yi
n
2
≤ c29
i=1
ηn,γ
h2γ
1
n
+
+
2
2
n hn
n
n
!
.
Step 5. We show the central limit theorem for
n
1 X
√
f (Xi ) −
n
i=1
Z
f 2 (x) dx .
R
It suffices to consider the case when the Xi are dependent. Note that
Z
P0 f (Xi ) −
f 2 (x) dx
R
= E [f (Xi )|F0 ] − E [f (Xi )|F−1 ]
i−1
∞
P
P
Z
i
λaj εi−j +ι
λaj εi−j h
ι
1
ιλai ε0
ιλai ε00
j=i+1
j=0
=
E
φ(λ) e
e
−e
dλ F0 ,
2π
R
where {ε0i : i ∈ Z} is an independent copy of {εi : i ∈ Z}.
26
2
Then
P0 f (Xi ) −
Z
2
f (x) dx
R
ι
Z
2
≤
φ(λ) e
ZR
≤
∞
P
λaj εi−j +ι
j=i+1
|φ(λ)||eιλai ε0 − e
i−1
P
j=0
ιλai ε00
λaj εi−j
| dλ
R
Z
0
eιλai ε0 − eιλai ε0 dλ
2
|φ(λ)||eιλai ε0 − φε (λai )| dλ
RZ
|λ|γ |φ(λ)| dλ
≤ 2|ai |γ
RZ
|λ|γ |φε (λ)|2 dλ.
≤ c30 |ai |γ
≤2
2
2
R
Now by Lemma 1 in Wu (2006), we can easily obtain
n
1 X
√
f (Xi ) −
n
i=1
Z
d
f 2 (x) dx −→ N (0, σ 2 )
R
for some σ 2 ∈ (0, ∞). Using similar arguments as in Step 3, we can show that
sup
n
where Sn =
Hence,
n
P
i=1
σ2 =
f (Xi ) −
R
Rf
2 (x) dx
1
E |Sn |4 < ∞,
n2
.
lim n−1 Var (Sn ). This completes the proof.
n→∞
Proof of Theorem 2.2. The proof of this theorem is similar to that of Theorem 2.1. We only
need to modify the Step 3 in the proof of Theorem 2.1 and consider the dependent case. Using
the same notations as in the proof of Theorem 2.1 and then Lemma 7.2,
Z
√
√
2
b
n E |Dn | ≤ n
|K(λh
dλ
n )|E |φn (λ) − φ(λ)|
R
Z
i
h
1
ιλX1
ιλX1 2
b
+√
|K(λh
)|E
|e
−
E
e
|
dλ
n
n R
Z
Z
c1
c1
b
≤√
|K(λ)|
dλ + √
|λ|2γ |φε (λa0 )|2 dλ.
nhn R
n R
This completes the proof.
At the end, we give two useful lemmas which are related to the sufficient and necessary conditions for the assumptions in Theorems 2.1 and 2.2.
Lemma 7.3 Let φε (λ) = E [eιλε1 ] for all λ ∈ R. For any γ ∈ [0, 1], if E |ε1 |2γ < ∞, then there
exists a positive constant cγ,2 such that
E |eιλε1 − φε (λ)|2 ≤ cγ,2 |λ|2γ .
Moreover, if E |ε1 |4γ < ∞, there exists a positive constant cγ,4 such that
E |eιλε1 − φε (λ)|4 ≤ cγ,4 |λ|4γ .
27
Proof. This first inequality follows from
E |eιλε1 − φε (λ)|2 = 1 − |φε (λ)|2 ≤ 1 − [E cos(λε1 )]2 ≤ c1 |λ|2γ .
For the second inequality,
E |eιλε1 − φε (λ)|4
2
= E 1 − eιλε1 φε (−λ) − e−ιλε1 φε (λ) + |φε (λ)|2
= 1 + φε (2λ)φ2ε (−λ) + φε (−2λ)φ2ε (λ) − 3|φε (λ)|4
= 1 + 2E cos(2λε1 ) E 2 cos(λε1 ) − E 2 sin(λε1 ) + 4E sin(2λε1 )E cos(λε1 )E sin(λε1 )
2
− 3 E 2 cos(λε1 ) + E 2 sin(λε1 )
λε1 2
= 1 + 2 1 − 2E sin2 (λε1 ) (1 − 2E sin2 (
)) − E 2 sin(λε1 )
2
λε
λε1
1
+ 8E sin(λε1 )(1 − 2 sin2 (
)) E sin(λε1 )[1 − 2E sin2 (
)]
2
2
2
λε1 2
− 3 (1 − 2E sin2 (
)) + E 2 sin(λε1 )
2
4γ
≤ c2 |λ| .
This completes the proof.
Lemma 7.4 Let φε (λ) = E [eιλε1 ] for all λ ∈ R. If ε1 has non-degenerate distribution, i.e., it is
not equal to a constant almost surely, then in the inequalities
E |eιλε1 − φε (λ)|2 ≤ cγ,2 (|λ|2γ ∧ 1)
and
E |eιλε1 − φε (λ)|4 ≤ cγ,4 (|λ|4γ ∧ 1),
the range γ ∈ (0, 1] is optimal.
Proof. By Cauchy-Schwarz inequality, we only need to show that the range γ ∈ (0, 1] is optimal
for the first inequality. Suppose E |eιλε1 −φε (λ)|2 = 1−|φε (λ)|2 ≤ c(|λ|2γ ∧1) for some γ > 1. Since
ε2 is an independent copy of ε1 , let φε1 −ε2 (λ) = E [eιλ(ε1 −ε2 ) ], then 1 − φε1 −ε2 (λ) = 1 − |φε (λ)|2 ≤
c(|λ|2γ ∧ 1) for some γ > 1.
Note that
E |eιλ(ε1 −ε2 ) − 1|2 = 1 − φε1 −ε2 (λ) + 1 − φε1 −ε2 (−λ) = 2(1 − φε1 −ε2 (λ)).
Therefore, for λ close enough to 0,
1
2
|φε1 −ε2 (x + λ) − φε1 −ε2 (x)| ≤ E |eιλε1 −ε2 − 1| ≤ E |eιλε1 −ε2 − 1|2
≤ c1 |λ|γ .
Since γ > 1, φ0ε1 −ε2 (x) = 0 for all x ∈ R. That is, |φε (x)|2 = φε1 −ε2 (x) = 1 for all x ∈ R or ε1
equals a constant almost surely. So the γ must be less than or equal to 1.
Acknowledgement
We would like to thank the two referees and the Associate Editor for their valuable comments.
We would also like to thank Professor Xia Chen for very helpful discussions. F. Xu is partially
supported by National Natural Science Foundation of China (Grant No.11401215), Shanghai Pujiang Program (14PJ1403300), 111 Project (B14019) and Natural Science Foundation of Shanghai
(16ZR1409700).
28
References
[1] Ahmad IA. 1979. Strong consistency of density estimation by orthogonal series methods for
dependent variables with applications. Ann. Inst. Statist. Math. 31: 279–288.
[2] Bailey WN. 1935. General Hypergeometric Series. Cambridge Tracts in Mathematics and
Mathematical Physics.
[3] Bickel JP and Ritov Y. 1988. Estimating integrated squared density derivatives: Sharp best
order of convergence estimates. Sankhyā Ser. A 50: 381–393.
[4] Bondon P and Palma W. 2007. A class of antipersistent processes. J. Time Series Anal. 28:
261-273.
[5] Devroye L and Györfi L. 1985. Nonparametric Density Estimation: The L1 View. New York:
Wiley.
[6] Faÿ G, Moulines E, Roueff F and Taqqu MS. 2009. Estimators of long-memory: Fourier versus
wavelets. J. Econometrics 151: 159-177.
[7] Gauss F. 1866. Disquisitiones generales circa sericm infinitam, Ges. Werke 3: 123-163 and
207-229.
[8] Giné E and Nickl R. 2008. A simple adaptive estimator of the integrated square of a density.
Bernoulli 14: 47–61.
[9] Gretton A, Borgwardt KM, Rasch MJ, Schölkopf B and Smola A. 2012. A kernel two-sample
test. J. Mach. Learn. Res. 13: 723–773.
[10] Hall P and Marron JS. 1987. Estimation of integrated squared density derivatives. Statist.
Probab. Lett. 6: 109–115.
[11] Hipel KW and McLeod AI. 1994. Time Series Modelling of Water Resources and Environmental Systems. Elsevier.
[12] Honda T. 2000. Nonparametric density estimation for a long-range dependent linear process.
Ann. Inst. Statist. Math. 52: 599–611.
[13] Hsing T. 1999. On the asymptotic distributions of partial sums of functionals of infinitevariance moving averages. Ann. Probab. 27: 1579–1599.
[14] Källberg D, Leonenko N and Seleznjev O. 2012. Statistical inference for Rényi entropy functionals. Lecture Notes in Comput. Sci. 7260: 36–51.
[15] Källberg D, Leonenko N and Seleznjev O. 2014. Statistical estimation of quadratic Rényi
entropy for a stationary m-dependent sequence. Journal of Nonparametric Statistics 26: 385–
411.
[16] Krishnamurthy A, Kandasamy K, Poczós B and Wasserman L. 2015. On estimating L22 divergence. Proceedings of the 18th International Conference on Artificial Intelligence and Statistics
(AISTATS), San Diego, CA, USA. JMLR: W&CP 38.
[17] Laurent B. 1996. Efficient estimation of integral functionals of a density. Ann. Statist. 24:
659–681.
[18] Laurent B. 1997. Estimation of integral functionals of a density and its derivatives. Bernoulli
3: 181–211.
29
[19] Leonenko N, Pronzato L and Savani V. 2008. A class of Rényi information estimators for
multidimensional densities. Ann. Statist. 36: 2153–2182. Corrections, Ann. Statist. 38 (2010),
3837–3838.
[20] Leonenko N and Seleznjev O. 2010. Statistical inference for the ε-entropy and the quadratic
Rényi entropy. J. Multivariate Anal. 101: 1981–1994.
[21] Mittnik S, Rachev ST and Paolella MS. 1998. Stable paretian modeling in finance: Some
empirical and theoretical aspects. In : Adler, R. J., Feldman, R.E., Taqqu, M.S., eds. A
Practical Guide to Heavy Tails. Birkhäuser: 79–110.
[22] Nadaraya EA. 1989. Nonparametric Estimation of Probability Densities and Regression
Curves. Kluwer Academic Pub.
[23] Parzen E. 1962. On estimation of a probability density and mode. Ann. Math. Stat. 31: 1065–
1079.
[24] Penrose M and Yukich JE. 2013. Limit theory for point processes in manifolds. Ann. Appl.
Probab. 23: 2161–2211.
[25] Rényi A. 1970. Probability Theory Amsterdam: North-Holland.
[26] Rosenblatt M. 1956. Remarks on some nonparametric estimates of a density function, Ann.
Math. Statist. 27: 832–837.
[27] Sang H and Sang Y. 2017. Memory properties of transformations of linear processes. Stat.
Inference Stoch. Process. 20: 79-103.
[28] Schimek MG. 2000. Smoothing and Regression : Approaches, Computation, and Application.
John Wiley & Sons.
[29] Scott DW. 2015. Multivariate Density Estimation: Theory, Practice, and Visualization 2.
John Wiley & Sons.
[30] Shannon CE. 1948. A mathematical theory of communication. Bell Syst. Tech. Jour. 27:
379–423, 623–656.
[31] Silverman BW. 1986. Density Estimation for Statistics and Data Analysis. London: Chapman
and Hall.
[32] Tran LT. 1992. Kernel density estimation for linear processes. Stochastic Process. Appl. 41:
281–296.
[33] Wand MP and Jones MC. 1995. Kernel Smoothing. London: Chapman and Hall.
[34] Wu WB. 2006. Unit root testing for functionals of linear processes. Econometric Theory 22:
1–14.
[35] Wu WB, Huang Y and Huang Y. 2010. Kernel estimation for time series: An asymptotic
theory. Stochastic Process. Appl. 120: 2412–2431.
[36] Wu WB and Mielniczuk J. 2002. Kernel density estimation for linear processes. Ann. Statist.
30: 1441–1459.
[37] Zhou Z. 2014. Inference of weighted V-statistics for non stationary time series and its applications. Ann. Statist. 42: 87–114.
30
| 10math.ST
|
A fast marching algorithm for the factored eikonal equation
Eran Treistera,∗, Eldad Habera,b
a Department
of Earth and Ocean Sciences, The University of British Columbia, Vancouver, BC, Canada.
of Mathematics, The University of British Columbia, Vancouver, BC, Canada.
b Department
arXiv:1607.00973v3 [cs.CE] 27 Aug 2016
Abstract
The eikonal equation is instrumental in many applications in several fields ranging from computer vision to
geoscience. This equation can be efficiently solved using the iterative Fast Sweeping (FS) methods and the
direct Fast Marching (FM) methods. However, when used for a point source, the original eikonal equation
is known to yield inaccurate numerical solutions, because of a singularity at the source. In this case, the
factored eikonal equation is often preferred, and is known to yield a more accurate numerical solution. One
application that requires the solution of the eikonal equation for point sources is travel time tomography.
This inverse problem may be formulated using the eikonal equation as a forward problem. While this problem
has been solved using FS in the past, the more recent choice for applying it involves FM methods because
of the efficiency in which sensitivities can be obtained using them. However, while several FS methods are
available for solving the factored equation, the FM method is available only for the original eikonal equation.
In this paper we develop a Fast Marching algorithm for the factored eikonal equation, using both first and
second order finite-difference schemes. Our algorithm follows the same lines as the original FM algorithm
and requires the same computational effort. In addition, we show how to obtain sensitivities using this FM
method and apply travel time tomography, formulated as an inverse factored eikonal equation. Numerical
results in two and three dimensions show that our algorithm solves the factored eikonal equation efficiently,
and demonstrate the achieved accuracy for computing the travel time. We also demonstrate a recovery
of a 2D and 3D heterogeneous medium by travel time tomography using the eikonal equation for forward
modeling and inversion by Gauss-Newton.
Keywords: eikonal equation, Factored eikonal equation, Fast Marching, First arrival, Travel Time
Tomography, Gauss Newton, Seismic imaging.
1. Introduction
The eikonal equation appears in many fields, ranging from computer vision [30, 31, 33, 11], where it is
used to track evolution of interfaces, to geoscience [19, 12, 22, 14, 24] where it describes the propagation of
the first arrival of a wave in a medium. The equation has the form
|∇τ |2 = κ(~x)2 ,
(1.1)
where | · | is the Euclidean norm. In the case of wave propagation, τ is the travel time of the wave and
κ(~x) is the slowness (inverse velocity) of the medium. The value of τ is usually given at some sub-region.
For example, in this work we assume the wave propagates from a point source at location ~x0 , for which the
travel time is 0, and hence τ (~x0 ) = 0.
Equation (1.1) is nonlinear, and as such may have multiple branches in its solution. One of these branches,
which is the one of interest in the applications mentioned earlier, corresponds to the ”first-arrival” viscosity
solution, and can be calculated efficiently [5, 25]. One of the ways to compute it is by using the Fast Marching
∗ Corresponding
author
Email addresses: [email protected] (Eran Treister), [email protected] (Eldad Haber)
Preprint submitted to Elsevier
August 30, 2016
Figure 1: The l2 norm of the approximation error |∇τ0 − Dτ0 | around a source point at [0.5,0.5], where τ0 is the distance
function and D is a central difference gradient operator on a mesh with hx = hy = 0.01.
(FM) methods [36, 30, 31], which solve it directly using first or second order schemes in O(n log n) operations.
These methods are based on the monotonicity of the solution along the characteristics. Alternatively, (1.1)
can be solved iteratively by Fast Sweeping (FS) methods, which may be seen as Gauss-Seidel method for
(1.1). First order accurate solutions of (1.1) can be obtained very efficiently in 2d Gauss-Seidel sweeps
in O(n) operations, where d is the dimension of the problem [35, 38]. An alternative for the mentioned
approaches is to use FS to solve a Lax-Friedrichs approximation for (1.1), which involves adding artificial
viscosity to the original equation [10]. This approach was suggested for general Hamilton-Jacobi equations,
and is simple to implement. In [23], such Lax-Friedrichs approximation is obtained using FS up to third
order accuracy using the weighted essentially non-oscillatory (WENO) approximations to the derivatives.
For a performance comparison between some of the mentioned solvers see [7].
In some cases, the eikonal equation (1.1) is used to get a geometrical-optics ansatz of the solution of the
Helmholtz equation in the high frequency regime [12, 15, 19, 17]. This is done using the Rytov decomposition
of the Helmholtz solution: u(~x) = a(~x) exp(iωτ (~x)), where a(~x) is the amplitude and τ (~x) is the travel time.
This approach involves solving (1.1) for the travel-time and solving the transport equation
1
1
∇τ · ∇a + a∆τ = (∇τ · ∇a + ∇ · (a∇τ )) = 0
2
2
(1.2)
for the amplitude [15, 19]. The resulting approximation includes only the first arrival information of the
wave propagation. Somewhat similarly, the work of [9] suggests using the eikonal solution to get a multigrid
preconditioner for solving linear systems arising from discretization of the Helmholtz equation.
In many cases in seismic imaging, the eikonal equation is used for modeling the migration of seismic
waves from a point source at some point ~x0 . In this case, when solving (1.1) numerically by standard finite
differences methods, the obtained solution has a strong singularity at the location of the source, which leads
to large numerical errors [22, 6]. Figure 1 illustrates this phenomenon by showing the approximation error
for the gradient of the distance function, which is the solution of (1.1) for κ = 1. It is clear that the largest
approximation error for the gradient is located around the source point, and that its magnitude is rather
large. In addition, it is observed that although τ may have more singularities in other places, the singularity
at the source is more damaging and polluting for the numerical solution [22, 6].
A rather easy treatment to the described phenomenon is suggested in [21, 6], and achieved using a factored
version of (1.1). That is, we define
τ = τ0 τ1 ,
(1.3)
where τ0 is the distance function, τ0 = k~x − ~x0 k2 , from the point source—the analytical solution for (1.1) in
the case where κ(~x) = 1 is a constant. Indeed, at the location of the source, the function τ0 is non-smooth.
However, the computed factor τ1 is expected to be very smooth at the surrounding of the source, and can be
approximated up to high order of accuracy [18]. Plugging (1.3) into (1.1) and applying the chain rule yields
the factored eikonal equation
τ02 |∇τ1 |2 + 2τ0 τ1 ∇τ0 · ∇τ1 + τ12 = κ(~x)2 .
(1.4)
2
Similarly to it original version, Equation (1.4) can be solved by fast sweeping methods in first order accuracy
[6, 16, 18], or by a Lax-Friedrichs scheme up to third order of accuracy [19, 15, 18]. The recent works
[18, 20] suggest hybrid schemes where the factored eikonal is solved at the neighborhood of the source and
the standard eikonal, which is computationally easier, is solved in the rest of the domain.
One geophysical tool that fits the scenario described earlier is travel time tomography. One way to
formulate it is by using the eikonal equation as a forward problem inside an inverse problem [29]. To solve
the inverse problem, one should be able to solve (1.1) accurately for a point source, and to compute its
sensitivities efficiently. The works of [13, 34] computes the tomography by FS, and require an FS iterative
solution for computing the sensitivities. The more recent [14] uses the FM algorithm for forward modeling
using the non-factored eikonal equation, because this way the sensitivities are obtained more efficiently by
a simple solution of a lower triangular linear system. [3] suggests to use FS for forward modeling using the
factored equation, but also efficiently obtain the sensitivities by approximating them using FM with the
non-factored eikonal equation.
In this paper, we develop a Fast Marching algorithm for the factored eikonal equation (1.4), based on
[30, 31]. As in [31], our algorithm is able to solve (1.4) using first order or second order schemes, in guaranteed
O(n log n) running time. When using our method for forward modeling in travel time tomography, one
achieves both worlds: (1) have an accurate forward modeling based on the factored eikonal equation, and
(2) obtain the sensitivities of the (factored) forward modeling efficiently, by solving lower triangular linear
systems in O(n) operations. Computationally, this is one of the most attractive ways to solve the inverse
problem, since the cost of the inverse problem can be governed by the cost of applying the sensitivities.
Our paper is organized as follows. In the next section we briefly review the FM method in [30, 31],
including some of its implementation details. Next, we show our extension to the FM method for the factored eikonal equation—in both first and second order of accuracy—and provide some theoretical properties.
Following that, we discuss the derivation of sensitivities using FM and briefly present the travel time tomography problem. Last, we show some numerical results that demonstrate the effectiveness of the method in
two and three dimensions, for both the forward and inverse problems.
2. The Fast Marching algorithm
We now review the FM algorithm of [30, 31] in two dimensions. The extension to higher dimensions is
straightforward. The FM algorithm is based on the Godunov upwind discretization [25] of (1.1). In two
dimensions, this discretization is given by
−y
+y
−x
+x
|∇τ |2 ≈ max{Dij
τ, −Dij
τ, 0}2 + max{Dij
τ, −Dij
τ, 0}2 = κ(~xij )2 , ~xij ∈ Ωh ,
(2.5)
τ
−τ
τ
−τ
+x
−x
τ = i+1,jh i,j are the backward and forward first
where in the simplest form Dij
τ = i,j hi−1,j and Dij
derivative operators, respectively. In principal, one can replace these operators with ones of higher order of
accuracy.
The FM algorithm solves (2.5) in a sophisticated way, exploiting the fact that the upwind difference
structure of (2.5) imposes a unique direction in which the information propagates—from smaller values of τ
to larger values. Hence, the FM algorithm rests on solving (2.5) by building the solution outwards from the
smallest τ value. It assumes that some initial value of τ is given at some region of Ωh (or a point ~x0 ) and it
propagates outwards from this initial region, by updating the next smallest value of τ at each step.
To apply the rule above, let us define three disjoint sets of variables: the known variables, the front variables
(which are sometimes called the trial variables) and the unknown variables. These three sets together contain
all the grid points in the problem. For simplicity, let us assume that we solve (2.5) for a point source. That
is, a source is located at point ~x0 , for which τ (~x0 ) = 0. Initially, known is chosen as an empty set, front is
set to contain only ~x0 , and unknown has the rest of the variables for which τ is set to infinity. At each step
we choose the point ~xij in front with minimal value of τ (~xij ) and move it to known . Next, we move all of
its neighbors which are in unknown to front , and solve (2.5) for all neighbors which are not in known . This
3
Algorithm: Fast Marching
Initialize:
τij = ∞ for all ~xij ∈ Ωh , τ (~x0 ) = 0,
known ← ∅, front ← {~x0 }.
while front 6= ∅ do
1. Find the minimal entry in front :
~ximin ,jmin = arg min~xij {τij : ~xij ∈front }
2. Add ~ximin ,jmin to known and take it out of front :
front ←front \{~ximin ,jmin } ; known ←known ∪{~ximin ,jmin }.
3. Add the unknown neighborhood of ~ximin ,jmin to front :
Nmin = {~ximin −1,jmin , ~ximin +1,jmin , ~ximin ,jmin −1 , ~ximin ,jmin +1 }\known
front ←front ∪ Nmin .
4. Foreach ~xij ∈ Nmin
Update τij by solving the quadratic (2.5), using only entries in known .
End
end
Figure 2: A minimum heap and its implementation using array.
way, we set all variables to be in known in n steps, and the algorithm finishes. A precise description of the
algorithm is given in Algorithm 1.
In [30] it was proved that Algorithm 1 produces a viable viscosity solution to (2.5) when using first order
approximations for the first derivatives. Furthermore, it is proved that the values of τ in the order of which
the points are set to known in Step 2 are monotonically increasing.
2.1. Efficient implementation using minimum heap
Algorithm (1) has two main computational bottlenecks in Steps 1 and 4, which are repeated n times. For
d−1
a d-dimensional problem, the set front contains a d-1 dimensional manifold of points, of size O(n d ). To
find the minimum of front efficiently, a minimum heap data structure is used [30, 31]. A minimum heap is
a binary tree with a property that a value at any node is less than or equal to the values at its two children.
Consequently, the root of the tree holds the minimal value. The simplest implementation of such a tree is
done by a sequential array of nodes, using the rule that if a node is located at entry k, then its two children
are located at entries 2k and 2k + 1 (the first element of the array is indexed by 1, and is the root of the
tree). Equivalently, the parent of a node at entry k is located at entry bk/2c. Figure 2 shows an example of
a minimum heap and its implementation using array. Generally, each element in the array can hold many
properties, and one of these has to be defined as a comparable “key”, which is used in the heap for sorting.
In our case, each node holds a point ~x in the mesh, and its value τ (~x) as a key.
4
In its simplest form, the minimum heap structure supports two basic operations: insert(element,key)
and getMin(). For example, this is the case in the C++ standard library implementation of the minimum
heap structure. To apply getMin(), we remove the first element in the array, and take the last element in
the array and push it in the first place. Then, to maintain the property of the heap, we repeatedly replace
this value with its smaller child (the smaller of the two) until it reaches down in the tree to the point where
it is smaller than its two children or it has no children. The insert(element,key) operation first places
a new element at the next empty space of the array. Then, it propagates this element upwards, each time
replacing it with its parent until it reaches a point where the element is either the root or its key is larger
than its parent’s key. Both of the described operations are performed in O(log m) complexity where m is
the number of elements in the heap.
In Algorithm 1, the set front is implemented using a min-heap. Steps 1-2 are trivially implemented using
getMin(), and insert(element,key) is used in Step 3. However, Step 4 of Algorithm 1 requires updating
values which are inside the heap but are not at the root. This operation is not supported in the standard
definition of either priority queue or minimum heap. Indeed, the papers [30, 31] use a variant of a priority
queue, which includes back-links from points ~xij to their corresponding locations inside the heap, and a
more “software-engineering friendly” implementation of this idea is suggested in [2], where those back-links
are incorporated within the heap implementation, without any relation to the mesh. However, although
this way an update of a key inside the heap can still be implemented in O(log m), it requires a specialized
implementation of the heap, and encumbers the operations described earlier to maintain these back-links. In
our implementation, we bypass this need for back-links, and implement Step 4 by reinserting elements to the
heap if they are indeed smaller than their value in the heap. In Step 1, we simply ignore entries which are in
known already. If the algorithm is indeed monotone, like the first order version in [30], this implementation
detail will not change the result of the algorithm. The downside of this change is that it enables front to
grow more than in the back-linked version. However, even if it grows four times compared to the back-linked
version, then the heap tree is just two nodes higher, making the difference in running time insignificant.
2.2. Second order Fast Marching
Solving the eikonal equation based on a first order discretization in (2.5) provides guaranteed monotonicity
and stability. However, it also provides a less accurate solution because of the added viscosity that is
associated with the first order approximation. To get a more accurate FM method, [31] suggests to use a
second order upwind approximation in (2.5), e.g.
−3τi + 4τi+1 − τi+2
3τi − 4τi−1 + τi−2
, Di+ τ =
.
(2.6)
2h
2h
However, in some cases, the scheme may revert to first order approximations from certain directions. The
obvious case for that is when there are not enough known points for the high order stencil. This case occurs
for example when the given initial region contains only one point. Another condition for using second order
operators is given in [31]:
τi−1 ≥ τi−2 or τi+1 > τi+2 ,
(2.7)
Di− τ =
where the left condition is used for backward operators and the right one for forward operators. If (2.7) is
not satisfied, the algorithm reverts to first order operators. Later we show that this condition guarantees
the monotonicity of the non-factored FM solution using second order scheme.
3. Fast Marching for the factored eikonal equation
Let us rewrite the factored eikonal equation (1.4) in a squared form, which is closer (1.1):
|τ0 ∇τ1 + τ1 ∇τ0 |2 = κ(~x)2 .
(3.8)
This writing is the key for deriving the FM algorithm for (1.4). Similarly to the Godunov upwind scheme in
(2.5), we discretize (3.8) for τ1 using a derivative operator D̂ instead of D
h
i
−y
+y
−x
+x
max{D̂ij
τ1 , −D̂ij
τ1 , 0}2 + max{D̂ij
τ1 , −D̂ij
τ1 , 0}2 = κ(~xij )2 , ~xij ∈ Ωh .
(3.9)
5
For example, the backward first order factored derivative operator is given by
−x
D̂ij
τ1 = (τ0 )ij
(τ1 )i,j − (τ1 )i−1,j
+ (p0 )ij (τ1 )ij ,
h
(3.10)
0
where τ0 and p0 = ∂τ
∂x are known. From this point we apply the Algorithm 1 as it is. We hold the values of
τ0 τ1 in front , and in Step 4 we update (τ1 )ij with the solution of (3.9).
Initialization: For the non factored equation, Algorithm 1 is initialized by τ (~x0 ) = 0 at the point
source. In the factored equation, this is trivially fulfilled by definition, because at the source τ0 (~x0 ) = 0.
Still, τ1 (~x0 ) should not be chosen arbitrarily since its value is used in the finite difference approximations
when evaluating its neighbors. Examining (3.8) at the source yields τ1 (~x0 )2 |∇τ0 |2 = τ1 (~x0 )2 = κ(~x0 )2 , since
we choose τ0 such that |∇τ0 |2 = 1, independently of κ. In some cases in the literature, i.e., [6], the value
κ(~x0 ) is absorbed in τ0 , such that |∇τ0 |2 = κ(~x0 )2 . Then τ1 (~x0 ) should be chosen as 1. This is obviously
equivalent for computing τ , however, it is much more convenient to choose τ0 independently of κ if one wants
to obtain the sensitivities of the FM algorithm (for more details, see Section 4).
Second order discretization: Similarly to the non-factored equation, the second order upwind approximations (2.6) can be used in (3.9)-(3.10) for τ1 . Again, we revert to the first order approximation in
cases where the additional point needed for the second order approximation is not in known . We note that
unlike the non-factored case, the solution τ1 is in most cases very smooth at the source (expected to be close
to constant or linear). So, when we initialize the algorithm with the value of τ1 at the point source and
revert to a first order approximation for the neighbors, we do not introduce large discretization errors. In
the non-factored case, the second derivative of τ is singular at the source, so using first order approximation
there significantly pollutes the rest of the solution.
3.1. Solution of the piecewise quadratic equation
We now describe how to solve both the non-factored and factored piecewise quadratic equations (2.5)
and (3.9) respectively. This is required in Step 4 of Algorithm 1. Solving such an equation consists of the
following four steps:
1. Determine the order of approximation for each derivative in (2.5)/(3.9) (Only required for high order
schemes).
2. Determine which directions to choose (backward or forward) for each dimension (x, y or z).
3. Solve the quadratic equation in (2.5)/(3.9), assuming all terms are positive.
4. Make sure that the solution is valid, such that all max terms in (2.5)/(3.9) are indeed held with positive
values. If not, some terms should be dropped, and the quadratic problem with the remaining terms is
solved again.
Let us first consider solving the non-factored first order (2.5), for which the Step 1 is not relevant. In this
+
−
case, Step 2 is simple: for each max{Dij
τ, −Dij
τ, 0} term, the smaller of the two values of τ from both sides
(forward or backward) is guaranteed to give a higher finite difference derivative. Furthermore, in Step 4, if
some of the terms turn out negative after Step 3, then we can drop terms from (2.5) in decreasing order of
the τ values, until a valid solution is reached. The same is true for a first order factored version in (3.9).
However, using second order schemes (selectively) imposes additional complications on the solution of
the piecewise quadratic equations (2.5) and (3.9). There are many options for order of accuracy vs directions
in Steps 1-2, and in addition it is not clear in which order to drop terms in Step 4 if negative terms are
detected. Obviously, one can check all possibilities, but such an option may be costly in high dimensions.
To simplify this we follow [31], and in Step 1 we use the second order approximation if the extra point is
available in known and fulfils the condition (2.7), and revert to first order approximation if not. Then, in
Step (2) we determine the choice of directions considering the non-factored first order approximation (2.5).
That is, if (τ0 τ1 )i−1 < (τ0 τ1 )i+1 , then we choose the backward upwind direction; otherwise we choose the
forward direction. That is done correspondingly for each dimension.
Once Steps 1-2 are done, (3.9) reduces to a piecewise quadratic equation of the form
X
2
max{αk (τij − βk ), 0} = κ(~xij )2 ,
(3.11)
k
6
where αk ≥ 0, βk ≥ 0 are non-negative constants that are coming from the finite difference approximations.
For example, assuming that k = 1 corresponds to the x coordinate, then (3.10) would correspond to α1 =
(τ ) (τ )
(τ0 )i,j
+ (p0 )i,j and β1 = 0 i,jhα11 i−1,j . In Step 3 we simply ignore the max{·, 0} function and solve the
h
equation assuming all terms are positive. We solve a simple quadratic function and choose the larger one
of its two solutions for τij . If all chosen derivative terms are positive, the solution is valid; otherwise, we
reduce the terms in (3.11) in decreasing order of βk , each time solving (3.11) with the remaining terms until
a valid solution is reached. In three dimensions for example, this involves at most three quadratic solves.
Algorithm 2 summarizes the solution of the piecewise quadratic equation.
Algorithm: Solution of the piecewise quadratic equation
for each dimension x,y,... do
% Choosing direction, forward or backward.
if both forward and backward neighboring points are in known then
Choose the direction with smaller neighboring τ .
else
Otherwise, choose the direction in known .
end
% Choosing order of approximation, 1st or 2nd.
if next neighboring point is in known then
Use second order approximation.
else
Use first order approximation.
end
end
% Now all coefficients of αk and βk of Equation (3.11) are known.
Calculate (τ1 )i,j by solving Equation (3.11).
while the solution (τ1 )i,j is not valid do
Remove the term with largest βk from the remaining terms in (3.11).
Calculate (τ1 )i,j by solving (3.11) with the remaining terms.
end
3.2. The monotonicity of the obtained solution
It is known that the solution of (1.1) is monotone in the direction of the characteristics. We now show
how to enforce the monotonicity of our solution using the FM method for the factored eikonal equation. To
set the stage, we first consider the FM method for the original non-factored equation.
In [30] the non-factored first order discretization (2.5) is considered. In this case, each newly calculated
value τij is guaranteed to be larger than its known neighbors at the time of the calculation. To show this
clearly, consider for example that the backward derivative is chosen in the x direction. Then,
τi−1,j = τi,j − h
τi,j − τi−1,j
−x
= τi,j − hDi,j
τ,
h
(3.12)
−x
so since Di,j
τ ≥ 0 in the solution of (2.5), we have that τi,j ≥ τi−1,j . This means that τi,j is greater or equal
to its known neighbors. This property insures the monotonicity of the solution. The proof for this appears in
[30], but here we can simplify it because unlike [30], we only calculate entries using known values. We state
the following lemma:
Lemma 1. Let τ be the result of Algorithm 1, for solving the (non-factored) first order equation (2.5). Then
the values of τ are monotonically non-decreasing in the order in which they are set to known .
7
Proof 1. Denote by ~xk an element that is set to known at Step 2 of the k-th iteration of Algorithm 1. Assume
by contradiction that there exists two elements ~xp and ~xk , such that τ (~xp ) > τ (~xk ) and p < k. Without loss
of generality, assume that k is the earliest iteration that this condition is fulfilled. Let k̄ < k be the iteration
in which the value of τ (~xk ) is updated in the last time and it is entered to front . We know that ~xk̄ is a
neighbor of ~xk . By the algorithm, we know that at the k̄-th iteration ~xp is already set to known , otherwise
~xk would have been chosen to known at the p-th iteration instead of ~xp . By the assumption, we know that
τ (~xp ) ≤ τ (~xk̄ ), because otherwise ~xk would not have been the earliest element to violate the monotonicity.
By the property in (3.12), we know that τ (~xk̄ ) ≤ τ (~xk ), and hence we reach τ (~xp ) ≤ τ (~xk ), which contradicts
our assumption.
Furthermore, the lemma above can be extended for a Fast Marching solution of any equation (2.5) such
that the discretization operator D satisfies a monotonicity condition:
−x
+x
Dij
τ ≥ 0 ⇒ τij ≥ τi−1,j and − Dij
τ ≥ 0 ⇒ τij ≥ τi+1,j .
(3.13)
The next corollary can be proved using the same arguments as in Lemma 1:
Corollary 1. Let τ be the result of Algorithm 1, for solving the Godunov upwind equation (2.5) using
operators D which satisfy (3.13). Then the values of τ are monotonically non-decreasing in the order in
which they are set to known .
The condition (3.13) and the corollary above is violated when the second order operators (2.6) are used
in (2.5). However, if we look at a single violation, then it is of order h2 . To show this, we examine the
backward difference derivative using Taylor expansion:
∂τ
−x
+ O(h2 ) = τi,j − hDi,j
τ + O(h2 ),
(3.14)
τi−1,j = τi,j − h
∂x i,j
−x
where Di,j
is given in (2.6) (the same arguments can be derived for the forward difference derivative).
−x
Assuming again that Di,j
τ > 0, this means that each newly calculated τi,j is generally greater than its
−x
known neighbors, but may violate that up to magnitude O(h2 ). Note that if Di,j
τ is sufficiently bounded
2
∂ τ
away from zero and the second derivative ∂x
2 is bounded in [xi−1 , xi ], then τi,j > τi−1,j will be satisfied.
To correct this and obtain a monotone solution using (2.6), one may impose the condition (2.7) for using
the second order scheme. If it is not satisfied, the scheme reverts to the first order scheme, which satisfies
(3.13). If (2.7) is satisfied, then (2.6) does satisfy (3.13), because
−x
2hDij
τ = 3τij − 4τi−1,j + τi−2,j < 3τij − 3τi−1,j .
(3.15)
Note that the condition (2.7) is suggested in [31] but the monotonicity guarantee of the second order scheme
is not examined.
We now examine the monotonicity of the obtained factored solution τ0 τ1 when using first order operators
in (3.9). Suppose that we are calculating (τ1 )ij using the backward operator (3.10) in the x direction in
(3.9). We again start with a Taylor expansion
∂τ1
2
2
0
(τ0 τ1 )i−1,j =
(τ0 )ij − h ∂τ
+
O(h
)
(τ
)
−
h
+
O(h
)
1
ij
∂x i,j
∂x i,j 2
∂τ0
1
(3.16)
= (τ0 τ1 )ij − h(τ0 )ij ∂τ
−
h(τ
)
+ O(h )
1
ij
∂x
∂x
i,j
=
−x
(τ0 τ1 )ij − hD̂ij
τ1 + O(h2 ),
i,j
(τ1 )i,j −(τ1 )i−1,j
1
where the last equality is obtained by placing ∂τ
+O(h). This expansion shows that if
∂x i,j =
h
the monotonicity is not obtained, i.e., (τ0 τ1 )i,j − (τ0 τ1 )i−1,j < 0, then the non-factored derivative is negative,
−x
−x
Dij
(τ0 τ1 ) < 0, while the factored derivative is non-negative D̂ij
τ1 ≥ 0 (otherwise it is not chosen in (3.9)).
This means that the monotonicity may be violated only up to an error of O(h2 ). This holds for both first
8
and second order upwind approximations. In fact, (3.2) shows that this is a result of using the chain rule
rather than the order of discretization of the operators D̂, since the monotonicity condition involves the
value (τ0 )i−1,j , while it does not appear in the discretization scheme. In any case, the magnitude of the error
in the monotonicity violation is either of the same or of higher order as the error in τ1 , using first or second
2
−x
order schemes. Again, if Di,j
τ1 is sufficiently bounded away from zero and ∂∂xτ21 is bounded in [xi−1 , xi ],
then the monotonicity (τ0 τ1 )i,j > (τ0 τ1 )i−1,j will be satisfied.
Nevertheless, in our algorithm we may enforce the monotonicity of the obtained solution by reverting to
the non-factored operators in cases where the monotonicity is not satisfied, or, the factored and non-factored
−x
−x
schemes do not agree in sign, for example: D̂ij
τ1 ≥ 0, but Dij
(τ0 τ1 ) < 0. Note that in this case the
numerical derivative is approximately zero, hence the direction of the characteristic is almost parallel to the
y direction. We apply this change using the same order of derivative which the algorithm chooses to use.
That is, if the algorithm chooses a first or second order factored stencil, we revert to a standard first or
second order stencil, respectively. Following Corollary 1, this guarantees the monotonicity of the solution,
because we enforce the condition (3.13) at all stages of the algorithm. We note that experimentally, this
small correction does not influence the accuracy of the solution obtained with our algorithm in both first
and second order schemes in two and three dimensions.
4. Calculation of sensitivities and travel time tomography
Travel time tomography is a useful tool in some Geophysical applications. One way to obtain it is by
using the eikonal equation as a forward problem inside an inverse problem [29]. To solve the inverse problem,
one should be able to solve (1.1) accurately, and to compute its sensitivities. The works of [13, 34] computes
the tomography by FS, and require an FS iterative solution for computing the sensitivities. When using
the FM algorithm for forward modeling, those are obtained more efficiently by a simple solution of a lower
triangular linear system [14, 3]. More explicitly, let us denote by boldface all the discretized values of the
mentioned functions on a grid, and suppose that we set m to be the vector of the values of κ(~x)2 on this grid.
By solving (3.9), we get a function τ 1 (m) for the values of τ1 on the grid. We wish to get a linearization for
τ 1 (m), such that we can predict its change following a small change in m. That is, we wish to be able to
apply an approximation
τ 1 (m + δm) ≈ τ 1 (m) + Jδm,
(4.17)
where J is the sensitivity matrix (or Jacobian) defined by
Jij = (∇m τ 1 )ij =
∂(τ 1 )i
.
∂mj
(4.18)
To obtain the sensitivity we first rewrite (3.9) in implicit form
f (m, τ 1 ) = (D̂x τ 1 )2 + (D̂y τ 1 )2 − m = 0,
(4.19)
where D̂x = diag (τ 0 )Dx + diag (p0 ) and D̂y = diag (τ 0 ) · Dy + diag (q0 ) are the matrices that apply the
finite difference derivatives that are chosen by the FM algorithm when applied for m. p0 and q0 are the
analytical derivatives of τ0 with respect to x and y on the grid respectively, and diag (x) denotes a diagonal
matrix whose diagonal elements are those of the vector x. We note that in the points where no derivative is
chosen in the solution of (3.9), a zero row is set in the corresponding operator D̂. Also, at the row of the
point source, we set each of D̂x and D̂y to have only one diagonal non-zero element, which equals to the
values of p0 and q0 at the source. This way, (4.19) is exactly fulfilled for D̂x and D̂y and τ 1 .
To obtain the sensitivity, we apply the gradient operator to both sides of (4.19), yielding (∇τ 1 f )(∇m τ 1 )+
∇m f = 0, and define [8]:
J(m) = ∇m τ 1 = −(∇τ 1 f )−1 (∇m f ).
(4.20)
J = (diag (2D̂x τ 1 )D̂x + diag (2D̂y τ 1 )D̂y )−1 ,
(4.21)
This results in
9
following ∇m f = −I, and since the operators D̂ do not depend on m (we defined τ0 and its derivatives so it
does not depend on κ).
The matrix (4.21) can be multiplied with any vector efficiently given the order of variables (i, j) in which
the FM algorithm set their values as known . To apply J on an arbitrary vector x, i.e. calculate e = Jx,
a linear system Ae = x can be solved with A = J−1 (note that A is a sparse matrix). The equations of
this linear system, which correspond to the rows of A, can be approached and solved sequentially in the
FM order of variables. Since the FM algorithm uses only known variables for determining each new variable,
then when looking at each row i of A, the non-zero entries in that row (except i) correspond to variables
that where in known when τ i was determined during the FM run. Therefore, if all those variables are known
except i, then the i-th equation has only one unknown (ei ) and can be trivially solved. In other words, if
we permute A according to the FM order, we get a sparse lower triangular matrix, and the corresponding
system can be solved efficiently in one forward substitution sweep in O(n) operations. For the non-factored
equation one may use (4.21) with non-factored operators Dx and Dy instead of the factored ones [14].
4.1. Travel time tomography using Gauss-Newton
Assume that we have several sources and receivers set on an open surface, and for each source we have
traveltime data diobs given in the location of the receivers. Based on these observations we wish to compute
the unknown slowness model of the ground underneath. The inverse problem for this process, called travel
time tomography, may be given by
(n
)
s
X
> i
i
2
kP τ (m) − dobs k + αR(m) ,
min
φ(m) =
min
(4.22)
mL <m<mH
mL <m<mH
i=1
where
|∇τ i |2 = m(~x) τ i (~xi ) = 0 i = 1, . . . , ns
(4.23)
Here τ i is the travel time from the point source ~xi , and m(~x) = κ(~x)2 is the squared slowness model as in
(1.1), only now it is unknown. The operator P> is a projection to the set of receivers that gather the wave
information. Here we assume that the information from all sources is available on all the receivers, i.e., the
projection operator P does not change between sources. R(m) is a regularization term and α > 0 is its
balancing parameter. The parameters mL and mH are positive lower and upper bounds needed for keeping
the slowness of the medium physical. We note that the observations diobs can be obtained manually from
recorded seismic data or by automatic time picking—for more information see [28] and references therein.
Without the regularization term R(m), the problem (4.22) is ill-posed, i.e., many solutions m may fit
the predicted travel time to the measured data [37, 32]. For this reason, in most cases we cannot expect to
exactly recover the true model, but wish to recover a reasonable model by adding prior information using the
regularization term R(m). This term aims to promote physical or meaningful solutions that we may expect
to see in the recovered model. For example, in seismic exploration, one may expect to recover a layered
model of the earth subsurface, hence may choose R to promote smooth or piecewise-smooth functions like
the total variation regularization term [26].
There are several ways to solve (4.22), and most of them are gradient-based. Here we focus on GaussNewton. This method is computationally favorable here, since its cost is governed by the application of
sensitivities, which are easy to obtain using FM. Given an approximation m(k) at the k-th iteration, we
place (4.17) into (4.22) and get
min
δm
ns
1X
kP> diag (τ i0 ) τ i1 (m(k) ) + Ji δm − diobs k2 + αR(m(k) + δm),
2 i=1
(4.24)
where Ji is the sensitivity of τ i1 at m(k) . Minimizing this approximation for δm leads to computing the
gradient
∇m φ(m(k) ) =
ns
X
(Ji )> diag (τ i0 )P P> diag (τ i0 )τ i1 − diobs + α∇m R(m(k) ).
i=1
10
(4.25)
We then approximately solve the linear system
Hδm = −∇m φ(m(k) )
where
H=
n
X
(Ji )> diag (τ i0 )PP> diag (τ i0 )Ji + α∆m R(m(k) ).
i=1
The linear system is solved using the conjugate gradient method where only matrix vector products are
computed. Finally, the model is updated, m ← m + µδm where µ ≤ 1 is a line search parameter that is
chosen such that the objective function is decreased at each iteration.
5. Numerical results: solving the eikonal equation
In this section we demonstrate the FM algorithm using first or second order upwind discretization for
solving the factored eikonal equation (1.4). We demonstrate both the accuracy of the obtained solution,
and the computational cost of calculating it using the FM algorithm. The accuracy of the algorithm is
demonstrated by two error norms: one in the maximum norm l∞ , and one is the mean l2 norm defined by
the standard l2 norm of the error divided by the square root of the total number of variables. Similarly to
[31], we show these two measures to demonstrate the accuracy of the second order scheme. Showing the l∞
norm of the error for this scheme may result in only first order accuracy, because at some points our second
order FM algorithm reverts to first order operators, which may be picked by the l∞ norm.
To demonstrate the efficiency of the computation, we measure the time in which the algorithm solves
each test. We also show this timing in terms of work-units, where each work unit is defined by the time
that it takes to evaluate the equation (1.1) using given central difference gradient stencils (without memory
allocation time). We note that the more reliable timings appear for the large scale examples.
We use analytical examples for media where there is a known analytical solution for a point source
located at ~x0 . The first two appear in [6]. We show results for two and three dimensions. Our code
is written in Julia language [4] version 0.4.5, and all our tests were calculated on a laptop machine using
Windows 10 64bit OS, with Intel core-i7 2.8 GHz CPU with 32 GB of RAM. Our code is publicly available in
https://github.com/JuliaInv/FactoredEikonalFastMarching.jl. We do not enforce the monotonicity
in the results below, but those can be enforced in our package. The three test cases are listed below.
Test case 1: Constant gradient of squared slowness. In this test case we set:
κ2 (~x) = s20 + 2a~e1 · (~x − ~x0 ),
(5.26)
where ~e1 = (1, 0) is a unit vector, and · is the standard dot product. The parameters a, s0 , the domain and
the source location are chosen differently in 2D and 3D. The corresponding exact solution is given by
1
τexact (~x) = S̄ 2 σ − a2 (σ 3 ),
6
(5.27)
where
S̄ 2 (~x)
=
σ 2 (~x)
=
s20 + a~e1 · (~x − ~x0 )
−1
q
2
4
2
2
2k~x − x~0 k2 .
S̄ + S̄ − a k~x − x~0 k
Figures 3(a) and 3(d) show the model κ for this test case with the chosen parameters for 2D.
11
(5.28)
(5.29)
(a) Test case 1: Constant gradient of (b) Test case 2: Constant gradient of vesquared slowness.
locity.
(c) Test case 3: Gaussian factor
(d) Test case 1: Contours of the solution. (e) Test case 2: Contours of the solution. (f) Test case 3: Contours of the solution.
Figure 3: The 2D slowness model κ(~
x) of the three test cases and the corresponding contours of the 2D solutions.
Test case 2: Constant gradient of velocity. In this test case we set:
κ(~x) =
1
+ a~e1 · (~x − ~x0 )
s0
−1
,
(5.30)
where again ~e1 = (1, 0), · is the dot product, and the parameters a, s0 , the domain and the source location
are chosen differently in 2D and 3D. The exact solution is given by
1
1
2
2
(5.31)
τexact (~x) = acosh 1 + s0 a κ(~x)k~x − ~x0 k .
a
2
Figures 3(b) and 3(e) show the model κ for this test case with the chosen parameters for 2D.
Test case 3: Gaussian factor. In this test case we choose a function for τ1exact and multiply it by τ0 to get
τexact . We choose τ1 as a Gaussian function centered around a point x~1 :
τ1exact (~x) =
1
1
exp −(~x − ~x1 )T Σ(~x − ~x1 ) + ,
2
2
(5.32)
where Σ is a 2 × 2 or 3 × 3 positive diagonal matrix. As before, the parameters ~x1 and Σ, the domain and
the source location x~0 are chosen differently in 2D and 3D. Here κ(~x) is defined by (3.8), with τ0 being the
distance function. Figures 3(c) and 3(f) show the model κ for this test with the chosen parameters for 2D.
5.1. Two dimensional tests
Now we show results for the two dimensional versions of the tests mentioned above. For all tests in 2D
we choose the domain to be [0, 4] × [0, 8], while h = hx = hy varies from large to small.
Test case 1: Constant gradient of squared slowness. For this 2D setting, we use the parameters a = −0.4,
s0 = 2.0, and the source location is ~x0 = (0, 4). Table 1 summarizes the results for this test. On the first
order section we see a typical first order convergence rate in both error norms. As the mesh size increases
by two in each direction, the errors drop by a factor of two. In the second order section we see the typical
behavior of the FM algorithm. At some points, first order operators are used, and hence the error at those
12
h
n
1/40
1/80
1/160
1/320
1/640
1/1280
161 × 321
321 × 641
641 × 1281
1281 × 2561
2561 × 5121
5121 × 10241
1st order
error in τ
[3.71e-03, 9.42e-04]
[1.85e-03, 4.69e-04]
[9.22e-04, 2.34e-04]
[4.60e-04, 1.17e-04]
[2.30e-04, 5.83e-05]
[1.15e-04, 2.92e-05]
time(work)
0.05s(217)
0.19s (199)
0.85s (217)
3.89s (266)
16.4s (278)
76.6s (316)
2nd order
error in τ
time(work)
[9.33e-05, 9.26e-06] 0.05s (202)
[3.30e-05, 2.21e-06] 0.20s (209)
[1.14e-05, 5.32e-07] 0.85s (218)
[4.06e-06, 1.28e-07] 3.84s (262)
[1.47e-06, 3.12e-08] 17.1s (289)
[5.18e-07, 7.64e-09] 77.5s (320)
Table 1: Results for 2D constant gradient of squared slowness (test case 1). The error measures are in the [l∞ , mean l2 ] norms.
locations dominates the l∞ norm. Still, we observe much better convergence compared to the first order l∞ ,
only it is not of second order. In the mean l2 norm we see typical second order convergence—as the mesh
size increases by two in each direction, the errors drop by a factor of four. In any case, the errors in the
second order columns are much smaller than those in the first order columns.
In terms of computational cost, the 2D FM algorithm exhibits favorable timings and work counts. Except
the small cases, the cost of the algorithm is comparable to 200-300 function evaluations using standard
difference stencils. This is maintained for all the considered mesh sizes. The difference in the computational
cost between using first and second order schemes is only about 10% of execution time.
Test case 2: Constant gradient of velocity. For this 2D setting, we use the parameters a = 1.0, s0 = 2.0,
and the location of the source is again at ~x0 = (0, 4). Table 2 summarizes the results for this test. The
results here are almost identical to the previous test case. The first order columns show typical first order
convergence in both error norms. The second order columns show better convergence and exhibits second
order convergence in the mean l2 norm column. The computational costs columns show timings which are
almost identical to the previous test case.
h
n
1/40
1/80
1/160
1/320
1/640
1/1280
161 × 321
321 × 641
641 × 1281
1281 × 2561
2561 × 5121
5121 × 10241
1st order
error in τ
[2.66e-02, 1.01e-02]
[1.32e-02, 5.05e-03]
[6.59e-03, 2.52e-03]
[3.29e-03, 1.26e-03]
[1.65e-03, 6.28e-04]
[8.22e-04, 3.14e-04]
time(work)
0.05s (205)
0.21s (221)
0.87s (223)
3.80s (259)
16.2s (274)
73.8s (304)
2nd order
error in τ
time(work)
[4.86e-04, 2.90e-04] 0.05s (236)
[1.67e-04, 7.38e-05] 0.20s (206)
[5.18e-05, 1.85e-05] 0.86s (221)
[1.90e-05, 4.61e-06] 3.88s (265)
[6.58e-06, 1.15e-06] 16.6s (280)
[2.28e-06, 2.86e-07] 74.6s (307)
Table 2: Results for the 2D constant gradient of velocity (test case 2). The error measures are in the [l∞ , mean l2 ] norms.
Test case 3: Gaussian factor. For this setting, we use the parameters Σ = diag (0.1, 0.4), ~x1 = (4/3, 2)
(floored to the closest grid point), and the source is located in the point ~x0 = (1, 2). Table 3 summarizes the
results for this test. Again we see first order convergence at the first order columns in both norms. On the
second order columns we again see faster convergence, and in the mean l2 norm column we see convergence
rate that is close to second order—the error decreases by a factor of about 3.9 when the mesh size increases
by a factor of 2 in each dimension. Again we see similar behavior in the computational cost columns.
We now wish to better illustrate the difference between the accuracy of the first order scheme and the
second order scheme. First, Figure 4 shows contours of the exact and approximate solutions in certain
regions of the domain for the second and third test cases. It is clear that the first order approximation is less
accurate than the second order one. Next, Figure 5 shows plots of the errors in Tables 1-3 in logarithmic
scales for both h and the error norms, where the order of convergence determines the slope of the lines. It
is clear that in all cases, using the second order scheme we get second order convergence in mean l2 norm,
and a bit more than first order convergence in the l∞ norm.
13
h
n
1/40
1/80
1/160
1/320
1/640
1/1280
161 × 321
321 × 641
641 × 1281
1281 × 2561
2561 × 5121
5121 × 10241
1st order
error in τ
[6.15e-03, 3.86e-03]
[3.07e-03, 1.93e-03]
[1.54e-03, 9.67e-04]
[7.68e-04, 4.83e-04]
[3.84e-04, 2.42e-04]
[1.92e-04, 1.21e-04]
time(work)
0.05s(205)
0.21s (221)
0.87s (223)
3.80s (259)
16.2s (275)
73.8s (304)
2nd order
error in τ
time(work)
[1.60e-04, 5.94e-05] 0.05s(236)
[3.85e-05, 1.56e-05] 0.20s (206)
[1.08e-05, 4.03e-06] 0.86s (221)
[3.18e-06, 1.04e-06] 3.88s (265)
[9.59e-07, 2.66e-07] 16.6s (280)
[2.99e-07, 6.88e-08] 74.6s (307)
Table 3: Results for the 2D Gaussian factor test case (test case 3). The error measures are in the [l∞ , mean l2 ] norms.
(a) Constant gradient of velocity.
(b) Gaussian factor.
Figure 4: Contours of small regions in the exact, first order accurate and second order accurate travel times τ using h = 0.1.
The exact solution appears in black line. The first order approximation appears in dotted red line and the second order
approximation appears in a dashed blue line mostly right with the exact solution.
5.2. Three dimensional tests
We now show results for the same type of tests in three dimensions. For all the 3D tests we choose the
domain to be [0, 0.8] × [0, 1.6] × [0, 1.6], and h = hx = hy = hz varies.
Test case 1: Constant gradient of squared slowness. For the 3D version of this test case we use the parameters
a = −1.65, and s0 = 2.0, and the source is located at (0, 0.8, 0.8). Table 4 summarizes the results for this
test case. Again, like in two dimensions, the first order version of FM yields first order convergence rate in
both error norms. When using the second order scheme we get a super-linear convergence rate in the l∞
column, and second order convergence in the mean l2 column. We note that in 3D the FM algorithm reverts
to first order scheme on 2D manifolds, where the derivative in each dimension switches sign, and not on 1D
curves as in 2D.
In terms of computational cost, it is obvious that the 3D problem is much more expensive than the 2D
one. The computational cost in seconds per grid-point in 3D is about 3 times higher than the corresponding
cost in 2D. That is because the treatment of each grid point is more expensive (more neighbors and more
derivative directions), and the number of grid points that are processed inside the heap is much larger (a
2D manifold of points compared to a 1D curve). As a result, when we normalize the timing by the cost of a
3D “work-unit” (evaluation of (1.1) in 3D), the cost grows a little when the mesh-size grows. Still, solving
the problem requires 200-500 work units. Again, using the first and second order schemes requires similar
computational effort in our 3D implementation of the FM algorithm.
Test case 2: Constant gradient of velocity. For the 3D version of this test case we use the parameters a = 1.0,
and s0 = 2.0, and the source is located at (0, 0.8, 0.8). Table 5 summarizes the results for this test case. As
in the previous case, we get first order convergence when using the first order scheme, in both error norms.
Again, when using the second order scheme we get a super-linear convergence rate in the l∞ column, and
second order convergence in the mean l2 column.
14
(a) 2D constant
squared slowness.
gradient
of
the (b) 2D constant gradient of velocity.
(c) 2D Gaussian factor.
Figure 5: The accuracy of the FM approximations in logarithmic scales for the 2D cases. Red plots are used for first order
approximations, blue plots for second order approximations; dotted lines for l∞ error norm and solid for mean l2 norm. Black
circles denote a reference for exact second order convergence rate.
h
n
1/20
1/40
1/80
1/160
1/320
17 × 33 × 33
33 × 65 × 65
65 × 129 × 129
129 × 257 × 257
257 × 513 × 513
1st order
error in τ
[5.41e-03, 1.46e-03]
[2.64e-03, 7.05e-04]
[1.30e-03, 3.46e-04]
[6.41e-04, 1.72e-04]
[3.19e-04, 8.55e-05]
time(work)
0.04s(236)
0.30s (230)
2.88s (332)
28.7s (427)
264s (481)
2nd order
error in τ
time(work)
[5.63e-4 ,1.49e-04]
0.04s(234)
[2.00e-04 ,3.52e-05] 0.32s (235)
[6.99e-05 ,7.82e-06] 2.90s (334)
[2.51e-05 ,1.68e-06] 29.0s (432)
[8.78e-06 ,3.53e-07] 272s (497)
Table 4: Results for the 3D constant gradient of squared slowness test case (test case 1). The error measures are in the
[l∞ , mean l2 ] norms.
Test case 3: Gaussian factor. For this 3D test case we use the parameters Σ = diag (0.2, 0.4, 0.1), ~x1 =
(0.4, 1.6
x0 = (0.2, 0.4, 0.4).
3 , 0.4) (floored to the closest grid point), and the source is located in the point ~
Table 5 summarizes the results for this test case. The results are similar to the previous test case in both
the convergence (first/second order using l∞ /l2 norms) and computational costs in seconds and work units.
Again we wish to demonstrate the order accuracy of the FM approximations using the first and second
order schemes. Figure 6 shows the results in Tables 4-6 in logarithmic scales. Like in 2D, we observe second
order convergence rate when the error is measure in mean l2 norm. However, because the second order
stencil reduces to first order stencil in two dimensional manifolds, the error in l∞ norm is higher in 3D than
it is in 2D.
6. Numerical results: travel time tomography.
In this section we demonstrate a solution of travel time tomography using synthetic travel time data dobs
for a 2D and SEG/EAGE salt model given in [1] and presented in Figure 7(a), using a 256 × 128 grid that
represents an area of approximately 13.5km × 4.2km. We choose 51 equally distanced sources locations on
the open surface (that is, they are located every 5 pixels on the top row), and 256 receivers (located in every
pixel on the top row). We note that to have a reasonable solution using the first arrivals for the inverse
problem under this setup, the velocity in the interior has to be larger than that on the surface. This is to
guarantee that the first arrival rays obtained on the surface actually come from the interior but not only
travel along the surface. To dobs we add white Gaussian noise with standard deviation of 0.01 × mean(|dobs |).
To fit the model to the data, we minimize (4.22) using Gauss-Newton (we perform 10 iterations, where
in each we apply 8 CG steps for the Gauss Newton direction problem). We use the general-propose inversion
package [27], which is freely available in https://github.com/JuliaInv/jInv.jl, together with our FM
15
h
n
1/20
1/40
1/80
1/160
1/320
17 × 33 × 33
33 × 65 × 65
65 × 129 × 129
129 × 257 × 257
257 × 513 × 513
1st order
error in τ
[1.35e-02, 5.04e-03]
[6.24e-03, 2.44e-03]
[3.00e-03, 1.20e-03]
[1.47e-03, 5.99e-04]
[7.30e-04, 2.99e-04]
time(work)
0.04s(237)
0.31s (234)
2.86s (330)
27.6s (411)
263s (481)
2nd order
error in τ
time(work)
[2.34e-03, 9.36e-04] 0.04s(255)
[5.12e-04, 1.72e-04] 0.32s (236)
[1.70e-04, 3.82e-05] 2.89s (334)
[5.42e-05, 9.33e-06] 28.9s (430)
[1.95e-05, 2.29e-06] 271s (496)
Table 5: Results for the 3D constant gradient of velocity test case (test case 2). The error measures are in the [l∞ , mean l2 ]
norms.
h
n
1/20
1/40
1/80
1/160
1/320
17 × 33 × 33
33 × 65 × 65
65 × 129 × 129
129 × 257 × 257
257 × 513 × 513
1st order
error in τ
[7.53e-03, 3.26e-03]
[3.69e-03, 1.56e-03]
[1.83e-03, 7.62e-04]
[9.11e-04, 3.77e-04]
[4.54e-04, 1.87e-04]
time(work)
0.04s (230)
0.33s (245)
2.77s (319)
26.4s (393)
267s (487)
2nd order
error in τ
time(work)
[3.65e-04, 1.27e-04] 0.04s (229)
[9.95e-05, 2.85e-05] 0.34s (253)
[3.22e-05, 7.50e-06] 2.80s (323)
[1.06e-05, 2.06e-06] 27.2s (405)
[3.54e-06, 5.66e-07] 276s (504)
Table 6: Results for the 3D Gaussian factor test case (test case 3). The error measures are in the [l∞ , mean l2 ] norms.
(a) 3D constant gradient of the squared
slowness.
(b) 3D constant gradient of velocity.
(c) 3D Gaussian factor.
Figure 6: The accuracy of the FM approximations in logarithmic scales for the 3D cases. Red and blue plots are used for first
and second order approximations, respectively; dotted lines for l∞ error norm and solid for mean l2 norm. Black circles denote
an exact second order convergence rate.
package mentioned earlier. For that, we first generate an initial slowness model m(0) = mref , whose velocity
model shown in Figure 7(b). This corresponds to a velocity field with a constant gradient in the y direction,
similarly to the model in Figure 3(b). To bound m from above and from below throughout the minimization,
we invert for an auxiliary variable m0 and use the following scalar bounded bijective mapping that prevents
m from being below mL or above mH :
mH − mL
2
mH + mL
0
0
mbound (m ) =
· tanh
· m −
+ 1 + mL .
2
mH − mL
2
That is, instead of minimizing (4.22) as is, we minimize
(n
)
s
X
0
> i
0
i
2
0
min
φ(m ) = min
kP τ (mbound (m )) − dobs k + αR(m ) ,
0
0
m
m
(6.33)
i=1
subject to the same constraints in (4.23). For the regularization R use a simple discrete central-differences
16
(a) True velocity model.
(b) Initial velocity model mref
(c) Recovered velocity model.
Figure 7: 2D travel time tomography experiment, grid size 256 × 128. Velocities are given in km/sec.
(a) Initial data using mref .
(b) Final predicted data.
(c) Initial residual: |dref − dobs |
(d) The final residual: |dpred − dobs |
Figure 8: Initial and final data and residuals in the source-receiver domain.
Laplacian, and apply it for m0 ; that is
R(m0 ) =
1 0
(m − m0ref )> ∆h (m0 − m0ref ),
2
where m0ref is the model such that mbound (m0ref ) = mref . For ∆h we use Neumann boundary conditions,
since those lead to an effect of an automatic salt flooding, which is a popular way to treat salt bodies. We
set the regularization parameter to be α = 0.5. We note that other choices of mref and regularization terms
may definitely be suitable here, but are beyond the scope of this paper. Figure 7(c) shows the result model
of the Gauss Newton minimization, and Figure 8 presents the initial and final data and data residuals. In
particular, Figure 8(d) shows that the final residual mostly contains the added Gaussian noise. Figure 9
shows that the misfit was indeed reduced throughout the iterations, until the reduction stalls and the misfit
reflect the noise level.
To demonstrate our algorithm in 3D, we use a 3D version of the same SEG/EAGE model, presented in
Figure 10(a), using a 256 × 256 × 128 grid that represents a volume of 13.5km × 13.5km × 4.2km. We choose
144 equally distanced sources locations on the open surface, located every 23 pixels on the top surface, and
256 × 256 receivers located on the top surface. We use the same parameters as in the 2D experiment (bound
function, regularization, initial 3D model, added noise to the data, number of iterations etc.). Figure 10(b)
shows the result of the inversion. Similarly to the 2D case, the top part of the model is recovered quite well,
while a “salt flooding” effect is evident in the bottom part of the model. We performed the inversion using a
machine with two Intel(R) Xeon(R) E5-2670 v3 processors with 128 GB of RAM. Using 24 cores, we applied
the inversion in approximately 15 hours, and the highest memory footprint of the algorithm reached around
30GB.
17
Figure 9: Convergence history of the inversion.
(a) True velocity model.
(b) Recovered model.
Figure 10: 3D travel time tomography experiment, grid size 256 × 256 × 128. Velocities are given in km/sec.
7. Conclusions
In this paper we developed a Fast Marching algorithm for the factored eikonal equation, which in many
cases yields a more accurate solution of the travel time than the original equation. Similarly to the original
FM algorithm, our version solves the factored problem by exploiting the monotonicity of the solution along
the characteristics. Our algorithm is capable of solving the problem using both first and second order
schemes. The advantages of our algorithm are (1) its favorable guaranteed O(n log n) running time, and (2)
the easily computed sensitivity matrices for solving the inverse (factored) eikonal equation.
8. Acknowledgement
The research leading to these results has received funding from the European Union’s - Seventh Framework
Programme (FP7/2007-2013) under grant agreement no 623212 - MC Multiscale Inversion.
References
References
[1] F. Aminzadeh, B. Jean, and T. Kunz, 3-D salt and overthrust models, Society of Exploration
Geophysicists, 1997.
[2] J. A. Bærentzen, On the implementation of fast marching methods for 3D lattices, Tech. Report
IMM-TR-2001-13, Informatics and Mathematical Modelling, Technical University of Denmark, DTU,
Richard Petersens Plads, Building 321, DK-2800 Kgs. Lyngby, 2001.
18
[3] A. Benaichouche, M. Noble, and A. Gesret, First arrival traveltime tomography using the fast
marching method and the adjoint state technique, in 77th EAGE Conference Proceedings., 2015.
[4] J. Bezanzon, S. Karpinski, V. Shah, and A. Edelman, Julia: A fast dynamic language for
technical computing, in Lang.NEXT, Apr. 2012.
[5] M. G. Crandall and P.-L. Lions, Viscosity solutions of Hamilton-Jacobi equations, Transactions of
the American Mathematical Society, 277 (1983), pp. 1–42.
[6] S. Fomel, S. Luo, and H. Zhao, Fast sweeping method for the factored eikonal equation, Journal of
Computational Physics, 228 (2009), pp. 6440–6455.
[7] P. A. Gremaud and C. M. Kuster, Computational study of fast methods for the eikonal equation,
SIAM journal on scientific computing, 27 (2006), pp. 1803–1816.
[8] E. Haber, Computational Methods in Geophysical Electromagnetics, vol. 1, SIAM, 2014.
[9] E. Haber and S. MacLachlan, A fast method for the solution of the Helmholtz equation, Journal of
Computational Physics, 230 (2011), pp. 4403–4418.
[10] C. Y. Kao, S. Osher, and J. Qian, Lax–friedrichs sweeping scheme for static Hamilton-Jacobi
equations, Journal of Computational Physics, 196 (2004), pp. 367–391.
[11] R. Kimmel and J. A. Sethian, Computing geodesic paths on manifolds, Proceedings of the National
Academy of Sciences, 95 (1998), pp. 8431–8435.
[12] S. Leung, J. Qian, and R. Burridge, Eulerian Gaussian beams for high-frequency wave propagation,
Geophysics, 72 (2007), pp. SM61–SM76.
[13] S. Leung, J. Qian, et al., An adjoint state method for three-dimensional transmission traveltime
tomography using first-arrivals, Communications in Mathematical Sciences, 4 (2006), pp. 249–266.
[14] S. Li, A. Vladimirsky, and S. Fomel, First-break traveltime tomography with the double-square-root
eikonal equation, Geophysics, 78 (2013), pp. U89–U101.
[15] S. Luo and J. Qian, Factored singularities and high-order lax–friedrichs sweeping schemes for pointsource traveltimes and amplitudes, Journal of Computational Physics, 230 (2011), pp. 4742–4755.
[16]
, Fast sweeping methods for factored anisotropic eikonal equations: multiplicative and additive
factors, Journal of Scientific Computing, 52 (2012), pp. 360–382.
[17] S. Luo, J. Qian, and R. Burridge, Fast Huygens sweeping methods for Helmholtz equations in
inhomogeneous media in the high frequency regime, Journal of Computational Physics, 270 (2014),
pp. 378–401.
[18] S. Luo, J. Qian, and R. Burridge, High-order factorization based high-order hybrid fast sweeping
methods for point-source eikonal equations, SIAM Journal on Numerical Analysis, 52 (2014), pp. 23–44.
[19] S. Luo, J. Qian, and H. Zhao, Higher-order schemes for 3d first-arrival traveltimes and amplitudes,
Geophysics, 77 (2012), pp. T47–T56.
[20] M. Noble, A. Gesret, and N. Belayouni, Accurate 3-d finite difference computation of traveltimes
in strongly heterogeneous media, Geophysical Journal International, 199 (2014), pp. 1572–1585.
[21] A. Pica et al., Fast and accurate finite-difference solutions of the 3d eikonal equation parametrized
in celerity, 67th Ann. Internat. Mtg, Soc. of Expl. Geophys, (1997), pp. 1774–1777.
[22] J. Qian and W. W. Symes, An adaptive finite-difference method for traveltimes and amplitudes,
Geophysics, 67 (2002), pp. 167–176.
19
[23] J. Qian, Y.-T. Zhang, and H.-K. Zhao, A fast sweeping method for static convex Hamilton-Jacobi
equations, Journal of Scientific Computing, 31 (2007), pp. 237–271.
[24] N. Rawlinson and M. Sambridge, Wave front evolution in strongly heterogeneous layered media
using the fast marching method, Geophysical Journal International, 156 (2004), pp. 631–647.
[25] E. Rouy and A. Tourin, A viscosity solutions approach to shape-from-shading, SIAM Journal on
Numerical Analysis, 29 (1992), pp. 867–884.
[26] L. I. Rudin, S. Osher, and E. Fatemi, Nonlinear total variation based noise removal algorithms,
Physica D: Nonlinear Phenomena, 60 (1992), pp. 259–268.
[27] L. Ruthotto, E. Treister, and E. Haber, jInv – a flexible Julia package for PDE parameter
estimation, Submitted, (2016).
[28] C. Saragiotis, T. Alkhalifah, and S. Fomel, Automatic traveltime picking using instantaneous
traveltime, Geophysics, 78 (2013), pp. T53–T58.
[29] A. Sei, W. W. Symes, et al., Gradient calculation of the traveltime cost function without ray tracing,
in 65th Ann. Internat. Mtg., Soc. Expl. Geophys., Expanded Abstracts, 1994, pp. 1351–1354.
[30] J. A. Sethian, A fast marching level set method for monotonically advancing fronts, Proceedings of
the National Academy of Sciences, 93 (1996), pp. 1591–1595.
[31]
, Fast marching methods, SIAM review, 41 (1999), pp. 199–235.
[32] E. Somersalo and J. Kaipio, Statistical and computational inverse problems, Applied Mathematical
Sciences, 160 (2004).
[33] A. Spira and R. Kimmel, An efficient solution to the eikonal equation on parametric manifolds,
Interfaces and Free Boundaries, 6 (2004), pp. 315–328.
[34] C. Taillandier, M. Noble, H. Chauris, and H. Calandra, First-arrival traveltime tomography
based on the adjoint-state method, Geophysics, 74 (2009), pp. WCB1–WCB10.
[35] Y.-H. R. Tsai, L.-T. Cheng, S. Osher, and H.-K. Zhao, Fast sweeping algorithms for a class of
Hamilton-Jacobi equations, SIAM journal on numerical analysis, 41 (2003), pp. 673–694.
[36] J. N. Tsitsiklis, Efficient algorithms for globally optimal trajectories, Automatic Control, IEEE Transactions on, 40 (1995), pp. 1528–1538.
[37] C. R. Vogel, Computational methods for inverse problems, vol. 23, SIAM, Philadelphia, 2002.
[38] H. Zhao, A fast sweeping method for eikonal equations, Mathematics of computation, 74 (2005),
pp. 603–627.
20
| 5cs.CE
|
1
arXiv:1803.06790v2 [math.ST] 28 Mar 2018
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
Eugene Katsevich
Department of Statistics
Stanford University
[email protected]
Aaditya Ramdas
Departments of Statistics and EECS
University of California, Berkeley
[email protected]
03.30.18
Abstract
The false discovery rate (FDR) has become a popular Type-I error criterion for multiple testing, but it is not
without its flaws. Indeed, (a) controlling the mean of the false discovery proportion (FDP) does not preclude large
FDP variability, and (b) committing to an error level q before observing the data limits its use in exploratory data
analysis. In this paper, we take a step towards addressing both of the above drawbacks by proving uniform FDP
bounds for a variety of existing FDR procedures. In particular, many such procedures proceed by examining a
path of potential rejection sets ∅ = R0 ⊆ R1 ⊆ · · · ⊆ Rn ⊆ [n], assigning some algorithm-dependent estimate
d k ) to each one, and choosing the final rejection set Rk∗ via k∗ = max{k : FDP(R
d k ) ≤ q}. The resulting
FDP(R
∗
inferential guarantee applies only to the set Rk , guaranteeing the FDP of this set will be small on average. In
this paper, we prove that for a wide variety of such procedures (including Benjamini-Hochberg, accumulation
d bounds the FDP to within a small explicit
tests, knockoffs, AdaPT, and STAR), under independent p-values, FDP
constant factor calg (α), uniformly across the entire path, with probability 1 − α. In each case calg (α) scales
FDP(Rk )
has exponentially decaying tails. The consequences of this
like log( α1 ), meaning that the quantity supk FDP(R
d
k)
statement are two-fold. First of all, they imply that existing FDR procedures also have FDP bounded with high
probability by a small constant multiple of the target FDR level q. This constant is close to 2 for several procedures
at the 95% confidence level. Secondly, our bounds open up a middle ground between fully simultaneous inference
(guarantees for all 2n possible rejection sets), and fully selective inference (guarantees only for Rk∗ ). They allow
the scientist to spot one or more suitable rejection sets (select post-hoc on the algorithm’s trajectory) by picking
d k ) and the uniform upper band on
data-dependent sizes or error-levels, after examining the entire path of FDP(R
FDP. This post hoc mode of inference addresses both aforementioned drawbacks of FDR. Our bounds apply to
online FDR procedures as well, providing simultaneous high-probability uniform FDP bounds at arbitrary datadependent query times for arbitrary online procedures, including alpha-investing, LORD and SAFFRON. Our
analysis unifies existing martingale and empirical process viewpoints on FDR algorithms.
Keywords: false discovery rate, FDR, multiple testing, post hoc confidence bounds, false discovery exceedence, FDX, simultaneous selective inference, optional spotting, uniform martingale concentration.
1
Introduction
The false discovery rate (FDR) [1] is a popular Type-I error criterion for multiple testing problems. Given a set
of hypotheses H1 , . . . , Hn and a rejection set R ⊂ [n] obtained based on a set of p-values p1 , . . . , pn , the false
discovery proportion is defined by
V
|R ∩ H0 |
FDP(R) ≡
≡ ,
(1)
|R|
R
where H0 ⊂ [n] is the set of nulls. By convention, FDP ≡ 0 when R = ∅ (we use the ≡ symbol for definitions).
The false discovery rate is defined as the expectation of the false discovery proportion: FDR ≡ E [FDP].
2
1.1
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
Offline versus online multiple testing
Multiple testing may occur in a batch (or offline) context, where all p-values are available ahead of time, or in an
online context, where p-values arrive in a (potentially infinite) stream. Given an FDR target level q, e.g. q = 0.1,
the usual goal in the offline context is to find one set of rejections R for which
FDR ≡ E [FDP(R)] ≤ q.
(2)
In the online context, one tests a possibly infinite sequence of hypotheses, maintaining a rejection set Rk ⊇
Rk−1 at each time k, by irrevocably deciding whether to reject Hk immediately after seeing pk , but before seeing
pk+1 , . . . . These decision are made to ensure that for any k ∈ N, we have
FDRk ≡ E [FDP(Rk )] ≤ q.
(3)
In the batch setting, the most widely adopted FDR procedure is the Benjamini-Hochberg procedure [1], which
guarantees the desired bound (2) if the p-values are independent or positively dependent [2]. While the BH procedure satisfies certain optimality criteria when the hypotheses are treated as exchangable and there is no structure
to exploit [3], this fact is not true in the presence of prior information or structural constraints. Recognition of this
fact has led to the development of more powerful procedures that are applicable in a wide variety of settings, and
we will discuss many such examples in this work, like ForwardStop for sequential model selection [4], Selective
SeqStep for variable selection using knockoffs [5], multilayer grouped constructions like p-filter [6, 7, 8], accumulation tests for ordered testing [9], and interactive human-in-the-loop protocols like AdaPT [10] and STAR [11].
The online FDR literature began with Foster and Stine’s alpha-investing algorithm [12], with more recent work
proposing some generalizations [13], and also more powerful alternatives like LORD [14, 15] and SAFFRON [16].
1.2
Criticism of the FDR, and alternatives
The FDR and associated methodologies have gained broad acceptance across many application domains. This approach is attractive compared to controlling the family-wise error rate (FWER), defined as FWER = Pr{V > 0},
because it often affords more power while retaining a meaningful type-I error guarantee. Especially in modern
large-scale testing applications, requiring the absence of false discoveries with high probability is much too stringent. Despite the advantages of FDR, this error rate is not without its pitfalls. Its two primary weaknesses, which
several authors have highlighted already, are as follows.
FDR control gives no guarantee on FDP variability. The FDR by definition only controls the expectation of
the FDP. Hence, the data set at hand might have an FDP much higher than the nominal FDR target level. In this
paper, we show that even for independent p-values, some FDR procedures may exhibit high FDP variability. This
criticism also applies to the online context.
In order to remedy this issue, it has been proposed that the FDP be controlled with high probability, instead of
just on average:
Pr{FDP > γ} ≤ α.
The quantity on the left hand side is sometimes called the false discovery exceedance (FDX). Bounding the FDP
with high probability provides a high-confidence guarantee for a given dataset as opposed to a guarantee on average
across potential data sets. Several related step-down procedures have been proposed for FDX control in the offline
setting [17, 18, 19, 20, 21, 22], and a rather conservative Bonferroni-like approach by Javanmard and Montanari
(JM) [14] has been proposed in the online setting.
The standard multiple testing paradigm, and FDR control in particular, provides little or no flexibility.
Whether controlling FDR or FDX, the standard multiple testing paradigm involves choosing a target level q (or
bound γ) before looking at the data and formally cannot be changed afterwards. Even for a prespecified target
level q, FDR procedures do not allow for post hoc analyses. In the offline context, the rejection set R cannot be
modified (shrunk or expanded or altered otherwise) in a post hoc fashion without invalidating the guarantee. In the
3
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
online setting, we have no guarantees on FDPk at random times in the past, since the FDR guarantee (3) holds at
all fixed times but not uniformly over time.
This rigid multiple testing paradigm has limited applicability for exploratory data analysis, in which additional
flexibility is desirable, as eloquently argued by Goeman and Solari (GS) [23], as well as by earlier authors [24, 25].
GS argue that this inflexibility clashes with the practice of exploratory data analysis, where it is common to re-run
the algorithm with multiple target levels to get a set of the desired size, or to choose a subset of the rejection set
that makes sense based on the user’s domain knowledge.
To allow more flexibility in multiple testing, procedures have been designed [26, 27, 28, 29, 23, 30] to provide
high-probability bounds FDPα simultaneously for all subsets R ⊂ [n]:
Pr FDP(R) ≤ FDPα (R) for all R ≥ 1 − α.
(4)
While FDR and FDX are instances of selective inference, where inferential guarantees are made only on the
selected (i.e. rejected) set of hypotheses, statement (4) represents simultaneous inference, providing guarantees
on all 2n possible rejection sets. Goeman and Solari (GS) [23] argue that simultaneous inference paves the way
for a more flexible paradigm for multiple testing, called exploratory multiple testing. This paradigm reverses the
traditional roles of the data analyst and multiple testing algorithm. Instead of the data analyst providing an FDR
target level and the algorithm returning a rejection set, the algorithm provides simultaneous FDP bounds for all
sets, and then the data analyst may choose a rejection set however she wishes. For this reason, bounds like (4)
are also called post hoc bounds, bringing multiple testing procedures more in line with the practice of exploratory
data analysis. In addition to resolving the issue of a priori selection of the target level, exploratory multiple testing
allows data analysts to apply their domain knowledge, intuition, or any other outside information to examine
multiple candidate rejection sets and choose the one that makes the most sense for the given application.
1.3
Preview: Uniform FDP bounds for existing FDR procedures
Despite the weaknesses of FDR, this criterion and procedures controlling it are used widely in practice. Moreover, as described in Section 1.1, much progress has been made in designing FDR procedures in a variety of
settings. Moreover, while the FDR criterion itself does not provide any guarantees on FDP variability, at least
under independent (or weakly dependent) p-values, many of these procedures have been empirically found to have
well-behaved FDP. In particular, it is well-known the the BH procedure has a tight FDP histogram under independence. Hence, one of the goals of this paper is to formally study the FDX performance of existing FDR procedures
under independent p-values.
To get a handle on this question, we study a construction common to many offline FDR procedures, including
BH and nearly all others mentioned in Section 1.1. These procedures can be each viewed as constructing a path of
rejection sets
∅ = R0 ⊆ R1 ⊆ · · · ⊆ Rn ⊆ [n].
(5)
d k ) of false discovery proportion in Rk is defined. FiThen, for each k, an algorithm-dependent estimate FDP(R
nally, the algorithm rejects the set Rk∗ , where
d k ) ≤ q}.
k ∗ = max{k : FDP(R
(6)
It turns out that for all these procedures, when the p-values are independent, one can summarize a proof that they
control the FDR in a single line. One begins by constructing an appropriate backwards filtration Ωk with respect
to which k ∗ can be viewed as a stopping time, and then identifying a nonnegative backwards supermartingale Mk
with E [Mn ] ≤ 1, such that
FDP(Rk )
V (Rk )
=
≤ Mk .
(7)
d k)
FDP(R
Vb (Rk )
It is often an art to identify Mk , Ωk , but after doing so, the proof of the fact that FDR = E [FDP(Rk∗ )] ≤ q reduces
to the following simple argument that applies inequality (7) at stopping time k ∗ :
"
#
(ii)
FDP(Rk∗ ) (i)
FDP(Rk∗ )
≤E
≤ E [Mk∗ ] ≤ 1,
(8)
E
d k∗ )
q
FDP(R
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
4
03.30.18
where inequality (i) follows by definiton of k ∗ , and inequality (ii) follows by the optional stopping theorem. One
take-away message from the above proof is that, for a variety of existing procedures, proving that FDR ≤ q actually
amounts to establishing the fact that
"
#
FDP(Rk∗ )
E
≤ 1.
(9)
d k∗ )
FDP(R
The central technical claim of this paper is that, in fact, a similar kind of statement can be made uniformly over
the entire path.
Main result: For a variety of FDR procedures working with independent p-values, FDP(Rk ) is only a constant
d k ) (or a small variant thereof), uniformly over k with high probability. More precisely,
factor larger than FDP(R
we prove for several FDR algorithms that
(
)
FDP(Rk )
≤ calg (α) ≥ 1 − α,
Pr sup
(10)
d k)
k FDP(R
where calg (α) are explicit algorithm-dependent constants.
By integrating the bound (10) across α, we also obtain results of the form
"
#
FDP(Rk )
E sup
≤ Calg .
d k)
k FDP(R
(11)
Here, Calg are algorithm-dependent constants that are not only independent of the number of hypotheses n, but
remarkably, are less than 2 in most cases. If we compare inequality (11) with statement (9), we find that by paying
a factor of about 2 or less, we move from a bound at the point k ∗ to a uniform bound across the whole path.
1.4
Towards “simultaneous selective inference”: from stopping times to spotting times
Here, we briefly summarize some interpretations and implications of our bounds. To begin, it is possible to interpret
the bounds (10) and (11) in the context of the stopping time k ∗ in definition (6), or in fact also for any other post
d k∗ ) ≤ q. For such an index k ∗ , our bounds imply that
hoc means of selecting k ∗ for which FDP(R
Pr{FDP(Rk∗ ) ≤ calg (α) · q} ≥ 1 − α,
(12)
E [FDP(Rk∗ )] ≤ Calg · q.
(13)
and also that
For this reason, we call calg (α) an “FDP multiplier” and Calg an “FDR multiplier”. Note that inequality (12) states
something slightly surprising about existing FDR procedures: though designed to control the expected FDP at one
point, they automatically control FDX uniformly to within a multiplicative factor of the nominal level. In addition,
inequality (13) is useful for proving FDR guarantees in contexts where k ∗ has a complex definition (e.g. cannot be
described as the stopping time with respect to a backwards filtration). This partially alleviates the FDP variability
issue for FDR procedures, at least under independence.
Note that statements (12) and (13) apply to the context where a target FDR level q is specified in advance.
However, an equally important interpretation of the claim (10) is to view it as a uniform upper confidence bound
(or confidence envelope) for the unknown FDP along the path of the algorithm:
n
o
d k ) for all k ≥ 1 − α.
Pr FDP(Rk ) ≤ calg (α)FDP(R
(14)
d k ), we see that inequality (14) is similar to inequality (4), except it is a
Defining FDPα (Rk ) = calg (α)FDP(R
uniform statement over the path of an FDR procedure instead of over all subsets. Hence, statement (14) also
represents a solution to the issues concerning FDR laid out in Section 1.2: the uniformity of the bound provides
flexibility and its FDX formulation avoids the FDP variability problem of traditional FDR methods.
5
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
For these reasons, we refer to statements (10) and (11) as “optional spotting” bounds, where to spot is to Select
Posthoc On the Trajectory of the algorithm (the last term refers to the fact that each FDR algorithm effectively
uses the information available at hand to pick sets of shrinking sizes from the full set to the empty set, carving out
a trajectory of at most n + 1 sets out of the 2n possible sets).
Spotting is a variant of GS’s vision for exploratory multiple testing, where the user is first presented with the
d k ), calg (α)FDP(R
d k )) and the user has the option of “spotting” one
full trajectory of n + 1 triplets (Rk , FDP(R
or more sets whose combination of size, content, and high-probability FDP bound is to their liking. Reporting
the results on only some spotted triplets is statistically valid and mathematically principled, since the underlying
bounds are uniform and simultaneous over all n + 1 sets. The term optional spotting is also used to explicitly
contrast of our bounds (10) and (11) to the standard FDR guarantee (9), usually proved via optional stopping
arguments. Moving from stopping to spotting resolves both aforementioned drawbacks of FDR procedures at
the cost of a small multiplicative constant. Since this approach represents a middle path between fully selective
inference (guarantees on the single rejection set Rk∗ ) and fully simultaneous inference (guarantees for all possible
2n rejection sets), we term our approach as “simultaneous selective inference”. Indeed, the FDP is controlled
simultaneously, but only over the sets that were selected by the FDR algorithm. This approach is a promising path
towards flexible, powerful, and computationally tractable multiple testing procedures.
1.5
Summary of contributions and paper outline
Some contributions of our paper are briefly summarized below, after which we briefly describe the organization of
the rest of this paper.
• For several existing FDR procedures (or minor modifications thereof), we find tight and explicit constants
d process, assuming
bounding with high probability the extent to which the FDP process exceeds the FDP
independent p-values.
• We accomplish this by developing a simple yet versatile proof technique to obtain tight non-asymptotic
bounds for the probability that the stochastic process V (Rk ) hits the boundary c(α)Vb (Rk ).
• In the batch context, the corollary (12) shows that existing FDR procedures (or minor modifications thereof)
also control FDX to within a constant factor c(α) of the nominal target level q. For BH, c(α) ≈ 2.1
for α = 0.05, justifying theoretically the common observation that the FDP of BH is well-behaved under
independent p-values.
• In the online context, we prove statement (10) for a broad class of online procedures, covering all existing
online FDR algorithms and more. This provides high-probability bounds on FDP for any data-dependent
query times in the present or past.
• The reformulation (14) provides a uniform confidence envelope for the FDP process, echoing prior works
like GS [23] and GW [29]. This points towards an “optional spotting” approach to multiple testing that
addresses three common criticisms of traditional FDR methods.
• Since our bounds are fairly tight, the differences in FDP multipliers across procedures reflects differences in
the inherent variability of FDP for these procedures. Using this observation, we can quantitatively study the
trade-off between power and FDP quantiles.
Having overviewed our results and their implications, the remainder of the paper will focus on describing
the corresponding multiple testing procedures in more detail (Section 2), formally stating these results (Section 3),
discussing their implications for multiple testing (Section 4) and beyond (Section 5), exploring potential extensions
to the correlated setting (Section 6), and proving these results (Section 7). We conclude the paper in Section 8.
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
6
2
03.30.18
FDR controlling procedures: a unified review
The purpose of this section is two-fold: (a) to recap existing classes of FDR procedures for the uninitiated (especially since much rapid progress has been made in the past few years), and (b) to present the variety of procedures
using a somewhat unified viewpoint. Readers that are familiar with the literature may still benefit from the unified
view that we present here, which has not been provided elsewhere to the best of our knowledge. The impatient
reader may skip to Section 3, where we present our main results.
To begin, we note that in both the batch and the online contexts, a path (5) of rejection sets is constructed.
This path may be finite or infinite, depending on the context. For batch multiple testing, the path represents a
sequence of increasingly liberal rejection sets, among which one member (indexed by k ∗ ) is finally chosen as the
reported rejection set Rk∗ that satisfies the desired FDR bound (2). In this setting, the path is based on an ordering
π(1), . . . , π(n) of the hypotheses, constructed so that non-nulls are more likely to occur early in the ordering.
This ordering may be specified in advance or constructed data-adaptively. On the other hand, in the case of online
testing, the FDR guarantee is on each set Rk on the path (recall the definition (3)), and the ordering π is dictated
by the order in which hypotheses arrive in the stream, which might not be in the user’s control.
In the batch setting, in order to choose which set to reject, several algorithms adopt a common paradigm: the
FDP for each set Rk is estimated via
b
d k ) ≡ V (Rk ) ,
(15)
FDP(R
|Rk |
where Vb is a (possibly data-dependent) estimate of Vk ≡ |H0 ∩ Rk |. Then, the rejection set Rk∗ is defined via
d k ) ≤ q}.
k ∗ ≡ max{k : FDP(R
(16)
d Several
Batch FDR procedures vary based on the definitions of the ordering π, the path Rk and the estimate FDP.
examples are detailed in this section, and our theoretical guarantees will apply to all of them, such as
(a) Benjamini-Hochberg (BH) [1],
(b) accumulation tests [9], Selective SeqStep [5], and Adaptive SeqStep [31] for ordered testing, or
(c) AdaPT [10] and STAR [11] for interactive testing.
We first discuss batch procedures mentioned in items (a), (b) and (c), and then the case of online testing.
2.1
The BH procedure
In this procedure, π is the ascending ordering of the p-values, i.e. hypotheses are considered in order of increasing
p-value. Despite the discreteness of the path, it is easier to parameterize it using a continuous p-value threshold
t ∈ [0, 1]. Then, the rejection set corresponding to t is
Rt ≡ {j ∈ [n] : pj ≤ t},
d is based on the estimate
and FDP
Vb (Rt ) ≡ nt.
P
Note that E [Vt ] ≡ E [Rt ∩ H0 ] = j∈H0 Pr{pj ≤ t} = |H0 |t. Hence, Vb (Rt ) is an upwardly biased estimate of
Vt , the bias being fairly small if |H0 | ≈ n.
2.2
Ordered testing with a fixed, pre-defined order
As in BH, the hypothesis order π for ordered testing is chosen so that non-nulls are likely to occur near the
beginning of the ordering. However, unlike BH, the ordering π is specified in advance based on prior knowledge,
7
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
meaning that the hypotheses are ordered as Hπ(1) , Hπ(2) , . . . , Hπ(n) . Given such a predefined ordering π, two
kinds of rejection sets have been considered:
Rk ≡ {π(j) : j ≤ k};
(R1)
Rk ≡ {π(j) : j ≤ k, pπ(j) ≤ p∗ },
(R2)
for some pre-specified constant p∗ . The first choice corresponds to rejecting the first k hypotheses in the ordering
regardless of p-value, whereas the second corresponds to rejecting those among the first k with p-values below
a threshold p∗ . The definition of Rk clearly impacts the choice of Vb , so the two choices above lead to different
constructions of Vb .
Case (R1). For this choice of rejection sets, a family of accumulation tests has been proposed [9]. An accumulation function is a nondecreasing function h : [0, 1] → [0, ∞) that integrates to 1:
Z 1
h(u)du = 1.
0
For example, for λ ∈ (0, 1), the SeqStep [5] accumulation function is
h(pj ) ≡
I(pj > λ)
.
1−λ
1
Other choices include ForwardStop [4], which used h(u) = log( 1−p
), and HingeExp [9], which used h(pj ) =
1
1−λ
1−λ log 1−pj I(pj > λ)). In all cases, we note that Eu∼U [0,1] [h(u)] = 1, which inspires the definition
Vb (Rk ) ≡ B +
k
X
h(pπ(j) ),
(17)
j=1
for some constant B, often chosen as B = supu∈[0,1] h(u) if the latter quantity is finite. From the defining
properties of h, it follows that Vb (Rk ) is an upwardly biased estimator for Vb . The extra additive term B is a
regularization employed to ensure exact finite-sample FDR control. This FDR control does not hold without
regularization due to the unstable behavior of Vb near the beginning of the path; we return to this issue in Section
3. As a minor note, the original work on accumulation tests [9] required a slight modification of the definition
(15): they use the denominator 1 + |Rk | instead of |Rk |. For consistency with the other procedures discussed in
this paper, we deviate from the original authors’ proposal, and refer to the accumulation test as the procedure with
d defined via equations (15) and (17). We explicitly specify if we wish to refer to the original accumulation test
FDP
with 1 + |Rk | in the denominator.
Case (R2). The first procedure proposed for this case was Selective SeqStep [5], and a generalization called
Adaptive SeqStep [31] was later proposed. Selective SeqStep is defined via
k
X
p∗
Vb (Rk ) ≡
1+
I(pπ(j) > p∗ ) .
(18)
1 − p∗
j=1
Note that if p∗ is small, then even non-nulls are likely to contribute to Vb (Rk ), making the procedure more conservative. To fix this issue, Adaptive SeqStep [31] instead defines
k
X
p
∗
Vb (Rk ) ≡
1+
I(pπ(j) > λ) ,
(19)
1−λ
j=1
where λ > p∗ . For example, we might define p∗ = 0.1 and λ = 0.5.
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
8
2.3
03.30.18
Interactive algorithms (ordered testing with an adaptive order)
Recently, it was noted that ordered testing procedures described in the previous subsection can be improved and
generalized to a variety of settings, by constructing the ordering π data-adaptively, and possibly interactively with
a scientist-in-the-loop. Two examples are noted below.
AdaPT. Working in the context of definitions (R2) and (18) with p∗ = 0.5, Lei and Fithian note that only part
of the information in the p-values is being used at the inference step (namely I(p > 1/2)), so the “left over”
information may be used to help construct the ordering. This idea was inspired by the concept of data carving
[32]. More precisely, this left over information is captured by the masking function g(p) = min(p, 1 − p). Indeed,
note that min(u, 1 − u) and I(u > 1/2) are independent for uniformly distributed u. They showed that this
information g(pj ) can be used freely in order to determine the ordering π while maintaining the FDR guarantee
based on definition (18). Based on this idea, they proposed AdaPT (adaptive p-value thresholding), where the
ordering is constructed based on g(pj ) as well as arbitrary side information xj attached to each hypothesis. AdaPT
builds up the ordering π backwards, starting from π(n), at each step k using the information
Gk ≡ σ({xj , g(pj )}j∈[n] , {pπ(j) }j>k ),
(20)
to determine π(k). Note that in addition to the masked p-values and the side information, the AdaPT algorithm
gives access to unmasked p-values as well for hypotheses that have been excluded from the rejection set.
STAR. Despite the many benefits of AdaPT, it cannot guarantee that the final rejection set obeys any userspecified constraints or desired structural properties, and further, one was restricted to using the pair min(p, 1 − p)
and I(p > 1/2). This motivated the development of STAR (selectively traversed accumulation rules) [11], that
generalized the notion of data carving to any accumulation function (17), by showing that for any non-decreasing
h, a corresponding masking function g may be constructed. Like AdaPT, STAR also builds up the ordering from
back to front, using the information in the filtration (20), and also guaranteeing access to unmasked p-values for
hypotheses outside the rejection set. However, unlike AdaPT which uses definition (R2), the STAR procedure uses
definition (R1), allowing it to guarantee arbitrary user-defined structural constraints on the rejected sets (in fact,
these constraints can themselves be data-dependent).
We refer the reader to the original works for more intuition for how interaction is enabled by masking p-values.
In particular, the original authors describe many new problems for which we can control the FDR while having a
human-in-the-loop, possibly gaining power due to the human’s ability to use their intuition, prior knowledge and
flexible modeling capabilities. Here, we just point out that our bounds and resulting implications apply to both
AdaPT and STAR.
2.4
Online testing
Batch ordered testing procedures attempt to ensure that non-nulls occur early in the ordering (either using just
the raw p-values like in BH, or using just prior information like in ordered testing, or using a combination of
prior information, covariates and masked p-values in the case of AdaPT and STAR). In the online context, two
distinct situations can arise. If the order of the hypotheses is in the experimenter’s control, then once more it is
advantageous to attempt to use any available side information to have the non-nulls appear early in the ordering.
However, if the entire order of hypotheses is not in the experimenter’s control, then non-nulls might not necessarily
occur early, and further the rejection decision for the Hk must be made without knowing the outcomes of future
experiments. Hence, in general, online multiple testing procedures must proceed differently from aforementioned
batch experiments, since they may prefer not to reject most of the first k hypotheses while accepting the rest, and
may not wish to test all hypotheses at a fixed level. Instead, online procedures produce a sequence of levels αj at
which to test hypotheses, resulting in rejection sets of the form
Rk ≡ {π(j) : j ≤ k, pπ(j) ≤ αj }.
(21)
9
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
These levels αj are chosen based on the outcomes of past experiments. Foster and Stine’s alpha-investing procedure [12] and follow-up works [13, 14] are built on the analogy of testing a hypothesis at level αj as spending
wealth. One pays a price to test each hypothesis, and gets rewarded for each rejected hypothesis. In particular,
generalized alpha-investing (GAI) [13] is defined as follows. The procedure begins with wealth W0 ≤ α. At step
j, the algorithm pays a price of ϕj to test the jth hypothesis, and gets a reward ψj if this hypothesis is rejected:
Wj ≡ Wj−1 − ϕj + Rj ψj , where Rj = I(pj ≤ αj ). The rules for defining ϕj and ψj must be chosen so that
Wj ≥ 0 at all times j. A particular GAI procedure called LORD [14] chooses ϕj = αj , and ψj ≤ α.
Until recently, online FDR procedures were formulated without reference to an estimate of FDP. However, it
d defined by
was recently noted [15] that LORD implicitly keeps track of an FDP
Vb (Rk ) ≡
k
X
αj ,
(22)
j=1
d j ) ≤ q for all j. This places
reformulating the requirement that Wj ≥ 0 for all j as the requirement that FDP(R
LORD in the same framework as the batch algorithms considered in this paper. Moving beyond LORD, an adaptive
algorithm called SAFFRON [16] was recently proposed, which is defined by equation (21) and
Vb (Rk ) ≡
k
X
j=1
αj
I(pπ(j) > λj ).
1 − λj
(23)
SAFFRON improves upon the naive estimate (22) by correcting for the proportion of nulls, making it the online
analog of the Storey-BH procedure [33], or of Adaptive SeqStep (19). Like the levels αj , the constants λj may
also be chosen based on the outcomes of prior experiments. Interestingly, the original alpha-investing algorithm
is recovered by picking λj = αj , and the LORD estimate (22) is recovered by choosing λj = 0 (however, the
SAFFRON algorithm does not reduce to the LORD algorithm when λj → 0, because for nonzero λj , SAFFRON
requires λj ≥ αj for all j). Simulations [16] demonstrate that SAFFRON with constant λ = 1/2 successfully
adapts to the null proportion and enjoys higher power than LORD as a result.
2.5
Tying together ordered, interactive and online testing algorithms
By examining the procedures in the preceding sections, it is clear that they share many similar features. Here, we
make these similarities explicit by situating all ordered, interactive, and online testing algorithms in a common
framework. The first part of the framework is the definition (21) of the rejection sets, considered before in the
online context. Next, we note that (R1) and (R2) are recovered as special cases of definition (21), corresponding
to αk = 1 and αk = p∗ respectively. Lastly, we note that each definition of Vb can be written as
Vb (Rk ) ≡ a +
k
X
hj (pπ(j) )
(24)
j=1
for some additive regularization a and some functions hj . We call hj generalized accumulation functions, because
for uniform [0, 1] random variables U , they satisfy
Z
1
hk (u)du = αk = Pr{U ≤ αk } .
E [h(U )] = Eu∼U [0,1] [hk (u)] =
(25)
0
The above property allows us to generalize all earlier constructions, which satisfy the intuitive property of hk
obtained by comparing the leftmost and rightmost expressions in equation (25). Table 1 summarizes how the
different ordered, interactive and online testing procedures fall into this framework.
Having introduced several different classes of FDR procedures and presented them in a unified fashion, we
now turn to stating our uniform FDP bounds for these procedures.
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
10
Algorithm
Accumulation test
STAR
Selective SeqStep
AdaPT
Adaptive SeqStep
LORD
SAFFRON
a
a ≡ supu h(u)
a ≡ supu h(u)
p∗
a ≡ 1−p
∗
p∗
a ≡ 1−p
∗
p∗
a ≡ 1−λ
a≡0
a≡0
αk
αk ≡ 1
αk ≡ 1
αk ≡ p∗
αk ≡ p∗
αk ≡ p∗
(defined via GAI)
(defined via GAI)
hk
hk ≡ h
hk ≡ h
p∗
hk (pk ) ≡ 1−p
I(pk > p∗ )
∗
p∗
I(pk > p∗ )
hk (pk ) ≡ 1−p
∗
p∗
hk (pk ) ≡ 1−λ
I(pk > λ)
hk (pk ) ≡ αk
αk
hk (pk ) ≡ 1−λ
I(pk > λk )
k
03.30.18
π
fixed
adaptive
fixed
adaptive
fixed
(stream)
(stream)
Table 1: Ordered, interactive and online testing algorithms.
3
Main results
Before presenting our central theorems, we briefly discuss three subtle issues: additive regularization, reversal of
interactive procedures, and the use of arbitrary online rules. We will return to the first issue in more detail in the
next section, after presenting the main results.
Regularization. As noted earlier, past work has often used some additional regularization (denoted by the additive quantity a in the definition of Vb ) to ensure exact FDR control for many of the procedures under consideration.
Despite the fact that regularization is not necessary for FDR control for the other procedures—BH, LORD, and
SAFFRON—the behavior of Vb is nevertheless unstable near the beginning of the path. For example, consider BH
under the global null. Then,
(
)
)
(
Pn
FDP(Rt )
1
j=1 I(pj ≤ t)
Pr sup
>x = ,
> x = Pr sup
d
nt
x
t∈[0,1] FDPa=0 (Rt )
t∈[0,1]
as shown by Robbins in 1954 [34], which is in sharp contrast to the exponentially decaying tails that we derive
below using regularization. Without regularization, we would achieve an impractical
and nearly trivial
h
i confidence
FDP(Rt )
bound of c(α) = 20 for α = 0.05. Moreover, integrating across x gives E supt∈[0,1] FDP
d a=0 (Rt ) = ∞. This
motivates us to consider all procedures with regularization. In fact, we will prove bounds for Vb defined in equation
(24) with arbitrary a > 0, not just those dictated by Table 1. The only exception is BH, for which we will
take a = 1 by default. Let BH+, LORD+, and SAFFRON+ denote the regularized counterparts of the original
procedures. Specifically, note that
VbBH+ (Rt ) = 1 + nt.
(26)
Reversal of STAR and AdaPT. As discussed in Section 2.3, STAR and AdaPT build up the ordering π from
back to front, at each step choosing the least promising hypothesis to remove from the rejection set. The p-values
are masked until they are removed from the rejection set, at which point they are unmasked. We find reversed
versions of STAR and AdaPT more amenable to our theoretical analysis. Let rSTAR and rAdaPT be procedures
analogous to STAR and AdaPT, but which build up the ordering π from front to back; i.e. π must be predictable
with respect to the filtration
Gkr ≡ σ({xj , g(pj )}j∈[n] , {pπ(j) }j≤k ).
(27)
Hence, rAdaPT and rSTAR begin with R0 = ∅, and then at each step choose the most promising hypothesis to
add to the ordering. In this case, p-values are masked until they are added to the ordering, after which they are
unmasked. Note the difference between the filtrations (27) and (20); the p-values are unmasked upon entering the
path, rather than upon exiting the path. We do not anticipate this to change the behavior of these procedures by
much at all, and neither the original nor reversed procedures will uniformly dominate each other from the perspective of power. Indeed, as discussed in the original papers, the masked p-values contain most of the information in
the p-values, and the additional information obtained from unmasking a p-value is fairly incremental (to see this,
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
11
03.30.18
note that in the special case of h(pj ) = I(pj > 1/2), only a single bit per p-value is hidden at the start, and this is
all that is gained by unmasking). It is unlikely that there is a major difference between a top-down (full to empty)
and a bottom-up (empty to full) ordering, since the original and the reversed procedures both receive all masked
p-values right at the start, and both learn one unmasked p-value per step.
Arbitrary levels αk for online procedures. Whereas alpha-investing, LORD and SAFFRON each correspond
to a particular prescribed set of rules that decide the levels αk , our theoretical results hold whenever αk is any
predictable sequence, whether or not it was derived from any of the aforementioned online FDR algorithms. A
particularly extreme example would occur when αk is set to an absolute constant for all k; such a rule will not
guarantee control of the FDR, but our results will provide uniform bounds on the achieved FDP of such an online
algorithm at any time.
3.1
Optional spotting theorems
We first state our results for BH+, and then for ordered testing, interactive testing, and online testing. The results
below are illustrated in Figure 1, which shows the FDP multipliers (with α = 0.05) from Theorems 1, 2, 3, and 4
as well as the FDR multipliers from Corollary 1.
Theorem 1. (FDP multipliers; BH+). Suppose the null p-values are independent and uniformly distributed. Then,
the BH+ procedure (26) satisfies statement (10) for all α ≤ 0.31, with
cBH+ (α) ≡
log( α1 )
.
log 1 + log( α1 )
(28)
We remark that the range α ≤ 0.31 covers most practically relevant choices, but we conjecture that the bound
actually holds for all α. Next, we state the bounds for ordered testing procedures.
Theorem 2. (FDP multipliers; ordered testing). Suppose the null p-values are independent and superuniformly
distributed, i.e. Pr{pj ≤ s} ≤ s for all j ∈ H0 and s ∈ [0, 1].
1. Accumulation test: Given any (possibly unbounded) accumulation function h, and any additive regularization a, statement (10) holds for all α, with
log( α1 )
chacc (α) ≡
a log
R
1
0
−1 .
(29)
αh(u)/a du
Moreover, for any accumulation function bounded by B, we have
log( α1 )
chacc (α) ≤ cB
acc (α) ≡
a log
1−
1−αB/a
B
−1 .
(30)
For the original formulation of the accumulation test (additive regularization B, with 1 + |Rk | in the ded we have
nominator of FDP),
1
log(
)
1
1 +
α .
c̃B
(31)
acc (α) ≡
B
log B−α
B−1
2. Adaptive SeqStep, Selective SeqStep: Noting that the generalized accumulation function used by Adaptive
p∗
p∗
, statement (10)
SeqStep is bounded by B = 1−λ
, and that of Selective SeqStep is bounded by B = 1−p
∗
holds for all α, with:
log( α1 )
.
(32)
cB
(α)
≡
AS
B/a
a log 1 + 1−αB
12
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
When the ordering is not fixed, but is interactively determined, we have the following result.
Theorem 3. (FDP multipliers; interactive setting). If the null p-values are independent of each other and of the
non-null p-values, and the null p-values have nondecreasing densities, then statement (10) holds for all α with
B
B
B
cB
rAdaPT (α) ≡ cAS (α) and crSTAR (α) ≡ cacc (α). The definitions of B for rAdaPT and rSTAR mimic those for
Selective SeqStep and accumulation test respectively.
Lastly, shifting to the online FDR setting, we derive the following guarantee.
Theorem 4. (FDP multipliers; online setting). Let p1 , p2 , . . . be a stream of p-values adapted to a filtration Gk .
If the null p-values are superuniformly distributed conditional on the past, meaning that
Pr{pk ≤ s|Gk−1 } ≤ s for each k ∈ H0 and each s ∈ [0, 1],
(33)
then the bound (10) holds with the following constants:
1. LORD+. If αk is predictable with respect to Gk , then we have
cLORD+ (α) ≡
log( 1 )
α 1 .
log( )
a log 1 + a α
2. SAFFRON+. If αk and λk are predictable with respect to Gk , with αk ≤ λk for all k, and supk
B < ∞, then we have cSAFFRON+ (α) ≡ cB
AS (α).
(34)
αk
1−λk
≤
d and not necessarily
Importantly, the names LORD+ and SAFFRON+ refer to the formulae used to calculate FDP,
to the rule used to derive the sequence αk . Indeed, one can theoretically use the LORD (or any other) rule to
d
derive αk , but give a post hoc uniform guarantee on the achieved FDP using the SAFFRON+ formula for FDP.
The following FDR multipliers can be calculated by integrating the above high-probabilty tail bounds.
Corollary 1. (FDR multipliers). For each of the above procedures except BH+, the bound (11) holds with
Z 1
Calg ≡
calg (α)dα.
0
R 0.31
For BH+, we instead have CBH+ = 0 cBH+ (α)dα + 0.69cBH+ (0.31). The above calculations typically yield
constants smaller than 2; for example CBH+ = 1.62, CLORD+ = 1.42 for a = 1, and CAS = 1.92 for B = 2 and
a = 1. See Figure 1 for more examples of FDR multipliers.
Remark 1 (A connection to prior work). A bound of the kind stated in Corollary 1 was first proved by [8] for
Selective SeqStep, with a = 1 and p∗ = 0.5, in the context of the multilayer knockoff filter. The goal of that
work was to prove FDR control for a predefined target level q with respect to multiple hypothesis groupings;
the uniform bound was necessary because the more complicated form of k ∗ in that context precluded the use of
standard martingale arguments.
Remark 2 (Assumptions for different procedures). The optional spotting theorems rely on slightly different assumptions on the distributions of the p-values. Note that Theorems 1 and 2 make assumptions only on the distribution (pj )j∈H0 , not requiring independence between null and non-null p-values. This is in contrast to FDR
results established before [1, 9] for the corresponding procedures, which do require independence between null
and non-null p-values. On the other hand, our results for ordered, interactive and online procedures result assume
dependence conditions that are nearly identical to the conditions that appear in the original works. Further, note
that Theorem 1 requires uniform null p-values, whereas the other theorems have looser restrictions on the distributions of null p-values. For ordered testing, we allow superuniform null p-values, an assumption that is at least
as general as those made for the original methods. For the interactive methods, we assume that null p-values have
an increasing density, which is the same assumption made by STAR and similar to the one made by AdaPT. Finally, our assumption of superuniformity conditional on the past for online procedures is similar to those made in
previous works [12, 14, 16].
13
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
Remark 3 (The constants for BH+). In Theorem 1, we require that α ≤ 0.31. However, strong numerical evidence
shows that the bound (10) is valid for BH+ with cBH+ (α) for all α. The restriction on α is an artifact of our proof
and does not represent an intrinsic breaking point of the bound. Despite this limitation in our proof, the range
α ≤ 0.31 comfortably includes any confidence level that would be used in practice. This limitation leads to an
estimate of CBH+ = 1.62, that is slightly more conservative compared to what we observed in simulations, which
suggest that in reality, CBH+ ≈ 1.35. Independently, we note that our constant being smaller than 2 seems morally
related to Eq. (12) in work by Goeman et al. [35], where the authors were interested specifically in FDP-quantiles
of BH (but did not derive any closed form bounds).
Figure 1: FDP multipliers for α = 0.05 (red) and FDR multipliers (salmon) for each method for a = 1. Where
applicable, multipliers for original regularizations shown using dashed lines.
3.2
Tightness of the bounds
Here, we investigate the tightness of our bounds using numerical simulations. For all simulations, we chose
n = 500. We simulated independent test statistics Xj = µj +j , where j ∼ N (0, 1), for j = 1, . . . , n, and derived
p-values pj = 1 − Φ(Xj ). Here, µj = µI(j ∈ H1 ). We chose a signal strength of µ = 4; other signal strengths
produced qualitatively similar figures. We simulated four sparsity settings, corresponding to |H1 | = 0, 10, 25, 50.
The locations of the signals were either chosen randomly, or biased towards the beginning. For each simulation
k)
setting, we generated 10,000 Monte Carlo realizations of supk FDP(R
.
d
FDP(R
k)
Figure 2 shows the theoretical FDP multipliers (in black) and empirical realizations (in shades of blue) for four
batch procedures: BH+, accumulation test with SeqStep accumulation function (p∗ = 0.5), Selective SeqStep with
p∗ = 0.5, and Adaptive SeqStep with p∗ = 0.05 and λ = 0.5. The regularization was chosen to be a = 1 for
each method. We generated the signals to be biased towards the beginning, because the three ordered methods are
designed with this assumption built in and for BH the order does not matter anyways.
We find that across the range of sparsity settings, the bounds are fairly tight. For BH+, the bounds get looser
as the number of non-nulls increases, which reflects the fact that the BH method does not correct for the non-null
proportion. For relatively small non-null proportions (common in modern large-scale testing applications), this
14
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
Figure 2: (Batch) The theoretically derived FDP bounds are generally tight for all four batch procedures considered.
does not have a large effect. For the other three methods, the performance is not affected much by the sparsity
level, because the SeqStep accumulation function and its variants are relatively insensitive to non-nulls as long
as their p-values are small enough. While our bounds c(α) are smooth in all cases, we see that the Monte Carlo
curves can be jagged, especially for SeqStep and Selective SeqStep. In these cases, this jaggedness reflects the
d
discreteness of FDP.
Figure 3 shows the results of a similar simulation for the online FDR methods considered here: LORD+
and SAFFRON+. For both LORD+ and SAFFRON+, we choose αk = 0.05 for all k, and for SAFFRON+ we
additionally set λk = 0.5 for all k. Again, we choose a = 1. Unlike batch ordered testing, the signals for online
testing do not necessarily occur near the beginning, so in addition we considered the situation when signals are
randomly located. Like BH+, we find that the bounds for LORD+ get looser as the sparsity fraction increases.
While for randomly located signals the difference is not very large, for signals near the beginning we see that the
LORD+ bound becomes quite conservative. This is because the “local” non-null fraction near the beginning is
large even if the overall non-null fraction is not, resulting in much looser bounds. On the other hand, we find that
the SAFFRON+ bounds are fairly tight across sparsity levels and signal locations.
4
Implications for existing FDR procedures and new spotting methods
The optional spotting theorems allow us to view existing FDR procedures in a new light. In addition, they point the
way towards a new spotting paradigm for exploratory multiple testing. In this section, we consider both of these
implications of our results.
15
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
Figure 3: (Online) The theoretically derived FDP bounds are generally tight, except for LORD+ when signals are
near the beginning. SAFFRON+ adapts to the non-null proportion and thus does not suffer from this issue.
4.1
New FDX bounds for existing FDR procedures
d k∗ ) ≤ q. As discussed in the introduction, the
Batch FDR procedures choose the rejection set Rk∗ so that FDP(R
optional stopping theorems immediately imply that in this setting,
Pr{FDP(Rk∗ ) ≤ calg (α) · q} ≥ 1 − α.
(35)
Hence, existing FDR procedures also control FDX at level calg (α) · q. To our knowledge, these are the first finitesample FDX bounds for FDR procedures.
Of course, the practical implications of this result depend on the multipliers calg (α). Figure 1 depicts these
multipliers for regularization a = 1 at confidence α = 0.05. We see from the figure that the constants are
remarkably small for BH+, SAFFRON+, and LORD+, all near 2. Hence, under independent p-values, the FDP of
BH+ is at most about twice the nominal level q with probability 95%. This provides a theoretical foundation for the
common knowledge that the FDP distribution of BH is tightly concentrated around its mean under independence.
On the other hand, note that the multipliers are somewhat larger for accumulation test and Selective SeqStep, both
being slightly larger than 4. As we discuss below, this reflects the fact that these procedures have intrinsically more
variable FDP than BH.
We must bear in mind that the aforementioned multipliers are obtained for regularization a = 1, which is
not necessarily the same as the original regularization. For those procedures where regularization was originally
employed, the corresponding multipliers are depicted in Figure 1 using dashed lines. We see from the discrepancies
between the multipliers for original regularization and a = 1 that the regularization necessary for FDR control says
16
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
little about the consequences for FDX control. For BH, LORD, and SAFFRON, no regularization is necessary for
FDR control, so we do not have FDP or FDR multipliers for those procedures. For Adaptive SeqStep, only a small
amount of regularization is necessary, so the resulting multipliers are very large. On the other hand, accumulation
tests require the most regularization for FDR control, and thus the multipliers are small. For both FDR and FDX
control, regularization is especially necessary for the case when few rejections are made. Our results suggest that
FDR methods with little or no regularization should be used with extreme caution (or not at all) in low-signal
scenarios. On the other hand, regularization does not have a large effect on an FDR procedure if the signal is
strong enough that there are many discoveries (which is anyways the case in which FDR is most applicable as an
error criterion). Hence, in many cases our results are relevant even for procedures like BH defined originally with
no regularization.
4.2
Quantifying FDP variability using the constants c(α)
As we have observed, different FDR procedures have different FDP multipliers. For example, the multiplier for
Selective SeqStep (for p∗ = 0.5) is about twice as large as the multiplier for BH+. Moreover, as demonstrated in
Section 3.2, our bounds are fairly tight. Hence, examining how calg (α) behaves for different procedures can help
provide insight about the intrinsic variability of FDP for these procedures. More precisely, (35) shows that calg (α)
is related to the quantiles of FDP. However, especially when comparing FDP quantiles between two procedures
controlling FDR at the same level (thus with roughly similar FDP means), the FDP quantiles are directly tied to
FDP variability. To illustrate this, consider the case of accumulation tests.
Li and Barber [9] show that any accumulation function bounded by B leads to the same FDR guarantee.
However, there are many accumulation functions with the same value of B. For example, the left panel of Figure
4 shows three such functions (truncated and renormalized if necessary to have B = 2). The right panel of this
figure has the corresponding FDP multipliers, and we observe that steeper accumulation functions have larger FDP
multipliers. This reflects the fact that for a uniform random variable U , h(U ) is more variable if h is steeper,
d
since this random variable has more mass at its extremes. This translates to an increase in the variance of FDP.
d
Since accumulation tests stop when FDP(Rk ) ≤ q, this in turn increases FDP variance. In particular, the SeqStep
accumulation function is the “steepest" among all accumulation functions with a fixed maximum B, corresponding
to the fact that
hSeqStep = arg max Var[h(U )].
h:h≤B
Hence, the SeqStep accumulation function leads to the highest-variance accumulation test. On the other hand,
Li and Barber showed that the SeqStep accumulation function also leads to the highest-power accumulation test!
Indeed, they proved that the asymptotic power of an accumulation test increases as the quantity Epj ∼F1 [h(pj )]
decreases, where F1 is the distribution of non-null p-values (assuming a two-groups model). Since non-null pvalues tend to be smaller, accumulation tests will be powerful to the extent that their accumulation functions put
d This makes steeper accumulation functions more powerful.
little mass near the origin, resulting in a smaller FDP.
This demonstrates that for accumulation tests controlling FDR at a fixed level, there exists a trade-off between
FDP variability and power. This is not surprising if we recall that the FDP variability in this context is a proxy
for FDP quantiles. It is well known that there is a trade-off between mean FDP and power; a similar trade-off
exists involving FDP quantiles, which are simply another way to measure the location of the FDP distribution.
This conclusion adds more nuance to Li and Barber’s conclusion that SeqStep accumulation functions are the most
powerful for a fixed FDR level. If we switch from considering the mean of the FDP to its upper quantiles, then
other accumulation functions are likely to more favorably trade off FDP with power. We leave this direction of
inquiry for future work.
We have used accumulation tests as a case study for what insight we can glean by inspecting the bounds calg (α).
The same approach can be used to make other comparisons between either different settings of the same method or
between different methods. As another example of the former, note that several of the methods under consideration
are indexed by a parameter B. If we inspect each of the FDP multipliers c(α) for these methods, we see that these
multipliers are all increasing in B. Hence, B can be interpreted as a parameter controlling the FDP variability for a
given method. In parallel to what we saw above, increasing B also has the dual effect of increasing power. Hence,
choosing B involves a trade-off as well. For brevity’s sake, we do not elaborate on this in detail but encourage
17
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
Figure 4: Three accumulation functions and the corresponding FDP multipliers. Note that the steeper the accumulation function, the larger the FDP multiplier.
the reader to think about B in the context of accumulation tests and Adaptive SeqStep. We can also use the
multipliers c(α) to make comparisons across procedures. For example, we saw that Selective SeqStep has a larger
FDP multiplier than BH. This reflects the fact that Selective SeqStep relies on “one-bit p-values" I(pj > 0.5) to
d This results in a coarse estimate of FDP, which we see reflected in the increased multiplier.
define FDP.
4.3
“Simultaneous selective inference” via spotting
Our optional spotting theorems can provide more information than just about FDP(Rk∗ ) for k ∗ defined using
d k ) is a confidence envelope for the entire FDP
a fixed q. Indeed, we have proved that FDP(Rk ) = c(α)FDP(R
curve; recall (14). This puts our bounds into the realm of the uniform FDX literature, where many other confidence
envelopes for FDP have been proposed, including [26, 27, 29, 23, 30]. Some works [26, 29, 23] directly construct
bounds on the FDP of all subsets, whereas other [27] bound FDP only across a path of candidate rejection sets.
The latter kind of FDP bounds can be extended to any other sets, as pointed out implicitly by GW [29] and GS
[23], and explicitly by more recent work [30]. In particular, this implies that our uniform bounds on FDP(Rk ) can
also be extended to uniform bounds on all subsets.
Most of the existing literature on uniform bounds for FDP takes place in the unstructured setting, where hypotheses are considered in order of increasing p-values. Hence, the existing confidence envelopes are most comparable with our confidence envelope for BH+. We have found that confidence envelopes like those of [29] can
be competitive with or even tighter than those we derive for BH+. We do not show these comparisons here, since
it is not the goal of this work to derive the tightest possible confidence envelopes for FDP. As another point of
comparison, the closest result in the online setting is that of Javanmard and Montanari (JM) [14]. JM consider a
truncated version of GAI rules that satisfy a uniform FDX bound like Pr{supk FDPk ≥ γ} ≤ α. Their result is
similar in spirit to our result in Theorem 4 about LORD+, but there are some subtle differences. Their results, like
most other FDX bounds, are pre-hoc, meaning that given a γ, α ∈ (0, 1), their procedure produces a sequence of
rejections satisfying the desired FDX guarantee. Our guarantees are post hoc, meaning that they would apply to
any sequence of rejections produced by any online algorithm, that may or may not have been designed for FDR or
FDP control.
On the other hand, we are not currently aware of confidence envelopes that explicitly leverage side information
18
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
in the form of covariates, a priori hypothesis orderings, and so on. If this side information is informative, we expect
that better confidence envelopes can be developed than those derived in the unstructured setting. From this point
d k ) we derived for all procedures except BH represent the first
of view, the envelopes FDP(Rk ) = c(α)FDP(R
uniform FDX bounds in structured settings. We leave the uniform FDX performance of these envelopes for future
work.
As observed by Goeman and Solari before us, uniform bounds on FDP open up the possibility to reconcile data
exploration with multiple testing, via post hoc inference. Our bounds suggest that FDR procedures themselves
d Using prior information
can be reinterpreted as vehicles for simultaneous inference based on the path of FDP.
if applicable, the FDR procedures we analyze can be viewed as carving out a path of candidate rejection sets,
d is small. We hope that
from which the user can spot a set of appropriate size and content, while keeping FDP
restricting the user’s attention to a smaller collection of promising candidate rejection sets (simultaneous selective
inference) can keep the exploration more focused while enjoying better statistical and computational properties
than simultaneous inference.
In the previous section, we found trade-offs between power and FDP variability. For a procedure controlling
d (lower FDP
d permits more discoveries), and the
FDR at a fixed target level q, the power is represented by FDP
FDP variability is represented by c(α). The same trade-off is present in the confidence envelope FDP(Rk ) =
d k ), where these two quantities come together. Hence, good confidence envelopes must also navigate
c(α)FDP(R
d
the trade-off between c(α) and FDP.
To investigate this phenomenon further, we conducted a numerical experiment to examine the effect of accumulation function on the FDP confidence envelope. We consider the same simulation setup as in Section 3.2, with
50 non-nulls, except we decrease the signal amplitude to µ = 2.5. The reason for this is that the differences in
d are more pronounced when the signal is weaker, so there is a nontrivial trade-off between power and FDP
FDP
variability. To consider an extreme case, if the non-null p-values all equaled zero, then the accumulation function
h(pj ) = I(pj > 0) would simultaneously maximize power and minimize FDP variability.
d and FDP for three accumulation functions. In the left panel, the three paths of FDP
d
Figure 5: Comparing FDP
d
are plotted, along with the mean FDP. Flatter accumulation functions lead to larger FDP estimates. In the right
d k ) are shown for the three accumulation functions, along with the pointwise FDP
panel, FDP(Rk ) = c(α)FDP(R
d with HingeExp providing the best FDP curve.
quantiles. We see the nontrivial trade-off between c(α) and FDP,
In Figure 5, we compare the SeqStep, HingeExp, and ForwardStop accumulation functions. The first panel
d curves, averaged over 10,000 simulations. For comparison, the black curve is the average true
depicts the FDP
19
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
d then HingeExp in a close second,
FDP. As expected, the SeqStep accumulation function leads to the smallest FDP,
d
and then ForwardStop has the largest FDP. This trend reflects how steep each accumulation function is. The second
d k ). As a reference, the black curve represents the pointwise 1−α
panel shows the average FDP bounds c(α)FDP(R
d but the largest c(α),
quantiles for the true FDP. Here, we see the trade-off in action. SeqStep has the smallest FDP
and as a result leads to the worst FDP bound. ForwardStop and HingeExp both give better bounds than SeqStep,
with HingeExp being a little better over the most interesting range of k, around 25-50. Hence, HingeExp strikes
the best balance between power and FDP variability.
5
5.1
Implications beyond standard multiple testing
From multiple testing to model selection
All of the results presented here have been formulated in the multiple testing context, where we have p-values
available for a set of hypotheses. However, these results also apply to model selection and variable selection
contexts as well. In particular, G’Sell et. al. [4] proposed the ForwardStop accumulation function as a means to
the end of model selection with FDR control. Using selective inference techniques, valid p-values may be obtained
at each step of forward step-wise regression, and thus an accumulation test may be applied to select a model. As
another example, the knockoff filter [5] is a recent proposal for FDR control in the variable selection context.
Their idea is to create a knockoff variable for each original variable, and then use these knockoffs as controls for
the originals. Instead of p-values, the knockoff filter produces knockoff statistics Wj for each variable j. However,
the methodology is based on the fact that sign(Wj ) are independent fair coin flips for null j, in effect producing a
set of “one-bit p-values”. Then, Selective SeqStep with p∗ = 0.5 may be applied to these one-bit p-values to select
a set of variables with FDR control. Hence, our uniform FDX bounds apply to the knockoff filter as well.
It is interesting to note that the knockoff filter has been observed to produce somewhat highly variable results,
due to the coarse nature of one-bit p-values. We encountered this phenomenon in our theoretical results, since the
FDP multiplier c(α) for Selective SeqStep was one of the largest among the procedures considered.
5.2
Extensions to the multilayer setting using the p-filter
An intermediate statistical guarantee between FDR control and uniform FDX control is multilayer FDR control,
introduced by Barber and Ramdas [6]. This criterion applies when the rejected set R will be interpreted according
to multiple groupings (or layers) that are specified in advance. For example, these authors consider an application to
neuroscience, in which the hypotheses are indexed by brain location and by time. Then, specific voxels discovered
to be activated at specific times may be interpreted by grouping by time (finding voxels activated at any time
during the experiment) or by location (finding times at which any voxels were active). Multilayer FDR control is
the requirement that FDRm ≤ qm for each layer m, where qm are pre-specified FDR target levels.
Barber and Ramdas proposed the p-filter methodology, where p-value thresholds at each layer are chosen such
that an estimate of FDP at each layer m is below the target threshold qm , with extensions [7] proposed to handle
weights, adaptivty, overlapping or incomplete groups, etc.
Porting the p-filter into the variable selection setting, Katsevich and Sabatti introduce the multilayer knockoff
filter [8], which allows (group) knockoff statistics to be used at each layer instead of p-values. The proof of the
multilayer knockoff filter proceeds by decoupling the layers from each other via supremum bounds like inequality
(11), which was proved for Selective SeqStep and conjectured for BH+. However, the purpose of these bounds
was, like in most classical settings, as a tool to guarantee FDR control of the underlying procedure. In this paper,
we prove the conjecture for BH+, extend our proof technique to a multitude of other procedures and settings,
provide the post hoc spotting interpretation of these bounds in terms of uniform guarantees on the FDP, and more.
Now that our paper derives such supremum bounds for a much larger variety of procedures, we note that one
may use the same p-filter framework to arbitrarily “stack" any of these FDR procedures at different layers (meaning
that we could apply a different procedure at each layer) while maintaining multilayer FDR control.
20
6
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
Future work: the effect of correlation
Note that all of the results in this paper have assumed independence of p-values in one form or another. Many of
the procedures considered here also only have guarantees under independence, though BH is a notable exception.
Aside from online testing applications, independent p-values are hard to come by in practice, so more robust
guarantees are necessary. BH is known to control FDR under the PRDS criterion [2], a form of positive dependence
that contains no information about the strength of the dependence. Indeed, a set of p-values that are all equal to
each other almost surely satisfy the PRDS condition. However, it is known that while the mean of FDP might not
change much as dependence increases, the variance of the FDP will increase [36, 37]. Hence, high-probability
bounds on FDP under dependence are likely to use criteria other than PRDS to capture this dependence.
Figure 6: FDP multipliers under local correlation are displayed for several algorithms. For all subplots, negative
correlation does not seem to hurt, but strong positive correlations violate our theoretical bounds.
As a first step in this direction, Figure 6 shows the result of numerical simulations under AR(1) dependence for
the six procedures considered before, all with a = 1. These curves all correspond to the global null case, which
leads to the largest FDP multipliers. As we vary the correlation parameter ρ from −0.9 to 0.9, we can see how the
empirical quantiles change. We observe that higher correlations lead to larger FDP quantiles, as expected, but under
moderate correlations these quantiles are still below or at least near the theoretical bounds under independence.
Figure 7 shows how the expectation bounds change with ρ. Interestingly, for all methods besides Selective
SeqStep and accumulation test, the FDR multipliers are U-shaped, whereas for the two aforementioned methods,
the empirical quantiles are monotonically increasing in ρ. The bounds for BH+ and LORD+ are remarkably robust
to local correlation, whereas the other bounds grow somewhat more rapidly for large ρ. Investigating theoretically
how our bounds change under correlation (and how to precisely quantify correlation) is left for future work.
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
21
03.30.18
Figure 7: FDR bounds under local correlation are displayed for several algorithms. Most procedures seem less
affected by negative correlation than by strong positive correlations. Crosses indicate multipliers for ρ = 0, the
case with independent p-values as considered in the theorems of this paper.
7
Proofs of the four optional spotting theorems
The method we employ to obtain all our FDP multipliers is to investigate the properties of the stochastic process
FDP(Rk ). Below, we prove the optional spotting theorems; the proof of Corollary 1 will follow from the fact that
the expectation of a random variable is the integral of its quantile function. We start with the BH+ procedure, for
which we must proceed differently since the hypotheses are ordered according to the p-values, while for all other
d
settings, the hypothesis order is either pre-specified or is in some sense “orthogonal” to FDP.
7.1
FDP multipliers for BH+
Proof of Theorem 1. Let Vt ≡
(
P
FDP(Rk )
Pr sup
≥x
d k)
k FDP(R
j∈H0
)
I(pj ≤ t). Then, for any fixed x > 1 we have
(
Vt
= Pr sup
≥x
1
+
nt
t∈[0,1]
)
= Pr{Vt ≥ x + xnt, for some t ∈ [0, 1]} .
(36)
Hence, the quantity of interest is the probability that the stochastic process Vt hits the linear boundary x + xnt.
This event is illustrated in Figure 8.
22
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
Figure 8: Stochastic process view of (36), with the solid black line representing Vt , the dashed red line representing
the linear boundary x + xnt, and the dashed black line being the diagonal that emphasizes our knowledge that
V (0) = 0 and V (1) = n.
Since the null p-values were assumed to be independent and uniformly distributed, we deduce that the distribution of Vt is that of an unscaled empirical process for n0 = |H0 | uniform random variables. If we were to
replace Vt by the unscaled empirical process of n ≥ n0 uniform random variables, denoted Vt0 , it can only increase
the probability in equation (36). However, it is somewhat difficult to obtain non-asymptotic bounds for hitting
probabilities of empirical processes like Vt0 . Instead, we claim that for x ≥ 1.5, replacing Vt0 with a rate n Poisson
process Nt only further increases the hitting probability:
Pn
i.i.d.
Lemma 1. Let Vt0 ≡ j=1 I(pj ≤ t) for pj ∼ U [0, 1]. Let Nt be a rate n Poisson process. Then, for x ≥ 1.5,
Pr{Vt0 ≥ x + xnt, for some t ∈ [0, 1]} ≤ Pr{Nt ≥ x + xnt, for some t ∈ [0, 1]} .
(37)
d
See the appendix for the proof of this lemma. To understand why the lemma holds, note that Vt0 = Nt |N1 = n;
i.e. the distribution of the empirical process is the same as that of the Poisson process, conditioned on observing
exactly n events at time t = 1. Hence, Lemma 1 states that for sufficiently large x, a Poisson process is less likely
to hit the line t 7→ x + xnt if we know that it is equal to its mean at time t = 1. This makes intuitive sense because
the process needs to be far above its mean to hit this line for large x, which is less likely to happen if it must be
equal to its mean at time t = 1.
Now, we may use the martingale properties of Poisson processes to bound the probability a Poisson process
hits a linear boundary:
Lemma 2 ([8]). If Nt is a rate n Poisson process, then for any x > 1, we have
Pr{Nt ≥ x + xnt for some t ∈ [0, 1]} ≤ exp(−xθx ),
where θx is the unique positive root of the equation
eθ = 1 + θx.
(38)
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
23
03.30.18
To complete the derivation of the FDP multiplier for BH+, suppose we choose x such that α = exp(−xθx ).
Then, plugging this choice of x into the definition of θx implies that
x=
− log α
− log α
− log α
=
=
= cBH+ (α),
θx
log(1 + θx x)
log(1 − log α)
(39)
which proves the bound (28). Hence, statement (10) holds for all α corresponding to x ≥ 1.5, which translates to
holding for all α ≤ 0.31, since cBH+ (0.31) = 1.5.
7.2
An exponential tail inequality lemma
The optional spotting theorems for each of the ordered, interactive and online testing procedures will be derived in
the next subsection, as consequences of the following lemma.
Lemma 3. Consider an ordered testing procedure (batch or online), as described in Section 2.5. Suppose there
exists a filtration
Fk ⊇ σ(H0 , {π(j)}j≤k , {pπ(j) , αj }j≤k,π(j)∈H0 ).
(40)
such that for all π(k) ∈ H0 , we have
Pr pπ(k) ≤ αk Fk−1 ≤ αk and E hk (pπ(k) ) Fk−1 ≥ αk
almost surely.
(41)
Then, for each x > 1,
(
FDP(Rk )
Pr sup
≥x
d k)
k FDP(R
)
≤ exp(−aθx x),
(42)
where θx is defined in the following four cases:
1. If hk = h for some accumulation function h, αk = 1, π(k) is pre-specified (i.e. nonrandom), and pπ(k) ⊥
⊥
Fk−1 for all π(k) ∈ H0 , then θx is the unique positive root of the equation
Z 1
exp(−θxh(u))du = exp(−θ).
(43)
0
2. If hk = h for some accumulation function h bounded by B and αk = 1, then θx is the unique positive root
of the equation
1 − exp(−θxB)
= 1.
(44)
exp(−θ) +
B
3. If hk (p) = 0 for all p ≤ αk , and hk (p) ≤ B for all k, p, then θx is the unique positive root of the equation
exp(θ) −
1 − exp(−θxB)
= 1.
B
(45)
4. If hk (pk ) = αk , then θx is the unique positive root of equation (38).
Proof. Fix any arbitrary x > 1 and θ > 0. Since we make no assumptions whatsoever on the non-null p-values,
we first restrict our attention to only the nulls as follows:
(
)
(
)
FDP(Rk )
V (Rk )
Pr sup
≥ x = Pr sup
≥x
b (Rk )
d k)
k FDP(R
k V
k
k
X
X
= Pr
I(pπ(j) ≤ αj )I(π(j) ∈ H0 ) ≥ ax + x
hj (pπ(j) ), for some k
j=1
j=1
k
k
X
X
≤ Pr
I(pπ(j) ≤ αj )I(π(j) ∈ H0 ) ≥ ax + x
hj (pπ(j) )I(π(j) ∈ H0 ), for some k .
j=1
j=1
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
24
03.30.18
Now, we may rearrange terms and employ the Chernoff exponentiation trick, to conclude that:
(
)
k
X
FDP(Rk )
Pr sup
≥ x = Pr sup exp θ [I(pπ(j) ≤ αj ) − xhj (pπ(j) )]I(π(j) ∈ H0 ) ≥ exp(aθx)
d k)
k
k FDP(R
j=1
≡ Pr sup Zk ≥ exp(aθx) .
k
We claim that if θ = θx , then Zk is a supermartingale with respect to Fk . Note that this supermartingale and
filtration is unrelated to the supermartingale Mk and filtration Ωk mentioned in the introduction of this paper
that are typically used to prove FDR control for these procedures (indeed, Zk , Fk are forward supermartingales
and filtrations, while Mk , Ωk are backwards supermartingales and filtrations). If Zk is indeed a supermartingale as
claimed, then the conclusion of the lemma would follow from the maximal inequality for positive supermartingales:
Pr sup Zk ≥ exp(aθx) ≤ exp(−aθx)E [Z0 ] = exp(−aθx),
(46)
k
as desired. Hence the rest of this proof will focus on showing the supermartingale property of Zk . Note first of all
that Zk is adapted to Fk by assumption (40). Hence, it suffices to show that
Zk
Fk−1 = E exp [θ(I(pπ(k) ≤ αk ) − xhk (pπ(k) ))]I(π(k) ∈ H0 ) Fk−1 ≤ 1.
E
Zk−1
Clearly, this inequality holds for any k such that π(k) 6∈ H0 . For k such that π(k) ∈ H0 , we find that
E exp θ(I(pπ(k) ≤ αk ) − xhk (pπ(k) )) Fk−1
= E I(pπ(k) ≤ αk ) exp θ(1 − xhk (pπ(k) )) Fk−1
+ E I(pπ(k) > αk ) exp −θxhk (pπ(k) )) Fk−1 .
(47)
To show that the above quantity is at most one, we consider the four cases defined in the statement of the lemma.
Case 1:
Since αk = 1, the second term of equation (47) equals zero. Since hk = h, π(k) is fixed, and
pπ(k) ⊥
⊥ Fk−1 , the first term simplifies to
E exp θx (1 − xh(pπ(k) )) Fk−1 = E exp θx (1 − xh(pπ(k) ))
≤ exp(θx )EU ∼U [0,1] [exp(−θx xh(U ))]
= 1,
The inequality holds because pπ(k) is superuniformly distributed by assumption and u 7→ exp(−θx xh(u)) is
a nonincreasing function (since h is nondecreasing by definition), and the last step holds because θx satisfies
equation (43) by definition.
Case 2:
Again, the second term of equation (47) equals zero because αk = 1. We may bound the first term as:
1 − exp(−θx xB)
E exp θx (1 − xh(pπ(k) )) Fk−1 ≤ exp(θx )E 1 −
h(pπ(k) ) Fk−1
B
1 − exp(−θx xB)
≤ exp(θx ) 1 −
B
= 1.
In the first line, we used the fact that hk (pπ(k) ) = h(pπ(k) ) ≤ B, and we bounded the convex function z 7→
1−exp(−θx xB)
exp(−θ
z. In the second step, we used the assumption
x xz) on [0,B] with the line z 7→ 1 −
B
E h(pπ(k) ) Fk−1 ≥ αk = 1, and in the third line we used the definition of θx .
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
25
03.30.18
Because hk (p)I(p ≤ αk ) = 0, the first term of equation (47) simplifies to
E I(pπ(k) ≤ αk ) exp θx (1 − xhk (pπ(k) )) Fk−1 = exp(θx )Pr pπ(k) ≤ αk Fk−1 .
Case 3:
To bound the second term, we write
E I(pπ(k) > αk ) exp −θx xhk (pπ(k) )) Fk−1
1 − exp(−θx xB)
Fk−1
≤ E I(pπ(k) > αk ) 1 −
hk (pπ(k) )
B
1 − exp(−θx xB)
= Pr pπ(k) > αk Fk−1 −
E hk (pπ(k) )I(pπ(k) > αk ) Fk−1
B
1 − exp(−θx xB)
= Pr pπ(k) > αk Fk−1 −
E hk (pπ(k) ) Fk−1
B
1 − exp(−θx xB)
αk .
≤ Pr pπ(k) > αk Fk−1 −
B
The inequality in the second line follows from the same convexity argument as in Case 2, the fourth
line is a
consequence of the fact that hk (p)I(p ≤ αk ) = 0, and the last line follows from E hk (pπ(k) ) Fk−1 ≥ αk .
Combining the results of the previous two equations, we obtain that
E I(pπ(k) ≤ αk ) exp θx (1 − xhk (pπ(k) )) Fk−1 + E I(pπ(k) > αk ) exp −θx xhk (pπ(k) )) Fk−1
1 − exp(−θx xB)
≤ exp(θx )Pr pπ(k) ≤ αk Fk−1 + Pr pπ(k) > αk Fk−1 −
αk
B
1 − exp(−θx xB)
= (exp(θx ) − 1)Pr pπ(k) ≤ αk Fk−1 + 1 −
αk
B
1 − exp(−θx xB)
αk
≤ (exp(θx ) − 1)αk + 1 −
B
1 − exp(−θx xB)
= αk exp(θx ) − 1 −
+1
B
= 1.
The inequality in the fourth line follows from the first part of assumption (41), and the last equality follows from
the definition of θx .
Case 4:
Since hk (p) = αk , equation (47) simplifies to:
exp(θx (1 − xαk ))Pr pπ(k) ≤ αk Fk−1 + exp(−θx xαk )Pr pπ(k) > αk Fk−1
= exp(−θx xαk )((exp(θx ) − 1)Pr pπ(k) ≤ αk Fk−1 + 1)
≤ exp(−θx xαk )((exp(θx ) − 1)αk + 1).
Noting that θx satisfies (38), we see that the above expression can be bounded by one:
exp(−θx xαk ) (θx xαk + 1) ≤ 1,
as desired, where the inequality follows because the function z 7→ e−z (z + 1) is decreasing and takes the value 1
at z = 0.
7.3
Proofs of Theorems 2, 3, and 4
Now, we may derive Theorems 2, 3, and 4 as corollaries of Lemma 3.
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
26
03.30.18
Proof of Theorem 2. For all ordered procedures, the ordering π is pre-specified, so we may assume without loss of
generality that π(j) = j. First we note that the filtration
Fk = σ(H0 , {pj }j≤k,j∈H0 )
(48)
satisfies the required condition (40). Next, we consider accumulation tests and Adaptive SeqStep separately. Since
Selective SeqStep is a special case of Adaptive SeqStep, the corresponding bound will follow.
1. Accumulation test: For any k ∈ H0 , since αk = 1, the first part of the requirement (41) is trivially
satisfied. To show the second part, we use the independence assumption in Theorem 2 and the assumed
superuniformity of null p-values to derive
Z 1
E [hk (pk ) | Fk−1 ] = E [h(pk )] ≥ EU ∼U [0,1] [h(U )] =
h(u)du = 1 = αk .
0
The inequality follows because pk is superuniform and h is nondecreasing.
Hence, the assumptions of Lemma 3, case 1 are satisfied. Hence, the bound (42) holds with θx satisfying
equation (43). In order to prove the bound (29), we simply note that for any α ∈ (0, 1), we may choose x
such that exp(−aθx x) = α, from which it follows that
x=
log α
log α
− log α
=
=
= chacc (α),
R1
R1
aθx
a log 0 exp(−θx xh(u))du
a log 0 αh(u)/a du
as desired.
If the accumulation function is bounded by B, then the assumptions of Lemma 3, case 2 are also satisfied,
so the bound (42) holds with θx satisfying equation (44). In order to prove the bound (30), we note that for
any α ∈ (0, 1), we may choose x such that exp(−aθx x) = α, from which it follows that
x=
log α
log α
− log α
=
= cB
=
acc (α),
1−exp(−θx xB)
1−αB/a
aθx
a log 1 −
a
log
1
−
B
B
as desired. The proof of bound (31) is similar to that of (30), so we omit it here for brevity.
2. Adaptive SeqStep: We claim that the assumptions of Lemma 3 are satisfied with Fk defined as in equation
(48). Indeed, fix k ∈ H0 . Then, the assumed superuniformity of null p-values implies that
Pr{pk ≤ αk | Fk−1 } = Pr{pk ≤ p∗ } ≤ p∗ = αk
and
p∗
E [hk (pk ) | Fk−1 ] = E
I(p > λ) ≥ p∗ = αk .
1−λ
Hence, by case 3 of Lemma 3, it follows that (42) holds for θx satisfying (45). For fixed α ∈ (0, 1), define
x such that exp(−aθx x) = α. Then,
x=
− log α
− log α
− log α
=
= cB
=
AS (α),
1−exp(−θ
xB)
1−αB/a
x
aθx
a log 1 +
a log 1 + B
B
which proves the bound (32).
Proof of Theorem 3. This proof is similar to that of Theorem 2, exception the filtrations must be somewhat more
complicated to accommodate the interactivity of rAdaPT and rSTAR. For both of these procedures, define the
filtration
Fk = σ(H0 , {xj , g(pj )}j∈[n] , {pπ(j) }j≤k ).
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
27
03.30.18
Since π is predictable with respect to the filtration Gkr (27) by the definition of rSTAR and Fk ⊃ Grk , it follows
that π is also predictable with respect to Fk . Hence, requirement (40) holds for Fk as defined above. Now, let
π(k) ∈ H0 .
For rAdaPT, the first part of condition (41) holds because
Pr pπ(k) ≤ αk Fk−1 = E Pr pπ(k) ≤ p∗ g(pπ(k) ), π(k) π(k) ≤ PrU ∼U [0,1] {U ≤ p∗ | g(U )} = p∗ = αk .
The first equality follows from the independence assumption on the p-values, the first inequality follows from the
predictability of π(k) and from the assumption that the null p-values have increasing densities (see the proof of
[11, Proposition 1]). Similarly,
p∗
E hk (pπ(k) ) Fk−1 = E E
I(pπ(k) > p∗ ) g(pπ(k) ), π(k) π(k)
1 − p∗
p∗
≥
PrU ∼U [0,1] {U > p∗ | g(U )} = p∗ = αk .
1 − p∗
For rSTAR, the first part of condition (41) is trivially satisfied. To derive the second part, we write
E hk (pπ(k) ) Fk−1 = E E h(pπ(k) ) g(pπ(k) ), π(k) π(k) ≥ EU ∼U [0,1] [h(U )|g(U )] = 1 = αk .
The justification for this derivation is similar to that for rAdaPT, noting in addition that EU ∼U [0,1] [h(U )|g(U )] = 1
by construction.
Having established that the assumptions of Lemma 3 are satisfied for rAdaPT and rSTAR, the rest of the proof
is as in the proof of Theorem 2.
Proof of Theorem 4. Let Gk be the filtration defined in the statement of Theorem 4. Define
Fk ≡ σ(Gk , H0 ).
(49)
Then, assumption (40) holds for LORD+ and SAFFRON+ because pk are adapted to Fk and αk are predictable
with respect to Fk by assumption. The first part of (41) holds for both online procedures considered by assumption
(33). For LORD+, the second part holds because hk (pk ) = αk . For SAFFRON+, the second part holds because
for k ∈ H0 ,
αk
E [hk (pk ) | Fk−1 ] = E
I(pk > λk ) Fk−1 ≤ αk ,
1 − λk
where the last step follows from the predictability of αk and λk and the assumption (33).
Hence, LORD+ and SAFFRON+ satisfy the assumptions of Lemma 3. The remainder of the proof corresponds
exactly to analogous parts of the earlier proofs described for BH+ and Adaptive SeqStep, respectively.
8
Conclusion
We have shown in this paper that under independence, for a variety of procedures in the literature, the estimated
d (or a minor variant thereof) actually bounds the unknown FDP not just on average for the rejected set, but also
FDP
up to a small constant factor with high probability over the entire path of rejected sets proposed by each procedure.
This nontrivial and somewhat surprising result gives new guarantees for old procedures and sheds light on their
performance in terms of FDX. In particular, we can interpret c(α) as quantitative reflections of the variability in
FDP because our bounds are fairly tight. For example, our results support the empirical observations that BH has
a tight FDP distribution under independence, while echoing the fact that knockoff filter (i.e. Selective SeqStep)
can suffer from high FDP variance. In addition, our bounds have produced new insights into how the behavior
of accumulation tests’ FDP depends on the accumulation function. The FDP and FDR multipliers are only one
d By investigating this
half of the story, since decreasing these multipliers comes at a cost of increasing FDP.
trade-off, we have added nuance to Li and Barber’s [9] conclusion that the SeqStep accumulation function has the
highest asymptotic power by pointing out that it also has the highest-variance FDP, and hence that other choices of
accumulation function might better navigate this trade-off.
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
28
03.30.18
In addition to helping us understand existing FDR procedures, our results also open up a new perspective on
these procedures as tools for “simultaneous selective inference”. Instead of picking a single rejected set out of
2n , we can use the same machinery to computationally-efficiently carve out a path of n + 1 rejected sets based
d for each of them, along with a uniform
on structural constraints or prior knowledge, provide an estimated FDP
confidence band for the unknown FDP, and allow the user to choose one or more rejection sets in hindsight,
providing simultaneous guarantees over all such rejection sets.
Exploring such bounds and trade-offs is an important direction, since we believe that control of the FDP is as
important or more so than control of the FDR. Procedures with FDX control provide more interpretable guarantees
in practice, and those with uniform FDX control may pave the way to a new exploratory paradigm of multiple
testing à la Goeman and Solari.
Finally, the proof technique we developed in this paper is versatile enough to cover a large portion of the
currently available FDR procedures. Like Genovese and Wasserman [28], we employ a stochastic process approach
to analyze the FDP. However, while GW’s bounds are asymptotic, we have used martingale arguments instead to
obtain tight, non-asymptotic bounds. Perhaps these proof techniques may be extended further to apply to other
multiple testing scenarios as well.
9
Acknowledgements
E.K. would like to thank Chiara Sabatti for her generous and valuable feedback on this work and on the manuscript
itself. E.K. also thanks David Siegmund, Emmanuel Candes, Anya Katsevich, and Michael Sklar for helpful
discussions. E.K.’s work was supported by the Fannie and John Hertz Foundation. A.R. acknowledges several
fruitful conversations and email exchanges with Ohad Feldheim, Jim Pitman and Jon Wellner. Both authors are
also grateful to the organizers of the MCP 2017 conference, where this collaboration began.
References
[1] Y. Benjamini and Y. Hochberg. Controlling the false discovery rate: a practical and powerful approach to
multiple testing. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 57(1):289–300,
1995.
[2] Y. Benjamini and D. Yekutieli. The control of the false discovery rate in multiple testing under dependency.
The Annals of Statistics, 29(4):1165–1188, 2001.
[3] E. Arias-Castro and S. Chen. Distribution-free multiple testing. Electronic Journal of Statistics, 11(1):1983–
2001, 2017.
[4] M. G. G’Sell, S. Wager, A. Chouldechova, and R. Tibshirani. Sequential selection procedures and false discovery rate control. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 78(2):423–
444, 2016.
[5] R. F. Barber, E. J. Candès, et al. Controlling the false discovery rate via knockoffs. The Annals of Statistics,
43(5):2055–2085, 2015.
[6] R. F. Barber and A. Ramdas. The p-filter: multilayer false discovery rate control for grouped hypotheses.
Journal of the Royal Statistical Society: Series B (Statistical Methodology), 2016.
[7] A. Ramdas, R. F. Barber, M. J. Wainwright, and M. I. Jordan. A unified treatment of multiple testing with
prior knowledge. arXiv preprint arXiv:1703.06222, 2017.
[8] E. Katsevich and C. Sabatti. Multilayer knockoff filter: Controlled variable selection at multiple resolutions.
arXiv preprint arXiv:1706.09375, 2017.
[9] A. Li and R. F. Barber. Accumulation tests for FDR control in ordered hypothesis testing. Journal of the
American Statistical Association, pages 1–13, 2017.
29
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
03.30.18
[10] L. Lei and W. Fithian. AdaPT: An interactive procedure for multiple testing with side information. arXiv
preprint arXiv:1609.06035, 2016.
[11] L. Lei, A. Ramdas, and W. Fithian. STAR: A general interactive framework for FDR control under structural
constraints. arXiv preprint arXiv:1710.02776, 2017.
[12] D. P. Foster and R. A. Stine. α-investing: a procedure for sequential control of expected false discoveries.
Journal of the Royal Statistical Society: Series B (Statistical Methodology), 70(2):429–444, 2008.
[13] E. Aharoni and S. Rosset. Generalized α-investing: definitions, optimality results and application to public
databases. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 76(4):771–794, 2014.
[14] A. Javanmard and A. Montanari. Online rules for control of false discovery rate and false discovery exceedance. The Annals of Statistics, 2017.
[15] A. Ramdas, F. Yang, M. J. Wainwright, and M. I. Jordan. Online control of the false discovery rate with
decaying memory. In Advances In Neural Information Processing Systems, 2017.
[16] A. Ramdas, T. Zrnic, M. J. Wainwright, and M. I. Jordan. SAFFRON: an adaptive algorithm for online
control of the false discovery rate. arXiv preprint arXiv:1802.09098, 2018.
[17] E. L. Korn, J. F. Troendle, L. M. McShane, and R. Simon. Controlling the number of false discoveries:
application to high-dimensional genomic data. Journal of Statistical Planning and Inference, 124(2):379–
398, 2004.
[18] E. L. Lehmann and J. Romano. Generalizations of the familywise error rate. The Annals of Statistics, pages
1138–1154, 2005.
[19] J. Romano and A. Shaikh. On stepdown control of the false discovery proportion. In Optimality, pages
33–50. Institute of Mathematical Statistics, 2006.
[20] J. P. Romano and M. Wolf. Control of generalized error rates in multiple testing. The Annals of Statistics,
pages 1378–1408, 2007.
[21] W. Guo, L. He, and S. K. Sarkar. Further results on controlling the false discovery proportion. The Annals of
Statistics, 42(3):1070–1101, 2014.
[22] S. Delattre and E. Roquain. New procedures controlling the false discovery proportion via Romano-Wolf’s
heuristic. The Annals of Statistics, 43(3):1141–1177, 2015.
[23] J. Goeman and A. Solari. Multiple testing for exploratory research. Statistical Science, pages 584–597, 2011.
[24] J. F. Troendle. Stepwise normal theory multiple test procedures controlling the false discovery rate. Journal
of Statistical Planning and Inference, 84(1-2):139–158, 2000.
[25] H. Finner and M. Roters. On the false discovery rate and expected type I errors. Biometrical Journal,
43(8):985–1005, 2001.
[26] M. J. van der Laan, S. Dudoit, and K. S. Pollard. Multiple testing. Part III. Procedures for control of the
generalized family-wise error rate and proportion of false positives. 2004.
[27] N. Meinshausen. False discovery control for multiple tests of association under general dependence. Scandinavian Journal of Statistics, 33(2):227–237, 2006.
[28] C. Genovese and L. Wasserman. A stochastic process approach to false discovery control. The Annals of
Statistics, pages 1035–1061, 2004.
[29] C. R. Genovese and L. Wasserman. Exceedance control of the false discovery proportion. Journal of the
American Statistical Association, 101(476):1408–1417, 2006.
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
30
03.30.18
[30] G. Blanchard, P. Neuvial, and E. Roquain. Post hoc inference via joint family-wise error rate control. arXiv
preprint arXiv:1703.02307, 2017.
[31] L. Lei and W. Fithian. Power of ordered hypothesis testing. In International Conference on Machine Learning, pages 2924–2932, 2016.
[32] W. Fithian, D. Sun, and J. Taylor. Optimal inference after model selection. arXiv preprint arXiv:1410.2597,
2014.
[33] J. Storey, J. Taylor, and D. Siegmund. Strong control, conservative point estimation and simultaneous conservative consistency of false discovery rates: a unified approach. Journal of the Royal Statistical Society:
Series B (Statistical Methodology), 66(1):187–205, 2004.
[34] H. Robbins. A one-sided confidence interval for an unknown distribution function. In Annals of Mathematical
Statistics, volume 25, pages 409–409, 1954.
[35] J. Goeman, R. Meijer, T. Krebs, and A. Solari. Simultaneous control of all false discovery proportions in
large-scale multiple hypothesis testing. arXiv preprint arXiv:1611.06739, 2016.
[36] A. B. Owen. Variance of the number of false discoveries. Journal of the Royal Statistical Society: Series B
(Statistical Methodology), 67(3):411–426, 2005.
[37] B. Efron. Large-scale inference, volume 1 of Institute of Mathematical Statistics (IMS) Monographs. Cambridge University Press, Cambridge, 2010. Empirical Bayes methods for estimation, testing, and prediction.
[38] G. D. Anderson, R. W. Barnard, K. C. Richards, M. K. Vamanamurthy, and M. Vuorinen. Inequalities for
zero-balanced hypergeometric functions. Transactions of the American Mathematical Society, pages 1713–
1723, 1995.
A
Technical lemmas: From empirical to Poisson processes
Lemma 4. Let {Nt }t≥0 be a Poisson process with rate n. For x ≥ 1.5,
#
"
#
"
Nt
Nt
≥ x N1 = n ≤ P sup
≥x .
P sup
t∈[0,1] 1 + nt
t∈[0,1] 1 + nt
(50)
Proof. Let us define
τ = inf t :
Nt
≥x .
1 + nt
We claim that it suffices to show that for x ≥ 1.5,
P[N1 = n|τ ] ≤ P[N1 = n] for all τ ≥ 0.
Indeed, it would then follow that
"
#
Nt
P sup
≥ x N1 = n = P [τ ≤ 1|N1 = n]
t∈[0,1] 1 + nt
P[τ ≤ 1]P[N1 = n|τ ≤ 1]
P[N1 = n]
≤ P[τ ≤ 1]
"
#
Nt
= P sup
≥x .
t∈[0,1] 1 + nt
=
(51)
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
31
Note that for a given x, x(1 + nt) > n for t >
only consider
1
x
− n1 . Hence, statement (51) is trivial for τ >
τ≤
03.30.18
1
x
− n1 , so we need
1
1
− .
x n
Define
f (λ, y) = e−λ
λy
.
Γ(y + 1)
(52)
(53)
This function f : [0, ∞) → R is equal to the probability mass function of the Poisson with parameter λ when
y ∈ {0, 1, 2, . . . }. We have
P[N1 = n|τ ] = P[N1 = n|Nτ = dx(1 + nτ )e]
= P[N1−τ = n − dx(1 + nτ )e]
= f (n(1 − τ ), n − dx(1 + nτ )e)
(54)
≤ f (n(1 − τ ), n − x(1 + nτ ))
≡ g(τ, x).
The inequality follows by Lemma 6 because n − x(1 + nτ ) ≤ n(1 − τ ) − x ≤ n(1 − τ ) − 1.5. Define
(n(1 − τ ))n−x(1+nτ )
h(τ, x) = log g(τ, x) = log exp(−n(1 − τ ))
Γ(1 + n − x(1 + nτ ))
(55)
= −n(1 − τ ) + (n − x(1 + nτ )) log(n(1 − τ )) − log Γ(1 + n − x(1 + nτ )).
Note that
∂
n − x(1 + nτ )
h(τ, x) = n − xn log(n(1 − τ )) −
+ xnψ(1 + n − x(1 + nτ ))
∂τ
1−τ
1
n − x(1 + nτ )
+ xn log(1.5 + n − x(1 + nτ )) −
≤ n − xn log(n(1 − τ )) −
1−τ
1.5 + n − x(1 + nτ )
≡ r(τ, x).
(56)
The inequality follows by Lemma 5.
To prove the inequality (51), it suffices to show that
(a) g(0, x) ≤ P[N1 = n] for all x ≥ 1.5;
(b) r(τ, 1.5) ≤ 0 for each τ ≤
(c)
∂
∂x r(τ, 1.5)
1
1.5
≤ 0 for each τ ≤
− n1 ;
1
1.5
− n1 ;
(d) r(τ, x) is concave in x for each τ .
1
Indeed, note that (c) and (d) imply that for each τ ≤ 1.5
− n1 , r(τ, x) is a decreasing function of x for x ≥ 1.5.
Hence, for each x ≥ 1.5 and each τ , we have r(τ, x) ≤ r(τ, 1.5) ≤ 0, where the last inequality follows from (b).
Hence, by inequality (56), ∂h
∂τ (τ, x) ≤ r(τ, x) ≤ 0 for each x ≥ 1.5. This means that h(τ, x) is decreasing in τ
for each x ≥ 1.5, so g(τ, x) is decreasing in τ for each x ≥ 1.5, from which it follows that g(τ, x) ≤ g(0, x) ≤
P[N1 = n], where the last inequality follows by (a).
Proof of (a).
We have
g(0, x) = f (n, n − x) ≤ f (n, n − 1) = f (n, n) = P[N1 = n].
where the inequality follows by Lemma 6 and the equality f (n − 1, n) = f (n, n) holds because
f (n, n − 1)
=
f (n, n)
nn−1
Γ(n)
nn
Γ(n+1)
=
Γ(n + 1)
= 1.
nΓ(n)
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
32
03.30.18
Proof of (b). We have
n − 1.5(1 + nτ )
r(τ, 1.5) = n − 1.5n log(n(1 − τ )) −
1−τ
1
+ 1.5n log(1.5 + n − 1.5(1 + nτ )) −
1.5 + n − 1.5(1 + nτ )
n − nτ − 1.5 − 0.5nτ
1
= n − 1.5n log(n(1 − τ )) −
+ 1.5n log(n(1 − 1.5τ )) −
1−τ
n(1 − 1.5τ )
1 − 1.5τ
1.5 + 0.5nτ
1.5
= 1.5n log
+
−
1−τ
1−τ
1 − 1.5τ
1 − 1.5τ
0.5τ
1
1
= n 1.5 log
+
+ 1.5
−
1−τ
1−τ
1−τ
1 − 1.5τ
1 − 1.5τ
0.5τ
≤ n 1.5 log
+
1−τ
1−τ
0.5τ
0.5τ
= n 1.5 log 1 −
+
1−τ
1−τ
0.5τ
0.5τ
≤ n −1.5
+
1−τ
1−τ
0.25τ
= −n
1−τ
≤ 0.
Proof of (c).
We have
∂r
1 + nτ
(τ, x) = −n log(n(1 − τ )) +
+ n log(1.5 + n − x(1 + nτ ))
∂x
1−τ
1.5 + n − x(1 + nτ ) + x(1 + nτ )
(1 + nτ )xn
−n
−
1.5 + n − x(1 + nτ )
(1.5 + n − x(1 + nτ ))2
1 + nτ
= −n log(n(1 − τ )) +
+ n log(1.5 + n − x(1 + nτ ))
1−τ
(1 + nτ )xn
1.5 + n
−
−n
.
1.5 + n − x(1 + nτ )
(1.5 + n − x(1 + nτ ))2
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
33
03.30.18
Plugging in x = 1.5, we have
∂r
1 + nτ
1.5(1 + nτ )n
1.5 + n
(τ, 1.5) = −n log(n(1 − τ )) +
+ n log(n(1 − 1.5τ )) −
−n
∂x
1−τ
n(1 − 1.5τ )
(n(1 − 1.5τ ))2
1 − 1.5τ
1.5
1
τ
1.5
1
1.5
= n log
−
+
+n
−
(1 + nτ ) −
2
1−τ
1−τ
1−τ
1 − 1.5τ
n (1 − 1.5τ )
(1 − 1.5τ )2
1 − 1.5τ
τ
1.5τ
1
1.5
≤ n log
+
−
+
−
1−τ
1−τ
1 − 1.5τ
1−τ
1 − 1.5τ
1 − 1.5τ
τ
1.5τ
≤ n log
+
−
1−τ
1−τ
1 − 1.5τ
−0.5τ
τ
1.5τ
≤n
+
−
.
1−τ
1−τ
1 − 1.5τ
1.5τ
0.5τ
−
.
=n
1−τ
1 − 1.5τ
τ (0.75τ − 1)
≤n
(1 − τ )(1 − 1.5τ )
≤ 0,
where the last inequality follows because τ ≤
Proof of (d).
1
1.5
−
1
n
≤ 4/3.
Modulo terms linear in x and the scaling factor n, r(τ, x) is equal to
x
x log (1.5 + n − x(1 + nτ )) −
.
1.5 + n − x(1 + nτ )
We claim that the first term is concave in x and the second term is convex, from which it will follow that their
difference is concave. By linear transformations, the concavity of the first term will follow from the concavity
x
of x log(1 − x), which follows because its first derivative log(1 − x) − 1−x
is decreasing in x. Again by linear
x
1
transformations, the convexity of the second term will follow from the convexity of 1−x
= −1 + 1−x
on x < 1,
which is clear.
Lemma 5. Let ψ(x) = Γ0 (x)/Γ(x) be the digamma function. Then, ψ is increasing, and for x ≥ 1,
ψ(x) ≤ log(x) −
1
1
≤ log(x + 0.5) −
≤ log x.
2x
x + 0.5
(57)
Proof. The fact that ψ is increasing is well-known. The first inequality follows directly from [38, Theorem 3.1].
To prove the second inequality, write
2
0.5
0.5 1 0.5
log(x + 0.5) − log(x) = log 1 +
≥
−
,
x
x
2 x
and the conclusion follows because
2
0.5 1 0.5
1
1
1
1
x − 0.5
−
≥
−
⇐⇒
−
≥
⇐⇒ x ≥ 1/6.
x
2 x
x + 0.5 2x
2x 8x2
2x(x + 0.5)
To prove the third inequality, write
0.5
0.5
log(x + 0.5) − log(x) = log 1 +
≤
,
x
x
and the conclusion follows because
0.5
1
≤
⇐⇒ x ≥ 0.5.
x
x + 0.5
Towards “simultaneous selective inference”:
post hoc bounds on the false discovery proportion
34
Lemma 6. Let f be as defined in equation (53). Then,
∂
f (λ, y) ≥ 0
∂y
for all y ≤ λ − 1.
Proof. To prove this, it suffices to show that for y ≤ λ − 1,
0≤
∂
∂
log f (λ, y) =
(−λ + y log λ − log Γ(y + 1)) = log λ − ψ(y + 1).
∂y
∂y
Indeed, by Lemma 5, for y ≤ λ − 1 we have ψ(y + 1) ≤ ψ(λ) ≤ log λ.
03.30.18
| 10math.ST
|
Achievability Bounds for T-Fold Irregular
Repetition Slotted ALOHA Scheme in the Gaussian
Multiple Access Channel
Nikolay Matveev‡, Kirill Andreev∗, Alexey Frolov∗† and Andrey Turlikov
‡
∗ Skolkovo Institute of Science and Technology
Moscow, Russia
† Institute for Information Transmission Problems
arXiv:1802.07332v1 [cs.IT] 20 Feb 2018
Russian Academy of Sciences
Moscow, Russia
‡ State University of Aerospace Instrumentation, St. Petersburg, Russia
[email protected], [email protected], [email protected], [email protected]
Abstract—We address the problem of massive random access
for an uncoordinated Gaussian multiple access channel (MAC).
The performance of T-fold irregular repetition slotted ALOHA
(IRSA) scheme for this channel is investigated. The main difference of this scheme in comparison to IRSA is as follows: any
collisions of order up to T can be resolved with some probability
of error introduced by Gaussian noise. First, we generalize the
density evolution method for T -fold IRSA and noisy channel
and find optimal degree distributions for different values of T .
Then we perform analysis and find minimal Eb /N0 for a fixed
length, rate and packet loss probability. The scheme is shown to
work closer to finite length random coding bound proposed by
Y. Polyanskiy, than existing solutions.
I. I NTRODUCTION
Existing wireless networks are designed with the goal of
increasing a spectral efficiency in order to serve human users.
Next generation of wireless networks will face a new challenge
in the form of machine-type communication. Analysts predict
that the number of devices connected to the network will
exceed 50 millions by 2020. The main challenges are as
follows: (a) huge number (billions) of autonomous devices
connected to one access point, (b) low energy consumption,
(c) short data packets. This problem has attracted attention of
3GPP standardization committee under the name of mMTC
(massive machine-type communication).
Let us describe the system model. There are Ktot ≫ 1
users, of which only K are active in each time instant.
Communication proceeds in a frame-synchronized fashion
(this can be implemented with use of beacons). The length
of each frame is n. Each active user has k bits to transmit
during a frame. The main goal is to minimize the energy-perbit spent by each of the users. We are interested in grant-free
access (5G terminology), i.e. active users transmit their data,
without any resource requests.
This paper deals with construction of low-complexity random coding schemes for the Gaussian multiple-access channel
(GMAC) with equal-power users, i.e.
y=
Ktot
X
si xi + z,
(1)
i=1
where xi ∈ Rn is a codeword transmitted by the i-th user, si
is an activity indicator for the i-th user, i.e. si = 1 if the i-th
user is active and si = 0 otherwise. z ∼ N (0, I) is an additive
white Gaussian noise (AWGN). Following [1] we assume all
the users to use the same message set [M ] , {1, . . . , M } and
the same codebook C = {x(ω)}M
ω=1 of size M . Let ωi denote
the message of the i-th user. To transmit the message ωi the
user will use a codeword xi = x(ωi ). We require in addition
that ||x(ω)||22 ≤ nP , which means a natural power constraint.
Decoding is done up to permutation of messages. We only
require the decoder to output a set L(y) = (ω1 , ω2 , . . . , ωJ ) ∈
[M ]J , J ≤ K. Thus in accordance to [1] we decouple the user
identification problem and the data transmission problem. The
probability of error (per user) is defined as follows
Pe =
max
|(s1 ,s2 ,...,sKtot )|=K
Ktot
1 X
si Pr(Wi 6∈ L(y)).
K i=1
Let us emphasize the main differences from the classical
setting. Almost all well-known low-complexity coding solutions for the traditional MAC channel (e.g. [2]) assume
coordination between the users. Due to the gigantic number
users we assume them to be symmetric, i.e. the users use the
same codes and equal powers.
We continue the line of work started in [1], [3], [4]. In
[1] the bounds on the performance of finite-length codes
for GMAC are presented. In [3] Ordentlich and Polyanskiy
describe the first low-complexity coding paradigm for GMAC.
The improvement (it terms of required Eb /N0 ) was given
in [4]. The overall scheme can be called T-fold irregular
repetition slotted ALOHA (IRSA, [5], [6]) scheme for GMAC.
The main difference of this scheme in comparison to IRSA is
as follows: any collisions of order up to T can be resolved
with some probability of error introduced by Gaussian noise.
users
+
+
+
+
In this paper we investigate the potential capabilities of T fold IRSA. The authors of [4] suggested to split the error
probability into three terms: interference cancellation (IC)
error, preamble error and decoder error. Then a union bound
was applied. We were able to write joint density evolution
rules, which include all the terms. In what follows we show,
that this change allows us to better predict the error probability.
We also note the following thing. In [4] in each slot preambles
were used to identify active users. So the packet consists of
preamble and data part. The authors used finite length random
coding bound [1] to estimate the probability of decoding
failure in the data part. We emphasize, that preambles are
not needed to characterize the potential capabilities (derive
achievability bounds) of T -fold IRSA scheme as we do not
need to identify users, that are active in a slot to use IC
algorithm.
Our contribution is as follows. Achievability bounds for T fold IRSA are derived. To achieve this goal we generalized the
density evolution method for T -fold IRSA and noisy channel
and find optimal degree distributions for different values of T .
Then we perform analysis and find minimal Eb /N0 for a fixed
length, rate and packet loss probability. The scheme is shown
to work closer to finite length random coding bound proposed
by Y. Polyanskiy, than existing solutions.
slots
Fig. 1: Tanner graph representation
The algorithm stops when all the slots are either decoded or
empty (see Algorithm 1 for more details).
Algorithm 1 Interference cancellation decoder
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
while slot set not empty do
Remove empty slots
Select slot with the minimal collision order t
if t ≤ T then
Resolve the collision
Subtract decoded packets from all the slots
Update collision indices for all slots
end if
Remove decoded slot from slot set
end while
II. P RELIMINARIES
A. System model
The main features of the scheme are as follows:
• transmission is performed in a frame-synchronized fashion. The length of a frame is n channel uses;
• the frame consists of M slots of size ñ = n/M channel
uses;
• the user chooses a message ω, then encodes it and obtains
a codeword x(ω) of length ñ;
• users repeat their codewords in multiple slots. The distribution of the repetition count is denoted by D[i]. This
distribution is the same for all the users;
• the number of repetitions r and the r slots in which to
send are chosen based on the message ω (see [4]): as
ω is distributed uniformly on [M ] the slots are chosen
uniformly at random (definitely without repetitions) from
M existing slots.
B. Interference cancellation decoder
Decoding algorithm is based on successive cancellation approach. The algorithm is iterative. The slot with the minimum
collision order is selected at every step. If the order of collision
is less or equal to T we resolve the collision. As a result some
messages (in collision) are decoded successfully, some of the
messages are decoded incorrectly. Then all successfully decoded messages are removed from other slots (if the message
was transmitted by user more than once), the slot itself is
marked as resolved and collision indices are updated for all
other slots. We note, that we can always find where the replicas
were transmitted as these positions are chosen based on the
data (so in contrast to [5] we do not need to store the pointers).
Remark 1: We note, that during the slot decoding the errors
may occur, i.e. some of the packets (codewords) may be
decoded incorrectly. In what follows we assume, that we can
always detect the error packets (the packets include control
information).
C. Tanner graph representation
The transmission and decoding processes can be described
with the use of a bipartite graph, which is called the Tanner
graph [7]. The vertex set of the graph consists of the set of
user nodes V = {v1 , v2 , . . . , vK } which correspond to the set
of users and the set of slot nodes C = {c1 , c2 , . . . , cM } which
correspond to signals received in slots. The user node vi and
the slot node cj are connected with an edge if and only if the
i-th users transmitted a packet in the j-th slot.
D. Degree distributions
P
P
i−1
i
denote the
Let L(x) =
i λi x
i Li x and λ(x) =
user node degree distributions from node and edge degree
perspective, respectively. We recall (see e.g. [8]), that Li and
λi denote respectively the fractions of user nodes of degree i
and the fraction of edges incident to user nodes of degree i.
′
Also recall, that λ(x) = L′ (x)/L
(1). In our case L
Pi = D[i].
P
i−1
i
Analogously, let R(x) =
i ρi x
i Ri x and ρ(x) =
denote the slot node degree distributions from node and edge
degree perspective, respectively.
Let G = K/M . Let us consider the j-th slot. Each
user chooses ′ this slot ′for transmission independently with
probability LM(1) = GLK(1) . Thus,the slot node
distribution
GL′ (1)
. In the limit
(from node perspective) is Bino K, K
K → ∞ this distribution becomes a Poisson distribution. In
′
what follows we use R(x) = ρ(x) = e−GL (1)(1−x) .
xℓ
xl
yl+1
User
node
III. D ENSITY EVOLUTION
we note, that L′ (1) is actually the average number of transmissions.
Let us introduce a notation. By Pe (ñ, k, P, t) we denote
the error probability per user, calculated with use of random
coding bound [1]. Now let us write the density evolution
rules. By xl and yl we denote the probability that an outgoing
message from the user node and slot node, respectively, are
erased during the l-th iteration. We start with initial condition
x0 = 1, which means, that the user messages are erased at the
beginning and we observe only the noisy signal sums in slots.
…
xl
xl
r-1
…
yl
i
…
i-1
yℓ
xl
yl
yℓ
yℓ
yl
Fig. 2: Density evolution rules
16
Random coding bound
Reference T = 2
Reference T = 4
10 decoding iterations
100 decoding iterations
14
12
T=1
T=2
10
Eb/N0, dB
Following [1] we want to minimize the required energy-perbit Eb /N0 . For this purpose we suggest a density evolution
method, which helps us to choose the system parameters.
Similar to [5], [6] we consider the ensemble of Tanner
graphs G(K; M ; cλ(x); ρ(x)) corresponding to the multipleaccess scheme with K users, M slots, and the degree distributions λ(x) and ρ(x). We are interested in the decoding
performance averaged over the ensemble G(K; M ; λ(x); ρ(x))
in the limit as K,M → ∞.
The major difference of our approach in comparison to [5],
[6] is that we
• consider co-called T -fold IRSA, i.e. collisions of order up
to T can be resolved in slot node with some probability
of error introduced by Gaussian noise;
• take into account a noisy channel (AWGN channel to be
precise);
• take into account finite length effect as the slots have
small length;
• take into account a transmit energy. Assume we use a
strategy with L(x) = x2 . In this case we spent 2 times
more energy while transmitting in comparison to L(x) =
x strategy;
Let us fix the slot length ñ, the number of information bits
k to be sent by each user, the maximal number of iterations
ℓ, the average transmit power P (linear scale) and L(x). Then
the average energy per information bit can be calculated as
follows
ñP L′ (1)
Eb
=
,
N0
2k
User
node
Slot
node
T=4
8
6
4
2
0
0
100
200
300
User count
400
500
600
Fig. 3: Minimum Eb /N0 required to achieve less than 5%
error rate as a user count function
because of finite length effects in the slot. So in what follows
we do not consider infinite number of iterations and fix ℓ.
Proof:
Consider the l-th iteration. Let us consider the slot node
of degree r. We want to calculate the erasure probability of
the outgoing message yl+1 based on incoming messages (with
erasure probabilities xl ). The probability can be calculated as
follows
min(r,T )−1
X
r−1 t
1−
(1 − Pe (ñ, k, P, t))
xl (1 − xl )r−1−t .
t
t=1
Each term in the sum corresponds to a probability of
collision of order t. We are interested only in collisions of
order t ≤ T . The probability of correct outgoing message
decoding is given by 1 − Pe (ñ, k, P, t). We only need to apply
the sum rule of total probability to obtain the needed result.
Recall, that ρr is the probability, that the outgoing edge is
connected to a slot node of degree r.
rX
max
h min(r,T
The rule in user node coincides with the rules from [5], [6].
X)−1
r−1
yl+1 = 1 −
ρr
(1 − Pe (ñ, k, P, t))
t
r=1
t=1
i
IV. N UMERICAL RESULTS
× xtl (1 − xl )r−1−t
We choose the same system parameters as in [1], [3], [4]
xl = λ(yl ), 1 ≤ l < ℓ.
for honest comparison.
xℓ = L(yℓ ).
Parameter
Description
4
n
=
3
×
10
Frame length
Remark 2: We note, that
k = 100
Number of information bits
lim xℓ > 0.
P
=
0.05
Maximum error probability allowed
e
ℓ→∞
K
25
50
100
150
200
250
300
350
400
450
500
550
DE for T = 2, 100 iterations
1.0000x
0.4447x + 0.5553x2
0.1261x + 0.8739x2
0.1155x + 0.8845x2
0.1226x + 0.8774x2
0.1252x + 0.8748x2
0.1591x + 0.6886x2 + 0.1524x3
0.1776x + 0.5895x2 + 0.2329x3
0.1860x + 0.5429x2 + 0.2711x3
0.2005x + 0.4861x2 + 0.3135x3
0.2048x + 0.4620x2 + 0.3333x3
0.2140x + 0.4269x2 + 0.3590x3
0.2218x + 0.3982x2 + 0.3776x3
+0.0015x4 + 0.0005x5
+0.000x6 + 0.0001x7
M
79
61
66
97
128
159
181
205
232
258
286
312
Eb /N0 , dB
1.63
3.09
3.89
4.64
5.46
6.32
7.19
8.03
8.88
9.73
10.60
11.48
339
12.37
K
25
50
100
150
200
250
300
350
400
450
500
550
600
DE for T = 4, 100 iterations
1.0000x
1.0000x
0.6260x + 0.3740x2
0.3190x + 0.6810x2
0.3228x + 0.6772x2
0.3279x + 0.6721x2
0.3259x + 0.6741x2
0.3333x + 0.6667x2
0.3298x + 0.6702x2
0.3349x + 0.6651x2
0.3381x + 0.6619x2
0.3386x + 0.6436x2 + 0.0177x3
0.3445x + 0.6077x2 + 0.0477x3
M
25
41
47
43
57
71
85
99
113
127
140
153
166
Eb /N0 , dB
0.53
1.09
2.84
3.41
4.04
4.75
5.49
6.24
7.03
7.82
8.63
9.46
10.30
K
25
50
100
150
200
250
300
350
400
450
500
DE for T = 1, 100 iterations
0.0920x + 0.9080x2
1.0000x2
1.0000x2
0.0009x + 0.4560x2 + 0.5431x3
0.1138x + 0.0931x2 + 0.7930x3
0.1371x + 0.0146x2 + 0.8482x3
0.1450x + 0.8550x3
0.1466x + 0.8534x3
0.1531x + 0.7928x3 + 0.0540x4
0.1610x + 0.7137x3 + 0.1253x4
0.1647x + 0.6666x3 + 0.1687x4
0.1700x + 0.6125x3
+0.1955x4 + 0.0219x5
M
68
106
195
203
251
309
368
427
482
535
590
Eb /N0 , dB
3.71
4.45
5.79
6.77
7.66
8.54
9.43
10.35
11.28
12.21
13.16
700
15.07
600
600
A. Optimization procedure
The goal is to find the optimal slot count M and a
polynomial L(x) in order to minimize the Eb /N0 under the
maximum error probability allowed.
The optimization procedure is conducted separately for
every user count and consists of two sub-procedures. The first
one is a constrained local minimum search with respect to L
coefficients and the number of slots used in the system. The
constraints are formed by
• Li ≥ 0 ∀i = 1, . . . , Lmax ,
• L(1) = 1
The error probability is minimized under fixed Eb /N0 at this
step. As soon as constrained local minimum search procedures
can search the local minimum only, one need to run multiple
optimization procedures starting from different random initial
points within constraints.
The second sub-procedure is to find a minimum Eb /N0 that
satisfies maximum allowed error probability Pe . We expect the
error probability at the optimal configuration to be a monotonic
function of Eb /N0 and use a binary search.
The numerical experiment results show that L(x) behaves
smoothly when varying the number of users. This means that
a global minimum is found at every optimization point. Note,
that the error probability has multiple local minimums, because
the slot count changes sharply at several points.
B. Simulation results
Interference cancellation algorithm was tested via Gaussian
MAC Monte Carlo simulations. The result of a single run is a
set of slots and the number of simultaneous transmissions (or
collision index) for each slot. Each user selects the number of
transmissions in accordance to L(x) and then selects particular
non-coinciding slots from uniform distribution during each
run. The same Eb /N0 is assumed for all slots.
The decoding is done in accordance to Algorithm 1. The
only thing we need to explain is how we resolve the collisions.
Error probability is set to 1 if the number of simultaneous
transmissions within some slot exceeds the threshold (T ∈
{1, 2, 4}). If the order of collision is less or equal to T ,
then the error probability is calculated independently for each
transmitted message in a slot in accordance to finite length
random coding bound from [1].
R EFERENCES
[1] Y. Polyanskiy, “A perspective on massive random-access,” in Information
Theory (ISIT), 2017 IEEE International Symposium on. IEEE, 2017, pp.
2523–2527.
[2] B. Rimoldi and R. Urbanke, “A rate-splitting approach to the gaussian
multiple-access channel,” IEEE Transactions on Information Theory,
vol. 42, no. 2, pp. 364–375, 1996.
[3] O. Ordentlich and Y. Polyanskiy, “Low complexity schemes for the
random access gaussian channel,” in 2017 IEEE International Symposium
on Information Theory (ISIT), June 2017, pp. 2528–2532.
[4] A. Vem, K. R. Narayanan, J. Cheng, and J.-F. Chamberland, “A userindependent serial interference cancellation based coding scheme for the
unsourced random access gaussian channel,” in Proc. IEEE Information
Theory Workshop (ITW), 2017, pp. 1–5.
[5] G. Liva, “Graph-based analysis and optimization of contention resolution
diversity slotted aloha,” IEEE Transactions on Communications, vol. 59,
no. 2, pp. 477–487, February 2011.
[6] K. R. Narayanan and H. D. Pfister, “Iterative collision resolution for
slotted aloha: An optimal uncoordinated transmission policy,” in 2012
7th International Symposium on Turbo Codes and Iterative Information
Processing (ISTC), Aug 2012, pp. 136–139.
[7] R. Tanner, “A recursive approach to low complexity codes,” IEEE
Transactions on information theory, vol. 27, no. 5, pp. 533–547, 1981.
[8] T. Richardson and R. Urbanke, Modern coding theory.
Cambridge
university press, 2008.
| 7cs.IT
|
1
Hybrid Algorithm for Multi-Objective Optimization
by Greedy Hypervolume Maximization
arXiv:1506.05424v1 [cs.NE] 17 Jun 2015
Conrado S. Miranda, Fernando J. Von Zuben, Senior Member, IEEE
Abstract—This paper introduces a high-performance hybrid
algorithm, called Hybrid Hypervolume Maximization Algorithm
(H2MA), for multi-objective optimization that alternates between
exploring the decision space and exploiting the already obtained
non-dominated solutions. The proposal is centered on maximizing
the hypervolume indicator, thus converting the multi-objective
problem into a single-objective one. The exploitation employs
gradient-based methods, but considering a single candidate efficient solution at a time, to overcome limitations associated with
population-based approaches and also to allow an easy control
of the number of solutions provided. There is an interchange between two steps. The first step is a deterministic local exploration,
endowed with an automatic procedure to detect stagnation.
When stagnation is detected, the search is switched to a second
step characterized by a stochastic global exploration using an
evolutionary algorithm. Using five ZDT benchmarks with 30
variables, the performance of the new algorithm is compared to
state-of-the-art algorithms for multi-objective optimization, more
specifically NSGA-II, SPEA2, and SMS-EMOA. The solutions
found by the H2MA guide to higher hypervolume and smaller
distance to the true Pareto frontier with significantly less function
evaluations, even when the gradient is estimated numerically.
Furthermore, although only continuous decision spaces have been
considered here, discrete decision spaces could also have been
treated, replacing gradient-based search by hill-climbing. Finally,
a thorough explanation is provided to support the expressive gain
in performance that was achieved.
Index Terms—Exploration-exploitation algorithm; Gradientbased optimization; Hypervolume maximization; Multi-objective
optimization.
I. I NTRODUCTION
ULTI-OBJECTIVE optimization (MOO) is a generalization of the standard single-objective optimization to
problems where multiple criteria are defined and they conflict
with each other [1]. In this case, there can be multiple optimal
solutions with different trade-offs between the objectives.
Since the optimal set can be continuous, an MOO problem is
given by finding samples from the optimal set, called Pareto
set. However, we also wish that the projection of the obtained
samples of the Pareto set into the objective space be welldistributed along the Pareto frontier, which is the counterpart
for the Pareto set, so that the solutions present more diverse
trade-offs.
The current state-of-the-art for MOO relies on the use of
evolutionary algorithms for finding the desired samples [2].
One of these algorithms is the NSGA-II [3], which performs
non-dominance sorting, thus ordering the proposed solutions
according to their relative dominance degree, and dividing the
M
C. S. Miranda and F. J. Von Zuben are with the School of Electrical and
Computer Engineering, University of Campinas (Unicamp), Brazil. E-mail:
{conrado,vonzuben}@dca.fee.unicamp.br
solution set in subsequent frontiers of non-dominated solutions. NSGA-II also uses crowding distance, which measures
how close the nearby solutions are, to maintain diversity in
the objective space. Another well-known algorithm is the
SPEA2 [4], where the solutions have a selective pressure to
move towards the Pareto frontier and also to stay away from
each other.
These algorithms are based on heuristics to define what
characterizes a good set of solutions. However, the hypervolume indicator [5] defines a metric of performance for a set
of solutions, thus allowing a direct comparison of multiple
distinct sets of solutions [6], with higher values indicating
possible better quality. The hypervolume is maximal at the
Pareto frontier and increases if the samples are better distributed along the frontier [7]. Due to these properties, it
represents a good candidate to be maximized in MOO, being
explicitly explored in the SMS-EMOA [8], where solutions
that contribute the least to the hypervolume are discarded.
On the other hand, local search methods have been successful in single-objective optimization (SOO) due to their
efficiency in finding a local optimum for some problems [9],
[10], so that research has been performed to try to adapt
these methods for MOO problems. For instance, [11] defined a
method for finding all minimizing directions in a MOO problem, but the proposed algorithm achieved low performance on
usual benchmark functions.
Alternatively, instead of adapting the single-objective methods to work on MOO problems, we can create a SOO problem
associated with the MOO one, such that a good solution for the
single-objective case is a good solution for the multi-objective
case. Since the hypervolume is able to describe how good a
population is, based on a single indicator, the MOO problem
can be converted into the maximization of the population’s
hypervolume.
Based on this idea, [12] proposed a method to compute the
hypervolume’s gradient for a given population, so that the optimal search direction for each individual could be established.
However, [13] showed that adjusting the population through
integration of the hypervolume’s gradient not always work,
with some initially non-dominated points becoming dominated
and others changing very little over the integration.
In this paper, we introduce an algorithm for maximizing
the hypervolume by optimizing one point at a time, instead of
adjusting a whole population at once. The algorithm alternates
between exploring the space for non-dominated solutions and,
when they are found, exploiting them using local search
methods to maximize the populations’ hypervolume when
only this active point can be moved. Therefore, once the
hypervolume has converged, which is guaranteed to happen
2
because the problem is bounded, the point is fixed in all further
iterations. We found that this restriction is enough to overcome
the issues presented in [13] when using the hypervolume’s
gradient. The proposed algorithm, called Hybrid Hypervolume
Maximization Algorithm (H2MA), is a hybrid one, since
it is composed of global exploration and local exploitation
procedures, properly managed to be executed alternately.
Results over the ZDT benchmark [14] show that the new
algorithm performs better than the state-of-the-art evolutionary
algorithms, both in terms of total hypervolume and distance
to the Pareto frontier. Moreover, the algorithm was able to
work deterministically in most of the benchmark problems,
which makes it less susceptible to variations due to random
number generation. Due to the high quality of the solutions
found in less function evaluations than what is achieved by the
current state-of-the-art, we consider that the new algorithm is
a viable choice for solving MOO problems. Moreover, since
a single solution is introduced at a time, the user is able to
stop the algorithm when the desired number of solutions is
found, while evolutionary algorithms must evolve the whole
population at the same time.
This paper is organized as follows. Section II introduces the
concepts of multi-objective optimization required, including
the hypervolume indicator, and discusses the problems with
the gradient-based approach for hypervolume maximization
introduced in [13]. Section III provides the details of the new
H2MA algorithm, and Section IV shows the comparison with
the state-of-the-art algorithms. Finally, Section V summarizes
the results and discusses future research direction.
II. M ULTI -O BJECTIVE O PTIMIZATION
H YPERVOLUME I NDICATOR
AND THE
A multi-objective optimization problem is described by its
decision space X and a set of objective functions fi (x) : X →
Yi , i ∈ {1, . . . , M }, where Yi ⊆ R is the associated objective space for each objective function [15]. Due to the
symmetry between maximization and minimization, only the
minimization problem is considered here. Each point x in
the decision space has a counterpart in the objective space
Y = Y1 ×· · ·×YM given by y = f (x) = (f1 (x), . . . , fM (x)).
Since there are multiple objectives, a new operator for
comparing solutions must be used, since the conventional
“less than” operator < can only compare two numbers. This
operator is denoted the dominance operator and is defined as
follows.
Definition 1 (Dominance). Let y and y ′ be points in Y, the
objective space. Then y dominates y ′ , denoted y ≺ y ′ , if yi <
yi′ for all i.
From this definition, a point y that dominates another
point y ′ is better than y ′ in all objectives. Thus, there is
no reason someone would choose y ′ over y, and it can be
discarded, as occurs in many multi-objective optimization
algorithms [15]. Note that there are other definitions of the
dominance operator [6], where one considers the inequality ≤
instead of the strict inequality < used here. However, equality
in some of the coordinates may be an issue when using the
hypervolume indicator, such as when taking its derivative [12].
This is why the strict version is used in this work.
Using the dominance, we can define the set of points
characterized by the fact that no other point can have better
performance in all objectives.
Definition 2 (Pareto Set and Frontier). The Pareto set is
defined by the set of all points in the decision space that
are not dominated by any other point in the decision space,
when using the objectives. That is, the Pareto set is given
by P = {x ∈ X | ∄x′ ∈ X : f (x′ ) ≺ f (x)}. The Pareto
frontier is the associated set in the objective space, given by
F = {f (x) | x ∈ P}.
A. The Hypervolume Indicator
In order to define the hypervolume indicator [5], we must
first define the Nadir point, which is a point in the objective
space that is dominated by every point in a set.
Definition 3 (Nadir Point). Let X = {x1 , . . . , xN } ∈ X N be
a set of points in the decision space. Let z ∈ RM be a point
in the objective space. Then z is a valid Nadir point if, for all
x ∈ X and i ∈ {1, . . . , M }, we have that fi (x) < zi . Using
Definition 1, this can be written as f (x) ≺ z.
Again, it is possible to allow equality in the definition of the
Nadir point, just like in the definition of dominance. However,
when equality is allowed, it is possible for some point to have
a null hypervolume, which can guide to undesired decisions
when using the hypervolume as a performance metric, since
such points would not contribute to the hypervolume and
would be replaced by other points. Using the definition of
a Nadir point, we can define the hypervolume for a set of
points.
Definition 4 (Hypervolume). Let X = {x1 , . . . , xN } ∈ X N
be a set of points in the decision space. Let z ∈ RM be a valid
Nadir point in the objective space. Then the hypervolume can
be defined as:
Z
H(X; z) =
1[∃x ∈ X : f (x) ≺ y ≺ z]dy,
(1)
RM
where 1[·] is the indicator function.
The hypervolume measures how much of the objective space
is dominated by a current set X and dominates the Nadir point
z. Fig. 1 shows an example of the hypervolume for a set of
three non-dominated points. For each point, the shaded region
represents the area dominated by the given point, with colors
combining when there is overlap.
B. Gradient of the Hypervolume
As stated earlier, since the hypervolume provides such a
good indicator of performance in multi-objective problems, it
can be used to transform the multi-objective problem into a
single-objective one, characterized by the maximization of the
hypervolume.
Although such approach proved to be successful when using
evolutionary algorithms as the optimization method [8], the
same did not happen when using the hypervolume’s gradient
3
1
0.9
0.8
y2
0.7
0.6
0.5
0.4
0.3
0.2
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
y1
Figure 1: Example of hypervolume. The non-dominated solutions in the objective space are shown in black circles, and
the Nadir point is shown in the black square. For each nondominated solution, the region between it and the Nadir point
is filled, with colors combining when there is overlap, and the
total hypervolume is given by the area of the shaded regions.
Best viewed in color.
to perform the optimization [13]. However, it is well-known
that gradient methods have been successful in single-objective
optimization [9], [10], thus suggesting that they should be a
reasonable choice for multi-objective optimization devoted to
maximizing the hypervolume, since the hypervolume operator
is well-defined almost everywhere in the objective space.
The hypervolume’s gradient for a set of points was introduced in [12], and it can be used to compute the optimal
direction in which a given point should move to increase the
hypervolume associated with the current set of non-dominated
solutions. Although the hypervolume is not a continuously differentiable function of its arguments, since dominated points
do not contribute to the hypervolume and thus have null
gradient, the gradient can be computed whenever any two
points have different values for all objectives.
Based on this motivation, [13] used the hypervolume’s
gradient as a guide for adjusting a set of points by numerical
integration, that is, performing a small step in the direction
pointed by the gradient. Even though the algorithm was able
to achieve the Pareto set in some cases, it failed to converge
to efficient points when some points got stuck along the
iterative process, either because their gradients became very
small or because they became dominated by other points. Once
dominated, these points do not contribute to the hypervolume
and remain fixed. This causes a major issue to using the
hypervolume gradient in practice, since dominated points can
be discarded, because there is no possibility to revert them
to non-dominated points anymore, and the points with small
gradients remain almost stagnant.
If we analyze Eq. (1), we can see that points at the border
in the objective space are the only ones that can fill some
portions of the objective space. On the other hand, points that
are not at the border have less influence in the hypervolume,
since part of the area dominated by them is also dominated
by some other points. In the analysis presented in [13], it is
clear that the cases where some points got stuck had higher
gradients for the border points in the objective space, which
led to the dominance or decrease of contribution of some or
all central points.
To make this idea clearer, consider the example in Fig. 1.
If the point located at (0.75, 0.25) decreases its value on the
second objective, it can increase the population’s hypervolume. Moreover, it is the only point that can do so without
competition for that portion of the space, since it is the point
with the largest value for the first objective. The same holds
for the point at (0.25, 0.75) and the first objective.
However the point located at (0.5, 0.5) has to compete
with the other two points to be the sole contributor for some
regions. Therefore, its effect on the hypervolume is smaller,
which leads to a smaller gradient. Furthermore, if less area is
dominated by the middle point alone, which can occur during
the points adjustment as the middle one moves less, then its
influence becomes even smaller and it can become dominated.
It is important to highlight that this behavior does not
happen always, but can occur along the iterative process, as
shown in [13]. This leads to the base hypothesis for the algorithm developed in this paper: when using the hypervolume’s
gradient for optimization, the competition for increasing the
hypervolume among points should be avoided.
III. H YBRID H YPERVOLUME M AXIMIZATION A LGORITHM
From the discussion in Section II-B, one can see that the
major problem when optimizing the hypervolume directly
using its gradient may be the competition among points.
Therefore, our proposed algorithm optimizes a single solution
at a time, avoiding this competition.
Theoretically, the algorithm can be described by choosing a
new point that maximizes the hypervolume when taking into
account the previous points, such that its recurring equation
can be written as:
xt = arg max H(Xt−1 ∪{x}), Xt = Xt−1 ∪{xt }, t ∈ N, (2)
x∈X
where the initial set is given by X−1 = {}.
Since a single point is being optimized at a time, the optimization becomes simpler and, as we will show in Section IV,
requires less function evaluations. Moreover, one could argue
that maintaining the previous set fixed reduces the flexibility
allowed in comparison with a set where all the points are being
concurrently adjusted. Although this may be true, we will also
show in Section IV that the proposed algorithm performs well
despite this loss of flexibility.
The algorithm described in Eq. (2) is theoretically ideal,
since finding the maximum is hard in practice. Therefore, the
actual algorithm proposed is shown in Fig. 2. This algorithm
performs exploration of the objective space until a new solution that is not dominated by the previous candidate solutions
is found. When it happens, the hypervolume of the whole set
is larger than the hypervolume when considering only previous
candidate solutions.
The new candidate solution is then exploited to maximize
the total hypervolume and, after convergence, is then added to
the existing set. It is important to highlight that the exploitation
phase cannot make the solution become dominated, since
that would reduce the hypervolume in comparison with the
initial condition. Therefore, the problem of points becoming
4
Input: Objectives f
Input: Design space X
Input: Nadir point z
Output: Set of candidate solutions X
function H YBRID G REEDYO PTIMIZER(f, X , z)
Regions, X ← C REATE I NITIAL R EGION(f,X )
while not stop condition and |Regions| > 0 do
R ← Regions.pop()
⊲ Removes the region with the
largest volume
x0 ← E XPLORE D ETERMINISTIC(f,X , R, X)
if x0 is valid then
x ← E XPLOIT (f, X , x0 , X, z)
N ewRegions ← C REATE R EGIONS(R,x, f )
Regions ← Regions ∪ N ewRegions
X ← X ∪ {x}
end if
end while
while not stop condition do
x0 ← E XPLORE S TOCHASTIC(f,X , X)
x ← E XPLOIT (f, X , x0 , X, z)
X ← X ∪ {x}
end while
return X
end function
Figure 2: Hybrid algorithm that performs deterministic and
stochastic exploration until a suitable solution is found, and
then exploits it.
Input: Objectives f
Input: Design space X
Input: Current exploration region R
Input: Set of candidate solutions X
Output: New initial condition x0
function E XPLORE D ETERMINISTIC(f, X , R, X)
x0 ← M EAN(R.X)
Minimize kR.mid − f (x)k from x0 until some candidate x is
not dominated by X
if found non-dominated x then
x0 ← x
else
x0 ← some invalid state
end if
return x0
end function
Figure 3: A deterministic exploration is performed based on
some region.
dominated during the exploitation is avoided. Furthermore, the
exploitation is a traditional single-objective optimization, so
that gradient methods can be used if the decision set X is
continuous or hill-climbing methods can be used for discrete
X.
Once finished the exploitation, the algorithm begins the
exploration phase again. The exploration can be deterministic,
based on regions of the objective space defined by previous
solutions, or stochastic, where a stochastic algorithm, such as
an evolutionary algorithm, is used to find the new candidate.
When a non-dominated candidate is found, the algorithm turns
to exploitation again.
We highlight that the deterministic exploitation algorithm
proposed is based on the definition of these regions, but other
deterministic methods can be used. However, the algorithm
must be able to establish when it is not able to provide
Input: Objectives f
Input: Design space X
Output: Set of candidate solutions X
Output: Initial exploration region R
function C REATE I NITIAL R EGION(f,X )
X ← {}
x0 ← X .mean
⊲ Gets the average candidate
for i = 1, . . . , |f | do
x ← M INIMIZE (fi, x0 , X )
X ← X ∪ {x}
end for
R ← C REATE R EGION(X,f )
return {R}, X
end function
Figure 4: The initial region is created from the points that
minimize each objective individually.
Input: Current explored region R
Input: Current solution x
Input: Objectives f
Output: New exploration regions N ewRegions
function C REATE R EGIONS(R, x, f )
N ewRegions ← {}
for X ′ in C OMBINATIONS(R.X, |R.X| − 1) do
R′ ← C REATE R EGION(X ′ ∪ {x}, f )
N ewRegions ← N ewRegions ∪ {R′ }
end for
return N ewRegions
end function
Figure 5: New exploration regions are created by combining
the current solution with the previous region.
further improvements, so that the change to the stochastic
global exploration can be made. In the algorithm shown in
Fig. 2, regions that do not provide a valid initial condition are
discarded without creating new regions, so that eventually the
algorithm can switch to the stochastic global exploration.
The algorithm for deterministic exploration is shown in
Fig. 3. It combines the points used to create a given region in
order to produce an initial condition and tries to minimize the
distance between its objective value and a reference point.
Once a non-dominated point is found, it is returned for
exploitation. Although this simple optimization provided good
results without requiring many function evaluations, other
methods can be used to perform this exploration. Alternatively,
one can also perform a stochastic exploration instead of a
deterministic one, but this may have negative effects on the
performance if the information provided by the output (region
R) is not used, since a global search would be required.
The first region is created by finding points that minimize
each objective separately, as shown in Fig. 4. This establishes
that the initial region will have a number of candidate solutions
associated with it equal to the number of objectives, so that
the solutions are at the border of the region.
When new regions are created after exploitation, we ignore
the solutions that created the region, one at a time, and replace
it with the proposed new solution, as shown in Fig. 5, to create
a new region. This guarantees that the number of solutions for
each region is kept equal to the number of objectives.
Finally, Fig. 6 shows how a region is created. If a region
5
f (x12 )
f (x2 )
1
R
f (x′ )
0.6
y 12
0.4
H′
0.8
f2 (·)
f2 (·)
0.8
f (x1 )
0.2
0.4
f (x′ )
f (x∗ )
0.2
0
H∗
R2
0.6
0.4
0.2
0
f (x2 )
1
0.6
0.8
1
1.2
0
R1
f (x1 )
0
0.2
0.4
0.6
0.8
1
1.2
f1 (·)
f1 (·)
(b) Exploitation
(a) Deterministic exploration
Figure 7: Deterministic exploration and exploitation steps of the new algorithm in an example problem. The Pareto frontier is
shown in the blue line, and the regions used by the deterministic exploration are shown in yellow.
Input: Objectives f
Input: Set of candidate solutions X
Output: Exploration region R
function Q
C REATE R EGION(X,f )
|
V = |f
i=1 (maxx∈X fi (x) − minx∈X fi (x))
if V > 0 then
R.X ← X
R.mid ← M EAN({f (x) | x ∈ X})
R.V ← V
else
R ← null element such that {R} ≡ {}
end if
return R
end function
Figure 6: An exploration region is created from a set of
candidates if the region have some volume.
does not have a volume, then at least one objective for two
solutions is the same. Although we could allow such region
to exist without modifying the rest of the algorithm, these
regions tend to not provide good candidates for exploitation
and delay the change to stochastic global exploration. Furthermore, one can even prohibit regions with volume smaller
than some known constant, as they probably will not provide
good exploitation points, and the change to stochastic global
exploration happens earlier.
Fig. 7 shows a step of the algorithm in an example problem
with two objectives. The deterministic exploration receives a
region R, composed of the points x1 and x2 . The mean of the
points that compose the region is given by x12 = (x1 + x2 )/2
and its evaluation in the objective space is shown in Fig. 7a.
The mean objective of the points that compose the region is
also computed and is shown as y 12 = (f (x1 ) + f (x2 ))/2. The
deterministic exploration is then defined by the problem
min kf (x) − y 12 k,
x∈X
f (x2 ), this should guide the search to the non-dominated
region of the space.
While performing this optimization, some intermediary
points are evaluated, either while computing the numeric
gradient or after performing a gradient descent step. The deterministic exploration stops as soon as a non-dominated point
is found, which is given by f (x′ ) in the example in Fig. 7a.
Note that this example shows f (x12 ) as being dominated by
f (x1 ) and f (x2 ), but it can also be non-dominated. In this
case, x′ = x12 and no optimization step for the problem in
Eq. (3) is performed. Supposing no non-dominated point f (x′ )
is found during the deterministic exploration, the region is
simply discarded, without performing an exploitation step.
Using the point x′ , whose f (x′ ) is non-dominated, provided
by the deterministic or stochastic exploration, the exploitation
is performed. Fig. 7b shows the hypervolume contributions
for the initial point x′ and the optimal point x∗ , which
maximizes the total hypervolume as in Eq. (2). Since x′ is
non-dominated, its hypervolume contribution H ′ is positive
and the hypervolume gradient relative to the objectives is
non-zero. After finding x∗ and if x′ was provided by the
deterministic exploration, new regions must be created to allow
further exploration. Therefore, according to Fig. 5, the regions
R1 = (x1 , x∗ ) and R2 = (x2 , x∗ ) are created for further
exploration.
This finalizes a step of the algorithm, which is repeated until
the given stop condition is not met. As at most one point is
found by each step, the stop condition can be defined based
on the number of desired points.
Note that all the methods used in this algorithm assume that
the optimization, either for exploitation or for minimizing one
objective alone, requires an initial condition. This is true for
hill climbing or gradient methods, but the algorithm can easily
be modified if the optimization does not require it.
(3)
which uses x12 as the initial condition for the optimization.
Since y12 is guaranteed to be non-dominated by f (x1 ) and
IV. E XPERIMENTAL R ESULTS
To compare the algorithm proposed in Section III, called
Hybrid Hypervolume Maximization Algorithm (H2MA), with
6
22
1
NSGA−II
SPEA2
SMS1
SMS2
H2MA
0.5
21
0
20
−0.5
H
log10 P
19
18
−1
−1.5
−2
17
−2.5
NSGA−II
SPEA2
SMS1
SMS2
H2MA
16
15
−0.5
0
0.5
1
1.5
2
# evaluations
−3
2.5
4
x 10
(a) Hypervolume
−3.5
−0.5
0
0.5
1
1.5
2
# evaluations
2.5
4
x 10
(b) P-distance. Zero values not shown.
Figure 8: 0th, 25th, 50th, 75th, and 100th percentiles every 2000 evaluations for the all algorithms on ZDT1.
Table I: Benchmark problems used for evaluation. See the
Appendix for Eqs. (4) to (8).
Problem
ZDT1
ZDT2
ZDT3
ZDT4
ZDT6
Objectives
Eq. (4)
Eq. (5)
Eq. (6)
Eq. (7)
Eq. (8)
X
[0, 1]n
[0, 1]n
[0, 1]n
[0, 1] × [−5, 5]n−1
[0, 1]n
Nadir point
(2, 11)
(2, 11)
(2, 11)
(2, 2 + 50(n − 1))
(2, 11)
the existing algorithms, the ZDT family of functions [14] was
chosen. These functions define a common benchmark set in
the multi-objective optimization literature, since they define a
wide range of problems to test different characteristics of the
optimization algorithm. All functions defined in [14] have a
continuous decision space X , except for the ZDT5 which has
a binary space. In this paper, only the continuous test functions
were used to evaluate the performance of the new algorithm,
and their equations are shown in the appendix.
Table I provides a summary of the evaluation functions,
their decision spaces, and the Nadir points used to compute the
hypervolume. The Nadir points are defined by upper bounds
of the objectives, which guarantees that the hypervolume
computation is always valid, plus one, since not adding an
extra value would mean that points at the border of the frontier
would have no contribution to the hypervolume and would be
avoided. In all instances, a total of n = 30 variables were
considered, as common in the literature. The evolutionary
algorithms’ and evaluation functions’ implementations were
given by the PaGMO library [16].
We compare our algorithm with existing state-of-the-art
multi-objective optimization algorithms, namely NSGA-II [3],
SPEA2 [4], and SMS-EMOA [8]. All of them used a population size of 100 individuals. Tests have shown that this size is
able to provide a good performance due to balance between
exploration of the space and exploitation of the individuals,
with much less individuals not providing good exploration and
much more not providing good exploitation. The SMS-EMOA
can use two methods for selecting points in dominated fronts:
the least hypervolume contribution or the domination count.
Both methods were tested, with labels SMS1 and SMS2,
respectively, in the following figures. Note that this method
only applies for the dominated fronts, since the domination
count is zero for all points in the non-dominated front and
the least contributor method must be used. Furthermore, the
SMS-EMOA algorithm’s performance presented in this paper
uses a dynamic Nadir point, which is found by adding one to
the maximum over all points in each objective, since using
the Nadir points presented in Table I created a very high
selective pressure, which in turn led to poor exploration and
performance.
Since the decision space and objectives are continuous, the
exploitation and deterministic exploration methods may resort
to a gradient-based algorithm. In this paper, we used the LBFGS-B method implemented in the library SciPy [17], which
is able to handle the bounds of X and is very efficient to find
a local optimum. As the other algorithms being compared are
evolutionary algorithms, which can only access the objective
functions by evaluating them at given points, the gradient for
the L-BFGS-B is computed numerically to avoid an unfair
advantage in favor of our algorithm.
For the stochastic global exploration, we used an evolutionary algorithm with non-dominance sorting and removal
based on the number of dominating points. The population
had a minimum size of 20 and was filled with the given set
of previous solutions X. If less than 20 points were provided,
the others were created by randomly sampling the decision
space X uniformly. Once a new point is introduced to the
non-dominated front, it is returned for exploitation because it
increases the hypervolume when added to the previous solutions X. The size of this population was chosen experimentally
to provide a good enough exploration of the space toward
the initial conditions for the exploitation. This size is smaller
than the population size for the pure evolutionary algorithms
7
22
1
NSGA−II
SPEA2
SMS1
SMS2
H2MA
21
0.5
20
0
19
−0.5
H
log10 P
18
17
16
−1
−1.5
−2
15
−2.5
14
NSGA−II
SPEA2
SMS1
SMS2
H2MA
13
12
−0.5
0
0.5
1
1.5
2
# evaluations
−3
−3.5
−0.5
2.5
0
0.5
1
1.5
2
2.5
# evaluations
4
x 10
(a) Hypervolume
4
x 10
(b) P-distance. Zero values not shown.
Figure 9: 0th, 25th, 50th, 75th, and 100th percentiles every 2000 evaluations for the all algorithms on ZDT2.
23
1
NSGA−II
SPEA2
SMS1
SMS2
H2MA
0.5
22
0
21
−0.5
H
log10 P
20
19
−1
−1.5
18
−2
NSGA−II
SPEA2
SMS1
SMS2
H2MA
17
16
−0.5
0
0.5
1
# evaluations
1.5
2
−2.5
2.5
4
x 10
(a) Hypervolume
−3
−0.5
0
0.5
1
1.5
2
# evaluations
2.5
4
x 10
(b) P-distance. Zero values not shown.
Figure 10: 0th, 25th, 50th, 75th, and 100th percentiles every 2000 evaluations for the all algorithms on ZDT3.
because the pure evolutionary algorithm need diversity to
explore and exploit all of its population, but the stochastic
part of the H2MA is already initialized with good and diverse
candidate solutions provided by the exploitation procedure,
reducing its exploration requirements.
Besides computing the solutions’ hypervolume, which is the
metric that the H2MA is trying to maximize and that provides
a good method for comparing solutions, we can compute
the distance between the achieved objectives and the Pareto
frontier, since the Pareto frontiers for the ZDT functions are
known. This defines the P-distance, which is zero, or close to
zero due to numerical issues, for points at the frontier.
Figs. 8, 9, 10, 11, and 12 present the results for the problems ZDT1, ZDT2, ZDT3, ZDT4, and ZDT6, respectively.
A maximum of 20000 function evaluations was considered,
and the graphs show the 0th, 25th, 50th, 75th, and 100th
percentiles for each performance indicator over 100 runs of
the algorithms. Since the P-distance is shown in log-scale,
some values obtained by our proposal are absent or partially
present, because they have produced zero P-distance.
From ZDT1 to ZDT4, the H2MA never ran out of regions
to explore, so the stochastic exploration was not used and all
runs have the same performance. For the function ZDT6, the
first objective, given by Eq. (8a), causes some problems to the
deterministic exploration.
During the creation of the first region for this problem, the
mean point is used as initial condition for optimizing each
objective, as shown in Fig. 4. However, the first objective for
ZDT6 has null derivative when x1 = 0.5. In this case, even
traditional higher-order methods would not help, since the first
non-zero derivative of f1 (x) is the sixth. As the first objective
does not change in this case and it also has local minima that
8
4
3000
2900
NSGA−II
SPEA2
SMS1
SMS2
H2MA
2
2800
0
2700
−2
H
log10 P
2600
2500
2400
−4
−6
2300
−8
2200
NSGA−II
SPEA2
SMS1
SMS2
H2MA
2100
2000
−0.5
0
0.5
1
1.5
2
# evaluations
−10
−12
−0.5
2.5
0
0.5
1
1.5
2
2.5
# evaluations
4
x 10
(a) Hypervolume
4
x 10
(b) P-distance. Zero values not shown.
Figure 11: 0th, 25th, 50th, 75th, and 100th percentiles every 2000 evaluations for the all algorithms on ZDT4.
20
1
18
0.8
16
0.6
NSGA−II
SPEA2
SMS1
SMS2
H2MA
0.4
H
log10 P
14
12
0.2
0
10
−0.2
8
−0.4
NSGA−II
SPEA2
SMS1
SMS2
H2MA
6
4
−0.5
0
0.5
1
# evaluations
1.5
2
−0.6
2.5
4
x 10
(a) Hypervolume
−0.8
−0.5
0
0.5
1
1.5
2
# evaluations
2.5
4
x 10
(b) P-distance. Zero values not shown.
Figure 12: 0th, 25th, 50th, 75th, and 100th percentiles every 2000 evaluations for the all algorithms on ZDT6.
are very hard to overcome, the algorithm quickly switches to
using stochastic exploration. Once new regions have candidate
points, the algorithm is able to exploit them.
Besides this issue in the deterministic exploration of the
problem ZDT6, the local minima of the first objective makes
some candidate solutions be sub-optimal, increasing the Pdistance as shown in Fig. 12b. Nonetheless, the achieved Pdistance is better than the evolutionary algorithms and the 75th
percentile is zero. Moreover, Figs. 8b, 9b, 10b, and 11b show
that the candidate solutions are always on the Pareto frontier
for the problems ZDT1 to ZDT4. This allows the user to stop
the optimization at any number of evaluations, even with very
few function evaluations, and have a reasonable expectation
that the solutions found are efficient.
When we evaluate the hypervolume indicator, we see that,
for the problems ZDT1, ZDT2, ZDT4, and ZDT6, the per-
formance of the H2MA is much better, even for the last
one using stochastic exploration. Moreover, the H2MA’s worst
hypervolume was always better than the best hypervolume for
all evolutionary algorithms and it was able to get closer to the
maximum hypervolume possible with relatively few function
evaluations, being a strong indication of its efficiency.
For the problem ZDT3, whose hypervolume performance
is shown in Fig. 10a, the H2MA was generally better than
the evolutionary algorithms. The Pareto frontier for ZDT3 is
composed of disconnected sets of points, which was created
to test the algorithm’s ability to deal with discontinuous
frontiers. Since the exploitation algorithm used for the results
is gradient-based, it is not able to properly handle discontinuous functions, which is the case of the hypervolume on
discontinuous frontiers. However, the deterministic exploration
method is able to find points whose exploitation lay on the
9
4
2
x 10
1.8
1.6
1.4
Numeric
1.2
1
0.8
0.6
0.4
ZDT1
ZDT2
ZDT3
ZDT4
0.2
0
0
100
200
300
400
500
600
700
Analytic
Figure 13: Comparison between the number of function
evaluations required to achieve the same hypervolume using
numeric or analytic gradient. The dotted line represents a 30fold improvement.
different parts of the Pareto frontier, providing the expected
diversity.
Fig. 13 shows a comparison between the number of function
evaluations required by the numeric and the analytic gradient
to achieve the same hypervolume on the problems ZDT1 to
ZDT4. The analytic method for computing the hypervolume’s
gradient is described in [12]. The comparison for ZDT6 is not
shown due to its different scale, since many function evaluations are used in the global stochastic exploration because the
deterministic exploration fails to find regions.
As expected, using the analytic gradient causes a 30-fold
improvement in comparison to the numeric gradient, since
the number of decision variables is 30. However, the gain
is not linear. This can be explained by the difference in
behavior during the deterministic exploration: the first nondominated point found is used to perform the exploitation,
even if this point was found during the computation of the
numeric gradient. For ZDT1 and ZDT4, this causes the new
points found by the numeric gradient to be very close to the
original points, reducing its performance and increasing the
improvement of using the analytic gradient.
Moreover, a similar effect makes the ZDT3 performance to
have a lower improvement when using the analytic gradient.
Since the Pareto frontier for ZDT3 is discontinuous and this
causes a discontinuity in the hypervolume, these large changes
can be seen by the numeric gradient because small changes
in the variables can have large effects on the hypervolume,
pulling the solution if the difference is significant, while
the analytic gradient is not able to provide such knowledge.
Nonetheless, the analytic gradient presents at least a 15-fold
improvement over the numeric one over the ZDT3.
A. Analysis of the H2MA’s performance
As shown in Section IV, the proposed H2MA is able to
surpass the state-of-the-art in multi-objective optimization,
based on evolutionary algorithms. Therefore, it is important
to analyze the algorithm and to discuss why this improvement
happened.
Evolutionary algorithms perform a guided exploration, with
new individuals created based on existing high-performing
individuals, which allows them to escape local minima but
reduces the convergence speed. On the other hand, traditional
optimization algorithms tend to find local minima quickly, but
the optimal point achieved depends on the minima’s regions
of attraction.
These two kind of algorithms have complementary natures,
which makes them good candidates for creating a hybrid
algorithm: the evolutionary algorithm explores the space and
provides initial conditions for the local optimization, which
then finds minima quickly. Although this does create better
results, it only explains the performance on the ZDT6 problem,
since the other problems did not enter the stochastic phase.
In order to understand the algorithm’s behavior, we must
keep in mind that each new point added by the algorithm is
solving a very different problem. Since the previous points that
are considered during the hypervolume optimization change
as more points are added, the objective surface for each new
point is different from the previous ones and takes into account
the already achieved portion of the hypervolume. To visualize
this, supposed that the hypervolume’s gradient is defined over
previously found points and one previous solution is used as
the initial condition for the gradient-based exploitation to find
a new point to be added to the solution set. Although the
initial condition was a local optimum for a previous problem,
it is not a local optimum to the current problem, because any
small change that creates a non-dominated point will improve
the total hypervolume. Therefore, we do not need to worry
about the new optimization converging to a previous solution
point because the problem landscape is different and different
local minima will be found, increasing the total hypervolume.
The deterministic exploration is only required because the
hypervolume’s gradient is not defined at the border of the
hypervolume, so a new independent point must be found.
This explains the performance improvement over ZDT1 to
ZDT4, because every added point improves the hypervolume
as much as it can do locally, so that an improvement is guaranteed to happen. Evolutionary algorithms, on the other hand,
use function evaluations without guarantees of improvement
of the total hypervolume, since dominated solutions can be
found.
Moreover, although a local optimum found during exploitation may not be an efficient solution due to irregularities in
the objective surface, the experiments show that this is not the
case most of the time, since the P-distance of the solutions
found are generally zero. This result is expected, since the
hypervolume is maximal when computed over points in the
Pareto set, and the performance on all ZDT problems provide
support to this claim.
We must highlight that we are not saying that evolutionary
algorithms should not be used at all, but that they should
10
be applied whenever traditional optimization methods are not
able to solve the problem. This is the case of the ZDT6,
for instance, where an evolutionary algorithm was required
to provide initial conditions for the exploitation. We consider
very important to have alternative methods that are better on
a subset of the problems and to use them when a problem
from such subset is present. This is exactly what the H2MA
does: when the traditional optimization is not able to find an
answer, which indicates that the problem is outside of the
subset with which it can deal, an evolutionary algorithm, which
is able to handle a superset class of problems, is used until
the problem becomes part of the subset again, establishing a
switching behavior that takes advantage of both algorithms.
V. C ONCLUSION
This paper proposed the Hybrid Hypervolume Maximization
Algorithm (H2MA) for multi-objective optimization, which
tries to maximize the hypervolume one point at a time. It first
tries to perform deterministic local exploration and, when it
gets stuck, it switches to stochastic global exploration using
an evolutionary algorithm. The optimization algorithm used
during deterministic optimization is problem-dependent and
can be given by a gradient-based method, when the decision
space is continuous, or a hill-climbing method, when the
decision space is discrete. Here we have explored solely
continuous decision spaces.
The algorithm was compared with state-of-the-art algorithms for multi-objective optimization, namely NSGA-II,
SPEA2, and SMS-EMOA on the ZDT problems. Despite using
numeric gradient for the objective functions, which increases
the number of function calls, the algorithm consistently provided a higher hypervolume for the same number of function
evaluations when compared to the aforementioned evolutionary algorithms. Only for the ZDT3 the performance was
slightly reduced due to the discontinuous nature of the Pareto
frontier, which causes a discontinuity in the hypervolume, not
properly handled by gradient-based methods.
Moreover, for all problems except for ZDT6, all the solutions found by the algorithm were over the Pareto frontier,
which makes them efficient solutions. For the ZDT6, the
median case also had all solutions over the Pareto frontier,
but the use of the stochastic exploration not always guided
to a solution at the Pareto frontier. Nonetheless, the obtained
solutions were better than those provided by the evolutionary
algorithms. Moreover, the solutions provided for ZDT1 to
ZDT4 achieved high performance using only the deterministic
part of the algorithm.
Evolutionary algorithms usually have better performance
when their populations are larger, so that diverse individuals
can be selected for crossover. However, most of the time
people do not require many options, so the H2MA presents
itself as an alternative choice for finding a good set of solutions
at a lower computational cost in most problems, although it
does not limit the computational burden and the number of
points found. If the problem has more reasonable objectives
than ZDT6, which was designed with an extreme case in
mind, we can expect that many points will be found by the
deterministic mechanisms, which makes the algorithm more
reliable. Moreover, the solutions found should be efficient,
which is characterized by a low P-distance, and diverse on
the objectives, which is characterized by a larger hypervolume
when only efficient solutions are considered.
Future work should focus on using surrogates to reduce the
number of evaluations [18], [19], [20]. Although the H2MA
is very efficient on its evaluations, the numeric gradient may
consume lots of evaluations and be unreliable for complicated
functions, as their implementation can cause numerical errors
larger than the step used. Using a surrogate, the gradient can
be determined directly and less evaluations are required.
Another important research problem is to find a new algorithm for computing the hypervolume, since existing algorithms are mainly focused on computing the hypervolume
given a set of points [21]. Since the solution set is constructed
one solution at a time in the H2MA, a recursive algorithm that
computes the hypervolume of X ∪ {x} given the hypervolume
of X should reduce the computing requirement.
A PPENDIX
ZDT1:
f1 (x) = x1
ZDT2:
(4a)
f2 (x) = g(x)h(f1 (x), g(x))
n
9 X
xi
g(x) = 1 +
n − 1 i=2
p
h(f1 (x), g(x)) = 1 − f1 (x)/g(x)
(4b)
(4c)
(4d)
f1 (x) = x1
(5a)
f2 (x) = g(x)h(f1 (x), g(x))
n
9 X
xi
g(x) = 1 +
n − 1 i=2
(5b)
h(f1 (x), g(x)) = 1 − (f1 (x)/g(x))2
(5c)
(5d)
ZDT3:
f1 (x) = x1
f2 (x) = g(x)h(f1 (x), g(x))
n
9 X
xi
g(x) = 1 +
n − 1 i=2
s
f1 (x)
f1 (x)
h(f1 (x), g(x)) = 1 −
− sin(10πf1 (x))
g(x)
g(x)
(6a)
(6b)
(6c)
(6d)
ZDT4:
f1 (x) = x1
(7a)
f2 (x) = g(x)h(f1 (x), g(x))
(7b)
n
X
(x2i − 10 cos(4πxi ))
g(x) = 1 + 10(n − 1) +
i=2
(7c)
p
h(f1 (x), g(x)) = 1 − f1 (x)/g(x)
(7d)
11
ZDT6:
f1 (x) = 1 − exp(−4x1 ) sin6 (6πx1 )
(8a)
f2 (x) = g(x)h(f1 (x), g(x))
!0.25
n
X
xi
g(x) = 1 + 9
n−1
i=2
(8b)
h(f1 (x), g(x)) = 1 − (f1 (x)/g(x))2
[21] N. Beume, C. M. Fonseca, M. López-Ibáñez, L. Paquete, and J. Vahrenhold, “On the complexity of computing the hypervolume indicator,”
Evolutionary Computation, IEEE Transactions on, vol. 13, no. 5, pp.
1075–1082, 2009.
(8c)
(8d)
ACKNOWLEDGMENT
Conrado S. Miranda received his M.S. degree on Mechanical Engineering
and his B.S. in Control and Automation Engineering from the University of
Campinas (Unicamp), Brazil, in 2014 and 2011, respectively. He is currently a
Ph.D. student at the School of Electrical and Computer Engineering, Unicamp.
His main research interests are machine learning, multi-objective optimization,
neural networks, and statistical models.
The authors would like to thank CNPq for the financial
support.
R EFERENCES
[1] K. Miettinen, Nonlinear Multiobjective Optimization. Springer US,
1999.
[2] K. Deb, Multi-Objective Optimization Using Evolutionary Algorithms.
John Wiley & Sons, 2001.
[3] K. Deb, A. Pratap, S. Agarwal, and T. A. M. T. Meyarivan, “A fast
and elitist multiobjective genetic algorithm: NSGA-II,” Evolutionary
Computation, IEEE Transactions on, vol. 6, no. 2, pp. 182–197, 2002.
[4] E. Zitzler, M. Laumanns, and L. Thiele, “SPEA2: Improving the strength
Pareto evolutionary algorithm,” 2001.
[5] E. Zitzler, D. Brockhoff, and L. Thiele, “The hypervolume indicator
revisited: On the design of Pareto-compliant indicators via weighted
integration,” in Evolutionary multi-criterion optimization.
Springer,
2007, pp. 862–876.
[6] E. Zitzler, L. Thiele, M. Laumanns, C. M. Fonseca, and V. G.
Da Fonseca, “Performance Assessment of Multiobjective Optimizers: An
Analysis and Review,” Evolutionary Computation, IEEE Transactions
on, vol. 7, no. 2, pp. 117–132, 2003.
[7] A. Auger, J. Bader, D. Brockhoff, and E. Zitzler, “Theory of the
hypervolume indicator: optimal µ-distributions and the choice of the
reference point,” in Proceedings of the tenth ACM SIGEVO workshop
on Foundations of genetic algorithms. ACM, 2009, pp. 87–102.
[8] N. Beume, B. Naujoks, and M. Emmerich, “SMS-EMOA: Multiobjective selection based on dominated hypervolume,” European Journal of
Operational Research, vol. 181, no. 3, pp. 1653–1669, 2007.
[9] D. Bertsekas, Nonlinear programming. Athena Scientific, 1999.
[10] D. Luenberger and Y. Ye, Linear and Nonlinear Programming. Springer
US, 2008.
[11] P. A. N. Bosman, “On gradients and hybrid evolutionary algorithms
for real-valued multiobjective optimization,” Evolutionary Computation,
IEEE Transactions on, vol. 16, no. 1, pp. 51–69, 2012.
[12] M. Emmerich and A. Deutz, “Time Complexity and Zeros of the
Hypervolume Indicator Gradient Field,” in EVOLVE-A Bridge between
Probability, Set Oriented Numerics, and Evolutionary Computation III.
Springer, 2014, pp. 169–193.
[13] V. A. S. Hernández, O. Schütze, and M. Emmerich, “Hypervolume
Maximization via Set Based Newton’s Method,” in EVOLVE-A Bridge
between Probability, Set Oriented Numerics, and Evolutionary Computation V. Springer, 2014, pp. 15–28.
[14] E. Zitzler, K. Deb, and L. Thiele, “Comparison of Multiobjective Evolutionary Algorithms: Empirical Results,” Evolutionary Computation,
vol. 8, no. 2, pp. 173–195, 2000.
[15] K. Deb, “Multi-objective optimization,” in Search methodologies.
Springer, 2014, pp. 403–449.
[16] F. Biscani, D. Izzo, and C. H. Yam, “A global optimisation toolbox for massively parallel engineering optimisation,” arXiv preprint
arXiv:1004.3824, 2010.
[17] E. Jones, T. Oliphant, P. Peterson et al., “SciPy: Open source scientific
tools for Python,” 2001–, [Online; accessed 2015-05-10]. [Online].
Available: http://www.scipy.org/
[18] Y. Jin, “A Comprehensive Survey of Fitness Approximation in Evolutionary Computation,” Soft computing, vol. 9, no. 1, pp. 3–12, 2005.
[19] J. Knowles and H. Nakayama, “Meta-Modeling in Multiobjective Optimization,” in Multiobjective Optimization. Springer, 2008, pp. 245–284.
[20] I. Voutchkov and A. Keane, “Multi-objective Optimization Using Surrogates,” in Computational Intelligence in Optimization. Springer, 2010,
pp. 155–175.
Fernando J. Von Zuben received his Dr.E.E. degree
from the University of Campinas (Unicamp), Campinas, SP, Brazil, in 1996. He is currently the header
of the Laboratory of Bioinformatics and Bioinspired
Computing (LBiC), and a Full Professor at the
Department of Computer Engineering and Industrial
Automation, School of Electrical and Computer Engineering, University of Campinas (Unicamp). The
main topics of his research are computational intelligence, natural computing, multivariate data analysis,
and machine learning. He coordinates open-ended
research projects in these topics, tackling real-world problems in the areas
of information technology, decision-making, pattern recognition, and discrete
and continuous optimization.
| 9cs.NE
|
arXiv:1609.01763v2 [math.GT] 28 Nov 2016
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY
HYPERBOLIC GROUPS
LEONID POTYAGAILO AND WEN-YUAN YANG
Abstract. In this paper, we study the Hausdorff dimension of the Floyd and
Bowditch boundaries of a relatively hyperbolic group, and show that for the
Floyd metric and shortcut metrics respectively, they are are both equal to a
constant times the growth rate of the group.
In the proof, we study a special class of conical points called uniformly
conical points and establish that, in both boundaries, there exists a sequence
of Alhfors regular sets with dimension tending to the Hausdorff dimension and
these sets consist of uniformly conical points.
1. Introduction
1.1. Main results. The main goal of the paper is to calculate the Hausdorff dimension of the limit set of a geometrically finite action of a finitely generated group
G on a compactum X. Every action G y X we consider is a convergence action,
i.e. the induced action on the space of the distinct triples is discontinuous. We say
that G y X is minimal if X coincides with the limit set ΛX G (or ΛG if X is fixed)
of the action, which is the set of the accumulation points of every orbit Gx (x ∈ X).
A point ξ ∈ X is called conical if there exists a sequence of elements gn ∈ G
(n ≥ 1) such that the closure of {(gn ξ, gn η) : n ≥ 1} in X 2 is disjoint from the
diagonal ∆(X 2 ) = {(x, x) : x ∈ X} for any η ∈ X \ ξ. If, in addition, the set of
−1
elements {gn gn+1
: n ≥ 1} is in a uniformly bounded distance from the identity,
then ξ is called uniformly conical. A quantitative version of an L-uniformly conical
point for L ≥ 0 is given in Definition 2.2.
The action of a subgroup H < G on X is parabolic if H fixes a point p ∈ X,
called parabolic fixed point. The parabolic action is bounded parabolic if H acts
properly and cocompactly on X \ {p}. We will always assume that the action of the
whole group G is non-parabolic so there is no a global fixed point.
A minimal non-parabolic action G y X is called geometrically finite (or relatively hyperbolic) if every point x ∈ X is either conical or bounded parabolic (cf.
Definition 2.2). The stabilizer of a parabolic point is a maximal parabolic subgroup
of G. We denote by P the set of maximal parabolic subgroups and call it peripheral
system for the action. A group is called relatively hyperbolic with respect to P if
G admits a geometrically finite action on X with the peripheral system P. If the
compactum X on which G acts is metrizable then the action is geometrically finite
if and only if the induced action on the space of distinct pairs is co-compact (we say
in this case that the action on X is 2-cocompact) [11]. If the opposite is not stated
Date: April 11, 2016.
2000 Mathematics Subject Classification. Primary 20F65, 20F67.
This research is supported by the ERC starting grant GA 257110 RaWG and by ANR grant
BLAN2011BS0101304.
1
2
LEONID POTYAGAILO AND WEN-YUAN YANG
we will always assume that a relatively hyperbolic group is finitely generated and
so X is metrizable.
Let G be a group with a finite generating set S. Assume that 1 ∈
/ S and S = S −1 .
Consider the word metric dS on G. Denote B(n) = {g ∈ G : dS (1, g) ≤ n} for n ≥ 0.
The growth rate δG,S of G relative to S is the limit
log ]B(n)
.
n
Recall that Floyd completion of a group G generated by S is the Cauchy completion of the Cayley graph G (G, S) equipped with the distance ρoλ obtained by
rescaling the length of an edge e ∈ G (G, S) by a scalar function λd(e,o) for a fixed
λ ∈]0, 1[ and a basepoint o ∈ G. The distance ρoλ is called Floyd distance at o, and
we use the notation ρ if o and λ are clear from the context (see Subsection 2.2 for
more details). We denote by Gλ and ∂λ G the corresponding Floyd completion and
its boundary respectively. By V. Gerasimov’s theorem [12, Proposition 3.4.6] for
every finitely generated relatively hyperbolic group the space ∂λ G is the universal
pullback space for every geometrically finite action of G y X in the sense that
there exists an equivariant continuous mapping F : ∂λ G → X (called Floyd map).
A. Karlsson proved that the action of G on the compact space Gλ is a convergence
action [19]. Let ∂λc G (resp. ∂λuc G) denote the set of all (resp. uniformly) conical
points for the action. We denote by Hdimρ the Hausdorff dimension with respect
to ρ = ρλ,o . The first main result of the paper is the following.
δG,S = lim
n→∞
Theorem 1.1. Let G be a relatively hyperbolic group with a finite generating set
S. There exists a constant 0 < λ0 < 1 such that
Hdimρ (∂λ G) = Hdimρ (∂λc G) = Hdimρ (∂λuc G) = −
δG,S
log λ
for any λ ∈ [λ0 , 1).
Remark. Note that for a hyperbolic group the Floyd metric is bilipschitz equivalent to the visual metric on the Gromov boundary (with appropriate choices of
parameters). Even though this result seems to be a folklore, we have not found the
corresponding reference in the literature. We provide a proof of it in the Appendix. As a consequence the result of M. Coornaert [6] for the hyperbolic groups is a
partial case of Theorem 1.1.
Note that the action of G on the Floyd boundary ∂λ G is not necessarily geometrically finite, as it is shown in [29] for Dunwoody’s inaccessible groups. In particular
the Floyd boundary is not in general homeomorphic to the limit set ΛG. So it is
natural to ask if an analogous result to Theorem 1.1 is true for ΛG.
Consider a minimal geometrically finite action of G on a compact X = ΛG. It is
shown in [13] that the Floyd metric ρ transferred by the Floyd map F : ∂λ G → ΛG
is a metric on ΛG, called shortcut metric, and is denoted by ρ̄ (see subsection 2.2).
Our next goal is to calculate the Hausdorff dimension Hdimρ̄ of ΛG with respect
to ρ̄. Denote by Λuc G the set of uniformly conical points of ΛG. The following
theorem provides the same conclusion for the shortcut metric as in the case of the
Floyd metric.
Theorem 1.2. Let G be a group with a finite generating set S acting geometrically
finitely on a compactum X = ΛG. Then there exists a constant 0 < λ0 < 1 such
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS
3
that
Hdimρ̄ (ΛG) = Hdimρ̄ (Λuc G) = −
δG,S
log λ
for any λ ∈ [λ0 , 1).
The above theorems implies the following.
Corollary 1.3. For any shortcut metric ρ̄, the Hausdorff dimension of the limit
set of every relatively hyperbolic action of a group G is constant and is equal to
Hdimρ (∂λ G) = Hdimρ̄ (ΛG) = −
δG,S
log λ
for any λ ∈ [λ0 , 1[ where λ0 ∈]0, 1[ is a fixed number.
We say that a metric space X is Ahlfors Q-regular for a constant Q > 0 if there
exists a Borel measure µ on X such that the following holds
µ(B(x, r)) rQ
for any open ball B(x, r) centered at x ∈ X of radius r > 0, where the symbol
denotes the bilipschitz equivalence between two quantities: C −1 rQ ≤ µ(B(x, r)) ≤
CrQ for a uniform constant C.
Our next main result shows that the Hausdorff dimension of the Floyd boundary
and of the limit set of a relatively hyperbolic action can be well-approximated by
a sequence of Ahlfors regular subsets.
Theorem 1.4. Let G be a finitely generated relatively hyperbolic group with a finite
generating set S. Then there exists a sequence of Ahlfors Qi -regular subsets Xi in
∂λ G or ΛG such that Xi consists of uniformly conical points, 0 < Qi < δG,S and
Qi → δG,S as i → ∞.
The proof of Theorem 1.4 is based on the existence of an L-transitional geodesic
tree T = T (L) ⊂ G (Lemma 3.7) depending on a parameter L 0. Every vertex
of T is a central point of a geodesic interval, whose size depends on L, and which
belongs to a neighbourhood of a left coset (horosphere) gP where P ∈ P (see
Subsection 2.4). We show that the endpoints of such a tree are L-uniformly conical
(Lemma 2.12). However it is not true in general that every uniformly conical point
appears as an endpoint of an L-transitional tree for a bounded L (see the discussion
after Lemma 2.12). The proof of Theorem 1.4 shows that the Hausdorff dimension
of the endpoints of L-transitional trees well approximate the Hausdorff dimension
of the Floyd boundary (or the limit set) if L → ∞. We recapitulate all these facts
in the following.
Corollary 1.5. There exists a sequence Ti of Li -transitional trees such that Xi =
∂Ti are Ahlfors Qi -regular spaces from the statement of Theorem 1.4.
Our next result given in Section 5 generalizes the result of [14] that the geodesics
of the Floyd metrics are approximated by so called tight paths. Considering generalized tight paths (see Definition 5.9) we show that they approximate the geodesics
with respect to the shortcut metric defined on the limit set.
uc
uc
Let ∂L,o
G and Λuc
L,o G denote subsets of uniformly conical points in ∂λ G and
uc
Λ G depending on the above parameter L (see Subsection 2.4 for the precise
definitions). The following result is central in this Section.
4
LEONID POTYAGAILO AND WEN-YUAN YANG
Theorem 1.6 (Proposition 5.13). Under the assumptions of Theorem 1.1 there
exists 0 < λ0 < 1 such that for any L > 0 and λ ∈ [λ0 , 1[ we have
uc
ρλ,o (ξ, η) L λn , ∀ ξ 6= η ∈ ∂L,o
G
and
ρ̄λ,o (ξ, η) L λn , ∀ ξ 6= η ∈ Λuc
L,o G,
where n = d(o, [ξ, η]).
1.2. Historical remarks and motivations. We provide here a short history of
the study of the Hausdorff dimension of the limit set of various convergence actions:
Kleinian, hyperbolic and relatively hyperbolic.
The identification of the Hausdorff dimension with the critical exponent of Poincaré
series was first established by S. Patterson [21]. He introduced a probability measure on the limit set of the convex-cocompact Fuchsian groups, and proved that up
to a constant it is equal to the Hausdorff measure. D. Sullivan generalized this result and constructed such measures (called since then Patterson-Sullivan measures)
on the limit sets of geometrically finite Kleinian groups acting on the hyperbolic
space Hn of dimension n [25]. To finish the discussion of the case of Kleinian
groups, we note the result of C. Bishop and P. Jones who proved in [1] that for
a non-elementary Kleinian group acting on the hyperbolic 3-space the Hausdorff
dimension of the conical limit set is equal to the critical exponent of the Poincaré
series (compare with our Theorems 1.1 and 1.2). The latter results were generalized by F. Paulin [22] to discrete groups of isometries of Riemannian manifolds of
strictly negative curvature.
M. Coornaert has generalized the results of Patterson-Sullivan to the class of
word-hyperbolic groups [6]. In particular he proved that the Hausdorff dimension
of the (Gromov) boundary of such a group with respect to the visual metric is equal
to the critical exponent of the Poincaré series.
A natural question arises whether Coornaert’s theorem holds for the class of
relatively hyperbolic groups. However it was shown by M. Burger and S. Mozes
that if G is a closed subgroup of the isometry group of a CAT(−1) space X and the
parabolic subgroups of G are not amenable then the critical exponent is infinite [4,
Proposition 1.6]. Such an example of a relatively hyperbolic group whose parabolic
subgroups contain non-cyclic free subgroups was constructed by D. Gaboriau and
F. Paulin [10, Example 1, p. 189]. By [22] it then follows that the Hausdorff
dimension of the limit set for the action of such a group with respect to the visual
metric is infinite too. So in order to generalize Coornaert’s theorem to the class of
relatively hyperbolic groups one must replace the visual metric by a different one.
The Floyd metric obtained by a rescaling procedure of the word metric is a natural candidat as it extends to the Floyd compactification of a group. Furthermore by
a theorem of V. Gerasimov there exists an equivariant and continuous map from the
Floyd boundary ∂λ G to the limit set of any relatively hyperbolic action of G [12].
In particular, if G is hyperbolic, the Floyd and Gromov boundaries are bilipschitz
equivalent for some exponential Floyd function.
M. Bourdon has observed (private communication) that the Hausdorff dimension
of the Floyd boundary of a relatively hyperbolic group, calculated with respect to
the Floyd metric obtained with the exponential rescaling function λn (λ ∈ (0, 1)) is
δG,S
always upper bounded by − log
λ (cf. Lemma 4.1). However the question whether it
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS
5
admits a lower strictly positive bound, which is equal to the same constant remained
open. This was our first motivation giving rise to Theorem 1.1. Theorem 1.2 is
then obtained by transferring the Floyd metric from the Floyd boundary ∂λ G to
the limit set ΛG of the geometrically finite action using the above Gerasimov’s map.
The lower bound estimate for the Hausdorff dimension in Theorems 1.1 and
1.2 follow from Theorem 1.4 providing the approximation of the boundary points
by Ahlfors regular subsets Xi . These subsets entirely consist of uniformly conical
points which are the space of ends of subtrees of the Cayley graph of G. Note
that the idea of such an approximation by trees is quite standard in both settings:
hyperbolic (see e.g. [18, 6.1]) or Kleinian (see [1]). However these constructions of
trees essentially use the hyperbolicity of the ambiant space. The latter property
is not true for a relatively hyperbolic group: the Cayley graph is not in general
hyperbolic and the relative Cayley graph is hyperbolic but the action on the set of
vertices is not proper. The approximating trees constructed in the paper admits
certain periodicity allowing us to obtain a Patterson-Sullivan measure µ on Xi also
having periodic properties. Theorem 1.4 then shows that these measures converge to
the Hausdorff measure on a subset of uniformly conical points and whose dimension
coincides with the full Hausdorff dimension of the ambiant space.
Acknowledgments. The authors are deeply grateful to Marc Bourdon for several
discussions motivating our initial interest to the subject of the paper, we owe the
proof of Lemma 4.1 to him as well as the references to the papers [4] and [10].
We also thank Misha Kapovich who indicated us that the bilipschitz equivalence
between the Floyd and the visual metrics needs to be justified (see the Remark
above).
During the work on this paper the first author was partially supported by the
ANR grant DiscGroup (BLAN 2011 BS01 013 04). He is also very grateful to
the Max-Planck-Institut für Mathematik in Bonn for the hospitality and support
during his research stay at the Institute where a part of the work was done.
The second author is grateful to the CNRS for providing him a research fellowship
at the University of Lille 1, and was partially supported by the ERC starting grant
GA 257110 RaWG when he was a postdoc in Orsay.
2. Preliminaries
2.1. Notations and Conventions. Let (Y, d) be a geodesic metric space. Given
a subset X and a number r ≥ 0, let Nr (X) = {y ∈ Y : d(y, X) ≤ r}. For x ∈ Y
denote B(x, r) = Nr ({x}). Sometimes, we will write Bd (x, r) to emphasize the
metric d.
Given a point y ∈ Y and a subset X ⊂ Y , let ProjX (y) be the set of points x
in X such that d(y, x) = d(y, X). The projection of a subset A ⊂ Y to X is then
ProjX (A) = ∪a∈A ProjX (a).
We always consider a rectifiable path α in Y with arc-length parametrization.
Denote by `(α) the length of α, and by α− , α+ the initial and terminal points of α
respectively. Let x, y ∈ α be two points which are given by parametrization. Then
denote by [x, y]α the parametrized subpath of α going from x to y. We also denote
by [x, y] a choice of a geodesic in Y between x, y ∈ Y .
A path α is called a c-quasi-geodesic for c ≥ 1 if the following holds
`(β) ≤ c · d(β− , β+ ) + c
6
LEONID POTYAGAILO AND WEN-YUAN YANG
for any rectifiable subpath β of α.
Let α, β be two paths in Y . Denote by α · β (or simply αβ) the concatenated
path provided that α+ = β− .
A path α going from α− to α+ induces a first-last order as follows. Given a
property (P), a point z on α is called the first point satisfying (P) if z is among
the points w on α with the property (P) such that `([α− , w]α ) is minimal. The last
point satisfying (P) is defined in a similar way (replacing [α− , w]α by [w, α+ ]α ).
Let f, g be real-valued functions with domain understood in the context. Then
f ≺ci g means that there is a constant C > 0 depending on parameters ci such that
f < Cg, and ci is defined similarly. We use the symbol ci if both inequalities
are true. For simplicity, we omit ci if they are some universal constants.
Denote by k · k the diameter of a set in a metric space. Recall the notion of
Hausdorff measures in a metric space.
Definition 2.1. Let X be a subset in a metric space (Y, d). Given numbers , s ≥ 0,
define
∞
X
[
s
Hs (X) = inf{
kUi k : X ⊂
Ui , Ui ⊂ Y, kUi k ≤ }.
i=1
Define Hs (X) = lim Hs (X), the s-dimensional Hausdorff measure of X. The
→0
Hausdorff dimension of X is defined as follows,
Hdimd (X) = inf{s ≥ 0 : Hs (X) = 0} = sup{s ≥ 0 : Hs (X) = ∞}.
By convention, set inf ∅ = sup{s ∈ R≥0 } = ∞. Thus, Hdimd X ∈ [0, ∞]. Note that
Hs (X) may be zero for s = Hdimd X.
2.2. Floyd boundary and relative hyperbolicity. Let G be a group with a
finite generating set S. Assume that 1 ∈
/ S and S = S −1 . Let G (G, S) be the
Cayley graph of G with respect to S. Denote by dS (or simply by d if there is no
ambiguity) the word metric on G (G, S).
Fix 0 < λ < 1 and a basepoint o ∈ G. We define a Floyd metric ρλ,o as follows.
The Floyd length lλ,o (e) of an edge e in G (G, S) is λn , where n = d(o, e). The Floyd
length lλ,o (γ) of a path γ is the sum of Floyd lengths of its edges. This induces a
length metric ρλ,o on G (G, S), which is the infimum of Floyd lengths of all possible
paths between two points.
Let Gλ be the Cauchy completion of G with respect to ρλ,o . The complement
∂λ G of G (G, S) in Gλ is called Floyd boundary of G. The ∂λ G is called non-trivial
if ]∂λ G > 2. We refer the reader to [9], [12], [13], [19] for more details.
By construction, the following equivariant property holds
(1)
ρλ,o (x, y) = ρλ,go (gx, gy)
for any g ∈ G. The Floyd metrics with different basepoints are related by a biLipschitz inequality:
(2)
0
λd(o,o ) ≤
0
ρλ,o (x, y)
≤ λ−d(o,o )
ρλ,o0 (x, y)
for any two points o, o0 ∈ G.
We now recapitulate few standard definitions concerning geometrically finite
convergence actions which will be often used further.
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS
7
Definition 2.2. Let X be a compact metrizable space on which G admits a minimal
and non-trivial convergence action by homeomorphisms.
(1) A point ξ ∈ X is called conical if there exists a sequence of elements gn ∈ G
(n ≥ 1) such that the closure of {gn (ξ, η) : n ≥ 1} in X 2 is disjoint from
the diagonal ∆(X 2 ) = {(x, x) : x ∈ X} for any η ∈ X \ ξ.
−1
If, in addition, there exists L > 0 such that d(1, gn gn+1
) ≤ L, then ξ
is called L-uniformly conical (or uniformly conical if the constant L is not
important).
(2) A point ξ ∈ X is called bounded parabolic if the stabilizer Gξ of ξ in G is
infinite, and acts properly and co-compactly on X \ ξ. The subgroup Gξ is
called maximal parabolic.
(3) A convergence group action of G on X is called geometrically finite if every
limit point ξ ∈ X is either a conical point or a bounded parabolic point.
As it was mentioned in the Introduction a pair (G, P) is relatively hyperbolic if
G admits a geometrically finite group action on a compact metrizable space X such
that P coincides with the collection of maximal parabolic subgroups (peripheral
system). Using the relative Cayley graph one can construct the limit set ΛG of the
action with the boundary of this graph [3]. We will often call Bowditch boundary
the limit set ΛG of a geometrically finite action. Bowditch proved that if G is
finitely generated then ΛG up to an equivariant homeomorphism depends only on
the pair (G, P) [2]. We also note the same result still holds in general case when G
is not finitely generated [16, Corollary 6.1.e].
The following result establishes the following universal pullback property of the
Floyd boundary.
Proposition 2.3. [12] Suppose (G, P) is a relatively hyperbolic pair. Then there exists
0 < λ0 < 1 such that for any λ ∈ [λ0 , 1) there exists a continuous G-equivariant
surjective map (called Floyd map):
Fλ : ∂λ G → ΛG.
Let Gp be the stabilizer of a parabolic point p ∈ X for the action G y X = ΛG.
Denote by Λ∂λ G (Gp ) and ∂λ Gp the limit set of Gp for its action on the Floyd
boundary ∂λ G of G, and the Floyd boundary of Gp respectively. The following
result precisely describes the kernel of the Floyd map.
Proposition 2.4. [13] Under the assumption of Proposition 2.3, the following holds
Fλ−1 (p) = Λ∂λ G (Gp ) = ∂λ Gp
for any parabolic point p in ΛG. Moreover, Fλ−1 (p) consists of one point if p is a
conical point.
We equip ΛG with a shortcut metric as follows: let
ω = {(η, ξ) ∈ ∂λ G × ∂λ G : Fλ (ξ) = Fλ (η)}
be the relation on ∂λ G given by the Floyd map Fλ : ∂λ G → ΛG. For any ξ, η ∈ Gλ ,
define a pseudo-distance ρ̃λ,o (ξ, η) on Gλ to be
(3)
n
X
ρ̃λ,o (ξ, η) = inf {
ρλ,o (ξi , ηi ) : (ηi , ξi+1 ) ∈ ω, 1 ≤ i < n, ξ1 = ξ, ηn = η}.
n≥1
i=1
8
LEONID POTYAGAILO AND WEN-YUAN YANG
We have
(4)
∀ξ, η ∈ Gλ : ρ̃λ,o (ξ, η) ≤ ρλ,o (ξ, η),
and it is a maximal pseudo-metric on Gλ ×Gλ satisfying this inequality. It is shown
g := ΛG t G (G, S) (called attractor sum) is compact. The
in [12] that the space ΛG
g is convergence such that its restriction on G (G, S) is the identity
action G y ΛG
and on ΛG it coincides with the initial action. Furthermore the Floyd map Fλ
extends to an equivariant continuous map (denoted by the same symbol):
g
Fλ : Gλ → ΛG
such that Fλ |G ≡ id. Pushing forward ρ̃λ,o with Fλ , we obtain a shortcut pseudog:
metric on ΛG
g : ρ̄λ,o (x, y) = ρ̃λ,o (F −1 (x), F −1 (y)),
(5)
∀x, y ∈ ΛG
λ
λ
g (see [13, Section 3] for details). By the
which turns out to be a real metric on ΛG
above construction, one can easily see that the shortcut metrics ρ̄λ,o satisfy the
properties (1) and (2) too.
Convention 2.5. Since now on we will always suppose that λ ∈ [λ0 , 1) where λ0 is
given by Proposition 2.3. We omit the index λ in lλ,o , ρλ,o and ρ̄λ,o if λ is given in
the context.
Finally, we recall the following Visibility Lemma.
Lemma 2.6 (Visibility lemma). [19] There is a function ϕ : R≥0 → R≥0 such that for
any v ∈ G and any geodesic γ in G (G, S), we have if lv (γ) ≥ κ, then d(v, γ) ≤ ϕ(κ).
Remark. The same result is valid for quasi-geodesics or more general Θ-geodesics
where Θ : N → G is a polynomial distortion function [13, Lemma 5.1].
2.3. Floyd geodesics. In this subsection, we provide a few basic tools to study
Floyd geodesics.
We say that a path α : Z → G (G, S) ends at ξ ∈ ∂λ G if ξ = limn→∞ α(n).
Denote in this case α+ = ξ, and α− = limn→−∞ α(n). It follows from Lemma 2.6
that every geodesic ray ends at a point of the Floyd boundary. Moreover, Gλ is a
geodesic metric space and is a visual boundary: any two distinct points ξ, η ∈ Gλ
are connected by a bi-infinite word geodesic belonging to the Cayley graph [13,
Proposition 2.4].
We note that a Floyd geodesic between ξ, η does not necessarily belong to the
graph (e.g. an example of such situation is given by the Floyd geodesic [n, +∞] ∪
[−∞, −n] between −n and n for the group Z + Z). A method to overcome this
problem was proposed in [14]. It consists in introducing a special type of paths
called tight paths (see Definition 5.1 in section 5) situating in the Cayley graph
which will approximate well the Floyd geodesics. To provide a certain development
of this method we will need the following preliminary statements.
Lemma 2.7. [14, Lemma 7.2] For any l > 0, there exists 0 < λ0 < 1 such that the
following property hold for any λ ∈ [λ0 , 1).
Let x, y ∈ G (G, S) such that d(x, y) ≤ l, and p be a path with α− = x such
that `(α) ≥ d(x, y) + 1. Then lλ,o (α) > ρλ,o (x, y). In particular, the ρλ,o -geodesic
between x, y is a geodesic in G (G, S)
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS
9
We consider the following shortening procedure introduced in [14]: consider
two points x, y ∈ Gλ , we take a sequence of paths γn in G (G, S) such that (γn )− →
x, (γn )+ → y and
lλ,o (γn ) → ρλ,o (x, y).
For every l > 0 we can choose λ0 ∈]0, 1[ such that γn is an l-local geodesic. Indeed,
if a segment between two points of γn at distance at most l is not a geodesic, then it
can be replaced by a geodesic. Applying this procedure several times, we obtain a
l-local geodesic, still denoted by γn , whose Floyd length is not increased by Lemma
2.7 (see Lemma 5.8 for more details).
The following lemma states that word geodesic rays are also Floyd and shortcut
geodesics.
Lemma 2.8. Let o ∈ G be a base point and γ be a geodesic ray with γ− = o. Then
for any v ∈ γ we have
lλ,o ([v, x]γ ) = ρ̄λ,o (v, y),
and
lλ,o ([v, x]γ ) = ρλ,o (v, x),
where x = γ+ ∈ ∂λ G and y = F (x) ∈ ΛG where F is the Floyd map given in
Proposition 2.3.
Proof. We only prove the result for the shortcut metric. In the case of Floyd
metric a straightforward calculation shows that the geodesic ray γ as well as every
its subray is also a ρλ,o -Floyd geodesic.
By definition (3) of a shortcut metric, for any n ∈ N, there exist pairs (ηi , ξi+1 ) ∈
ω where 1 ≤ i < m such that
ρ̄λ,o (v, y) ≥
X
ρλ,o (ξi , ηi ) −
1≤i≤m
1
,
2n
where ξ1 = v, ηm = x. Every geodesic ray [o, η1 ] is also a Floyd geodesic so we can
1
. It follows
choose η̃1 ∈ [o, η1 ] such that ρλ,o (η̃1 , η1 ) ≤ 2n
ρ̄λ,o (v, y) ≥ ρλ,o (v, η̃1 ) −
1
.
n
Choose w ∈ [v, y]γ such that d(v, w) = d(v, η̃1 ) = m. Then the following is true:
(6)
ρλ,o (v, η̃1 ) ≥ ρλ,o (v, w).
Indeed, connect v and η̃1 by a curve α. There exists a point u = α(t0 ) such
that d(v, u) = m and choose a sub-curve α0 = [v, u]α containing m edges. Since
γ is a word geodesic, for the k-th edge e ∈ α0 and the k-th edge e1 ∈ [v, w]γ we
have lλ,o (e) ≥ lλ,o (e1 ) (k ∈ {0, ..., m}). Then lλ,o (α) ≥ lλ,o (α0 ) ≥ ρλ,o (v, w). So (6)
follows.
10
LEONID POTYAGAILO AND WEN-YUAN YANG
x
η
1
η
1
u
w
α
v= ξ
1
o
We have
1
1
≥ ρλ,o (v, w) − ≥
2n
n
1
λn
− .
≥ lλ,o ([v, x]γ ) −
1−λ n
Passing to the limit we obtain ρ̄λ,o (v, y) ≥ lλ,o ([v, x]γ ) = ρλ,o (v, x). Since
ρλ,o (v, x) ≥ ρ̄λ,o (v, y), we conclude that lλ,o ([v, γ+ ]γ ) = ρ̄λ,o (v, γ+ ).
ρ̄λ,o (v, y) ≥ ρλ,o (v, η̃1 ) −
2.4. Transitional paths and uniformly conical points. In this subsection we
shall give a description of uniformly conical points in ΛG using the geometry of
Cayley graph.
Let (G, P) be a relatively hyperbolic pair. Denote P = {gP : g ∈ G, P ∈ P̃},
where P̃ is a maximal set of non-conjugate subgroups in P. Following [13] we call
the elements of P horospheres.
Definition 2.9. Fix , R > 0. Let γ be a path in G (G, S) and v ∈ γ a vertex. Given
X ∈ P, we say that v is (, R)-deep in X if
γ ∩ B(v, R) ⊂ N (X).
If v is not (, R)-deep in any X ∈ P, then v is called an (, R)-transition point of γ.
The following lemma together with Lemma 2.6 will be invoked several times.
Lemma 2.10.
(1) For any c ≥ 1, R > 0, there exists = (c), κ = κ(, R) > 0
such that for any c-quasi-geodesic γ and an (, R)-transitional point v in γ,
we have
ρv (γ− , γ+ ) ≥ ρ̄v (γ− , γ+ ) > κ.
(2) For any c ≥ 1, κ, > 0 there exists R = R(c, κ, ) > 0 such that for any
c-quasi-geodesic γ and a point v ∈ γ with ρ̄v (γ− , γ+ ) > κ, we have that v
is an (, R)-transitional point of γ.
Proof. Let us first prove (2). Suppose not, then ∃c ≥ 1, κ, > 0 : ∀n, ∃ c-quasigeodesics γn and vn ∈ γn such that vn is (, n)-deep and ρ̄vn ((γn )− , (γn )+ ) > κ.
Up to a normalization we may assume that vn = v = γn (0). Then γn (] − n, n[) ⊂
N (Xn ) for Xn ∈ P. By compactness of geodesics in the Tikhonoff topology, we
obtain a limit horocycle α such that α± = q and every part of α belongs to γn for
sufficiently large n (see [14, Prop. 5.2.3] for more details). Then the diameter of
∂(γn ∩ α) with respect to the distance ρv tends to 0. As γn are geodesics whose
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 11
all interior points are in the graph we must have ρ̄v ((γn )− , (γn )+ ) → 0 which is a
contradiction.
(1). By [13, Corollary 3.9] there exists a constant = (c) such that for every
X ∈ P any c-quasi-geodesic with endpoints in X lies in N (X) (all horospheres
are uniformly quasi-convex). For the constants c and = (c) the statement now
follows from [15, Corollary 5.10], following a similar argument as above.
We introduce a special class of paths, which plays an important role in the
present study.
Definition 2.11. Given , R, L > 0, a path γ in G (G, S) is called (, R, L)-transitional
(or simply transitional if the choice of the constants is not important) if for any
point v ∈ γ, there exists an (, R)-transitional point w ∈ γ such that `([v, w]γ ) ≤ L.
We say that an infinite path γ in G (G, S) is eventually (, R, L)-transitional if
there exists v ∈ γ such that [v, γ+ )γ is (, R, L)-transitional.
We fix the constant = (1) > 0 given by Lemma 2.10.1. The following lemma
characterizes uniformly conical points as the endpoints of transitional geodesic rays.
Lemma 2.12. Let (G, P) be a relatively hyperbolic pair. There exists R > 0 for
which tplhe following property is true:
a point ξ ∈ ΛG is uniformly conical if and only if some (or any) geodesic ray
ending at ξ is eventually an (, R, L)-transitional geodesic ray for some L > 0.
Proof of “ ⇒ ”. Since G acts geometrically finitely on ΛG, it follows from [26,
Theorem 1C] that there exists δ > 0 such that for any conical point ξ ∈ ΛG, there
exists a sequence (gn ) ⊂ G such that for all points η ∈ (ΛG ∪ G) \ ξ one has
ρ̄1 (gn ξ, gn η) > δ. Denote r0 := ϕ(δ/2), where ϕ is given by Lemma 2.6.
Assume that ξ is an L-uniformly conical point for some L > 0. Let γ = [γ− , ξ[
be a geodesic ray ending at ξ and (gn ) ⊂ G be the above sequence taken for the
−1
pair (γ− , ξ). Then ρ̄1 (gn ξ, gn γ− ) = ρ̄gn−1 (ξ, γ− ) > δ/2 and d(1, gn gn+1
) ≤ L for
−1
all n ≥ 1. By Lemma 2.6, γ ∩ B(gn , r0 ) 6= ∅ for n ≥ 1. Let vn ∈ γ such that
d(vn , gn−1 ) < r0 . By the inequality (2) such that ρ̄vn (γ− , ξ) > κ where κ = λr0 · δ/2
is a uniform constant. Moreover, d(vn , vn+1 ) ≤ L + 2r0 .
Hence, Lemma 2.10.2 gives rise to a uniform constant R for which vn are all
(, R)-transitional for n ≥ 1.
Proof of “ ⇐ ”. Let γ be an (, R, L)-transitional geodesic ray at ξ = γ+ , and
vn (n ≥ 0) a sequence of (, R)-transitional points in γ such that d(vn , vn+1 ) ≤ L
and vn → ξ. Then ρ̄vn (γ− , ξ) ≥ κ, where κ > 0 is given by Lemma 2.10.1. Denote
gn := vn−1 . Then ρ̄1 (gn γ− , gn ξ) ≥ κ. In other words, {(gn γ− , gn ξ)} lies outside a
uniform neighborhood of the diagonal ∆(ΛG2 ).
−1
Since the action is convergence the point ξ is conical. As d(1, gn gn+1
) ≤ L it is
uniformly conical.
Remarks.
(1) The proof of the “ ⇐ ” direction equally applies to a conical
point in Floyd boundary ∂λ G without assuming the geometrical finiteness
of the action.
(2) The existence of the uniform constant δ > 0 which measures the size of
a compact fundamental set for the co-compact action of G on the set of
distinct pairs was only used to prove the implication “ ⇒ ” (in order to
get a uniform constant R). The existence of such a constant implies that
the action of G on a metrizable space ΛG is 2-cocompact; the converse
12
LEONID POTYAGAILO AND WEN-YUAN YANG
statement that a 2-cocompact and non-elementary convergence action is
geometrically finite is shown in [11], and its proof does not request the
metrisability of the space X = ΛG.
(3) As a corollary we see that for each L > 0 the set of L-uniformly conical
points is G-invariant, although this is not clear at all from the dynamical
definition.
Corollary 2.13. Let = (1) > 0 given by Lemma 2.10.1. For any R, L > 0, an
(, R, L)-transitional geodesic ray ends at a uniformly conical point ξ ∈ ∂λ G.
As another consequence of the proof, we have the following result.
Corollary 2.14. Let G y X be a geometrically finite action. Then there exists a
constant L > 0 such that for any conical point ξ ∈ X there is a sequence of elements
gn ∈ G such that for any geodesic γ ending at ξ, we have
[v, ξ[γ ⊂ ∪n≥1 B(gn , L)
for some v ∈ γ.
Remark. In the setting of Kleinian groups, this property is used to define uniformly conical points, cf. [24]. Here we do not need to assume that G acts geometrically finitely on ∂λ G. Also the corollary holds for “quasi-geodesics” instead
of “geodesics”.
We setup some notations for future discussions about uniformly conical points.
Let , R be given by Lemma 2.12. Denote by Λuc
L G the set of uniformly conical
points ξ ∈ ΛG such that there exists an (, R, L)-transitional geodesic ray γ ending
at ξ. It is obvious that Λuc
L G is a G-invariant set.
Fixing a basepoint o ∈ G, denote by Λuc
L,o G the set of all uniformly conical points
ξ ∈ Λuc
L G where a geodesic γ between o and ξ is (, R, L)-transitional.
uc
uc
Clearly, G·Λuc
L,o G = ΛL G. Thus, the set ΛL,o G can be thought as a fundamental
uc
domain for the action of G on the set ΛL G.
uc
G and ∂Luc G on the
Similarly, we define the set of uniformly conical points ∂L,o
Floyd boundary ∂λ G. By Proposition 2.4, there exists one-to-one correspondence
uc
between Λuc
L G and ∂L G.
2.5. Contracting property. Recall that k · k denotes the diameter of a set in a
metric space.
Definition 2.15. For c ≥ 1, a subset X is called c-contracting in a metric space Y
if there exists µc , Dc > 0 such that the following holds
(7)
kProjX (γ)k < Dc
for any c-quasi-geodesic γ in Y with Nµc (X) ∩ γ = ∅.
A collection of c-contracting subsets is referred to as a c-contracting system if
µc , Dc depends only on c.
A system X has a bounded intersection property if for any > 0 there exists
R = R() > 0 such that
kN (X) ∩ N (X 0 )k < R
for any two distinct X, X 0 ∈ X.
In what follows, our discussion applies to the Cayley graph of a relatively hyperbolic group (G, P) with a finite generating set S. In particular, we are interested
in the contracting system with bounded intersection given by the following lemma.
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 13
Lemma 2.16. [13] Let P = {gP : g ∈ G, P ∈ P̃}, where P̃ is a complete set of
conjugacy representatives in P. There exists R : R>0 → R>0 such that the collection
P is a c-contracting system with the R-bounded intersection for each c ≥ 1.
Proof. The contracting property is proven in [13, Proposition 8.5], and the bounded
intersection is in [13, Corollary 5.7].
The following lemma will be often used further on.
Lemma 2.17. Let P be the collection of horospheres in (2.16). For any c ≥ 1, there
exist c = (c) > 0 such that for every c-quasi-geodesic γ in G (G, S) and ≥ c we
have:
∀ R ≥ 0, ∃ L = L(, R) > 0 such that the condition max{d(γ− , X), d(γ+ , X)} <
, for some X ∈ P, implies that every point z ∈ γ satisfying d(z, γ− ), d(z, γ+ ) > L
is (c , R)-deep in X.
Proof. The result is proved in [27, Lemma 2.8] for geodesics. We provide below a
proof to precise the choice of the constants.
By Lemma 2.16, let µc , Dc be the constants such that for any X ∈ P, for any cquasi-geodesic outside Nµc (X), the diameter of its projection to X is upper bounded
by Dc .
Set c := c(2µc + Dc ) + c. If a c-quasi-geodesic has two endpoints in Nµc (X) for
X ∈ P, then it lies in Nc (X). Indeed, if x, y ∈ γ satisfy
max{d(x, X), d(y, X)} ≤ µc
and ]x, y[γ ∩Nµc (X) = ∅, then by Lemma 2.16, d(x, y) ≤ 2µc + Dc . Since γ is
c-quasi-geodesic we have `([x, y]γ ) ≤ c and [x, y]γ ⊂ Nc (X).
Set L = c(2 + Dc ) + c + R for ≥ c . We first claim γ ∩ Nµc (X) 6= ∅. Otherwise,
we obtain using projection the following
2L ≤ `(γ) ≤ cd(γ− , γ+ ) + c ≤ c(2 + Dc ) + c.
This gives a contradiction by the choice of L. Thus, there exist the entry point x
and the exit point y of γ in Nµc (X).
By the same argument one obtains
max{`([γ− , x]γ ), `([y, γ+ ]γ )} ≤ c( + µc + Dc ) + c < L.
Since min{d(z, γ− ), d(z, γ+ )} > L, we have z ∈ [x, y]γ . Then we obtain
min{d(x, z), d(z, y)} ≥ L − ( + µc + Dc ) > R.
By definition of c , we have [x, z]γ ⊂ Nc (X) and [z, y]γ ⊂ Nc (X). So z is
(c , R)-deep in X.
Remark. By the proof, we actually have c > µc , where µc is uniform for every
X ∈ P by Lemma 2.16.
In what follows, we take constants , R as in Convention 2.18.
Convention 2.18 (About c , Rc ). When talking about (c , Rc , L)-transitional c-quasigeodesics, or (c , Rc )-transitional and (c , Rc )-deep points in a c-quasi-geodesic, we
assume without explicitely specifying the quantifiers:
(1) c = (c) > µc to satisfy Lemmas 2.10 and 2.17, where µc is given by
Definition 2.15.
(2) Rc > R(), where R is given by Lemma 2.16.
14
LEONID POTYAGAILO AND WEN-YUAN YANG
Besides the peripheral cosets (horospheres), transitional quasi-geodesics provide
another source of contracting subsets.
Lemma 2.19 (Transitional geodesic is contracting). For any L ≥ 0, any (, R, L)transitional geodesic γ is 1-contracting.
Remark. The same argument also applies (with natural changes for the constants)
to show that if γ is c-quasi-geodesic then it is contracting for any c ≥ 1. For our
purposes, we only need to consider the case when c = 1.
Proof. Let κ = κ(, R) given by Lemma 2.10 and φ given by Lemma 2.6. By Lemma
2.6, there exists D0 = φ(κ/4) > 0 such that for any v ∈ G, a geodesic segment
outside the ball B(v, D0 ) has lv -Floyd length less than κ/4.
Let D = 2(L + 2D0 + 1) and µ = φ(κ/2). Let β be a geodesic such that
β ∩ Nµ (γ) = ∅. Let x, y ∈ Projγ (β) such that d(x, y) = kProjγ (β)k. We are going
to prove that d(x, y) ≤ D. Suppose by contradiction that d(x, y) > D.
Assume that x, y are projection points of x̃, ỹ ∈ β respectively. Observe that
(8)
2d(z, [x, x̃]) ≥ d(z, x), 2d(z, [y, ỹ]) ≥ d(z, y),
for any z ∈ [x, y]γ . We only prove the first inequality; the second one is completely
analogous. Let m ∈ [x, x̃] such that d(z, m) = d(z, [x, x̃]). Note that d(m, z) +
d(m, x̃) ≥ d(x, x̃) by the shortest point property. Since d(x, x̃) = d(x, m) + d(m, x̃)
we obtain d(m, z) ≥ d(x, m). Then d(z, x) ≤ d(z, m) + d(m, x) ≤ 2d(z, m) implying
(8).
~x
~y
β
m
m
D0
x
z
v
γ
z
y
Since d(x, y) > D, there exists z ∈ [x, y]γ such that
min{d(z, x), d(z, y)) > D/2 = L + 2D0 + 1.
Since γ is (, R, L)-transitional, one of the intervals [x, z]γ or [z, y]γ contains an
(, R)-transitional point v such that min{d(x, v), d(y, v)} > 2D0 . Hence by (8),
min{d(v, [x, x̃]), d(v, [y, ỹ]))} > D0 . By the choice of D0 = φ(κ/4), we have
max{ρv (x, x̃), ρv (y, ỹ)} < κ/4.
From the other hand, v is (, R)-transitional, so ρv (x, y) ≥ κ by Lemma 2.10. Hence,
ρv (x̃, ỹ) > κ/2 and thus d(v, β) ≤ µ which is impossible.
For a c-quasigeodesic we denote by c = (c), Rc = R(c ) any numbers satisfying Convention 2.18 (in particular 1 and R1 correspond to geodesics). In the
following Proposition we will establish a ”thinness” of a triangle whose two sides
are transitional geodesics.
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 15
Proposition 2.20 (Transitional triangle is thin). For any L, c > 0 and the there exist
constants D = D(c), M = M (L, c), L0 = L0 (L, c) > 0 with the following properties.
Let α1 , α2 be (1 , R1 , L)-transitional geodesic rays issuing at o and ending at
ξ 6= η ∈ ΛG respectively. Then for any c-quasi-geodesic γ with γ− ∈ α1 , γ+ ∈ α2 ,
the following holds.
(1) γ is (c , Rc , L0 )-transitional.
(2) If the length of γ is sufficiently large then there exists an (c , Rc )-transitional
point z ∈ γ such that d(z, α1 ∪ α2 ) ≤ D and d(z, αi ) ≤ M for i = 1, 2.
(3) Let d(o, [ξ, η]) denote the distance from o to a geodesic between ξ, η. If
min{d(γ− , o), d(γ+ , o)} 0. Then |d(o, [ξ, η]) − d(o, γ)| ≤ M .
Remark. Note that in (2) D is a uniform constant not depending on L, this will
play a crucial role in establishing Lemma 5.13 below.
Proof. Let κ = κ(c , Rc ) given by Lemma 2.10 and D = φ(κ/2), where φ is given
by Lemma 2.6. The constant L0 will be computed below.
(1) Given a point x in γ, assume that x is (, Rc )-deep in some X ∈ P. Let
x− , x+ be the entry and exit points of γ in Nc (X) respectively.
Observe first that x− , x+ are (c , Rc )-transitional in γ. Indeed, if not, there
exists Y ∈ P such that x− is (c , Rc )-deep in Y. Then Y 6= X by the choice of x− as
the entry point of γ in Nc (X). Since d(x, x− ) ≥ Rc , we have kN (X) ∩ N (Y )k ≥
Rc > R(c ) by Convention 2.18. This contradicts to Lemma 2.16.
To find a constant L0 we will check when the opposite inequality:
(9)
min{`([x, x− ]γ ), `([x, x+ ]γ )} > L0
is not valid. We have `([x− , x+ ]γ ) ≥ 2L0 . Since x− , x+ are (c , Rc )-transitional, by
Lemma 2.10,
min{ρx− (γ− , γ+ ), ρx+ (γ− , γ+ )} > κ.
By the triangle inequality
κ
,
2
and the same for ρx+ . Then max{d(x− , α1 ∪ α2 ), d(x+ , α1 ∪ α2 )} ≤ D = φ(κ/2).
For concreteness consider the case that
max{ρx− (γ− , o), ρx− (o, γ+ )} ≥
(10)
d(x− , α1 ), d(x+ , α2 ) ≤ D;
the other cases are similar and even easier.
Project x− , x+ to x0− , x0+ ∈ X so that d(x− , x0− ), d(x+ , x0+ ) ≤ c . So
d(x0− , α1 ), d(x0+ , α2 ) ≤ c + D
and ND+c (X) ∩ αi 6= ∅ (i = 1, 2).
Let w ∈ X be a projection point of o to X. We claim that
(11)
d(w, αi ) ≤ D2 := max{D + c + D1 , µ1 + D1 } (i = 1, 2),
where µ1 , D1 > 0 are given for 1-contracting X ∈ P such that (7) holds.
Indeed if, first, o ∈ Nµ1 (X) then there is nothing to prove. If not, there are two
more cases: if αi ∩Nµ1 (X) = ∅, then by the contracting property we have d(w, αi ) ≤
D + c + D1 ; otherwise the projection on X of the maximal connected subcurve of
γ, situated outside of Nµ1 (X) and containing o, gives d(w, αi ) ≤ µ1 + D1 . So (11)
follows.
16
LEONID POTYAGAILO AND WEN-YUAN YANG
Let L0 = L(c + D + D2 , R1 + L) given by Lemma 2.17. Set
L0 = 2c(D2 + D + L0 + 2c ) + c2 .
(12)
Since γ is a c-quasi-geodesic, we have
d(x0− , x0+ ) ≥ `([x− , x+ ]γ )/c − c − 2c
≥ 2L0 /c − c − 2c
≥ 4(c + D + D2 + L0 ).
Since max{d(x0− , α1 ), d(x0+ , α2 )} ≤ c + D, we obtain from (11)
kα1 ∩ Nc +D+D2 (X)k ≥ d(x0− , w), kα2 ∩ Nc +D+D2 (X)k ≥ d(x0+ , w).
We have d(x0− , w) + d(x0+ , w) ≥ d(x0− , x0+ ) ≥ 4(c + D + D2 + L0 ). Thus
max kαi ∩ Nc +D+D2 (X)k ≥ 2(c + D + D2 + L0 ).
i=1,2
Hence, αi contains a subcurve of length at least 2L0 such that its endpoints lie
in Nc +D+D2 (X). By the choice of L0 and Lemma 2.17, αi contains an (1 , R1 +L)deep point in X. This gives a contradiction, as αi is (1 , R1 , L)-transitional. So for
the value of L0 chosen in (12) the inequality (9) is not valid. The statement (1) is
proved.
(2) By the statement (1) γ is (c , Rc , L0 )-transitional. Lemma 2.19 implies that γ
is contracting. By the projection argument (used to prove (11)) we have a constant
D3 = D3 (c , Rc , L0 ) > 0 such that for any projection point v of o to γ we have
d(v, αi ) ≤ D3 for i = 1, 2.
Remark. We need a new constant D3 (and not D2 used above) since we project
now on γ and not on a horosphere.
Recall that D = φ(κ/2). By Lemma 2.6, for any z ∈ G, a geodesic segment
outside B(z, D) has lz -Floyd length less than κ/2.
The curve γ is quasigeodesic and its length is sufficently large. So by continuity
of the distance function d(v, x) (x ∈ γ) we find a point z 0 such that D + D3 + L0 ≤
d(v, z 0 ) ≤ D + D3 + L0 + 1. Since γ is (c , Rc , L0 )-transitional by Definition 2.11
there exists an (c , Rc )-transitional point z ∈ γ for which d(z 0 , z) ≤ l([z, z 0 ]γ ) ≤ L0 .
We obtain
(13)
D + D3 ≤ d(v, z) ≤ D + D3 + 2L0 + 1.
Then
d(z, αi ) ≤ d(z, v) + d(v, αi ) ≤ M,
where M = 2L0 + 2D3 + D + 1.
To prove the first claim of (2) assume for definiteness that z ∈ [v, γ+ ]. Lemma
2.10 yields ρz (v, γ+ ) ≥ κ.
Let z2 ∈ α2 such that d(v, z2 ) ≤ D3 . We have d(z, [v, z2 ]) ≥ d(z, v) − d(z2 , v) ≥
D + D3 − D3 = D. By Lemma 2.6 ρz (v, z2 ) < κ/2 and so ρz (z2 , γ+ ) ≥ ρz (v, γ+ ) −
ρz (v, z2 ) ≥ κ/2. Lemma 2.6 gives
d(z, α2 ) ≤ D.
The statement (2) is proved.
(3) Since ξ, η are distinct, by Lemma 2.6, there exists n0 = n0 (ξ, η), r = r(ξ, η) >
0 such that if
min{d(γ− , o), d(γ+ , o)} > n0 ,
then d(o, γ) ≤ r. In the proof of the statement (2), we projected o to a point v in
γ, and found an (, R)-transitional point z ∈ γ such that d(v, z) ≤ M .
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 17
Since d(o, z) ≤ M + r (and these constants do not depend on γ) up to increasing
n0 , by Lemma 2.6 we have
max{ρ̄z (ξ, γ− ), ρ̄z (η, γ+ )} ≤ κ/4.
The point z is (, R)-transitional, thus ρ̄z (γ− , γ+ ) ≥ κ, and so ρ̄z (ξ, η) ≥ κ/2.
Consequently d(z, [ξ, η]) ≤ D which yields:
d(o, [ξ, η]) ≤ d(o, z) + d(z, [ξ, η]) ≤ d(o, γ) + M + d(z, [ξ, η])
≤ d(o, γ) + M + D.
By symmetry, we obtain d(o, γ) ≤ d(o, [ξ, η]) + M + D.
Since D is a uniform constant not depending on L we put M := M +D. Then the
statements of (2) and (3) are both valid for the same constant M. The Proposition
is proved.
The claim (3) of the Proposition and Lemma 2.12 imply:
Corollary 2.21. Suppose (G, P) is a relatively hyperbolic pair. Then for any L > 0,
uc
there exists M = M (L) such that for any ξ, η ∈ Λuc
L,o G or ξ, η ∈ ∂L,o G, the distance
d(o, [ξ, η]) is comparable with the distance d(o, γ) where γ is a c-quasi-geodesic with
the endpoints on the corresponding geodesic rays converging to ξ and η..
3. Patterson-Sullivan measures on ends of a geodesic tree
In this section, we shall construct an iterated transitional tree having several
nice properties which will allow us to carry out the Patterson’s construction on this
tree. The space of ends of the tree equipped with the Patterson-Sullivan measure
will give rise to an Ahlfors regular subset of the boundary.
3.1. Iterated Transitional Trees. Let (G, P) be a relatively hyperbolic pair and
G (G, S) the Cayley graph of G with respect to S. The existence of large transitional
trees is established in [27, Theorem 5.9]. The main difference of the construction
below is that these trees will be equipped with certain periodicity. By this reason
we call them iterated transitional trees. We start by recalling several results from
[27].
Definition 3.1 (Partial Cone). For , R ≥ 0, the partial cone Ω,R (g) at g ∈ G is
the set of elements h ∈ G such that there exists a geodesic γ = [1, h] containing g
and one of the following holds.
(1) d(1, h) ≤ d(1, g) + 2R,
(2) γ contains an (, R)-transitional point v such that d(v, g) ≤ 2R.
For ∆ ≥ 0, n ≥ 0, define
A(g, n, ∆) = {h ∈ G : n − ∆ ≤ d(1, h) − d(1, g) < n + ∆},
for any g ∈ G. For simplicity we write A(n, ∆) := A(1, n, ∆). For r, , R, ∆ > 0,
define
Ω,R (g, n, ∆) = Ω,R (g) ∩ A(g, n, ∆),
for any g ∈ G, n ≥ 0.
For fixed , R > 0, two partial cones Ω,R (g), Ω,R (g 0 ) are of same type if
g 0 g −1 · Ω,R (g) = Ω,R (g 0 ).
By abuse of language, we say that g, g 0 have the same partial cone types.
The following result generalizes the result of Cannon [5] for hyperbolic groups.
18
LEONID POTYAGAILO AND WEN-YUAN YANG
Lemma 3.2 (Finiteness of partial cone types). [27, Lemma B.1] There exist , R0 > 0
such that for any R > R0 , there are at most M = M (, R) types among all (, R)partial cones {Ω,R (g) : g ∈ G}.
The following is a key technical result in [27, Lemma 5.8].
Lemma 3.3. There exist , R, ∆, θ, L0 > 0 with the following property.
For any L > L0 there exists a subset Ĝ of G such that
](Ω,R (g, L, ∆) ∩ Ĝ) > θ · exp(L · δG,S ), 1 ∈ Ĝ
(14)
for any g ∈ Ĝ.
Convention 3.4 (, R, ∆). Until the end of Section 3, the constants , R, ∆ > 0 are
given by Lemmas 3.2 and 3.3, and satisfy Convention 2.18.
The following terminology comes from the paper [1] which was certainly very
motivating for us.
Definition 3.5 (Iterated Tree Set). For given L > 0, an L-iterated tree set T in G
is a union of a sequence of sets Ti (i ≥ 0) in G defined inductively as follows.
Let T0 = {1}. Assume that Ti is defined for i ≥ 0. The children T (x) of x ∈ Ti
is a subset in Ω,R (x, L, ∆). Then Ti+1 is the union of children of all x ∈ Ti .
Recall that a subset Z of a metric space space (X, d) is called C-separated if the
distanced(z1 , z2 ) ≥ C for every pair of distinct points {z1 , z2 } ⊂ Z. The following
fact is elementary.
Lemma 3.6. Let (X, d) be a proper metric space on which a group G ⊂ Isom(X)
acts properly on X. For any orbit Go (o ∈ X) and C > 0 there exists a constant
θ = θ(Go, C) > 0 with the following property.
For any finite set Y in Go, there exists a C-separated subset Z ⊂ Y such that
]Z ≥ θ · ]Y .
Proof. Let Z be a maximal C-separated set in Y . We have Y ⊂ NC (Z). Since the
action of G on (X, d) is proper, any ball of radius C contains at most N points in
Go. The result follows for θ := 1/N .
An (, R, L)-transitional geodesic tree T rooted at o in G (G, S) is a tree subgraph
with a distinguished vertex x such that every branch in T originating at x is a
(, R, L)-transitional geodesic in G (G, S).
In order to obtain a useful theory of Patterson-Sullivan measures, certain symmetry on the iterated tree set is required. This is the content of the following.
Lemma 3.7 (Existence of iterated transitional trees). There exist constants L0 , C0 , t0 , n0 >
0 such that for L > L0 , C > C0 there are θ = θ(C) and L0 = L0 (L) and an iterated
tree set T parameterized by (, R, L) with the following properties:
(1)
(2)
(3)
(4)
(5)
x−1 T (x) = y −1 T (y) for any x ∈ Tt , y ∈ Tt+n0 and t ≥ t0 ,
x−1 T (x) = y −1 T (y) for any x, y ∈ Tt and t ≥ t0 .
]T (x) ≥ θ · exp(δG,S L) for any x ∈ T .
T (x) is C-separated for any x ∈ T .
there exists an (, R, L0 )-transitional geodesic tree T rooted at 1 in G (G, S)
such that the vertex set T 0 contains T , and lies in NL0 (T ).
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 19
Proof. Set L0 = ∆, and all other constants will be defined in the proof. We divide
the proof into 3 steps for the reader convenience.
Step 1: At this step we construct the iterated tree set T with properties (1-3).
The construction proceeds by an induction argument. Set T0 = {1} to start.
Let M be the number of (, R)-partial cone types in G given by Lemma 3.2, and
Ĝ the set by Lemma 3.3. Then there exists T1 ⊂ Ω,R (1, L, ∆) ∩ Ĝ such that every
element in T1 has the same partial cone type and the inequality (14) holds for g = 1
where the constant κ is divided by M . By Lemma 3.6, we can also require that
T1 is C-separated, where κ is further decreased and depends on C (given in Step 3
below).
Fix some x1 ∈ T1 . Up to dividing θ by M again, we choose Y to be a subset
of Ω,R (x1 , L, ∆) ∩ Ĝ such that the inequality (14) holds for Y and every element
in Y has the same partial cone type. By the same reason, we can choose Y to be
C-separated. Since all x ∈ T1 are of same type as x1 , we could define
T (x) := xx−1
1 Y ⊂ Ω,R (x, L, ∆).
Then all elements in the union T2 := ∪x∈T1 T (x) have the same partial cone types.
We note that Y is chosen to be contained in Ĝ, but T2 may not be in Ĝ.
We repeat the same argument to construct Ti for i ≥ 3, with a sequence of
divisions of κ. By construction, all elements in the constructed Ti are of the same
partial cone type. Since there are finitely many partial cone types, we obtain that
there are 1 ≤ t0 , n0 ≤ M so that x−1 T (x) = y −1 T (y) for any x ∈ Tt , y ∈ Tt+n0
and t ≥ t0 . This also implies that the division of θ stops after at most n0 times,
and thus θ in the inequality (14) can be chosen uniformly for all T (x) where x ∈ Ti
(i ≥ 1). The set T satisfies the properties (1 − 3).
Step 2: Using the iterated tree set T , we will now construct a geodesic graph
T.
Without loss of generality assume that κ < 1. The root of T is T0 = {1}. Assume
that Ti is defined for i ≥ 0 and for each terminal vertex x ∈ Ti , denote by γx the
geodesic [1, x] in Ti . We choose a geodesic [x, y] for each y ∈ T (x). Since Ti (x) is a
subset in Ω,R (x, L, ∆) ∩ Ĝ, we set
(15)
Ti+1 = ∪x∈Ti (∪y∈T (x) γx · [x, y]),
where γx ·[x, y] is a geodesic in G (G, S). Inductively, we get the limit T = limi→∞ Ti .
By construction, each geodesic ray originating at 1 is (, R, L0 )-transitional for L0 :=
L + 2R + ∆. By construction we have T ⊂ T 0 ⊂ NL0 (T ).
Step 3: We now prove that T is a geodesic tree rooted at 1 in G (G, S). Indeed, if
not, there exist two distinct geodesics α1 , α2 in T with the same endpoints x, w ∈ T
such that the length of α1 , α2 is minimal among all such choices. Assume that x is
closer to 1 than w. Consider two points yi ∈ αi ∩ T (x) for i = 1, 2. By the choice
of α1 , α2 , we have y1 6= y2 . Then by construction, d(yi , w) ≥ L − ∆ for i = 1, 2.
Moreover, there exists an (, R)-transitional point z1 ∈ α1 such that d(y1 , z1 ) ≤ 2R.
Let D0 = φ(κ), where φ is given by Lemma 2.6 and κ = κ(, R) by Lemma 2.10.
There exists z2 ∈ α2 such that d(z1 , z2 ) ≤ D0 and then d(y1 , z2 ) ≤ 2R+D0 . We can
choose ỹ2 ∈ α2 such that d(x, ỹ2 ) = d(x, y1 ). Hence d(z2 , ỹ2 ) = |d(x, z2 )−d(x, ỹ2 )| =
|d(x, z2 ) − d(x, y1 )| ≤ 2R + D0 . It follows that d(y1 , ỹ2 ) ≤ 2(2R + D0 ).
Since y1 , y2 lie in the annulus A(x, L, ∆), we get |d(x, y1 ) − d(x, y2 )| ≤ 2∆, and
then d(y2 , ỹ2 ) = |d(x, ỹ2 )−d(x, y2 )| ≤ 2∆. It follows that d(y1 , y2 ) ≤ 2(2R+D0 +∆).
20
LEONID POTYAGAILO AND WEN-YUAN YANG
Choosing now the constant C to be greater than
(16)
C0 := 2(2R + φ(κ/2) + ∆)
we obtain that T (x) is C-separated in Ω,R (x, L, ∆), and d(y1 , y2 ) ≥ C0 > 2(2R +
D0 + ∆) which is a contradiction. Thus, T is a rooted geodesic tree.
Remarks.
(1) By Lemma 2.12 the boundary of the tree T (in ΛG or in ∂λ G)
constructed above consists of uniformly conical points.
(2) The constant C0 in (16) is bigger than we really need in the above proof
(it is enough to replace ϕ(κ/2) by the smaller term D0 = ϕ(κ))) but we do
need such a constant in the next lemma.
In the next two lemmas, we shall derive more properties of the sets T and T
constructed in Lemma 3.7. To this end, we recall the notion of Poincaré series.
For a subset X ⊂ G and a point o ∈ G, set
X
ΘX (s, o) =
exp(−sd(o, g)), s ≥ 0.
g∈X
Define the critical exponent of ΘX (s, o) to be
(17)
δX,S = lim sup
n→∞
log ](B(o, n) ∩ X)
,
n
where S is a fixed finite symmetric generating set of G, and B(o, n) is the ball in
the word metric of radius n centered at o.
It is elementary fact that ΘX (s, o) converges for s > δX,S , and diverges for
s < δX,S .
Recall that the bilipschitz equivalence const between two functions means that
they are comparable up to a constant (see Section 2). We have the following.
Lemma 3.8. Under the same assumptions as in Lemma 3.7, we have
ΘT (s, x) L ΘT (s, y)
for any x, y ∈ T and s ≥ 0, whenever one of the series converges.
Proof. Let Ω(x) be a cone at x ∈ T , which is the union of y ∈ T such that the
unique geodesic [1, y] in the geodesic tree T contains x.
Claim. The Poincaré series of T is bilipschitz equivalent to that of any cone at a
vertex in T :
(18)
ΘT (s, 1) L ΘΩ(x) (s, x),
for any x ∈ T .
Proof of the Claim. It follows from Lemma 3.7.(1) that after a finite time t0 , the
set T is periodic with a fixed period n0 . So it is enough to show (18) for x ∈ T such
that t0 ≤ d(1, x) ≤ n0 + t0 . By Lemma 3.7.2, the cones based at points y ∈ Tt have
the same type where t = d(1, x). Thus the number an of points in T situated at
the distance n from 1 is at most C · bn,x . Here C is the number of elements in the
ball B(1, t), and b(n, x) is the number of elements of Ω(x) at the distance n from
x. The same argument works in the opposite sense. The Claim follows.
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 21
To complete the proof of the lemma, by (18), it suffices to establish the following
ΘT (s, 1) L ΘT \Ω(x) (s, x),
(19)
as 18 and 19 would imply ΘT (s, 1) L ΘT (s, x), ∀ x ∈ T.
For y ∈ T \ Ω(x), let o be the farrest point to 1 such that o ∈ T and [1, o] ⊂
[1, x] ∩ [1, y], where the geodesics [1, x] and [1, y] are in the geodesic tree T . The
point o will be referred to as the branch point of [1, x] and [1, y].
By Lemma 3.7, [1, x], [1, y] are (, R, L0 )-transitional geodesics, where L0 = L0 (L).
By Proposition 2.20, [x, y] is transitional and so is contracting by Lemma 2.19.
Claim. There exists a uniform constant D = D(L) > 0 such that d(o, [x, y]) ≤ D.
Proof of the Claim. Let z ∈ [x, y] be the projection of o to a geodesic [x, y] in the
Cayley graph G (G, S). By the contracting property of [x, y] it follows from the
inequality 11 that there exists D1 = D1 (, R, L0 ) such that
max{d(z, [o, x]), d(z, [o, y])} ≤ D1 .
So, let x1 ∈ [o, x], y1 ∈ [o, y] such that d(z, x1 ) ≤ D1 and d(z, y1 ) ≤ D1 .
Set d(o, z) = d, then
min{d(o, x1 ), d(o, y1 )} ≥ d − D1
(20)
Let w ∈ [o, x1 ] ∩ T (o) where T (o) ⊂ Ω,R (o, L, ∆). Then d(o, w) < L + ∆.
Furthermore since x ∈ T there exists an (, R)-transitional point x2 ∈ [o, x] such
that d(w, x2 ) ≤ 2R, and so d(o, x2 ) ≤ L + ∆ + 2R. Using (20) we deduce
(21)
d(x2 , [x1 , y1 ]) ≥ d(x1 , o) − d(x2 , o) − 2D1 ≥ K,
where K = d − 3D1 − L − ∆ − 2R.
We affirm that
K ≤ φ(κ/2),
(22)
where κ and φ are universal constants given by Lemmas 2.10 and 2.6. respectively.
Indeed, suppose (22) is not true, then d(x2 , [x1 , y1 ]) ≥ φ(κ/2). By Lemma 2.6 we
have ρx2 (x1 , y1 ) ≤ κ/2. Since x2 is transitional, Lemma 2.10 yields ρx2 (o, x1 ) ≥ κ.
It follows ρx2 (o, y1 ) ≥ κ/2, and thus ∃ x̃2 ∈ [o, y] : d(x2 , [o, y]) = d(x2 , x̃2 ) ≤
φ(κ/2).
z
x
y
x1
w"
w’
w
x2
~
x
2
o
1
y
1
22
LEONID POTYAGAILO AND WEN-YUAN YANG
Following the argument of Step (3) of Lemma 3.7 we choose a vertex w0 ∈
[o, y] such that d(o, w0 ) = d(o, w). Then we have d(x̃2 , w0 ) = |d(o, x̃2 ) − d(o, w)| ≤
d(x̃2 , w) ≤ 2R + φ(κ/2). Then d(w0 , w) ≤ d(w0 , x̃2 ) + d(x̃2 , w) ≤ 2(2R + φ(κ/2)).
Let w00 ∈ T (o) ∩ [o, y]. Since w ∈ Ω,R (o, L, ∆) we have d(w00 , w0 ) ≤ |d(o, w0 ) −
d(o, w00 )| ≤ 2∆. Indeed |(d(o, w0 ) = d(o, w)) − L| ≤ ∆ and also |d(o, w00 ) − L| ≤ ∆.
Therefore for the vertices w00 , w ∈ T (o) ⊂ T we have d(w, w00 ) ≤ 2(φ(κ/2) + R + ∆).
This is impossible by (16). The obtained contradiction implies that K ≤ φ(κ/2)
and by definition of K (see (21)), we have
d(o, z) = d ≤ D = D1 + φ(κ/2) + L + ∆ + 2R.
The claim is proved.
The second claim implies
d(o, x) + d(o, y) ≥ d(x, y) ≥ d(o, x) + d(o, y) − 2D.
(23)
Given o ∈ [1, x) ∩ T , we denote by Yo the set of elements y ∈ T \ Ω(x) such that
o ∈ [1, x] is the branch point of [1, y] and [1, x] in T . The argument of the first
Claim also yields
ΘΩ(o) (s, o) L ΘYo (s, o).
Then (18) and (23) imply
X
exp(−sd(x, y)) L exp(−sd(o, x)) · ΘT (s, 1),
y∈Yo
for every o ∈ [1, x) ∩ T . By construction of T in Lemma 3.7 the sequence of points
[1, x) ∩ T has the property that any two consecutive points has a distance between
L − ∆ and L + ∆. Summing up over all o ∈ [1, x) ∩ T , we get
X
X
exp(−sd(x, y)) L
exp(−sk) · ΘT (s, 1) L ΘT (s, 1),
y∈T \Ω(x)
0≤k<d(1,x)
which proves (19). The Lemma is proved.
Lemma 3.9. Under the same assumptions in Lemma 3.7, the Poincaré series ΘT (s, 1)
is divergent at s = δT,S . Furthermore, lim δT,S = δG,S .
L→∞
Proof. It is inspired by the proof of Proposition 4.1 in [8]. Consider the annulus
set in T ,
AT (g, n, 3∆0 ) := A(g, n, 3∆0 ) ∩ T,
where ∆0 := ∆ + L + 2R and n ≥ 0. Observe that there exists c > 0 such that
(24)
c−1 · ]AT (g 0 , n, 3∆0 ) ≤ ]AT (g, n, 3∆0 ) ≤ c · ]AT (g 0 , n, 3∆0 )
for any g, g 0 ∈ T and n ≥ 0. Indeed, this is a direct consequence of Lemma 3.7 that
T has certain periodicity. Moreover, we claim that
Claim. The following inequality holds
]AT (1, n + m, 3∆0 ) ≤ c · ]AT (1, n, 3∆0 ) · ]AT (1, m, 3∆0 ),
for n, m ≥ 0.
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 23
Proof of the Claim. For h ∈ AT (n + m, 3∆0 ), we connect 1 and h by a geodesic
[1, h] in T . Assume that d(1, h) = m + n + 3∆1 for some |∆1 | ≤ ∆0 . Let z ∈ [1, h]
such that d(1, z) = n + 3/2 · ∆1 . Note that z might not be in T . However, by
Lemma 3.7.4), there exists w ∈ T such that d(z, w) ≤ ∆ + L + 2R = ∆0 and then
d(w, h) ≤ m + 3∆0 . This implies that w ∈ A(1, n, 3∆0 ) and h ∈ AT (w, m, 3∆0 ).
The conclusion thus follows from (24).
Define an = c · ]AT (1, n, 3∆0 ). The above Claim implies that an+m ≤ an am .
log an
So the sequence (log an )n is subadditive. Then by Fekete Lemma lim
=
n→∞
n
n log a
o
P
n
inf
: n ≥ 1 . Since (an )n is non-decreasing we have an ≤ 0≤i≤n ai ≤ nan .
n
So
P
n log a
o
log 0≤i≤n ai
log an
n
= lim
= inf
:n≥1 .
δT,S = lim sup
n→∞
n
n
n
n→∞
It follows that ]AT (1, n, ∆0 ) ≥ c−1 exp(nδT,S ) for n ≥ 1. Observe that
X
ΘT (s, 1) L,∆
]AT (1, n, ∆0 ) · exp(−sn), s ≥ 0,
n≥0
whenever both parts are finite. Thus, ΘT (s, 1) is divergent at s = δT,S .
To prove the second statement we estimate the lower bound of δT,S . By Lemma
3.7, we notice that
](B(1, i(L + ∆)) ∩ T ) ≥ θi · exp(i · δG,S · L),
for i ≥ 0. This implies that
δT,S ≥
log ]BT (1, i(L + ∆)) ∩ T
L · δG,S + log θ
≥
.
i(L + ∆)
L+∆
We obtain lim δT,S ≥ δG,S . Since δT,S ≤ δG,S (∀L), the lemma follows.
L→∞
3.2. Patterson-Sullivan measures on the space of ends of an iterated transitional tree. In this and next subsections, for any L 0, let T and T be the
iterated tree set and transitional tree respectively given by Lemma 3.7. At the
same time, assume that they satisfy Lemmas 3.8 and 3.9.
We denote by the common notation ∂T the limit set of T in either the Bowditch
boundary ΛG or in the Floyd boundary ∂λ G. In this subsection, we shall construct
a Patterson-Sullivan measure on ∂T .
Consider the set M(Te) of finite Borel measures on Te := T ∪∂T , which is endowed
with the weak-convergence topology. Then µn → µ for µn ∈ M(Te) if and only if
lim inf µn (U ) ≥ µ(U ) for any open set U ⊂ Te. Note that a set of uniformly bounded
n→∞
measures in M(Te) is relatively compact.
We first construct a family of measures {µsv }v∈T ⊂ M(Te) supported on T . Set
X
1
µsv =
exp(−sd(v, g)) · Dirac(g),
ΘT (s, 1)
g∈T
where s > δT,S and v ∈ T . By Lemma 3.8, the measures {µsv }v∈T are bounded by
a uniform constant depending on L.
24
LEONID POTYAGAILO AND WEN-YUAN YANG
By Lemma 3.9, for any v ∈ T , ΘT (s, v) is divergent at s = δT,S . Choose
si → δT,S such that µsvi converge in M(Te). The limit measures µv = lim µsvi are
called Patterson-Sullivan measures at v. Clearly, {µv }v∈G are absolutely continuous
with respect to each other.
In the sequel, we will write PS-measures as shorthand for Patterson-Sullivan
measures.
A horofunction co-cycle Bξ : G × G → R at conical points ξ ∈ ΛG or ξ ∈ ∂λ G
was studied in [27]. The precise definition is not relevant here, but we have the
following estimation.
Lemma 3.10. [27, Lemma 2.20] For any L > 0 there exists C = C(L) > 0 such that
the following holds.
Fix ξ ∈ ∂T . For any x, y ∈ G, there is a neighbourhood V of ξ in Gλ or G ∪ ΛG
such that the following property holds:
|Bξ (x, y) − Bz (x, y)| < C, ∀z ∈ V ∩ G,
where Bz (x, y) := d(z, x) − d(z, y).
Remarks. (on the proof) The above statement is proved in [27, Lemma 2.20] for
a conical point of the Bowditch boundary, where the constant C is universal (not
depending on L). In our setting by Lemma 3.7 there exists an (, R, L0 )-transitional
ray in the tree T ending at ξ in ∂T . Then by Lemma 2.12 the constant R is uniform
for every ξ ∈ ∂T . So the same proof as [27, Lemma 2.20] works to produce a
constant C = C(L).
We have to warn the reader that the constant C > 0 cannot be made uniform
for all conical points for the action G y ∂λ G on the Floyd boundary as the action
is not necessarily geometrically finite (see the discussion after Lemma 2.12).
End of remarks.
With the help of Lemma 3.10, the following can be proven exactly as Théorème
5.4 in [6].
Lemma 3.11. PS-measures {µg }g∈T on ∂T satisfy the following property,
dµg
(ξ) L exp(−δT,S Bξ (g, h)),
dµh
for µh -a.e. points ξ ∈ ∂T and any g, h ∈ T .
(25)
3.3. Shadow Lemma. We shall establish a shadow lemma for {µg }g∈T on ∂T .
Definition 3.12 (Shadow). The shadow Πr (g) at g ∈ T is the set of points ξ ∈ ∂T
such that there exists SOME geodesic [1, ξ] in T intersecting B(g, r).
Lemma 3.13 (Shadow Lemma). There exists r0 > 0 such that the following holds
exp(−δT,S d(1, g)) ≺ µ1 (Πr (g)) ≺r exp(−δT,S d(1, g))
for any r > r0 and g ∈ T .
Remark. In [27] the Shadow lemma was proved for the whole group G. The current
lemma describes the shadows of the points g ∈ T in terms of δT,S .
Proof. By Lemmas 3.10 and 3.11, there exists C1 = C1 (L), C2 = C2 (L) > 0 such
that the following holds
dµ1
(ξ) ≤ C2 exp(−δT,S d(1, g))
(26)
C1 exp(−δT,S d(1, g)) ≤
dµg
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 25
for µ1 -a.e. points ξ ∈ ∂T . So in order to estimate µ1 (Πr (g)) we can do it for
µg (Πr (g)).
Claim. Given any > 0, there is a constant r0 > 0 such that the following holds
µg (∂T \ Πr (g)) <
for all g ∈ T and r > r0 .
Proof of the Claim. Note that Πr (g) is a closed set. We consider the convex cone
C(∂T \ Πr (g)) of ∂T \ Πr (g), which consists of all geodesic rays in T originating at
1 and terminating at a point in ∂T \ Πr (g). Let V be the set of vertices of T in
C(∂T \ Πr (g)).
For any x ∈ V , consider the branch point o of [1, x] and [1, g] in T (defined in
the proof of Lemma 3.8). Since x ∈
/ C(Πr (g)), we have d(g, o) > r. By a similar
argument to that of Lemma 3.8, we get
X
X
ΘV (s, g) =
exp(−sd(x, g))
exp(−sk) · ΘT (s, 1).
x∈V
r≤k≤d(1,g)
So,
µsg (V ) =
ΘV (s, g)
ΘT (s, 1)
X
exp(−sk),
r≤k≤d(1,g)
which tends to 0 when r → ∞ and s > δT,S .
Thus, the µsg -measure of the open set V ∪ (∂T \ Πr (g)) can be arbitrarily small
for r large enough, and so is µg (∂T \ Πr (g)). This proves the claim.
By Lemma 3.8, we have that {µg (∂T )}g∈T are lower and upper bounded by a
uniform constant depending on L. Let η1 = 1/2 inf{µg (∂T ) : g ∈ T } > 0 and
η2 = sup{µg (∂T ) : g ∈ T } < ∞. By the above Claim, there is a constant r0 > 0
such that the following holds
(27)
η1 < µg (Πr (g)) < η2 , ∀r > r0 ,
for all g ∈ T . So (26) implies that
η1 C1 exp(−δT,S d(1, g)) ≤ µ1 (Πr (g)) ≤ η2 C2 exp(−δT,S d(1, g)),
for all g ∈ G. The Lemma is proved.
By Shadow Lemma to estimate the PS-measure of balls we need to compare it
with that of shadows a in the boundary ∂T . Below, we use the symbol bsc to denote
the integer part of s ∈ R. Denote by Bρλ,1 (ξ, t) (resp. Bρλ,1 (ξ, t)) the ball in ∂T
around ξ ∈ ∂T of radius t with respect to the metric ρλ,1 (resp. ρ̄λ,1 ).
Lemma 3.14 (Shadows Balls). Let r0 given by Lemma 3.13. There exists 0 <
λ0 < 1 such that for any λ ∈ [λ0 , 1) and L > 0, there exist r = r(L, λ) > r0 and
C = C(L, λ) > 0 with the following property.
For any ξ ∈ ∂T and 0 < t < λ, the following holds
(28)
Bρλ,1 (ξ, C −1 t) ⊂ Πr (g) ⊂ Bρλ,1 (ξ, Ct),
and
(29)
Bρ̄λ,1 (ξ, C −1 t) ⊂ Πr (g) ⊂ Bρ̄λ,1 (ξ, Ct),
where g ∈ [1, ξ] is chosen such that d(1, g) = blogλ tc.
26
LEONID POTYAGAILO AND WEN-YUAN YANG
Proof. Let λ0 be given by Proposition 2.3. For any λ ∈ [λ0 , 1[, we consider the
family of Floyd metrics {ρλ,v }v∈ G .
For any 0 < t < λ, let g ∈ [1, ξ] such that d(1, g) = blogλ tc. Thus,
λd(1,g)+1 < t ≤ λd(1,g) .
By construction of T (see Lemma 3.7.5), we know that [1, ξ] is (, R, L)-transitional.
So there exists an (, R)-transitional point z in [1, ξ] such that d(z, g) ≤ L. By
Lemma 2.10.2, there exists κ = κ(λ, , R) such that ρ̄λ,z (1, ξ) ≥ κ. By property (2),
we have ρλ,g (1, ξ) ≥ ρ̄λ,g (1, ξ) ≥ κ · λL .
Set 2C1 = κ · λL and r = max{φλ (C1 ), r0 } where the function φ is given in
Lemma 2.6. Let η ∈ Bρλ,1 (ξ, C1 t). By property (2), it follows that ρλ,g (η, ξ) ≤
λ−d(g,1) ρλ,1 (η, ξ) ≤ C1 . Then ρλ,g (1, η) ≥ C1 and by Lemma 2.6, we have d(g, [1, η]) ≤
r. So η ∈ Πr (g). This proves the first inclusions of (28) and (29) for C = C1
Let η ∈ Πr (g) so that d(g, [1, η]) ≤ r for some geodesic [1, η]. Consequently,
there exists w ∈ [1, η[ such that d(1, w) = d(1, g) and d(g, w) ≤ 2r. By Lemma 2.7
d(1,g)
any segment of [1, ξ] is a Floyd geodesic with respect to ρλ,1 , so ρλ,1 (ξ, g) = λ1−λ .
Let α be a word geodesic between w and g. Every edge of α is in the word distance
at most d(1, g) − 2r from 1. So the Floyd length of α is at most 2r · λd(1,g)−2r . We
obtain
ρ̄λ,1 (ξ, η) ≤ ρλ,1 (ξ, η) ≤ ρλ,1 (g, ξ) + ρλ,1 (w, η) + ρλ,1 (g, w)
1
≤ 2( 1−λ
+
r
λ2r )
· λd(g,1) .
1
r
+
). Then ρ̄λ,1 (ξ, η) ≤ ρλ,1 (ξ, η) ≤ C2 t and so the second
1 − λ λ2r
inclusions of (28) and (29) follow.
Setting C = max{C1 , C2 } we complete the proof of the lemma.
Let C2 = 2λ−1 (
3.4. Proof of Theorem 1.4. We recapitulate the main results of the previous
Subsections in the following.
Proposition 3.15. There exists λ0 > 0 such that for every λ ∈ [λ0 , 1) and L 0,
there exist an L-iterated tree set T and a PS-measure µ1 on ∂T satisfying:
(30)
µ1 (Bρλ,o (ξ, t)) λ,L t−δT ,S / log λ ,
for any ξ ∈ ∂T and 0 < t < λ.
Proof. The existence of the tree T is proved in Lemma 3.7. Lemmas 3.13 and 3.14
and direct calculations imply that ∂T is Alhfors Q-regular for Q = −δT,S / log λ
(see the definition in the Introduction). Hence, (30) follows.
By Lemma 2.12 ∂T consists of uniformly conical points, so Proposition 3.15
implies the first claim of the Theorem. The statement Qi → δG,S (i → ∞) is
proved in Lemma 3.9. Theorem 1.4 is proved.
4. Proofs of Theorems 1.1 and 1.2
We consider the Floyd metric on ∂λ G and shortcut metric on ΛG, where the
corresponding Theorems 1.1 and 1.2 are proved with the same argument.
The following lemma giving the upper bound for the Hausdorff dimension is due
to Marc Bourdon. We notice that it is a general fact which is true for a finitely
generated group G without assuming that it is relatively hyperbolic.
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 27
Lemma 4.1 (M. Bourdon, oral communication). For every λ ∈ (0, 1), the Hausdorff dimension Hdimρλ,1 of ∂λ G (respectively Hdimρ̄λ,1 of ΛG) with respect to
the Floyd metric ρλ,1 (respectively to the shortcut metric ρ̄λ,1 is upper bounded by
−δG,S / log λ.
Proof. To give an upper bound, it suffices to prove that Hs (∂λ G) = 0 for any fixed
s > −δG,S / log λ.
Define Sn = {g ∈ G : d(1, g) = n}. For any g ∈ Sn , define the cone Ωg := {ξ ∈
∂λ G, g ∈ [1, ξ]}, where [1, ξ] is a geodesic between 1 and ξ.
For any ξ ∈ ∂λ G, consider a point x ∈ [1, ξ]∩Sn . By Lemma 2.8 the sub-ray [x, ξ)
λn
is a ρλ,1 -Floyd geodesic. So ρλ,1 (x, ξ) = 1−λ
for any ξ ∈ ∂λ G. Thus, {Ωg : g ∈ Sn }
2λn
.
is an ε-covering of ∂λ G, where ε := 1−λ
δ
G,S
−tn
For any t ∈] − log
for n ≥ 1.
λ , s[, we have −t log λ > δG,S and so ]Sn ≺t λ
We obtain for all n ≥ 1:
X
Hs (∂λ G) ≤
εs ≺ λ(s−t)n
g∈Sn
δ
G,S
which then tends 0 as n → ∞. Thus, Hs (∂λ G) = 0 for any s > − log
λ . The lemma
is proved.
So the upper bound on the Hausdorff dimension of ∂λ G and ΛG in Theorems
1.1, 1.2 is proved. In the remainder of proofs, we aim to establish the lower bound
for the Hausdorff dimension.
Taking into account Proposition 3.15: there exists a universal λ0 > 0 such that
for each L 0, there exist an L-iterated tree T and a PS-measure µ1 on ∂T such
that (30) holds and δT,S → δG,S as L → ∞.
The following lemma shows that PS-measures constructed in Section 3 are actually the Hausdorff measures on ∂T with respect to the Floyd metric ρλ,1 restricted
on ∂T .
Lemma 4.2. Let µ1 be a PS-measure on ∂T in ΛG or ∂λ G. Denote σ = −δT,S / log λ.
Then we have
Hσ (A) L µ1 (A).
for any subset A ⊂ ∂T .
Proof. In the proof, we assume that ∂T is a subset of the Bowditch boundary. The
proof for ∂T ⊂ ∂λ G is similar.
P
Let B be an ε-covering of A for ε > 0. Then µ1 (A) ≤ B∈B µ1 (B). Let ε → 0.
By Proposition 3.15, we obtain that µ1 (A) ≺L Hσ (A).
For the other inequality, we need to make use of the following well-known covering
result. Let B be a metric ball of radius rad(B) in a proper metric space X. Denote
by 5B the union of all balls of radius 2 · rad(B) intersecting B so that k5Bk ≤
10 · rad(B). Then by [20, Theorem 2.1] for a family of balls B in X with uniformly
bounded radii there exists a sub-family B 0 ⊂ B of pairwise disjoint balls such that
the following holds
[
[
(31)
B⊂
5B.
B∈B
B∈B0
Note that µ1 , Hσ are Radon measures. Then for any τ > 0 there exists a compact
set K and an open set U such that K ⊂ A ⊂ U and Hσ (U \ K) < τ, µ1 (U \ K) < τ .
28
LEONID POTYAGAILO AND WEN-YUAN YANG
Set 0 := ρ1 (K, ΛG \ U ) > 0. For any 0 < < 0 , let B be an ε-covering of K.
By (31) and Proposition 3.15, there exists a sub-family B 0 of B such that
X
X
(10 · rad(B))σ ≺L µ1 (U ).
(k5Bk)σ ≤
Hσ (K) ≤
B∈B0
B∈B0
The condition τ → 0 yields Hσ (A) ≺ µ1 (A).
Remark. We note that µ1 is unique in the following sense: if µ1 , µ01 are two PSmeasures, then dµ1 /dµ01 is bounded from up and below.
Lemma 4.2 proves that the Hausdorff dimension of ∂T is equal to σ. Since ∂T
is a subset of the set of uniformly conical points in ∂λ G and ΛG, the dimension
σ = −δT,S / log λ of ∂T gives a lower bound of Hdimρλ,1 (∂λuc G) and Hdimρ̄λ,1 (Λuc G).
Letting L → ∞, we have δT,S → δG,S by Proposition 3.15. So,
Hdimρλ,1 (∂λuc G) ≥ −δG,S / log λ
and
Hdimρλ,1 (Λuc G) ≥ −δG,S / log λ.
The proofs of Theorems 1.1 and 1.2 are complete.
5. Tight Paths and Floyd metrics
In this section, we shall develop a detailed understanding of shortcut geodesics
via a class of well-controlled paths called (generalized) tight paths.
5.1. Tight paths. It is well-known that in hyperbolic spaces, a sufficiently “long”
local geodesic becomes globally a quasi-geodesic. This property in general fails for
the Cayley graph of a relatively hyperbolic group. V. Gerasimov and L. Potyagailo
proposed in [14] a notion of tight paths as a generalization of local geodesics to the
relative setting. The following definition is a small modification of it.
Definition 5.1. For c ≥ 1, l > 0, a path γ is called (c, l)-tight path if for any two
points x, y ∈ γ with d(x, y) ≤ l the subpath [x, y]γ is a c-quasi-geodesic.
Remark. This definition is a partial case of [14, Definition 6.1] where a local quasigeodesicity is requested outside of the horospheres only and an additional condition
is assumed for the horospheres. So if a path is tight in the sense of Definition 5.1
it is also tight in the sense of [14, Definition 6.1] but not necessarily vice versa. In
particular we can use all results proven in [14]. In addition, the above definition
implies that every subpath of a tight path is a tight path itself, which is not always
true in the general case. This stability of the tightness for subpaths will be often
used below.
We also stress that the above definition does not coincide with the standard
notion of local (quasi-)geodesicity when the assumption that the length of a subpath
(and not its diameter) is small implies its (quasi-)geodesicity.
In what follows, to reduce cumbersome quantifiers, we continue to use Convention
2.18 without explicit mention on the constants , R, which depend on the parameter
c > 0 in tight paths.
We recall the following result about tight paths proved in [14].
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 29
Lemma 5.2. For any c ≥ 1, there exist κ = κ(c), l0 = l0 (c) > 0 with the following
property.
Let γ be a (c, l)-tight path for l ≥ l0 . Then ρv (γ− , γ+ ) ≥ κ for any (, R)transitional point v ∈ γ.
Comments on the proof. The statement that ρv (γ− , γ+ ) ≥ κ is first established in
[14, Proposition 6.7] for a special sub-sequence of transitional vertices v ∈ γ. Then
it is shown in the proof of [14, Theorem B] that the tigthtness of a path implies it
for every transitional vertex (up to decreasing the constant κ).
We call below a sequence of points zi = γ(ti ) of a length-parametrized path γ
well-ordered if ti > ti−1 (ti ∈ Z).
The following lemma is an intermediate step in the proof of Proposition 5.6 below
which is the main result of this subsection.
Lemma 5.3 (Transitional tight path is quasi-geodesic). For any c, L ≥ 0, there exist
l0 = l0 (L), c0 = c0 (c) ≥ 1 with the following property.
Let γ be a (c, l)-tight path for l ≥ l0 . Assume that γ is an (, R, L)-transitional
path, where , R satisfy Convention 2.18. Then γ is a c0 -quasi-geodesic.
Proof. By Lemma 5.2, there exists κ = κ(c) ≥ 0 such that ρx (γ− , γ+ ) ≥ κ for every
(, R)-transitional point x ∈ γ. Set D0 = φ(κ/2). Choose l0 ≥ 2(L + D0 ).
Since any subpath of γ is (c, l)-tight, it is enough to prove that there exists a
linear bound for `(γ) with respect to d(γ− , γ+ ). Let α be a geodesic with the same
endpoints as γ. The idea of proof is to find two sequences of well-ordered points in
γ and α respectively which are uniformly close.
Since γ is (, R, L)-transitional, there exists a maximal set of (, R)-transitional
well-ordered points {zi : 1 ≤ i ≤ n} in γ such that
`([zi , zj ]γ ) ≥ 2D0
for i 6= j and
`([zi , zi+1 ]γ ) ≤ 2(L + D0 )
for 1 ≤ i < n. Indeed, let z1 be the first (, R)-transitional point in γ. Suppose zi
is chosen for i ≥ 1. If `([zi , γ+ ]) ≤ 2(L + D0 ) then zi+1 = γ+ . Consider the point
z in [zi , γ+ ]γ such that `([zi , z]γ ) = L + 2D0 . If z is (, R)-transitional in γ, then
set zi+1 = z. Otherwise, there exists an (, R)-transitional point zi+1 such that
`([z, zi+1 ]γ ) ≤ L and `([zi , zi+1 ]γ ) ≤ 2(L + D0 ).
By Lemma 5.2, there exists κ > 0 such that
ρzi+1 (zi , γ+ ) ≥ κ
for any 1 ≤ i < n. By Lemma 2.6, there exists w1 ∈ α such that d(z1 , w1 ) ≤ D0 .
We now choose other wi inductively for i ≥ 1.
Suppose wi ∈ α is chosen such that d(zi , wi ) ≤ D0 . Since d(zi+1 , zi ) ≥ 2D0 ,
we obtain [zi , wi ] ∩ B(zi+1 , D0 ) = ∅. By the choice of D0 = φ(κ/2), we know
that for any v ∈ G, any geodesic outside B(v, D0 ) has lv -length at most κ/2.
So ρzi+1 (wi , zi ) ≤ κ/2 and then ρzi+1 (wi , γ+ ) ≥ κ/2. Thus there exists wi+1 ∈
[wi , α+ ]α such that d(zi+1 , wi+1 ) ≤ D0 . Clearly, the obtained points wi are wellordered on α.
As l0 > 2(L+D0 ), [zi , zi+1 ]γ is a c-quasi-geodesic by the tightness property. Since
wi are well-ordered on α, we see that γ is a c0 -quasi-geodesic for c0 := c + 2D0 .
The following lemma will be often used further.
30
LEONID POTYAGAILO AND WEN-YUAN YANG
Lemma 5.4 (Bounded overlap). For c ≥ 1 and (, R) given by Convention 2.18,
there exist K0 , l0 > 0 with the following property.
Let γ be a (c, l)-tight path for l ≥ l0 . Assume that β1 , β2 are two maximal
connected segments of γ such that (βi )− , (βi )+ ∈ N (Xi ) for some Xi ∈ P with
i = 1, 2. Then `(β1 ∩ β2 ) ≤ K0 . In particular, the endpoints of βi are (, R)transitional for i = 1, 2.
Proof. By Definition 5.1 a subpath of a tight path is itself tight. Then by [14,
Proposition 7.6] it follows that there exists l0 > 0 such that for all l ≥ l0 the
elements of P are uniformly quasi-convex with respect to the system of (c, l)-tight
paths. This implies that there exists a uniform constant ε = ε(, c) > 0 such that
βi ⊂ Nε (Xi ) for i = 1, 2. By Lemma 2.16 we find a constant R = R(ε) > 0 such that
kNε (X) ∩ Nε (X 0 )k ≤ R for every X, X 0 ∈ P. Assume that l0 > cR + c. Since βi
are l-local c-quasi-geodesic for l > l0 , it follows that `(β1 ∩ β2 ) ≤ K0 := cR + c.
Remark. By the bounded intersection of P, this lemma holds trivially if γ is a
quasi-geodesic. However, the tight path γ above is a local quasi-geodesic only.
Let γ be a (c, l)-tight path. Let = (c) given by Convention 2.18 and K0
given by Lemma 5.4. For K > K0 , we consider all maximal connected segments
βi in γ (1 ≤ i ≤ m) such that `(βi ) ≥ K and (βi )− , (βi )+ ∈ N (Xi ) for some
Xi ∈ P. Consequently, Xi 6= Xj for i 6= j. These (βi , Xi ) shall be referred to as
(, K)-components of γ.
We stress that by the argument of Lemma 5.4 the segment βi belongs to N (Xi )
for a uniform > 0 and unique Xi .
We now introduce a modification of a tight path to make the obtained path a
quasi-geodesic.
Definition 5.5 (Truncation of a tight path). Let γ be a (c, l)-tight path for c ≥
1, l > 0. Consider all (, K)-components (βi , Xi ) (1 ≤ i ≤ m) for a fixed K > 2K0 ,
where K0 > 0 is given by Lemma 5.4.
Set y1 = (β1 )− , x2 = (β1 )+ . If βi ∩ βi−1 = ∅ for i ≥ 2, denote yi = (βi )− , xi+1 =
(βi )+ ; otherwise, set yi = xi−1 , xi+1 = (βi )+ . Replace [yi , xi+1 ]γ by a geodesic
segment [yi , xi+1 ] for each i ≥ 1.
The path γ̄ obtained in this way is called a K-truncation of γ.
Remark. The following observation is elementary and useful: every βi produces an
(, K/2)-deep point in Xi in the truncation path γ̄. Consequently, if γ̄ does not
contain an (, R)-deep point, then d((βi )− , (βi )+ ) ≤ 2R for all βi .
The following lemma is the main result of this subsection. It provides a further
generalization of Lemma 5.3 to the truncated tight paths.
Lemma 5.6 (Truncation is quasi-geodesic). For any c ≥ 1, there exist l0 = l0 (c), K =
K(c), c0 = c0 (c) > 0 with the following property. For any l ≥ l0 , the K-truncation
of a (c, l)-tight path is a c0 -quasi-geodesic.
Proof. Let K > 2K0 be a fixed integer, where K0 is given by Lemma 5.4. Let γ̄ be
the K-truncation of a (c, l)-tight path γ. Keeping the notations as in Definition 5.5,
we have by Lemma 5.4 that yi , xi+1 for 1 ≤ i ≤ m are (, R)-transitional points in
γ. Furthermore since [xi , yi ]γ contains no (, K)-components for 1 ≤ i < m, we see
that [xi , yi ]γ is an (, R, L)-transitional path for L := K/2. By Lemma 5.3, there
exist l0 = l0 (L), c0 = c0 (c) ≥ 1 such that [xi , yi ]γ is a c0 -quasi-geodesic.
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 31
The geodesic [yi , xi+1 ] belongs to the -neighbourhood N (Xi ) for some Xi ∈ P
where we assume that yi is the entry point of γ ∩ N (X) and xi+1 is the exit point
of it. We will now show that γ is a quasi-geodesic in a neighbourhood of yi .
Since yi is (, R)-transitional, ρyi (xi , xi+1 ) ≥ κ where κ is given by Lemma 5.2.
Hence, d(yi , [xi , xi+1 ]) ≤ D0 := φ(κ) where φ is the function given by Lemma
2.6. Connect the point xi with an arbitrary point z ∈ [yi , xi+1 ] by a geodesic
β. By the triangle inequality we have d(xi , yi ) + d(yi , xi+1 ) ≤ d(xi , xi+1 ) + 2D0 .
Since d(yi , z) + d(z, xi+1 ) = d(yi , xi+1 ) we obtain d(xi , yi ) + d(yi , z) ≤ d(xi , xi+1 ) −
d(z, xi+1 ) + 2D0 ≤ d(xi , z) + 2D0 . Finally
`([xi , z]γ̄ )
= `([xi , yi ]γ ) + `([yi , z])
≤ c0 d(xi , yi ) + c0 + d(yi , z)
≤ c0 d(xi , z) + c0 + 2c0 D0 .
So [xi , z]γ̄ is a c1 -quasi-geodesic for c1 := c0 (2D0 + 1).
We have that any subpath γ̃ of the truncated path γ̄ is the union of three
types of c1 -quasi-geodesic subpaths: a) γi = [xi , yi ]γ , b) βi = [yi , xi+1 ]γ and c)
δ = [a, b]. Both vertices of the intervals of types a) and b) are transitional on the
corresponding tight path γ, and every γi is (, R, L)-transitional whereas βi is an
(, K)-component. The path γ̃ can contain at most two intervals δ of type c) such
that one of the endpoints of δ coincides with an endpoint of γ̃ and is an interior
point of a geodesic truncation of γ̄.
Repeating the argument of Lemma 5.3 consider a maximal well-ordered subset
V of the transitional vertices {vj ∈ γ̃} in the set W := γ̃ ∩ {yi , xi+1 : 1 ≤ i ≤ m}
such that d(vj , vj+1 ) ≥ 2D0 . We connect the endpoints of γ̃ by a geodesic α. Then
0
0
for each vj ∈ V , there exists vj0 ∈ α such that d(vj , vj0 ) ≤ D0 and vj0 ∈ [vj−1
, vj+1
]α .
Since V is maximal in W , for any w ∈ W there exists v ∈ V such that d(v, w) ≤ 2D0 .
If W = {w1 , w2 , · · · , wn }, there exists a well-ordered set W 0 = {w10 , w20 , · · · , wn0 } of
vertices in α such that d(wi , wi0 ) ≤ 3D0 . Then [γ̃− , w1 ]γ̃ , [wi , wi+1 ]γ and [wn , γ̃+ ]γ̃
are all c1 -quasi-geodesics by the above argument. We have
Pn−1
`(γ̃) ≤ c1 (d(γ̃− , w1 ) + i=1 d(wi , wi+1 ) + d(wn , γ̃+ )) + c1
Pn−1
0
≤ c1 (3D0 + d(γ̃− , w10 ) + i=1 (d(wi0 , wi+1
) + 3D0 ) + d(wn0 , γ̃+ ) + 3D0 ) + c1
0
0
≤ c d(γ̃− , γ̃+ ) + c ,
where c0 := (1 + 3D0 )c1 . The Lemma is proved.
Convention 5.7. For any c ≥ 1, we will assume further on that l0 , K > 0 satisfy
both Lemmas 5.2 and 5.4.
5.2. Shortcut metrics and generalized tight paths. Recall that a Floyd geodesic in the Floyd completion does not in general belongs to the Cayley graph and
the shortening procedure described in subsection 2.3 allows one to approximate
them by local geodesics in the graph. Furthermore the following lemma shows that
this approximation can be done using the tight paths:
Lemma 5.8. [14, Corollary 7.8] For any l > 0 there exists λ0 ∈]0, 1[ such that for
every λ ∈]λ0 , 1[ if the Floyd geodesic γ ⊂ Gλ (with respect to the metric ρλ,o )
joining two distinct points x, y in Gλ does not belong to the Cayley graph G (G, S),
then for ε > 0 there exists a tight path γ̃ ⊂ G (G, S) such that |lλ,o (γ̃) − lλ,o (γ)| < ε.
32
LEONID POTYAGAILO AND WEN-YUAN YANG
The goal of this subsection is to extend this result to the geodesics with respect
to the shortcut metrics {ρ̄λ,o }o∈G on ΛG (see Section 2.2). For this purpose we
generalize the notion of a tight path as follows.
Definition 5.9 (Generalized tight paths and truncations). Let γ be a finite sequence
of (c, l)-tight paths γi in G (G, S) (1 ≤ i ≤ n) such that (γi )+ , (γi+1 )− ∈ N (Xi ) for
some Xi ∈ P where Xi 6= Xj (1 ≤ i 6= j < n).
We say that γ is a (c, l)-generalized tight path if for each pair of entry and
exit points yi , xi+1 of γi and γi+1 respectively in N (Xi ) we have d(yi , xi+1 ) ≥ l
(1 ≤ i < n).
Fix K > 0. For n > 1, consider the K-truncation γ̄i of [xi , yi ]γi where 1 ≤ i ≤ n.
The path
γ̃ = γ̄1 · [y1 , x2 ] · γ̄2 · · · [yn−1 , xn ] · γ̄n
is called the K-truncation of a generalized (c, l)-tight γ.
Remark. Note that a generalized tight path is possibly not connected. If it is
connected, then it is a tight path in Definition 5.1.
Lemma 5.10 (Generalized truncation is quasi-geodesic). For any c ≥ 1, there exist
l0 , K, c0 ≥ 1 such that for any l > l0 , the K-truncation of a (c, l)-generalized tight
path is a c0 -quasi-geodesic.
Proof. Let K = K(c) be given by Lemma 5.6. Let γ̃ be the K-truncation of a
generalized (c, l)-tight path γ. We keep the notations of Definition 5.9. If n = 1,
the proof is finished by Lemma 5.6. Assume that n ≥ 2.
By Lemma 5.6, there exists c1 > 0 such that each γ̄i is a c1 -quasi-geodesic for
each 1 ≤ i < n. We prove below that γ̄i and γ̄i+1 have bounded projection to
N (Xi ) where Xi ∈ P.
By Lemma 2.16, X ∈ P is c1 -contracting and there exist µc1 , Dc1 > 0 such that
the (7) holds. By Convention 2.18, we have ≥ µc1 . Let z be the entry point of γ̄i
in N (Xi ).
Claim. There exists a constant C > 0 such that d(z, yi ) ≤ C.
Proof of Claim. Since yi is the entry point of γi in N (Xi ), there exists an (, K)component β of γi such that z ∈ β. Consider the tight subpath [β+ , yi ]γi and
its K-truncation β1 . By the argument in Lemma 5.6, the path [z, β+ ] · β1 is a
c2 -quasi-geodesic for some c2 > 0.
Since Xi is quasi-convex, there exists ε = ε(, c2 ) > 0 such that any c2 -quasigeodesic with two endpoints in N (Xi ) lies in Nε (Xi ). This implies that β1 ⊂
Nε (Xi ). However, there exists no (, K)-components in [β+ , yi ]γi . Indeed, if not,
0
there exists an (, K)-component β 0 in [β+ , yi ]γi and Y ∈ P such that β±
∈ N (Y )
0
0
and d(β− , β+ ) > K > l0 . Since yi is the entry point of γi in N (Xi ), we have
0
0
0
Y 6= Xi . Since β±
∈ Nε (Xi ), we get d(β−
, β+
) ≤ R := R(max{, ε}) by Lemma
2.16. This is a contradiction as l0 > R(max{, ε}). The same reasoning shows that
d(z, β+ ) ≤ R.
Let L = L(ε, 1) be given by Lemma 2.17. If d(z, yi ) > 2L + R, there exists an
interior point in [β+ , yi ]γi which is (c , 1)-deep in Xi . This is a contradiction, since
yi is the entry point of [β+ , yi ]γi in N (Xi ) and ≥ c by Convention 2.18. Hence
we proved that d(z, yi ) ≤ C := 2L + R. The claim is proved.
By the contracting property 2.16 we see that ProjXi (γ̄i ) ≤ τ := 2(Dc1 + ) + C.
The same is true for ProjXi (γ̄i+1 ). Then γ̃ satisfies the following properties:
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 33
(1) Each γ̄i is a c1 -quasi-geodesic,
(2) max{ProjXi (γ̄i ), ProjXi (γ̄i+1 )} ≤ τ ,
(3) d(xi , yi+1 ) > l.
(These properties imply that γ̃ is (l, c1 , c1 , τ )-admissible in the sense of [28, Section
3]). Therefore by Corollary 3.3 in [28], there exist l0 , c0 > 0 such that for any l > l0
the truncation γ̃ is a c0 -quasi-geodesic. The lemma is proved.
Remark. An alternative way to prove the above Lemma is to use the arguments of
Proposition 6.1.1 in [15] to prove that γ̃ is a curve whose distortion is a quadratic
polynomial, then it follows from Proposition 7.2.2 in [15] that γ̃ is linearly distorted.
Proposition 5.11 (Approximation by generalized tight paths). For any l ≥ 0, there
exists 0 < λ0 < 1 such that the following property holds for any λ ∈ [λ0 , 1).
For any ξ 6= η ∈ ΛG, there exists a sequence of generalized (1, l)-tight paths γn
with (γn )− ∈ [o, ξ], (γn )+ ∈ [o, η] such that
lim d(o, (γn )− ) = lim d(o, (γn )+ ) = ∞
n→∞
n→∞
and
lim lλ,o (γn ) = ρ̄λ,o (ξ, η).
n→∞
Proof. By definition of the shortcut metric (3), for any ε > 0, there are finitely
many pairs (ηi , ξi+1 ) ∈ ω where 1 ≤ i < m such that
X
(32)
ρ̄λ,o (ξ, η) ≥
ρλ,o (ξi , ηi ) − ε/3,
1≤i≤m
where ξ1 := ξ, ηm := η. If m = 1, the proof is completed by Lemma 5.8. Assume
that m ≥ 2.
Let κ = min{ρo (ηi , ξi+1 ) : 1 ≤ i < m} > 0. For each 1 ≤ i < m, there exists
Xi ∈ P such that ηi , ξi+1 ∈ ∂λ (Xi ) where ∂λ (Xi ) is the topological boundary of Xi
in ∂λ G.
First we claim that one can choose ξ˜1 , η̃m and η̃i , ξ˜i+1 ∈ Xi for each 1 ≤ i < m
such that the following two conditions hold,
ε
(1) max{ρo (ξ˜i , ξi ), ρo (η̃i , ηi )} ≤ min{κ/4, 6m
} for 1 ≤ i ≤ m.
˜
(2) If there exists a path α between η̃i , ξi+1 for 1 ≤ i < m such that `(α) ≤ 3l,
then it has lo -length at most κ/4.
Indeed, (1) is true for ξ˜i and η̃i sufficiently close to ξi and ηi respectively. To prove
(2), let R = min{d(1, ξ˜i ), d(1, η̃i ) : 1 ≤ i ≤ m}. We have d(1, α) ≥ R − 3l. So for
sufficiently large R the statement (2) follows from the visibility lemma 2.6.
By Lemma 5.8, we can connect ξ˜i , η̃i by a (1, l)-tight path γi for 1 ≤ i ≤ m such
that (γi )− = ξ˜i and (γi )+ = η̃i and
ε
(33)
|ρλ,o (ξ˜i , η̃i ) − lλ,o (γi )| ≤
.
6m
By the condition (1) above, (32) and (33) , the following holds
X
5ε
(34)
ρ̄λ,o (ξ, η) ≥
lλ,o (γi ) − .
6
1≤i≤m
Let yi , xi+1 be the entry and exit points of γi and γi+1 in N (Xi ) respectively. If
d(yi , xi+1 ) ≥ l for all 1 ≤ i < m, then we are done: {γi } give the generalized tight
path. Otherwise, assume that d(xj+1 , yj ) ≤ l for some 1 ≤ j < m.
34
LEONID POTYAGAILO AND WEN-YUAN YANG
Observe that max{d(η̃j , yj ), d(ξ˜j+1 , xj+1 )} ≥ l + 1. Indeed, if not, it follows that
η̃j , ξ˜j+1 are connected by a path of length at most 3l. By the above condition
(2), we have ρo (η̃j , ξ˜j+1 ) ≤ κ/4. By the condition (1), we have ρo (ηj , ξj+1 ) ≤
3κ/4. We arrive at a contradiction with the definition of κ. Thus, we proved that
max{d(η̃j , yj ), d(ξ˜j+1 , xj+1 )} ≥ l + 1. By Lemma 2.7, we obtain the following
lλ,o ([yj , η̃j ]γ ) + lλ,o ([ξ˜j+1 , xj+1 ]γ ) ≥ lλ,o ([yj , xj+1 ]).
j
j+1
which yields
lλ,o (γj ) + lλ,o (γj+1 ) ≥ lλ,o ([ξ˜j , yj ]γj ) + lλ,o ([yj , xj+1 ]) + lλ,o ([xj+1 , η̃j+1 ]γj+1 )
≥ ρλ,o (ξ˜j , η̃j+1 ).
This implies that we can drop the pair (ηj , ξj+1 ) in (32) such that the corresponding
inequality in (34) still holds. Precisely, choose a (1, l)-tight path αj between ξ˜j , η̃j+1
such that
ε
|lλ,o (αj ) − ρλ,o (ξ˜j , η̃j+1 )| ≤
.
6m
ε
. It follows by (34),
So lλ,o (γj ) + lλ,o (γj+1 ) ≥ lλ,o (αj ) −
6m
X
ε
5ε
−
.
ρ̄λ,o (ξ, η) ≥
lλ,o (γi ) + lλ,o (αj ) −
6
6m
1≤i≤m;i6=j,j+1
Consider the new set of (1, c)-tight paths γi (i 6= j, j + 1) and αj . Repeat the
above argument for those j for which d(xj+1 , yj ) ≤ l. Since m is finite, for every
ε > 0 we obtain a generalized tight path γ such that ρ̄λ,o (ξ, η) ≥ lλ,o (γ) − . The
Proposition is proved.
5.3. Floyd and shortcut metrics on uniformly conical points. A priori, the
shortcut metrics as quotient of the Floyd metrics might be distorted in a unexpected
way. The main result of this subsection is to show that this distortion is not severe
for uniformly conical points.
Fix a basepoint o ∈ G. Recall that, in Section 2.4, Λuc
L,o G denotes the set
of uniformly conical points ξ ∈ ΛG for which there exists an (, R, L)-transitional
uc
geodesic ray between o and ξ. Similarly, denote by ∂L,o
G the set of uniformly conical
points in ∂λ G based at o. By Proposition 2.4, there exists one-to-one correspondence
uc
between Λuc
L,o G and ∂L,o G.
The following is a version of Proposition 2.20 for generalized tight paths.
Proposition 5.12. There exist l0 , D > 0 such that for any L > 0, there exists
M = M (L) > 0 with the following property.
Denote α1 = [o, ξ] and α2 = [o, η] for ξ 6= η ∈ CL G. Let γ be a generalized (1, l)tight path for some l ≥ l0 with γ− ∈ α1 and γ+ ∈ α2 . If d(o, γ− ), d(o, γ+ ) 0,
then there exists z ∈ γ such that d(z, α1 ∪ α2 ) ≤ D and d(z, αi ) ≤ M for i = 1, 2.
Moreover, |d(o, z) − d(o, [ξ, η])| ≤ M .
Proof. Let l0 , c0 ≥ c, K > 0 given by Lemma 5.10 such that the K-truncation γ̃ of
a generalized (c, l)-tight γ for l ≥ l0 is a c0 -quasi-geodesic. By Proposition 2.20,
there exists an (c0 , Rc0 )-transitional point z in γ̃ such that the conclusion of this
Proposition holds. If the point z lies on γ, then we are done. So below, we assume
that z ∈
/ γ, and then have two cases to consider as follows.
Let L1 = L(c , Rc0 ) be given by Lemma 2.17, where c0 ≥ c ≥ 1 and they all
satisfy Convention 2.18. Assume that l0 ≥ 2L1 .
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 35
Case 1. The point z lies in some (c , K)-component β of a (c, l)-tight path γi .
Then max{d(β− , γ), d(β+ , γ)} ≤ L1 . Indeed, if not, then min{d(z, β− ), d(z, β+ )} ≥
L1 . By applying Lemma 2.17 for the geodesic β, we have z ∈ γ̃ is (1 , Rc0 )-deep in
Xj . This is a contradiction, as z is an (c0 , Rc0 )-transitional point in γ̃.
Case 2. The point z lies in some [yj , xj+1 ] for some j, where [yj , xj+1 ] is given in
Definition 5.9 of a generalized tight path. By the same reasoning as above, we apply
Lemma 2.17 for the geodesic [(β)− , (β)+ ]. Then min{d(z, yj ), d(z, xj+1 )} ≤ L1 .
Thus, we proved that z has a distance at most 2L1 to a point in γ. The conclusion
follows as a consequence of Proposition 2.20.
The main result of this subsection is the following.
Proposition 5.13 (Visual Floyd/shortcut metric). There exists 0 < λ0 < 1 such that
the following holds for any L > 0 and λ ∈ [λ0 , 1).
We have
uc
ρλ,o (ξ, η) L λn , ∀ ξ 6= η ∈ ∂L,o
G
and
ρ̄λ,o (ξ, η) L λn , ∀ ξ 6= η ∈ Λuc
L,o G
where n = d(o, [ξ, η]).
Proof. Let us consider the shortcut metric case only. The Floyd metric case is
similar and even easier.
Let α1 , α2 be two (1 , R1 , L)-transitional geodesic rays originating at o and terminating at ξ, η respectively.
Let l0 , D > 0 be given by Proposition 5.12, and we choose λ0 ∈]0, 1[ verifying
Proposition 5.11 for l = l0 . Then by Propositions 5.12 and 5.11 there exists M =
M (L) > 0 such that the following holds:
(1) For each k > 0, there exists a sequence of generalized (1, l0 )-tight paths γk
with (γk )− ∈ α1 , (γk )+ ∈ α2 and such that (γk )− → ξ, (γk )+ → η and
|lλ,o (γk ) − ρ̄λ,o (ξ, η)| ≤ 1/k.
(35)
(2) There exists zk ∈ γk such that d(zk , α1 ∪ α2 ) ≤ D and d(zk , αi ) ≤ M for
i = 1, 2. Moreover, |d(o, zk ) − d(o, [ξ, η])| ≤ M .
Denote uk := (γk )− and vk := (γk )+ .
Upper bound. Choose xk ∈ α1 and yk ∈ α2 such that max{d(zk , xk ), d(zk , yk )} ≤
M . Then for every point t ∈ [xk , zk ] ∪ [zk , yk ] we have
d(o, t) ≥ d(o, zk ) − M ≥ n − 2M.
Hence
max{lλ,o ([xk , zk ]), lλ,o ([yk , zk ])} ≤ M · λn−2M .
We also have
min{d(o,xk ),d(o,yk )}
≤
max{lλ,o ([xk , ξ]α1 ), lλ,o ([yk , η]α2 )} ≤ λ
1−λ
λn−2M
1−λ
.
It follows that
ρ̄λ,o (ξ, η) ≤ lλ,o ([xk , ξ]α1 ) + lλ,o ([yk , η]α2 ) + lλ,o ([xk , zk ]) + lλ,o ([yk , zk ])
(36)
1
≤ 2λn−2M ( 1−λ
+ M ).
1
Let C1 := 2λ−2M ( 1−λ
+ M ). Then ρ̄λ,o (ξ, η) ≤ C1 λn .
36
LEONID POTYAGAILO AND WEN-YUAN YANG
Lower bound. Since d(zk , α1 ∪ α2 ) ≤ D, there exists wk ∈ α1 ∪ α2 such that
d(zk , wk ) ≤ D. Assume that wk ∈ α2 for concreteness.
By Lemma 2.8, any segment of α2 is a Floyd geodesic. Since vk → η and
d(o, wk ) ≤ n + D, we can assume that wk ∈ [o, wk ]α2 for all k 0. So
ρλ,o (vk , wk ) = lλ,o ([wk , vk ]α2 )
=
λd(o,wk )
1−λ
≥
λd(o,zk )+D
1−λ
λd(o,vk )
1−λ
−
−
λd(o,vk )
1−λ .
We have
ρλ,o (vk , wk ) ≤ lλ,o (γk ) + lλ,o ([wk , zk ]).
Since d(zk , wk ) ≤ D we have
lλ,o ([wk , zk ]) ≤ D · λd(o,zk )−D .
Thus
lλ,o (γk ) ≥ ρλ,o (vk , wk ) − lλ,o ([wk , zk ])
≥ lλ,o ([vk , wk ]α2 ) − lλ,o ([wk , zk ])
(37)
D
D
)λd(o,zk )
λD
D
D
)
λD
λ
−
≥ ( 1−λ
λ
≥ ( 1−λ
−
−
λd(o,vk )
1−λ
· λM · λn −
λd(o,vk )
1−λ .
Since D does not depend on L by Lemma 5.12, there exists 1 > λ0 > 0 such that
(38)
D
λD
λD
D
0
− D ≥
− D > 0,
1−λ λ
1 − λ0
λ0
λD
0
− λDD ) · λM > 0. Note that d(o, vk ) → ∞ as
for any λ ∈ [λ0 , 1). Let C2 := ( 1−λ
0
0
k → ∞. By (35) and (37), passing to the limit when k → ∞, we obtain
ρ̄λ,o (ξ, η) ≥ C2 λn ,
for any ξ 6= η ∈ Λuc
L,o G and any L > 0. The proof is then complete.
Remark.
(1) The fact that the constant D does not depend on L is crucial for
the choice of λ0 in (38).
(2) This lemma gives an asymptotic formula for two uniformly conical points
with respect to Floyd metric and shortcut metric. This could be used to
give an alternative proof of Lemma 3.14, but cannot be derived from (the
proof of) Lemma 3.14.
6. Appendix: Visual metrics and Floyd metrics are bilipschitz
equivalent
The aim of the Appendix is to give a short proof that the visual Gromov metric
νa,o and the Floyd metric ρλ,o on the boundary ∂X of a δ-hyperbolic graph (X, d)
are bilipschitz equivalent for some choice of parameters a and λ. This fact, mentioned in the Introduction, is often considered as a folklore, however we have not
found a complete proof of it in the literature (see e.g. [18, Lemma 7.2.1] and the
key inequality after it, or [6, Formula (1.3)], in both cases the fact is stated without
proof).
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 37
Recall the definition of the Gromov visual metric ν on ∂X. For a real parameter
a > 0 set δa,o (ξ, η) = e−a(ξ|η) , where (ξ|η) denotes the Gromov product for the
basepoint o. Let
(39)
νa,o (ξ, η) = inf
n
nX
o
δa,o (ci−1 , ci ) : ci ∈ Cξ,η ,
i=0
where Cξ,η is the set of chains of points in ∂X such that c0 = ξ and cn = η.
ln 2
then νa,o is a metric on ∂X satisfying the following inequality [17,
If a <
6δ
Proposition 7.10]:
(3 − 2eaδ ) · δa,o (ξ, η) ≤ νa,o (ξ, η) ≤ δa,o (ξ, η)
(40)
Remark. We note that an inequality similar to (40) where the metric ν is replaced
by the Floyd metric (which is our goal now) is formally stated in [23, page 5] but
no justification is given.
For the Gromov product the following inequality is true [7, Lemma 2.7]:
e−4aδ · δa,o (ξ, η) ≤ e−ad(o,[ξ,η]) ≤ δa,o (ξ, η),
(41)
where d(o, [ξ, η]) is the distance in X from o to the union of all geodesics between
ξ and η. The inequalities (40) and (41) imply:
νa,o (ξ, η) C1 e−a·d(o,[ξ,η])
(42)
for the constant C1 = max{e4aδ , 3 − 2eaδ }.
The following proposition provides the bilipschitz equivalence between the visual
metric and the Floyd metric on the boundary of X.
Proposition 6.1. Let (X, d) be a δ-hyperbolic graph. There exist a constant a0 such
that for any a ∈]0, a0 ] there exists a constant C for which
∀ξ, η ∈ ∂X : νa,o (ξ, η) C ρλ,o (ξ, η),
(43)
where λ = e−a .
Proof. By (39) and (42) we have
n n
o
X
νa,o (ξ, η) = inf
λd(o,[ci−1 ,ci ]) : ci ∈ Cξ,η ≤ C1 ρλ,o (ξ, η).
Cξ,η
i=0
Indeed ρλ,o is obtained by taking infimum of the expression (39) over the subset of
Cξ,η given by the sets of vertices of paths between ξ and η (see Section 2.2).
To prove the opposite inequality, we need to use the δ-thin triangle property.
Consider a geodesic triangle with vertices o, x, y in X. There exists a δ-center c on
[x, y] such that d(c, [o, x]) ≤ δ and d(c, [o, y]) ≤ δ. For notational simplicity we
ignore a small uniform difference between different hyperbolicity constants (see e.g.
[17, Proposition 2.21]), and denote all of them by δ > 0. Since |d(o, c)−d(o, [x, y])| is
uniformly upper bounded to simplify the notations again we assume that d(o, c) =
d(o, [x, y]).
38
LEONID POTYAGAILO AND WEN-YUAN YANG
Choose x0 ∈ [o, x] and y 0 ∈ [o, y] so that max{d(c, x0 ), d(c, y 0 )} ≤ δ. We have
min{d(o, [x0 , c]), d(o, [y 0 , c]) ≥ d(o, [x, y]) − δ. Hence the Floyd length lλ,o ([x0 , c]) of
[x0 , c] is at most
0
δ · λd(o,[x ,c]) ≥ δ · λd(o,[x,y])−δ ,
and similarly for lλ,o ([y 0 , c]). Since [x0 , x] and [y 0 , y] are Floyd ρλ,o -geodesics (Lemma
2.8) we have
0
λd(o,x )
λd(o,[x,y])−δ
0
lλ,o ([x , x]) =
≤
,
1−λ
1−λ
0
and the same for lλ,o ([y , y]). Summing all up we obtain the following estimation
for the Floyd length of [x, y]:
ρλ,o (x, y) ≤ lλ,o ([x0 , x]) + lλ,o ([x0 , c]) + lλ,o ([y 0 , c]) + lλ,o ([y 0 , y])
d(o,[x,y])
≤ 2 λ 1−λ + 2δ · λd(o,[x,y])−δ
≤ C2 · λd(o,[x,y]) ,
n 2
2δ o
for the constant C2 = max
, δ . Passing to the limits when x → ξ ∈ ∂X
1−λ λ
and y → η ∈ ∂X, and using (42) we obtain (43) for the constant C = max{C1 , C2 }.
References
1. C. Bishop and P. Jones, Hausdorff dimension and kleinian groups, Acta Math. 3 (1995),
no. 179, 1?39.
2. B. Bowditch, Convergence groups and configuration spaces, Geometric Group Theory Down
Under (J. Cossey, C.F. Miller, W.D. Neumann, M. Shapiro, eds.), pp. 23–54, de Gruyter,
1999.
3.
, Relatively hyperbolic groups, Int. J. Algebra Comput. (2012), no. 22, p1250016.
4. M. Burger and S. Mozes, CAT(-1)-spaces, divergence groups and their commensurators, Journal of the American Mathematical Society (1996), no. 1, 57–93.
5. J. Cannon, The combinatorial structure of cocompact discrete hyperbolic groups, Geom. Ded.
(1984), no. 2, 123–148.
6. M. Coornaert, Mesures de Patterson-Sullivan sure le bord d’un espace hyperbolique au sens
de Gromov, Pac. J. Math. (1993), no. 2, 241–270.
7. M. Coornaert, T. Delzant, and A. Papadopoulos, Géométrie et théorie des groupes. les groupes
hyperboliques de Gromov, Lecture Notes in Mathematics, Springer-Verlag, 1990.
8. F. Dal’bo, M. Peigné, J.C. Picaud, and A. Sambusetti, On the growth of quotients of Kleinian
groups, Ergodic Theory and Dynamical Systems 31 (2011), no. 3, 835–851.
9. W. Floyd, Group completions and limit sets of Kleinian groups, Inventiones Math. 57 (1980),
205–218.
10. D. Gaboriau and F. Paulin, Sur les immeubles hyperboliques, Geom. Dedicata (2001), 153–197.
11. V. Gerasimov, Expansive convergence groups are relatively hyperbolic, Geom. Funct. Anal. 19
(2009), 137–169.
12.
, Floyd maps for relatively hyperbolic groups, Geom. Funct. Anal. (2012), no. 22, 1361
– 1399.
13. V. Gerasimov and L. Potyagailo, Quasi-isometries and Floyd boundaries of relatively hyperbolic groups, J. Eur. Math. Soc. 15 (2013), 2115 – 2137.
14.
, Non-finitely generated relatively hyperbolic groups and floyd quasiconvexity, Groups,
Geometry and Dynamics 91 (2015), 95–135.
15.
, Quasiconvexity in the relatively hyperbolic groups, Journal für die reine und angewandte Mathematik (Crelle journal) 710 (2016), 95–135.
16.
, Similar relatively hyperbolic actions of a group, International Mathematics Research
Notices (2016), no. 7, 2068–2103.
17. E. Ghys and P. de la Harpe, Sur les groupes hyperboliques d’après Mikhael Gromov, Progress
in Math., Birkaüser, 1990.
HAUSDORFF DIMENSION OF BOUNDARIES OF RELATIVELY HYPERBOLIC GROUPS 39
18. M. Gromov, Hyperbolic groups, Essays in group theory (S Gersten, editor), vol. 1, pp. 75–263,
Springer New York, 1987.
19. A. Karlsson, Free subgroups of groups with non-trivial Floyd boundary, Comm. Algebra. 31
(2003), 5361–5376.
20. P. Mattila, Geometry of sets and measures in Euclidean spaces: Fractals and rectifiability,
Cambridge Studies in Advanced Mathematics, 44, Cambridge University Press, Cambridge,
1995.
21. S. Patterson, The limit set of a Fuchsian group, Acta Mathematica (1976), no. 1, 241–273.
22. F. Paulin, On the critical exponent of a discrete group of hyperbolic isometries, Differential
Geom. Appl. (1997), 231236.
23. M. Puls, The first lp -cohomology of some groups with one end, arXiv:math/0509171, 2005.
24. B. Stratmann, The exponent of convergence of Kleinian groups; on a theorem of Bishop
and Jones, Proceedings of the Conference ‘Fractal Geometry and Stochastics’ (Friedrichroda,
Germany, 2003); Progress in Probability, vol. 57, p. 93?107, 2004.
25. D. Sullivan, The density at infinity of a discrete group of hyperbolic motions, Publ. Math.
IHES (1979), 171–202.
26. P. Tukia, Conical limit points and uniform convergence groups, J. Reine. Angew. Math. 501
(1998), 71–98.
27. W. Yang, Patterson-Sullivan measures and growth of relatively hyperbolic groups, Preprint,
2013.
28.
, Growth tightness for groups with contracting elements, Math. Proc. Cambridge Philos. Soc 157 (2014), 297 – 319.
, Peripheral structures of relatively hyperbolic groups, Journal für reine und ange29.
wandte Mathematik 689 (2014), 101–135.
Leonid Potyagailo, UFR de Mathématiques, Université de Lille 1, 59655 Villeneuve
d’Ascq cedex, France
E-mail address: [email protected]
Beijing International Center for Mathematical Research (BICMR), Beijing University, No. 5 Yiheyuan Road, Haidian District, Beijing, China
E-mail address: [email protected]
| 4math.GR
|
arXiv:1711.02837v1 [stat.ML] 8 Nov 2017
Revealing structure components of the retina by deep
learning networks
Qi Yan, Zhaofei Yu, Feng Chen
Center for Brain-Inspired Computing Research, Department of Automation, Tsinghua University
{q-yan15,yuzf12}@mails.tsinghua.edu.cn
[email protected]
Jian K. Liu
Institute for Theoretical Computer Science, Graz University of Technology
[email protected]
Abstract
Deep convolutional neural networks (CNNs) have demonstrated impressive performance on visual object classification tasks. In addition, it is a useful model
for predication of neuronal responses recorded in visual system. However, there
is still no clear understanding of what CNNs learn in terms of visual neuronal
circuits. Visualizing CNN’s features to obtain possible connections to neuronscience underpinnings is not easy due to highly complex circuits from the retina
to higher visual cortex. Here we address this issue by focusing on single retinal ganglion cells with a simple model and electrophysiological recordings from
salamanders. By training CNNs with white noise images to predicate neural responses, we found that convolutional filters learned in the end are resembling to
biological components of the retinal circuit. Features represented by these filters
tile the space of conventional receptive field of retinal ganglion cells. These results suggest that CNN could be used to reveal structure components of neuronal
circuits.
1 Introduction
Deep convolutional neural networks (CNNs) have been a powerful model for numerous tasks related
to system identification [1]. By training a CNN with a large set of targeted images, it can achieve
the human-level performance for visual object recognition. However it is still a challenge for understanding the relationship between computation and the underlying structure components learned
within CNNs [2, 3]. Thus, visualizing and understanding CNNs are not trivial [4].
Inspired by neuroscience studies, a typical CNN model consists of a hierarchical structure of layers
[5], where one of the most important properties for each convolutional (conv) layer is that one can
use a conv filter as a feature detector to extract useful information from inputed images after the
previous layer [6, 7]. Therefore, after learning, conv filters are meaningful. The features captured by
these filters can be represented in the original natural images [4]. Often, one typical feature shares
some similarities with part of natural images from the training set. These similarities are obtained
by using a very large set of specific images. The benefit of this is that features are relative universal
for one category of objects, which is good for recognition. However, it also causes the difficulty
of visualization or interpretation due to the complex nature of natural images, i.e., the complex
statistical structures of natural images [8]. As a result, the filters and features learned in CNNs are
often not obvious to be interpreted [9].
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
On the other hand, researchers begin to adapt CNNs for studying the target questions from neuroscience. For example, CNNs have been used to model the ventral visual pathway that has been
suggested as a route for visual object recognition starting from the retina to visual cortex and reaching inferior temporal (IT) cortex [10–12]. The prediction of neuronal responses recorded in monkey
in this case has a surprisingly good performance. However, the final output of this CNN model is
representing dense computations conducted in many previous conv layers, which may or may not be
related to the neuronscience underpinnings of information processing in the brain. Understanding
these network components of CNN are difficult given the IT cortex part is sitting at a high level of
our visual system with abstract information, if any, encoded [13]. In principle, CNN models can
also be applied to early sensory systems where the organization of underlying neuronal circuitry is
relatively more clear and simple. Thus one expect knowledge of these neuronal circuitry could provide useful and important validation for such models. For instance, a recent study employs CNNs to
predict neural responses of the retinal ganglion cells to white noise and natural images [14].
Here we move a step further in this direction by relating CNNs with single RGCs. Specifically,
we used CNNs to learn to predict the responses of single RGCs to white noise images. In contrast
to the study by [14] where one single CNN model was used to model a population of RGCs, in the
current study, our main focus is based on single RGCs to revealing the network structure components
learned by CNNs. Our aim is to study what kind of possible biological structure components in the
retina can be learned by CNNs. This concerns the research focus of understanding, visualizing and
interpreting the CNN components out of its black box.
To the end, by using a minimal model of RGC, we found the conv filters learned in CNN are essentially the subunit components of RGC model. The features represented by these filters are fallen into
the receptive field of modeled RGC. Furthermore, we applied CNNs to analyze biological RGC data
recorded in salamander. Surprisingly, the conv filters are resembling to the receptive fields of bipolar
cells that sit in the previous layer of RGC and pool their computations to a downstream single RGC.
2 Methods
2.1 RGC model and data
A simulated RGC is modeled in Fig. 1 as previously [15]. The model cell has five subunits that
each filter, similar to a conv filter in DNN, convolves the incoming stimulus image and then applies
an nonlinearity of threshold-quadratic rectification. The subunit signals are then polled together by
the RGC. The polled signal is applied with a threshold-linear output nonlinearity with a positive
threshold at unity to make spiking sparse.
The biological data of RGCs were recorded in salamander as described in [15]. Briefly RGC spiking
activity were obtained by multielectrode array recordings as in [16]. The retinas was optically stimulated with spatiotemporal white noise images, temporally updated at a rate of 30 Hz and spatially
arranged in a checkerboard layout with stimulus pixels of 30x30 mµ.
2.2 CNN model
We adopt a CNN model containing two convolution layers and a dense layer as in [14]. Several
sets of parameters in convolution layers, including the number of layers, the number and size of
convolution filters were explored. The predication performance is robust against these changes of
parameters. Therefore we used a filter size as of 15 × 15 to compare our results with those in [14].
The major difference between our model with that in [14] is that our CNN is for studying of single
RGCs.
For the RGC model, we used a data set consisting of 600k training samples of white noise images,
and additional set of samples for testing. The training labels are a train of binary spikes with 0 and
1 generated by the model. For the biological RGCs recorded in salamander, we used the same data
sets as in [15]. Briefly there are about 40k training samples and labels with the number of spikes
as in [0 5] for each image. The test data have 300 samples, which are repeatedly presented to the
retina for about 200 trials. The average firing rate of this test data is compared to the CNN output
for performance calculation.
2
A
B
Subunit model
CNN model
convolution
convolution
dense
labels
Bernoulli
process
# Spikes
Images
Output
nonlinearity
Subunits
C
Subunit model
RF
Spikes
Subunit
nonlinearity
D Conv filters
CNN model
Features
Figure 1: The CNN filters are resembling to the subunits in RGC model. (A) An illustration of RGC
model structure. Note there are five subunits playing the role of conv filters. (B) An illustration of
CNN model that trains the same set of images to predicate the labels, here spikes, of all images. (C)
Receptive fields (RFs) of modeled RGC and predication by CNN model. (D) Visualizing of CNN
model components of both conv filters and average features represented by each filter.
3 Results
Here we focus on single RGC that has the benefit to clarify the network structure components of
CNNs. Recently, a variation of non-negative matrix factorization was used to analyze the RGC’
responses to white noise images and identify a number of subunits, resembling to the biopolar cells,
within the receptive field of each RGC [15]. With this picture in mind, here we address the question
that what types of network structure components can be revealed by CNNs when they are used to
model the single RGC response.
A previous study [14] focused on predicating neural response of RGC at the population level with
one CNN model, and claimed that the features represented by conv filters are resembling to the
receptive fields of bipolar cells (BCs). However, a careful examination reveals that this connection
between CNN feature map and BCs is weak since the number of conv filters in the CNN is much
less than that of BCs in the RGC population from the retina. By using a CNN model, one expect to
reveal a more clear picture of this connection between CNNs and the retina.
Here, we set up a single RGC model with conv subunits as in Fig. 1(A), which is resembling to a
2-layer CNN with one conv layer of subunits and one dense layer of single RGC. By training a CNN
as in Fig. 1(B) with a set of white noise images to predicate the target labels as the simulated spikes
generated by this RGC model, we found that the CNN model can predicate the RGC model response
well with Pearson correlation coefficient (CC) up to 0.70 similar to the study by [14].
Interestingly, we also found the CNN model can predicate the receptive field (RF) well as in
Fig. 1(C). Furtherer more, the conv filters learned by CNN are the exact subunits employed in
the RGC model as shown in Fig. 1(C). A subset of the conv filters, that can be termed as effective
filters, start from random shapes and converge to the exact subunits. Although we set up the filter
size as 15x15 pixels, the resulting effective filters are sparse represented with a 6x6 pixel size. The
rest of the filters are still random and close to zero. Therefore, these results show that CNN parameters are highly redundant. Such a redundancy of parameters, including conv filters, units/neurons
and connections of conv and dense layers, is widely observed for deep learning models [17–19].
All together, These results suggest that the CNN model can identify the underlying hidden network
structure components within the RGC model by only looking at the input stimulus images and the
output response in terms of the number of spikes.
3
A
BConv filters
Data
RF
100 m
Features
CNN
100 m
100 m
C
Data
Firing rate (Hz)
CNN
40
20
0
0
2
4
6
8
10
Time (sec.)
Figure 2: The CNN reveals subunit structures in biological RGC data. (A) Receptive fields of the
sample cell and CNN predication. (B) Visualizing of CNN model components of both conv filters
and average features represented by each filter. (C) Neural response predicated well by CNN model
visualized by RGC data spike rasters (upper) and CNN spike rasters (middle) and their average firing
rates.
To further characterizing these structure components in details, we use a CNN to learn the biological
RGC data with the similar images of white noise and the spiking responses. Similar to the results of
the RGC model above, the outputs of CNN model can recover the receptive field of data very well
as in Fig. 2(A). We also found that the learned conv filters converge to a set of localized subunits
whereas the rest of filters are noisy and close to zero as in Fig. 2(B). The size of these localized
filters is comparable to that in bipolar cells around 100 mµ [15].
In addition, the features represented by these localized conv filers are also localized. Given the
example RGC is a OFF type cell that response to the dark part of images strongly, most features
have similar OFF peaks resulted from the OFF BC-like filters. These OFF features tile the space
of receptive field of RGC. Interestingly, there are some features with ON peaks, which play a role
as inhibition in the retinal circuit. A few features have some complex structures mixed with OFF
and ON peaks, which are mostly resulted from the less localized filters. However, if the filters are
pure noise, the resulting features are pure noise without any structure embedded. Besides filters and
features, the CNN model generates a good predication of RGC response as in Fig. 2(C) with the CC
up to 0.75. These observations are similar across different RGCs recorded.
4 Discussion
Here by focusing on single RGCs, we shown that CNN can learn their parameters in an interpretable
fashion. Both filters and features are close to the biological underpinnings within the retinal circuit.
With the benefits of relative well-understood neuronal circuit of the retina ganglion cells, our preliminary results give a strong evidence that the building-blocks of CNNs are meaningful when they
are applied to neuroscience for revealing network structure components. Our results extend the previous studies [11, 14] that focus on predication of neural responses. Furtherer more, our approach is
suitable to address other difficult issues of deep learning, such as transfer learning, since the domain
4
of images seen by single RGCs is local and less complicated than those global structures of entire
natural images.
5 Acknowledgements
Q.Y., Z.Y. and F.C. were supported in part by the National Natural Science Foundation of China
under Grant 61671266, 61327902, in part by the Research Project of Tsinghua University under
Grant 20161080084, in part by National High-tech Research and Development Plan under Grant
2015AA042306. J.K.L. was partially supported by the Human Brain Project of the European Union
#604102 and #720270.
References
[1] Yann Lecun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444, 2015.
[2] Leslie N. Smith and Nicholay Topin.
arXiv:1611.00847v3, 2016.
Deep convolutional neural network design patterns.
[3] Adam H. Marblestone, Greg Wayne, and Konrad P. Kording. Toward an integration of deep learning and
neuroscience. Frontiers in Computational Neuroscience, 10, sep 2016.
[4] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European
Conference on Computer Vision, pages 818–833, 2014.
[5] Y Lecun, K Kavukcuoglu, and C Farabet. Convolutional networks and applications in vision. In IEEE
International Symposium on Circuits and Systems, pages 253–256, 2010.
[6] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. Computer Science, 2014.
[7] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional
neural networks. In International Conference on Neural Information Processing Systems, pages 1097–
1105, 2012.
[8] E. P. Simoncelli and B. A. Olshausen. Natural image statistics and neural representation. Annual Review
of Neuroscience, 24(24):1193, 2001.
[9] Matthew D. Zeiler, Graham W. Taylor, and Rob Fergus. Adaptive deconvolutional networks for mid and
high level feature learning. In International Conference on Computer Vision, pages 2018–2025, 2011.
[10] Daniel Yamins, Ha Hong, Charles Cadieu, and James J. Dicarlo. Hierarchical modular optimization of
convolutional networks achieves representations similar to macaque it and human ventral stream. Advances in Neural Information Processing Systems, pages 3093–3101, 2013.
[11] D. L. K. Yamins, H. Hong, C. F. Cadieu, E. A. Solomon, D. Seibert, and J. J. DiCarlo. Performanceoptimized hierarchical models predict neural responses in higher visual cortex. Proceedings of the National Academy of Sciences, 111(23):8619–8624, may 2014.
[12] Charles F. Cadieu, Ha Hong, Daniel L. K. Yamins, Nicolas Pinto, Diego Ardila, Ethan A. Solomon,
Najib J. Majaj, and James J. Dicarlo. Deep neural networks rival the representation of primate it cortex
for core visual object recognition. Plos Computational Biology, 10(12):e1003963, 2014.
[13] Daniel L K Yamins and James J DiCarlo. Using goal-driven deep learning models to understand sensory
cortex. Nature Neuroscience, 19(3):356–365, feb 2016.
[14] Lane McIntosh, Niru Maheswaranathan, Aran Nayebi, Surya Ganguli, and Stephen Baccus. Deep learning
models of the retinal response to natural scenes. In Advances in Neural Information Processing Systems
29. 2016.
[15] Jian K. Liu, Helene M. Schreyer, Arno Onken, Fernando Rozenblit, Mohammad H. Khani, Vidhyasankar
Krishnamoorthy, Stefano Panzeri, and Tim Gollisch. Inference of neuronal functional circuitry with spiketriggered non-negative matrix factorization. Nature Communications, 8(1), jul 2017.
[16] Jian K. Liu and Tim Gollisch. Spike-triggered covariance analysis reveals phenomenological diversity of
contrast adaptation in the retina. PLOS Computational Biology, 11(7):e1004425, jul 2015.
5
[17] Misha Denil, Babak Shakibi, Laurent Dinh, MarcAurelio Ranzato, and Nando de Freitas. Predicting
parameters in deep learning. In C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q.
Weinberger, editors, Advances in Neural Information Processing Systems 26, pages 2148–2156. Curran
Associates, Inc., 2013.
[18] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with
pruning, trained quantization and huffman coding. Fiber, 56(4):3–7, 2015.
[19] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient
neural network. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances
in Neural Information Processing Systems 28, pages 1135–1143. Curran Associates, Inc., 2015.
6
| 1cs.CV
|
arXiv:1711.04354v1 [math.AC] 12 Nov 2017
HYPERGRAPH ENCODINGS OF ARBITRARY TORIC IDEALS
SONJA PETROVIĆ, APOSTOLOS THOMA, MARIUS VLADOIU
Abstract. Relying on the combinatorial classification of toric ideals using their
bouquet structure, we focus on toric ideals of hypergraphs and study how they
relate to general toric ideals. We show that hypergraphs exhibit a surprisingly
general behavior: the toric ideal associated to any general matrix can be encoded
by that of a 0/1 matrix, while preserving the essential combinatorics of the original ideal. We provide two universality results about the unboundedness of degrees
of various generating sets: minimal, Graver, universal Gröbner bases, and indispensable binomials. Finally, we provide a polarization-type operation for arbitrary
positively graded toric ideals, which preserves all the combinatorial signatures and
the homological properties of the original toric ideal.
Introduction
Toric ideals associated to 0/1 matrices occupy a special place in the world of
toric ideals due to their applications to biology [23], algebraic statistics [21], integer
programming [8], matroid theory [14], combinatorics [18]. They are, by definition,
incidence matrices of hypergraphs, and in the special case when there are only two
nonzero entries in each column they are the incidence matrices of graphs. There is
an abundant literature for toric ideals of graphs, including [16, 20, 25, 26], and the
book [27], while toric ideals of hypergraphs were recently studied in [11] and [18].
Over the last decades, the rich combinatorial structure of toric ideals has been
an anchor for many investigations of their properties. With this literature in mind
in [19] we defined and studied the bouquet ideal of an integer matrix A and its
relationship to the combinatorial and algebraic structure of the toric ideal encoded
by A. Bouquets of a matrix naturally arise from an oriented matroid that records
the dependencies from the Gale transform of A. Specifically, in [19] we showed that
the bouquets of a matrix capture the essential combinatorics of the corresponding
toric ideal, leading to a classification of toric ideals with respect to the bouquet
structure. In addition, the bouquet construction represented the key step for a
combinatorial classification of strongly robust toric ideals and provides a technique
for constructing infinitely many examples of the prominent classes of toric ideals
with rich combinatorics, due to equality of various bases, that include unimodular
([2]), generic ([17]) and robust ideals ([3]).
In light of these results, here we focus on toric ideals of 0/1 matrices and study
their relationship to general toric ideals. Similarly to how the usual reference to
2010 Mathematics Subject Classification. 14M25, 13P10, 05C65, 13D02.
Key words and phrases. Toric ideals, Graver basis, universal Gröbner basis, hypergraphs, minimal generating sets, resolutions.
1
(Markov, Gröbner, Graver) bases of matrices refers to bases of corresponding toric
ideals, throughout the paper we refer to bases of hypergraphs when we think of bases
of toric ideals of incidence matrices of those hypergraphs. We refer the reader to
[22, 19] for general definitions of the above-mentioned bases together with the circuits
and indispensable binomials, to [22, 6] for some equivalent algebraic descriptions,
and to [22, 7] for the general hierarchy among them. As in [19], we will denote the
Graver basis of the matrix A by Gr(A), its set of circuits by C(A), and we recall
that since any element of KerZ (A) can be written as a (conformal) sum of elements
from Gr(A) then one can say that the essential combinatorial information of the
toric ideal of A is provided by Gr(A). For the remainder of the introduction, we
summarize and put into perspective our two main results.
The first main result says, roughly speaking, that the essential combinatorics of
arbitrary toric ideals is encoded by almost 3-uniform hypergraphs, that is, hypergraphs with edges of cardinality at most 3 and having at least one such edge:
Theorem 1 [Theorem 3.2: Hypergraph encodings of arbitrary toric ideals]. Given
any integer matrix A, there exists an almost 3-uniform hypergraph H = (V, E) such
that the toric ideal of the hypergraph IH is strongly robust and there is a bijective
correspondence between the elements of KerZ (A), Gr(A), and C(A) and KerZ (H),
Gr(H), and C(H), respectively.
Thus toric ideals of almost 3-uniform hypergraphs, that is toric ideals of 0/1matrices with at most three 1’s on each column are “as complicated” as any arbitrary
toric ideals with respect to its essential combinatorics, the Graver basis. Recall that
if H is a 0/1 matrix, the toric ideal IH is positively graded, that is, KerZ (H) ∩ Nn =
{0}. In interpreting this result, one has to take into account that by ‘arbitrary toric
ideals’ we do not mean arbitrarily positively graded toric ideals, but arbitrary in the
complete sense. In general, a non-positively graded toric ideal has infinitely many
minimal generating sets of different cardinalities and no indispensable binomials,
see [5, Theorems 4.18, 4.19], and its Graver basis represents the best combinatorial
information one could have on such a toric ideal. However, the fact that Graver
basis of any toric ideal is in a (precise defined way) one-to-one correspondence to
the Graver basis of a hypergraph shouldn’t mislead the reader that the latter one is
easy to describe. Here, it is important to note that while the Graver bases of toric
ideals of graphs have very special form and support structure and are completely
understood, see [20, 25], those of hypergraphs can be quite complicated ([18]) and not
so many things are known. Under these circumstances, “controlling” the essential
combinatorics of an arbitrary toric ideal by toric ideals of hypergraphs, Theorem 1
is the best one could hope for.
Theorem 1 has some interesting consequences, two of which we single out as
universality results for almost 3-uniform hypergraphs. These are also presented in
Section 3 (Corollaries 3.3 and 3.4) along with comments relating them to [9].
The second main result provides a polarization-type operation, saying that if IA
is positively graded then there exists a stable toric ideal of a hypergraph H of same
combinatorial complexity as IA . Stable toric ideals, also introduced and extensively
studied in [19], are those whose all of the bouquets are non-mixed, or in other words
2
for which passing to the bouquet ideal preserves all combinatorial information and,
in the positively graded case, even the homological information.
Theorem 2 [Theorem 4.2: Stable hypergraphs and positively graded toric ideals].
Let IA be an arbitrary positively graded nonzero toric ideal. Then there exists a
hypergraph H such that there is a bijective correspondence between the Graver bases,
all minimal Markov bases, all reduced Gröbner bases, circuits, and indispensable
binomials of IA and IH . Moreover the minimal graded free resolution of IA can be
obtained from the corresponding minimal graded free resolution of IH and viceversa,
and thus all homological data of IA is inherited from IH .
Note by comparison to Theorem 1 that since the toric ideal is positively graded
all of the combinatorics, that is all the distinguished special sets, of the toric ideal is
encoded by the toric ideal of hypergraph. The great bonus offered by this construction is that one can pass from an arbitrary matrix A with KerZ (A) ∩ Nn = {0} to a
0/1 matrix AH by preserving all the homological data, thus this theorem might be
also regarded as a polarization-type operation, see Remark 4.3. Theorem 2 also has
some interesting consequences: in particular, it implies that any classification problem about arbitrary positively graded toric ideals involving equality of bases can be
reduced to a problem about a toric ideal of a hypergraph defined by a 0/1-matrix.
For example, see the conjecture of Boocher et al. [4, Question 6.1] discussed at the
end of Section 4 in Remark 4.4.
The technical backbone of our results are special types of bouquets for 0/1 matrices, called bouquets with bases, that encode the basic building blocks of Graver
bases elements for hypergraphs. Section 2 offers a more technical motivation for this
construction and studies bouquets of 0/1 matrices in detail. Two main definitions
are those of a bouquet with basis and a monomial walk (Definitions 2.1 and 2.7).
We classify such bouquets in Theorem 2.3. There are two technical applications,
which solve two problems for hypergraphs: Theorem 2.10 offers a structural result
generalizing [18], while Proposition 2.13 generalizes a result from [10] and [25] from
graphs to uniform hypergraphs. In particular, Theorem 2.10 and Corollary 2.5 represent the key ingredients in the construction of the hypergraph H from Theorem 1.
For the convenience of the reader, we begin by revisiting the bouquet constructions
and the basic terminology from [19] next.
1. Revisiting the bouquet construction
Let K be a field and A = [a1 , . . . , an ] ∈ Zm×n be an integer matrix of rank r with
the set of column vectors {a1 , . . . , an }. We recall that the toric ideal of A is the
ideal IA ⊂ K[x1 , . . . , xn ] given by
+
−
IA = (xu − xu : u ∈ KerZ (A)),
where u = u+ − u− is the unique expression of an integral vector u as a difference
of two non-negative integral vectors with disjoint support, see [22, Section 4]. To
every integer matrix A one can associate its Gale transform G(A), which is the
n × (n − r) matrix whose columns span the lattice KerZ (A). We will denote the set
of ordered row vectors of the matrix G(A) by {G(a1 ), . . . , G(an )}. The vector ai is
3
called free if its Gale transform G(ai ) is equal to the zero vector, which means that
i is not contained in the support of any minimal generator of the toric ideal IA , or
any element in the Graver basis. A vector which is not free will be called non-free.
Definition 1.1 ([19, Definition 1.1]). The bouquet graph GA of IA is the graph on
vertices {a1 , . . . , an } whose edge set EA consists of those {ai , aj } for which G(ai ) is
a rational multiple of G(aj ) and vice versa. The connected components of the graph
GA are called bouquets.
It follows from the definition that the free vectors of A form one bouquet, which
we call the free bouquet of GA . A bouquet B which is not free is called non-free.
The non-free bouquets are of two types: mixed and non-mixed. A non-free bouquet
is mixed if contains an edge {ai , aj } such that G(ai ) = λG(aj ) for some λ < 0, and
is non-mixed if it is either an isolated vertex or for all of its edges {ai , aj } we have
G(ai ) = λG(aj ) with λ > 0, see [19, Lemma 1.2]. By slight abuse of notation, we
identify vertices of GA with their labels; that is, ai will be used to denote vectors in
the context of A and MA , and vertices in the context of GA .
Example 1.2. Let A be the matrix
3 0
1 1
3 0
7 1
6 0
0
0
1
2
2
4
4
0
4
0
5
5
0
3
0
0
0
0
1
0
1
2
0
1
1
.
A basis for KerZ (A) is given by (1, 2, −3, 3, −3, −6, 0) and (1, 2, −3, −2, 1, 2, 0). Thus
G(a1 ) = (1, 1), G(a2 ) = (2, 2), G(a3 ) = (−3, −3), G(a4 ) = (3, −2), G(a5 ) = (−3, 1),
G(a6 ) = (−6, 2) and G(a7 ) = (0, 0). Therefore the bouquet graph GA with the
vertex set {a1 , . . . , a7 } consists of the four bouquets B1 , . . . , B4 depicted below.
−
•
a1
a3
•
−
+ •a
2
B1
a6
•
•
a4
+
•
a5
B3
B2
•
a7
B4
Here, B1 is mixed, B3 is non-mixed, B2 is non-mixed (since G(a4 ) 6= (0, 0)) and
B4 is the free bouquet (since G(a7 ) = (0, 0)).
For each bouquet B there exist some bouquet-index-encoding vectors cB and
aB which record the bouquet’s types and linear dependencies. For the complete
technical definition of cB we refer the reader to page 9 of [19]; here we simply
summarize how to compute these vectors and offer an example below. If the bouquet
B is free then we set cB ∈ Zn to be any nonzero vector such that supp(cB ) = {i :
ai ∈ B} and with the property that the first nonzero coordinate is positive. For a
non-free bouquet B of A, consider the Gale transforms of the elements in B. All
the elements are nonzero and pairwise linearly dependent, therefore there exists a
4
nonzero coordinate j in all of them. Let gj = gcd(G(ai )j | ai ∈ B) and fix the
smallest integer i0 such that ai0 ∈ B. Then cB is the vector in Zn whose i-th
coordinate is 0 if ai ∈
/ B, and is εi0 j G(ai )j /gj if ai ∈ B, where εi0 j represents the
sign of the integer G(ai0 )jP
. With cB introduced the vector aB (see [19, Definition
1.7]), is defined as aB = ni=1 (cB )i ai ∈ Zm , where (cB )i is the i-th component of
the vector cB .
Detecting the type of a non-free bouquet B is important for our further considerations and we can read this information from the corresponding cB . Indeed, if B is a
non-free bouquet of A, then B is a mixed bouquet if and only if the vector cB has a
negative and a positive coordinate, see [19, Lemma 1.6]. Hence the non-free bouquet
is non-mixed if and only if the vector cB has all nonzero coordinates positive, taking
into account that by definition the first nonzero coordinate of cB is positive.
The matrix AB whose column vectors are the vectors aB corresponding to the
bouquets of A is called the bouquet matrix of A. There is a surprising general
connection between the matrix A and its bouquet matrix AB summarized in the
following theorem.
Theorem 1.3 ([19, Theorem A]). Let A = [a1 , . . . , an ] ∈ Zm×n and its bouquet
matrix AB = [aB1 , . . . , aBs ]. There is a bijective correspondence between the elements of KerZ (A) in general, and Gr(A) and C(A) in particular, and the elements
of KerZ (AB ), and Gr(AB ) and C(AB ), respectively. More precisely, this correspondence is defined as follows: for u = (u1 , . . . , us ) ∈ KerZ (AB ) then B(u) =
cB1 u1 + · · · + cBs us ∈ KerZ (A).
Example 1.4 (Example 1.2, continued). Let A be the matrix from Example 1.2. Let
us compute the bouquet-index-encoding vectors aBi s and cBi s. For cB1 we can choose
j to be either 1 or 2, while i0 = 1. Set j = 1, then g1 = gcd(G(a1 )1 , G(a2 )1 , G(a3 )1 )
and the nonzero coordinates of cB1 are
G(a1 )1
G(a2 )1
G(a3 )1
(cB1 )1 = ε11
= 1, (cB1 )2 = ε11
= 2, (cB1 )3 = ε11
= −3.
g1
g1
g1
Thus the corresponding bouquet vector is cB1 = (1, 2, −3, 0, 0, 0, 0) and consequently
aB1 = 1a1 + 2a2 − 3a3 = (3, 3, 0, 3, 0). Similarly one obtains cB2 = (0, 0, 0, 1, 0, 0, 0),
aB2 = a4 = (4, 4, 0, 4, 0), cB3 = (0, 0, 0, 0, 1, 2, 0), aB3 = 1a5 + 2a6 = (5, 5, 0, 5, 0),
cB4 = (0, 0, 0, 0, 0, 0, 1) and aB4 = a7 . Therefore we obtain the bouquet matrix
AB = [aB1 , aB2 , aB3 , aB4 ] ∈ Z5×4 . Now, the correspondence from Theorem 1.3 works
as follows. For example, to the vector (2, 1, −2, 0) ∈ KerZ (AB ) corresponds the
vector
B((2, 1, −2, 0)) = 2cB1 + 1cB2 − 2cB3 + 0cB4 = (2, 4, −6, 1, −2, −4, 0) ∈ KerZ (A).
Note that the bouquet ideal IAB is also defined by the matrix
3 4 5 0
,
0 0 0 1
and taking into account that aB4 is a free vector, IA is just the extension of the
defining ideal of the monomial curve parametrized by (t3 , t4 , t5 ) in the polynomial
ring in four indeterminates.
5
A subbouquet of GA is an induced subgraph of GA . One easily sees that subbouqets are cliques and maximal ones are the bouquets defined above. A is said to
have a subbouquet decomposition if there exists a family of subbouquets B1 , . . . , Bt
that is pairwise vertex-disjoint, the union of whose vertices equals {a1 , . . . , an }. A
subbouquet decomposition always exists if we consider, for example, the subbouquet decomposition induced by all of the bouquets. We define vectors cB and aB
for each subbouquet B as we did for the bouquets. The matrix associated to such a
subbouquet decomposition is called a subbouquet matrix. Crucially, Theorem 1.3 is
true even if we replace bouquets with proper subbouquets which form a subbouquet
decomposition of A.
Example 1.5 (Example 1.2, continued). The family of subbouquets B1′ , . . . , B5′
depicted below represents a subbouquet decomposition of A. Note that all of the
non-free bouquets B1′ , . . . , B4′ are non-mixed.
a6
•
•
a1
+ •a
2
′
B1
•
a3
•
a4
B′2
B′3
+
•
a5
B′4
•
a7
B′5
There is a natural question of whether there is an inverse construction to the one
described in Theorem 1.3: given a set of vectors a1 , . . . , as and c1 , . . . , cs that can
act as bouquet-index-encoding vectors can we construct a toric ideal IA whose s
(sub)bouquets are encoded by the given vectors? The answer is yes, comprised in
the next theorem, and will be often exploited in the next sections.
Theorem 1.6 ([19, Theorem B]). Let {a1 , . . . , as } ⊂ Zm be an arbitrary set of
vectors. Let c1 , . . . , cs be any set of primitive vectors, with ci ∈ Zmi for some
mi ≥ 1, each having full support and a positive first coordinate. Then, there exists a
matrix A with the subbouquet decomposition B1 , . . . , Bs , such that the ith subbouquet
is encoded by the following vectors: aBi = (ai , 0, . . . , 0) and cBi = (0, . . . , ci , . . . , 0),
where the support of cBi is precisely in the ith block, of size mi .
As a consequence of Definition 1.1 and the fact there are two types of non-free
bouquets, one has the following straightforward combinatorial classification of toric
ideals. A toric ideal may have: 1) all of its non-free bouquets non-mixed; 2) all of its
non-free bouquets mixed, or 3) the non-free bouquets are either mixed or non-mixed.
In the first case, the Theorem below says that all of the combinatorial information
is preserved when passing from IA to IAB . In addition, homological properties are
also preserved; see [19, Theorem 3.11].
Theorem 1.7 ([19, Theorem C]). Let IA be a stable toric ideal, that is all of the
non-free bouquets of A are non-mixed. Then the bijective correspondence between
the elements of KerZ (A) and KerZ (AB ) given by u 7→ B(u), is preserved when we
restrict to any of the following sets: Graver basis, circuits, indispensable binomials,
minimal Markov bases, reduced Gröbner bases (universal Gröbner basis).
6
This result holds in a more general setting, as explained in detail in [19, Section
3]. For example, even if IA is not stable, one cand find a (maximal) subbouquet decomposition such that all of its subbouquets are non-mixed, and whose subbouquet
matrix AB has the property that the bijective correspondence between KerZ (A) and
KerZ (AB ) from Theorem 1.7 is preserved. In particular, Example 1.5 gives such a
maximal subbouquet decomposition of the matrix A from Example 1.2.
The second class of toric ideals, described above, and whose bouquets are all
mixed have the following nice property.
Proposition 1.8 ([19, Corollary 4.4]). Suppose that every non-free bouquet of A is
mixed. Then IA is strongly robust, i.e. the following sets coincide:
• the Graver basis of A,
• the universal Gröbner basis of A,
• any reduced Gröbner basis of A,
• any minimal Markov basis of A.
However, despite the fact that almost all the examples of strongly robust ideals,
previously known in the literature, had all of the bouquets mixed, there are examples of strongly robust ideals which have both types of non-free bouquets, see [19,
Example 4.3. b)]. This lead us to raise the following question: Is it true that a
strongly robust ideal has at least one mixed bouquet?, see [19, Question 4.6]. Sullivant answered this question in the affirmative in [24] for the particular case of
codimension 2 strongly robust toric ideals.
2. Hypergraphs and bouquets with bases
As this section consists of three parts, we begin with a ‘roadmap’ to point out
the main definitions and results. As described in the Introduction, this section is
concerned with toric ideals of 0/1 matrices, which are, by definition, incidence matrices of hypergraphs. To understand this class of toric ideals, we need two main
definitions: a bouquet with basis in Definition 2.1, lacking in the previous literature and a monomial walk in Definition 2.7, recovering the meaning of monomial
walks on graphs from [20], [27]. The first two subsections are motivated by [18,
Problem 6.2], and they, essentially, partially solve its generalization. Theorem 2.3
classifies bouquets with bases, which are either free or mixed subbouquets, while the
running example shows its implications on identifying and constructing elements in
the Graver basis of a hypergraph. Proposition 2.10 shows that certain hypergraphs
based on a sunflower are bouquets with bases and, in particular, recovers the main
structural result of [18]. As another application of bouquets with bases, Proposition 2.13 generalizes a result by [10] and [25] for graphs to uniform hypergraphs and
showing that the universal Gröbner and Graver bases differ for complete uniform
hypergraphs with enough vertices. However, in general, there exist hypergraphs
that do not admit such a subbouquet structure. In this case, we provide general
constructions in Sections 3 and 4.
Let H = (V, E) be a finite hypergraph on the set of vertices V = {x1 , . . . , xm }
with edge set E = {E1 , . . . , En }, where each Ei is a subset of {x1 , . . . , xm }. When
7
|Ei | = 2 for all i and Ei 6= Ej for all i 6= j, we have a finite simple graph, and
we will specialize to this case to recall results about graphs. We denote by α E the
support (column) vector of an edge E, and thus the toric ideal IH is the toric ideal
α E1 , . . . , α En ]. Hereafter, for ease of notation, various bases of IH
of the matrix [α
will be referred to as bases of H; the reader may simply keep in mind that the
underlying toric matrix is the incidence matrix of the hypergraph H; for example,
Gr(H) := Gr(AH ) where AH is the vertex-edge incidence matrix of H.
2.1. Bouquets with bases. Let H = (V, E) be a hypergraph and U ⊂ V . We
define the multiset UE = {E ∩ U|E ∈ E, E ∩ U 6= ∅} and the set EU = {E ∈
E|E ∩ U 6= ∅}. Note that the multiset UE and the set EU have the same number
of elements: EU is the set of edges of H that intersect the vertex set U, while UE
consists of their restrictions to U.
Definition 2.1. The set of edges EU of H is called a bouquet with basis U ⊂ V if
the toric ideal of the multi-hypergraph (U, UE ) is principal, generated by an element
+
−
ec −ec with supp(c) = EU , and such that the first nonzero coordinate of c = c+ −c−
is positive. Here the toric ideal of (U, UE ) is contained in the polynomial ring with
variables indexed by the non-empty subsets ei := Ei ∩ U.
n
Moreover, for a bouquet with basis EU , we define the vector cEU ∈
PZ such that
(cEU )E = cE for any E ∈ EU and 0 otherwise, and the vector aEU = E∈EU cE α E ∈
Zm . Here cE denotes the coordinate of c corresponding to the edge E, and note
from the definition of c that cE 6= 0 for all E ∈ EU .
The following examples capture the subtleties of this definition and its similarity
to Definition 1.1.
Example 2.2. a) Consider the hypergraph H = (V, E) from Figure 1, with the set
of vertices V = {x, v1 , . . . , v22 } and whose set of edges E consists of the following 20
edges: E1 = {x, v1 , v2 }, E2 = {x, v3 , v4 }, E3 = {x, v5 , v6 }, E4 = {v1 , v3 , v5 }, E5 =
{v2 , v4 , v6 }, E6 = {x, v7 , v8 }, E7 = {x, v9 , v10 }, E8 = {x, v11 , v12 }, E9 = {x, v13 , v14 },
E10 = {v7 , v8 , v9 }, E11 = {v10 , v11 , v13 }, E12 = {v12 , v14 }, E13 = {x, v15 , v16 }, E14 =
{x, v17 , v18 }, E15 = {x, v19 , v20 }, E16 = {x, v21 , v22 }, E17 = {v16 , v18 , v20 }, E18 =
{v15 , v19 }, E19 = {v17 , v19 , v21 }, E20 = {v20 , v22 }.
The first bouquet with basis EU1 has five edges, the basis U1 is the set {v1 , . . . , v6 },
the vector c1 corresponds to the binomial generator e1 e2 e3 − e4 e5 of the toric ideal
of (U1 , U1E ), and thus cEU1 = (1, 1, 1, −1, −1, 0, . . . , 0) ∈ Z20 and
aEU1 = α E1 + α E2 + α E3 − α E4 − α E5 = (3, 0, . . . , 0) ∈ Z23 .
The second bouquet with basis EU2 has seven edges, the basis U2 is the set
{v7 , . . . , v14 }, the vector c2 corresponds to the binomial generator e6 e7 e8 e9 −e10 e11 e12
of the toric ideal of (U2 , U2E ), and thus the encoding vectors are
cEU2 = (0, 0, 0, 0, 0, 1, 1, 1, 1, −1, −1, −1, 0, . . . , 0) ∈ Z20
and
aEU2 =
9
X
i=6
α Ei −
12
X
α Ej = (4, 0, . . . , 0) ∈ Z23 .
j=10
8
The third bouquet with basis EU3 has eight edges, the basis U3 = {v15 , . . . , v22 },
the vector c3 corresponds to the binomial generator e13 e14 e215 e16 − e17 e18 e19 e20 of the
toric ideal of (U3 , U3E ), and thus cEU3 = (0, . . . , 0, 1, 1, 2, 1, −1, −1, −1, −1) ∈ Z20
and
20
X
αE15 + α E16 −
aEU3 = α E13 + α E14 + 2α
α Ej = (5, 0, . . . , 0) ∈ Z23 .
j=17
On the other hand, if A = AH is the incidence matrix of the hypergraph H with
columns α E1 , . . . , α E20 , then GA has three mixed non-free bouquets B1 , B2 , B3 . More
precisely, the first bouquet B1 corresponds to the vectors α E1 , . . . , α E5 , the second
bouquet B2 to the vectors α E6 , . . . , α E12 , and the third bouquet B3 to the vectors
α E13 , . . . , α E20 . Moreover, it turns out that cBi = cEUi and aBi = aEUi for all i.
v2
v4
•
v1
v8
•
•
•
•
v7
•
v3
v5
•
x
x
•
•
•
•
•
•
v10
v13
v11
•
v9
v16
v15
•
v12
•
v14
x
•
•
•
•
•
•
v17
v19
v21
•
•
•
v18
v20
v22
v6
Figure 1. Bouquets with bases EU1 , EU2 , EU3 .
(b) The tetrahedron is an example of a bouquet with basis, and the basis can be
chosen to be any facet. Let V = {v1 , v2 , v3 , v4 } and E = {E1 = {v2 , v3 , v4 }, E2 =
{v1 , v3 , v4 }, E3 = {v1 , v2 , v4 }, E4 = {v1 , v2 , v3 }}. If we consider as basis the set U1 =
{v1 , v2 , v3 } then U1E = {e1 = {v2 , v3 }, e2 = {v1 , v3 }, e3 = {v1 , v2 }, e4 = {v1 , v2 , v3 }}
+
−
and the toric ideal of (U1 , U1E ) is generated by the element ec − ec = e1 e2 e3 − e24 ,
with c = (1, 1, 1, −2). Therefore, cEU1 = (1, 1, 1, −2) and aEU1 = α E1 + α E2 +
αE4 = (0, 0, 0, 3). The same example has four different representations as a
α E3 − 2α
bouquet with basis U, each one of them having pairwise distinct vectors cEU and
aEU , respectively. More precisely, if U2 = {v1 , v2 , v4 } then cEU2 = (1, 1, −2, 1) and
aEU2 = (0, 0, 3, 0), if U3 = {v1 , v3 , v4 } then cEU3 = (1, −2, 1, 1) and aEU3 = (0, 3, 0, 0),
and if U4 = {v2 , v3 , v4 } then cEU4 = (2, −1, −1, −1) and aEU4 = (−3, 0, 0, 0). Note
αEi ) are zero vectors for all i.
that IH is the zero ideal and the Gale transforms G(α
Therefore, the bouquet graph of IH has only the free bouquet B consisting of all
vertices α E1 , . . . , α E4 . In particular, by definition of cB and aB we observe that we
can choose these vectors to be any of the 4 pairs of vectors obtained before as cEU
and aEU .
Given the apparent similarity with Definition 1.1 and in light of the two situations discussed in Example 2.2, one would expect a certain relationship between a
bouquet with basis of a hypergraph H and a bouquet of its incidence matrix. Even
more, one may ask whether cEU and aEU , the encoding vectors of a bouquet with
basis, are analogous to cB and aB as defined in Section 1. The following result
clarifies the ‘bouquet with basis’ terminology. As we will see, a bouquet with basis
corresponds to a subbouquet of the incidence matrix of the hypergraph, where the
9
correspondence is the natural one associating to a set of edges of a hypergraph the
set of corresponding support column vectors of its incidence matrix. Thus, from
now on, by abuse of notation, we will identify the bouquet with basis EU with the
corresponding subbouquet of the incidence matrix. Furthermore, we also note that
cEU matches the definition of cB from Section 1, and the same holds for aEU .
Theorem 2.3. A bouquet with basis of the hypergraph H = (V, E) is either a free
subbouquet or a mixed subbouquet of the incidence matrix of H.
Proof. Let EU be a bouquet with basis for some U ⊂ V , and set EU = {E1 , . . . , Es }.
Furthermore, denote by M ∈ Zm×n the incidence matrix of the hypergraph (V, E).
We may arrange M so that its rows are indexed first by vertices in U and then
vertices in V \U, while its columns are indexed first by the edges E1 , . . . , Es and next
by the remaining edges, if any. Note that the submatrix of M corresponding to the
rows indexed by U and the first s columns, denoted by MU , is the incidence matrix of
the multi-hypergraph (U, UE ), while the submatrix of M corresponding to the rows
indexed by U and the rest of the columns is 0. Finally, denote by G = (gij ) ∈ Zn×r
the Gale transform of M, and according to the labeling of the columns of M its first
αE1 ), . . . , G(α
α Es ). By definition of the Gale transform, column j of G
s rows are G(α
is in the kernel of M. Therefore (g1j , . . . , gsj ) ∈ Ker MU and since EU is a bouquet
with basis then (g1j , . . . , gsj ) is a multiple of c = (cE1 , . . . , cEs ). Thus gij = λj cEi for
αEi ) = cEi (λ1 , λ2 , . . . , λr ) for all i ∈ {1, . . . , s}.
all i, j and implicitly we obtain G(α
Then it follows at once that α E1 , . . . , α Es belong to the same subbouquet B of M.
In addition, since supp(c) = EU then cEi 6= 0 for all i ∈ {1, . . . , s}. Thus there are
α Ei ) = 0 for all i or G(α
αEi ) 6= 0 for all i. In the first case we
two possibilities: G(α
obtain that B is a free subbouquet, while in the second B is a non-free subbouquet.
Moreover, the toric ideal of (U, UE ) being positively graded implies that the vector
cEU has at least one positive and one negative coordinate, and by the description
of a non-free mixed bouquet B in terms of cB from Section 1, we obtain that B is
mixed.
Let us point out that a bouquet with basis can be a proper subbouquet of a
bouquet. Indeed, let A′ be the submatrix of the incidence matrix of the hypergraph
H from Example 2.2(a) corresponding to the first twelve columns. As it was already
noticed the first five edges form a bouquet with basis, and the other seven edges
form another bouquet with basis. In contrast, one can easily see that the bouquet
graph of A′ has one mixed bouquet - consisting of all column vectors of A′ , and thus
the two bouquets with bases are proper subbouquets.
The previous theorem shows that bouquets with bases are always subbouquets.
The natural converse question arises: Can there exist (sub)bouquets of the incidence
matrix of a hypergraph which are not bouquets with bases? The answer is yes, and
as an example consider the complete graph K4 on four vertices, whose incidence
matrix is the submatrix corresponding to the first 6 columns of A from [19, Example
1.4]. As it was shown there K4 has three non-mixed bouquets, each one being an
edge. On the other hand, K4 does not have any bouquets with bases, since by
10
Theorem 2.3 this would imply the existence either of a mixed non-free subbouquet
or a free subbouquet.
The main consequence of Theorem 2.3 is that if the edge set of a hypergraph
H can be partitioned into bouquets with bases, then the toric ideal IH is easier to
describe, via Theorem 1.3. The following example captures this remark.
Example 2.4 (Example 2.2(a), continued). The hypergraph H has three bouquets
with bases EU1 , EU2 , EU3 which partition E. Therefore, the subbouquet ideal of AH is
given by the toric ideal of the matrix whose columns are aEU1 = (3, 0, . . . , 0), aEU2 =
(4, 0, . . . , 0), aEU3 = (5, 0, . . . , 0) ∈ Z23 , which is the same as the toric ideal of the
monomial curve A = (3 4 5). Computing with [1] we obtain that the Graver basis of
IA consists of seven elements (4, −3, 0), (1, −2, 1), (3, −1, −1), (2, 1, −2), (5, 0, −3),
(1, 3, −3), (0, 5, −4). Therefore, by Theorem 1.3, the Graver basis of the toric ideal
of the hypergraph consists of seven elements. For example, (1, −2, 1) corresponds
to the following Graver basis element of IH :
cEU1 −2cEU2 + cEU3 = (1, 1, 1, −1, −1, −2, −2, −2, −2, 2, 2, 2, 1, 1, 2, 1, −1, −1, −1, −1)
and it encodes the binomial
2
2
2
2
E1 E2 E3 E10
E11
E12
E13 E14 E15
E16 − E4 E5 E62 E72 E82 E92 E17 E18 E19 E20 .
This binomial corresponds to the primitive monomial walk (see Definition 2.7) depicted in Figure 2, where the three copies of the vertex x should be identified, but
are shown separately for better visibility.
•
•
•
•
•
• x
•
•
x •
•
•
•
•
•
•
•
•
x •
•
•
•
•
•
•
•
Figure 2. A primitive monomial walk of H
In particular, combining Theorem 2.3 with Proposition 1.8 we obtain the following:
Corollary 2.5. If the set of edges of a hypergraph H can be partitioned into bouquets
with bases then IH is strongly robust.
Remark 2.6. As a second application of Proposition 1.8 we give a new class of
hypergraphs which satisfy the conclusion of Corollary 2.5, and whose building blocks
are not necessarily bouquets with bases. Let H = (V, E) be a hypergraph such that
there exists U ⊂ V with the property that U ∩ E 6= ∅ for all E ∈ E and every vertex
of U belongs to exactly two edges. Denote by U = {v1 , . . . , vt } ⊂ {v1 , . . . , vm } = V ,
E = {E1 , . . . , En } and let B be a non-free bouquet of AH (if such a B does not exist
then IH = 0 and we are done). This implies that there exists an i ∈ [n] such that
α Ei ∈ B. By definition of H there exists vj ∈ Ei ∩ Ek with j ≤ m and k 6= i. Then
11
the vector u ∈ Zm , whose only nonzero coordinate is 1 on the j-th position satisfies
the following equalities
u · α Ei = 1, u · α Ek = 1, u · α El = 0 for all l 6= i, k.
α Ei ) +
Therefore the vector cik := (u · α E1 , . . . , u · α En ) has support {i, k} and G(α
α Ek ) = 0, see [19, Remark 1.3]. Since B is non-free, G(α
α Ei ) = −G(α
α Ek ) 6= 0, and
G(α
thus B is mixed. Therefore by Proposition 1.8 we obtain the desired conclusion, and
in particular we recover [11, Proposition 4.5]. Imposing in the definition of H that
U = V , thus making H a 2-regular hypergraph, we also recover [11, Proposition
4.2].
Of course, general hypergraphs do not admit a partition of their edge sets into
bouquets with bases, let alone mixed subbouquets, as it was shown earlier for K4 .
Infinitely many such examples can be constructed; see Section 4 for details.
2.2. Sunflowers. In this subsection we identify some interesting examples of bouquets with bases, namely, the so-called ‘sparse bouquets’ from [18] (of which there
was no formal definition!). These hypergraphs are built on sunflowers. The sunflower is highly structured and useful in the hypergraph literature; for example, it
is guaranteed to occur in hypergraphs with large enough edge sets, independently
of the size of the vertex set (see e.g. [13].)
General properties of bouquets with bases studied in the previous subsection allow us to not only recover the theorems about existence of Graver basis elements,
but also to: 1) describe completely their Graver basis elements by identifying the
bouquet ideal, thus the aB ’s, 2) show that these sunflowers are actually strongly
robust, and 3) identify Graver basis elements of any hypergraphs which have sunflowers as subhypergraphs. In addition, unlike [18], we do not specialize to uniform
hypergraphs.
In Section 1 it was recalled in Theorem 1.3 that the bouquet graph of A encodes
the Graver basis of IA . On the other hand, [18] showed that the Graver basis of IH
is encoded by primitive monomial walks on the hypergraph. The two concepts are
consolidated in the following definition.
Definition 2.7. Let (Eblue , Ered ) be a multiset collection of edges of H = (V, E). We
denote by degblue (v) and degred (v) the number of edges of Eblue and Ered containing
the vertex v, respectively. We say that (Eblue , Ered ) are balanced on U ⊂ V if
degblue (v) = degred (v) for each vertex v ∈ U.
The vector aEblue ,Ered = (degblue (v) − degred (v))v∈V is called the vector of imbalances of (Eblue , Ered ) and its support is contained in the complement of U in V . If
aEblue ,Ered = 0 then we say that (Eblue , Ered) is a monomial walk 1.
Q
Q
Every monomial walk encodes a binomial fEB ,ER = E∈EB E − E∈ER E in IH .
A monomial walk (EB , ER ) is said to be primitive if there do not exist proper submultisets E ′ B ⊂ EB and E ′ R ⊂ ER such that (E ′ B , E ′ R ) is also a monomial walk. The
1For
completeness, note that the support of a monomial walk, considered as a multi-hypergraph,
was called a monomial hypergraph in [18], but this definition does not make an appearance in our
results. Instead, we focus on bouquets with bases and the corresponding UE and EU .
12
toric ideal IH is generated by binomials corresponding to primitive monomial walks,
see [18, Theorem 2.8].
Remark 2.8. Let H = (V, E) be a hypergraph, U ⊂ V such that EU is a bouquet
with basis. Since the toric ideal of the (multi)hypergraph (U, UE ) is principal gen+
−
erated by ec − ec with supp(c) = EU the primitive monomial walk (UEblue , UEred )
encoded by (U, UE ) is the following: UEblue is the multiset consisting of the edges
e1 , ei2 , . . . , eir with multiplicities c1 , ci2 , . . . , cir , respectively, while UEred is the multiset consisting of the edges eir+1 , . . . , eit with multiplicities −cir+1 , . . . , −cit , respectively. Here c1 , ci2 , . . . , cir are the positive coordinates of c, while cir+1 , . . . , cit are
the remaining coordinates of c, all negative. If we consider Eblue to be the multiset
collection of edges E1 , Ei2 , . . . , Eir with multiplicities c1 , ci2 , . . . , cir respectively, and
Ered the multiset collection of edges Eir+1 , . . . , Eit with multiplicities −cir+1 , . . . , −cit
respectively then (Eblue , Ered ) are balanced on U. Moreover, notice that the vector of
imbalances aEblue ,Ered equals aEU , and as explained before cEU determines (Eblue , Ered).
For example, considering the bouquet with basis EU3 from Example 2.2(a), the toric
ideal of (U3 , U3E ) was generated by e13 e14 e215 e16 − e17 e18 e19 e20 , and thus the corresponding Eblue and Ered are depicted with the corresponding colors in the rightmost
part of Figure 2.
Recall that a matching on a hypergraph H = (V, E) is a subset M ⊂ E of pairwise
disjoint edges. A matching is called perfect if it covers all the vertices of the hypergraph. A hypergraph is said to be connected if its primal graph is connected, where
the primal graph has the same vertex set as the hypergraph and an edge between
any two vertices contained in the same hyperedge.
A hypergraph H = (V, E) is a sunflower if, for some vertex set C, Ei ∩ Ej = C
for all edges Ei , Ej ∈ E, i 6= j and C ( E for all edges E ∈ E. The set of vertices of
C is called the core of the sunflower, and each Ei is called a petal. A matched-petal
sunflower2 is a hypergraph consisting of a sunflower and a perfect matching on the
non-core vertices. Note that the set of edges of a matched-petal sunflower partitions
into the edges of sunflower, i.e. petals, and the edges of the matching, while its set of
vertices is just the set of the vertices of the sunflower. A matched-petal sunflower H
is called connected if the (multi)hypergraph H − C is connected, where C represents
the core vertices. Here, H − C is the (multi)hypergraph consisting of the restricted
sunflower: (V \ C, E ′) where E ′ = {E \ C : E ∈ E}, and the edges of the perfect
matching of H.
A matched-petal partitioned-core sunflower is a hypergraph H consisting of a collection of vertex-disjoint sunflowers S1 , S2 , . . . , Sl and a perfect matching on the
union of non-core vertices, that is ∪i (Si \ Ci ). A matched-petal partitioned-core
sunflower is called connected if H − C is connected, where C = ∪i Ci . A matchedpetal relaxed-core sunflower is a hypergraph H consisting of a collection of sunflowers
2 Let us also relate this definition to [18], where the authors defined a monomial sunflower :
the multi-hypergraph with aEblue ,Ered = 0 whose support is a matched-petal sunflower. However,
there is an important distinction: not every matched-petal sunflower is the support of a monomial
sunflower. (As a monomial sunflower is an example of a monomial hypergraph, this definition also
isn’t used in this manuscript.)
13
S1 , S2 , . . . , Sl with cores C1 , . . . , Cl respectively, which may only intersect at their
cores, and a perfect matching on the union of the non-core vertices of the sunflowers.
A matched-petal relaxed-core sunflower is called connected if the (multi)hypergraph
H − C is connected, where C = ∪i Ci .
Remark 2.9. The definitions above resemble various monomial walks (based on
various types of sunflowers) introduced in [18]. In contrast, here we allow hypergraphs to be non-uniform and consider the supporting sunflowers as sets instead of
multisets.
As far as the standard terminology is concerned, note that a matched-petal sunflower is a particular case of matched-petal partitioned-core sunflower, which in turn
is a particular case of matched-petal relaxed-core sunflower. Each of the first two
subhypergraphs from Figure 1 are connected matched-petal sunflowers, while the
third one is not. If we identify the vertices x1 , x2 and x3 from Figure 3 then the
hypergraph consisting of the three depicted connected matched-petal sunflowers is
a non-connected matched-petal sunflower.
Theorem 2.10. Every connected matched-petal relaxed-core sunflower is a bouquet
with basis, where the basis consists of the non-core vertices.
Proof. Let H be a connected matched-petal relaxed-core sunflower consisting of l
sunflowers S1 , . . . , Sl with Ci being the core vertices of sunflower Si for all i, and
let U = ∪li=1 Si \ Ci . Note that by definition the set of edges E of H partitions into
the set of petals of the sunflowers S1 , . . . , Sl , labeled E1 , . . . , Et , and the edges of
the matching, labeled Et+1 , . . . , Ek . In order to prove that EU is a bouquet with
basis, we note first that the vector c = (1, 1, . . . , 1, −1, . . . , −1) ∈ Zk corresponds to
a binomial in the toric ideal of (U, UE ), where the number of 1’s equals the number
of petals, while the number of −1’s equals the number of edges of the matching.
Assume that u = (u1 , . . . , ut , ut+1 , . . . , uk ) is an arbitrary vector corresponding to a
binomial in the toric ideal of (U, UE ). It remains to prove that u1 = · · · = ut and
ut+1 = · · · = uk = −u1 . Let E, E ′ ∈ UE be two different edges, restrictions of two
petals and thus non-empty, and let v ∈ E, v ′ ∈ E ′ .
Since H is connected, there exists a path in the primal graph of (U, UE ) from v to
v ′ , that is v = v0 , v1 , . . . , vr = v ′ . We construct inductively a sequence of edges of UE :
let s1 be the largest number such that vs1 ∈ E := E1 . Since E, E ′ are restrictions
of petals then E ∩ E ′ = ∅, and thus s1 < r. By definition of s1 , there exists an
edge of UE , say E2 , such that vs1 ∈ E1 ∩ E2 . Let s2 be the largest number such that
vs2 ∈ E2 . If s2 = r then E ′ = E2 and we stop, otherwise we continue. In this way
we obtain a sequence of edges E = E1 , . . . , Ep = E ′ of UE for some p ≥ 2.
Denote by uij the coordinate of u corresponding to Ej for all j = 1, . . . , p (that is
|uij | is the exponent of Ej in the binomial). Since the binomial corresponding to u
is in the toric ideal of (U, UE ), every vertex of U is balanced. Moreover, since every
vertex of U belongs to exactly two edges, vsj being balanced implies that uij+1 = −uij
for all j < p. Thus we obtain uij = (−1)j+1 u1 for all j = 1, . . . , p, and in particular,
if Ej is the restriction of a petal then Ej+1 is an edge of a matching. Therefore,
we obtain that for any distinct edges E, E ′ of UE the corresponding coordinates of
14
u are either equal or negatives of each other, with equality holding if and only if
E, E ′ are simultaneously either restrictions of petals or edges of the matching M.
Hence we get that u = (u1 , . . . , u1 , −u1 , . . . , −u1 ), which implies that the toric ideal
+
−
of (U, UE ) is principal, and generated by ec − ec . Thus EU is a bouquet with basis,
as desired.
Since the connected components of a matched-petal relaxed-core sunflower are
connected matched-petal relaxed-core sunflowers, then by Theorem 2.10 and Corollary 2.5 we obtain the following:
Proposition 2.11. Let H be a matched-petal relaxed-core sunflower. Then IH is
strongly robust.
In particular, one recovers [18, Theorem 4.12] and implicitly [18, Proposition 4.5]
and [18, Proposition 4.9]. To see this, we first identify the subbouquets of a matchedpetal relaxed-core sunflower H and their corresponding a-vectors. If H1 , . . . , Ht are
the connected components of H, that is matched-petal relaxed-core sunflowers with
the sets of core vertices C1 , . . . , Ct , then denote by C = {v1 , . . . , vs } the union ∪i Ci
of core vertices of H. Moreover, we label the edges such that petals are labeled first,
while the edges of the matching are labeled last. By Theorem 2.10, H1 , . . . , Ht are
bouquets with bases, the bases are the sets of non-core vertices, and the vectors
aHi can be computed as vectors of imbalances induced by cHi , as explained in
Remark 2.8. Therefore, if we label the rows of the incidence matrix of H first
according to the vertices from C, then it follows from Remark 2.8 that for each j
the vector aHj has at most the first s components nonzero and for each i = 1, . . . , s
the i-th coordinate of aHj is equal to dij , where dij is the number of petals of Hj
containing the vertex vi . Now it is obvious via Theorem 1.3 that IH 6= 0 if and only
if ID 6= 0, where D is the matrix (dij ) ∈ Zs×t , and their Graver basis are in bijective
correspondence, and this is essentially the content of [18, Theorem 4.12].
Example 2.12. Consider H to be the matched-petal relaxed-core sunflower whose
three connected components H1 , H2 , H3 are depicted below, and with cores Ci =
{xi } for i = 1, 2, 3. Note that each Hi is a connected matched-petal sunflower, and
thus a bouquet with basis by Theorem 2.10. If xi 6= xj for all i 6= j then the union of
core vertices is C = {x1 , x2 , x3 } and the matrix (dij ) ∈ Z3×3 is the displayed matrix
D1 . We recall from the previous paragraph that dij is the number of petals of Hj
containing xi . Thus in this case IH = 0.
3 0 0
3 4 0
0 4 0
D1 =
D2 =
D3 = 3 4 5
0 0 5
0 0 5
If x1 = x2 6= x3 then the union of core vertices is C = {x1 , x3 } and we obtain the
corresponding matrix D2 ∈ Z2×3 . Therefore in this case IH 6= 0 is principal. Finally,
if x1 = x2 = x3 then C = {x1 } and the matrix D3 is given above.
Specializing the previous discussions to a matched-petal sunflower H with t ≥ 1
connected components one obtains the following classification. The toric ideal IH =
0 if and only if t = 1, that is H is a connected matched-petal sunflower. The toric
15
•
•
•
•
•
• x1
•
•
x2 •
•
•
•
•
•
•
•
•
•
x3 •
•
•
•
•
•
•
•
•
Figure 3. Matched-petal sunflowers
ideal IH 6= 0 if and only if t > 1, in which case the subbouquet ideal is just the toric
ideal of the monomial curve (d1 . . . dt ), where di represents the number of petals of
the i-th connected component, containing the core.
2.3. Graver basis elements of hypergraphs. In cases of interest, it may happen
that all (or almost all) bouquets are singletons, seemingly implying that the bouquet
construction does not offer anything. However, this is not the case since by [22,
Proposition 4.13] if we restrict to a submatrix C of A obtained by deleting some
of the columns of A then the Graver basis, universal Gröbner basis and circuits
of C are included in the corresponding ones of A. Based on this and the bouquet
techniques we construct in Proposition 2.13 an element in the Graver basis of the
uniform complete hypergraph with large enough number of vertices, which is not in
the universal Gröbner basis.
In [10, 25] it was shown that the universal Gröbner basis and the Graver basis of
the toric ideal of the complete graph Kn are identical for n ≤ 8 and differ for n ≥ 9.
As one application of the bouquet technique, we show that the universal Gröbner
basis and the Graver basis of the toric ideal of the complete d-uniform hypergraph
Knd differ for n ≥ (d+1)2 , by giving a non trivial example of an element in the Graver
basis which does not belong to the universal Gröbner basis. First note that all of
the bouquets of the complete d-uniform hypergraph Knd , for large n, are singletons.
Then, restricting to a subhypergraph H of Knd , as explained before, we can apply
the bouquet techniques to find an element from the Graver basis of H that belongs
also to the Graver basis of Knd .
We consider a hypergraph Hd+1 = (V, E) with (d + 1)2 vertices V = {vij |1 ≤ i, j ≤
d + 1} and (d + 2)(d + 1) edges: let Ej = {vk1 |1 ≤ k ≤ d + 1, k 6= j} for 1 ≤ j ≤ d + 1
and Eij = {vik |1 ≤ k ≤ d + 1, k 6= j} for 1 ≤ i, j ≤ d + 1. The hypergraph H4 , for
d = 3, is depicted in Figure 4.
Note that Hd+1 is a subhypergraph of Knd for n ≥ (d + 1)2 . In addition the set of
edges of Hd+1 partitions into d+1 bouquets with bases EU1 , . . . , EUd+1 and d+1 single
edges E1 , . . . , Ed+1 . Here EUi = {Eij |1 ≤ j ≤ d + 1} and Ui = {vij |2 ≤ j ≤ d + 1}
for all i = 1, . . . , d + 1, while the principal
generator of the toric ideal of the hyQ
pergraph (Ui , UiE ) is the binomial j6=1 eij − ed−1
i1 , see Definition 2.1. Moreover,
the incidence matrix of Hd+1 has the rows indexed by the vertices in the following
way: v11 , v21 , . . . , vd+1,1 , v12 , v22 , . . . , vd+1,d+1 and the columns indexed by the edges
in the following way E11 , E12 , . . . , E1,d+1 , E21 , E22 , . . . , Ed+1,d+1 , E1 , . . . , Ed+1 . With
16
respect to this labeling of the edges and similarly to the computations from Example 2.2(b) we obtain that cEU1 = (d − 1, −1, . . . , −1, 0, . . . , 0), cEU2 = (0, d −
1, −1, . . . , −1, 0, . . . , 0), and so on cEUd+1 = (0, . . . , 0, d − 1, −1, . . . , −1, 0) are vectors of Z(d+2)(d+1) , where 0 ∈ Zd+1 represents the vector with all coordinates zero.
2
By Remark 2.8 the vector aEUi ∈ Z(d+1) has all coordinates zero except the one
corresponding to the vertex vi1 , which equals −d. For the singleton subbouquets
E1 , . . . , Ed+1 , the encoding vectors are: cE1 = (0, . . . , 0, 1, 0, . . . , 0), . . . , cEd+1 =
(0, . . . , 0, 0, . . . , 0, 1), while aEi = α Ei for all i = 1, . . . , d + 1.
v13
v14
v12
v24
v44
v11
v21
v43
v23
v41
v31
v22
v32
v42
v34
v33
Figure 4. H4
For the rest of this subsection we use for simplicity the binomial representation
for an element in the toric ideal of a hypergraph instead of the vector representation.
It is easy to see that the following three types of binomials belong to the toric ideal
of the complete d-uniform hypergraph Knd : a) d + 1 binomials of the form
Y
Y
d−1
gi := Eid−1
Eij − Ei1
Ej ,
j6=1
j6=i
b) d + 1 binomials of the form
YY
hi :=
Eij − Eld
i6=l j6=1
and c) one binomial of the form
YY
i
Eij −
j6=1
Y
d−1
Ei1
,
i6=l
Y
Ei
i
Y
d−1
Ei1
.
i
Furthermore, it can be shown that they are all elements of the Graver basis of Knd ,
although we will prove next just for the last binomial.
Proposition 2.13. The universal Gröbner basis of the toric ideal of the complete
d-uniform hypergraph Knd differs from the Graver basis for n ≥ (d + 1)2 .
Proof. First we prove that the binomial
Y Y
YY
d−1
Eij −
Ei
Ei1
f=
i
j6=1
i
17
i
does not belong to the universal Gröbner basis of the toric ideal of the complete
d-uniform hypergraph Knd . We argue by contradiction. Assume that there exists a
monomial order > such that the binomial
f belongs to the reduced
basis
Q Gröbner
Q d−1
d−1 Q
with respect to the order >. Since Ei1
E
divides
properly
E
E
j6=i j
i i
i i1 , it
follows that
Y
Y
d−1
(1)
Eid−1
Eij > Ei1
Ej .
j6=1
j6=i
d−1 Q
Indeed, if this is not the case then in< (gi ) = Ei1
E . From the previous
Q
Q d−1j6=i j
divisibility it follows that: 1) if in< (f ) = Qi Q
Ei i Ei1 then we contradict that f
is in the Gröbner basis, or 2) if in< (f ) = i j6=1 Eij then we contradict that f is
reduced.QThus
(1).
Q we obtain the inequality
Q Q
Also i6=l j6=1 Eij divides i j6=1 Eij , and similarly to the proof of (1) it can
be shown that
Y
YY
d−1
(2)
Ei1
.
Eij < Eld
i6=l
i6=l j6=1
Taking the product of inequalities (1) when i runs from 1 to d + 1 and canceling
common terms we obtain
YY
Y Y
d−1
Eij >
Ei
Ei1
.
i
j6=1
i
i
Similarly, taking the product of inequalities (2) we obtain
Y Y
YY
d−1 d
Eij )d < ( Ei
Ei1
) ,
(
i
a contradiction.
In order to prove that
enough to prove via [22,
its subhypergraph Hd+1 .
with bases EU1 , . . . , EUd+1
2
AB ∈ Z(d+1) ×2(d+1)
j6=1
i
i
the binomial f belongs to the Graver basis of Knd it is
Proposition 4.13] that f belongs to the Graver basis of
The partition of the edges of Hd+1 into d + 1 bouquets
and d + 1 single edges E1 , . . . , Ed+1 induces the matrix
−d 0
0 −d
AB = [aEU1 , . . . , aEUd+1 , aE1 , . . . , aEd+1 ] =
0
0
0
0
... 0 0
... 0 1
..
.
. . . −d 1
... 0 0
1 ...
0 ...
..
.
1 ...
0 ...
1
1
,
0
0
where aEi = α Ei for all i, the first d + 1 rows are indexed after the vertices
v11 , . . . , vd+1,1 and the last row represents block matrix 0 ∈ Zd(d+1)×2(d+1) . As it
was noticed in the comments prior to this proposition the binomial f corresponds
to the vector
v = (1 − d, 1, . . . , 1, 1 − d, 1, . . . , 1, . . . , 1 − d, 1, . . . , 1, −1, −1, . . . , −1) ∈ Ker(AHd+1 ).
18
Applying Theorem 1.3 we have a bijective correspondence between Ker(AB ) and
Ker(AHd+1 ) and given by
B(u1 , . . . , u2d+2 ) =
d+1
X
cEUi ui +
i=1
d+1
X
cEi ud+1+i .
i=1
Therefore replacing the formulas obtained before for cEUi and cEi , for all i =
1, . . . , d + 1, we obtain that v = B(u) where u = (1, . . . , 1, −1, . . . , −1) ∈ Z2d+2
with equally many 1’s and −1’s. By Theorem 1.3 it is enough to prove that u belongs to the Graver basis of AB to conclude that v (and thus f ) is in the Graver
basis of Hd+1 . Assume by contradiction u = u+ − u− is not in the Graver basis of
AB . Then there exists 0 6= w ∈ Ker(AB ) such that w+ ≤ u+ and w− ≤ u− , with at
least one inequality strict. In terms of coordinates this can be restated: wi ∈ {0, 1}
for all i = 1, . . . , d + 1 and wi ∈ {−1, 0} for all i = d + 2, . . . , 2d + 2, and with at
least one coordinate zero. If w+ = u+ then since w ∈ Ker(AB ) we obtain w− = u− ,
a contradiction. Otherwise there exists i ∈ {1, . . . , d + 1} such that wi = 0, and
without loss of generality we may assume that w1 = 0. Since w ∈ Ker(AB ) then
wd+3 + · · · + w2d+2 = 0, and thus wd+3 = . . . = w2d+2 = 0. If wd+2 = 0 then w = 0,
a contradiction, so we necessarily have wd+2 = −1. This leads to −dwi = 1 for all
i = 1, . . . , d + 1 a contradiction to w ∈ Z2d+2 . Therefore u belongs to the Graver
basis of AB , and we are done.
3. Complexity of hypergraphs
It is well-known that the elements of the Graver basis of a toric ideal of a graph
are of a rather special form: the exponents of each variable in an element of the
Graver basis is either one or two, see, for example, [20] and references therein, and
are completely determined by their support. In other words, one can not find two
elements in the Graver basis of a toric ideal of a graph with the same support. In
[18] it was shown that exponents in the elements of the Graver basis of a toric ideal
of a hypergraph can be arbitrarily high, and are not uniquely determined by their
supports, see Example 2.4. Thus the natural question is: how complicated are the
Graver basis, the universal Gröbner basis or a Markov basis of a hypergraph?
The bouquet ideal technique can be used to prove that toric ideals of hypergraphs
are as complicated as toric ideals in general. Namely, for any toric ideal, there
exists a toric ideal of a hypergraph with “worse” Graver basis, universal Gröbner
basis and Markov bases, as well as circuits. “Worse” means that the corresponding
set for the toric ideal of a hypergraph has at least the same cardinality and elements
of higher degrees than the corresponding elements of the toric ideal. In [9, Theorem
2.1] the authors show that for a particular matrix A, namely the one corresponding
to the no-three-way interaction model in statistics, elements in a minimal Markov
basis can be arbitrarily complicated as two of the three dimensions of the underlying
table grow to infinity. Such matrices are, in fact, incidence matrices of 3-uniform
hypergraphs. In contrast, Theorem 3.2 implies that the complexity of the whole
Graver basis (resp. minimal Markov or Universal Gröbner basis) of any matrix A
can be captured by an almost 3-uniform hypergraph, that is a hypergraph whose
19
edges have cardinality at most 3. In Section 4 there will be a stronger connection
but for matrices A defining positively graded toric ideals.
We start first with an example, which shows the details of the general construction.
Example 3.1. Let
A=
−1 −1 2 2
−2 2 −1 0
,
be the matrix whose columns are denoted by a1 , . . . , a4 . We construct a hypergraph H = (V, E) such that the bouquet graph of H has four bouquets with bases
EU1 , . . . , EU4 , and with the property that the corresponding subbouquet ideal (associated to the vectors aEU1 , . . . , aEU4 ) is IA . The latter will follow since the vectors aEU1 , . . . , aEU4 are “essentially” the vectors a1 , . . . , a4 , in the sense that each
aEUi ∈ Z|V | is just the natural embedding of ai in Z|V | with the other coordinates 0.
The construction of H is carried out in three steps.
Step 1. Every non-zero entry of the matrix A is used to construct a sunflower,
which will be the building blocks of the desired hypergraph. Precisely, for a positive
entry λ of the matrix A we consider a 3-uniform sunflower with one core vertex
and |λ| petals, while for a negative entry λ we consider the sunflower with one core
vertex and |λ| + 1 petals, which is almost 3-uniform, meaning that only one petal
has two vertices, the other have three. In the particular case of our example the
matrix A has three different nonzero entries −1, −2, 2, and thus we have two almost
3-uniform sunflowers and one 3-uniform sunflower pictured below, see Figure 5.
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Figure 5. The sunflowers corresponding to matrix entries −1 (left),
2 (center), and −2 (right).
Step 2. Next we construct for each column aj a connected matched-petal
partitioned-core sunflower Hj . For this we use the sunflowers constructed in Step
1 in the following way: if aij > 0 we use the previously constructed sunflower,
otherwise the sunflower obtained by deleting the petal with two vertices from the
corresponding sunflower. Finally, Hj consists of these disjoint 3-uniform sunflowers,
and a perfect matching on the union of non-core vertices. Then we add to Hj the
deleted petals with two vertices to obtain Hj′ . For example, since a1 = (−1, −2) we
take the sunflowers corresponding to −1 and −2 from Step 1 (see the leftmost picture of Figure 6), then delete the petals with two vertices to construct the connected
matched-petal partition-core sunflower H1 (see the middle picture of Figure 6) and
finally, add the two petals of cardinality 2 to obtain H1′ (the rightmost picture of
Figure 6).
20
•
•
•
•
v2 (11) v1 (11) u(11)
v1
v2 (11) v1 (11) u(11)
•
+
v2 (21)
•
−→
v1 (21)
•
•
v2 (21)
•
−→
v1 (21)
•
•
•
v2
•
u(21)
•
•
v3 (21)
v4 (21)
•
•
•
u(21)
v1
v1 (21)
•
•
•
v2 (21)
•
•
•
v2 (11) v1 (11) u(11)
•
•
•
v3 (21)
v4 (21)
•
u(21)
v2
v3 (21)
v4 (21)
Figure 6. The bouquet with basis H1′
v2 (12) v1 (12) u(12)
•
•
•
v1
v2 (13)
•
•
•
v1 (14)
•
• v1
•
• v2
•
v4 (22)
•
• v1
v1 (22)
•
•
v2 (14)
v1 (13)
•
v2 (22)
v3 (22)
v4 (13)
•
•
v3 (13)
•
•
•
•
v2 (23) v1 (23) u(23)
•
v3 (14)
v4 (14)
v2
Figure 7. The other 3 bouquets with basis corresponding to a2 , a3 , a4 .
For sake of completeness, we have labeled the vertices of sunflowers from Figure 6
according to the general construction of Theorem 3.2, and we denote their edges by
e0 (11) = {u(11), v1}, e1 (11) = {u(11), v1(11), v2 (11)}, e0 (21) = {u(21), v2}, e1 (21) =
{u(21), v1(21), v2 (21)}, e2 (21) = {u(21), v3(21), v4(21)}, e1 = {v2 (11), v1 (21)}, e2 =
{v2 (21), v3(21)}, e3 = {v4 (21), v1 (11)}. One should note that there is not a unique
way to choose the blue edges (in the picture e2 , e3 , e4 ) such that they form a matching
of the non-core vertices of the connected matched-petal partitioned-core sunflower
H1 . Also, we see that the hypergraph H1′ constructed is a bouquet with basis the set
U1 of all vertices except v1 , v2 . Similarly one repeats the construction for each one
of the column vectors of A, that is a2 , a3 , a4 , obtaining the hypergraphs depicted in
Figure 7. Note that all four hypergraphs H1′ , . . . , H4′ obtained through the previous
construction are bouquets with basis U1 , . . . , U4 , where each Ui is just the set of
vertices of Hi′ except v1 , v2 .
Step 3 Finally, consider the hypergraph H = (V, E) with |V | = 28 and |E| = 26,
obtained from H1′ , . . . , H4′ by taking the vertex set to be the union of the set of
vertices of H1′ , . . . , H4′ , and the set of edges is the (disjoint) union of the set of edges
of H1′ , . . . , H4′ . Furthermore, we compute the subbouquet vectors of H1′ , . . . , H4′ . For
this, we analyze in detail only the first hypergraph H1′ which is a bouquet with basis
U1 , constructed in Figure 6, the others being computed similarly. We assume that
the column vectors of the incidence matrix of H are indexed such that the first eight
correspond, in this order, to the edges e1 (11), e0 (11), e1(21), e2 (21), e0 (21), e1 , e2 , e3 ,
while the first two rows are indexed by the vertices v1 and v2 . Then cH′1 =
21
(1, −1, 1, 1, −2, −1, −1, −1, 0) ∈ Z26 and
aH′1 = e1 (11) − e0 (11) + e1 (21) + e2 (21) − 2e0 (21) − e1 − e2 − e3 = (−1, −2, 0) ∈ Z28 .
Analogously, with respect to a similar order for the rest of the edges of the other
three bouquets, we obtain the following cH′2 = (0, 1, −1, 1, 1, −1, −1, −1, 0, 0), cH′3 =
(0, 0, 1, 1, 1, −1, −1, −1, −1, 0) and cH′4 = (0, 0, 0, 1, 1, −1, −1) ∈ Z26 . Therefore
the other three bouquet vectors are aH′2 = (−1, 2, 0), aH′3 = (2, −1, 0) and aH′4 =
(2, 0, 0) ∈ Z28 . Since the set of edges of H can be partitioned into four bouquets with
bases H1′ , . . . , H4′ , the ideal IH is strongly robust, by Corollary 2.5. By Theorem 1.3,
the bijective correspondence between the Graver basis of IA and the Graver basis of
IH applies in the following way: to the vector u = (1, 3, 4, −2) ∈ Gr(A) corresponds
the vector B(u)
(1, −1, 1, 1, −2, −1, −1, −1|3, −3, 3, 3, −3, −3, −3|4, 4, 4, −4, −4, −4, −4|2, 2, −2, −2).
There are seven elements in Gr(H).
For the next construction we may assume that the matrix A has no zero column
or row. Indeed, if, say, a1 = 0, then IA = (IA′ , 1 − x1 ), where IA′ ⊂ K[x2 , . . . , xn ],
and if, say, the first row is zero, then IA = IA′ , where A′ = [a′1 , . . . , a′n ], with
a′j = (a2j , . . . , amj ) ⊂ Zm−1 for all j.
Theorem 3.2. Given any integer matrix A without any zero row or zero column,
there exists an almost 3-uniform hypergraph H = (V, E) such that:
(1) There is a bijective correspondence, u 7→ B(u), between KerZ (A), Gr(A) and
C(A) and KerZ (H), Gr(H) and C(A), respectively.
P
(2) For every u ∈ Gr(A) we have |u|1 ≤ |B(u)|1 , where |u|1 = ni=1 |ui | represents the 1-norm of the vector u.
(3) The toric ideal IH is strongly robust.
Proof. Let A = [a1 , . . . , an ] ∈ Zm×n . We will construct a hypergraph H whose
subbouquet ideal is IA , and with the property that all its non-free subbouquets
are mixed. This will imply at once conditions (1) and (3), by Theorem 1.3 and
Corollary 2.5. To this end, let {v1 , . . . , vm } be a set of vertices, and for each nonzero
entry of the matrix aij we introduce the following new vertices:
(1) if aij > 0 the set {v1 (ij), v2 (ij), · · · , v2aij (ij)} of 2aij vertices,
(2) if aij < 0 the set {u(ij), v1(ij), v2 (ij), · · · , v−2aij (ij)} of −2aij + 1 vertices,
as well as the following edges:
(3) if aij > 0 the aij edges es (ij) = {vi , v2s−1 (ij), v2s (ij)}, where 1 ≤ s ≤ aij ,
(4) if aij < 0 the −aij + 1 edges es (ij) = {u(ij), v2s−1 (ij), v2s (ij)}, where 1 ≤
s ≤ −aij and e0 (ij) = {vi , u(ij)}.
Note that for each aij 6= 0 the hypergraph Sij , on the set of vertices defined in
item (1) (item (2), respectively) with the set of edges defined in item (3) (item
(4), respectively) is a sunflower with the core vi (u(ij), respectively). Furthermore, each sunflower Sij is either a 3-uniform sunflower if aij > 0 or an almost
3-uniform sunflower if aij < 0 (by almost we mean that all the edges have three
vertices except e0 (ij), which has only two vertices). In addition, for any fixed j the
22
sunflowers S1j , . . . , Smj are vertex disjoint by definition. For each nonzero column
aj = (a1j , . . . , amj ) we construct first a connected matched-petal partitioned-core
sunflower Hj = (Vj , Ej ) on the set of vertices
[
[
Vj =
{v1 (ij), v2 (ij), . . . , v2aij (ij), vi }∪
{v1 (ij), v2 (ij), . . . , v−2aij (ij), u(ij)},
i:aij >0
i:aij <0
with the core vertices
Cj =
[
{vi } ∪
i:aij >0
[
{u(ij)}.
i:aij <0
For this is enough to describe the construction of H1 = (V1 , E1 ), the others being
similar. Consider all nonzero entries of the column a1 = (a11 , . . . , am1 ). To each
nonzero ai1 we associate the sunflower Si1 if ai1 > 0, or the sunflower denoted by
Si1 \ {e0 (i1)} obtained from Si1 by removing the edge e0 (i1), if ai1 < 0. Note that
Si1 \ {e0 (i1)} is a 3-uniform sunflower with core the set {u(i1)}, and its vertex set is
obtained from the vertex set of Si1 by removing vi . The matched-petal partitionedcore sunflower H1 with vertex set V1 , consists of the vertex-disjoint sunflowers Si1
(for those i with ai1 > 0) and Si1 \ {e0 (i1)} (for those i with ai1 < 0), and with the
following perfect matching on the set Vj \ Cj , of non-core vertices:
{v2 (11), v3(11)}, . . . , {v2|a11 |−2(11), v2|a11 |−1 (11)}, {v2|a11 | (11), v1 (21)},
{v2 (21), v3 (21)}, . . . , {v2|am1 |−2 (m1), v2|am1 |−1 (m1)}, {v2|am1 | (m1), v1 (11)},
where for convenience of notation we assumed that all ai1 6= 0 (see for an example
the blue edges from Figures 6, 7). This perfect matching on the non-core vertices
ensures that Hj = (Vj , Ej ) is a connected matched-petal partitioned-core sunflower.
Then we “extend” the hypergraph Hj to the hypergraph Hj′ = (Vj′ , Ej′ ), where
Vj′ = Vj ∪ {vi : aij < 0} and Ej′ = Ej ∪ {e0 (ij) : aij < 0}, which turns out to be
a bouquet with basis the set Uj = Vj′ \ {vi : aij 6= 0}. Alternatively, we can write
Hj′ = EUj for all j = 1, . . . , n. Finally, the hypergraph
H = (V, E)Swe are looking for
Sn
′
′
′
is obtained from H1 , . . . , Hn as follows V = j=1 Vj , and E = nj=1 Ej′ . Note that
since A has no zero row, then {v1 , . . . , vm } ⊂ V . Since each Hj′ is a bouquet with
basis, then it follows immediately from construction that H has n bouquets with
bases. Moreover, the resulting bouquet vector of Hj′ = EUj is aEUj = (aj , 0) ∈ Z|V |
and thus we obtain that the subbouquet ideal of IH is IA , as desired.
The following two results were inspired by the universality results of [9]. In some
sense, the following two corollaries strengthen [9, Theorem 1.2] and [9, Corollaries
2.1,2.2]. There, the motivation from algebraic statistics was to show that, for a particular 0/1 matrix of a 3-uniform hypergraph encoding a model on three-dimensional
contingency tables, as two of the dimensions of the table grow, there cannot be a
universal upper bound on the degrees of minimal Markov bases. To show this, it
is enough for [9] to construct one element of given arbitrarily high degree. The results below differ in three ways: 1) the underlying incidence matrices of the almost
3-uniform hypergraphs are 0 − 1 matrices with at most three ones on each column;
2) our statement holds true for all binomials in any minimal Markov basis, and not
only for one binomial; 3) since the toric ideal constructed is strongly robust then the
23
statement is also valid for Graver basis, indispensable binomials, universal Gröbner
basis respectively. The reader should note that the two classes of matrices considered in [9] and considered here have in common only the fact that they are 0 − 1
matrices with at most three ones on each column.
Corollary 3.3. For any vectors d1 , . . . , dr ∈ Zm there exists an almost 3-uniform
hypergraph H with its toric ideal strongly robust such that all of the elements of
the minimal Markov basis of H restricted to some of its entries cover the whole set
{d1 , . . . , dr }.
α 1 , . . . , α r ] ∈ Zp×r be a unimodular matrix such that all of its
Proof. Let D = [α
bouquets are not free, and consider the vectors c′i = (1, di ) = (1, di1 , . . . , dim ) ∈ Zm+1
for all i = 1, . . . , r. By definition the vectors c′1 , . . . , c′r are primitive and thus
we can apply Theorem 1.6 in order to obtain a generalized Lawrence matrix A =
[a1 , . . . , an ] ∈ Zq×n such that its subbouquet ideal equals ID , and whose Graver basis
equals the set of circuits and implicitly the universal Gröbner basis. In addition,
α i , 0, . . . , 0) and
we know that the subbouquet Bi is encoded by the vectors aBi = (α
cBi = (0, . . . , c′i , . . . , 0), for all i = 1, . . . , r. Since D is unimodular applying [22,
Proposition 8.11] we have that the Graver basis elements of D are vectors with
coordinates only 0, 1, −1, which in turn implies via the one-to-one correspondence
of Theorem 1.3
r
X
′
′
B((u1 , . . . , ur )) =
cBi u′i = (u′1 , u′1d1 , . . . , u′r , u′r dr ) ∈ Zn
i=1
that the Graver basis elements of A are vectors whose coordinate nonzero blocks are
either c′i or −c′i .
Finally, we apply for the matrix A the construction of Theorem 3.2 to obtain the
hypergraph H which has n bouquets with bases H1′ , . . . , Hn′ , and whose toric ideal is
strongly robust. Note from the proof of Theorem 3.2 that the nonzero coordinates of
the vectors cH′i corresponding to the edges of the perfect matching on the non-core
vertices are equal either to 1 or −1. Since the subbouquet ideal of IH is equal to
IA it follows from the description of the Graver basis elements of A that via the
one-to-one correspondence
n
X
(u′1 , u′1d11 , . . . , u′1 d1m , . . . , u′r , u′r dr1 , . . . , u′r drm ) = (u1 , . . . , un ) 7→
cH′j uj
j=1
the Graver basis elements of H contain as subvectors at least one of ±d1 , . . . , ±dr .
Indeed, from the construction of A we have that u′i ∈ {−1, 0, 1} and we just substitute them in the previous displayed formula. Since IH is strongly robust then
the Graver basis equals the minimal Markov basis and thus we have the desired
conclusion.
For the final part of statement, that is all of the vectors d1 , . . . , dr appear as a
support of some elements from the minimal Markov basis of H we need to make
a choice on the unimodular matrix D to ensure that in the Graver basis of ID
appear all of the variables. For example, we may choose D to be the incidence
matrix of a complete bipartite graph Kp,ℓ with p, ℓ ≥ 2 if r equals its number of
24
edges. Otherwise, we consider a complete bipartite graph Kp,ℓ with the number
of edges q > r, let D ∈ Zp×q be its incidence matrix and consider the q vectors
c′1 , . . . , c′r , c′r , . . . , c′r with the vector c′r being repeated q − r + 1 times.
Taking d1 = · · · = dr = d in the previous corollary and switching from the vector
+
−
notation of an element v ∈ KerZ (H) to the binomial notation xv − xv ∈ IH we
obtain the following
Corollary 3.4. For any vector d ∈ Zn there exists an almost 3-uniform hypergraph
H with its toric ideal strongly robust such that all binomials in the minimal generating set of IH satisfy the following: one of its monomials restricted to a suitable
subset of variables has multi-degree d+ and its other monomial restricted to a suitable
subset of variables has multi-degree d− .
4. Hypergraphs encode all positively graded toric ideals
This section presents a correspondence between positively graded (general) toric
ideals and stable toric ideals of hypergraphs. Namely, given a positively graded IA ,
Theorem 4.2 constructs a stable toric ideal of a hypergraph H = H(A) that has
the same combinatorial complexity and whose homological properties are preserved.
In particular, the Graver bases of the ideals IA and IH have the same number of
elements, and the same holds for their universal Gröbner and Markov bases, as well
as indispensable binomials and circuits. Even more is true: if IA has ℓ different
minimal Markov bases then IH has also ℓ different minimal Markov bases, and the
same holds for the reduced Gröbner bases. Contrast this with Section 3, where an
arbitrary toric ideal may have infinitely many minimal Markov bases.
For the remainder of this section, fix the following notation:
0 1 ... ... 1
1 0
1 . . . 1
. .
.
n×n
.
.
Σn :=
.. . . . . . . .. ∈ Z
1 . . . 1
0 1
1 ... ... 1 0
and
εk,n := (1| {z
· · · 1} 0 · · · 0) ∈ Zn , where 1 ≤ k < n.
k times
The following example illustrates a general construction which is the basis for the
proof of Theorem 4.2.
Example 4.1. Given a non-negative integer matrix D, we seek a procedure to create
a 0/1 matrix A such that there is a bijective correspondence between distinguished
sets of binomials of ID and IA , namely, each of: Graver basis, circuits, indispensable
binomials, minimal Markov bases, reduced Gröbner bases (universal Gröbner basis).
Let
1 3 2 0 1
D = (dij ) = 3 2 1 3 2 ∈ N3×5 .
3 0 2 2 1
25
Let δi := maxl {dli } be the maximum entry in column i of D, and ji := min{l :
dli = δi } be the index of the first row where δi appears. For the given matrix
D, these values are as follows: δ1 = 3, δ2 = 3, δ3 = 2, δ4 = P
3 and δ5 = 2; and
5
j1 = 2, j2 = 1, j3 = 1, j4 = 2 and j5 = 2. Define δ :=
i=1 (δi + 1) = 18
and l = 3 − |{j1 , . . . , j5 }| = 1, the number of rows that do not contain any columnmaximum entry δi . We construct a 0-1 matrix A = MH of size 19 × 18, the incidence
matrix of a hypergraph H, such that its subbouquet ideal will be ID . Here, 19 = δ+l
and 18 = δ. The matrix A is constructed from D in two steps.
Step 1. Every column-maximum entry δi defined above will determine a set of
horizontal blocks of the matrix A as follows. For each row index k ∈ {ji }5i=1 , consider
the set of all δi ’s appearing on the k-th row of D. For each such δi , construct a
(δi + 1) × δ block matrix by concatenating (horizontally) the following 5 sub-blocks:
block i shall consist of the matrix Σδi +1 , while for each l 6= i, block l shall consist of
δi + 1 copies of εdkl ,δi +1 .
For example, the first row of D contains two column-maximum entries, δ2 and
δ3 . The first entry will generate a row-block of size (3 + 1) × 18 inside A, while the
second entry will generate a row-block of size (2 + 1) × 18. First, the entry δ2 = 3
requires concatenating five sub-blocks: the first one consisting of 4 = δ2 + 1 copies
of εd11 ,4 = (1 0 0 0); the second one, Σ4 ; the third one consisting of 4 copies of
εd13 ,4 = (1 1 0 0); the fourth one, 4 copies of εd14 ,4 = (0 0 0 0); and the fifth, 4
copies of εd15 ,4 = (1 0 0 0). This resulting row-block of A is displayed below, with
the distinguished sub-block corresponding to Σ4 colored in blue.
1 0 0 0 0 1 1 1 1 1 0 0 0 0 0 1 0 0
1 0 0 0 1 0 1 1 1 1 0 0 0 0 0 1 0 0
1 0 0 0 1 1 0 1 1 1 0 0 0 0 0 1 0 0 .
1 0 0 0 1 1 1 0 1 1 0 0 0 0 0 1 0 0
Similarly, the second entry, δ3 = 2,
1 0 0 0 1 1 1
1 0 0 0 1 1 1
1 0 0 0 1 1 1
will generate the following 3 × 18 row-block:
0 0 1 1 0 0 0 0 1 0 0
0 1 0 1 0 0 0 0 1 0 0 .
0 1 1 0 0 0 0 0 1 0 0
The remaining row indices are j1 = j4 = j5 = 2, thus we consider the second row
of D. It contains three δi values: δ1 , δ4 , δ5 , and thus gives rise to three row-blocks of
similar structure. These are the third, fourth and fifth row-blocks of the matrix A
in Equation (3).
Step 2. For each row k of D containing none of the column-maxima δi ’s, that
is, for all row indices k ∈ {1, 2, 3} \ {ji }5i=1 , we will create one additional row of
A. This row will be obtained by concatenating (in this order) the following vectors
εdk1 ,δ1 +1 , εdk2 ,δ2 +1 , εdk3 ,δ3 +1 , εdk4 ,δ4 +1 , εdk5 ,δ5 +1 . For example, the third row of D does
not contain any δi and thus the row of A obtained from concatenating the vectors
εd31 ,4 = (1 1 1 0), εd32 ,4 = (0 0 0 0), εd33 ,3 = (1 1 0), εd34 ,4 = (1 1 0 0) and
εd35 ,3 = (1 0 0) is
1 1 1 0 0 0 0 0 1 1 0 1 1 0 0 1 0 0 .
26
Finally, the matrix A = [a1 , . . . , a18 ] is obtained by concatenating (vertically) the
two row-blocks generated by the first row of D, the three row-blocks generated by
the second row of D, and the row generated by the third row of D; see Equation (3).
Note that the submatrix of A generated by the first row of D has 4 + 3 = 7 rows, the
one generated by the second row of D has 4 + 4 + 3 = 11 rows, and the one generated
by the third row of D has 1 row. The matrix A has 5 non-mixed bouquets such
that a1 , a2 , a3 , a4 belong to the first bouquet B1 , a5 , a6 , a7 , a8 to B2 , a9 , a10 , a11 to
B3 , a12 , a13 , a14 , a15 to B4 , and a16 , a17 , a18 to B5 . Moreover all nonzero components
of cB1 , . . . , cB5 ∈ Z18 are 1, and thus aB1 = a1 + · · · + a4 ∈ Z19 is the transposed
vector of
1 1 1 1 1 1 1 3 3 3 3 3 3 3 3 3 3 3 3 ,
where the first block has 7 coordinates, the second one 11 coordinates, and the last
one 1 coordinate. Similarly, aB2 = a5 + · · · + a8 is the transposed vector of
3 3 3 3 3 3 3 2 2 2 2 2 2 2 2 2 2 2 0 ,
and so on. Thus IAB = ID , and since IA is a stable toric ideal the desired bijective
correspondence follows from Theorem 1.7.
(3) A = MH =
1
1
1
1
1
1
1
0
1
1
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
1
0
1
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
1
1
0
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
1
1
1
0
0
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0
1
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0
1
1
0
1
1
1
1
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
1
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0
1
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
1
0
1
1
1
1
1
1
1
0
0
0
0
0
0
0
1
1
1
1
1
0
1
1
1
1
1
1
0
0
0
0
0
0
0
1
1
1
1
1
1
0
1
1
1
1
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0
1
1
1
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
0
0
Theorem 4.2. Let ID be an arbitrary positively graded nonzero toric ideal. Then
there exists a hypergraph H such that there is a bijective correspondence between the
Graver bases, all minimal Markov bases, all reduced Gröbner bases, circuits, and
indispensable binomials of ID and IH . Furthermore, all of the homological data of
IH is inherited by ID and viceversa.
27
Proof. Since ID is a positively graded toric ideal, we may assume by [15, Corollary
7.23] that D = (dij ) ∈ Nm×n . Furthermore, every column of D is nonzero, since,
otherwise, if the j-th column is 0, then xj − 1 ∈ ID , a contradiction to ID being
positively graded. We will construct the incidence matrix A of a hypergraph H, such
that its subbouquet ideal is equal to ID . For this define column-maximum
entries
P
δi := max{dji : j = 1, . . . , m} for all i = 1, . . . , n and set δ = n + ni=1 δi . Note
that δi > 0 for every i = 1, . . . , m. Moreover, denote by ji := min{k : dki = δi } for
all i = 1, . . . , n and denote by l = m − #{ji : 1 ≤ i ≤ n}.
Then the 0-1 matrix A ∈ N(δ+l)×δ consists of several blocks, concatenated vertically, of the following two types: 1) For each k ∈ {j1 , . . . , jn }, and for each columnmaximum entry δi located on the k-th row of D, construct δi + 1 rows of A by
concatenating (horizontally) n block matrices. Here, the i-th block is Σδi +1 , while
for each l 6= i the l-th block consists of δi + 1 copies of the row sub-vector εdkl ,δl +1 .
2) For each row index k ∈ [m] \ {j1 , . . . , jn } of D, that is, each row not containing
any column-maximum entry δi , construct a row of A by by concatenating (in this
order) the following vectors
(εdk1 ,δ1 +1 |εdk2 ,δ2 +1 | . . . |εdkn ,δn +1 ).
Assume now that the columns of A are labeled a1 , . . . , aδ . We prove that the first
δ1 + 1 column vectors of A belong to the same subbouquet B1 , the next δ2 + 1 belong
to the same subbouquet B2 , and so on, until the last δn + 1 column vectors belong to
the same subbouquet Bn . We will prove this only for the first δ1 + 1 columns of A,
the other cases being similar. By the definition of A we have that in the submatrix
of A determined by the columns a1 , . . . , aδ1 +1 there exist integers i1 , . . . , i1 + δ1 = i2
such that the matrix corresponding to these rows and the columns a1 , . . . , aδ1 +1 is
Σδ1 +1 . We consider the vectors ci,i+1 ∈ Zδ+l for every i = i1 , . . . , i1 + δ1 − 1, whose
only nonzero coordinates are 1 on position i, and −1 on position i + 1. Then the covector (ci1 ,i1 +1 · a1 , . . . , ci1 ,i1 +1 · aδ ) = (−1, 1, 0, . . . , 0) has support of cardinality two
and G(a1 ) = G(a2 ). Here, we used the fact that in the horizontal block containing
Σδ1 +1 (and corresponding to the rows i1 , . . . , i2 ) all the entries corresponding to the
rows i1 , i1 + 1 and columns 3, . . . , δ are identical, by definition. Similarly, using all
the vectors ci,i+1 for i = i1 , . . . , i2 − 1, we obtain G(a1 ) = · · · = G(aδ1 +1 ), and
thus they all belong to the same subbouquet B1 , which is either free or non-mixed.
In any case, the vector cB1 = (1, 0, . . . , 0) ∈ Zδ with 1 = (1, . . . , 1) ∈ Zδ1 +1 and
aB1 = a1 +· · ·+aδ1 +1 . Analogously, we have cB2 = (0, 1, . . . , 0) with 1 = (1, . . . , 1) ∈
Zδ2 +1 and aB2 = aδ1 +2 + · · · + aδ1 +δ2 +2 , and so on, until cB2 = (0, 0, . . . , 1) with
1 = (1, . . . , 1) ∈ Zδn +1 and aBn = aδ−δn + · · · + aδ . One can easily see from the
construction of A that IAB = ID . In addition, since ID 6= 0 then at least one
subbouquet is non-free, and therefore IH = IA is a stable toric ideal. Applying now
Theorem 1.7 and [19, Theorem 3.11] we obtain the desired conclusions.
Remark 4.3. In conclusion, the previous theorem can be regarded as a polarizationtype operation for positively graded toric ideals by comparison to the properties of
the classical polarization for monomial ideals, see [12, Corollary 1.6.3]. Indeed, if
ID ⊂ S = K[x1 , . . . , xn ] and IH ⊂ T = K[y1 , . . . , yδ ] are the ideals from the above
28
construction since ID is the subbouquet ideal of IH then height(ID ) = height(IH ).
Applying now [19, Theorem 3.11] we have pd(T /IH ) = pd(S/ID ), the ND-graded
Betti numbers of ID are obtained from the NAH -graded Betti numbers of IH and
viceversa, and thus S/ID is Cohen-Macaulay (respectively Gorenstein) if and only
if T /IH is Cohen-Macaulay (respectively Gorenstein).
Remark 4.4. Moreover, combinatorial classification problems of arbitrary positively graded toric ideals whose different sets of bases are equal can be reduced
to problems about toric ideals of hypergraphs. For example, in [4] Boocher et. al
proved that, for robust toric ideals of graphs, the Graver basis is a minimal generating set. In other words, robust toric ideals of graphs are strongly robust. They ask
if this property is true in general for any robust ideal. To prove such a statement,
it is enough to prove it only for toric ideals of hypergraphs. Indeed, if IA is any
robust toric ideal then Theorem 4.2 shows that IH is robust. Then if one can prove
that robust ideals of hypergraphs are strongly robust then it follows, again from
Theorem 4.2, that IA is also strongly robust.
References
[1] 4ti2 team, 4ti2 - a software package for algebraic, geometric and combinatorial problems on
linear spaces, available at www.4ti2.de (2007).
[2] D. Bayer, S. Popescu, B. Sturmfels, Syzygies of unimodular Lawrence ideals, J. Reine Angew.
Math. 534, 169–186 (2001).
[3] A. Boocher, E. Robeva, Robust toric ideals, J. Symbolic Computation, 68(1), 254–264 (2015).
[4] A. Boocher, B. C. Brown, T. Duff, L. Lyman, T. Murayama, A. Nesky and K. Schaefer,
Robust graph ideals, Annals of Combinatorics 19(4), 641–660 (2015).
[5] H. Charalambous, A. Thoma, M. Vladoiu, Minimal generating sets of lattice ideals, Collectanea Mathematica 68, 377–400 (2017).
[6] H. Charalambous, A. Thoma, M. Vladoiu, Markov complexity of monomial curves, Journal
of Algebra 417, 391–411 (2014).
[7] H. Charalambous, A. Thoma, M. Vladoiu, Markov bases and generalized Lawrence liftings,
Annals of Combinatorics 19(4), 661–669 (2015).
[8] J. De Loera, R. Hemmecke, S. Onn, R. Weismantel, N-fold integer programming, Discrete
Optimization 5, 231–241 (2008).
[9] J. De Loera, S. Onn, Markov bases of three-way tables are arbitrarily complicated, J. Symbolic
Computation 41, 173–181 (2006).
[10] J. De Loera, B. Sturmfels, R. Thomas, Gröbner bases and triangulations of the second hypersimplex, Combinatorica 15(3), 409–424 (1995).
[11] E. Gross, S. Petrović, Combinatorial degree bound for toric ideals of hypergraphs, Internat.
J. Algebra Comput. 23(6), 1503–1520 (2013).
[12] J. Herzog, T. Hibi, Monomial Ideals, GTM 260. Springer 2010.
[13] S. Jukna, Extremal Combinatorics-with Applications in Computer Science, Texts in Theoretical Computer Science. Springer, New York (2001).
[14] M. Lason, M. Michalek, On the toric ideal of a matroid, Adv. in Mathematics 259, 1–12
(2014).
[15] E. Miller, B. Sturmfels, Combinatorial Commutative Algebra, Grad. Texts in Math., vol. 227,
Springer-Verlag, New York, (2005).
[16] H. Ohsugi, T. Hibi, Toric ideals generated by quadratic binomials, J. Algebra 218, 509–527
(1999).
[17] I. Peeva, B. Sturmfels, Generic lattice ideals, J. Amer. Math. Soc., Vol. 11(2), 363–373 (1998).
29
[18] S. Petrović, D. Stasi, Toric algebra of hypergraphs, J. Algebraic Combinatorics, Volume 39,
Issue 1, pp 187-208 (2014).
[19] S. Petrović, A. Thoma, M. Vladoiu, Bouquet algebra of toric ideals, arXiv:1507.02740v3.
[20] E. Reyes, Ch. Tatakis, A. Thoma, Minimal generators of toric ideals of graphs, Adv. Appl.
Math. 48, 64–78 (2012).
[21] F. Santos, B. Sturmfels, Higher Lawrence configurations, J. Combin. Theory Ser. A 103,
151–164 (2003).
[22] B. Sturmfels, Gröbner Bases and Convex Polytopes, University Lecture Series, No 8, MAS,
R.I. (1995).
[23] B. Sturmfels, S. Sullivant, Toric ideals of phylogenetic invariants, Journal of Computational
Biology 12, 204–228 (2005).
[24] S. Sullivant, Strongly robust toric ideals in codimension 2, arXiv:1610.07476v1, 2016.
[25] Ch. Tatakis, A. Thoma, On the universal Gröbner bases of toric ideals of graphs, J. Combin.
Theory Ser. A 118, 1540–1548 (2011).
[26] R. Villarreal, Rees algebras of edge ideals, Comm. Algebra 23, 3513–3524 (1995).
[27] R. Villarreal, Monomial Algebras, Second Edition, Monographs and Research Notes in Mathematics, Chapman and Hall/CRC, 2015.
Sonja Petrović, Department of Applied Mathematics, Illinois Institute of Technology, Chicago 60616, USA
Apostolos Thoma, Department of Mathematics, University of Ioannina, Ioannina
45110, Greece
E-mail address: [email protected]
Marius Vladoiu, Faculty of Mathematics and Computer Science, University of
Bucharest, Str. Academiei 14, Bucharest, RO-010014, Romania, and
Simion Stoilow Institute of Mathematics of Romanian Academy, Research group
of the project PN-II-RU-TE-2012-3-0161, P.O.Box 1–764, Bucharest 014700, Romania
E-mail address: [email protected]
30
| 0math.AC
|
The Probability of Causation
arXiv:1706.05566v1 [math.ST] 17 Jun 2017
Philip Dawid∗
Monica Musio†
Rossella Murtas‡
Dedicated to the memory of Stephen Elliott Fienberg
27 November 1942–14 December 2016
June 20, 2017
Abstract
Many legal cases require decisions about causality, responsibility or blame, and these may
be based on statistical data. However, causal inferences from such data are beset by subtle
conceptual and practical difficulties, and in general it is, at best, possible to identify the “probability of causation” as lying between certain empirically informed limits. These limits can
be refined and improved if we can obtain additional information, from statistical or scientific
data, relating to the internal workings of the causal processes. In this paper we review and
extend recent work in this area, where additional information may be available on covariate
and/or mediating variables.
Key words: Balance of probabilities, Causes of effects, Counterfactual, Group to individual
inference, Mediator, Potential response, Sufficient covariate
1
Introduction
Many legal proceedings hinge on the assignment of blame or responsibility for some undesirable
outcome. In a civil case a patient may sue a pharmaceutical company for damage caused as a side
effect of one of its products; or a state health department may bring an action against a tobacco
company for not disclosing information it had on the health risks of smoking, thus (it is claimed)
leading to unnecessary deaths.
In many such cases there will be no dispute about the facts at issue. The patient took the drug
and suffered the side effect. The tobacco company admits to having withheld the information.
What remains at issue is the causal relation between the established facts. But any attempt
to understand this quickly throws us into the philosophical quagmire of counterfactual reasoning,
where we have to consider what might have happened in circumstances known to be false. Whether
or not the patient’s suit succeeds will depend on whether or not she can prove, to the appropriate
legal standard (e.g., “on the balance of probabilities”) that she would not have developed the same
outcome, had she not taken the pharmaceutical product. The damages that the tobacco company
are liable for will depend on an assessment of how many lives could have been saved, had they
made the information available.
In this article we concentrate on cases of the first kind, where it is desired to assess whether or
not the same outcome would have occurred had the putative causal event been absent. Typically,
epidemiological evidence will be admitted as to the frequency of the adverse outcome, both in
patients who have, and in patients who have not, been exposed to the product. Although such
evidence is clearly of relevance, it is less clear exactly how. A common procedure is to compute
∗ University
of Cambridge
of Cagliari
‡ University of Cagliari
† University
1
the relative risk (RR), obtained by dividing the frequency of the outcome among those exposed
to the product by the corresponding frequency among those not so exposed. And it is frequently
asserted that a relative risk exceeding 2 is enough to prove that the “probability of causation”
(PC) exceeds 0.5, and thus to establish a causal link “on the balance of probabilities”.
In this paper we consider such arguments in greater detail. We explain why it is difficult to
establish a precise value for the probability of causation on the basis of epidemiological or other
scientific evidence, which at best can only provide interval bounds for PC. We also show how it is
possible to reduce the uncertainty about PC by collecting and taking account of additional data,
so shedding some light on the internal workings of the causal black box.
The work presented here builds on a long-standing interaction and collaboration with Stephen
Fienberg, as represented in particular by the recent papers Dawid et al. (2014); Dawid et al.
(2015); Dawid et al. (2016b). This formed a very small component part of Steve’s many highly
significant contributions to the correct use of Statistics in the Law, which work, substantial as it
was, itself constituted only a very small component of his numerous original and highly influential
contributions over an enormous range of statistical topics. He was a delightful friend and a
stimulating colleague and companion; he is sorely missed.
2
Group to individual analysis: The basic problem
Epidemiology attempts to discover causal relationships, such as between a certain exposure, E
and a certain response, R. Such a relationship can be intuitively represented by a diagram such
as that of Figure 1. (There is a formal semantics underlying such diagrams, but we shall not delve
into that here).
E
R
Figure 1: Simple cause-effect relationship
Epidemiological experts are frequently called on to testify in cases revolving around the assignment of individual responsibility. However, such testimony can only concern the behaviour of
groups of individuals, and its relevance to any individual case is at best indirect. The general issue
of “group-to-individual” (“G2i”) inference has recently become a centre of attention both within
and outside the Law (Faigman et al. 2014).
In our context, the important distinction to be made is between inference about “Effects of
Causes” (“EoC”), and inference about “Causes of Effects” (“CoE”). The former, concerning the
likely outcome of a contemplated intervention or exposure, in a specific or generic individual, is
the focus of most scientific investigation and experimentation, including epidemiological studies.
However, it is the latter, concerning the possible causes of an observed effect in an individual case,
that is of most relevance for the Law. The distinction between EoC and CoE is crucial, but all
too often goes unremarked and unappreciated.
We illustrate the subtle relationship between these two forms of inference by means of an
example (Dawid 2011; Dawid et al. 2014).
Effects of Causes (EoC) Ann has a headache. She is wondering whether to take aspirin. Will
that cause her headache to disappear?
Causes of Effects (CoE) Ann had a headache and took aspirin. Her headache went away. Was
that caused by the aspirin?
To assist in addressing these questions, we assume that we have data from a very large, wellconducted, randomised trial of the success of aspirin in curing headaches, whose subjects can be
assumed to be similar to Ann in the way they respond. This yields percentages as in the following
table:
2
No aspirin
Aspirin
Recover
12
30
Not recover
88
70
Total
100
100
Table 1: Results of aspirin trial
3
A framework for analysis
Effects of Causes (EoC) Ann can argue as follows “If I take the aspirin, I will be just like the
subjects in the treatment group, so I can assess my probability of recovery in this case as
30%. Similarly, If I don’t take the aspirin, my probability of recovery in that case can be
assessed to be 12%. Other things being equal, it would be better to take the aspirin.” While
this argument does, not, perhaps address the question exactly as posed, it does give Ann
fully adequate guidance to address and solve her EoC decision problem. What is at issue
was the as yet unknown event of recovery, and the data are directly relevant for assessing
uncertainty about that unknown outcome (for each of the two decisions that Ann might
take).
Causes of Effects (CoE) Things are very different for the CoE question. Indeed, there is now
no unknown event: We know that Ann definitely took the aspirin, and she definitely recovered. What is at issue here is not uncertainty about an as yet unknown event, but
uncertainty about a relationship between known events: the existence, or otherwise, of a
causal relationship between taking the aspirin and recovering.
3.1
Counterfactuals and potential outcomes
One way of thinking about whether a certain outcome was caused by a certain exposure is to ask the
question: “Would the same outcome still have occurred, even if the exposure had been absent?”.
Because this question is posited on a hypothesis that is known to be false, and thus counter to
known facts, it is termed a counterfactual query. There is much discussion of counterfactuals in
the philosophical literature (see e.g. Menzies (2014)), and they raise many perplexing problems. In
particular, is it even meaningful to regard a counterfactual as having a definite (even if unknown)
truth value?
Supposing that we can talk of the truth value of counterfactual, we might reinterpret the
assertion “The exposure caused the outcome” as:
“Both the exposure and the outcome occurred; and had (counterfactually) the exposure
not occurred, neither would the outcome have occurred.”
With this interpretation, and taking account of the known facts, the truth value of the assertion
“The exposure caused the outcome” becomes identified with the (typically unknown) truth value
of the counterfactual proposition “In the counterfactual world where the exposure did not occur,
the outcome did not occur”.
Some comments
(i). There is an implicit “ceteris paribus” condition in the above interpretation. In the counterfactual world being considered, we are changing the truth value of the exposure, but nothing
else—or, at any rate, nothing else occurring before, or simultaneously with, the exposure.
We must of course allow for the outcome to change, and we might possibly also allow changes
to other, later, events (perhaps intermediate between the exposure and the outcome), in the
light of the change to the exposure. Since there are various ways in which our counterfactual
comparison world can be constructed, there is a resultant ambiguity in the truth value of
the counterfactual assertion. Lewis (1973) argues that we should use the “closest possible
3
world” to the real world, subject to the counterfactual change in the exposure—but this still
leaves a great element of vagueness.
(ii). There may be some choice as to what the counterfactual value of the exposure would have
been, if it had been other than it actually was. If Ann had not taken two aspirins, she might
have taken none at all, or taken three aspirins, or taken two paracetamol tablets, or gone to
bed with a wet towel around her head. The counterfactual query only becomes precise when
we have fully specified the “counterfactual foil” for comparison with the factual exposure.
(iii). There will typically be a variety of different exposures that we can consider. For example,
Ted has been a smoker for 20 years, has lived by a busy main road which is subject to
high air pollution, and has developed lung cancer. Then we could ask “Was Ted’s cancer
caused by his smoking?”, “Was it caused by his air pollution exposure?”, “Was it caused
by his smoking and the air pollution jointly?”, etc., etc. Each of these questions involves
consideration of a different counterfactual world for comparison with the real world, and
there is no reason why we could not regard all of them as true simultaneously—which leads
to some difficulties in understanding what might be the “actual cause” of Ted’s cancer
(see Halpern (2016) for a thoughtful analysis of the elusive concept of “actual cause”). In
particular, we should not express “Ted’s smoking caused his cancer” as “The cause of Ted’s
cancer was his smoking”—though we might say “A cause of Ted’s cancer was his smoking”.
The former phrasing suggests, misleadingly, that there is a variable, “The cause of Ted’s
cancer”, that can take different values (e.g., smoking, air pollution, natural causes). But
as these putatuve causes are not mutually exclusive, there can be no such variable. And
when we can assign probabilities to “Ted’s smoking caused his cancer”, “Air pollution caused
Ted’s cancer”, etc., it is perfectly acceptable for these to sum to more than 1. (For a specific
example of this, see Hutchinson et al. (1991)).
The standard framework of probability theory does not easily admit argument about counterfactual worlds and uncertain relationships. To formulate such questions meaningfully, we need to
expand the universe of discourse. The way this has traditionally been done in statistics is by the
introduction of so-called “potential outcomes” (Neyman 1935; Rubin 1978). We consider a binary
exposure, or cause, variable E, and a binary response, outcome, or effect, variable R. For each
value e = 0 or 1 of E, we introduce a variable Re , conceived as the value that R would assume,
were it the case that E = e. If in the actual world E = 1 (say), then the observed outcome
will be R = R1 ; in this case we have no way of observing R0 , which thus becomes a (necessarily
unobservable) counterfactual variable. Similarly, if in the actual world E = 0, then R = R0 ,
and R1 becomes counterfactual and unobservable. It is typically assumed, explicitly or implicitly,
that the bivariate variable R := (R0 , R1 ) is determined before the value of E is, and does not
depend on how that value of E came to be determined or known (for example, whether by external
intervention, or by observation alone). The effect of setting or observing E = 1 (say) is just to
uncover the pre-existing value of R1 , and identify R with R1 . In particular, although the pair
(R0 , R1 ) is modelled probabilistically just like any other bivariate random variable, there is the
epistemic problem that there are no circumstances in which it would be possible to observe both
its constituent values. This has been termed “the fundamental problem of statistical causality”
(Holland 1986).
The framing and analysis of causal questions in terms of potential outcomes has become the
industry standard in statistical causality. For the investigation of “effects of causes”, it has been
strongly argued that this approach is both unnecessary and potentially misleading (Dawid 2000;
Dawid 2015). However, for studying causes of effects there is currently no alternative to the use of
potential outcomes to model counterfactual possibilities. Nevertheless, the logical difficulty that
the bivariate quantity R can never be fully observed creates subtle problems that must not be
ignored.
4
3.2
The probability of causation
Armed with the above machinery and notation, we can now formulate an expression for the
probability of causation.
We interpret the expression “the exposure E = 1 causes the outcome R = 1” as the event “EC:
R1 = 1, R0 = 0”. This encodes the property that the potential response to E = 1 is 1, while at
the same time the potential response to E = 0 is 0. However, although we have thus expressed
the causality relation as an event, in what appears to be a standard probabilistic framework, the
fundamental problem of statistical causality implies that this is an event whose truth value can
never be determined.
Suppose now that (for Ann) we have observed E = 1 and R = 1 (or, equivalently, E = 1
and R1 = 1). Since we must condition on all known facts, the appropriate expression for the
probability of causation (in Ann’s case) is
PC
=
Pr(EC | E = 1, R = 1)
=
Pr(R0 = 0 | E = 1, R1 = 1).
(1)
However, merely establishing a notational expression for the probability of causation does not, of
itself, solve the problem of how to evaluate this quantity, involving as it does the probability of
an unobservable event.
The approach we shall take is to regard the pair (R0 , R1 ) as having a joint distribution (jointly
with all other variables in the problem), but to make no assumptions about that distribution, other
than any constraints imposed by empirically observable information. Typically it will not then
be possible to identify that joint distribution fully, but only to constrain it within limits—though
those limits will depend on the precise form of the available information.
There is a further point that needs to be taken into account. The variables in (1) are those
pertaining specifically to Ann, and the joint distribution Pr appearing in it is thus a distribution
over those specific variables. But who is the analyst assigning this distribution? It might be Ann
herself, or some onlooker. These individuals will typically have different background information,
and so different distributions. In particular, an onlooker might reasonably infer, merely from
observing that Ann decided to take the aspirin (E = 1), that her headache is particularly bad,
and consequently that her potential responses R0 and R1 are both likely to be poorer than if he did
not have this information. For Ann, however, the severity of her headache is known background
information that is already taken into account in her own distribution, and the mere fact that
she is contemplating taking the aspirin gives her no additional information about her potential
responses. That is to say, for Ann’s own probability distribution we can reasonably invoke the
“no confounding” assumption,
E ⊥⊥ R,
(2)
which expresses probabilistic independence between the fact of her exposure, E. and the pair R
of her potential responses. In this case expression (1) simplifies to
PC = Pr(R0 = 0 | R1 = 1).
(3)
However, for the onlooker who is not initially aware of the severity of Ann’s headache, it may be
inappropriate to accept (2), and hence (3).
In most of the sequel (with the exception of § 5) we shall assume that the analyst has sufficient
background information about Ann for (2) to be true (to a good enough approximation), and hence
use (3) as the expression for the probability of causation. Furthermore, because we shall wish to use
external data from other individuals to assist in estimating the analyst’s probability distribution
Pr for Ann, we shall also need to assume that the background information the analyst has on those
individuals is essentially the same as what he1 has for Ann — a condition that can sometimes
be met by restricting the set of external individuals considered. Even so, these assumptions,
necessary for much of our analysis, are highly restrictive, and will often be inappropriate. Careful
consideration must be given to their suitability before applying our results.
1 We refer to the analyst in the masculine, to distinguish him from the subject, Ann—while not precluding that
they could be the same individual.
5
4
Basic analysis
We start by considering the case that the only available empirical information is that presented
in Table 1.
Assuming the analyst can regard the individuals in the trial as similar to Ann, he might argue:
“When Ann takes aspirin, she puts herself in the same position as those external individuals
in the treatment group, of which Pr(R = 1 | E = 1) = 30% recovered. Because of the no
confounding assumptions (both in the data and for Ann), this means that her potential response
to taking aspirin, R1 , has a 30% probability of taking the value 1 (Recover). I can thus identify
my marginal probability (for Ann), Pr(R1 = 1) = 0.3. Similarly, I can evaluate my marginal
probability Pr(R0 = 1) = 0.12. However, because of the fundamental problem of statistical
causality, there is no further empirical information available relating to my joint distribution of
(R0 , R1 ).”
Armed with this limited information, the analyst can attempt to fill in the entries in Table 2,
specifying his bivariate probability distribution for (R0 , R1 ). The margins are fully determined by
the above empirically informed probabilities, but since he has no empirical information directly
relevant to the joint probability Pr(R0 = 0, R1 = 1), he simply enters an unknown (and unknowable) value x in the associated internal cell of the table. In combination with the known marginal
probabilities, this allows completion of the body of the table, as shown.
Table 2: Joint distribution of potential responses R0 and R1 .
R0
0
1
R1
0
0.88 − x x − 0.18 0.70
1
x
0.30 − x 0.30
0.88
0.12
1
Now although the precise value of x remains unknown, from the table we can extract partial
information about it, using the fact that any probability must be nonnegative. Applying this
principle to the four entries in the body of the table, we obtain: x ≤ 0.88, x ≥ 0.18, x ≥ 0,
x ≤ 0.30. Combining these and omitting redundancies gives the inequality: 0.18 ≤ x ≤ 0.30. This
interval bound is the full information available about x on the basis of the empirical data.
Now because of assumption (2), PC = Pr(R0 = 0 | R1 = 1) = x/0.30. Thus the analyst can
assert:
0.6 ≤ PC ≤ 1.
(4)
Even under the restrictive assumptions that have been made, this wide interval bound is all that
can be deduced, from the empirical data, about the probability of causation in this particular case.
The above logic can be applied in the general case, leading to the following interval bound in
terms of empirically estimable quantities (under our no-confounding assumptions):
1
Pr(R = 0 | E = 0)
max 0, 1 −
≤ PC ≤ min 1,
,
(5)
RR
Pr(R = 1 | E = 1)
where
RR :=
Pr(R = 1 | E = 1)
Pr(R = 1 | E = 0)
(6)
is the (causal) risk ratio. We note that RR > 2 implies PC > 0.5, which might be taken to prove
causation “on the balance of prbabilities”. However, if RR < 2 we can not deduce from (5) that
PC < 0.5 (to reject causation, on the balance of probabilities).
6
5
Allowing for unobserved confounding
When we cannot assume (2) we cannot simply apply the above theory directly, since the very
fact that Ann decided to take the aspirin gives the analyst some indirect information about Ann’s
state of health—in the light of which it would no longer be appropriate to consider her as similar
to the individuals in the experimental study, for which this information was not available. Tian
and Pearl (2000) showed that in this situation we can still obtain bounds for PC, so long as, in
addition to the experimental data, we also have observational data on individuals whose behaviour
(in the opinion of the analyst) can be considered similar to that of Ann, in that they have the
same dependence as she does between their decision E to take the aspirin and their pair R of
potential responses. Tian and Pearl (2000) then derive the following interval bounds:
1 − Pr(R0 = 1) − Pr(E = 0, R = 0)
Pr(R = 1) − Pr(R0 = 1)
≤ PC ≤ min 1,
, (7)
max 0,
Pr(E = 1, R = 1)
Pr(E = 1, R = 1)
where Pr(R0 = 1) denotes the recovery probability in the experimental control group, who were
externally assigned to have E set to 0; while the other probabilities are obtained from the observational data (interestingly, we do not require experimental data on those who were assigned
treatment).
For example, suppose that, in addition to the experimental data of Table 1, we have observational data in which Pr(E = 0, R = 1) = 0.12. We then find that the lower bound in (7) is
1—allowing us to establish causation with certainty.
6
Peeking into the box
So far we have considered the relationship between exposure and response as a causal “black box”,
without attempting to understand or model its hidden mechanisms in any further detail. If we are
willing to make some assumptions about these mechanisms, or access external information about
them, we can make better causal inferences from our data—even in the absence of any further
information about Ann.
6.1
Monotonicity and beyond
One assumption commonly made is monotonicity. In a context such as ours, this would say that, if
recovery would have occurred without the treatment, it would certainly occur with the treatment.
That is to say, R0 = 1 ⇒ R1 = 1. This implies in particular Pr(R0 = 1) ≤ Pr(R1 = 1), or
equivalently RR ≥ 1. This is a constraint on the margins of a table such as Table 1, and so
could be falsified; but even when this marginal constraint is satisfied, there is typically no further
information to help us decide whether monotonicity holds or not. Nevertheless it is frequently
taken as a reasonable requirement, and would add one further constraint on the entries in Table 1:
Pr(R0 = 1, R1 = 0) = 0. Since in that table Pr(R0 = 1, R1 = 0) = x − 0.18, we can then compute
the exact value x = 0.18, so that PC = 0.18/0.30 = 0.6. Here (as in general), monotonicity allows
as to replace the interval given by (5) by its lower endpoint 1 − 1/RR. In this case, RR < 2 does
imply PC < 0.5.
Monitonicity, while appealing, is a strong, perhaps overstrong, assumption about the internal
workings of the causal black box. Moreover, because it is a property of the pair of values (R0 , R1 )
considered jointly, the fundamental problem of statistical causality implies that we can never
expect to obtain empirical evidence to let us decide whether or not monitonicity does in fact hold.
For that reason we take a different approach in the rest of this paper, focusing on using additional,
empirically observable, information to shed some light on internal causal mechanisms, and thus
improve causal inferences. In such cases, although we typically cannot obtain an exact point value
for the probability of causation, we may be able to narrow the interval bounds on it. We shall
particularly consider cases in which we might have information on additional covariates and/or
mediators that influence the causal pathway.
7
7
Covariate information
Suppose that, in the experimental data, we can also observe an additional covariate, X — an
individual characteristic that can vary from person to person, and can affect that person’s response.
This situation is represented by the diagram of Figure 2. For simplicity, we suppose that X can
E
R
X
Figure 2: Covariate
only take a discrete set of values. We can then estimate, from the data, the response probability,
Pr(R = 1 | E = e, X = x), among those individuals having covariate value X = x and exposure
level e (= 0 or 1).
In those cases that we are also able to measure X for Ann, say X = x∗ , we can simply restrict
the experimental subjects to those having the same covariate value (who are thus like Ann in all
relevant respects), and apply the analysis of § 4. We can then use formula (5), after first replacing
all the probabilities in (5) and (6) by their conditional versions, given X = x∗ .
However, it turns out that we can make use of the additional covariate information in the
experimental data, even when we cannot observe X for Ann. For this case, it is shown in Dawid
(2011) that we have the interval bounds
∆
Γ
≤ PC ≤ 1 −
Pr(R = 1 | E = 1)
Pr(R = 1 | E = 1)
(8)
where
∆
=
X
Pr(X = x) × max {0, Pr(R = 1 | E = 1, X = x) − Pr(R = 1 | E = 0, X = x)}
x
Γ
=
X
Pr(X = x) × max {0, Pr(R = 1 | E = 1, X = x) − Pr(R = 0 | E = 0, X = x)} ,
x
and that this interval is always contained in that given by (5), which ignores the additional
covariate information.
Example 7.1 Suppose that the covariate X is binary, taking values 0 and 1 with equal probability.
Suppose further that, from the experimental data, we obtain the following probabilities:
Pr(R = 1 | E = 1, X = 0)
=
0.12
Pr(R = 1 | E = 0, X = 0)
=
0.24
Pr(R = 1 | E = 1, X = 1)
=
0.60
Pr(R = 1 | E = 0, X = 1)
=
0.12.
These imply marginal recovery probabilities (not taking the value of X into account):
Pr(R = 1 | E = 1)
=
0.36
Pr(R = 1 | E = 0)
=
0.18.
(i). If we had not observed X, either in the data or in Ann, we would use the marginal values,
to obtain a relative risk of 2, and interval bounds
0.5 ≤ PC ≤ 1.
8
(ii). If we observe X both in the data and in Ann, we obtain
0.8 ≤ PC ≤ 1
if Ann has X = 1, and
0 ≤ PC ≤ 1
if Ann has X = 0.
(iii). Finally, if we have observed X in the data but not in Ann, the relevant interval becomes
0.67 ≤ PC ≤ 1,
an improvement on that of (i)
2
8
Sufficient covariate
In § 7 above it was assumed that the assignment of exposure E was unrelated to the covariate
X. Here we relax this, and allow X to influence both the exposure E and the response R.
But we assume that there are no further, unobserved, variables that might act to confound this
relationship—X is a “sufficient covariate” (Guo and Dawid 2010; Dawid 2015) for assessing the
causal effect of E on R. The relevant diagram is now that of Figure 3. We assume that the
E
R
X
Figure 3: Sufficient covariate
identical same structure holds for the individuals in the data and for Ann. X is observed in the
data, but possibly not for Ann.
When X is observed for Ann, say X = x∗ , we can again apply the basic inequalities (5), after
replacing the probabilities in (5) and (6) by their conditional versions, given X = x∗ .
When X is not observed for Ann, the interval bounds for PC become:
X
1
×
max{0, {Pr(R = 1 | E = 1, X = x) − Pr(R = 1 | E = 0, X = x)}
Pr(R = 1 | E = 1)
x
× Pr(X = x | E = 1)
≤ PC ≤
1−
X
1
×
max{0, {Pr(R = 1 | E = 1, X = x) − Pr(R = 0 | E = 0, X = x)}
Pr(R = 1 | E = 1)
x
× Pr(X = x | E = 1).
Now in this case we can alternatively use the “back-door formula” (Pearl 2009):
X
Pr(R0 = 1) =
Pr(R = 1 | E = 0, X = x) × Pr(X = x)
x
9
(9)
(10)
to recreate the experimental response rate among the controls—and so compute the interval (7).
However, this would be tantamount to ignoring the additional information on X in the data. It
can be shown that the interval given by (9) is always contained in that given by (7).
Example 8.1 Suppose X is binary, and from the data we obtain the following probabilities:
Pr(X = 1)
=
0.5
Pr(E = 1 | X = 0)
=
0.8
Pr(E = 1 | X = 1)
=
0.2
Pr(R = 1 | E = 1, X = 1)
=
0.2
Pr(R = 1 | E = 0, X = 1)
=
0.8
Pr(R = 1 | E = 1, X = 0)
=
0.8
Pr(R = 1 | E = 0, X = 0)
=
0.2.
Then we obtain the following lower bounds for the probability of causation (the upper bound
being 1 in all cases):
When Ann is observed to have X = 1:
When Ann is observed to have X = 0:
When Ann’s value for X is not observed:
PC ≥ 0
PC ≥ 0.75
PC ≥ 0.71
Ignoring X and using the Tian and Pearl (2000) bounds :
PC ≥ 0.53.
2
9
Complete mediator
We now turn to consider the case, represented by the diagram in Figure 4, that the causal effect
of E on R is completely mediated by some variable M , assumed for simplicity to be binary. That
is to say, E affects M and M effects R, and the combination of these two pathways constitutes the
totality of the effect of E on R. We assume that there is no further confounding of either of these
E
M
R
Figure 4: Complete mediator
relationships, either in the data or for Ann. We shall consider the case that all three variables
(E, M, R) are observed in the data, but that we only observe E and R for Ann.
Note that our assumptions imply the conditional independence property R ⊥⊥ E | M , so that
X
Pr(R = r | E = e) =
Pr(R = r | M = m)Pr(M = m | E = e),
(11)
m
which should hold, at least approximately, in the data. In any event, we should estimate Pr(R = r | E = e)
using formula (11), rather than directly. (We could also have separate experimental data on the
E → M and M → R relationships, and apply (11).)
It is shown in Dawid et al. (2016a) that (with this understanding) the additional information
on the mediator does not affect the lower bound in (5). However, we obtain the following improved
10
upper bound:
1
Pr(R = 1 | E = 1)
× min{
Pr(M = 0 | E = 0) Pr(R = 0 | M = 0) + Pr(R = 0 | M = 1) Pr(M = 0 | E = 1),
Pr(M = 1 | E = 1) Pr(R = 0 | M = 0) + Pr(R = 0 | M = 1) Pr(M = 1 | E = 0),
Pr(M = 0 | E = 0) Pr(R = 1 | M = 1) + Pr(R = 1 | M = 0) Pr(M = 0 | E = 1),
Pr(M = 1 | E = 1) Pr(R = 1 | M = 1) + Pr(R = 1 | M = 0) Pr(M = 1 | E = 0) }.
(12)
Example 9.1 Suppose we obtain the following values from the data:
Pr(M = 1 | E = 1)
=
0.25
Pr(M = 1 | E = 0)
=
0.025
Pr(R = 1 | M = 1)
=
0.9
Pr(R = 1 | M = 0)
=
0.1.
Marginalising over M using (11), we obtain the same values as in Table 1. On applying (12),
we get 0.60 ≤ PC ≤ 0.76; whereas without taking account of the mediator M the bounds were
0.6 ≤ PC ≤ 1.
2
We can elaborate the above analysis by allowing for an additional sufficient covariate X, as
represented by Figure 5, that can modify all the above marginal and conditional probabilities, and
conditional on which M is a complete mediator.
E
M
R
X
Figure 5: Complete mediator with covariate
We assume that X is always observed in the data. If we observe X = x for Ann, we simply
replace all probabilities by their versions further conditioned on X = x, yielding bounds l(x) ≤
PC ≤ u(x), say. If we do not observe X in Ann, the lower bound is E{l(X) | E = 1, R = 1)}, and
the upper bound is E{u(X) | E = 1, R = 1)}.
10
Partial mediator
In the case of a partial mediator, as represented by Figure 6, there us an additional “direct effect”
of E on R, that is not mediated by M (again supposed binary, for simplicity). So now we have
to consider the effect of E on M , and the joint effect of E and M on R. Again we suppose that
there is no further confounding of these relationships, either in the data or for Ann, and consider
the case that all three variables (E, M, R) are observed in the data, but that we only observe E
and R for Ann.
In Murtas et al. (2016) it is shown that, once again, the lower bound for PC is unaffected by
11
E
M
R
Figure 6: Partial mediator
the additional information on M . However, we are able to abtain a new upper bound u, where
Pr(R = 1 | E = 1) × u =
min {Pr(R = 0 | E = 0, M = 0), Pr(R = 1 | E = 1, M = 0)}
× min {Pr(M = 0 | E = 0), Pr(M = 0 | E = 1)}
+ min {Pr(R = 0 | E = 0, M = 0), Pr(R = 1 | E = 1, M = 1)}
× min {Pr(M = 0 | E = 0), Pr(M = 1 | E = 1)}
+ min {Pr(R = 0 | E = 0, M = 1), Pr(R = 1 | E = 1, M = 0)}
× min {Pr(M = 1 | E = 0), Pr(M = 0 | E = 1)}
+ min {Pr(R = 0 | E = 0, M = 1), Pr(R = 1 | E = 1, M = 1)}
× min {Pr(M = 1 | E = 0), Pr(M = 1 | E = 1)} .
(13)
Example 10.1 Suppose that, from the data, we estimate the following probabilities:
Pr(M = 1 | E = 0)
=
0.27
Pr(M = 1 | E = 1)
=
0.019
Pr(R = 1 | E = 0, M = 0)
=
0.02
Pr(R = 1 | E = 0, M = 1)
=
0.835
Pr(R = 1 | E = 1, M = 0)
=
0.685
Pr(R = 1 | E = 1, M = 1)
=
0.857.
Marginalising over M , these imply
Pr(R = 1 | E = 0)
=
0.24
Pr(R = 1 | E = 1)
=
0.69.
We then obtain 0.65 ≤ PC ≤ 0.81 when accounting for the mediator, compared with 0.65 ≤ PC ≤ 1
when ignoring it.
2
However, accounting for a partial mediator may not always yield improved bounds.
Example 10.2 Suppose the probabilities estimated from the data are as follows:
Pr(M = 1 | E = 0)
=
0.96
Pr(M = 1 | E = 1)
=
0.74
Pr(R = 1 | E = 0, M = 0)
=
0.02
Pr(R = 1 | E = 0, M = 1)
=
0.33
Pr(R = 1 | E = 1, M = 0)
=
0.91
Pr(R = 1 | E = 1, M = 1)
=
0.73.
These imply
Pr(R = 1 | E = 0)
=
0.32
Pr(R = 1 | E = 1)
=
0.78.
The lower bound is PC ≥ 0.59. The upper bound obtained from (13) is PC ≤ 0.95. However, just
ignoring M , and applying (5), we get upper bound PC ≤ 0.88, which should theregore be used in
preference.
2
12
Again, we can elaborate the problem by including a sufficient covariate X, conditional on
which we have the above structure, as represented by Figure 7. Let the appropriate lower and
upper bounds, computed with all probabilities conditioned on X = x, be l(x) and u(x). Then
in the absence of information about X for Ann, the relevant lower and upper bounds for PC are
E{l(X) | E = 1, R = 1}, E{u(X) | E = 1, R = 1}.
E
M
R
X
Figure 7: Partial mediator with covariate
11
Discussion
We have assumed throughout that the available data are sufficiently extensive to allow essentially
perfect estimates of the required probabilities. Of course this will never be the case. Dawid et al.
(2016b) consider how to make suitable inferences about PC from limited data, using a Bayesian
approach.
We have presented some simple cases, involving covariates and mediators, where having additional information about the internal structure of the causal link between exposure and response
enables us to refine our causal inferences. Other problems might involve a more complex collection
of variables and pathways, perhaps expressed as a directed acyclic graph. It should be possible to
extend our analysis to such more general cases. Information on different causal subsystems might
be obtained from a variety of scientific investigations. But caution must be exercised in justifying
the transfer of such information from generic scientific data to the individual case. Additional
uncertainty over the correct representation of the problem must also be taken into account.
To return, finally, to the issue facing a court of law that has to return a decision on causality:
we hope we have shown that this is not a simple matter! We can only express our sympathy with
a judge who has to make such a decision, on the “balance of probabilities,” when the best possible
conclusion, based on available scientific evidence and the kind of analysis we have presented here,
is that the probability of causation lies between 0.4 and 0.8.
References
Dawid, A. P. (2000). Causal inference without counterfactuals (with Discussion). Journal of the
American Statistical Association, 95, 407–48.
Dawid, A. P. (2011). The role of scientific and statistical evidence in assessing causality. In
Perspectives on Causation, (ed. R. Goldberg), pp. 133–147. Hart Publishing, Oxford.
Dawid, A. P. (2015). Statistical causality from a decision-theoretic perspective. Annual Review of
Statistics and its Application, 2, 273–303.
DOI:10.1146/annurev-statistics-010814-020105.
Dawid, A. P., Faigman, D. L., and Fienberg, S. E. (2014). Fitting science into legal contexts:
Assessing effects of causes or causes of effects? (with Discussion and authors’ rejoinder). Sociological Methods and Research, 43, 359–421. DOI:10.1177/0049124113515188.
Dawid, A. P., Faigman, D. L., and Fienberg, S. E. (2015). On the causes of effects: Response to
Pearl. Sociological Methods and Research, 44, 165–74. DOI:10.1177/0049124114562613.
13
Dawid, A. P., Murtas, R., and Musio, M. (2016a). Bounding the probability of causation in
mediation analysis. In Topics on Methodological and Applied Statistical Inference, (ed. T. D.
Battista, E. Moreno, and W. Racugno), pp. 75–84. Springer.
Dawid, A. P., Musio, M., and Fienberg, S. E. (2016b). From statistical evidence to evidence of
causality. Bayesian Analysis, 11, 725–52.
Faigman, D. L., Monahan, J., and Slobogin, C. (2014). Group to individual (G2i) inference in
scientific expert testimony. University of Chicago Law Review, 81, 417–80.
Guo, H. and Dawid, A. P. (2010). Sufficient covariates and linear propensity analysis. Journal
of Machine Learning Research Workshop and Conference Proceedings, 9, 281–8. Proceedings
of the Thirteenth International Workshop on Artificial Intelligence and Statistics (AISTATS)
2010, Chia Laguna, Sardinia, Italy, May 13–15, 2010, edited by Y. W. Teh and D. M. Titterington.
http://jmlr.csail.mit.edu/proceedings/papers/v9/guo10a/guo10a.pdf.
Halpern, J. Y. (2016). Actual Causality. MIT Press.
Holland, P. W. (1986). Statistics and causal inference (with Discussion). Journal of the American
Statistical Association, 81, 945–970.
Hutchinson, T. A., Dawid, A. P., Spiegelhalter, D. J., Cowell, R. G., and Roden, S. (1991).
Computerized aids for the probabilistic assessment of drug safety II: An expert system. Drug
Information Journal, 25, 41–8.
Lewis, D. K. (1973). Counterfactuals. Blackwell, Oxford.
Menzies, P. (2014). Counterfactual theories of causation. The Stanford Encyclopedia of Philosophy
(Spring 2014 Edition), Edward N. Zalta (ed.). https://plato.stanford.edu/archives/
spr2014/entries/causation-counterfactual/.
Murtas, R., Dawid, A. P., and Musio, M. (2016). New bounds for the probability of causation in
mediation analysis. arXiv:1706.04857.
Neyman, J. (1935). Statistical problems in agricultural experimentation (with Discussion). Journal
of the Royal Statistical Society, Supplement, 2, 107–80.
Pearl, J. (2009). Causality: Models, Reasoning and Inference, (second edn). Cambridge University
Press, Cambridge.
Rubin, D. B. (1978). Bayesian inference for causal effects: The rôle of randomization. Annals of
Statistics, 6, 34–68.
Tian, J. and Pearl, J. (2000). Probabilities of causation: Bounds and identification. Annals of
Mathematics and Artificial Intelligence, 28, 287–313.
14
| 10math.ST
|
arXiv:1609.06418v1 [math.ST] 21 Sep 2016
Multiple Testing via Relative Belief Ratios
Michael Evans and Jabed Tomal
Department of Statistical Sciences
University of Toronto
Abstract: Some large scale inference problems are considered based on using the
relative belief ratio as a measure of statistical evidence. This approach is applied
to the multiple testing problem. A particular application of this is concerned
with assessing sparsity. The approach taken to sparsity has several advantages
as it is based on a measure of evidence and does not require that the prior be
restricted in any way.
Key words and phrases: multiple testing, sparsity, statistical evidence, relative
belief ratios, priors, checking for prior-data conflict, testing for sparsity, relative
belief multiple testing algorithm.
1
Introduction
Suppose there is a statistical model {fθ : θ ∈ Θ} and the parameter of interest
is given by Ψ : Θ → Ψ (we don’t distinguish between the function and its
range to save notation) where Ψ is an open subset of Rk . Our concern here
is with assessing individual hypotheses H0i = {θ : Ψi (θ) = ψ0i }, namely, H0i
is the hypothesis that the i-th coordinate of ψ equals ψ0i . If it is known that
ψi = Ψi (θ) = ψ0i , then the effective dimension of the parameter of interest
is k − 1 which results in a simplification of the model. For example, the ψi
may be linear regression coefficients and ψ0i = 0 is equivalent to saying that an
individual term in the regression is dropped and so simplifies the relationship.
Considering all these hypotheses separately is the multiple testing problem
and the concern is to ensure that, while controlling the individual error rate,
the overall error rate does not become too large especially when k is large. An
error means either the acceptance of H0i when it is false (a false negative) or the
rejection of H0i when it is true (a false positive). This problem is considered
here from a Bayesian perspective. A first approach is to make an inference
about the number of H0i that are true (or false) and then use this to control
the number of H0i that are accepted (or rejected). Problems arise when k is
large, however, due to an issue that arises with typical choices made for the
prior. This paper considers these problems and proposes solutions.
The problem of making inferences about sparsity is clearly related. For
suppose that there is a belief that many of the hypotheses H0i are indeed true
1
but there is little prior knowledge about which of the H0i are true. This is
effectively the sparsity problem. In this case it is not clear how to choose a
prior that reflects this belief. A common approach in the regression problem is
to use a prior that, together with a particular estimation procedure, forces many
of the ψi to take the value ψ0i . A difficulty with such an approach is that it is
possible that such an assignment is an artifact of the prior and the estimation
procedure and not primarily data-based. For example, the use of a Laplace prior
together with MAP estimation is known to accomplish this in certain problems.
It would be preferable, however, to have a procedure that was not dependent
on a specific form for the prior but rather worked with any prior and was based
on the statistical evidence for such an assignment contained in the data. The
methodology for multiple testing developed here accomplishes this goal.
The prior π on θ clearly plays a key role in our developments. Suppose
that π is decomposed as π(θ) = π(θ | ψ)πΨ (ψ) where π(θ | ψ) is the conditional
prior on θ given Ψ(θ) = ψ and πΨ is the marginal prior on ψ. If πΨ is taken
to be too concentrated about ψ0 , then there is the risk of encountering priordata conflict whenever some of the H0i are false. Prior-data conflict arises
whenever the true value of the parameter lies in a region of low prior probability,
such as the tails of the prior, and can lead to erroneous inferences. Consistent
methods for addressing prior-data conflict are developed in Evans and Moshonov
(2006), Evans and Jang (2011a) and a methodology for modifying the prior,
when a conflict is detected, is presented in Evans and Jang (2011b). Generally,
a check for prior-data conflict is made to ensure that the prior has been chosen
reasonably, just as model checking is performed to ensure that the model is
reasonable, given the observed data. The avoidance of prior-data conflict plays
a key role in our developments.
The assessment of the evidence for the truth of H0i is based on a measure of
the evidence that H0i is true as given by the relative belief ratio for ψi . Relative
belief ratios are similar to Bayes factors and, when compared to p-values, can be
considered as more appropriate measures of statistical evidence. For example,
evidence can be obtained either for or against a hypothesis. Moreover, there is
a clear assessment of the strength of this evidence so that when weak evidence
is obtained, either for or against a hypothesis, this does not entail acceptance or
rejection, respectively. Relative belief ratios and the theory of inference based
on these quantities, is discussed in Evans (2015) and a general outline is given
in Section 2. Section 3 considers applying relative belief to the multiple testing
problem. In Section 4 some practical applications are presented with special
emphasis on regression problems including the situation where the number of
predictors exceeds the number of observations.
There have been a number of priors proposed for the sparsity problem in
the literature. A common choice is the spike-and-slab prior due to George and
McCulloch (1993). Other priors are available that are known to produce sparsity
at least in connection with MAP estimation. For example, the Laplace prior
has a close connection with sparsity via the LASSO, as discussed in Tibshirani
(1996), Park and Casella (2008) and Hastie, Tibshirani and Wainwright (2015).
The horseshoe prior of Carvalho, Polson and Scott (2009) will also produce
2
sparsity through estimation. Both the Laplace and the horseshoe prior can
be checked for prior-data conflict but the check is not explicitly for sparsity,
only that the location and scaling of the prior is appropriate. This reinforces a
comment in Carvalho, Polson and Scott (2009) that the use of the spike-andslab prior represents a kind of ”gold standard” for sparsity. The problem with
the spike-and-slab prior, and not possessed to the same degree by the Laplace
or horseshoe prior, is the difficulty of the computations when implementing the
posterior analysis. Basically there are 2k mixture components to the posterior
and this becomes computationally intractable as k rises. Various approaches
can be taken to try and deal with this issue such as those developed in Rockova
and George (2014). Any of these priors can be used with the approach taken
here but our methodology does not require that a prior induces sparsity in any
sense. Indeed sparsity is induced only when the evidence, as measured by how
the data changes beliefs, points to this. This is a strong point of the approach
as sparsity arises only through an assessment of the evidence.
2
Inferences Based on Relative Belief Ratios
Suppose now that, in addition to the statistical model {fθ : θ ∈ Θ} there is a
prior π on Θ. After observing data x, the posterior distribution
of θ is then given
R
by the density π(θ | x) = π(θ)fθ (x)/m(x) where m(x) = Θ π(θ)fθ (x) dθ is the
prior predictive density of x. For a parameter of interest Ψ(θ) = ψ denote the
prior and posterior densities of ψ by πΨ and πΨ (· | x), respectively. The relative
belief ratio for a value ψ is then defined by RBΨ (ψ | x) = limδ→0 ΠΨ (Nδ (ψ )| x)/
ΠΨ (Nδ (ψ )) where Nδ (ψ ) is a sequence of neighborhoods of ψ converging (nicely)
to {ψ} as δ → 0. Quite generally
RBΨ (ψ | x) = πΨ (ψ | x)/πΨ (ψ),
(1)
the ratio of the posterior density to the prior density at ψ. So RBΨ (ψ | x) is
measuring how beliefs have changed concerning ψ being the true value from a
priori to a posteriori. When RBΨ (ψ | x) > 1 the data have lead to an increase
in the probability that ψ is correct and so there is evidence in favor of ψ, when
RBΨ (ψ | x) < 1 the data have lead to a decrease in the probability that ψ is
correct and so there is evidence against ψ, and when RBΨ (ψ | x) = 1 there is no
evidence either way. While there are numerous quantities that measure evidence
through change in belief, the relative belief ratio is perhaps the simplest and
it possesses many nice properties as discussed in Evans (2015). For example,
RBΨ (ψ | x) is invariant under smooth changes of variable and also invariant
to the choice of the support measure for the densities. As such all relative
belief inferences possess this invariance which is not the case for many Bayesian
inferences such as using a posterior mode (MAP) or expectation for estimation.
The value RBΨ (ψ0 | x) measures the evidence for the hypothesis H0 = {θ :
Ψ(θ) = ψ0 }. It is also necessary to calibrate whether this is strong or weak
evidence for or against H0 . Certainly the bigger RBΨ (ψ0 | x) is than 1, the more
evidence there is in favor of ψ0 while the smaller RBΨ (ψ0 | x) is than 1, the more
3
evidence there is against ψ0 . But what exactly does a value of RBΨ (ψ0 | x) = 20
mean? It would appear to be strong evidence in favor of ψ0 because beliefs have
increased by a factor of 20 after seeing the data. But what if other values of ψ
had even larger increases? A calibration of (1) is then given by the strength
ΠΨ (RBΨ (ψ | x) ≤ RBΨ (ψ0 | x) | x),
(2)
namely, the posterior probability that the true value of ψ has a relative belief
ratio no greater than that of the hypothesized value ψ0 . While (2) may look
like a p-value, it has a very different interpretation. For when RBΨ (ψ0 | x) < 1,
so there is evidence against ψ0 , then a small value for (2) indicates a large
posterior probability that the true value has a relative belief ratio greater than
RBΨ (ψ0 | x) and so there is strong evidence against ψ0 while only weak evidence
against if (2) is big. If RBΨ (ψ0 | x) > 1, so there is evidence in favor of ψ0 , then a
large value for (2) indicates a small posterior probability that the true value has
a relative belief ratio greater than RBΨ (ψ0 | x)) and so there is strong evidence
in favor of ψ0 , while a small value of (2) only indicates weak evidence in favor
of ψ0 . Notice that in {ψ : RBΨ (ψ | x) ≤ RBΨ (ψ0 | x)}, the ‘best’ estimate of ψ
is given by ψ0 because the evidence for this value is the largest in the set.
Various results have been established in Baskurt and Evans (2013) and Evans
(2015) supporting both (1), as the measure of the evidence, and (2) as a measure of the strength of that evidence. For example, the following simple inequalities, see Baskurt and Evans (2013), are useful in assessing the strength
as computing (2) can be avoided, namely, ΠΨ ({ψ0 } | x) ≤ ΠΨ (RBΨ (ψ | x) ≤
RBΨ (ψ0 | x) | x) ≤ RBΨ (ψ0 | x). So if RBΨ (ψ0 | x) > 1 and ΠΨ ({ψ0 } | x) is large,
there is strong evidence in favor of ψ0 while, if RBΨ (ψ0 | x) < 1 is very small,
then there is immediately strong evidence against ψ0 .
The use of RBΨ (ψ0 | x) to assess the hypothesis H0 also possesses optimality
properties. For example, let A be a subset of the sample space such that whenever x ∈ A, the hypothesis is accepted. If M (A | H0 ) denotes the prior predictive
probability of A given that H0 is true, then M (A | H0 ) is the prior probability
of accepting H0 when it is true. The relative belief acceptance region is naturally given by Arb (ψ0 ) = {x : RBΨ (ψ0 | x) > 1}. Similarly, let R be a subset
such that whenever x ∈ R, the hypothesis is rejected and let M (R | H0c ) denote
the prior predictive probability of R given that H0 is false. The relative belief
rejection region is then naturally given by Rrb (ψ0 ) = {x : RBΨ (ψ0 | x) < 1}.
The following result is proved in Evans (2015) as Proposition 4.7.9.
Theorem 1 (i) The acceptance region Arb (ψ0 ) minimizes M (A) among all
acceptance regions A satisfying M (A | H0 ) ≥ M (Arb (ψ0 ) | H0 ). (ii) The rejection region Rrb (ψ0 ) maximizes M (R) among all rejection regions R satisfying
M (R | H0 ) ≤ M (Rrb (ψ0 ) | H0 ).
To see the meaning of this result note that
M (A) = EΠΨ (M (A | ψ)) = EΠΨ (IH0c (ψ)M (A | ψ)) + ΠΨ (H0 )M (A | H0 )
≥ M (Arb (ψ0 )) = EΠΨ (IH0c (ψ)M (Arb (ψ0 ) | ψ)) + ΠΨ (H0 )M (Arb (ψ0 ) | H0 ).
4
Therefore, if ΠΨ (H0 ) = 0, then Arb (ψ0 ) minimizes EΠΨ (I{ψ0 }c (ψ)M (A | ψ)), the
prior probability that H0 is accepted given that it is false, among all acceptance
regions A satisfying M (A | ψ0 ) ≥ M (Arb (ψ0 ) | ψ0 ) and when ΠΨ (H0 ) > 0, then
Arb (ψ0 ) minimizes this probability among all acceptance regions A satisfying
M (A | H0 ) = M (Arb (ψ0 ) | H0 ). Also,
M (R) = EΠΨ (M (R | ψ)) = EΠΨ (IH0c (ψ)M (R | ψ)) + ΠΨ (H0 )M (R | H0 )
≤ M (Rrb (ψ0 )) = EΠΨ (IH0c (ψ)M (Rrb (ψ0 ) | ψ)) + ΠΨ (H0 )M (Rrb (ψ0 ) | H0 ).
Therefore, if ΠΨ ({ψ0 }) = 0, then Rrb (ψ0 ) maximizes EΠΨ (I{ψ0 }c (ψ)M (R | ψ)),
the prior probability that H0 is rejected given that it is false, among all rejection
regions R satisfying M (R | H0 ) ≤ M (Rrb (ψ0 ) | H0 ) and when ΠΨ (H0 }) > 0,
then Rrb (ψ0 ) maximizes this probability among all rejection regions R satisfying
M (R | ψ0 ) = M (Rrb (ψ0 ) | ψ0 ).
Note that it does not make sense to accept or reject H0 when RBΨ (ψ0 | x) =
1. Also, under i.i.d. sampling, M (Arb (ψ0 ) | H0 ) → 1 and M (Rrb (ψ0 ) | H0 ) →
0 as sample size increases, so these quantities can be controlled by design.
When ΠΨ ({ψ0 }) = 0, then M (Arb (ψ0 ) | H0 ) = 1 − M (Rrb (ψ0 ) | H0 ) and so
controlling M (Arb (ψ0 ) | H0 ) is controlling the ”size” of the test. In general,
EΠΨ (I{ψ0 }c (ψ)M (Rrb (ψ0 ) | ψ)) can be thought of as the Bayesian power of the
relative belief test. Note that it is reasonable to set either the probability of a
false positive or the probability of a true negative as part of design and so the
theorem is an optimality result with practical import. It is easily seen that the
proof of Theorem 1 can be generalized to obtain optimality results for the acceptance region Arb,q (ψ0 ) = {x : RBΨ (ψ0 | x) > q} and for the rejection region
Rrb,q (ψ0 ) = {x : RBΨ (ψ0 | x) < q}. The following inequality is useful in Section
3 in controlling error rates.
Theorem 2 M ( Rrb,q (ψ0 ) | ψ0 ) ≤ q.
Proof: By the Savage-Dickey result, see Proposition 4.2.7 in Evans (2015),
RBΨ (ψ0 | x) = m(x | ψ0 )/m(x). Now EM (· | ψ0 ) (m(x)/m(x | ψ0 )) = 1 and so by
Markov’s inequality, M ( Rrb,q (ψ0 ) | ψ0 ) = M ( m(x)/m(x | ψ0 ) > 1/q | ψ0 ) ≤ q.
There is another issue associated with using RBΨ (ψ0 | x) to assess the evidence that ψ0 is the true value. One of the key concerns with Bayesian inference
methods is that the choice of the prior can bias the analysis in various ways.
For example, in many problems Bayes factors and relative belief ratios can be
made arbitrarily large by choosing the prior to be increasingly diffuse. This phenomenon is associated with the Jeffreys-Lindley paradox and clearly indicates
that it is possible to bias results by the choice of the prior.
An approach to dealing with this bias is discussed in Baskurt and Evans
(2013). For, given a measure of evidence, it is possible to measure a priori
whether or not the chosen prior induces bias either for or against ψ0 . The bias
against ψ0 is given by
M ( RBΨ (ψ0 | x) ≤ 1 | ψ0 ) = 1 − M (Arb (ψ0 ) | ψ0 )
5
(3)
as this is the prior probability that evidence will not be obtained in favor of ψ0
when ψ0 is true. If (3) is large, subsequently reporting, after seeing the data,
that there is evidence against ψ0 is not convincing.
To measure the bias in favor of ψ0 choose values ψ00 6= ψ0 such that the
difference between ψ0 and ψ00 represents the smallest difference of practical importance. Then compute
MT ( RBΨ (ψ0 | x) ≥ 1 | ψ00 ) = 1 − M (Rrb (ψ0 ) | ψ00 )
(4)
as this is the prior probability that we will not obtain evidence against ψ0 when
ψ0 is false. Note that (4) tends to decrease as ψ00 moves further away from ψ0 .
When (4) is large, subsequently reporting, after seeing the data, that there is
evidence in favor of ψ0 , is not convincing. For a fixed prior, both (3) and (4)
decrease with sample size and so, in design situations, they can be used to set
sample size and so control bias. Notice that M (Arb (ψ0 ) | ψ0 ) can be considered
as the sensitivity and M (Rrb (ψ0 ) | ψ00 ) as the specificity of the relative belief
hypothesis assessment. These issues are further discussed in Evans (2015).
As RBΨ (ψ | x) measures the evidence that ψ is the true value, it can also be
used to estimate ψ. For example, the best estimate of ψ is clearly the value for
which the evidence is greatest, namely, ψ(x) = arg sup RBΨ (ψ | x). Associated
with this is a γ-credible region CΨ,γ (x) = {ψ : RBΨ (ψ | x) ≥ cΨ,γ (x)} where
cΨ,γ (x) = inf{k : ΠΨ (RBΨ (ψ | x) > k | x) ≤ γ}. Notice that ψ(x) ∈ CΨ,γ (x) for
every γ ∈ [0, 1] and so, for selected γ, we can take the ”size” of CΨ,γ (x) as a
measure of the accuracy of the estimate ψ(x). The interpretation of RBΨ (ψ | x)
as the evidence for ψ forces the sets CΨ,γ (x) to be the credible regions. For if
ψ1 is in such a region and RBΨ (ψ2 | x) ≥ RBΨ (ψ1 | x), then ψ2 must be in the
region as well as there is at least as much evidence for ψ2 as for ψ1 . A variety of
optimality results have been established for ψ(x) and CΨ,γ (x), see Evans (2015).
The view is taken here that any time continuous probability is used, then
this is an approximation to a finite, discrete context. For example, if ψ is a
mean and the response measurements are to the nearest centimeter, then of
course the true value of ψ cannot be known to an accuracy greater than 1/2
of a centimeter, no matter how large a sample we take. Furthermore, there
are implicit bounds associated with any measurement process. As such the
restriction is made here to discretized parameters that take only finitely many
values. So when ψ is a continuous, real-valued parameter, it is discretized to
the intervals . . . , (ψ0 − 3δ, ψ0 − δ], (ψ0 − δ, ψ0 + δ], (ψ0 + δ, ψ0 + 3δ], . . . for some
choice of δ > 0, and there are only finitely may such intervals covering the range
of possible values. It is of course possible to allow the intervals to vary in length
as well. With this discretization, then we can take H0 = (ψ0 − δ, ψ0 + δ].
3
Inferences for Multiple Tests
Consider now
Pk the multiple testing problem discussed in Section 1. Let ξ =
Ξ(θ) = k1 i=1 IH0i (Ψi (θ)) be the proportion of the hypotheses H0i that are
true and suppose that Ψi is finite for each i, perhaps arising via a discretization
6
as discussed in Section 2. Note that the discreteness is essential to realistically
determine what proportion of the hypotheses are correct, otherwise, under a
continuous prior on Ψ, the prior distribution of Ξ(θ) is degenerate at 0. In
an application it is desirable to make inference about the true value of ξ ∈
Ξ = {0, 1/k, 2/k, . . . , 1}. This is based on the relative belief ratio RBΞ (ξ | x) =
Π(Ξ(θ) = ξ | x)/Π(Ξ(θ) = ξ). The appropriate estimate of ξ is then the relative
belief estimate of Ξ, namely, ξ(x) = arg supξ RBΞ (ξ | x). The accuracy of this
estimate is assessed using the size of CΞ,γ (x) for some choice of γ ∈ [0, 1]. Also,
hypotheses such as H0 = {θ : Ξ(θ) ∈ [ξ0 , ξ1 ]}, namely, the proportion true
is at least ξ0 and no greater than ξ1 , can be assessed using the relative belief
ratio RB(H0 | x) = Π(ξ0 ≤ Ξ(θ) ≤ ξ1 | x)/Π(ξ0 ≤ Ξ(θ) ≤ ξ1 ) which equals
RBΞ (ξ0 | x) when ξ0 = ξ1 . The strength of this evidence can be assessed as
previously discussed.
The estimate ξ(x) can be used to control how many hypotheses are potentially accepted. For this select kξ(x) of the H0i as being true from among those
for which RBΨi (ψ0i | x) > 1. Note that it does not make sense to accept H0i as
being true when RBΨi (ψ0i | x) < 1 as there is evidence against this hypothesis.
So, if there are fewer than kξ(x) satisfying RBΨi (ψ0i | x) > 1, then fewer than
this number should be accepted. If there are more than kξ(x) of the relative
belief ratios satisfying RBΨi (ψ0i | x) > 1, then some method will have to be used
to select the kξ(x) which are potentially accepted. It is clear, however, that the
logical way to do this is to order the H0i , for which RBΨi (ψ0i | x) > 1, based on
their strengths ΠΨ (RBΨi (ψ0i | x) ≤ RBΨi (ψ0i | x) | x), from largest to smallest,
and accept at most the kξ(x) for which the evidence is strongest.
Note too that, if some of these strengths are indeed weak, there is no need
to necessarily accept these hypotheses. The ultimate decision as to whether
or not to accept a hypothesis is application dependent and is not statistical in
nature. The role of statistics is to supply a clear statement of the evidence and
its strength, while other criteria come into play when decisions are made. In any
case, it is seen that inference about ξ is being used to help control the number
of hypotheses accepted.
If, as is more common, control is desired of the number of false positives,
then the relevant parameter of interest is υ = Υ(θ) = 1−Ξ(θ), the proportion of
false hypotheses. Note that Π(Υ(θ) = υ) = Π(Ξ(θ) = 1 − υ) and so the relative
belief estimate of υ satisfies υ(x) = 1 − ξ(x). Following the same procedure, the
H0i with RBΨi (ψ0i | x) < 1 are then ranked via their strengths and at most
kυ(x) are rejected.
The consistency of the procedure just described, for correctly identifying
the H0i that are true and those that are false, follows from results proved in
Section 4.7.1 of Evans (2015) under i.i.d. sampling. In other words, as the
amount of data increases, ξ(x) converges to the proportion of H0i that are true,
each RB(ψ0i | x) converges to the largest possible value (always bigger than 1)
when H0i is true and converges to 0 when H0i is false, and the evidence in
favor or against converges to the strongest possible, depending on whether the
hypothesis in question is true or false.
We refer to this procedure as the multiple testing algorithm. Consider first
7
a somewhat artificial example where many of the computations are easy but
which fully demonstrates the relevant characteristics of the methodology.
Example 1. Location normal.
Suppose that there are k independent samples xij for 1 ≤ i ≤ k, 1 ≤ j ≤ n
where the i-th sample is from a N (µi , σ 2 ) distribution with µi unknown and
σ 2 known. It also assumed that prior knowledge about all the unknown µi
is reflected in the statement that the µ1 , . . . , µk are i.i.d. from a N (µ0 , λ20 σ 2 )
distribution. It is easy to modify our development to allow the sample sizes to
vary and to use a general multivariate normal prior, while the case where σ 2 is
unknown is considered in Section 4. This context is relevant to the analysis of
microarray data.
The value of (µ0 , λ20 ) is determined via elicitation. For this it is supposed
that it is known with virtual certainty that each µi ∈ (ml , mu ) for specified
values ml ≤ mu . Here virtual certainty is interpreted to mean that the prior
probability of this interval is at least 0.99 and other choices could be made.
It is also supposed that µ0 = (ml + mu )/2. This implies that λ0 = (mu −
ml )/(2σΦ−1 ((1 + 0.99)/2)). Following Evans and Jang (2011b), increasing the
value of λ0 implies a more weakly informative prior in this context and, as such,
decreases the possibility of prior-data conflict.
The posterior distributions of the µi are then independent with µi | x ∼
N (µi (x), (nλ20 +1)−1 λ20 σ 2 ) where µi (x) = (n+1/λ20 )−1 (nx̄i +µ0 /λ20 ). Given that
the measurements are taken to finite accuracy, it is not necessarily realistic to
test µi = µ0 . As such, a value δ > 0 is specified so that H0i = (µ0 −δ/2, µ0 +δ/2]
in a discretization of the parameter space into a finite number of intervals, each
of length δ, as well as two tail intervals. Then for T ∈ N there are 2T +1 intervals
of the form (µ0 + (t − 1/2)δ, µ0 + (t + 1/2)δ], for t ∈ {−T, −T + 1, . . . , T } that
span (ml , mu ), together with two additional tail intervals (−∞, µ0 − (T + 1/2)δ]
and (µ0 + (T + 1/2)δ, ∞) to cover the full range. The relative belief ratio for
the t-th interval for µi is then given by
RBi ((µ0 + (t − 1/2)δ, µ0 + (t + 1/2)δ] | x)
Φ((nλ20 + 1)1/2 (µ0 + (t + 1/2)δ − µi (x))/λ0 σ)−
Φ((nλ20 + 1)1/2 (µ0 + (t − 1/2)δ − µi (x))/λ0 σ)
=
Φ((t + 1/2)δ/λ0 σ) − Φ((t − 1/2)δ/λ0 σ)
(5)
with a similar formula for the tail intervals. When δ is small this is approximated
by the ratio of the posterior to prior densities of µi evaluated at µ0 + tδ. Then
RB(H0i | x) = RBi ((µ0 − δ/2, µ0 + δ/2] | x) gives the evidence for or against H0i
and the strength of this evidence is computed using the discretized posterior
distribution. Notice that (5) converges to ∞ as λ0 → ∞ and this is characteristic
of other measures of evidence such as Bayes factors. As discussed in Evans
(2015), this is one of the reasons why calibrating (1) via (2) is necessary.
As previously noted, it is important to take account of the bias in the prior.
To simplify matters the continuous approximation is used here as this makes
little difference for the discussion of bias concerning inference about µi (see
8
Tables 2 and 3). The bias against µi = µ0 equals
M (RBi (µ0 | x) ≤ 1 | µ0 ) = 2(1 − Φ(an (1)))
(6)
where
an (q) =
{(1 + 1/nλ20 ) log((nλ20 + 1)/q 2 )}1/2 , q 2 ≤ nλ20 + 1
0,
q 2 > nλ20 + 1.
Note that (6) converges to 2(1 − Φ(1)) = 0.32 as λ0 → 0 and to 0 as λ0 → ∞
and, for fixed λ0 , converges to 0 as n → ∞. So there is never strong bias against
µi = µ0 and this is as expected since the prior is centered on µ0 . The bias in
favor of µi = µ0 is measured by
√
√
M (RBi (µ0 | x) ≥ 1 | µ0 ± δ/2) = Φ( nδ/2σ + an (1)) − Φ( nδ/2σ − an (1)). (7)
As λ0 → ∞ then (7) converges to 1 so there is bias in favor of µi = µ0 and
this reflects what was obtained for the limiting value of (5). Also this decreases
with increasing δ and goes to 0 as n → ∞. So indeed bias of both types can
be controlled by sample size. Perhaps the most important take away from this
discussion, however, is that by using a supposedly noninformative prior with λ0
large, bias in favor of the H0i is being induced.
Consider first a simulated data set x when k = 10, n = 5, σ = 1, δ = 1, µ0 =
0, (ml , mu ) = (−5, 5), so that λ0 = 10/2Φ−1 (0.995) = 1.94 and suppose µ1 =
µ2 = · · · = µ7 = 0 with the remaining µi = 2. The relative belief ratio function
RBΞ (· | x) is plotted in Figure 1. In this case the relative belief estimate ξ(x) =
0.70 is exactly correct. Table 1 gives the values of the RBi (0 | x) together with
their strengths. It is clear that the multiple testing algorithm leads to 0 false
positives and 0 false negatives. So the algorithm works perfectly on this data
but of course it can’t be expected to do as well when the three nonzero means
move closer to 0. Also, it is worth noting that the strength of the evidence
in favor of µi = 0 is very strong for i = 1, 2, 3, 5, 6, 7 but only moderate when
i = 4. The strength of the evidence against µi = 0 is very strong for i = 8, 9, 10.
Note that the maximum possible value of RBi ((µ0 − δ/2, µ0 + δ/2] | x) here
is (2Φ(δ/2λ0 σ) − 1)−1 = 4.92, so indeed some of the relative belief ratios are
relatively large.
Now consider basically the same context but where k = 1000 and µ1 =
· · · = µ700 = 0 while the remaining 300 satisfy µi = 2. The relative belief ratio
RBΞ (· | x) is plotted in Figure 2. In this case ξ(x) = 0.47 which is a serious
underestimate. As such the multiple testing algorithm will not record enough
acceptances and so will fail.
This problem arises due to the independence assumption on the µi . For
the prior distribution of kΞ(θ) is binomial(k, 2Φ(δ/2λ0 σ) − 1) and the prior
distribution of kΥ(θ) is binomial(k, 2(1 − Φ(δ/2λ0 σ))). So the a priori expected
proportion of true hypotheses is 2Φ(δ/2λ0 σ) − 1 and the expected proportion of
false hypotheses is 2(1−Φ(δ/2λ0 σ)). When δ/2λ0 σ is small, as when the amount
of sampling variability or the diffuseness of the prior are large, then the prior on
9
Figure 1: A plot of the relative belief ratio of Ξ when n = 5, k = 10 and 7 means
equal 0 with the remaining means equal to 2 in Example 1.
i
µi
RBi (0 | x)
Strength
i
µi
RBi (0 | x)
Strength
1
0
3.27
1.00
6
0
3.00
1.00
2
0
3.65
1.00
7
0
3.43
1.00
3
0
2.98
1.00
8
2
2.09 × 10−4
4.25 × 10−5
4
0
1.67
0.37
9
2
3.99 × 10−4
8.11 × 10−5
5
0
3.57
1.00
10
2
8.80 × 10−3
1.83 × 10−3
Table 1: Relative belief ratios and strengths for the µi in Example 1 with k = 10.
Figure 2: A plot of the relative belief ratio of Ξ when n = 5, k = 1000 and 700
means equal 0 with the remaining means equal to 2 in Example 1.
10
Ξ suggests a belief in many false hypotheses. When k is small, relatively small
amounts of data can override this to produce accurate inferences about ξ or υ
but otherwise large amounts of data are needed which may not be available.
Given that accurate inference about ξ and υ is often not feasible, we focus
instead on protecting against too many false positives and false negatives. From
the discussion in Example 1 it is seen that it is possible to produce bias in favor
of the H0i being true simply by using a diffuse prior. If our primary goal is to
guard against too many false positives, then this biasing will work as we can
make the prior conditional probabilities of the events RB(ψ0i | x) < 1, given
that H0i is true, as small as desirable by choosing such a prior. This could
be seen as a way of creating a higher bar for a positive result. The price we
pay for this, however, is too many false negatives. As such we consider another
way of ”raising the bar”. Given that RB(ψ0i | x) is measuring evidence, a
natural approach is to simply choose constants 0 < qR ≤ 1 ≤ qA and classify
H0i as accepted when RB(ψ0i | x) > qA and rejected when RB(ψ0i | x) < qR .
The strengths can also be quoted to assess the reliability of these inferences.
Provided qR is greater than the minimum possible value of RB(ψ0i | x), and
this is typically 0, and qA is chosen less than the maximum possible value of
RB(ψ0i | x), and this is 1 over the prior probability of H0i , then this procedure
is consistent as the amount of data increases. In fact, the related estimates of ξ
and υ are also consistent. The price paid for this is that a hypothesis will not be
classified whenever qR ≤ RB(ψ0i | x) ≤ qA . Not classifying a hypothesis implies
that there is not enough evidence for this purpose and more data is required.
This approach is referred to as the relative belief multiple testing algorithm.
It remains to determine qA and qR . Consider first protecting against too
many false positives. The a priori conditional prior probability, given that
H0i is true, of finding evidence against H0i less than qR is M (RBi (ψ0i | X) <
qR | ψ0i ) = M (m(X | ψ0i )/m(x) < qR | ψ0i ) ≤ qR where the final inequality
follows from Theorem 2. Naturally, we want the probability of a false negative
to be small and so choosing qR small accomplishes this. The a priori probability
that a randomly selected hypothesis produces a false positive is
k
1X
M (RBi (ψ0i | X) < qR | ψ0i )
k i=1
(8)
which by Theorem 2 is bounded above by qR and so converges to 0 as qR → 0.
Also, for fixed qR , (8) converges to 0 as the amount of data increases. More
generally qR can be allowed to depend on i but when the ψi are similar in nature
this does not seem necessary. Furthermore, it is not necessary to weight the
hypotheses equally so a randomly chosen hypothesis with unequal probabilities
could be relevant in certain circumstances. In any case, controlling the value
of (8), whether by sample size or by the choice of qR , is clearly controlling for
false positives. Suppose there is proportion pF P of false positives that is just
tolerable in a problem. Then qR can be chosen so that (8) is less than or equal
to pF P and note that qR = pF P satisfies this.
11
0
0
Similarly, if ψ0i
6= ψ0i then M (RBi (ψ0i | X) > qA | ψ0i
) is the prior proba0
bility of finding evidence for H0i when ψ0i is the true value. For a given effect
0
size δ of practical importance it is natural to take ψ0i
= ψ0i ± δ/2. In typical
0
applications this probability becomes smaller the ”further” ψ0i
is from ψ0i and
so choosing qA to make this probability small will make it small for all alternatives. Under these circumstances the a priori probability that a randomly
selected hypothesis produces a false negative is bounded above by
k
1X
0
M (RBi (ψ0i | X) > qA | ψ0i
).
k i=1
(9)
As qA → ∞, or as the amount of data increases with qA fixed, then (9) converges
to 0 so the number of false negatives can be controlled. If there is proportion
pF N of false negatives that is just tolerable in a problem, then qA can be chosen
so that (9) is less than or equal to pF N .
The following optimality result holds for relative belief multiple testing.
Corollary 3 (i) Among all procedures where the prior probability of accepting
H0i , when it is true, is at least M (RBi (ψ0i | X) > qA | ψ0i ) for i = 1, . . . , k,
the relative belief multiple testing algorithm minimizes the prior probability that
a randomly chosen hypothesis is accepted. (ii) Among all procedures where
the prior probability of rejecting H0i , when it is true, is less than or equal to
M (RBi (ψ0i | X) < qR | ψ0i ), then the relative belief multiple testing algorithm
maximizes the prior probability that a randomly chosen hypothesis is rejected.
Proof : For (i) consider a procedure for multiple testing and let Ai be the set
of data values where H0i is accepted. Then by hypothesis M (RBi (ψ0i | X) >
qA | ψ0i ) ≤ M (Ai | ψ0i ) and by the analog of Theorem 1 M (Ai ) ≥ M (RBi (ψ0i | X)
> qA ). Applying this to a randomly chosen H0i gives the result. The proof of
(ii) is basically the same.
Applying the same discussion as after Theorem 1, it is seen that, under reasonable conditions, the relative belief multiple testing algorithm minimizes the prior
probability of accepting a randomly chosen H0i when it is false and maximizes
the prior probability of rejecting a randomly chosen H0i when it is false.
Consider now the application of the relative belief multiple testing algorithm
in the previous example.
Example 2. (Example 1 continued)
In this context (8) equals M (RBi (µ0 | x) < qR | µ0 ) = 2(1 − Φ(an (qR )) for
all i and so this is the value of (8). Therefore, qR is chosen to make this number
suitably small. Table 2 records values for (8). From this it is seen that for small
n there can be some bias against H0i (qR = 1) and so the prior probability
of obtaining false positives is perhaps too large. Table 2 demonstrates that
choosing a smaller value of qR can adequately control the prior probability of
false positives.
12
n
1
λ0
1
2
10
qR
1
1/2
1/10
1
1/2
1/10
1
1/2
1/10
(8)
0.239 (0.228)
0.041 (0.030)
0.001 (0.000
0.156 (0.146)
0.053 (0.045)
0.005 (0.004)
0.031 (0.026)
0.014 (0.011)
0.002 (0.002)
n
5
λ0
1
2
10
qR
1
1/2
1/10
1
1/2
1/10
1
1/2
1/10
(8)
0.143 (0.097)
0.051 (0.022)
0.006 (0.001)
0.074 (0.041)
0.031 (0.013)
0.005 (0.001)
0.013 (0.004)
0.006 (0.002)
0.001 (0.001)
Table 2: Prior probability a randomly chosen hypothesis produces a false positive, continuous and discretized ( ) versions, in Example 2.
n
1
λ0
1
2
10
qA
1.0
1.2
1.4
1.0
2.0
2.2
1.0
5.0
10.0
(9)
0.704 (0.715)
0.527 (0.503)
0.141 (0.000)
0.793 (0.805)
0.359 (0.304)
0.141 (0.000)
0.948 (0.955)
0.708 (0.713)
0.070 (0.000)
n
5
λ0
1
2
10
qA
1.0
2.0
2.4
1.0
3.0
4.5
1.0
10.0
22.0
(9)
0.631 (0.702)
0.302 (0.112)
0.095 (0.000)
0.747 (0.822)
0.411 (0.380)
0.084 (0.000)
0.916 (0.961)
0.552 (0.588)
0.080 (0.000)
Table 3: Prior probability a randomly chosen hypothesis produces a false negative when δ/σ = 1, continuous and discretized ( ) versions, in Example 2.
For false negatives consider (9) where
√
Φ( √nδ/2σ + an (qA ))−
Φ( nδ/2σ − an (qA )),
M (RBi (µ0 | x) > qA | µ0 ±δ/2) =
0,
2
≤
1 ≤ qA
2
nλ0 + 1
2
> nλ20 + 1.
qA
for all i. It is easy to show that this is monotone decreasing in δ and so it
is an upper bound on the expected proportion of false negatives among those
hypotheses that are actually false. The cutoff qA can be chosen to make this
number as small as desired. When δ/σ → ∞, then (9) converges to 0 and
increases to 2Φ(an (qA )) − 1 as δ/σ → 0. Table 3 records values for (9) when
δ/σ = 1 so that the µi differ from µ0 by one half of a standard deviation. There
is clearly some improvement but still the biasing in favor
p of false negatives is
readily apparent. It would seem that taking qA = nλ20 + 1 gives the best
results but this could be considered as quite conservative. It is also worth
remarking that all the entries in Table 3 can be considered as very conservative
when large effect sizes are expected.
Now consider the situation that led to Figure 2. For this k = 1000, n = 5
13
Decision
Accept µ = 0 using qA = 1.0
Reject µ = 0 using qR = 1.0
Not classified
Accept µ = 0 using qA = 3.0
Reject µ = 0 using qR = 0.5
Not classified
µ=0
666
34
0
419
9
272
µ=2
3
297
0
0
287
13
Table 4: Confusion matrices for Example 2 with k = 1000 when 700 of the µi
equal 0 and 300 equal 2.
and λ0 = 1.94 is the elicited value. From Table 2 with qR = 1.0, about 8% false
positives are expected a priori and from Table 3 with qA = 1.0, a worst case
upper bound on the a priori expected percentage of false negatives is about
75%. The top part of Table 4 indicates that with qR = qA = 1.0, then 4.9%
(34 of 700) false positives and 0.1% (3 of 300) false negatives were obtained.
With these choices of the cutoffs all hypotheses are classified. Certainly the
upper bound 75% seems far too pessimistic in light of the results, but recall
that Table 3 is computed at the false values µ = ±0.5. The relevant a priori
expected percentage of false negatives when µ = ±2.0 is about 3.5%. The
bottom part of Table 4 gives the relevant values when qR = 0.5 and qA = 3.0.
In this case there are 2.1% (9 of 428) false positives and 0% false negatives but
39.9% (272 out of 700) of the true hypotheses and 4.3% (13 out 300) of the false
hypotheses were not classified as the relevant relative belief ratio lay between qR
and qA . So in this case being more conservative has reduced the error rates with
the price being a large proportion of the true hypotheses don’t get classified.
The procedure has worked well in this example but of course the error rates can
be expected to rise when the false values move towards the null and improve
when they move away from the null.
What is implemented in an application depends on the goals. If the primary
purpose is to protect against false positives, then Table 2 indicates that this is
accomplished fairly easily. Protecting against false negatives is more difficult.
Since the actual effect sizes are not known a decision has to be made. Note
that choosing a cutoff is equivalent to saying that one will only accept H0i if
the belief in the truth of H0i has increased by a factor at least as large as qA .
Computations such as in Table 3 can be used to provide guidance but there is no
avoiding the need to be clear about what effect sizes are deemed to be important
or the need to obtain more data when this is necessary. One comforting aspect
of the methodology is that error rates are effectively controlled but there may
be many true hypotheses not classified.
The idea of controlling the prior probability of a randomly chosen hypothesis
yielding a false positive or a false negative via (8) or (9), respectively, can be
extended. For example, consider the prior probability that a random sample of
14
l from k hypotheses yields at least one false positive
X
1
at least one of RBij (ψ0ij | X) < qR
M
.
k
for j = 1, . . . , l | ψ0i1 , . . . , ψ0il
l
(10)
{i1 ,...,il }⊂{1,...,k}
In the context of the examples of this paper, and many others, the term in (10)
corresponding to {i1 , . . . , il } equals M (at least one of RBij (ψ0ij | X) < qR for
j = 1, . . . , l | ψ0 ). The following result, whose proof is given in the Appendix,
then leads to an interesting property for (10).
Lemma 4 For probability model (Ω, F, P ), the probability that at least one of
l ≤ k randomly selected events from {A1 , . . . , Ak } ⊂ F occurs is an upper
bound on the probability that at least one of l0 ≤ l randomly selected events from
{A1 , . . . , Ak } ⊂ F occur.
It then follows, by taking Ai = {x : RBi (ψ0i | x) < qR }, that (10) is an upper
bound on the prior probability that a random sample of l0 hypotheses yields at
least one false positive whenever l0 ≤ l. So (10) leads to a more rigorous control
over the possibility false positives, if so desired. A similar result is obtained for
false negatives.
4
Applications
One application of the relative belief multiple testing algorithm is to the problem
of inducing sparsity.
Example 3. Testing for sparsity.
The position taken here is that, rather than trying to induce sparsity through
an estimation procedure, an assessment of sparsity is made through an explicit
measure of the evidence as to whether or not a particular assignment is valid.
Virtues of this approach are that it is based on a measure of evidence and it is
not dependent on the form of the prior as any prior can be used. Also, it has
some additional advantages when prior-data conflict is taken into account.
Consider the context of Example 1. A natural approach to inducing sparsity
is to estimate µi by µ0 whenever RBi (µ0 | x) > qA . From the simulation it is
seen that this works extremely well when qA = 1 for both k = 10 and k = 1000.
It also works when k = 1000 and qA = 3, in the sense that the error rate is low,
but it is conservative in the amount of sparsity it induces in that case. Again
the goals of the application will dictate what is appropriate.
A common method for inducing sparsity is to use a penalized estimator
as in the LASSO introduced in Tibshirani (1996). It is noted there that the
LASSO is equivalent to MAP estimation when using a product of independent
Laplace priors. This aspect was pursued further in Park and Casella (2006)
which adopted a more formal Bayesian approach but still used MAP.
Consider
Laplace priors for the prior on µ,
√ then a product√ of independent
Pk
namely, ( 2λ0 σ)−k exp{−( 2/λ0 σ) i=1 |µi − µ0 |} where σ is assumed known
15
Decision
Accept µ = 0 using qA = 1.0
Reject µ = 0 using qA = 1.0
µ=0
227
473
µ=2
0
300
Table 5: Confusion matrices using LASSO with k = 1000 when 700 of the µi
equal 0 and 300 equal 2 in Example 3
and µ0 , λ0 are hyperparameters. Note that each Laplace prior has mean µ0
and variance λ20 σ 2 . Using the elicitation algorithm provided in Example 1, but
replacing the normal prior with a Laplace prior, leads to the assignment µ0 =
(ml + mu )/2, λ0 = (mu − ml )/2σG−1 (0.995) where G−1 (p) = 2−1/2 log 2p when
p ≤ 1/2, G−1 (p) = −2−1/2 log 2(1 − p) when p ≥ 1/2 and G−1 denotes the
quantile function of a Laplace distribution with mean 0 and variance 1. With
the specifications used in the simulations of Example 1, this leads to µ0 = 0 and
λ0 = 1.54 which implies a smaller variance than the value λ0 = 1.94 used with
the normal prior and so the Laplace prior is more concentrated about 0.
The posteriors for the µi are
√ independent with the density for µi proportional
to exp{−n(x̄i − µi )2 /2σ 2 − 2|µi − µ0 |/λ0 σ} giving the MAP estimator
√
√
x̄
<
µ
−
2σ/λ0 n√
x̄i + 2σ/λ0 n,
i
0
√
µiMAP (x) =
µ ,
µ0 − 2σ/λ0 n ≤ x̄√
2σ/λ0 n
i ≤ µ0 +
0 √
x̄i − 2σ/λ0 n,
x̄i > µ0 + 2σ/λ0 n.
The MAP estimate of
√ µi is sometimes forced to equal µ0 although this effect is
negligible whenever 2σ/λ0 n is small.
The LASSO induces sparsity through estimation by taking λ0 to be small.
By contrast the evidential approach, based on the normal prior and the relative
belief ratio, induces sparsity through taking λ0 large. The advantage to this
latter approach is that by taking λ0 large, prior-data conflict is avoided. When
taking λ0 small, the potential for prior-data conflict rises as the true values
can be deep√into the tails of the prior. For example, for the simulations of
Example 1 2σ/λ0 n = 0.183 which is smaller than the δ/2 = 0.5 used in the
relative belief approach with the normal prior. So it can be expected that the
LASSO will do worse here and this is reflected in Table 5 where there are far
too many false negatives. To improve this the value of λ0 needs to be reduced
although note that this is determined by an elicitation and there is the risk
of then encountering prior-data conflict. Another possibility is to implement
the evidential approach with the elicited Laplace prior and the discretization
as done with the normal prior and then we can expect results similar to those
obtained in Example 1.
It is also interesting to compare the MAP estimation approach and the
relative belief approach with respect to the conditional prior probabilities of µi
being assigned the value µ0 when the true value actually is µ0 . It
√is easily
√ seen
that, based on the Laplace prior, M (µiMAP (x) = µ0 | µ0 ) = 2Φ( 2/λ0 n) − 1
and this converges to 0 as n → ∞ or λ0 → ∞. For the relative belief approach
M (RBi (µ0 | x) > qA | µ0 ) is the relevant probability. With either the normal or
16
Laplace prior M (RBi (µ0 | x) > qA | µ0 ) converges to 1 both as n → ∞ and as
λ0 → ∞. In particular, with enough data the correct assignment is always made
using relative belief but not with MAP based on the Laplace prior.
While the Laplace and normal priors work equally with the relative belief
multiple testing algorithm, there don’t appear to be any advantages to using
the Laplace prior. One could argue too that the singularity of the Laplace
prior at its mode makes it an odd choice and there doesn’t seem to be a good
justification for this. Furthermore, the computations are harder with the Laplace
prior, particularly with more complex models. So using a normal prior seems
preferable overall.
An example with considerable practical significance is now considered.
Example 4. Full rank regression.
Suppose the basic model is given by y = β0 + β1 x1 + · · · + βk xk + z =
β0 + x0 β1:k + z where the xi are predictor variables, z ∼ N (0, σ 2 ) and the βi
and σ 2 are unknown. The main issue in this problem is testing H0i : βi = 0 for
i = 1, . . . , k to establish which variables have any effect on the response. The
prior distribution of (β, σ 2 ) is taken to be
β | σ 2 ∼ Nk+1 (0, σ 2 Σ0 ), 1/σ 2 ∼ gammarate (α1 , α2 ),
(11)
for some hyperparameters Σ0 and (α1 , α2 ). Note that this may entail subtracting
a known, fixed constant from each y value so that the prior for β0 is centered
at 0. Taking 0 as the central value for the priors on the remaining βi seems
appropriate when the primary concern is whether or not each xi is having any
effect. Also, it will be assumed that the observed values of the predictor variable
have been standardized so that for observations (y, X) ∈ Rn × Rn×(k+1) , where
X = (1, x1 , . . . , xk ), then 10 xi = 0 and ||xi ||2 = 1 for i = 1, . . . , k. The marginal
2
prior for βi is then {(α2 /α1 )σ0ii
}1/2 t2α1 where t2α1 denotes the t distribution
on 2α1 degrees of freedom, for i = 0, . . . , k. Hereafter, we will take Σ0 = λ20 Ik+1
although it is easy to generalize to more complicated choices.
The elicitation of the hyperparameters is carried out via an extension of
a method developed in Cao, Evans and Guttman (2014) for the multivariate
normal distribution. Suppose that it is known with virtual certainty, based on
our knowledge of the measurements being taken, that β0 + x0 β1:k will lie in
the interval (−m0 , m0 ) for some m0 > 0 for all x ∈ R where R is a compact
set centered at 0. On account of the standardization, R ⊂ [−1, 1]k . Again
‘virtual certainty’ is interpreted as probability greater than or equal to γ where
γ is some large probability like 0.99. Therefore, the prior on β must satisfy
2Φ(m0 /σλ0 {1 + x0 x}1/2 ) − 1 ≥ γ for all x ∈ R and this implies that
σ ≤ m0 /λ0 τ0 z(1+γ)/2
(12)
where τ02 = 1 + maxx∈R ||x||2 ≤ 1 + k with equality when R = [−1, 1]k .
An interval that will contain a response value y with virtual certainty,
given predictor values x, is β0 + x0 β1:k ± σz(1+γ)/2 . Suppose that we have
lower and upper bounds s1 and s2 on the half-length of this interval so that
17
s1 ≤ σz(1+γ)/2 ≤ s2 or, equivalently,
s1 /z(1+γ)/2 ≤ σ ≤ s2 /z(1+γ)/2
(13)
holds with virtual certainty. Combining (13) with (12) implies λ0 = m0 /s2 τ0 .
To obtain the relevant values of α1 and α2 let G (α1 , α2 , ·) denote the cdf of
the gammarate (α1 , α2 ) distribution and note that G (α1 , α2 , z) = G (α1 , 1, α2 z) .
Therefore, the interval for 1/σ 2 implied by (13) contains 1/σ 2 with virtual cer2
−1
tainty, when α1 , α2 satisfy G−1 (α1 , α2 , (1+γ)/2) = s−2
(α1 , α2 , (1−
1 z(1+γ)/2 , G
−2 2
γ)/2) = s2 z(1−γ)/2 , or equivalently
2
G(α1 , 1, α2 s−2
1 z(1+γ)/2 ) = (1 + γ)/2,
(14)
2
G(α1 , 1, α2 s−2
2 z(1−γ)/2 )
(15)
= (1 − γ)/2.
It is a simple matter to solve these equations for (α1 , α2 ) . For this choose an
initial value for α1 and, using (14), find z such that G(α1 , 1, z) = (1 + γ)/2,
2
which implies α2 = z/s−2
1 z(1+γ)/2 . If the left-side of (15) is less (greater) than
(1 − γ)/2, then decrease (increase) the value of α1 and repeat step 1. Continue
iterating this process until satisfactory convergence is attained.
The methods discussed in Evans and Moshonov (2006) are available for
checking the prior to see if it is contradicted by the data. Methods are specified
there for checking each of the components in the hierarchy, namely, first check
the prior on σ 2 and, if it passes, then check the prior on β. If conflict is found,
then the methods discussed in Evans and Jang (2011b) are available to modify
the prior appropriately.
Assuming that X is of rank k + 1, the posterior of (β, σ 2 ) is given by
β | y, σ 2 ∼ Nk+1 (β(X, y), σ 2 Σ(X)),
1/σ 2 | y ∼ gammarate ((n + 2α1 )/2, α2 (X, y)/2),
(16)
−1
where b = (X 0 X)−1 X 0 y, β(X, y) = Σ(X)X 0 Xb, Σ(X) = (X 0 X + Σ−1
and
0 )
2
0
0
α2 (X, y) = ||y − Xb|| + (Xb) (In − XΣ(X)X )Xb + 2α2 . Then the marginal
posterior for βi is given by βi (X, y) + {α2 (X, y)σii (X)/(n + 2α1 )}1/2 tn+2α1 and
the relative belief ratio for βi at 0 equals
− n+2α2 1 +1
1 +1
Γ (α1 )
Γ n+2α
βi2 (X, y)
2
1+
×
RBi (0 | X, y) = 2α1 +1
1
α2 (X, y)σii (X)
Γ
Γ n+2α
2
2
− 21
α2 (X, y)σii (X)
.
(17)
α22 λ20
Rather than using (17), however, the distributional results are used to compute the discretized relative belief ratios as in Example 1. For this δ > 0 is
required to determine an appropriate discretization and it will be assumed here
that this is the same for all the βi , although the procedure can be easily modified
if this is not the case in practice. Note that such a δ is effectively determined
18
by the amount that xi βi will vary from 0 for x ∈ R. Since xi ∈ [−1, 1] then
|xi βi | ≤ δ provided |βi | ≤ δ. When this variation is suitably small as to be
immaterial, then such a δ is appropriate for saying βi is effectively 0. Note that
determination of the hyperparameters and δ is dependent on the application.
Again inference can be made concerning ξ = Ξ(β, σ 2 ), the proportion of the
βi effectively equal to 0. As in Example 1, however, we can expect bias when the
amount of variability in the data is large relative to δ or the prior is too diffuse.
To implement the relative belief multiple testing algorithm the quantities (8) and
(9) need to be computed to determine qR and qA , respectively. The conditional
prior distribution of (b, ||y − Xb||2 ), given (β, σ 2 ), is b ∼ Nk+1 (β, σ 2 (X 0 X)−1 )
statistically independent of ||y − Xb||2 ∼ gamma((n − k − 1)/2, σ −2 /2). So
computing (8) and (9) can be carried out by generating (β, σ 2 ) from the relevant
conditional prior, generating (b, ||y − Xb||2 ) given (β, σ 2 ), and using (17).
To illustrate these computations the diabetes data set discussed in Efron,
Hastie, Johnstone and Tibshirani (2006) and Park and Casella (2008) is now
analyzed. With γ = 0.99, the values m0 = 100, s1 = 75, s2 = 200 were used
to determine the prior together with τ0 = 1.05 determined from the X matrix.
This lead to the values λ0 = 0.48, α1 = 7.29, α2 = 13641.35 being chosen for
the hyperparameters. Using the methods developed in Evans and Moshonov
(2006), a first check was made on the prior on σ 2 against the data and a tail
probability equal to 0.19 was obtained indicating there is no prior-data conflict
with this prior. Given no prior-data conflict at the first stage, the prior on β was
then checked and the relevant tail probability of 0.00 was obtained indicating
a strong degree of conflict. Following the argument in Evans and Jang (2011)
the value of λ0 was increased to choose a prior weakly informative with respect
to our initial choice and this lead to choosing the value λ0 = 5.00 and then the
relevant tail probability equals 0.32.
Using this prior, the relative belief estimates, ratios and strengths are recorded
in Table 6. From this it is seen that there is strong evidence against βi = 0
for the variables sex, bmi, map and ltg and no evidence against βi = 0 for
any other variables. There is strong evidence of in favor of βi = 0 for age and
ldl, moderate evidence in favor of βi = 0 for the constant, tc, tch and glu and
perhaps only weak evidence in favor of βi = 0 for hdl.
As previously discussed it is necessary to consider the issue of bias, namely,
compute the prior probability of getting a false positive for different choices of
qR and the prior probability of getting a false negative for different choices of
qA . The value of (8) is 0.0003 when qR = 1 and so there is virtually no bias in
favor of false positives and one can feel confident that the predictors identified
as having an effect do so. The story is somewhat different, however, when
considering the possibility of false negatives via (9). For example, with qA = 1,
then (9) equals 0.9996 and when qA = 100 then (9) equals 0.7998. So there
is substantial bias in favor of the null hypotheses and undoubtedly this is due
to the diffuseness of the prior. The implication is that we cannot be entirely
confident concerning those βi assigned to be equal to 0. Recall, that the first
prior proposed lead to prior-data conflict and as such a much more diffuse prior
was substituted. The bias in favor of false negatives could be mitigated by
19
Variable
Constant
age
sex
bmi
map
tc
ldl
hdl
tch
ltg
glu
Estimates
2
−4
−224
511
314
162
−20
167
114
496
77
RBi (0 | X, y)
2454.86
153.62
0.13
0.00
0.00
33.23
57.65
27.53
49.97
0.00
66.81
Strength
0.44
0.95
0.00
0.00
0.00
0.36
0.90
0.15
0.37
0.00
0.23
Table 6: Relative belief estimates, relative belief ratios and strengths for assessing no effect for the diabetes data in Example 4..
making the prior less diffuse. It is to be noted, however, that this is an exercise
that should be conducted prior to collecting the data as there is a danger that
the choice of the prior will be too heavily influenced by the observed data. The
real cure for any bias in an application is to collect more data.
Next we consider the application to regression with k + 1 > n.
Example 5. Non-full rank regression.
In a number of applications k + 1 > n and so X is of rank l < n. In this
situation, suppose {x1 , . . . , xl } forms a basis for L(x1 , . . . , xk ), perhaps after
relabeling the predictors, and write X = (1 X1 X2 ) where X1 = (x1 . . . xl ).
For given r = (X1 X2 )β1:k there will be many solutions β1:k . A particular solution is given by β1:k∗ = (X1 (X10 X1 )−1 0)0 r. The set of all solutions is then given
by β1:k∗ + ker(X1 X2 ) where ker(X1 X2 ) = {(−B 0 Ik−l )0 η : η ∈ Rk−l }, B =
(X10 X1 )−1 X10 X2 and the columns of C = (−B 0 Ik−l )0 give a basis for ker(X1 X2 ).
Given that sparsity is expected for the true β1:k , it is natural to consider the
solution which minimizes ||β1:k ||2 for β1:k ∈ β1:k∗ + L(C). Using β1:k∗ , and applying the Sherman-Morrison-Woodbury formula to C(C 0 C)−1 C 0 , this is given
by the Moore-Penrose solution
MP
β1:k
= (Ik − C(C 0 C)−1 C 0 )β1:k∗ = (Il B)0 ω1:l
(18)
0 −1
where ω1:l = (Il + BB ) (β1:l + Bβl+1:k ).
From (11) with Σ0 = λ20 Ik+1 , the conditional prior distribution of (β0 , ω1:l )
given σ 2 is β0 | σ 2 ∼ N (0, σ 2 λ20 ) independent of ω1:l | σ 2 ∼ Nl (0, σ 2 λ20 (Il +
MP
BB 0 )−1 ) which, using (18), implies β1:k
| σ 2 ∼ Nk (0, σ 2 Σ0 (B)), conditionally
independent of β0 , where
(Il + BB 0 )−1
(Il + BB 0 )−1 B
2
Σ0 (B) = λ0
.
B 0 (Il + BB 0 )−1 B 0 (Il + BB 0 )−1 B
With 1/σ 2 ∼ gammarate (α1 , α2 ), this implies that the unconditional prior of the
1/2
MP
2
i-th coordinate of β1:k
is λ20 α2 σii
(B)/α1
t2α1 .
20
Putting X∗ = (1 X1 + X2 B 0 ) gives the full rank model y | β0 , ω1:l , σ 2 ∼
0
Nn (X∗ (β0 , ω1:l
)0 , σ 2 In ). As in Example 4 then (β0 , ω1:l ) | y, σ 2 ∼ Nl (ω(X∗ , y),
2
2
σ Σ(X∗ )), 1/σ | y ∼ gammarate ((n + 2α1 )/2, α2 (X∗ , y)/2) where ω(X∗ , y) =
Σ(X∗ )X∗0 X∗ b∗ , b∗ = (X∗0 X∗ )−1 X∗0 y and
n
0
1
0
−2
Σ−1 (X∗ ) =
+
λ
,
0
0 (X1 + X2 B 0 )0 (X1 + X2 B 0 )
0 (Il + BB 0 )
α2 (X∗ , y) = ||y − X∗ b∗ ||2 + (X∗ b∗ )0 (In − X∗ Σ(X∗ )X∗0 )X∗ b∗ + 2α2 .
Now noting that (X1 + X2 B 0 )0 (X1 + X2 B 0 ) = (Il + BB 0 )X10 X1 (Il + BB 0 ), this
implies b0∗ = (ȳ, (Il + BB 0 )−1 b1 ), where b1 = (X10 X1 )−1 X10 y is the least-squares
estimate of β1:l , and
n + λ−2
0
0
0
Σ(X∗ ) =
0
(Il + BB 0 )X10 X1 (Il + BB 0 ) + λ−2
0 (Il + BB )
nȳ/(n + λ−2
0 )
ω(X∗ , y) = Σ(X∗ )X∗0 X∗ b∗ =
.
0
0
−1 −1
(Il + BB + λ−2
(X
) b1
1 X1 )
0
−1
,
−1
−1
Using (18), then β0 | y, σ 2 ∼ N (n(n + λ−2
ȳ, σ 2 (n + λ−2
) independent of
0 )
0 )
MP
2
MP
2 MP
β1:k | y, σ ∼ Nk (β
(X, y), σ Σ
(X)) where
Db1
E
EB
MP
β M P (X, y) =
,
Σ
(X)
=
B 0 Db1
B 0 E B 0 EB
0
−1 −1
with D = (Il + BB 0 + λ−2
)
and E = ((Il + BB 0 )(X10 X1 )(Il +
0 (X1 X1 )
0 −1
BB 0 ) + λ−2
(I
+
BB
))
.
The
marginal
posterior
for βiM P is then given by
l
0
MP
βiM P (X, y) + {α2 (X∗ , y)σii
(X)/(n + 2α1 )}1/2 tn+2α1 . Relative belief inferences
MP
for the coordinates of β1:k
can now be implemented just as in Example 4.
We consider a numerical example where there is considerable sparsity. For
this let X1 ∈ Rn×l be formed by taking the second through l-th columns of
the (l + 1)-dimensional Helmert matrix, repeating each row m times and then
normalizing. So n = m(l + 1) and the columns of X1 are orthonormal and
orthogonal to 1. It is supposed that the first l1 ≤ l of the variables giving rise to
the columns of X1 have βi 6= 0 whereas the last l − l1 have βi = 0 and that the
variables corresponding to the first l2 ≤ k − l columns of X2 = X1 B ∈ Rn×(k−l)
have βi 6= 0 whereas the last k − l − l2 have βi = 0. The matrix B is obtained
by generating
B1 0
B=
0 B2
i.i.d.
where B1 = (z1 /||z1 || · · · zl2 /||zl2 ||) with z1 , . . . , zl2 ∼ Nl1 (0, I) independent of
B2 = (zl2 +1 /||zl2 +1 || · · · zk−l−l2 /||zlk−l−l2 ||) with zl2 +1 , . . . , zk−l−l2 i.i.d. Nl−l (0, I).
Note that this ensures that the columns of X2 are all standardized. Furthermore, since it is assumed that the last l −l1 variables of X1 and the last k −l −l2
variables of X2 don’t have an effect, the form of B is necessarily of the diagonal
form given. For, if it was allowed that the last k − l − l2 columns of X2 were
21
k = 10
True Positive
True Negative
Total
k = 20
True Positive
True Negative
Total
k = 50
True Positive
True Negative
Total
k = 100
True Positive
True Negative
Total
Classified Positive
5
1
6
Classified Positive
7
0
7
Classified Positive
7
0
7
Classified Positive
7
0
7
Classified Negative
0
4
4
Classified Negative
0
13
13
Classified Negative
0
43
43
Classified Negative
0
93
93
Total
5
5
10
Total
7
13
20
Total
7
43
50
Total
7
93
100
Table 7: Confusion matrices for the numerical example in Example 5.
linearly dependent on the the first l1 columns of X1 , then this would induce a
dependence on the corresponding variables and this is not the intention in the
simulation. Similarly, if the first l2 columns of X2 were dependent on the last
l − l1 columns of X1 , then this would imply that the variables associated with
these columns of X1 have an effect and this is not the intention.
The sampling model is then prescribed by setting l = 10, l1 = 5, l2 = 2,
with βi = 4 for i = 1, . . . , 5, 11, 12 with the remaining βi = 0, σ 2 = 1, m = 2,
so n = 22 and we consider various values of k ≥ l. It is is to be noted that a
different data set was generated for each value of k. The prior is specified as
in Example 4 where the values λ20 = 4, α1 = 11, α2 = 12 were chosen so that
there will be no prior-data conflict arising with the generated data. Also, we
considered several values for the discretization parameter δ. A hypothesis was
classified as true if the relative belief ratio is greater than 1 and classified as
false if it is less than 1. Table 7 gives the confusion matrices with δ = 0.1. The
value δ = 0.5 was also considered but there was no change in the results.
One fact stands out immediately, namely, in all of these example only one
misclassification was made and this was in the full rank (k = 10) case where one
hypothesis which was true was classified as a positive. The effect sizes that exist
are reasonably large, and so it can’t be expected that the same performance will
arise with much smaller effect sizes, but it is clear that the approach is robust
to the number of hypotheses considered. It should also be noted, however, that
the amount of data is relatively small and the success of the procedure will only
improve as this increases. This result can, in part, be attributed to the fact that
a logically sound measure of evidence is being used.
22
5
Conclusions
An approach to the problem of multiple testing has been developed based on
the relative belief ratio. It is argued in Evans (2015) that the relative belief
ratio is a valid measure of evidence as it measures change in belief as opposed
to belief and, among the many possible candidates for such a measure, it is the
simplest with the best properties. One can expect that statistical procedures
based on valid measures of evidence will perform better than procedures that
aren’t as they possess a sounder logical basis. For the multiple testing problem
this is reflected in the increased flexibility as well as in the results.
It seems that an appropriate measure of evidence in a statistical problem
requires the specification of a prior. While this may be controversial to some, it
is to be noted that there are tools for dealing with the subjective nature of some
of the ingredients to a statistical analysis such as the sampling model and prior.
In particular, there is the process of checking for prior-data conflict after the
data is obtained and possibly modifying the prior based upon the idea of weak
informativity when such a conflict is encountered. Before the data is actually
collected, one can measure to what extent a particular prior will bias the results
based upon the particular measure of evidence used. If bias is encountered
several mitigating steps can be taken but primarily this will require increasing
the amount of data collected. These concepts play a key role in the multiple
testing problem.
Acknowledgements
Thanks to Professor Lei Sun for making her notes on multiple testing available.
Bibliography
Baskurt, Z. and Evans, M. (2013) Hypothesis assessment and inequalities for
Bayes factors and relative belief ratios. Bayesian Analysis, 8, 3, 569-590.
Cao, Y., Evans, M. and Guttman, I. (2014) Bayesian factor analysis via concentration. Current Trends in Bayesian Methodology with Applications, edited
by S. K. Upadhyay, U. Singh, D. K. Dey and A. Loganathan. CRC Press.
Carvalho, C. M., Polson, N. G. and Scott, J. G. (2009) Handling sparsity via
the horseshoe. Journal of Machine Learning Research W&CP 5: 73-80.
Efron, B., Hastie, T, Johnstone, I., and Tibshirani, R. (2006) Least angle regression. The Annals of Statistics, 32, 407-499.
Evans, M. (1997) Bayesian inference procedures derived via the concept of relative surprise. Communications in Statistics, 26, 1125-1143.
Evans , M. (2015) Measuring Statistical Evidence Using Relative Belief. Monographs on Statistics and Applied Probability 144, CRC Press.
Evans, M. and Jang, G. H. (2011a) A limit result for the prior predictive.
Statistics and Probability Letters, 81, 1034-1038.
23
Evans, M. and Jang, G. H. (2011b) Weak informativity and the information in
one prior relative to another. Statistical Science, 26, 3, 423-439.
Evans, M. and Moshonov, H. (2006) Checking for prior-data conflict. Bayesian
Analysis, 1, 4, 893-914.
George, E. I. and McCulloch, R. E. (1993). Variable selection via Gibbs sampling. Journal of the American Statistical Association, 88,881-889.
George, E. I. and McCulloch, R. E. (1997). Approaches for Bayesian variable
selection. Statistica Sinica, 7(2), 339–373.
Hastie, T., Tibshirani, R. and Martin Wainwright, M. (2015) Statistical Learning with Sparsity: The Lasso and Generalizations. Monographs on Statistics
and Applied Probability 143, CRC Press.
Park, R. and Casella, G. (2008) The Bayesian Lasso. Journal of the American
Statistical Association, 103, 681-686.
Rockova, V. and George, E. I. EMVS: The EM approach to Bayesian variable
selection. Journal of the American Statistical Association, 109, 506, 828 - 846.
Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal
of the Royal Statistical Society B., 58, 1, 267-288.
Appendix
Proof of Lemma 4: Let ∆(i) be the event that exactly i of A1 , . . . , Ak ∈ F
occur, so that ∪ki=1 Ai = ∪ki=1 ∆(i) and note that the ∆(i) are mutually disjoint.
When l < k,
Sl,k =
X
IAi1 ∪···∪Ail
{i1 ,...,il }⊂{1,...,k}
X
l−1
k−1
X k i
k
=
I∆(k−i) +
−
I∆(k−i)
l i=0
l
l
i=l
k−1
k−1
X i
k X
I∆(k−i)
I∆(k−i) −
=
l
l i=0
i=l
and Sk,k = IA1 ∪···∪Ak . Now consider
1
k
X
l {i1 ,...,il }⊂{1,...,k}
IAi1 ∪···∪Ail −
k −1
Sl,k
l
1
−
k −1
Sl,k
l−1
X
which equals
IAi1 ∪···∪Ail−1
k
l−1 {i1 ,...,il−1 }⊂{1,...,k}
(19)
Pk−1
If l = k, then (19) equals IA1 ∪···∪Ak − i=0 I∆(k−i) + I∆(1) = IA1 ∪···∪Ak −
Pk−2
k −1
I∆(k−l+1) +
i=0 I∆(k−i) which is nonnegative. If l < k, then (19) equals l−1
Pk−1 i k −1
i k −1
− l l
]I∆(k−i) which is nonnegative since an easy calcui=l [ l−1 l−1
lation gives that each term in the second sum is nonnegative. The expectation
of (19) is then nonnegative and this establishes the result.
24
| 10math.ST
|
1
Message passing-based joint CFO and
channel estimation in millimeter wave systems
with one-bit ADCs
arXiv:1803.09012v1 [cs.IT] 23 Mar 2018
Nitin Jonathan Myers, Student Member, IEEE, and
Robert W. Heath Jr., Fellow, IEEE.
Abstract
Channel estimation at millimeter wave (mmWave) is challenging when large antenna arrays are
used. Prior work has leveraged the sparse nature of mmWave channels via compressed sensing based
algorithms for channel estimation. Most of these algorithms, though, assume perfect synchronization and
are vulnerable to phase errors that arise due to carrier frequency offset (CFO) and phase noise. Recently
sparsity-aware, non-coherent beamforming algorithms that are robust to phase errors were proposed
for narrowband phased array systems with full resolution analog-to-digital converters (ADCs). Such
energy based algorithms, however, are not robust to heavy quantization at the receiver. In this paper,
we develop a joint CFO and wideband channel estimation algorithm that is scalable across different
mmWave architectures. Our method exploits the sparsity of mmWave MIMO channel in the angle-delay
domain, in addition to compressibility of the phase error vector. We formulate the joint estimation as
a sparse bilinear optimization problem and then use message passing for recovery. We also give an
efficient implementation of a generalized bilinear message passing algorithm for the joint estimation in
mmWave systems with one-bit ADCs. Simulation results show that our method is able to recover the
CFO and the channel compressively, even in the presence of phase noise.
Index Terms
Millimeter wave communication, wideband channel estimation, synchronization, compressed sensing, message passing, one-bit receivers
N. J. Myers ([email protected]) and R. W. Heath Jr. ([email protected]) are with the Wireless Networking and
Communications Group, The University of Texas at Austin, Austin, TX 78712 USA. This material is based upon work supported
in part by the National Science Foundation under Grant No. NSF-CCF-1527079, and by a gift from Huawei Technologies, Inc.
2
I. I NTRODUCTION
Millimeter wave communication introduces new challenges in the design of MIMO communication systems [1]. For instance, large antenna arrays at the transmitter (TX) and the receiver
(RX) are necessary to meet the link budget requirements [2]. As a result, the channel has a
higher dimension compared to what is typical in lower frequency MIMO systems, and must be
estimated more frequently thanks to the smaller coherence time [3]. Furthermore, cost and power
consumption are major issues at the larger bandwidths that accompany mmWave, primarily due
to high resolution ADCs [4]. Typical mmWave hardwares that limit power consumption at large
bandwidths introduce compression in the channel measurements. For example, the one-bit ADC
architecture [4] allows access to the output of every antenna at the expense of heavy quantization.
The compression of channel measurements and the use of large antenna arrays complicate signal
processing at mmWave.
Compressed sensing (CS) [5] [6] is an efficient technique to recover sparse high-dimensional
signals with few projections. As MIMO channel matrices at mmWave are sufficiently sparse when
expressed in an appropriate dictionary, applying tools from CS to mmWave channel estimation
can potentially reduce the training overhead. CS-based sparse channel estimation algorithms have
been proposed for various hardware architectures [7]–[9]. Recent developments in approximate
message passing [10] [11] have enabled channel estimation algorithms in low resolution receivers
[12]. Most CS-based channel estimation algorithms, however, assume perfect synchronization and
fail in practice because of the CFO and phase noise [13].
CFO and phase noise are hardware impairments that corrupt the phase of the channel measurements. The mismatch between the carrier frequencies of the local oscillators at the TX and
the RX results in CFO. Phase noise in the system arises due to short-term random fluctuations
in the frequency of the oscillators. Both these non-idealities are larger at mmWave due to the
high carrier frequency and ignoring them can result in significant channel estimation error [14].
Correcting for the CFO and then performing channel estimation seems like a possible solution.
The disadvantage, however, is that prior to beamforming or channel estimation, mmWave systems
operate at very low SNR, which can result in significant error in the CFO estimate. Prior work
has considered joint CFO and channel estimation [15] [16] in lower frequency systems. These
joint estimation algorithms, however, cannot be applied to typical mmWave systems due to
differences in the hardware architectures. Furthermore, they are not designed to incorporate the
3
sparse nature of mmWave channels. Therefore, there is a need to design either phase error robust
channel estimation algorithms, or joint CFO and channel estimation algorithms that can exploit
the sparsity of mmWave channel.
Recent work on phase error robust channel estimation is limited to narrowband systems and has
focussed on specific mmWave hardware. In [13], [17] and [18], phase error robust compressive
beamforming algorithms were proposed for the analog beamforming architecture. In [13], phase
tracking followed by phase error compensated compressive beamforming was proposed. The
compensation, however, was done prior beamforming and therefore suffers from low SNR. The
non-coherent algorithms in [17] and [18] are not robust to heavy quantization at the receiver.
These algorithms cannot be used in one-bit receivers because the energy information of the
channel measurements is completely lost due to one-bit quantization.
Recent joint CFO and sparse narrowband channel estimation algorithms in [14] and [19]
require high computational complexity when extended to wideband systems. In [14], we proposed
a joint CFO and narrowband channel estimation algorithm using third-order tensors [20]. We
also developed a sparsity-aware joint estimation algorithm for the one-bit ADC architecture in
[19]. The main idea underlying our approach in [19] was to use the lifting technique [21] along
with message passing for the joint estimation with one-bit channel measurements. Extending
the narrowband solutions in [14] or [19] to typical wideband mmWave systems would require
convex optimization over millions of variables, which may be prohibitive in a practical setting.
In this paper, we propose a sparse bilinear formulation of the joint CFO and wideband channel
estimation problem, and solve it using message passing. We assume that all the RF chains at
the TX or the RX are driven by the same reference oscillator. Hence, there is a unique CFO
and a phase noise process in our MIMO system model, that corrupt the channel measurements.
We also assume that there is perfect frame timing synchronization between the TX and the RX.
We summarize the main contributions of our work as follows.
•
We formulate the joint CFO and wideband channel estimation problem as a noisy quantized
sparse bilinear optimization problem. Our framework leverages the sparse nature of the wideband channel in the angle and delay domains, and also exploits the compressibility of the
phase error vector in the frequency domain.
•
To solve the non-convex problem at hand, we use the vector variance version of the Parametric
Bilinear Generalized Approximate Message Passing (PBiGAMP) algorithm [22] and optimize
4
it for fast joint estimation. The parameters of the sparse priors corresponding to the wideband
channel and the phase error vector are learned using an Expectation Maximization (EM)
algorithm.
•
We provide insights into the design of training matrices for joint CFO and channel estimation
using PBiGAMP. Specifically, we show that shifted Zadoff-Chu training proposed in [12] to
accelerate message passing cannot be used for joint estimation as it results in a continuum
of optimal solutions for the bilinear optimization problem. We explain the “CFO propagation
effect” to highlight the trade-off between fast message passing and identifiability in the sparse
bilinear problem.
•
We evaluate the performance of our joint estimation algorithm assuming a digital receiver
architecture with one-bit ADCs and compare it with the hypothetical full resolution case.
Simulation results show that the proposed approach is able to recover both the channel and
the CFO compressively with IID Gaussian and IID QPSK training matrices, even in the
presence of phase noise uncertainity.
Our algorithm is advantageous over the existing sparsity-aware methods for joint estimation or
phase error robust channel estimation in terms of the capability to efficiently handle frequency
selective channels and scalability to other mmWave architectures.
Notation: A is a matrix, a is a column vector and a, A denote scalars. Using this notation
AT , A and A∗ represent the transpose, conjugate and conjugate transpose of A. The matrices
|A| and |A|2 contain the element-wise magnitude and squared magnitude of the entries of A.
We use A(i) and A(j) to denote the ith row and j th column of A. We use diag (a) to denote
a diagonal matrix with entries of a on its diagonal. The scalar am denotes the mth element
of a. The symbol ⊗ is used to denote the kronecker product. vec (A) is a vector obtained by
stacking all the columns of A and vecm (A) denotes the mth element of vec (A). We define
Ai,j = veci A(j) . We use IN to denote the set {1, 2, 3, ..N }. The matrix UN ∈ CN ×N denotes
the unitary discrete Fourier transform matrix. N (m, R) is the probability density function of
complex Gaussian random vector with mean m and covariance R. We define e`,N ∈ RN ×1 as
the N dimensional canonical basis vector with its `th coordinate as 1.
5
II. S YSTEM AND C HANNEL M ODELS
In this section, we describe the underlying hardware architecture, CFO and phase noise model,
and the wideband mmWave channel model used for our simulations. In particular, we focus on the
digital receiver architecture with one-bit ADCs, to highlight the differences with the existing noncoherent algorithms. Nevertheless, our algorithm can be extended to other mmWave architectures
as the underlying joint estimation problem is bilinear in nature.
A. System Model
RF
Re()
Im()
Transmitter
Ntx
RF
Block
H
ADC
Digital
Baseband
Processing
Nrx
RF
Re()
Im()
f1
q bit
q bit
ADC
f2
Fig. 1. A MIMO system with local oscillators operating at f1 and f2 , and q-bit ADCs at the receiver. Each antenna is associated
with an RF chain and a pair of q-bit ADCs. In this work, we consider the extreme cases of q = 1 and q = ∞.
We consider a MIMO system with uniform linear array of Ntx antennas at the TX and Nrx
antennas at the RX, as shown in Fig. 1. We use linear arrays for a concise representation of the
simplifications involved in PBiGAMP; our framework can be extended to other array geometries
using appropriate array response vectors in the formulation. We do not impose constraints on the
number of RF chains or the resolution of the digital-to-analog converters (DACs) at the TX. The
resolution of the Nrx ADCs at the RX, however, is assumed to be limited. The baseband signal
at the TX is upconverted to the mmWave band, using a local oscillator at a carrier frequency
f1 . The transmitted RF signal propagates through the wireless channel and is downconverted at
the RX using a carrier frequency f2 , that slightly differs from f1 . Although the MIMO system
can have multiple RF chains, we assume that all the RF chains at a given end are driven by
the same reference oscillator. Even if the RF chains at a given end were driven by different
6
oscillators, achieving carrier synchronization locally is feasible [23]. After downconversion at
the RX, the output at each antenna is sampled using a pair of q bit ADCs, one each for the
in-phase and the quadrature phase components. We use Qq (.) to represent the q bit quantization
function corresponding to the ADCs. In this work, we consider the extreme case of q = 1 and
provide a performance comparison relative to q = ∞. The quantization functions for the two
cases are Q1 (x) = sign (real (x)) + j sign (imag (x)) and Q∞ (x) = x. Note that the functions
sign (·) , real (·) and imag (·) are applied element-wise on the vector.
The impact of CFO on channel estimation algorithms is more significant in one-bit receivers
than the full resolution ones. The mismatch in the carrier frequencies, i.e., |f2 − f1 | is typically
in the order of several parts per millions (ppms) of f1 or f2 . Due to the high carrier frequencies
at mmWave, even such small differences can significantly perturb the channel estimate when
ignored [13]. For a symbol duration of T seconds, we define the digital domain CFO as =
2π (f1 − f2 ) T . CFO results in unknown phase errors in the received samples that linearly increase
with time. Hence, the impact of CFO on standard channel estimation algorithms is determined
by the length of training. As one-bit receivers relatively need a longer training for channel
estimation when compared to the full resolution ones, channel estimation algorithms that ignore
phase errors are more vulnerable to the CFO in one-bit systems than the full resolution ones.
Phase noise in wireless systems arises due to jitter in the frequency of the oscillators. For
2
2
a phase noise variance of βtx
at the TX and βrx
at the RX, the phase noise variance in the
2
2
received samples can be approximated as β 2 = βrx
+ βtx
. The approximation is valid when
the 3 dB bandwidth of the phase noise power spectral density is significantly smaller than the
channel coherence bandwidth [24]. Let φk denote the phase error introduced in the k th received
sample, due to phase noise at the TX and the RX. As is common in prior work, we model
the phase errors using a Wiener process [25] in which the increments, i.e., φk − φk−1 , are IID
Gaussian random variables with zero mean and variance β 2 . As β 2 is proportional to f12 [13],
phase noise is higher at mmWave carrier frequencies for a given quality of oscillator.
Now, we describe the received signal model in the digital receiver architecture. Let t [n] ∈
CNtx ×1 be the nth transmit symbol satisfying the power constraint E [t∗ [n] t [n]] = P . The
discrete time baseband representation of the MIMO channel is assumed to be limited to L
taps. Let H [`] ∈ CNrx ×Ntx be the `th tap of the equivalent baseband channel, where ` ∈
{0, 1, 2, ..., L − 1}. We assume that perfect frame timing synchronization can be achieved using
the control channel. Our assumption can be justified in situations where the mmWave system
7
co-exists with a lower frequency system [26] that can perform timing synchronization. Frequency
synchronization, however, may not be achieved due to different offsets and phase noise processes
for each of these systems. With the timing synchronization assumption, the sampled baseband
vector in the nth symbol duration can be given by
Y(n) = Qq
ej(n+φn )
L−1
X
!
H [`] t [n − `] + V(n) ,
(1)
`=0
2
where V(n) ∼ N (0, σ INrx ) is additive white Gaussian noise. In this work, we develop an
algorithm to estimate and {H [`]}L−1
`=0 from the series of observations Y(n) . Our joint estimation
algorithm can be extended to any q-bit ADC architecture by defining appropriate output likelihood
functions in message passing.
B. Channel Model
We consider a clustered channel model for the frequency selective mmWave MIMO channel.
The channel consists of Ncs clusters with Mn rays in the nth cluster. Let γn,m , τn,m , θr,n,m and
θt,n,m denote the complex gain, delay, angle-of-arrival (AoA) and angle-of-departure (AoD) of
the mth ray in the nth cluster. We assume that the transmitted signal is bandlimited to 1/T Hz.
With ωr,n,m = π sin θr,n,m , ωt,n,m = π sin θt,n,m and the Vandermonde vector
T
aN (∆) = 1 , ej∆ , ej2∆ , · · · , ej(N −1)∆ ,
(2)
the `th tap of the wideband MIMO channel for a half wavelength spaced uniform linear array
is given by
Ncs X
Mn
X
τn,m
.
H [`] =
γn,m aNrx (ωr,n,m ) aNtx (ωt,n,m ) sinc ` −
T
n=1 m=1
∗
(3)
The wideband channel can be represented using Nrx Ntx L complex entries, and the matrix in (3)
is large in typical mmWave systems. The channel impulse response in (3) is represented using a
linear combination of bandlimited sinc(·) functions. Notice that each of these sinc(·) functions
is delayed by the normalized delay spread, i.e, τn,m /T and evaluated at periodic time instants
to obtain the discrete time representation in (3). Other filtering functions could also be used to
incorporate the effect of pulse shaping at the TX or filtering at the RX [27].
The mmWave MIMO channel is aproximately sparse in an appropriate dictionary due to the
propagation characteristics of the environment at mmWave frequencies. Compared to the lower
frequency channels, mmWave channels comprise of fewer clusters [4]. Each of the channel taps
8
H [`], is approximately sparse in the spatial Fourier basis at mmWave [12]. Furthermore, the
channel is approximately sparse along the time dimension as the delays of the propagation rays
are heavily clustered within the delay spread. As the delays τn,m may not necessarily be an
integer multiple of T , there is a leakage effect along the time dimension. Let C [`] ∈ CNrx ×Ntx
be the 2-D Fourier transform of H [`], such that
H [`] = UNrx C [`] U∗Ntx ,
∀` ∈ {0, 1, 2, ..., L − 1} .
(4)
The approximate sparsity of the mmWave MIMO channel along the angle and delay domains
[28] is directly reflected in the matrices {C [`]}L−1
`=0 . A higher resolution dictionary can be used
for the angle and delay dimensions to increase sparsity of the mmWave channel at the expense
of higher dimensionality and higher frame coherence [29]. Our GAMP based approach will be
robust to leakage effects that arise due to approximate sparsity.
III. D EMYSTIFYING JOINT CFO
AND CHANNEL ESTIMATION AT MM WAVE
In this section, we propose a sparse bilinear formulation for the joint estimation problem. We
also identify existing techniques to solve the problem and describe their limitations in terms of
scalability to other mmWave architectures and computational complexity.
A. Bilinear formulation
We derive a compact form for the received signal model in (1) for a SC-FDE system [3]. Let
T ∈ CNtx ×Np be a training block of length Np such that T(k) = t [k]. The TX transmits a training
sequence with a cyclic prefix of length L − 1, i.e., T(Np −L+2) , T(Np −L+3) , ..., T(Np ) , T . The
cyclic prefix padded transmission gets convolved with the frequency selective MIMO channel
before sampling at the receiver. As usual, the first L − 1 samples of the received block that
experience interference from the previous transmit block are discarded [3]. Let Y ∈ CNrx ×Np
be the received block obtained after discarding the first L − 1 received vectors. We define
an ` circulant delay matrix J` ∈ CNp ×Np , such that its first column is the canonical basis
vector e1+`,Np . We define a vector d (, β) ∈ CNp ×1 such that its nth entry has the phase error
corresponding to the received vector Y(n) , i.e., dn (, β) = ej(n+φn ) . Note that β is the standard
deviation of the incremental phase errors in the Wiener phase noise process. The received samples
corresponding to the transmit block defined by T can be expressed using (1) as
!
L−1
X
Y = Qq
H [`] TJ` diag (d (, β)) + V .
`=0
(5)
9
The phase errors in (5) are invariant along any column of the unquantized received block as
there is a unique CFO and a phase noise process in the system. The representation in (5) can
be further simplified to capture the structure in the phase errors and the channel.
We exploit the structure in the joint estimation problem using sparsity of the channel and the
phase error vector in appropriate dictionaries. A compact representation of (5) can be obtained
by following the same steps in [12], except for the diagonal matrix containing the phase errors.
We use b to denote the DFT of the phase error vector d (, β). With Z used to denote the
noiseless unquantized version of Y in (5) such that Y = Qq (Z + V), we have
Z=
L−1
X
H [`] TJ` diag (d (, β))
`=0
=
L−1
X
UNrx C [`] U∗Ntx TJ`
diag
U∗Np b
`=0
U∗Ntx TJ0
U∗ TJ
1
Ntx
= UNrx [C [0] C [1] C [2] . . . C [L − 1]]
.
..
{z
}
|
∆
=C
U∗Ntx TJL−1
|
{z
∆
diag U∗Np b .
(6)
}
=F
Nrx ×Ntx L
The matrix C ∈ C
is just a concatenation of the angle domain representation in (4)
corresponding to each tap of the MIMO channel and is approximately sparse. Furthermore, the
vector b, i.e., the DFT of d (, β) can be considered to be approximately sparse. The sparse
representation of the phase error vector is valid in practice as the spread of the oscillator’s
spectrum about the center frequency is relatively small compared to the bandwidth of the signal.
Now, we derive a sparse bilinear formulation for the joint estimation problem. Using (6), the
quantized received block Y in (5) can be expressed as
Y = Qq UNrx CFdiag U∗Np b + V .
(7)
We define the Np Nrx ×1 vectors y = vec (Y), z = vec (Z) and v = vec (V), and the Nrx Ntx L×1
vector c = vec (C). In vector notation, (6) can be written as
z = diag U∗Np b ⊗ aNrx (0) vec (UNrx CF) .
(8)
10
Notice that aNrx (0) is just the all ones vector in Nrx dimension. Using the property vec (PQR) =
RT ⊗ P vec (Q), the received vector y is expressed as
(9)
y = Qq diag U∗Np b ⊗ aNrx (0) FT ⊗ UNrx c + v .
We define the matrices G = U∗Np ⊗ aNrx (0) and A = (F ⊗ UNrx )T to rewrite (9) as
y = Qq (diag (Gb) Ac + v) .
(10)
Estimating the phase errors and the channel is equivalent to estimating b and c from y in (10).
The joint estimation problem in (10) can be observed to be a noisy quantized bilinear problem
in b and c, subject to the sparsity of b and c.
B. Limitations of existing techniques
1) CFO robust methods: Existing sparse channel estimation methods that are robust to CFO
discard the phase of the channel measurements. A hashing technique based non-coherent beam
alignment algorithm was proposed in [17] for analog beamforming systems. This method, however, assumes fine control over the phase shifters, which is not necessarily the case with mmWave
systems. The received signal strength (RSS) based method in [18] accounts for the limited phase
control and uses pseudo-random phase shifts for compressive beam-training. The solutions in
[17] and [18] assume a narrowband mmWave system and perform beam-alignment with just the
magnitude of the channel measurements. These methods, however, cannot be used in one-bit
receivers as energy detection with undithered one-bit ADCs is not feasible unless additional
circuit components are used. For example, Q1 (r) and Q1 (α r) are the same for any α > 0.
As the only information provided by undithered one-bit ADCs is phase quantized to 4 levels,
discarding it due to phase errors leaves no information.
2) Joint estimation using lifting: Lifting [21] [30] is a convex relaxation technique that
transforms a bilinear problem to a higher dimensional one and then recovers the original vectors
by solving the higher dimensional problem. We describe the lifting technique applied to the joint
estimation problem in (10). With M denoting the number of entries in y or z, i.e., M = Nrx Np ,
the mth entry of z can be given as
zm = vecm (diag (Gb) Ac)
= G(m) bA(m) c
(11)
(12)
11
T
= G(m) bcT A(m)
(13)
∀m ∈ IM .
(14)
= A(m) ⊗ G(m) vec bcT ,
We define a lifted variable x = vec bcT and a measurement matrix Φ ∈ CM ×Np Nrx Ntx L , such
that Φ(m) = A(m) ⊗ G(m) . Hence, the quantized measurements in (10) can be expressed as
y = Qq (Φx + n) .
(15)
The lifted vector x in (15) is sparse as it is just an outer product [20] of the sparse vectors b
and c. Several CS-based algorithms [6] [31] can be used to recover x from the possibly underdetermined noisy quantized system in (15). Using the SVD of the higher dimensional matrix
estimate, the vectors in the bilinear problem can be estimated upto a scale factor.
Lifting followed by the SVD was applied to joint CFO and narrowband channel estimation
for one-bit receivers in our previous work [19]. The main issue in extending our method in [19]
to wideband systems arises due to the large dimensionality of the lifted problem. For instance,
the dimension of x to perform joint CFO and channel estimation in wideband systems would
be Nrx Ntx Np L. Using lifting necessarily implies solving for millions of variables for typical
wideband mmWave systems, due to the large number of antennas and the need for additional
pilots to compensate for the heavy quantization in low resolution systems.
The limitations of the existing phase error robust and joint estimation solutions in terms
of architectural scalability and computational complexity, motivate the need to develop new
low complexity joint estimation algorithms that can be applied to wideband systems and low
resolution receivers.
IV. M ESSAGE PASSING BASED JOINT CFO
AND CHANNEL ESTIMATION
In this section, we give a brief introduction to PBiGAMP [22] and discuss its application
to the joint estimation in (10). We exploit the inherent structure in our problem to derive a
low complexity and memory efficient implementation of PBiGAMP for the joint estimation.
Furthermore, we explain the CFO propagation effect induced by special training matrices that
prevents further reduction in the computational complexity.
A. Introduction to PBiGAMP
The joint estimation problem in (10) can be solved using PBiGAMP [22] by considering b, c,
z and y of (8) and (10) as realizations of random vectors, say b, c, z and y. Let bi , ck , zm and
12
ym be the elements of these random vectors. In general, deriving the closed form Minimum
Mean-Squared Error (MMSE) estimates [32] of b and c is difficult as it requires marginalizing
the joint PDF of b and c conditioned on y = y. Using ideas from message passing, PBiGAMP
can obtain the MMSE estimates of both the vectors in the bilinear problem.
We explain message passing using the factor graph [33] in Fig. 2, that shows the dependency
between y, the random variables (b, c) and their prior distributions. The circular nodes in the
factor graph are called as variable nodes as they represent the random variables. The rectangular
nodes in the factor graph are called as factor nodes and they contain the prior distribution
of a random variable or the likelihood function associated with an observation. The messages
in message passing are essentially probability distributions, also called as beliefs. The idea
underlying message passing is to perform belief flows iteratively between the factors and the
variables until all the variable nodes reach a consensus on their marginal probability distributions.
As the factor graph for joint estimation is strongly connected, standard message passing can be
computationally intractable.
Using ideas from Approximate Message Passing [10], PBiGAMP simplifies the messages by
assuming a large number of variable nodes. Simulation results in [22] that show that PBiGAMP
outperforms lifting techniques in Section III-B2, for IID Gaussian measurement matrices, motivate applying it to our problem. Furthermore, PBiGAMP performs optimization over the same
number of variables in the problem, unlike lifting [21] that solves the problem in a higher
dimensional space. For the joint estimation problem in typical wideband mmWave systems,
PBiGAMP is memory efficient over lifting by several orders of magnitude.
B. PBiGAMP for joint estimation
In this section, we explicitly state PBiGAMP [22] for joint estimation in (10) and describe
the information contained in the factor nodes of Fig. 2. To be consistent with the notation used
in [22], we rewrite the random variable dependency corresponding to (11) in the tensor notation
as
zm =
Nb X
Nc
X
(i,k)
zm
bi ck ,
(16)
i=1 k=1
(i,k)
where Nb = Np , Nc = Nrx Ntx L, and zm
is an element of a third order tensor given by
(i,k)
zm
= Gm,i Am,k .
(17)
13
Messages
Likelihood
functions
Messages
Channel
variables
Priors
Priors
Synchronization
variables
Fig. 2. The factor graph of bilinear message passing for joint CFO and channel estimation. The rectangular nodes, called as
factors, contain the likelihood functions corresponding to the received samples or the sparse priors. Messages are sent between
the factor nodes and the variable nodes until the marginal probability distributions of the variables converge.
The output likelihood function in Fig. 2, denoted by pym |zm (ym | z) is given by
√
√
2sign(Re{ym })Re{z}
2sign(Im{ym })Im{z}
F
F
q=1
σ
σ
pym |zm (ym | z) =
,
ky −zk2
1 2 e− mσ2
q=∞
πσ
(18)
where F (·) is the cumulative distribution function of the standard normal distribution. The
sparsity of the vectors b and c is incorporated by assuming parametrized Bernoulli-Gaussian
distributions for their priors pb (b) and pc (c). For simplicity, it is assumed that each entry of
b is independent of the other and identically distributed as pb . Similarly, the entries of c are
assumed to be IID, with pc as the distribution. Furthermore, the vectors b and c are assumed to
be independent of each other. Let λb and λc denote the sparsity fraction of b and c. Let σb2 and
σc2 be the variances of the coefficients corresponding to the non-zero support of the vectors b
and c. With δ (x) used to represent the Dirac-delta function, the Bernoulli-Gaussian distributions
pb and pc can be given as
pb (x) = λb δ (x) + (1 − λb ) N (0, σb2 ),
(19)
pc (x) = λc δ (x) + (1 − λc ) N (0, σc2 ).
(20)
14
Definitions:
pzm |pm z | pb; ν p ,
pck |rk(c | rb; ν r ) ,
(b | qb; ν q ) ,
pbi |qi
pym |zm(ym | z) N (z;b
p,ν p )
pym |zm(ym | z 0 ) N (z 0 ;b
p,ν p )
r ,ν r )
R pc(c) N (c;b
0
0 r ,ν r )
c0 pc(c ) N (c ;b
q ,ν q )
R pb(b) N (b;b
0
0 q ,ν q )
b0 pb(b ) N (b ;b
(D1)
R
z0
(D2)
(D3)
Initializations:
∀m : sbm (0) = 0
∀i, k : choose b
bi (1), ν b (1), b
ck (1), ν c (1)
i
(I1)
(I2)
k
for t = 1, . . . Tmax
(i,∗)
∀m, i : zbm
(i,k)
b
ck (t)
k=1 zm
P Nb b
(i,k)
= i=1 bi (t)zm
P Nb b
(i,∗)
bi (t)b
zm (t)
= i=1
(t) =
∀m, k :
(∗,k)
zbm (t)
∀m :
(∗,∗)
zbm (t)
∀m : ν pm (t) =
P Nc
(R1)
(R2)
or
(∗,k)
ck (t)b
zm (t)
k=1 b
P Nc c
(∗,k)
zm (t)|2
k=1 νk (t)|b
P Nc
(i,∗)
P Nb
b
zm (t)|2 +
i=1 νi (t)|b
P Nb b P Nc c
(i,k)
p
p
∀m : νm (t) = ν m (t) + i=1 νi (t) k=1 νk (t)|zm |2
∀m :
(∗,∗)
pbm (t) = zbm (t)
∀m :
z (t) = var{z
νm
m
−
sbm (t−1)ν pm (t)
(R4)
(R5)
(R6)
p
| pm = pbm (t); νm
(t)}
(R7)
p
∀m : zbm (t) = E{zm | pm = pbm (t); νm
(t)}
(R8)
s (t) = (1 − ν z (t)/ν p (t))/ν p (t)
∀m : νm
m
m
m
(R9)
p
pbm (t))/νm
(t)
∀m : sbm (t) = (b
zm (t) −
P
−1
(∗,k)
M
s
r
zm (t)|2
∀k : νk (t) =
m=1 νm (t)|b
P
(∗,k)
∀k : rbk (t) = b
ck (t) + νkr (t) M
bm (t)b
zm (t)∗
m=1 s
P
P Nb b
(i,k) 2
M
s (t)
− νkr (t)b
ck (t) m=1 νm
|
i=1 νi (t)|zm
P
−1
(i,∗)
q
M
s
∀i : νi (t) =
zm (t)|2
m=1 νm (t)|b
P
(i,∗)
q
M
∀i : qbi (t) = b
bi (t) + νi (t) m=1 sbm (t)b
zm (t)∗
P
P
(i,k)
− ν q (t)b
bi (t) M ν s (t) Nc ν c (t)|zm |2
m=1
i
m
k=1
k
∀k : νkc (t+1) = var{ck | rk = rbk (t); νkr (t)}
if
(R3)
∀k : b
ck (t+1) = E{ck | rk = rbk (t); νkr (t)}
∀i : νib (t+1) = var{bi | qi = qbi (t); νiq (t)}
∀i : b
bi (t+1) = E{bi | qi = qbi (t); νiq (t)}
PM
P
(∗,∗)
(∗,∗)
(∗,∗)
zm (t) − zbm (t−1)|2 ≤ τstop M
zm (t)|2 ,
m=1 |b
m=1 |b
(R10)
(R11)
(R12)
(R13)
(R14)
(R15)
(R16)
(R17)
(R18)
stop
(R19)
end
b=b
b (t) , b
Output: b
c=b
c (t) .
TABLE I
T HE PB I GAMP A LGORITHM FROM [22]
The parameters governing pb and pc , however, are not known apriori and can be learned by
embedding PBiGAMP within the Expectation Maximization (EM) algorithm [11]. For a given
set of likelihood functions and prior distributions, the vector variance PBiGAMP algorithm [22]
to obtain the MMSE estimates of b and c in (10) is summarized in Table I.
The channel and the phase error vector can be derived using appropriate transformations
over the PBiGAMP estimates. The vector b
c obtained from PBiGAMP is just an estimate of
15
the vectorized version of C, the angle-delay domain representation of the wideband channel in
b ∈ CNrx ×Ntx L denote the angle-delay domain estimate of the wideband channel, such
(7). Let C
b =b
that vec C
c. It can be seen from (6) that C is just a concatenation of the angle domain
representation of all the L taps of the MIMO channel. Therefore, the `th tap of the antenna
domain MIMO channel can be derived as
h
i
b
b
b
b
H [`] = UNrx C(`Ntx +1) , C(`Ntx +2) , . . ., C(`Ntx +Ntx ) U∗ Ntx .
(21)
b derived from PBiGAMP is an estimate of the DFT of the phase error vector
The vector b
b is just a single tone frequency estimation problem. As the
d (, β). Estimating the CFO from b
phase noise is modelled as a Wiener process, we apply the Extended Kalman Filter (EKF) [34]
b to estimate .
over the time domain samples, i.e., U∗ b,
Np
In this work, we exploit the compressibility of the phase error vector in the DFT basis for
joint estimation. It is possible, however, to incorporate the statistics of the phase noise by
replacing the nodes corresponding to b in Fig. 2 with the phase error variables and adding
factors corresponding to the phase noise process. In such case, the message passing algorithm
must handle the non-linear dependence of z on the phase errors.
C. PBiGAMP: From theory to practice
As seen from Table I, the generic implementation of PBiGAMP is memory and computationally intensive, as it involves repeated operations over a third order tensor in (17). In this
section, we provide insights into the key equations of Table I and describe our low complexity
implementation for joint estimation.
The equations in PBiGAMP are essentially determined by the belief flows from the factor
nodes to the variable nodes, and vice versa. Without loss of generality, we describe the message
flow between the variable node b1 and the factor node y1 in Fig. 2. In standard message passing,
the message sent from a factor node to b1 essentially represents the PDF of b1 presumed by
that factor node. In a fully connected factor graph, the node b1 receives messages from all the
M factor nodes ({yi }M
i=1 ) in addition to its prior distribution pb . As b1 receives several beliefs
from different factors, the belief sent by b1 to y1 is just a normalized product of all the beliefs
received by b1 except the one from y1 . For generic priors and likelihood functions in the factor
graph, performing standard message passing can be difficult as the belief flows are flows of
PDFs that are functions.
16
PBiGAMP simplifies standard message passing using the central limit theorem (CLT) and the
Taylor series approximation. Notice that y1 receives beliefs from all its neighbouring variable
nodes and contains the likelihood function in itself. The belief sent from y1 to b1 is computed
by multiplying the likelihood function, with all the incoming beliefs to y1 except the one from
b1 , and then integrating over all the random variables except b1 . The multiplication followed
by integration essentially yields the PDF of b1 presumed by the factor node y1 . The integration
is often multidimensional and can be difficult to compute. If y1 depends on a large number
of independent variable nodes through a linear function, then the linear combination can be
approximated as a Gaussian random variable using the CLT [10]. In such case, the variable nodes
can send just the mean and variances of the PDFs to the factors and this information is sufficient
to compute the mean and variance of the Gaussian random variable, as seen in (R3) and (R5)
of Table I. The message sent from y1 to b1 can now be computed by multiplying the likelihood
at y1 with the compound Gaussian PDF, and marginalizing the product with respect to b1 . In
general, the likelihood function can be non-linear in nature and can yield a complicated PDF
of b1 . Using a second order Taylor series expansion for the log PDF, PBiGAMP [10] simplifies
the belief sent from y1 to b1 to a Gaussian whose mean and variance can be computed from
(R7)-(R9) of Table I. For q-bit ADCs, the closed form expressions for the conditional mean and
variance in (R7) and (R8) can be found in [35, Appendix A]. Unlike standard message passing,
PBiGAMP is computationally tractable as the messages contain just the mean and variances of
the PDFs.
After several approximate message flows between the factor nodes and the variable nodes,
PBiGAMP is expected to converge. The MMSE of b1 is computed as the expectation of the
effective marginal, i.e., the normalized product of all the M Gaussian PDFs received by b1 from
the factor nodes and the prior distribution on b1 . Notice that the normalization has to be done to
ensure that the PDF integrates to 1. The expectation step for the MMSE of b1 is given in (R18)
using (R11) and (R12) as the intermediate steps. Similarly, the expectation for the MMSE of c1
is given in (R16) using (R13) and (R14) as the intermediate steps. Thus, PBiGAMP provides
estimates of the vectorized sparse channel and the DFT of the phase error vector.
The generic implementation of PBiGAMP is computationally expensive primarily due to (R1),
(R2), (R5), (R12) and (R14) in Table I. It can be verified that each of these operations have a
2
complexity of O Nrx
Np2 Ntx L for a single PBiGAMP iteration to perform joint estimation. For
(i,∗)
instance, (R1) requires computing the scalar zbm (t) for every m ∈ IM and i ∈ INb , Therefore,
17
(R1) demands O(M Nb Nc ) computations, as each scalar computation requires Nc multiplications.
As M = Nrx Np , Nb = Np and Nc = Nrx Ntx L for the joint estimation, (R1) has a complexity of
2
O Nrx
Np2 Ntx L for every PBiGAMP iteration. By exploiting the structure in the joint estimation
problem, the complexity of PBiGAMP can be significantly reduced.
We describe our fast implementation of PBiGAMP for the joint estimation in the following
b (t) ∈ CNp ×1 , C
b (t) ∈
sub-sections. The tth iteration variables of PBiGAMP are defined as b
b (t) ∈ CNrx ×Np , such that bbi (t) , b
CNrx ×Ntx L and Z
ck (t) and zbm (t) are the ith , k th and mth entries
b (t) , vec(C(t))
b
b
of b
and vec(Z(t)).
1) Operations (R1)-(R3): For an io ∈ INp , we have
(io ,∗)
zbm
(t)
=
Nc
X
(io ,k)
zm
b
ck (t)
(22)
k=1
=
Nb X
Nc
X
i ,Np
(i,k) o
zm
ei
b
ck (t)
(23)
i=1 k=1
∗
io ,Np
b
= vecm UNrx C(t)Fdiag(U
e
)
Np
∗
b
= vecm UNrx C(t)Fdiag(U
)
,
Np (i )
o
(24)
(25)
where the compact form in (25) is obtained by going back from the tensor formulation to the
original bilinear model in (6). Likewise, for a ko ∈ INc , we have
(∗,ko )
zbm
(t) =
Nb X
Nc
X
(i,k)b
zm
bi (t)ekko ,Nc
(26)
i=1 k=1
b
= vecm UNrx ero ,Nrx (eco ,Ntx L )T Fdiag(U∗Np b(t))
b
= vecm UNrx (ro ) F(co ) diag(U∗Np b(t))
,
(27)
(28)
where (ro , co ) correspond to the row and column of the Nrx × Ntx L matrix version of eko ,Nc ,
with
ko = (co − 1)Nrx + ro .
(29)
(∗,∗)
∗ b
b
zbm
(t) = vecm UNrx C(t)Fdiag(U
b(t))
.
Np
(30)
Similarly, (R3) can be computed as
b
To evaluate (30), the product C(t)F
can be found using Nrx Ntx LNp computations. Using the
b
FFT over the resultant product, UNrx C(t)F
can be computed with an additional complexity
18
of O(Np Nrx logNrx ). Finally, the complexity to multiply the resultant Nrx × Np matrix with
b
the IFFT of b(t)
is O(Np logNp ) + Nrx Np . Therefore, the computational complexity of (30) is
2
O(Nrx Ntx Np L), unlike O(Nrx
Np2 Ntx L) of the generic implementation using (R1)-(R3).
(i,∗)
2) Operations (R4) , (R5): It can be noticed from (25) that zbm (t) is invariant with respect
to i and is given by
(i,∗)
(t)
zbm
1
b
p
vecm UNrx C(t)F .
=
Np
(31)
Using the invariance property in (31), a compact version of the first summand in (R4) can be
expressed as
Np
X
vib
2
(i,∗)
zbm
(t)
PNp
i=1
=
Np
i=1
vib
vecm
2
b
UNrx C(t)F
.
(32)
For the second summand in (R4), it can be shown from (28) that
1
(∗,ko )
b
zbm
(t) = √
⊗ aNrx (0) .
vecm F(co ) diag U∗Np b(t)
Nrx
(33)
(∗,k)
Furthermore, as co denotes the column number corresponding to ko (see (29)), zbm
(t) is
invariant ∀k ∈ IaNrx \I(a−1)Nrx , where a ∈ INtx L . To use this invariance for efficient computation
of the second summand in (R4), we define a row vector µc (t) ∈ R1×Ntx L containing the columnwise mean corresponding to Nrx × Ntx L matrix version of {vkc (t)}k∈INc as
X
1
µcn (t) =
vkc (t) .
Nrx
(34)
k∈InNrx \I(n−1)Nrx
With some algebraic manipulation, the second summand in (R4) can be simplified as
Nc
X
2
2
c
(∗,k)
∗ b
c
vk (t) zbm (t) = vecm µ (t) Fdiag(UNp b(t))
⊗ aNrx (0) .
(35)
k=1
To simplify the computations involved in (R5), we expand the summand using (17) as
Nb X
Nc
X
vib (t)vkc (t)
(i,k) 2
zm
=
Nb X
Nc
X
i=1 k=1
(36)
i=1 k=1
PNp
=
where (37) follows from (36) as |Gm,i | =
vib (t)vkc (t) |Gm,i Am,k |2
√1 ,
Np
i=1
N
c
vib (t) X
Np
2
vkc (t) ATk,m ,
(37)
k=1
∀m, i. Besides, as AT = F⊗UNrx , the entries of
AT are invariant within blocks of size Nrx ×Nrx . With arguments similar to the simplifications
involved in the second summand of (R4), (R5) can be efficiently evaluated as
PNp b
Nb X
Nc
X
v (t)
c
(i,k) 2
b
vi (t)vk (t) zm
= i=1 i vecm µc (t) |F|2 ⊗ aNrx (0) .
Np
i=1 k=1
(38)
19
(∗,k)
2
3) Operations (R11) , (R13): From (33), it can be observed that zbm (t)
is fixed for m ∈
IaNrx \ I(a−1)Nrx and k ∈ IbNrx \ I(b−1)Nrx , where a ∈ INp and b ∈ INtx L . The invariance of
(∗,k)
zbm (t)
2
with respect to m and k arise because of the kronecker product with aNrx (0) and the
column invariance in (29). To exploit this property in computing vkr (t) of (R11), we construct a
vector µz (t) ∈ RNp ×1 to contain the column-wise mean corresponding to the Nrx × Np matrix
s
version of {vm
(t)}m∈INp Nrx , i.e.,
µzk (t) =
1
Nrx
X
s
vm
(t)
.
(39)
m∈IkNrx \I(k−1)Nrx
−1
With the above definitions, a simplified version of vjr (t)
in (R11) can be given as
M
X
2
s
(∗,k) 2
∗ b
z
vm (t) zbm
Fdiag(UNp b) µ (t) ⊗ aNrx (0) .
= veck
(40)
m=1
From (31), we rewrite (viq (t))−1 in (R13) as
(viq (t))−1
=
M
X
s
vm
vecm
2
b
UNrx C(t)F
m=1
s
= v vec
2
b
UNrx C(t)F
,
∀i ∈ INp .
b
b = Fdiag U∗ b(t)
and consider the term in
4) Operations (R12) , (R14): We define F(t)
Np
the second summand of (R12) for k = ko . From (28) and (29), we have
M
M
X
X
(co )
(∗,ko )
∗
b
sbm (t)vecm UNrx (ro ) F(t)
sbm (t)b
zm (t) =
.
(41)
m=1
m=1
b
b ∈ CNrx ×Np defined such that sbm (t) = vecm S(t)
, (41) can be expressed as,
With S(t)
M
X
(∗,ko )
sbm (t)b
zm
(t)∗
D
E
(co )
b
b
= S(t), UNrx (ro ) F(t)
(42)
m=1
∗
∗
b
b (co )
= UNrx (ro ) S(t)
F(t)
d∗ .
b F(t)
= vecko U∗Nrx S(t)
The term in the third summand of (R12) can be given by
PNp b
Np
M
M
X
X
X
v (t) |Am,k |2
s
b
(i,k) 2
s
vm (t)
vi (t) zm
=
vm (t) i=1 i
Np
m=1
m=1
i=1
!
PNp b
i=1 vi (t)
T 2 s
veck A v .
=
Np
(43)
(44)
20
Once again exploiting invariance within AT , we efficiently compute the third summand in
(R12) as
M
X
s
vm
(t)
m=1
Np
X
vib (t)
(i,k) 2
zm
PNp
b
i=1 vi (t)
Np
=
i=1
!
veck
|F|2 µz (t) ⊗ aNrx (0) .
(45)
The term in the second summand of (R14) can be rewritten using (25) as
M
X
D
E
(i,∗)∗
∗
b
b
= S(t),
)
.
sbm (t)b
zm
UNrx C(t)Fdiag(U
Np (i)
(46)
m=1
b
For fast implementation of (46), we first compute the column wise inner product between S(t)
b
then perform a Fast Fourier
Transform (FFT). We construct g ∈ CNp ×1 ,
and UNrx C(t)F
and
b
b
. Hence, the second summand in (R14) simplifies
such that gk =
S(t)
, UNrx C(t)F
(k)
(k)
to
M
X
(i,∗)∗
sbm (t)b
zm
= veci UNp g .
(47)
m=1
The term in the third summand of (R14) can be expressed as
M X
Nc
X
s
(i,k)
vm
(t)vkc (t) zm
2
=
m=1 k=1
M X
Nc
X
s
vm
(t)vkc (t) |Gm,i Am,k |2
m=1 k=1
Nc
M
X
1 X s
2
=
vkc (t) ATk,m .
vm (t)
Np m=1
k=1
Using the compact form of
M X
Nc
X
(48)
PNc
k=1
vkc (t) ATk,m
s
(i,k)
vm
(t)vkc (t) zm
m=1 k=1
2
=
2
(49)
from (38), we rewrite (49) as
M
1 X s
vm (t)vecm µc (t) |F|2 ⊗ aNrx (0) .
Np m=1
(50)
Furthermore, exploiting the kroenecker product with aNrx (0), we have
M X
Nc
X
m=1 k=1
(i,k)
s
vm
(t)vkc (t) zm
2
=
Nrx c
µ (t) |F|2 µz (t).
Np
(51)
5) Complexity of the simplified operations: Using the complexity of matrix multiplications
and FFTs, the complexity of our simplifications is summarized in Table II. It can be noticed
that the overall complexity of PBiGAMP using our implementation is O (Nrx Np Ntx L), thereby
achieving a speedup factor of Np Nrx compared to the generic implementation. After all possible
simplifications of PBiGAMP for joint estimation, training design is possibly the only frontier
that can be exploited to further reduce the computational complexity.
21
Operation
Complexity
b (t) F
UNrx C
O (Nrx Ntx LNp )
b (t)
U∗Np b
O (Np logNp )
(R3), (R13)
O (Nrx Np )
(R4), (R5), (R11)
O (Ntx LNp )
(R12), (R14)
O (Nrx Ntx LNp )
TABLE II
C OMPLEXITY OF A SINGLE PB I GAMP
ITERATION USING A FAST IMPLEMENTATION .
D. Insights into training design
In this section, we explain how a reasonable training solution that allows a low complexity
implementation does not permit joint estimation. Structured training blocks that aid fast transforms can reduce the complexity of PBiGAMP operations involving multiplications with the
training matrix. Furthermore, such blocks also occupy a lower memory footprint relative to
unstructured ones of the same dimension. For example, training blocks that contain circulantly
shifted rows of a fixed Zadoff-Chu (ZC) sequence were proposed in [12] for fast channel
estimation using EM-GAMP. Similar to EM-GAMP algorithm, circulant training matrices also
b
aid fast matrix multiplications in PBiGAMP. For instance, the computation of C(t)F
in Table II
can be accelerated in every PBiGAMP iteration for the joint estimation. Shifted ZC training
blocks, however, result in a continuum of optimal solutions for bilinear optimization problem.
We illustrate the “CFO propagation effect” to show the trade-off between fast message passing
and identifiability in the joint estimation problem.
To explain the CFO propagation effect, we consider Np = Ntx L pilots and a circulant training
matrix T as per [12], for a narrowband system, i.e., L = 1. Because T∗ is also circulant, the
eigenvectors of T∗ are the columns of the DFT matrix UNtx . Let ΛT be a diagonal matrix
containing the conjugated eigenvalues of T∗ , such that T∗ UNtx = UNtx Λ∗T . With Λ defined as
diag (d (, β)), the noiseless unquantized received block in (6) can be given as
Z = UNrx CU∗ Ntx TΛ
= UNrx CΛT U∗ Ntx Λ .
(52)
(53)
For our analysis, we assume that there is no phase noise, i.e., β = 0 and choose the CFO () to
be an integer multiple of 2π/Np . We define the integer d = Np /2π and interpret Λ as a matrix
22
containing the eigenvalues of the d circulant delay matrix Jd ∈ CNp ×Np , i.e., Jd UNtx = UNtx Λ .
As Λ∗ = Λ− , it can be shown that JNtx −d UNtx = UNtx Λ∗ . Furthermore, as JNtx −d is a real
matrix and UNtx = U∗Ntx , we have U∗Ntx Λ = JNtx −d U∗Ntx . It follows from (53) that
Z = UNrx CΛT JNtx −d U∗ Ntx .
(54)
Comparing (54) with (53), it can be observed that the beamspace matrix C and a CFO of
2πd/Np , result in the same received samples as the beamspace matrix CΛT JNtx −d and zero
CFO. We call this as the CFO propagation effect and study its impact on the identifiability of
the channel and the CFO in joint estimation.
Now, we show that training blocks consisting of circulantly shifted ZC sequences propagate
the CFO into the channel in an inseparable manner. From the perfect autocorrelation property
of ZC sequences, it can be concluded that the diagonal entries of ΛT have constant modulus.
In such case, Λ−1
T is well defined and (54) can be rewritten as
∗
Z = UNrx CΛT JNtx −d Λ−1
T ΛT U Ntx
(55)
∗
= UNrx CΛT JNtx −d Λ−1
T U Ntx T
|
{z
}
(56)
∆
=C()
= UNrx C()U∗ Ntx TΛ0 .
(57)
Notice that Λ0 in (57) is just an identity matrix, also interpreted as a zero CFO perturbation. We
compare (57) with (53) and define a beamspace matrix matrix C() = CΛT JNtx −d Λ−1
T . It can
be noticed that C() has the same sparsity as that of C, as permutation and scaling operations
determined by JNtx −d and ΛT preserve sparsity. Therefore, if (C, ) is a solution to the joint
estimation problem, (C(), 0) is also a solution. In fact, any frequency that is an integer multiple
of 2π/Np can be propagated into the channel matrix so that (C(2πk/Np ), − 2πk/Np ) is a
solution for every k ∈ INp . Therefore, the circulantly shifted ZC training that aids fast message
passing results in a continuum of optimal solutions for the joint estimation problem.
V. S IMULATIONS
In this section, we provide simulation results for the proposed bilinear message passing based
joint CFO and channel estimation algorithm. We consider a hardware architecture in Fig. 1 that
uses a uniform linear array of antennas and choose Ntx = Nrx = 32. In this work, we assume that
the receiver is equipped with one-bit ADCs, i.e., q = 1. To obtain a performance benchmark for
23
one-bit receivers, we evaluate our algorithm for full resolution receivers, i.e., q = ∞, although
they may not be practical at large bandwidths due to high power consumption. We consider
a mmWave carrier frequency of 38 GHz and an operating bandwidth of W = 100 MHz [36],
which corresponds to a symbol duration of T = 10 ns.
We describe the simulation parameters of the clustered mmWave channel model in (3). We
consider L = 16 taps to model the mmWave channel in the 100 MHz bandwidth. Such model
is good enough for RMS delay spreads upto 33 ns, assuming L = 5W τrms . Our assumption
is reasonable as 95% of the measured RMS delay spreads at 38 GHz were found to be less
than 37.8 ns [37]. The measurements in [37] were made in a UMi-LoS environment using a
resolution of 2 ns. We assume Ncs = 4 clusters, each comprising of 10 rays and the complex ray
gains as IID standard normal random variables. Furthermore, the AoAs and AoDs of the rays
within a cluster are chosen from a laplacian distribution corresponding to an angle spread of
P
2
2
15◦ . The wideband channel is scaled so that E
` kH [`]kF = E kCkF = Nrx Ntx , where the
expectation is taken across several channel realizations. The channel matrix generated with the
aforementioned parameters is practical at mmWave and the angle-delay domain representation,
i.e., C, can be verified to be approximately sparse. Our joint estimation algorithm does not
require any knowledge about the sparsity order of C and learns it using the EM algorithm.
Our PBiGAMP based approach exploits compressibility of the phase error vector in the Fourier
basis. To evaluate the worst case performance of our algorithm, we choose a CFO that is
maximally off grid and within the practical limits. As the resolution of the DFT grid is 2π/Np ,
we choose = 45π/1024 for a training block of Np = 1024 pilots. The corresponding analog
domain CFO can be verified to be 2.2 MHz, which is about 58 ppm of the carrier frequency.
We set βtx = βrx = 0.047 rad so that the standard deviation of the effective Wiener phase noise
process is β = 0.067 rad. These parameters translate to a phase noise level of about −85 dBc/Hz
at 1 MHz offset for each of the TX and RX oscillators [25], and meet the specifications of a
38 GHz oscillator [38].
The performance of our joint estimation algorithm is evaluated using the Normalised Mean
Square Error (NMSE) of the channel estimate and the mean square error (MSE) of the CFO
estimate . For a given SNR, the variance of the IID Gaussian noise, i.e., σ 2 in (1), is chosen
such that
SNR = 10log10
kTk2F
Np σ 2
!
.
(58)
24
For our simulations, we consider training blocks comprising of IID QPSK entries, IID Gaussian
entries, and shifted ZC sequences proposed in [12]. For mmWave systems, the IID QPSK training
is more practical than the IID Gaussian one, as it can be generated using a TX architecture that
is as simple as analog-beamforming with 2-bit phase shifters.
A. NMSE of the channel estimate
Due to the bilinear nature of the problem, we can only estimate the wideband channel or
equivalently C upto a scale factor. Furthermore, any positive amplification of C results in the
same received block in one-bit receivers at high SNR. Therefore, the NMSE of the channel
estimate is defined as
NMSE = E
2
b
C − γC
kCk2F
b
where γ is a scalar such that γ = arg min C − aC
a
F
F
,
(59)
b The matrix γ C
b
for a given C and C.
can be considered as the normalized angle-delay domain estimate of the wideband channel. The
function E[.] denotes the empirical expectation and is taken across several channel and training
realizations. For a given realization of the channel and its estimate, we define the Normalized
2
b /kCk2 .
Squared Error (NSE) as C − γ C
F
F
For a sequence of Np = 1024 pilots, our joint estimation algorithm recovered the channel
within acceptable limits with a probability greater than 0.95, at a SNR of 0 dB using IID QPSK
training. The number of outliers in this case is determined by the phase-transition region of
PBiGAMP [22]. It can be observed from Fig. 3 that the probability of successful recovery
monotonically increases as a function of the training length and quickly approaches 1. To ignore
the effect of outliers, the channel NMSE and CFO MSE results we report are averages over
95% of the realizations for Np = 1024. In practice, the failure probability can be lowered by
increasing the training length or by designing a retransmission protocol that accounts for the
failure. As seen in Fig. 4, the NMSE monotonically decreases with the number of pilots. In
practical wireless systems, the choice of the number of pilots is determined by the channel
coherence time [3].
In Fig. 5, we plot the NMSE as a function of the SNR for various training sequences. It can
be observed that the reconstruction error is approximately the same for IID QPSK training and
IID Gaussian training matrices. For the one-bit case, the NMSE saturates at high SNR because
25
1
0.9
0.8
Pr(NSE 5 x)
0.7
0.6
0.5
0.4
0.3
Np
Np
Np
Np
0.2
0.1
= 512
= 1024
= 2048
= 4096
0
-15
-12.5
-10
-7.5
-5
-2.5
0
x(dB)
Fig. 3. The empirical CDF of the Normalized Squared Error of the channel estimate obtained using an IID QPSK training,
at a SNR of 0 dB in a one-bit receiver. The reconstruction performance in terms of the recovery probability and the mean
monotonically improve with the number of pilots.
-5
IID QPSK, 1-bit
IID QPSK, 1-bit
-6
-7
NMSE (dB)
-8
-9
-10
-11
-12
-13
-14
-15
512
1024
2048
3072
4096
5120
Number of pilots (Np )
Fig. 4. The NMSE of the channel estimate as a function of the training length, for an IID QPSK training at a SNR of 0 dB.
The NMSE monotonically decreases with the training length for one-bit and full resolution receivers. Due to the quantization
noise in one-bit receivers, NMSE in the one-bit case is higher than the full resolution one.
the recovery performance is limited by the quantization noise. As shown in Fig. 4, the channel
reconstruction error in one-bit receivers can be further decreased by using a higher number of
pilots for the training. It can be noticed from Fig. 5 that joint estimation with circulantly shifted
ZC sequences proposed in [12] performs poorly. The failure due to such structured matrices can
be attributed to the CFO propagation effect and confirms with our analysis in Section IV-D.
26
0
NMSE (dB)
-5
-10
-15
-20
-25
-10
IID QPSK, 1-bit
IID Gaussian , 1-bit
Shifted ZC, 1-bit
IID QPSK , 1-bit
IID Gaussian , 1-bit
Shifted ZC, 1-bit
-5
0
5
10
15
20
SNR(dB)
Fig. 5. NMSE of the channel estimate obtained with one-bit and full resolution channel measurements for Np = 1024. Joint
estimation is possible with IID Gaussian and IID QPSK training, but not with shifted ZC training due to the CFO propagation
effect discussed in Section IV-D. It can be observed that the NMSE for the one-bit case saturates at high SNR due to quantization
noise.
B. MSE of the CFO estimate
The CFO in our algorithm is obtained using an Extended Kalman Filter [34] on the inverse
b If b
DFT of b.
is the estimate of the CFO in the digital domain, the MSE of the CFO estimate
is given by E ( − b
)2 . For a sequence of 1024 pilots, the MSE of b
is shown as a function of
the SNR in Fig. 6, for IID QPSK and IID Gaussian training matrices. It can be noticed from
Fig. 6 that the MSE of the CFO estimate saturates even for the full resolution case because the
performance of the EKF is limited by the phase noise at high SNR [39]. Similar to the NMSE
of the channel estimate, the MSE of b
is expected to saturate at high SNR for one-bit receivers
due to quantization noise. The CFO estimation error, however, is determined by the phase noise
[39] because the CFO MSE for the one-bit case approaches that of the full resolution one. As
expected, the CFO MSE decreases with the number of pilots and is shown in Fig. 7.
C. Performance invariance with the CFO
In this section, we show that the performance of our joint estimation approach is invariant
to the CFO within practical limits. As an example, we evaluate our algorithm for different
values of the CFO in the range [−40 ppm, 40 ppm] of f1 . These limits were chosen according
to the IEEE 802.11ad specifications. From Fig. 8, we see that the reconstruction errors in the
27
10-1
IID
IID
IID
IID
MSE of CFO estimate
10-2
QPSK , 1-bit
Gaussian , 1-bit
QPSK , 1-bit
Gaussian , 1-bit
10-3
10-4
10-5
10-6
-10
-8
-6
-4
-2
0
5
10
SNR(dB)
Fig. 6. MSE of the CFO estimate as a function of SNR for a training length of 1024. A maximally off-grid CFO of = 45π/1024
was chosen to evaluate the performance. At high SNR, the MSE saturates for both the one-bit case and the infinite resolution
case due to phase noise [39].
MSE of the CFO estimate
10-2
IID QPSK, 1-bit
IID QPSK, 1-bit
10-3
10
-4
10-5
10-6
512
1024
2048
3072
4096
Number of pilots (Np )
Fig. 7. MSE of the CFO estimate as a function of the training length for a SNR of 0 dB. Here, the CFO in the system was fixed
to = 45π/1024, and the phase noise standard deviation to β = 0.067 rad. The MSE decreases with the number of pilots and
the performance gap between the one-bit case and the full-resolution case is negligible. The sharp decrease in the CFO MSE
can be attributed to the phase-transition effect of PBiGAMP [22].
channel is constant across the practical range of the CFO. The invariance arises due to the use
of Bernoulli-Gaussian prior for the synchronization variables, that achieves robustness against
off-grid leakage effects.
28
-6
IID QPSK, 1-bit
IID QPSK, 1-bit
-7
NMSE (dB)
-8
-9
-10
-11
-12
-40
-30
-20
-10
0
10
20
30
40
CFO in ppm of f1
Fig. 8. For a SNR of 0 dB, the plot shows the invariance of the channel NMSE over the practical range of the CFO. The phase
noise standard deviation was set to 0.067 rad, and 1024 pilots were used for the joint estimation. It can be observed from the
plot that our joint estimation algorithm is robust to leakage effects that arise due to an off-grid CFO.
VI. C ONCLUSIONS
Most sparsity-aware channel estimation algorithms for mmWave systems assume perfect synchronization and perform poorly in the presence of such errors. In this work, we propose a
message passing based algorithm that can leverage the sparsity of the wideband mmWave channel
in addition to the compressibility of the phase error vector. We exploit the structure in the joint
estimation problem to provide a low complexity implementation of a bilinear message passing
algorithm. Unlike the existing methods that are specific to certain hardware architectures or use
non-coherent techniques, our technique can be adapted to perform joint estimation with other
mmWave architectures.
Simulation results show that it is possible to perform joint CFO and channel estimation
using IID QPSK and IID Gaussian training matrices. The CFO propagation effect proposed
in this work shows that there is a trade-off between fast message passing using structured
training and identifiability in joint estimation. In our future work, we will also consider frame
synchronization while ensuring low complexity and scalability of our algorithm to different
mmWave architectures.
ACKNOWLEDGMENT
The authors would like to thank Jianhua Mo and Philip Schniter for helping them in getting
started with the GAMP algorithm.
29
R EFERENCES
[1] S. Rangan, T. S. Rappaport, and E. Erkip, “Millimeter-wave cellular wireless networks: Potentials and challenges,” in Proc.
of the IEEE, vol. 102, no. 3, pp. 366–385, 2014.
[2] A. Alkhateeb, J. Mo, N. Gonzalez Prelcic, and R. W. Heath Jr., “MIMO precoding and combining solutions for millimeter
wave systems,” IEEE Commun. Mag., vol. 52, no. 12, pp. 122–131, 2014.
[3] R. W. Heath, Introduction to Wireless Digital Communication: A Signal Processing Perspective. Pearson Education, 2017.
[4] R. W. Heath, N. Gonzalez-Prelcic, S. Rangan, W. Roh, and A. M. Sayeed, “An overview of signal processing techniques
for millimeter wave MIMO systems,” IEEE J. Sel. Topics Signal Process., vol. 10, no. 3, pp. 436–453, 2016.
[5] E. J. Candès and M. B. Wakin, “An introduction to compressive sampling,” IEEE Signal Process. Mag., vol. 25, no. 2,
pp. 21–30, 2008.
[6] P. T. Boufounos and R. G. Baraniuk, “One bit compressive sensing,” in Proc. of the IEEE 42nd Annual Conf. on Inform.
Sciences and Systems, 2008., 2008, pp. 16–21.
[7] Z. Marzi, D. Ramasamy, and U. Madhow, “Compressive channel estimation and tracking for large arrays in mm-wave
picocells,” IEEE J. Sel. Topics Signal Process., vol. 10, no. 3, pp. 514–527, 2016.
[8] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, “Channel estimation and hybrid precoding for millimeter wave
cellular systems,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831–846, 2014.
[9] J. Mo, P. Schniter, N. G. Prelcic, and R. W. Heath, “Channel estimation in millimeter wave MIMO systems with one-bit
quantization,” in Proc. of the Asilomar Conf. on Signals, Systems and Computers, 2014, pp. 957–961.
[10] D. L. Donoho, A. Maleki, and A. Montanari, “Message-passing algorithms for compressed sensing,” Proc. of the National
Academy of Sciences, vol. 106, no. 45, pp. 18 914–18 919, 2009.
[11] J. P. Vila and P. Schniter, “Expectation-maximization Gaussian-mixture approximate message passing,” IEEE Trans. on
Signal Process., vol. 61, no. 19, pp. 4658–4672, 2013.
[12] J. Mo, P. Schniter, and R. W. Heath Jr., “Channel estimation in broadband millimeter wave MIMO systems with few-bit
ADCs,” IEEE Trans. on Signal Process., 2017.
[13] H. Yan and D. Cabric, “Compressive sensing based initial beamforming training for massive MIMO millimeter-wave
systems,” in Proc. of the IEEE Global Conf. on Signal and Info. Process. (GlobalSIP), 2016, pp. 620–624.
[14] N. J. Myers and R. W. Heath Jr., “A compressive channel estimation technique robust to synchronization impairments,” in
Proc. of the 18th IEEE Intl. Workshop on Signal Process. Adv. in Wireless Commun. (SPAWC), 2017, pp. 1–5.
[15] J. Chen, Y.-C. Wu, S. Ma, and T.-S. Ng, “Joint CFO and channel estimation for multiuser MIMO-OFDM systems with
optimal training sequences,” IEEE Trans. on Signal Process., vol. 56, no. 8, pp. 4008–4019, 2008.
[16] W. Zhang, Q. Yin, W. Wang, and F. Gao, “One-shot blind CFO and channel estimation for OFDM with multi-antenna
receiver,” IEEE Trans. on Signal Process., vol. 62, no. 15, pp. 3799–3808, 2014.
[17] O. Abari, H. Hassanieh, M. Rodreguez, and D. Katabi, “Millimeter wave communications: From point-to-point links to
agile network connections,” in Proc. of the ACM Workshop on Hot Topics in Networks, 2016, pp. 169–175.
[18] M. E. Rasekh, Z. Marzi, Y. Zhu, U. Madhow, and H. Zheng, “Noncoherent mmwave path tracking,” in Proc. of the 18th
ACM Intl. Workshop on Mobile Computing Systems and Applications, 2017, pp. 13–18.
[19] N. J. Myers and R. W. Heath Jr., “Joint CFO and channel estimation in millimeter wave systems with one-bit ADCs,” in
Proc. of the 7th IEEE Intl. Workshop on Comput. Adv. in Multi-Sensor Adapt. Process. (CAMSAP), 2017, pp. 1–5.
[20] N. D. Sidiropoulos, L. De Lathauwer, X. Fu, K. Huang, E. E. Papalexakis, and C. Faloutsos, “Tensor decomposition for
signal processing and machine learning,” IEEE Trans. on Signal Process., vol. 65, no. 13, pp. 3551–3582.
30
[21] E. J. Candes, T. Strohmer, and V. Voroninski, “Phaselift: Exact and stable signal recovery from magnitude measurements
via convex programming,” Commun. on Pure and Applied Mathematics, vol. 66, no. 8, pp. 1241–1274, 2013.
[22] J. T. Parker and P. Schniter, “Parametric bilinear generalized approximate message passing,” IEEE J. Sel. Topics Signal
Process., vol. 10, no. 4, pp. 795–808, 2016.
[23] Y. Fu, C. Tellambura, and W. A. Krzymien, “Limited-feedback precoding for closed-loop multiuser MIMO OFDM systems
with frequency offsets,” IEEE Trans. on Wireless Commun., vol. 7, no. 11, 2008.
[24] V. Syrjala and M. Valkama, “Receiver DSP for OFDM systems impaired by transmitter and receiver phase noise,” in Proc.
of the IEEE Intl. Conf. on Commun. (ICC), 2011, pp. 1–6.
[25] A. Demir, A. Mehrotra, and J. Roychowdhury, “Phase noise in oscillators: A unifying theory and numerical methods
for characterization,” IEEE Trans. on Circuits and Systems I: Fundamental Theory and Applications, vol. 47, no. 5, pp.
655–674, 2000.
[26] N. G. Prelcic, A. Ali, V. Va, and R. W. Heath Jr., “Millimeter-wave communication with out-of-band information,” IEEE
Commun. Magazine, vol. 55, no. 12, pp. 140–146, 2017.
[27] K. Venugopal, A. Alkhateeb, N. G. Prelcic, and R. W. Heath, “Channel estimation for hybrid architecture-based wideband
millimeter wave systems,” IEEE Journal on Selected Areas in Commun., vol. 35, no. 9, pp. 1996–2009, 2017.
[28] J. Wang, “Beam codebook based beamforming protocol for multi-Gbps millimeter-wave WPAN systems,” IEEE Journal
on Selected Areas in Commun., vol. 27, no. 8, 2009.
[29] M. F. Duarte and R. G. Baraniuk, “Spectral compressive sensing,” Applied and Computational Harmonic Analysis, vol. 35,
no. 1, pp. 111–129, 2013.
[30] S. Ling and T. Strohmer, “Self-calibration and biconvex compressive sensing,” Inverse Problems, vol. 31, no. 11, p. 115002,
2015.
[31] S. Rangan, “Generalized approximate message passing for estimation with random linear mixing,” in Proc. of the IEEE
Intl. Symposium on Inform. Theory, 2011, pp. 2168–2172.
[32] S. M. Kay, Fundamentals of statistical signal processing.
Prentice Hall PTR, 1993.
[33] F. R. Kschischang, B. J. Frey, and H.-A. Loeliger, “Factor graphs and the sum-product algorithm,” IEEE Trans. on Inform.
Theory, vol. 47, no. 2, pp. 498–519, 2001.
[34] P. Zhao, L. Kuang, and J. Lu, “Carrier frequency offset estimation using extended Kalman filter in uplink OFDMA systems,”
in Proc. of the IEEE Intl. Conf. on Commun., 2006., vol. 6, pp. 2870–2874.
[35] C.-K. Wen, C.-J. Wang, S. Jin, K.-K. Wong, and P. Ting, “Bayes-optimal joint channel-and-data estimation for massive
MIMO with low-precision ADCs,” IEEE Trans. on Signal Process., vol. 64, no. 10, pp. 2541–2556.
[36] F. B. Mismar and B. L. Evans, “Machine learning approach to estimating mmWave signal measurements during handover,”
arXiv preprint arXiv:1710.01879, 2017.
[37] Aalto Univ., AT & T, BUPT, CMCC, Ericsson, Huawei, Intel, KT Corporation, Nokia, NTT DOCOMO, New York Univ.,
Qualcomm, Samsung, Univ. of Bristol, and Univ. of Southern California, “5G channel model for bands up to 100 GHz
(v2.3), Tech. Rep.” http://www.5gworkshops.com/5gcm.html, 2016, [Online].
[38] “MMIC Voltage Controlled Oscillator,” http://www.tlcprecision.com/Data%20Sheets/TLCO01981.pdf, 2013, [Online].
[39] O. H. Salim, A. A. Nasir, H. Mehrpouyan, W. Xiang, S. Durrani, and R. A. Kennedy, “Channel, phase noise, and frequency
offset in OFDM systems: Joint estimation, data detection, and hybrid Cramer-Rao lower bound,” IEEE Trans. on Commun.,
vol. 62, no. 9, pp. 3311–3325, 2014.
| 7cs.IT
|
FEAST: An Automated Feature Selection
Framework for Compilation Tasks
Pai-Shun Ting
arXiv:1610.09543v1 [cs.PL] 29 Oct 2016
Department of Electrical Engineering and Computer Science
University of Michigan
Ann Arbor, Michigan 48109, USA
Email: [email protected]
Chun-Chen Tu
Department of Statistics
University of Michigan
Ann Arbor, Michigan 48109, USA
Email: [email protected]
Pin-Yu Chen
IBM T. J. Watson Research Center
Yorktown Heights, New York 10598, USA
Email: [email protected]
Ya-Yun Lo
Adobe Systems
San Francisco, California 94103, USA
Email: [email protected]
Shin-Ming Cheng
Department of Computer Science and Information Engineering
National Taiwan University of Sciecne and Technology
Taipei 106, Taiwan
Email: [email protected]
Abstract—Modern machine-learning techniques greatly reduce
the efforts required to conduct high-quality program compilation,
which, without the aid of machine learning, would otherwise
heavily rely on human manipulation as well as expert intervention. The success of the application of machine-learning
techniques to compilation tasks can be largely attributed to the
recent development and advancement of program characterization, a process that numerically or structurally quantifies a
target program. While great achievements have been made in
identifying key features to characterize programs, choosing a
correct set of features for a specific compiler task remains an ad
hoc procedure. In order to guarantee a comprehensive coverage
of features, compiler engineers usually need to select excessive
number of features. This, unfortunately, would potentially lead
to a selection of multiple similar features, which in turn could
create a new problem of bias that emphasizes certain aspects
of a program’s characteristics, hence reducing the accuracy and
performance of the target compiler task. In this paper, we propose
FEAture Selection for compilation Tasks (FEAST), an efficient
and automated framework for determining the most relevant
and representative features from a feature pool. Specifically,
FEAST utilizes widely used statistics and machine-learning tools,
including LASSO, sequential forward and backward selection,
for automatic feature selection, and can in general be applied
to any numerical feature set. This paper further proposes an
automated approach to compiler parameter assignment for assessing the performance of FEAST. Intensive experimental results
demonstrate that, under the compiler parameter assignment
task, FEAST can achieve comparable results with about 18% of
features that are automatically selected from the entire feature
pool. We also inspect these selected features and discuss their
roles in program execution.
Index Terms—Compiler optimization, feature selection,
LASSO, machine learning, program characterization
I. I NTRODUCTION
Program characterization, a process to numerically or structurally quantify a target program, allows modern machinelearning techniques to be applied to compiler tasks, since most
machine-learning methods assume numerical inputs for both
training and testing data. Program characterization is usually
achieved by extracting from the target program a set of static
features and/or a set of dyanamic features. Static features can
be obtained directly from the source code or intermediate
TABLE I: List of all original 56 static features from cTuning
Compiler Collection [1].
ft1
ft2
ft3
ft4
ft5
ft6
ft7
ft8
ft9
ft10
ft11
# basic blocks in the
method
# basic blocks with a single
successor
# basic blocks with two
successors
ft29
# basic blocks with more
then two successors
# basic blocks with a single
predecessor
# basic blocks with two
predecessors
# basic blocks with more
then two predecessors
# basic blocks with a single
predecessor and a single
successor
# basic blocks with a single
predecessor and two successors
# basic blocks with a two
predecessors and one successor
# basic blocks with two
successors and two predecessors
ft32
ft30
ft31
ft33
ft34
ft35
ft36
ft37
ft38
ft39
ft12
# basic blocks with more
then two successors and
more then two
ft40
ft13
# basic blocks with # instructions less then 15
ft41
ft14
ft15
ft16
ft17
ft18
ft19
ft20
ft21
ft22
ft23
ft24
ft25
ft26
ft27
ft28
# basic blocks with # instructions in the interval
[15, 500]
# basic blocks with # instructions greater then 500
# edges in the control flow
graph
# critical edges in the control flow graph
# abnormal edges in the
control flow graph
# direct calls in the method
# conditional branches in
the method
# assignment instructions
in the method
# binary integer operations
in the method
# binary floating point operations in the method
#instructions in the method
Average of # instructions in
basic blocks
Average of # phi-nodes at
the beginning of a basic
block
Average of arguments for a
phi-node predecessors
# basic blocks with no phi
nodes
ft42
ft43
ft44
ft45
ft46
ft47
ft48
ft49
ft50
ft51
ft52
ft53
ft54
ft55
ft56
# basic blocks with phi
nodes in the interval [0, 3]
# basic blocks with more
then 3 phi nodes
# basic block where total #
arguments for all phi-nodes
is in greater then 5
# basic block where total #
arguments for all phi-nodes
is in the interval [1, 5]
# switch instructions in the
method
# unary operations in the
method
# instruction that do
pointer arithmetic in the
method
# indirect references via
pointers (”*” in C)
# times the address of a
variables is taken (”&” in
C)
# times the address of a
function is taken (”&” in
C)
# indirect calls (i.e. done
via pointers) in the method
# assignment instructions
with the left operand an
integer constant in the
method
# binary operations with
one of the operands an
integer constant in the
method
# calls with pointers as arguments
# calls with the # arguments is greater then 4
# calls that return a pointer
# calls that return an integer
# occurrences of integer
constant zero
# occurrences of 32-bit integer constants
# occurrences of integer
constant one
# occurrences of 64-bit integer constants
# references of a local variables in the method
# references (def/use) of
static/extern variables in
the method
# local variables referred in
the method
# static/extern variables referred in the method
# local variables that are
pointers in the method
# static/extern variables
that are pointers in the
method
# unconditional branches in
the method
representation of the target program, while the procurement
of dynamic features usually requires actually executing the
target program to capture certain run-time behavior.
With current intensive research on program characterization,
new features, both static and dynamic, are continuously being
proposed. An example of a static feature set is shown in Table
I, which lists all the 56 original static features extracted by
Milpost GCC from cTuning Compiler Collection [1], with
many of them being different yet non-independent features.
For example. ft. 8 implies ft. 2 and ft. 5, meaning that these
features are correlated. In a compiler task, determining which
features to use or to include for program characterization is
of considerable importance, since different features can have
different effects, and hence different resulting performance
on a specific compiler task. Intuitively, including as many
features as possible for program characterization seems to be a
reasonable approach when considering feature selection. This,
however, essentially increases the dimensionality of the feature
space, and thus potentially introduces extra computational
overhead. Also, some features may not be relevant to the target
compiler task, and therefore behave equivalently as noise in
program characterization, harming the resulting performance.
Furthermore, many features, though different, capture very
similar characteristics of a program. The similarities among
features can produce bias that overemphasizes certain aspects
of a program’s characteristics, and consequently lead to an
inaccurate result for a target compiler task. Due to the aforementioned reasons, many machine-learning-aided compiler
tasks still heavily rely on expert knowledge for determining an
appropriate set of features to use. This, unfortunately, hinders
the full automation of compiler tasks using machine learning,
which is originally deemed as a tool to lower the involvement
of field expertise and engineer intervention.
This paper proposes FEAture Selection for compiler Tasks
(FEAST), an automated framework of feature selection for a
target compiler task. FEAST incorporates into its framework a
variety of feature selection methods, including the well-known
LASSO [2], sequential forward and backward selection. Given
a compiler task and a list of feature candidates, FEAST first
samples a small set of available programs as training data, and
then uses the integrated feature selection methods to choose M
most appropriate or relevant features for this specific compiler
tasks. The remaining programs can then be handled using only
the chosen features.
To demonstrate the feasibility and practicality of FEAST, we
assess its performance on a proposed method to assignment of
compiler parameters. Modern compilers usually embed a rich
set of tuning parameters to provide hints and guidance for
their optimization processes. To obtain an optimal compiled
executable program, exhaustive trials over all combinations
of tuning parameters of the utilized compiler is required.
This is, in general, excessively time consuming and hence
infeasible due to its combinatorial nature. As many other
compiler tasks, in practice, expert intervention is frequently
triggered and heuristics based on expertise and experience
are adopted as a general guideline for tuning parameters [3].
Unfortunately, fine tuning compiler’s parameters may require
multiple compilation processes, and can take up to weeks or
months to complete for a moderate program size, entailing
a huge burden and software engineering. In this work, as a
test case for FEAST, we develop an automated method to
assigning “good” parameters to a pool of programs using
machine-learning techniques. We then show that using FEAST,
the dimension of the feature space can be greatly reduced
to pertaining relevant features, while maintaining comparable
resulting performance.
!" # #$ %& '
("%%
!" # #$ %& '
)
*
Fig. 1: Flow diagrams of the proposed methods.
This paper is organized as follows. Sec. II details the
proposed FEAST framework as well as the automatic compiler
parameter assignment process. Sec. III presents experimental
results with detailed analysis. Related work is discussed in
Sec. IV. Finally, Sec. V draws some conclusions and provides
envisions for future work.
II. FEAST
AND
C OMPILER PARAMETER A SSIGNMENT
This section details the mechanisms of FEAST, and presents
the proposed compiler parameter assignment method.
A. FEAST
Given K training programs and a set of p numerical
features, FEAST assumes a linear model for resulting performance and feature values:
y = Xβ + β0 1
(1)
where y ∈ RK×1 is the compiled programs’ performance
vector, whose i-th entry denotes the performance (measured in
some pre-defined metrics) of the i-th program in the set of total
K training programs. X ∈ RK×p is a matrix whose (i, j)-th
entry denotes the value of the j-th extracted feature of the i-th
program. β ∈ Rp×1 and β0 ∈ R are the coefficients describing
the linear relationship between the features and the resulting
performance. 1 is a K × 1 vector with its elements being
all 1s. FEAST utilizes three widely used feature selection
methods, namely LASSO, sequential forward selection (SFS)
and sequential backward selection (SBS), all of which pick M
most influential features out of the total p features. Specifically,
the elastic net approach for LASSO adopted in FEAST selects
features by first solving optimization problem [2]:
minβ̃
1
y − X̃ β̃
K
2
+ λ β̃
2
(2)
1
where X̃ = X 1 . The first p elements of the solution β̃ are
the coefficient estimates whose magnitudes directly reflect the
influence of the corresponding features. M selected features
are chosen as those with coefficients of largest magnitudes.
SFS and SBS are other well-known feature selection methods.
For SFS, we sequentially or greedily select the most relevant
feature from the training programs until we have selected M
features. For SBS, we sequentially exclude the most irrelevant
feature until there are M features left. We omit algorithmic
descriptions of SFS and SBS, and refer interested readers to
[4] for implementation details.
B. Compiler Parameter Assignment Algorithms
This section discusses the proposed method for compiler
parameter assignment algorithm and the application of FEAST
to this task.
Given a compiler with many parameters to set, finding an
optimal assignment of parameters that can best compile a
target program is a vexing problem. This paper proposes a
machine-learning algorithm that automatically assigns “good”
parameters to target programs based on the known optimal
assignment to the training programs. The proposed method
works in two schemes: active training scheme and passive
training scheme (see Fig. 1). In active training scheme, the
users can actively acquire the optimal compiler parameters
for a subset of programs, while in passive training scheme,
the users are given as prior knowledge a set of programs
whose optimal compiler parameters are known. A practical
example of active training scheme is that a company has no
prior knowledge about compiler parameter assignment, and
it has a very limited budget which only allows it to select
a small set of programs for full tuning. Remaining large
set of programs, however, has to be quickly and efficiently
compiled. For passive training scheme, a company has a small
set of programs with well-tuned compiler parameters, and
it would like to quickly find good compiler parameters for
other programs. Note that, in the active training scheme, our
proposed method can also automatically choose a good set of
candidate programs for full tuning. The remaining section is
dedicated to detailing the two preceding schemes with FEAST
application to them respectively.
Active training scheme. Since acquirement of dynamic
features can be very expensive due to the potential need
for multiple compilations and iterative tuning, we opt to use
Algorithm 1 Active Training Scheme
Input: n programs with p static features, number K of
training programs for optimization
Output: compiler parameter assignment for each untrained
program
Procedures:
1) Partition n programs into K clusters by K-means
clustering
2) For each cluster select one program having the least
sum of distances to other programs in the same cluster
as a training program.
3) Find the set of optimal compiler parameters of the
selected K programs.
4) Use FEAST to perform feature selection on the selected K programs by regressing their optimal execution time with respect to their static features.
5) Repartition the untrained programs based on the similarities computed by the selected features.
6) For each untrained program, assign the compiler parameters of the closest trained program to it based on
the selected features.
numerical static features in the proposed compiler parameter
assignment task. Also, we use the execution time of a compiled
program as the measure of performance for that program.
Given n programs with p static features, we are granted to
choose K ≤ n programs as training samples and acquire
the optimal compiler parameters for each training program
that optimize the execution time. In order to choose the K
programs, we first compute the similarity of each program
pair based on the Euclidean distance between the corresponding static feature vectors, and partition the programs into
K clusters using K-means clustering. For each cluster, we
select one program that has the least average distance to
other programs in the same cluster as a training program.
Exhaustive trials are then conducted for the training programs
to obtain their optimal compiler parameters and the associated
execution time. Given the K training programs as well as
their execution time, FEAST can then select M features that
are most influential to the training programs’ performance
(execution time in this case) . We then leverage the selected
features to recompute similarities and repartition the programs.
Lastly, each untrained program is assigned by the optimal
parameters of the most similar training program. See Alg. 1
for detailed algorithm.
Passive training scheme. Different from the active training
scheme, in addition to n programs with p static features, we
are also given K pre-selected training programs. Therefore,
the methodology of the passive training scheme is similar to
that of the active training scheme, except that the clustering
procedures described in active training scheme are no longer
required. See Alg. 2 for detailed algorithm.
Algorithm 2 Passive Training Scheme
Input: n programs with p static features, K given training
programs
Output: compiler parameter assignment for each untrained
program
Procedures:
1) Find the set of optimal compiler parameters of the K
given training programs.
2) Use FEAST to perform feature selection on the selected K programs by regressing their optimal execution time with respect to their static features.
3) Compute distances between programs based on the
selected features.
4) For each untrained program, assign the compiler parameters of the closest trained program based on the
selected features.
III. P ERFORMANCE E VALUATION OF C OMPLIER
PARAMETER A SSIGNMENT
We test our implementations using the PolyBench benchmark suite [5] that consists of n = 30 programs. The programs
are characterized using p = 56 static features from cTuning Compiler Collection [1]. For the two proposed training
schemes, K trained programs are used for feature selection and
compiler parameter assignment, and for each feature selection
method (LASSO, SFS, SBS), we select the M = 10 most
relevant features.
A. Performance Comparison of the Active and Passive Training Schemes
Fig. 2 shows the program execution time of active and
passive training schemes. The minimal execution time refers
to the optimal performance over 192 possible combinations
of compiler parameters of every program. We also show the
results for the case where no tunable compiler parameters are
enabled. The results of minimal-time and no-tuning-parameter
are regarded as baselines for comparing the proposed FEAST
methods, as their execution time does not depend on the
number K of training programs. Furthermore, to validate the
utility of FEAST, we also calculate the execution time using
all features, i.e., the case where FEAST is disabled.
For the three feature selection methods introduced in Sec.
II-A (LASSO, SFS, and SBS), we select the top M = 10
important features and use these selected features to compute
program similarities and assign compiler parameters. In this
setting, we only compare the execution time of the untrained
programs, and we omit the computation time required to obtain
the optimal compiler parameters associated with the training
programs. We will consider the overall execution time of the
trained and untrained programs shortly.
In Fig. 2, each untrained program is executed once and we
sum up the execution time to get the overall execution time
under various values of K of trained programs. It is observed
that the execution time of both training schemes converge to
38
34
32
34
Executation time (s)
Executation time (s)
36
32
No feature selection
Lasso
Forward selection
Backward selection
Minimal time
No tuning parameter
30
28
26
24
30
28
26
24
22
22
20
No feature selection
Lasso
Forward selection
Backward selection
Minimal time
No tuning parameter
0
5
10
15
20
25
20
30
0
5
K: number of trained program
10
15
20
25
30
K: number of trained program
(a) Active training scheme.
(b) Passive training scheme.
Fig. 2: Program execution time of the active and passive training schemes. Minimal time refers to exhaustive parameter
optimization on every program. It can be observed that the execution time of the three feature selection algorithms integrated
in FEAST are comparable to that of using all features, which strongly suggests that important features affecting program
execution are indeed identified by FEAST.
Lasso
38
36
Passive training
Active training
36
34
Execution time (sec)
34
Execution time (sec)
SFS
38
Passive training
Active training
32
30
28
26
32
30
28
26
24
24
22
22
20
20
0
5
10
15
20
25
30
K: # trained program
0
5
10
15
20
25
30
K: # trained program
Fig. 3: Comparison of active and passive training schemes. Error bars represent standard deviation. The average execution
time of active training is smaller that that of passive training. The variations in active training scheme are caused by random
initialization in K-means clustering procedure, whereas the variations in passive training scheme are caused by randomness in
the selection of training programs.
the minimal execution time as K increases. The curve of the
passive training scheme is smoother than that of the active
training scheme due to the fact that the former is an averaged
result over 1000 trials of randomly selected training programs.
The trends in Fig. 2 can be explained as follows: when
adopting passive training, every program has an equal chance
to be selected as a training program. As K grows, the set of
available optimal compiler parameters for training programs
increases as well, resulting in the decrease in average total
execution time. Active training, on the other hand, adopts
K-means clustering when selecting the training programs.
Given a fixed K, only K programs can be selected for
training and compiler parameter optimization. Therefore, for
small K (e.g.,. K = 1 or 2), we might select the programs
whose optimal compiler parameters do not fully benefit other
untrained programs.
The execution time of cases with FEAST enabled are shown
to be comparable to that of using all features, which strongly
suggests that FEAST can successfully select important features
affecting program execution, leading to dimensionality reduc-
-5000
00
5
20
200
400
50
00
600
800
0
-2000
300
0
-3000
00
0
4000
1000
00
-1
0
Program execution times (Nexec)
(a) Active training scheme.
0
-2000
200
0
0
10
-3000
4000
-40
0
0
00
-2
200
400
600
800
-4000
-5000
3000
2000
5
-6000
1000
2000
-1000
-3
00
-4000
3000
1000
10
0
-3000
40
00
100
-2
0
3000
-1
00
0
10
0
00
-2000
15
0
00
-1
00
0
4000
3000
-1000
K: number of trained program
0
-10
00
0
00
-40
-30
0
20
3000
15
1000
4000
0
00
2000
5000
20
25
200
-20
100
00
-50
3000
2000
K: number of trained program
00
4000
5000
5000
25
Passive with lasso feature selection
30
5000
-5
0
Active with lasso feature selection
30
-6000
1000
Program execution times (Nexec)
(b) Passive training scheme.
Fig. 4: Time reduction for active and passive training schemes using LASSO feature selection. The parameter Nexec specifies
the number of times a program is executed. The contour indicates a phase transition in time reduction. The figure suggests
that with Nexec large enough, the proposed compiler parameter assignment method can provide time saving when considering
overall execution time.
tion while still attaining satisfactory execution time reduction.
To gain more insights on the performance of active and
passive training schemes, we compare the execution time
of LASSO and SFS in Fig. 3. The comparison of SBS is
omitted since in practice SBS is computationally inefficient
due to its sequential feature removal nature, especially for
high-dimensional data. It is observed in Fig. 3 that the average
execution time of active training is smaller than that of
passive training, since for active training, we are able to select
representative programs as training samples for optimization.
These results indicate that the execution time can be further
reduced when we have the freedom to select K representative
programs from clustering as training samples for compiler
parameter assignment.
B. Overall Execution Time Comparison
In this section, we consider the overall execution time,
including time overhead imposed by the proposed algorithms
as well as the execution time of every untrained program.
To this end, we introduce a parameter Nexec, the number of
executions per program. The motivation behind introducing
Nexec is as follows: for programs such as matrix operations,
users may execute these core programs multiple times (i.e.,
Nexec times). Therefore, as will be demonstrated by the
analysis detailed in this section, the time overhead introduced
by the proposed algorithms will be compensated as Nexec
increases, since the time spent in optimizing the training
programs comprises relatively small portion of the overall time
cost with large Nexec. Time Reduction (TR) can therefore
be computed using the following formula:
K
TR = Nexec · Tnull − Nexec · Tauto − Texhaustive
,
where Nexec denotes the number of executions for each
program, Tnull represents the total time to run every program
with all tunable compiler parameters disabled, Tauto denotes
the total time to run every program compiled by using the proK
posed compiler parameter assignment method, and Texhaustive
denotes the computation time for finding the optimal compiler
parameters for K training programs.
Fig. 4 shows the contour plot of the overall time reduction
metric for both training schemes with LASSO. Results using
SFS and SBS are similar, and are omitted in this paper. If TR is
positive, it indicates the overall execution time is smaller than
that with all tunable compiler parameters disabled. It can be
observed that for each K, TR becomes positive when Nexec
exceeds certain threshold value, implying that if a user uses
the compiled programs repeatedly, the proposed method could
potentially provide great time saving. Also for each K, the
preceding threshold of active training is lower than that of
passive training, since programs to exhaustively optimize are
actively chosen by the proposed method.
C. Features Selected by FEAST
To investigate the key features affecting compiler execution
time, we inspect the selected features from FEAST. Table
II lists the top 10 selected features using various feature
selection methods integrated in FEAST. The features are
selected with 3-fold cross validation method for regressing
the optimal execution time of all 30 programs with respect to
their static program features. Based on the selected features,
we can categorize the selected features into three groups:
1) Control Flow Graph (CFG), 2) assignments/operations,
and 3) phi node. CFG features describe a programs’ control
flow, which can be largely influenced by instruction branches,
such as “if-else”, “if- if” and “for loop” statements. The
selected CFG features are reasonable as in our program testing
dataset, for loop contributes to the major part of programs’
control flow. In addition, assignment operations are essential to
matrix operations, and hence possess discriminative power for
distinguishing programs. Lastly, Phi node is a special operation
of Intermediate Representation (IR). It is designed for Static
Single Assignment form (SSA) which enables a compiler to
perform further optimization, and hence it is an important
factor for program execution.
IV. R ELATED W ORK
Recent rapid development of program characterization, a
process to quantify programs, allows the application of modern
machine-learning techniques to the field fo code compilation
and optimization. These machine-learning techniques provide
powerful tools that are widely used in various aspects of
compilation procedures. For example, Buse and Weimer use
static features to identify ”hot paths” (executional paths that
are most frequently invoked) of a target program by applying
logistic regression [6] without ever profiling the program,
Kulkarni et al. build an evolving neural-network model that
uses static features to help guide the inlining heuristics in compilation process [7], and Wang and O’Boyle exploit the use
of artificial neural networks (ANNs) as well as support vector
machine (SVM) for automatic code parallelization on multicore systems [8], among others. Many existing applications of
machine-learning-enabled compiler tasks use features, either
static or dynamic, selected by the designers, and hence heavily
rely on field expertise. This work provides a comprehensive
solution to this problem by using modern statistical methods
to select appropriate features for a specific case.
There has also been a vast amount of research dedicated
to designing suitable features for target applications. For
example, it is shown in [9] that, for compiler tasks that
cannot afford the time cost for procurement of dynamic
features, carefully designed graph-based static features can
achieve accuracy and performance comparable to dynamic
features in some applications, regardless the fact that it is
prevailingly believed dynamic features are preferred due to
insightful information they provide. Another example is the
compiler parameters assignment task by Park et al. [12]. In
their work, an SVM-based supervised training algorithm is
used to train a set of support vectors that can help estimate
the performance or reaction of an unknown program to a set
of compiler parameters. They use newly-defined graph-based
static features for training, which achieves high performance
comparable to that using dynamic features, but without the
need to invoke multiple compilations and profiling. While in
general, it is possible to design dedicated features for specific
tasks, the applicability of these dedicated features to other
applications remain questionable. In the scenario where there
are excessive number of numerical features that may or may
not fit a target task, FEAST can help in selecting the most
meaningful and influential features.
As to the task of compiler parameter tuning, recent work has
been focused on automation of parameter assignment process.
Compiler parameter tuning has long been a crucial problem
which attracts a vast amount of attention. Recent trends and
efforts on exploiting the power of modern machine-learning
techniques have achieved tremendous success in compiler
parameter tuning. Stephenson and Amarasinghe demonstrate
the potential of machine learning on automatic compiler
parameter tuning by applying ANNs and SVM to predict
a suitable loop-unrolling factor for a target program [10].
This work is relatively restricted, since it deals with a single compiler parameter. Agakov et al. propose a computeraided compiler parameter tuning method by identifying similar
programs using static features [3], where certain level of
expert intervention is still required. Cavazos et al. characterize
programs with dynamic features, and use logistic regression, a
classic example of conventional machine-learning algorithm,
to predict good compiler parameter assignment [11]. While
providing state-of-the-art performance, [11] requires dynamic
features, which can be expensive to acquire. In [12], graphbased features are used along with SVM for performance
prediction given a compilation sequence. This work uses
dedicated features for the machine-learning task, and further
implicitly utilizes excessive number of candidate assignments
of compiler parameters in order to find a good assignment,
resulting in a non-scalable algorithm. On the other hand,
the proposed compiler parameter assignment algorithm is a
comprehensive assignment algorithm that does not require
dynamic features or dedicated static features. Furthermore,
the training data that need full optimization can be set fixed.
The good assignment for an unseen target program is directly
derived from that of trained programs, implying its scalability
with the number of potential assignments.
V. C ONCLUSIONS
AND
F UTURE W ORK
In this work, we propose FEAST, an automated framework
for feature selection in compiler tasks that incorporates with
well-known feature selection methods including LASSO, sequential forward and backward selection. We demonstrate the
feasibility and applicability of FEAST by testing it on a proposed method for the task of compiler parameter assignment.
The experimental results show that the three feature selection
methods integrated in FEAST can select a representative small
subset of static features that, when used in the compiler
parameter assignment task, can achieve non-compromised performance. We also validate the effectiveness of the proposed
methods by experimentally demonstrating significant overall
execution time reduction of our method in a practical scenario
where each program is required to run multiple times. Lastly,
we discussed the roles of the features selected by FEAST,
which provides deep insights into compilation procedures. In
summary, our contributions are two-fold:
1) We integrate into FEAST with various modern machinelearning and optimization techniques for feature selection for compilation tasks.
2) We demonstrate the applicability of FEAST by experimentally showing that it can achieve comparable
performance in compiler parameter assignment tasks
with a very small set of selected static features.
For future work, we are interested in exploring the inherent
structural dependencies of codes in each program as additional
features for compiler parameter assignment. We are also interested in integrating the proposed compiler parameter assignment algorithm with recently developed automated community
detection algorithms, such as AMOS [13], to automatically
cluster similar programs for the proposed passive and active
training schemes.
R EFERENCES
[1] “cTuning
Compiler
Collection.”
[Online].
Available:
http://ctuning.org/wiki/index.php?title=CTools:CTuningCC
[2] H. Zou and T. Hastie, “Regularization and variable selection via the
elastic net,” Journal of the Royal Statistical Society: Series B (Statistical
Methodology), vol. 67, no. 2, pp. 301–320, 2005.
[3] F. Agakov, E. Bonilla, J. Cavazos, B. Franke, G. Fursin, M. F. O’Boyle,
J. Thomson, M. Toussaint, and C. K. Williams, “Using machine learning
to focus iterative optimization,” in Proceedings of the International
Symposium on Code Generation and Optimization. IEEE Computer
Society, 2006, pp. 295–305.
[4] M. Dash and H. Liu, “Feature selection for classification,” Intelligent
data analysis, vol. 1, no. 1, pp. 131–156, 1997.
[5] “Polybench
benchmark
suite.”
[Online].
Available:
http://web.cse.ohiostate.edu/∼ pouchet/software/polybench/
[6] R. P. Buse and W. Weimer, “The road not taken: Estimating path
execution frequency statically,” in Proceedings of the 31st International
Conference on Software Engineering. IEEE Computer Society, 2009,
pp. 144–154.
[7] S. Kulkarni, J. Cavazos, C. Wimmer, and D. Simon, “Automatic
construction of inlining heuristics using machine learning,” in Code
Generation and Optimization (CGO), 2013 IEEE/ACM International
Symposium on. IEEE, 2013, pp. 1–12.
[8] Z. Wang and M. F. O’Boyle, “Mapping parallelism to multi-cores: a
machine learning based approach,” in ACM Sigplan notices, vol. 44,
no. 4. ACM, 2009, pp. 75–84.
[9] J. Demme and S. Sethumadhavan, “Approximate graph clustering for
program characterization,” ACM Transactions on Architecture and Code
Optimization (TACO), vol. 8, no. 4, p. 21, 2012.
[10] M. Stephenson and S. Amarasinghe, “Predicting unroll factors using supervised classification,” in International symposium on code generation
and optimization. IEEE, 2005, pp. 123–134.
[11] J. Cavazos, G. Fursin, F. Agakov, E. Bonilla, M. F. O’Boyle, and
O. Temam, “Rapidly selecting good compiler optimizations using performance counters,” in International Symposium on Code Generation
and Optimization (CGO’07). IEEE, 2007, pp. 185–197.
[12] E. Park, J. Cavazos, and M. A. Alvarez, “Using graph-based program
characterization for predictive modeling,” in Proceedings of the Tenth
International Symposium on Code Generation and Optimization. ACM,
2012, pp. 196–206.
[13] P.-Y. Chen, T. Gensollen, and A. O. Hero III, “Amos: An automated
model order selection algorithm for spectral graph clustering,” arXiv
preprint arXiv:1609.06457, 2016.
TABLE II: Top 10 selected features from various methods integrated in FEAST. The number in the bracket indicates the feature
ranking for each method.
LASSO
Number of basic blocks
with a single predecessor
and a single successor (6)
Number of basic blocks
with a single predecessor
and two successors (7)
Number of basic blocks
with more then two
successors and more than
two predecessors (8)
Number of basic blocks
with number of instructions
in the interval [15, 500] (5)
Number of assignment
instructions in the
method (9)
SFS
Number of binary integer
operations in the method (1)
Number of direct calls in the
method (9)
Number of binary floating
point operations in the
method (2)
Number of basic blocks
with phi nodes in the
interval [0,3] (4)
Number of basic block
where total number of
arguments for all phi-nodes
is greater than 5 (10)
Number of assignment
instructions in the
method (10)
Average of number of
phi-nodes at the beginning
of a basic block (10)
Number of basic blocks
with more than 3 phi
nodes (4)
Number of basic block
where total number of
arguments for all phi-nodes
is greater than 5 (7)
Number of switch
instructions in the
method (3)
Number of binary integer
operations in the method (1)
Number of unary operations
in the method (2)
Number of unary operations
in the method (3)
Number of basic blocks in
the method (3)
SBS
Number of basic blocks
with a two predecessors
and one successor (8)
Number of basic blocks with
a two predecessors and one
successor (7)
Number of conditional
branches in the method (5)
Number of basic blocks with
two successors and two
predecessors (2)
Number of instructions
in the method (9)
Number of basic blocks with
more then two successors and
more than two predecessors (6)
Number of basic blocks with
number of instructions greater
then 500 (4)
Number of basic blocks with
more than 3 phi nodes (5)
Number of basic block
where total number of
arguments for all phi-nodes
is in greater than 5 (8)
Number of assignment
instructions with the left
operand an integer constant
in the method (6)
Number of binary operations
with one of the operands an
integer constant in the
method (1)
| 2cs.AI
|
Published as a conference paper at ICLR 2018
O N THE E XPRESSIVE P OWER OF OVERLAPPING
A RCHITECTURES OF D EEP L EARNING
Or Sharir & Amnon Shashua
The Hebrew University of Jerusalem
{or.sharir,shashua}@cs.huji.ac.il
arXiv:1703.02065v4 [cs.LG] 24 Feb 2018
A BSTRACT
Expressive efficiency refers to the relation between two architectures A and B,
whereby any function realized by B could be replicated by A, but there exists
functions realized by A, which cannot be replicated by B unless its size grows significantly larger. For example, it is known that deep networks are exponentially
efficient with respect to shallow networks, in the sense that a shallow network
must grow exponentially large in order to approximate the functions represented
by a deep network of polynomial size. In this work, we extend the study of expressive efficiency to the attribute of network connectivity and in particular to the
effect of ”overlaps” in the convolutional process, i.e., when the stride of the convolution is smaller than its filter size (receptive field). To theoretically analyze
this aspect of network’s design, we focus on a well-established surrogate for ConvNets called Convolutional Arithmetic Circuits (ConvACs), and then demonstrate
empirically that our results hold for standard ConvNets as well. Specifically, our
analysis shows that having overlapping local receptive fields, and more broadly
denser connectivity, results in an exponential increase in the expressive capacity of
neural networks. Moreover, while denser connectivity can increase the expressive
capacity, we show that the most common types of modern architectures already
exhibit exponential increase in expressivity, without relying on fully-connected
layers.
1
I NTRODUCTION
One of the most fundamental attributes of deep networks, and the reason for driving its empirical
success, is the “Depth Efficiency” result which states that deeper models are exponentially more
expressive than shallower models of similar size. Formal studies of Depth Efficiency include the
early work on boolean or thresholded circuits (Sipser, 1983; Yao, 1989; Håstad and Goldmann,
1991; Hajnal et al., 1993), and the more recent studies covering the types of networks used in
practice (Pascanu et al., 2013; Montúfar et al., 2014; Eldan and Shamir, 2016; Cohen et al., 2016a;
Cohen and Shashua, 2016; Telgarsky, 2016; Safran and Shamir, 2016; Raghu et al., 2016; Poole
et al., 2016). What makes the Depth Efficiency attribute so desirable, is that it brings exponential
increase in expressive power through merely a polynomial change in the model, i.e. the addition
of more layers. Nevertheless, depth is merely one among many architectural attributes that define
modern networks. The deep networks used in practice consist of architectural features defined by
various schemes of connectivity, convolution filter defined by size and stride, pooling geometry and
activation functions. Whether or not those relate to expressive efficiency, as depth has proven to be,
remains an open question.
In order to study the effect of network design on expressive efficiency we should first define ”efficiency” in broader terms. Given two network architectures A and B, we say that architecture A is
expressively efficient with respect to architecture B, if the following two conditions hold: (i) any
function h realized by B of size rB can be realized (or approximated) by A with size rA ∈ O(rB );
(ii) there exist a function h realized by A with size rA , that cannot be realized (or approximated)
by B, unless rB ∈ Ω(f (rA )) for some super-linear function f . The exact definition of the sizes rA
and rB depends on the measurement we care about, e.g. the number of parameters, or the number
of “neurons”. The nature of the function f in condition (ii) determines the type of efficiency taking
place – if f is exponential then architecture A is said to be exponentially efficient with respect to
1
Published as a conference paper at ICLR 2018
architecture B, and if f is polynomial so is the expressive efficiency. Additionally, we say A is
completely efficient with respect to B, if condition (ii) holds not just for some specific functions
(realizable by A), but for all functions other than a negligible set.
In this paper we study the efficiency associated with the architectural attribute of convolutions,
namely the size of convolutional filters (receptive fields) and more importantly its proportion to
their stride. We say that a network architecture is of the non-overlapping type when the size of the
local receptive field in each layer is equal to the stride. In that case, the sets of pixels participating
in the computation of each two neurons in the same layer are completely separated. When the stride
is smaller than the receptive field we say that the network architecture is of the overlapping type. In
the latter case, the overlapping degree is determined by the total receptive field and stride projected
back to the input layer – the implication being that for the overlapping architecture the total receptive
field and stride can grow much faster than with the non-overlapping case.
As several studies have shown, non-overlapping convolutional networks do have some theoretical
merits. Namely, non-overlapping networks are universal (Cohen et al., 2016a; Cohen and Shashua,
2016), i.e. they can approximate any function given sufficient resources, and in terms of optimization, under some conditions they actually possess better convergence guaranties than overlapping
networks. Despite the above, there are only few instances of strictly non-overlapping networks used
in practice (e.g. Sharir et al. (2016); van den Oord et al. (2016)), which raises the question of why
are non-overlapping architectures so uncommon? Additionally, when examining the kinds of
architectures typically used in recent years, which employ a mixture of both overlapping and nonoverlapping layers, there is a trend of using ever smaller receptive fields, as well as non-overlapping
layers having an ever increasing role (Lin et al., 2014; Springenberg et al., 2015; Szegedy et al.,
2015). Hence, the most common networks used practice, though not strictly non-overlapping, are
increasingly approaching the non-overlapping regime, which raises the question of why having just
slightly overlapping architectures seems sufficient for most tasks?
In the following sections, we will shed some light on these questions by analyzing the role of
overlaps through a surrogate class of convolutional networks called Convolutional Arithmetic Circuits (ConvACs) (Cohen et al., 2016a) – instead of non-linear activations and average/max pooling
layers, they employ linear activations and product pooling. ConvACs, as a theoretical framework to
study ConvNets, have been the focused of several works, showing, amongst other things, that many
of the results proven on this class are typically transferable to standard ConvNets as well (Cohen
and Shashua, 2016; 2017). Though prior works on ConvACs have only considered non-overlapping
architectures, we suggest a natural extension to the overlapping case that we call Overlapping ConvACs. In our analysis, which builds on the known relation between ConvACs and tensor decompositions, we prove that overlapping architectures are in fact completely and exponentially more efficient
than non-overlapping ones, and that their expressive capacity is directly related to their overlapping
degree. Moreover, we prove that having even a limited amount of overlapping is sufficient for attaining this exponential separation. To further ground our theoretical results, we demonstrate our
findings through experiments with standard ConvNets on the CIFAR10 image classification dataset.
2
OVERLAPPING C ONVOLUTIONAL A RITHMETIC C IRCUITS
In this section, we introduce a class of convolutional networks referred to as Overlapping Convolutional Arithmetic Circuits, or Overlapping ConvACs for short. This class shares the same architectural features as standard ConvNets, including some that have previously been overlooked by
similar attempts to model ConvNets through ConvACs, namely, having any number of layers and
unrestricted receptive fields and strides, which are crucial for studying overlapping architectures.
For simplicity, we will describe this model only for the case of inputs with two spatial dimensions,
e.g. color images, and limiting the convolutional filters to the shape of a square.
We begin by presenting a broad definition of
a Generalized Convolutional (GC) layer as a
fusion of a 1×1 linear operation with a pooling function – this view of convolutional layers
is motivated by the all-convolutional architecture (Springenberg et al., 2015), which replaces
all pooling layers with convolutions with stride
previous layer
H(in)
R
H(out)
(R)
x(1) · · · x
.. . .
.
. ..
.
x(R
R
H(in)
2
GC layer
2
D(out)
H (out) = H (in)
S
)
D(in)
2
2
2
y = g(W (1) x(1) + b(1) , . . . , W (R ) x(R ) + b(R ) )
Figure 1: An illustration of a GC Layer.
Published as a conference paper at ICLR 2018
input
representa(on hidden layer 1
GC
hidden
layer L-2
hidden
layer L-1
GC
R(1),S(1)
xi
GC
(output)
GC
R(L-2),S(L-2)
rep(i, d) = f✓d (xi )
M
D(1)
D(L-2)
D(L-1)
D(L)
hy (x1 , . . . , xN )
Figure 2: An illustration of a Generalized Convolutional Network.
greater than 1. The input to a GC layer is a 3order tensor (multi-dimensional array), having
width and height equal to H (in) ∈ N and depth
D(in) ∈ N, also referred to as channels, e.g. the
input could be a 2D image with RGB color channels. Similarly, the output of the layer has width
(in)
and height equal to H (out) ∈ N and D(out) ∈ N channels, where H (out) = HS for S ∈ N that is
referred to as the stride, and has the role of a sub-sampling operation. Each spatial location (i, j) at
the output of the layer corresponds to a 2D window slice of the input tensor of size R × R × D(in) ,
extended through all the input channels, whose top-left corner is located exactly at (i·S, j ·S), where
R ∈ N is referred to as its local receptive field, or filter size. For simplicity, the parts of window
((out)
slices extending beyond the boundaries have zero value. Let y ∈ RD
be a vector representing
(in)
(1)
(R2 )
the channels at some location of the output, and similarly, let x , . . . , x
∈ RD be the set of
vectors representing the slice, where each vector represents the channels at its respective location
inside the R × R window, then the operation of a GC layer is defined as follows:
2
2
2
y = g(W (1) x(1) + b(1) , . . . , W (R ) x(R ) + b(R ) ),
2
(out)
(in)
2
(out)
×D
where W (1) , . . . , W (R ) ∈ RD
and b(1) , . . . , b(R ) ∈ RD
are referred to as the
(out)
(out)
(out)
weights and biases of the layer, respectively, and g : RD
× · · · × RD
→ RD
is some
point-wise pooling function. See fig. 1 for an illustration of the operation a GC layer performs.
With the above definitions, a GC network is simply a sequence of L GC layers, where for l ∈
[L] ≡ {1, . . . , L}, the l’th layer is specified by a local receptive field R(l) , a stride S (i) , D(l) output
channels, parameters θ(l) , and a pooling function g (l) . For classification tasks, the output of the last
layer of the network typically has 1×1 spatial dimensions, i.e. a vector, where each output channel
y ∈ [Y ] ≡ [D(L) ] represents the score function of the y’th class, denoted by hy , and inference is
perform by y ∗ = arg maxy hy (X). Oftentimes, it is common to consider the output of the very
first layer of a network as a low-level feature representation of the input, which is motivated by
the observation that these learned features are typically shared across different tasks and datasets
over the same domain (e.g. edge and Gabor filters for natural images). Hence, we treat this layer
as a separate fixed “zeroth” convolutional layer referred to as the representation layer, where the
operation of the layer can be depicted as applying a set of fixed functions {fd : Rs → R}M
d=1 to the
window slices denoted by x1 , . . . , xN ∈ Rs , i.e. the entries of the output tensor of this layer are
given by {fd (xi )}d∈[M ],i∈[N ] . With these notations, the output of a GC network can be viewed as a
function hy (x1 , . . . , xN ). The entire GC network is illustrated in fig. 2.
Given a non-linear point-wise activation function σ(·) (e.g. ReLU), then settingall poolingfunctions
PR2 (i)
2
for c ∈
to average pooling followed by the activation, i.e. g(x(1) , . . . , x(R ) )c = σ
i=1 xc
[D(out) ], give rise to the common all-convolutional network with σ(·) activations, which served
as the initial motivation for our formulation. Alternatively, choosing instead a product pooling
QR2 (i)
2
function, i.e. g(x(1) , . . . , x(R ) )c = i=1 xc for c ∈ [D(out) ], results in an Arithmetic Circuit,
i.e. a circuit containing just product and sum operations, hence it is referred to as a Convolutional
Arithmetic Circuit, or ConvAC. It is important to emphasize that ConvACs, as originally introduced
by Cohen et al. (2016a), are typically described in a very different manner, through the language
of tensor decompositions (see app. A for background). Since vanilla ConvACs can be seen as an
alternating sequence of 1×1 convolutions and non-overlapping product pooling layers, then the two
formulations coincide when all GC layers are non-overlapping, i.e. for all l ∈ [L], R(l) = S (l) . If,
however, some of the layers are overlapping, i.e. there exists l ∈ [L] such that R(l) > S (l) , then our
formulation through GC layers diverges, and give rise to what we call Overlapping ConvACs.
3
Published as a conference paper at ICLR 2018
Given that our model is an extension of the ConvACs framework, it inherits many of its desirable
attributes. First, it shares most of the same traits as modern ConvNets, i.e. locality, sharing and pooling. Second, it can be shown to form a universal hypotheses space (Cohen et al., 2016a). Third, its
underlying operations lend themselves to mathematical analysis based on measure theory and tensor
analysis (Cohen et al., 2016a). Forth, through the concept of generalized tensor decompositions (Cohen and Shashua, 2016), many of the theoretical results proven on ConvACs could be transferred
to standard ConvNets with ReLU activations. Finally, from an empirical perspective, they tend to
work well in many practical settings, e.g. for optimal classification with missing data (Sharir et al.,
2016), and for compressed networks (Cohen et al., 2016b).
While we have just established that the non-overlapping GC Network with a product pooling function is equivalent to vanilla ConvACs, one might wonder if using overlapping layers instead could
diminish what these overlapping networks can represent. We show that not only is it not the case,
but prove the more general claim that a network of a given architecture can realize exactly the same
functions as networks using smaller local receptive fields, which includes the non-overlapping case.
Proposition 1. Let A and B be two GC Networks with a product pooling function. If the architecture
of B can be derived from A through the removal of layers with 1×1 stride, or by decreasing the local
receptive field of some of its layers, then for any choice of parameters for B, there exists a matching
set of parameters for A, such that the function realized by B is exactly equivalent to A. Specifically,
A can realize any non-overlapping network with the same order of strides (excluding 1×1 strides).
Proof sketch. This follows from two simple claims: (i) a GC layer can produce an output equivalent
to that of a GC layer with a smaller local receptive field, by “zeroing” its weights beyond the smaller
local receptive field; and (ii) GC layers with 1×1 receptive fields can be set such that their output is
equal to their input, i.e. realize the identity function. With these claims, the local receptive fields of
A can be effectively shrank to match the local receptive fields of B, and any additional layers of A
with stride 1×1 could be set such that they are realizing the identity mapping, effectively “removing”
them from A. See app. C.2 for a complete proof.
Proposition 1 essentially means that overlapping architectures are just as expressive as nonoverlapping ones of similar structure, i.e. same order of non-unit strides. As we recall, this satisfies the first condition of the efficiency property introduced in sec. 1, and does so regardless if we
measure the size of a network as the number of parameters, or the number of “neurons”1 . In the following section we will cover the preliminaries required to show that overlapping networks actually
lead to an increase in expressive capacity, which under some settings results in an exponential gain,
proving that the second condition of expressive efficiency holds as well.
3
A NALYZING E XPRESSIVE E FFICIENCY T HROUGH G RID T ENSORS
In this section we describe our methods for analyzing the expressive efficiency of overlapping ConvACs that lay the foundation for stating our theorems. A minimal background on tensor analysis
required to follow our work can be found in sec. 3.1, followed by presenting our methods in sec. 3.2.
3.1
P RELIMINARIES
In this sub-section we cover the minimal background on tensors analysis required to understand
our analysis. A tensor A ∈ RM1 ⊗···⊗MN of order N and dimension Mi in each mode i ∈ [N ] ≡
{1, . . . , N }, is a multi-dimensional array with entries Ad1 ,...,dN ∈ R for all i ∈ [N ] and di ∈ [Mi ].
For simplicity, henceforth we assume that all dimensions are equal, i.e. M ≡ M1 = . . . = MN .
One of the central concepts in tensor analysis is that of tensor matricization, i.e. rearranging its
entries to the shape of a matrix. Let P ∪· Q = [N ] be a disjoint partition of its indices, such that
P = {p1 , . . . , p|P | } with p1 < . . . < p|P | , and Q = {q1 , . . . , q|Q| } with q1 < . . . < q|Q| . The
matricization of A with respect to the partition P ∪· Q, denoted by JAKP,Q , is the M |P | -by-M |Q|
matrix holding the entries of A, such that for all i ∈ [N ] and di ∈ [M ] the entry Ad1 ,...,dN is placed
1
We take here the broader definition of a “neuron”, as any one of the scalar values comprising the output
array of an arbitrary layer in a network. In the case the output array is of width and height equal to H and C
channels, then the number of such “neurons” for that layer is H 2 · C.
4
Published as a conference paper at ICLR 2018
P|P |
P|Q|
in row index 1 + t=1 (dpt − 1)M |P |−t and column index 1 + t=1 (dqt − 1)M |Q|−t . Lastly, the
tensors we study in this article originate by examining the values of some given function at a set
of predefined points and arranging them in a tensor referred to as the grid tensor of the function.
Formally, let f : Rs ×. . .×Rs → R be a function, and let {x(1) , . . . , x(M ) ∈ Rs } be a set of vectors
called template vectors, then the grid tensor of f is denoted by A(f ) ∈ RM ⊗...⊗M and defined by
A(f )d1 ,...,dN = f (x(d1 ) , . . . , x(dN ) ) for all d1 , . . . , dN ∈ [M ].
3.2
B OUNDING THE S IZE OF N ETWORKS VIA G RID T ENSORS
We begin with a discussion on how to have a well-defined measure of efficiency. We wish to compare the efficiency of non-overlapping ConvACs to overlapping ConvACs, for a fixed set of M representation functions (see sec. 2 for definitions). While all functions realizable by non-overlapping
ConvACs with shared representation functions lay in the same function subspace (see Cohen et al.
(2016a)), this is not the case for overlapping ConvACs, which can realize additional functions outside the sub-space induced by non-overlapping ConvACs. We cannot therefore compare both architectures directly, and need to compare them through an auxiliary objective. Following the work
of Cohen and Shashua (2016), we instead compare architectures through the concept of grid tensors, and specifically, the grid tensor defined by the output of a ConvAC, i.e. the tensor A(h) for
h(x1 , . . . , xN ). Unlike with the ill-defined nature of directly comparing the functions of realized
by ConvACs, Cohen and Shashua (2016) proved that assuming the fixed representation functions
are linearly independent, then there exists template vectors x(1) , . . . , x(M ) , for which any nonoverlapping ConvAC architecture could represent all possible grid tensors over these templates,
given sufficient number of channels at each layer. More specifically, if Fij = fi (x(j) ), then these
template vector are chosen such that F is non-singular. Thus, once we fix a set of linearly independent representation functions, we can compare different ConvACs, whether overlapping or not,
on the minimal size required for them to induce the same grid tensor, while knowing such a finite
number always exists.
One straightforward direction for separating between the expressive efficiency of two network architectures A and B is by examining the ranks of their respective matricized grid tensors. Specifically,
Let A(h(A) ) and A(h(B) ) denote the grid tensors of A and B, respectively, and let (P, Q) be a partition of [N ], then we wish to find an upper-bound on the rank of JA(h(A) )KP,Q
as a function of its
size on one hand, while showing on the other hand that rank JA(h(B) )KP,Q can be significantly
greater. One benefit of studying efficiency through a matrix rank is that not only we attain separation bounds for exact realization, but also immediately gain access to approximation bounds by
examining the singular values of the matricized grid tensors. This brings us to the following lemma,
which connects upper-bounds that were previously found for non-overlapping ConvACs (Cohen and
Shashua, 2017), with the grid tensors induced by them (see app. C.1 for proof):
Lemma 1. Let hy (x1 , . . . , xN ) be a score function of a non-overlapping ConvAC with a fixed set of
M linearly independent and continuous representation functions, and L GC layers. Let (P, Q) be
a partition dividing the spatial dimensions of the output of the representation layer into two equal
parts, either along the horizontal or vertical axis, referred to as the “left-right” and “top-bottom”
partitions, respectively. Then, for any template vectors such that F is non-singular and for any
choice of the parameters of the network, it holds that rank (JA(hy )KP,Q ) ≤ D(L−1) .
Lemma 1 essentially means that it is sufficient to show that overlapping ConvACs can attain
ranks super-polynomial in their size to prove they are exponentially efficient with respect to nonoverlapping ConvACs. In the next section we analyze how the overlapping degree is related to the
rank, and under what cases it leads to an exponentially large rank.
4
T HE E XPRESSIVE E FFICIENCY OF OVERLAPPING A RCHITECTURES
In this section we analyze the expressive efficiency of overlapping architectures. We begin by defining our measures of the overlapping degree that will used in our claims, followed by presenting our
main results in sec. 4.2. For the sake of brevity, an additional set of results, in light of the recent
work by Cohen and Shashua (2017) on “Pooling Geometry”, is deferred to app. B.
5
Published as a conference paper at ICLR 2018
4.1
T HE OVERLAPPING D EGREE OF A N ETWORK
To analyze the efficiency of overlapping archiInput Layer
tectures, we will first formulate more rigorously
the measurement of the overlapping degree of
Total
a given architecture. As mentioned in sec. 1,
Layer L-1
Stride
we do so by defining the concepts of the total receptive field and total stride of a given
(l)
(l)
Layer L
layer l ∈ [L], denoted by TR and TS , reTotal
spectively. Both measurements could simply be Receptive
Field
thought of as projecting the accumulated local
receptive fields (or strides) to the the first layer,
as illustrated in fig. 3, which represent a type of Figure 3: Illustrating the total receptive field and
global statistics of the architecture. However, total stride attributes for the L’th layer, which
note that proposition 1 entails that a given ar- could be seen as the projected receptive field and
chitecture could have a smaller effective total stride with respect to the input layer. Together,
receptive field, for some settings of its parame- they capture the overlapping degree of a network.
ters. This leads us to define the α-minimal total
receptive field, for any α ∈ R+ , as the smallest effective total receptive field still larger than α,
(l,α)
which we denote by TR . The exact definitions of the above concepts are formulated as follows:
(Q
l
(i)
l≥1
(l)
(l)
(1)
(l)
i=1 S
TS ≡ TS (S , . . . , S ) ≡
(1)
1
l=0
Xl−1
(l)
(l)
(l−1)
(k−1)
TR ≡ TR (R(1) , S (1) , . . . , R(l) , S (l) ) ≡ R(l) · TS
+
R(k) − S (k) · TS
(2)
(l,α)
TR
≡
(l,α)
TR (R(1) , S (1) , . . . , R(l) , S (l) )
(l−1)
where we omitted the arguments of TS
≡
argmin
k=1
(l)
TR (t1 , S (1) , . . . , tl , S (l) )
(3)
∀i∈[l],S (i) ≤ti ≤R(i)
(l)
TR (t1 ,S (1) ,...,tl ,S (l) )>α
(k−1)
and TS
for the sake of visual compactness.
Notice that for non-overlapping networks the total receptive field always equals the total stride, and
that only at the end of the network, after the spatial dimension collapses to 1×1, does the the total
receptive field grow to encompass the entire size of the representation layer. For overlapping networks this is not the case, and the total receptive field could grow much faster. Intuitively, this means
that values in regions of the input layer that are far apart would be combined by non-overlapping
networks only near the last layers of such networks, and thus non-overlapping networks are effectively shallow in comparison to overlapping networks. Base on this intuition, in the next section we
analyze networks with respect to the point at which their total receptive field is large enough.
4.2
M AIN R ESULTS
With all the preliminaries in place, we are ready to present our main result:
Theorem 1. Assume a ConvAC with a fixed representation layer having M output channels and both
width and height equal to H, followed by L GC layers, where the l’th layer has a local receptive
field R(l) , a stride S (l) , and D(l) output channels. Let K ∈ [L] be a layer with a total receptive
(K)
(K)
(K)
field TR ≡ TR (R(1) , S (1) , . . . , R(K) , S (K) ), such that TR > H2 . Then, for any choice of
parameters, except a null set (with respect to the Lebesgue measure), and for any template vectors
such that F is non-singular, the following equality holds:
rank (JA(hy )KP,Q ) ≥ D
&
H
H−T (K,b /2c)
R
+1·
T
where (P, Q) is either the “left-right”
D ≡ min{M, D(K) , 12 min1≤l≤K D(l) }.
(K)
S
or
the
'
H
(K)
T
S
“top-bottom”
(4)
partitions
and
Proof sketch. Because the entries of the matricized grid tensors are polynomials in the parameters,
then according to a lemma by Sharir et al. (2016), if there is a single example that attains the above
6
Published as a conference paper at ICLR 2018
input
H
representa+on
>
H
2
Arbitrary Layers
Conv
Conv
stride
1x1
>
H
Conv
H
2
M
Output
M
Figure 4: A network architectures beginning with large local receptive fields greater than N/2 and at
least M output channels. According to theorem 1, for almost all choice of parameters we obtain a
function that cannot be approximated by a non-overlapping architecture, if the number of channels
H2
in its next to last layer is less than M 2 .
lower-bound on the rank, then it occurs almost everywhere with respect to the Lebesgue measure on
the Euclidean space of the parameters.
Given the last remark, the central part of our proof is simply the construction of such an example.
First we find a set of parameters for the simpler case where the first GC layer is greater than a
quarter of the input, satisfying the conditions of the theorem. The motivation behind the specific
construction is the pairing of indices from each side of the partition, such that they are both in the
same local receptive field, and designing the filters such that the output of each local application of
them defines a mostly diagonal matrix of rank D, with respect to these two indices. The rest of the
parameters are chosen such that the output of the entire network results in a product of the entries
of these matrices. Under matricization, this results in a matrix who is equivalent2 to a Kronecker
product of mostly diagonal matrices. Thus, the matricization rank is equal to the product of the
ranks of these matrices, which results in the exponential form of eq. 4. Finally, we extend the above
example to the general case, by realizing the operation of the first layer of the above example through
multiple layers with small local receptive fields. See app. C.1 for the definitions and lemmas we rely
on, and see app. C.3 for a complete proof.
Combined with Lemma 1, it results in the following corollary:
Corollary 1. Under the same setting as theorem 1, and for all choices of parameters of an overlapping ConvAC, except a negligible set, any non-overlapping ConvAC that realizes (or approximates)
the same grid tensor must be of size at least:
D
&
H
H−T (K,b /2c)
R
+1·
T
(K)
S
'
H
(K)
T
S
.
While the complexity of the generic lower-bound above might seem incomprehensible at first, its
generality gives us the tools to analyze practically any kind of feed-forward architecture. As an
example, we can analyze the lower bound for the well known GoogLeNet architecture (Szegedy
et al., 2015), for which the lower bound equals 3298 , making it clear that using a non-overlapping
architecture for this case is infeasible. Next, we will focus on specific cases for which we can derive
more intelligible lower bounds.
According to theorem 1, the lower bound depends on the first layer for which its total receptive field
is greater than a quarter of the input. As mentioned in the previous section, for non-overlapping
networks this only happens after the spatial dimension collapses to 1×1, which entails that both
the total receptive field and total stride would be equal to the width H of the representation layer,
and substituting this values in eq. 4 results simply in D – trivially meaning that to realize one nonoverlapping network by another non-overlapping network, the next to last layer must have at least
half the channels of the target network.
On the other extreme, we can examine the case where the first GC layer has a local receptive field R
greater than a quarter of its input, i.e. R > H/2. Since the layers following the first GC layer do not
affect the lower bound in this case, it applies to any arbitrary sequence of layers as illustrated in fig. 4.
For simplicity we will also assume that the stride S is less than H/2, and that H2 is evenly divided
2
Two matrices are equivalent if one could be converted to the other by elementary row or column operations.
7
Published as a conference paper at ICLR 2018
input
representa.on
Block 0
BxB Conv
Block L-1
2x2 Conv
BxB Conv
R: BxB
S: 1x1
H
2x2 Conv
output
R: 2x2
S: 2x2
D(1)
M
H
D(1)
D(L-1)
D(L-1)
Figure 5: The common network architecture of alternating B×B “conv” and 2×2 “pooling” layers.
If B ≤ H/5+1 and D(l) ≥ 2M for all 1 ≤ l < L, then the lower bound of theorem 1 for this
(2B−1)2
4
network results in M
by S. In this case the
H2
2S
.
H
2 -minimal
receptive field equals to
H
2
+ 1, and thus the lower bound results
in D . Consider the case of D = M and S = 1, then a non-overlapping architecture that satisfies
this lower bound is of the order of magnitude at which it could already represent any possible grid
tensor. This demonstrate our point from the introduction, that through a a polynomial change in the
architecture, i.e. increasing the receptive field, we get an exponential increase in expressivity.
Though the last example already demonstrates that a polynomially sized overlapping architecture
could lead to an exponential separation, in practice, employing such large convolutions is very resource intensive. The common best practice is to use multiple small local receptive fields of size
B × B, where the typical values are B = 3 or B = 5, separated by a 2 × 2 “pooling” layers,
i.e. layers with both stride and local receptive field equal to 2 × 2. For simplicity, we assume that
H = 2L for some L ∈ N. See fig. 5 for an illustration of such a network. Analyzing the above
network with theorem 1 results in the following proposition:
Proposition 2. Consider a network comprising a sequence of GC blocks, each block begins with a
layer whose local receptive field is B×B and its stride 1×1, followed by a layer with local receptive
field 2×2 and stride 2×2, where the output channels of all layers are at least 2M , and the spatial
dimension of the representation layer is H×H for H=2L . Then, the lower bound describe by eq. 4
for the above network is greater than or equal to:
τ (B, H) ≡ M
(2B−1)2
2
whose limits are limB→∞ τ (B, H) = M
ing B ≤
H
5
+ 1, then τ (B, H) ≥ M
·(1+ 2B−2
H )
H2
2
(2B−1)2
4
−2
=M
H2
2
−2
H−1
·(1+ 2B−1
)
and limH→∞ τ (B, H) = M
,
(2B−1)2
2
. Finally, assum-
.
Proof sketch. We first find a closed-form expression for the total receptive field and stride of each
of the B×B layers in the given network. We then show that for layers whose total receptive field
H
H
is greater than H
2 , its α-minimal total receptive field, for α= 2 , is equal to 2 +1. We then use the
above to find the first layer who satisfies the conditions of theorem 1, and then use our closed-forms
expressions to simplify the general lower bound for this case. See app. C.4 for a complete proof.
In particular, for the typical values of M = 64, B = 5, and H ≥ 20, the lower bound is at
least 6420 , which demonstrates that even having a small amount of overlapping already leads to an
exponential separation from the non-overlapping case. When B grows in size, this bound approaches
the earlier result we have shown for large local receptive fields encompassing more than a quarter
of the image. When H grows in size, the lower bound is dominated strictly by the local receptive
fields. Also notice that based on proposition 2, we could also derive a respective lower bound
for a network following VGG style architecture (Simonyan and Zisserman, 2014), where instead
of a single convolutional layer before every “pooling” layer, we have K layers, each with a local
receptive field of C × C. Under this case, it is trivial to show that the bound from proposition 2
holds for B = K · (C − 1) + 1, and under the typical values of C = 3 and K = 2 it once again
results in a lower bound of at least 6420 .
8
100.00
93.33
86.67
80.00
73.33
66.67
60.00
53.33
46.67
40.00
B=1
B=2
B=3
B=4
B=5
16
32
64
128
256
Number of Channels
512
1024
32
64
128
256
Number of Channels
512
1024
100.00
93.33
86.67
80.00
73.33
B=1
66.67
B=2
60.00
B=3
53.33
B=4
46.67
B=5
40.00
1.0e+03 4.1e+03 1.6e+04 6.6e+04 2.6e+05 1.0e+06 4.2e+06 1.7e+07
100.00
93.33
86.67
80.00
73.33
B=1
66.67
B=2
60.00
B=3
53.33
B=4
46.67
B=5
40.00
1.0e+03 4.1e+03 1.6e+04 6.6e+04 2.6e+05 1.0e+06 4.2e+06 1.7e+07 6.7e+07
2048
B=1
B=2
B=3
B=4
B=5
16
Train Accuracy (%)
100.00
93.33
86.67
80.00
73.33
66.67
60.00
53.33
46.67
40.00
Train Accuracy (%)
Train Accuracy (%)
Train Accuracy (%)
Published as a conference paper at ICLR 2018
Number of Parameters
2048
Number of Parameters
Figure 6: Training accuracies of standard ConvNets on CIFAR-10 with data augmentations, where
the results of spatial augmentations presented at the top row, and color augmentations at the bottom
row. Each network follows the architecture of proposition 2, with with receptive field B and using
the same number of channels across all layers, as specified by the horizontal axis of left plot. We
plot the same results with respect to the total number of parameters in the right plot.
5
E XPERIMENTS
In this section we show that the theoretical results of sec. 4.2 indeed hold in practice. In other
words, there exists tasks that require the highly expressive power of overlapping architectures, on
which non-overlapping architectures would have to grow by an exponential factor to achieve the
same level of performance. We demonstrate this phenomenon on standard ConvNets with ReLU
activations that follow the same architecture that was outlined in proposition 2, while varying the
number of channels and the size of the receptive field of the B×B “conv” layers. The only change
we made, was to replace the 2×2-“pooling” layers of the convolutional type, with the standard 2×2max-pooling layers, and using the same number of channels across all layers. This was done for the
purpose of having all the learned parameters located only at the (possibly) overlapping layers. More
specifically, the network has 5 blocks, each starting with a B×B convolution with C channels,
stride 1×1, and ReLU activation, and then followed by 2×2 max-pooling layer. After the fifth
“conv-pool”, there is a final dense layer with 10 outputs and softmax activations.
We train each of these networks for classification over the CIFAR-10 dataset, with two types of
data augmentation schemes: (i) spatial augmentations, i.e. randomly translating (up to 3 pixels in
each direction) and horizontally flipping each image, and (ii) color augmentations following Dosovitskiy et al. (2014), i.e. randomly adding a constant shift (at most ±0.3) to the hue, saturation,
and luminance, for each attribute separately, and in addition randomly sampling a multiplier (in
the range [0.5, 1.5]) just to the saturation and luminance. Though typically data augmentation is
only used for the purpose of regularization, we employ it for the sole purpose of raising the hardness of the regular CIFAR-10 dataset, as even small networks can already overfit and effectively
memorize its small dataset. We separately test both the spatial and color augmentation schemes
to emphasize that our empirical results cannot be explained simply by spatial-invariance type arguments. Finally, the training itself is carried out for 300 epochs with ADAM (Kingma and Ba,
2015) using its standard hyper-parameters, at which point the loss of the considered networks have
stopped decreasing. We report the training accuracy over the augmented dataset in fig. 6, where for
each value of the receptive field B, we plot its respective training accuracies for variable number
of channels C. The source code for reproducing the above experiments and plots can be found at
https://github.com/HUJI-Deep/OverlapsAndExpressiveness.
It is quite apparent that the greater B is chosen, the less channels are required to achieve the same
accuracy. Moreover, for the non-overlapping case of B=1, more than 2048 channels are required
to reach the same performance of networks with B>2 and just 64 channels under the spatial aug9
Published as a conference paper at ICLR 2018
mentations – which means effectively exponentially more channels were required. Even more so,
under the color augmentations, we were not able to train non-overlapping networks to reach even the
smallest overlapping network (B = 2 and C = 16). In terms of total number of parameters, there
is a clear separation between the overlapping and the non-overlapping types, and we once again
see more than an order of magnitude increase in the number of parameters between an overlapping
and non-overlapping architectures that achieve similar training accuracy. As a somewhat surprising
result, though based only on our limited experiments, it appears that for the same number of parameters, all overlapping networks attain about the same training accuracy, suggesting perhaps that
having the smallest amount of overlapping already attain all the benefits overlapping provides, and
that increasing it further does not affect the performance in terms of expressivity.
As final remark, we also wish to acknowledge the limitations of drawing conclusions strictly from
empirical experiments, as there could be alternative explanations to these observations, e.g. the
effects overlapping has on the optimization process. Nevertheless, our theoretical results suggests
this is less likely the case.
6
D ISCUSSION
The common belief amongst deep learning researchers has been that depth is one of the key factors in the success of deep networks – a belief formalized through the depth efficiency conjecture.
Nevertheless, depth is one of many attributes specifying the architecture of deep networks, and each
could potentially be just as important. In this paper, we studied the effect overlapping receptive
fields have on the expressivity of the network, and found that having them, and more broadly denser
connectivity, results in an exponential gain in the expressivity that is orthogonal to the depth.
Our analysis sheds light on many trends and practices in contemporary design of neural networks.
Previous studies have shown that non-overlapping architectures are already universal (Cohen et al.,
2016a), and even have certain advantages in terms of optimization (Brutzkus and Globerson, 2017),
and yet, real-world usage of non-overlapping networks is scarce. Though there could be multiple
factors involved, our results clearly suggest that the main culprit is that non-overlapping networks
are significantly handicapped in terms of expressivity compared to overlapping ones, explaining why
the former are so rarely used. Additionally, when examining the networks that are commonly used
in practice, where the majority of the layers are of the convolutional type with very small receptive
field, and only few if any fully-connected layers (Simonyan and Zisserman, 2014; Springenberg
et al., 2015; He et al., 2016), we find that though they are obviously overlapping, their overlapping
degree is rather low. We showed that while denser connectivity can increase the expressive capacity, even in the most common types of modern architectures already exhibit exponential increase in
expressivity, without relying on fully-connected layers. This could partly explain that somewhat surprising observation, as it is probable that such networks are sufficiently expressive for most practical
needs simply because they are already in the exponential regime of expressivity. Indeed, our experiments seems to suggests the same, in which we saw that further increases in the overlapping degree
beyond the most limited overlapping case seems to have insignificant effects on performance – a
conjecture not quite proven by our current work, but one we wish to investigate in the future.
There are relatively few other works which have studied the role of receptive fields in neural networks. Several empirical works (Li and Perona, 2005; Coates et al., 2011; Krizhevsky et al., 2012)
have demonstrated similar behavior, showing that the classification accuracy of networks can sharply
decline as the degree of overlaps is decreased, while also showing that gains from using very large
local receptive fields are insignificant compared to the increase in computational resources. Other
works studying the receptive fields of neural networks have mainly focused on how to learn them
from the data (Coates and Ng, 2011; Jia et al., 2012). While our analysis has no direct implications to those specific works, it does lay the ground work for potentially guiding architecture design,
through quantifying the expressivity of any given architecture. Lastly, Luo et al. (2016) studied
the effective total receptive field of different layers, a property of a similar nature to our total receptive field, where they measure the the degree to which each input pixel is affecting the output
of each activation. They show that under common random initialization of the weights, the effective total receptive field has a gaussian shape and is much smaller than the maximal total receptive
field. They additionally demonstrate that during training the effective total receptive field grows in
size, and suggests that weights should be initialized such that the initial effective receptive field is
10
Published as a conference paper at ICLR 2018
large. Their results strengthen our theory, by showing that trained networks tend to maximize their
effective receptive field, taking full potential of their expressive capacity.
To conclude, we have shown both theoretically and empirically that overlapping architectures have
an expressive advantage compared to non-overlapping ones. Our theoretical analysis is grounded on
the framework of ConvACs, which we extend to overlapping configurations. Though are proofs are
limited to this specific case, previous studies (Cohen and Shashua, 2016) have already shown that
such results could be transferred to standard ConvNets as well, using most of the same mathematical
machinery. While adapting our analysis accordingly is left for future work, our experiments on
standard ConvNets (see sec. 5) already suggest that the core of our results should hold in this case as
well. Finally, an interesting outcome of moving from non-overlapping architectures to overlapping
ones is that the depth of a network is no longer capped at log2 (input size), as has been the case in
the models investigated by Cohen et al. (2016a) – a property we will examine in future works
ACKNOWLEDGMENTS
This work is supported by Intel grant ICRI-CI #9-2012-6133, by ISF Center grant 1790/12 and by
the European Research Council (TheoryDL project).
R EFERENCES
Alon Brutzkus and Amir Globerson. Globally Optimal Gradient Descent for a ConvNet with Gaussian Inputs.
In International Conference on Machine Learning ICML, February 2017.
Adam Coates and Andrew Y Ng. Selecting Receptive Fields in Deep Networks. Advances in Neural Information Processing Systems, 2011.
Adam Coates, Andrew Y Ng, and Honglak Lee. An Analysis of Single-Layer Networks in Unsupervised
Feature Learning. International Conference on Artificial Intelligence and Statistics, pages 215–223, 2011.
Nadav Cohen and Amnon Shashua. Convolutional Rectifier Networks as Generalized Tensor Decompositions.
In International Conference on Machine Learning, May 2016.
Nadav Cohen and Amnon Shashua. Inductive Bias of Deep Convolutional Networks through Pooling Geometry. In International Conference on Learning Representations ICLR, April 2017.
Nadav Cohen, Or Sharir, and Amnon Shashua. On the Expressive Power of Deep Learning: A Tensor Analysis.
In Conference on Learning Theory COLT, May 2016a.
Nadav Cohen, Or Sharir, and Amnon Shashua. Deep SimNets. In Computer Vision and Pattern Recognition
CVPR, May 2016b.
Alexey Dosovitskiy, Jost Tobias Springenberg, Martin A Riedmiller, and Thomas Brox. Discriminative Unsupervised Feature Learning with Convolutional Neural Networks. In Advances in Neural Information Processing Systems NIPS, Deep Learning Workshop, 2014.
Ronen Eldan and Ohad Shamir. The Power of Depth for Feedforward Neural Networks. In Conference on
Learning Theory COLT, May 2016.
Wolfgang Hackbusch. Tensor Spaces and Numerical Tensor Calculus, volume 42 of Springer Series in Computational Mathematics. Springer Science & Business Media, Berlin, Heidelberg, February 2012.
András Hajnal, Wolfgang Maass, Pavel Pudlák, Mario Szegedy, and György Turán. Threshold Circuits of
Bounded Depth. J. Comput. Syst. Sci. (), 46(2):129–154, 1993.
Johan Håstad and Mikael Goldmann. On the Power of Small-Depth Threshold Circuits. Computational Complexity (), 1(2):113–129, 1991.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition.
Computer Vision and Pattern Recognition, pages 770–778, 2016.
Yangqing Jia, Chang Huang, and Trevor Darrell. Beyond spatial pyramids - Receptive field learning for pooled
image features. Computer Vision and Pattern Recognition, 2012.
Diederik Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In International Conference
on Learning Representations, 2015.
11
Published as a conference paper at ICLR 2018
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification with Deep Convolutional
Neural Networks. Advances in Neural Information Processing Systems, pages 1106–1114, 2012.
Fei-Fei Li and Pietro Perona. A Bayesian Hierarchical Model for Learning Natural Scene Categories. Computer
Vision and Pattern Recognition, 2:524–531, 2005.
Min Lin, Qiang Chen, and Shuicheng Yan. Network In Network. In International Conference on Learning
Representations ICLR, 2014.
Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard S Zemel. Understanding the Effective Receptive Field in
Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems, pages 4898–
4906, 2016.
Guido F Montúfar, Razvan Pascanu, Kyunghyun Cho, and Yoshua Bengio. On the Number of Linear Regions
of Deep Neural Networks. CoRR abs/1202.2745, stat.ML:2924–2932, 2014.
Razvan Pascanu, Guido Montufar, and Yoshua Bengio. On the number of response regions of deep feed forward
networks with piece-wise linear activations. arXiv.org, December 2013.
Ben Poole, Subhaneil Lahiri, Maithreyi Raghu, Jascha Sohl-Dickstein, and Surya Ganguli. Exponential expressivity in deep neural networks through transient chaos. In Advances in Neural Information Processing
Systems NIPS, Deep Learning Workshop, 2016.
Maithra Raghu, Ben Poole, Jon Kleinberg, Surya Ganguli, and Jascha Sohl-Dickstein. On the expressive power
of deep neural networks. arXiv.org, June 2016.
Itay Safran and Ohad Shamir. Depth Separation in ReLU Networks for Approximating Smooth Non-Linear
Functions. CoRR abs/1202.2745, cs.LG, 2016.
Or Sharir, Ronen Tamari, Nadav Cohen, and Amnon Shashua. Tensorial Mixture Models. arXiv.org, October
2016.
Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. CoRR abs/1202.2745, cs.CV, 2014.
Michael Sipser. Borel sets and circuit complexity. ACM, New York, New York, USA, December 1983.
Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for Simplicity:
The All Convolutional Net. In Workshop track of International Conference on Learning Representations,
2015.
Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan,
Vincent Vanhoucke, and Andrew Rabinovich. Going Deeper with Convolutions. In Computer Vision and
Pattern Recognition CVPR, 2015.
Yaniv Taigman, Ming Yang, Marc’Aurelio Ranzato, and Lior Wolf. DeepFace: Closing the Gap to HumanLevel Performance in Face Verification. In Computer Vision and Pattern Recognition CVPR. IEEE Computer Society, June 2014.
Matus Telgarsky. Benefits of depth in neural networks. In Conference on Learning Theory COLT, May 2016.
Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. CoRR
abs/1609.03499, 2016.
Andrew Chi-Chih Yao. Circuits and Local Computation. STOC, pages 186–196, 1989.
12
Published as a conference paper at ICLR 2018
input X
hidden layer 0
representa.on
1x1 conv
hidden layer L-1
pooling
1x1 conv
pooling
dense
(output)
rL-1
Y
xi
r0
M
rep(i, d) = f✓d (xi )
pool0 (j, ) =
⌦
↵
conv(j, ) = a0,j, , rep(j, :)
Y
rL-1
r0
conv0 (j 0 , )
poolL
1 ( )=
Y
convL
0
1 (j , )
j 0 covers space
j 0 2window
⌦
out(y) = aL,y , poolL
1 (:)
↵
Figure 7: The original Convolutional Arithmetic Circuits as presented by Cohen et al. (2016a).
A
BACKGROUND ON C ONVOLUTIONAL A RITHMETIC C IRCUITS
We base our analysis on the convolutional arithmetic circuit (ConvAC) architecture introduced by Cohen et al.
(2016a), which is illustrated by fig. 7, and can be simply thought of as a regular ConvNet, but with linear activations and product pooling layers, instead of the more common non-linear activations (e.g. ReLU)
and average/max pooling. More specifically, each point in the input space of the network, denoted by
X = (x1 , . . . , xN ), is represented as an N -length sequence of s-dimensional vectors x1 , . . . , xN ∈ Rs .
X is typically thought of as an image, where each xi corresponds to a local patches from that image. The first
layer of the network is referred to as the representation layer, consisting of applying M representation functions
fθ1 , . . . , fθM : Rs → R on each local patch xi , giving rise to M feature maps. Under the common setting,
where the representation functions are selected to be fθd (x) = σ(wdT x + bd ) for some point-wise activation
σ(·) and parameterized by θd = (wd , bd ) ∈ Rs × R, the representation layer reduces to the standard convolutional layer. Other possibilities, e.g. gaussian functions with diagonal covariances, have also been considered
in Cohen et al. (2016a). Following the representation layer, are hidden layers indexed by l = 0, . . . , L − 1,
each begins with a 1 × 1 conv operator, which is just an rl−1 × 1 × 1 convolutional layer with rl−1 input channels and rl output channels, with the sole exception that parameters of each kernel could be spatially unshared
(known as locally-connected layer (Taigman et al., 2014)). Following each conv layer is a spatial pooling, that
takes products of non-overlapping two-dimensional windows covering the output of the previous layer, where
for l = L − 1 the pooling window is the size of the entire spatial dimension (i.e. global pooling), reducing
its output’s shape to a rL−1 × 1 × 1, i.e. an rL−1 -dimensional vector. The final L layer maps this vector
with a dense linear layer into the Y network outputs, denoted by hy (x1 , . . . , xN ), representing score functions
classifying each X to one of the classes through: y ∗ = argmaxy hy (x1 , . . . , xN ). As shown in Cohen et al.
(2016a), these functions have the following form:
hy (x1 , . . . , xN ) =
M
X
Ayd1 ,...,dN
d1 ,...,dN =1
N
Y
fθdi (xi )
(5)
i=1
where Ay , called the coefficients tensor, is a tensor of order N and dimension M in each mode, which for the
sake of discussion can simply be seen as a multi-dimensional array, specified by N indices d1 , . . . , dN each
ranging in {1, . . . , M }, with entries given by polynomials in the network’s conv weights. A byproduct of eq. 5
is that for a fixed set of M representation functions, all functions represented by ConvACs lay in the same
subspace of functions.
B
C OMPARISON TO P OOLING G EOMETRY
From theorem 1 we learn that overlaps give rise to networks which almost always cannot be efficiently implemented by non-overlapping ConvAC with standard pooling geometry. However, as proven by Cohen and
Shashua (2017), a ConvAC that uses a different pooling geometry – i.e. the input to the pooling layers are not
strictly contiguous windows from the previous layer – also cannot be efficiently implemented by the standard
ConvAC with standard pooling geometry. This raises the question of whether overlapping operations are simply equivalent to a ConvAC with a different pooling geometry and nothing more. We answer this question in
two parts. First, a ConvAC with a different pooling geometry might be able to implement some function more
efficiently than ConvAC with standard pooling geometry, however, the reverse is also true, that a ConvAC with
standard pooling can implement some functions more efficiently than ConvAC with alternative pooling. In contrast, a ConvAC that uses overlaps is still capable to implement efficiently any function that a non-overlapping
ConvAC with standard pooling can. Second, we can also show that some overlapping architectures are exponentially efficient than any non-overlapping ConvAC regardless of its pooling geometry. This is accomplished
by first extending lemma 1 to this case:
13
Published as a conference paper at ICLR 2018
Lemma 2. Under the same conditions as lemma 1, if for all partitions P ∪· Q such that |P | = |Q| = N/2
it holds that rank (JA(hy )KP,Q ) ≥ T , then any non-overlapping ConvAC regardless of its pooling geometry
must have at least T channels in its next to last layer to induce the same grid tensor.
Next, in theorem 2 below show that some overlapping architectures can induce grid tensors whose matricized
rank is exponential for any equal partition of its indices, proving they are indeed exponentially more efficient:
Theorem 2. Under the same settings as theorem 1, consider a GC network whose representation layer is
followed by a GC layer with local receptive field H × H, stride 1 × 1, and D ≥ M output channels, whose
parameters are “unshared”, i.e. unique to each spatial location in the output of the layer as opposed to shared
across them, followed by (L − 1) arbitrary GC layers, whose final output is a scalar. For any choice of the
parameters, except a null set (with respect to the Lebesgue measure) and for any template vectors such that F
H2
is non-singular, then the matricized rank of the induced grid tensor is equal to M 2 , for any equal partition of
the indices. The exact same result holds if the parameters of the first GC layers are “shared” and D ≥ M · H 2 .
Proof sketch. We follow the same steps of our proof of theorem 1, however, we do not construct just one
specific overlapping network that attains a rank of D ≥ M · H 2 , for all possible matricizations of the induced
grid tensor. Instead, we construct a separate network for each possible matricization. This proves that with
respect to the Lebesgue measure over the network’s parameters space, separately for each pooling geometry,
the set of parameters for which the lower bound does not hold is of measure zero. Since a finite union of zero
measured sets is also of measure zero, then the lower bound with respect to all possible pooling geometries
holds almost everywhere, which concludes the proof sketch. See app. C.5 for a complete proof.
It is important to note that though the above theorem shows that pooling geometry on its own is less expressive
than overlapping networks with standard pooling, it does not mean that pooling geometry is irrelevant. Specifically, we do not yet know the effect of combining both overlaps and alternative pooling geometries together.
Additionally, many times sufficient expressivity is not the main obstacle for solving a specific task, and the
inductive bias induced by a carefully chosen pooling geometry could help reduce overfitting.
C
D EFERRED P ROOFS
In this section we present our proofs for the theorems and claims stated in the body of the article.
C.1
P RELIMINARIES
In this section we lay out the preliminaries required to understand the proofs in the following sections. We
begin with a limited introduction to tensor analysis, followed by quoting a few relevant known results relating
tensors to ConvACs.
We begin with basic definitions and operations relating to tensors. Let A ∈ RM1 ⊗···⊗MN be a tensor of order
N and dimension Mi in each mode i ∈ [N ] (where [N ] = {1, . . . , N }), i.e. Ad1 ,...,dN ∈ R for all i ∈ [N ]
(2)
(1)
and di ∈ [Mi ]. For tensors A(1) and A(2) of orders N (1) and N (2) , and dimensions Mi1 and Mi2 in each of
(1)
(2)
(1)
(2)
the modes i1 ∈ [N ] and i2 ∈ [N ], respectively, we define their tensor product A ⊗ A as the order
N (1) + N (2) tensor, where
(1)
(2)
A(1) ⊗ A(2)
= Ad1 ,...,d (1) · Ad (1) ,...,d (1) (2)
d1 ,...,d (1)
N
+N (2)
N
N
+1
N
+N
For a set of vectors v(1) ∈ RM1 , . . . , v(N ) ∈ RMN , the N ordered tensor A = v(1) ⊗· · ·⊗v(N ) is called an
elementary tensor, or rank-1 tensor. More generally, any tensor can be represented as a linear combination of
P
(Z,1)
rank-1 tensors, i.e. A = Z
⊗· · ·⊗v(Z,1) , known as rank-1 decomposition, or CP decomposition,
z=1 v
where the minimal Z for which this equality holds is knows as the tensor rank of A. Given a set of matrices
0
0
F (1) ∈RM1 ×M1 , . . . , F (N ) ∈RMN ×MN , we denote by F = (F (1) ⊗· · ·⊗F (N ) ) the linear transformation from
0
0
RM1 ⊗···⊗MN to RM1 ⊗···⊗MN , such that for any elementary tensor A, with notations as above, it holds that:
F(A) = F (1) (v(1) ) ⊗ · · · ⊗ F (N ) (v(N ) )
F(A) is defined for a general tensor A through its rank-1 decomposition comprising elementary tensors and
applying F on each of them, which can be shown to be equivalent to
F(A)k1 ,...,kN =
M1
X
d1 =1
MN
···
X
dN =1
14
Ad1 ,...,dN
N
Y
i=1
(i)
Fki ,di
(6)
Published as a conference paper at ICLR 2018
· = [N ] be a disjoint partition of its
A central concept in tensor analysis is that of tensor matricization. Let P ∪Q
indices, such that P = {p1 , . . . , p|P | } with p1 < . . . < p|P | , and Q = {q1 , . . . , q|Q| } withq1 < . . . <q|Q| .
Q|P |
The matricization of A with respect to the partition P ∪· Q, denoted by JAKP,Q , is the
t=1 Mpt -by
Q
|Q|
matrix holding the entries of A, such that for all i ∈ [N ] and di ∈ [Mi ] the entry Ad1 ,...,dN is
t=1 Mqt
Q|Q|
P|Q|
Q|P |
P |
placed in row index 1 + |P
t=1 (dqt − 1)
t=1 (dpt − 1)
t0 =t+1 Mqt0 .
t0 =t+1 Mpt0 and column index 1 +
Applying the matricization operator J·KP,Q on the tensor product operator results in the Kronecker Product, i.e.
for an N -ordered tensor A, a K-ordered tensor B, and the partition P ∪· Q = [N + K], it holds that
JA ⊗ BKP,Q = JAKP ∩[N ],Q∩[N ]
JBK(P −N )∩[K],(Q−N )∩[K]
where P − N and Q − N are simply the sets obtained by subtracting the number N from every element of P
or Q, respectively. In concrete terms, the Kronecker product for the matrices A ∈ RM1 ×M2 and B ∈ RN1 ×N2
results in the matrix A B ∈ RM1 N1 ×M2 N2 holding Aij Bkl in row index (i − 1)N1 + k and column index
(j − 1)N2 + l. An important property of the Kronecker product is that rank (A B) = rank (A) · rank (B).
Typically, when wish to compute rank (JAKP,Q ), we will first decompose it to a Kronecker product of matrices.
For a linear transform F, as defined above in eq.6, and a partition P ∪· Q, if F (1) , . . . , F (N ) are non-singular
matrices, then F is invertible and the matrix rank of JAKP,Q equals to the matrix rank of JF(A)KP,Q (see proof
in Hackbusch (2012)). Finally, we define the concept of grid tensors: for a function
f : Rs × · · · × Rs → R and a set of template vectors x(1) , . . . , x(M ) ∈ Rs , the N -order grid tensor A(f ) is
defined by (A(f ))d1 ,...,dN = f (x(d1 ) , . . . , x(dN ) ).
In the context of ConvACs, circuits and the functions they can realize are typically examined through the
matricization of the grid tensors they induce. The following is a succinct summary of the relevant known
results used in our proofs – for a more detailed discussion, see previous works (Cohen et al., 2016a; Cohen
and Shashua, 2016; 2017). Using the same notations from eq. 5 describing a general ConvAC, let Ay be
the coefficients tensor of order N and dimension M in each mode, and let fθ1 , . . . , fθM :Rs →R be a set
of M representation functions (see app. A). Under the above definitions, a non-overlapping ConvAC can be
said to decompose the coefficients tensor Ay . Different network architectures correspond to known tensor
decompositions: shallow networks corresponds to rank-1 decompositions, and deep networks corresponds to
Hierarchical Tucker decompositions.
In Cohen and Shashua (2017), it was found that the matrix rank of the matricization of the coefficients tensors
Ay could serve as a bound for the size of networks decomposing Ay . For the conventional non-overlapping
ConvAC and the contiguous “low-high” partition P = {1, . . . , N/2}, Q = {N/2 + 1, . . . , N } of [N ], the rank
of the matricization JAy KP,Q serves as a lower-bound on the number of channels of the next to last layer of
any network which decomposes the coefficients tensor A. In the common case of square inputs, i.e. the input
is of shape H × H and N = H 2 , it is more natural to represent indices by pairs (j, i) denoting the spatial
location of each “patch” x(j,i) , where the first argument denotes the vertical location and the second denotes the
horizontal location. Under such setting the equivalent “low-high” partitions are either the “left-right” partition,
i.e. P = {(j, i)|j ≤ H
}, Q = {(j, i)|j > H
}, or the “top-bottom” partition, i.e. P = {(j, i)|i ≤ H
}, Q =
2
2
2
H
{(j, i)|i > 2 }. More generally, when considering networks using other pooling geometries, i.e. not strictly
contiguous pooling windows, then for each pooling geometry there exists a corresponding partition P ∪· Q such
that rank (JAy KP,Q ) serves as its respective lower-bound.
Though the results in Cohen and Shashua (2017) are strictly based on the matricization rank of the coefficients tensors, they can be transferred to the matricization rank of grid tensors as well. Grid tensors were
first considered for analyzing ConvACs in Cohen and Shashua (2016). For a set of M template vectors
x(1) , . . . , x(M ) ∈ Rs , we define the matrix F ∈ RM ×M by Fij = fθj (x(i) ). With the above notations
in place, we can write the the grid tensor A(hy ) for the function hy (x1 , . . . , xN ) as:
A(hy )k1 ,...,kN = hy (x(k1 ) , . . . , x(kN ) )
=
M
X
Ayd1 ,...,dN
=
fθdi (x(ki ) )
i=1
d1 ,...,dN =1
M
X
N
Y
Ayd1 ,...,dN
d1 ,...,dN =1
N
Y
Fki ,di
i=1
⇒ A(hy ) = (F ⊗ · · · ⊗ F )(Ay )
If the representation functions are linearly independent and continuous, then we can choose the template vectors such that F is non-singular (see Cohen and Shashua (2016)), which according to the previous discussion on tensor matricization, means that for any partition P ∪· Q and any coefficients tensor Ay , it holds that
rank (JAy KP,Q ) = rank (JA(hy )KP,Q ). Thus, any lower bound on the matricization rank of the grid tensor
15
Published as a conference paper at ICLR 2018
translates to a lower bound on the matricization rank of the coefficients tensors, which in turn serves as lower
bound on the size of non-overlapping ConvACs. The above discussion leads to the proof of lemma 1 and
lemma 2 that were previously stated:
Proof of lemma 1 and lemma 2. For the proofs of the base results with respect to the coefficients tensor, see
Cohen and Shashua (2017). To prove it is possible to choose the template vectors such that F is non-singular,
see Cohen and Shashua (2016). To prove that if F is non-singular, then the grid tensor and the coefficients
tensor have the same matricization rank, see lemma 5.6 in Hackbusch (2012).
We additionally quote the following lemma regarding the prevalence of the maximal matrix rank for matrices
whose entries are polynomial functions:
Lemma 3. Let M, N, K ∈ N, 1 ≤ r ≤ min{M, N } and a polynomial mapping A : RK → RM ×N , i.e.
for every i ∈ [M ] and j ∈ [N ] it holds that Aij : Rk → R is a polynomial function. If there exists a point
x ∈ RK such that rank (A(x)) ≥ r, then the set {x ∈ RK |rankA(x) < r} has zero measure (with respect to
the Lebesgue measure over RK ).
Proof. See Sharir et al. (2016).
Finally, we simplify the notations of the GC layer with product pooling function for the benefit of following our
(l−1)
×R(l) ×R(l)
proofs below. We will represent the parameters of the l’th GC layer by {(w(l,c) ∈ RD
, b(l,c) ∈
(l)
(l)
(in)
(in)
(in)
RR ×R )}c∈[D(l) ] , where w(l,c) represents the weights and b(l,c) the biases. Let X ∈ RH ×H ×D
be the input to the layer and Y ∈ RD
Yc,u,v
(out)
×H (out) ×H (out)
be the output, then the following equality holds:
(in)
(l)
(l)
D
R
X (l,c)
YR
Y
(l,c)
bji +
wdji Xd,uS (l) +j,vS (l) +i
(7)
=
j=1 i=1
d=1
The above treats the common case where the parameters are shared across all spatial locations,
but sometimes we wish to consider the “unshared” case, in which there is are different weights
(l−1)
×R(l) ×R(l)
and biases for each location, which we denote by {(w(l,c,u,v) ∈ RD
, b(l,c,u,v) ∈
R(l) ×R(l)
R
)}c∈[D(l) ],u∈[H (out) ],v∈[H (out) ] .
With the above definitions and lemmas in place, we are ready to prove the propositions and theorems from the
body of the article.
C.2
P ROOF OF P ROPOSITION 1
Proposition 1 is a direct corollary of the following two claims:
(in)
(in)
(in)
(out)
(out)
(out)
×H
×H
be a function realized by a single GC
→ RD
Claim 1. Let f :RD ×H ×H
layer with R × R local receptive field, S × S stride, and D(out) output channels, that is parameterized by
(out)
{(w(c) , b(c) )}D
. For all R̃ ≥ R, a GC layer with R̃×R̃ local receptive field, S×S stride, and C output
c=1
(out)
channels, parameterized by {(w̃(c) , b̃(c) )}D
, could also realize f . The same is true for the unshared case
c=1
of both layers.
Proof. The claim is trivially satisfied by setting w̃(c) such that it is equal to w(c) in all matching coordinates,
while using zeros for all other coordinates. Similarly, we set b̃(c) to be equal to b(c) in all matching coordinates,
while using ones for all other coordinates.
(in)
(in)
(in)
(out)
(out)
(out)
×H
×H
Claim 2. Let f :RD ×H ×H
→ RD
be a function realized by a GC layer with
(out)
R×R local receptive field and 1×1 stride, parameterized by {(w(c) , b(c) )}D
. Then there exists an asc=1
signment to (w, b) such that f is the identity function f (X) = X. The same is true for the unshared case of
both layers.
Proof. From claim 1 it is sufficient to show the above holds for R = 1. Indeed, setting
(
1 d=c
(c)
wd =1[d=c] =
0 d 6= c
and b(c) ≡ 0 satisfies the claim.
16
Published as a conference paper at ICLR 2018
C.3
P ROOF OF T HEOREM 1
We wish to show that for all choices of parameters, except a null set (with respect to Lebesgue measure) the
grid tensor induced by the given GC network has rank satisfying eq. 4. Since the entries of the matricized grid
tensor are polynomial function of its parameters, then according to lemma 3, it is sufficient to find a single
example that achieves this bound. Hence, our proof is simply the construction of such an example.
Recall that the template vectors must hold that for the matrix F , defined by Fij = fj (x(i) ), where {fj }M
j=1
are the representation matrices, is a non-singular matrix. We additionally assume in the following claims that
the output of the representation layer is of width and height equal to H ∈ N, where H is an even number – the
claims and proofs can however be easily adapted to the more general case.
Assume a ConvAC as described in the theorem, with representation layer defined according to above, followed
by L GC layers, where the l’th layer has a local receptive field of R(l) , a stride of S (l) , and D(l) output channels.
We first construct our example, that achieves the desired matricization rank, for the simpler case where the first
layer following the representation layer has a local receptive field large enough, i.e. when it is larger than
H
. Recall that for the first layer the total receptive field is equal to its local receptive field. In the context of
2
theorem 1, this first layer satisfies the conditions necessary to produce the lower bound given in the theorem.
The specific construction is presented in the following claim, which relies on utilizing the large local receptive
field to match each spatial location in the left side of the input with one from the right side, such that for each
such pair, the respective output of the first layer will represent a mostly diagonal matrix. We then set the rest
of the parameters such that the output of the entire network is defined by a tensor product of mostly diagonal
matrices. Since the matricization rank of the tensor product of matrices is equal to the product of the individual
ranks, it results in an exponential form of the rank as is given in the theorem.
Claim 3. Assume a ConvAC as defined above, ending with a single scalar output. For all l ∈ [L], the param(l−1)
(l)
(l)
×R(l) ×R(l)
eters of the l-th GC layer are denoted by {(w(l,c) ∈ RD
, b(l,c) ∈ RR ×R )}c∈[D(l) ] . Let
h(x1 , . . . , xN ) be the function realized the output of network. Additionally define R ≡ R(1) , S ≡ S (1) and
D ≡ min{D(1) , M }. If R > H
, and the weights w(1,c) and biases b(1,c) of the first GC layer layer are set
2
to:
(
−α F −1 m,c c ≤ D and (j, i) ∈ {(1, 1), (ρ, τ )}
(1,c)
wmji =
0
Otherwise
β c ≤ D and (j, i) ∈ {(1, 1), (ρ, τ )}
(1,c)
bji = 1 c ≤ D and (j, i) 6∈ {(1, 1), (ρ, τ )}
0 Otherwise
where β =
2α
,
D
then there exists an assignment to α and the parameters of the other GC layers such that
H−R
H
rank (JA(h)KP,Q ) = Db S +1c·d S e , where P ∪· Q is either the “left-right” or “top-bottom” partition, and
(ρ, τ ) equals to (1, R) or (R, 1), respectively.
Proof. The proof for either the “left-right” or “top-bottom” partition is completely symmetric, thus it is
enough to prove the claim for the “left-right” case, where (ρ, τ ) = (1, R). We wish to compute the entry
A(h)d(1,1) ,...,d(H,H) of the induced grid tensor for arbitrary indices d(1,1) , . . . , d(H,H) . Let O ∈ RM ×H×H be
the 3-order tensor output of the representation layer, where Om,j,i = Fd(j,i) ,m for the aforementioned indices
and for all 1 ≤ i, j ≤ H and m ∈ [M ].
We begin by setting the parameters of all layers following the first GC layer, such that they are equal to computing the sum along the channels axis of the output of the first GC layer, followed by a global product of all of
the resulting sums. To achieve this, we can first assume w.l.o.g. that these layers are non-overlapping through
proposition 1. We then set the parameters of the second GC layer to w(2,c) = 1 and b(2,c) ≡ 0, i.e. all
ones and all zeros, respectively, which is equivalent to taking the sum along the channels axis for each spatial
location, followed by taking the products over non-overlapping local receptive fields of size R(2) . For the other
layers, we simply set them as to take just the output of(the first channel of the output of the preceding layer,
1 d=1
(l,c)
which is equal to setting their parameters to wdji =
and b(l,c) ≡ 0. Setting the parameters as
0 d 6= 1
described above results in:
!
(1)
R
M
Y D
X Y
X
(1,c)
(1,c)
(A(h))d(1,1) ,...,d(H,H) =
(8)
bji +
wmji Om,uS+j,vS+i
0≤uS<H c=1 j,i=1
|
0≤vS<H
17
m=1
{z
g(u,v,c,j,i)
}
Published as a conference paper at ICLR 2018
where we extended O with zero-padding for the cases where uS + j > H or vS + i > H, as mentioned in
sec. 2. Next, we go through the technical process of reducing eq. 8 to a product of matrices.
(1,c)
(1,c)
Substituting the values of wdji and bji with those defined in the claim, and computing the value of
g(u, v, c, j, i) results in:
P
−1
β−α M
F
c ≤ D and vS + R ≤ H and (j, i) ∈ {(1, 1), (1, R)}
m=1 F
m,c d(uS+j,vS+i) ,m
PM
−1
β
−
α
F
F
c
≤ D and vS + R > H and (j, i) = (1, 1)
m=1
m,c d(uS+j,vS+i) ,m
g(u, v, c, j, i) = β
c ≤ D and vS + R > H and (j, i) = (1, R)
c ≤ D and (j, i) 6∈ {(1, 1), (1, R)}
1
0
c>D
−1
β − α FF
c ≤ D and vS + R ≤ D and (j, i) ∈ {(1, 1), (1, R)}
d(uS+1,vS+i) ,c
−1
β
−
α
F
F
c ≤ D and vS + R > H and (j, i) = (1, 1)
d(uS+1,vS+i) ,c
= β
c ≤ D and vS + R > H and (j, i) = (1, R)
1
c ≤ D and (j, i) 6∈ {(1, 1), (1, R)}
0
c>D
from which we derive:
2
Dβ − αβ(1[d(uS+1,vS+1) ≤D] + 1[d(uS+1,vS+R) ≤D] ) vS + R ≤ H
+ α2 1[d(uS+1,vS+1) =d(uS+1,vS+R) ≤D]
f (u, v) ≡
g(u, v, c, j, i) =
Dβ 2 − αβ1
c=1 j,i=1
vS + R > H
[d(uS+1,vS+1) ≤D]
Q
where (A(h))d(1,1) ,...,d(H,H) = 0≤uS<H f (u, v).
(1)
D
X
R
Y
0≤vS<H
At this point we branch into two cases. If S divides R − 1, then for all u, v ∈ N such that vS + R ≤ H and
) depends only on the indices d(uS+1,vS+1) and
uS < H, the above expression for f (u, v) and f (u, v + R−1
S
d(uS+1,vS+R) , while these two indices affect only the aforementioned expressions. By denoting
(u,v)
Ad(uS+1,vS+1) ,d(uS+1,vS+R) = f (u, v) · f (u, v + R−1
), we can write it as:
S
Dβ 2 − 2αβ + α2 1[i=j] Dβ 2 − αβ
i, j ≤ D
Dβ 2 − αβ Dβ 2
i ≤ D and j > D
(u,v)
2
Aij
=
2
Dβ
−
αβ
i > D and j ≤ D
2
i, j > D
Dβ 2
where i, j ∈ [M ] stand for the possible values of d(uS+1,vS+1) and d(uS+1,vS+R) , respectively. Substituting
1/4
2α
, results in:
, as stated in the claim, and setting α = D
D
2
1[i=j] i, j ≤ D
4
i ≤ D and j > D
(u,v)
= D
Aij
2
i > D and j ≤ D
D
8
i, j > D
D
which means rank A(u,v) = D. Since (A(h))d(1,1) ,...,d(H,H) equals to the product
Q
(u,v)
, then JA(h)KP,Q equals to the Kronecker product of the matrices
0≤uS<H Ad
(uS+1,vS+1) ,d(uS+1,vS+R)
β=
0≤vS≤H−R
in {A(u,v) |0 ≤ uS < H, 0 ≤ vS ≤ H − R}, up to permutation of its rows and columns, which do not affect
its matrix rank. Thus, the matricization rank of A(h) satisfies:
H−R
(u,v)
H
|0≤uS<H,0≤vS≤H−R}|
rank (JA(h)K ) = D|{A
= Db S +1c·d S e
P,Q
which proves the claim for this case.
If S does not divide R − 1, then for all u, v ∈ N, such that vS + R ≤ H and uS < H, it holds that f (u, v)
depends only on the indices d(uS+1,vS+1) and d(uS+1,vS+R) , and they affect only f (u, v). Additionally, for
all u, v ∈ N, such that H < vS + R, vS < H and uS < H, it holds that f (u, v) depends only on the index
(u,v)
d(uS+1,vS+1) , and this index affects only f (u, v). Let us denote Ad(uS+1,vS+1) ,d(uS+1,vS+R) = f (u, v) for
vS + R ≤ H:
Dβ 2 − 2αβ + α2 1[i=j]
i, j ≤ D
Dβ 2 − αβ
i ≤ D and j > D
(u,v)
Aij
=
2
Dβ
−
αβ
i > D and j ≤ D
Dβ 2
i, j > D
18
Published as a conference paper at ICLR 2018
which by setting β =
2α
D
and α = 1 results in:
1[i=j]
2
(u,v)
Aij
= D
2
D
4
D
i, j ≤ D
i ≤ D and j > D
i > D and j ≤ D
i, j > D
(u,v)
which means rank A(u,v) = D. For vS + R > H, and we can define the vector ad(uS+1,vS+1) = f (u, v),
which by using the same values of α and β results in:
(
2
i≤D
(u,v)
ai
= D
4
i>D
D
By viewing a(u,v) as either a column or row vector, depending on whether d(uS+1,vS+1) ∈ P or
d(uS+1,vS+1) ∈ Q, respectively, it holds that JA(h)KP,Q equals to the Kronecker product of the matrices in
{A(u,v) } 0≤uS<H ∪ {a(u,v) } 0≤uS<H , up to permutations of its rows and columns, which do not affect
H−R<vS<H
0≤vS≤H−R
the rank. Since a(u,v) 6= 0 then rank a(u,v) = 1, which means the matricization rank of A(h) once again
holds:
rank (JA(h)KP,Q ) = D|{A
(u,v)
|0≤uS<H,0≤vS≤H−R}|
= Db
H−R
+1
S
c·d H
S e
In the preceding claim we have describe an example for the case where the total receptive of the first GC layer
is already large enough for satisfying the conditions of the theorem. In the following claim we extend this result
for the general case. This is accomplished by showing that a network comprised of just L GC layers with local
receptive fields {R(l) × R(l) }l∈[L] , strides {S (l) × S (l) }l∈[L] , and output channels {D(l) }l∈[L] , can effectively
compute the same output as the first GC layer from claim 3, for all inputs 3 .
Recall that the layer from claim 3 performs an identical transformation on each M × 1 × 1 patch from its input,
followed by taking the point-wise product of far-away pairs of transformed patches. Thus, the motivation
behind the specific construction we use, is to use the first of the L layers to perform this transformation, while
using half of its output channels for storing the transformed patch from the same location, and the other half for
storing a transformed patch, but from a location farthest to the right, constrained by its local receptive field. This
is equal to having one set of transformed patches sitting still, while another “shifted” set of transformed patches.
The other layers simply pass the the first half of the channels as is, using an identity operation as defined in
claim 2, while continuously shifting the other half of the channels more and more to the left, bringing faraway
patches closer together. Finally, at the last layer we take both halves and multiple them together.
Claim 4. Assume Φ is a ConvAC comprised of just L GC layers as described above, where the output of the
network is not limited to a scalar value. Assume the total stride of the L-th GC layer is greater than H/2, and
(L)
(L,α)
let TS and TR
be the total stride and the α-minimal total receptive field, respectively, for α = bH/2c + 1.
(L,α)
(L)
Let Ψ be a ConvAC comprised of a single GC layer with local receptive field R ≡ TR
, stride S ≡ TS ,
output channels D ≡ min{ 21 min1≤l<L D(l) , D(L) }, where its weights and biases are set to the following:
(
Am,c (j, i) ∈ {(1, 1) , (ρ, τ )}
(c)
wmji =
0
Otherwise
(
βc (j, i) ∈ {(1, 1) , (ρ, τ )}
(c)
bji =
1 Otherwise
for β ∈ RD , A ∈ RM ×D and (ρ, τ ) ∈ {(1, R), (R, 1)}. Then, there exists a set of weights to the layers
of Φ such that for every input X, the output of Φ is equivalent to the output of Ψ for channels ≤ D, and zero
otherwise.
Proof. The two possible cases for (ρ, τ ) are completely symmetric, thus it is enough to prove the claim just
for (ρ, τ ) = (1, R). Additionally, we can assume w.l.o.g. that ∀l, R(l) > 1, by setting any 1 × 1 layer to act
as pass-through according to claim 2, and also assume that the α-minimal total receptive field is exactly equal
to the total receptive field of the L-th layer, by applying claim 1 to realize an equivalent network with smaller
windows. Finally, the case for L = 1 is trivial, and thus we assume L > 1.
3
Notice that in this context, there is no representation layer, and the input can be any 3-order tensor.
19
Published as a conference paper at ICLR 2018
n
o
Let us set the parameters w(l,k) , b(l,k) of the layers of Φ as follows:
(l,k)
wdji
(l,k)
bji
−Ad,k
(l = 1) and (1 ≤ k ≤ D) and (j, i) = (1, 1)
−A
(l = 1) and (D < k ≤ 2D) and (j, i) = (1, R(l) )
d,k−D
(1 < l < L) and (1 ≤ k ≤ D) and (j, i) = (1, 1)
1[d=k]
= 1[d=k]
(1 < l < L) and (D < k ≤ 2D) and (j, i) = (1, R(l) )
1[d=k]
(l = L) and (1 ≤ k ≤ D) and (j, i) = (1, 1)
1
(l
= L) and (1 ≤ k ≤ D) and (j, i) = (1, R(l) )
[d=M
+k]
0
Otherwise
βk
(l = 1) and (1 ≤ k ≤ D) and (j, i) = (1, 1)
β
(l = 1) and (D < k ≤ 2D) and (j, i) = (1, R(l) )
k−D
(1 < l < L) and (1 ≤ k ≤ D) and (j, i) = (1, 1)
0
= 0
(1 < l < L) and (D < k ≤ 2D) and (j, i) = (1, R(l) )
0
(l = L) and (1 ≤ k ≤ D) and (j, i) = (1, 1)
0
(l
= L) and (1 ≤ k ≤ D) and (j, i) = (1, R(l) )
1
Otherwise
(l)
(l)
(l)
It is left to prove the above satisfies the claim. Let O(l) ∈ RD ×H ×H be the output of the l-th layer,
for l ∈ [0, . . . , L], where H (l) is the width and height of the output. We additionally assume that for indices
beyond the bounds of [D(l) ] × [H (l) ] × [H (l) ] the value of O(l) is zero, i.e. we assume zero padding when
applying the convolutional operation of the GC layer. We extend the definition for l = 0, by setting D(0) ≡ M
and H (0) ≡ H, where we identify O(0) with the input to the network Φ. Given the above, the output of the
first layer for k ∈ [D(1) ] and 0 ≤ u, v < H (1) , is as follows:
(1)
Ok,u+1,v+1
=
(1)
R
Y
b(1,k)
ji
j,i=1
+
(0)
D
X
(1,k) (0)
wdji Od,uSh +j,vSw +i
d=1
PM
(0)
Ad,k · Od,uSh +1,vSw +1
βk + d=1
PM
(0)
= βk−D + d=1 Ad,k−D · Od,uS
(1)
h +1,vSw +R
1
1≤k≤D
D < k ≤ 2D
Otherwise
We will show by induction that for 1 < l < L, k ∈ [D(l) ] and 0 ≤ u, v < H (l) the output of the l-th layer
(l)
Ok,u+1,v+1 always equals to:
(l)
Ok,u+1,v+1
1
Ok,uη(l) +1,vη(l) +1
1
= Ok,uη(l) +1,vη(l) +ξ(l)
1
k≤D
D < k ≤ 2D
Otherwise
Q
P
(k)
where η (l) = li=2 S (i) and ξ (l) = R(l) · η (l−1) + l−1
− S (l) ) · η (k−1) . It is trivial to verify that for
k=2 (R
l = 2 it indeed holds, since:
(2)
Ok,u+1,v+1
(1)
Ok,uS (2) +1,vS (2) +1
= O(1) (2)
(2)
(2)
k,uS +1,vS +R
1
20
k≤D
D < k ≤ 2D
Otherwise
Published as a conference paper at ICLR 2018
where η (2) = S (2) and ξ (2) = R(2) . Assume the claim holds up to l − 1, and we will show it also holds for l:
(l)
(l−1)
R
DX
Y
(l)
(l,k)
(l,k)
l−1
bji +
Ok,u+1,v+1 =
wdji O
(l)
(l)
d,uS
j,i=1
+j,vS
+i
d=1
(l−1)
k≤D
Ok,uS (l) +1,vS (l) +1
= O(l−1)(l)
D
< k ≤ 2D
(l)
(l)
k,uS +1,vS +R
1
Otherwise
(1)
Ok,(uS (l) )η(l−1) +1,(vS (l) )η(l−1) +1
Induction Hypothesis ⇒ = O(1) (l) (l−1)
+1,(vS (l) +R(l) −1)η (l−1) +ξ(l−1)
k,(uS )η
1
1
k≤D
Ok,uη(l) +1,vη(l) +1
1
= Ok,uη(l) +1,vη(l) +ξ(l) D < k ≤ 2D
1
Otherwise
k≤D
D < k ≤ 2D
Otherwise
Were we used the fact that η (l) = S (l) η (l−1) and ξ (l) = R(l) η (l−1) + ξ (l−1) − η (l−1) .
(L)
Finally, we show that Ok,u+1,v+1 for k ≤ D and 0 ≤ u, v < H (L) equals to the output of the single GC layer
specified in the claim:
(L−1)
(L)
DX
R
Y
(L,k)
(L−1)
(L,k)
(L)
bji
wdji Od,uS (L) +j,vS (L) +i
+
Ok,u+1,v+1 =
j,i=1
d=1
(L−1)
(L−1)
(1)
(1)
= Ok,uS (L) +1,vS (L) +1 · Ok,uS (L) +1,vS (L) +R(L)
= Ok,uη(L) +1,vη(L) +1 · Ok,uη(L) +1,vη(L) +ξ(L)
(0)
D
X
(0)
= βk +
Ad,k Od,uη(L) S (1) +1,vη(L) S (1) +1
d=1
· βk +
(0)
D
X
(0)
Ad,k Od,uη(L) S (1) +1, vη(L) +ξ(L) −1 S (1) +R(1) )
(
)
(0)
(0)
D
D
X
X
(0)
(0)
= βk +
Ad,k O
Ad,k O
β
+
k
(L)
(L)
d=1
d=1
d,uTS
+1,vTS
+1
d=1
(L)
d,uTS
(L)
+1,vTS
(L)
+TR
which is indeed equal to the single GC layer. For k > D, both the bias and the weights for the last layer are
(L)
zero, and thus Ok,u+1,v+1 = 1.
Finally, with the above two claims set in place, we can prove our main theorem:
Proof. (of theorem 1) Using claim 4 we can realize the networks from claim 3, for which the matricization
rank for either partition equals to:
H
l
m
H−T (K,b /2c)
R
+1· H
D
TS
TS
Since for any matricization [A (Ψ)]P,Q the entries of the matricization are polynomial functions with respect
to the parameters of the network, then, according to lemma 3, the set of parameters of Φ, that does not attain
the above rank, has zero measure. Since the union of zero measured sets is also of measure zero, then all
parameters except a set of zero measure attain this matricization rank for both partitions at once, concluding
the proof.
C.4
P ROOF OF P ROPOSITION 2
Following theorem 1, to compute the lower bound for the network described in proposition 2, we need to find
the first layer for which its total receptive field is greater than H/2, and then estimate its total stride and its
21
Published as a conference paper at ICLR 2018
α-minimal total receptive field, for α = bH/2c. In the following claims we analyze the above properties of the
given network:
Claim 5. The total stride and total receptive field of the l-th B × B layer in the given network, i.e. the
(2l − 1)-th GC layer after the representation layer, are given by the following equations:
(2l−1)
TS
(2l−1)
TR
(S (1) , . . . , S (2l−1) ) = 2l−1
(R(1) , S (1) , . . . , R(2l−1) , S (2l−1) ) = (2B − 1)2l−1 − B + 1
(2l−1)
Proof. From eq. 1 it immediately follows that TS
(S (1) , . . . , S (2l−1) ) = 2l−1 . From eq. 2, the 2 × 2 stride
(2l)
2 layers do not contribute to the receptive field as R
− S (2l) = 0, which results in the following equation:
(2l−1)
TR
(R(1) , S (1) , . . . , R(2l−1) , S (2l−1) ) = B · 2l−1 +
l−1
X
(B − 1)2i−1
i=1
= B · 2l−1 + (B − 1)(2l−1 − 1)
= (2B − 1)2l−1 − B + 1
Claim 6. The α-minimal total receptive field for the l-th B × B layer in the given network, for α ∈ N and
2l−1 ≤ α < 2l − 1, always equals (α + 1).
Proof. From eq. 3, the following holds:
(2l−1,α)
TR
=
(2l−1)
argmin
TR
∀i∈[l],1≤t2i−1 ≤B
(t1 , 1, 2, 2, t3 , 1, . . . , t2l−1 , 1)
(2l−1)
TR
(t1 ,1,2,2,t3 ,1,...,t2l−1 ,1)>α
=
t2l−1 · 2l−1 +
argmin
∀i∈[l],1≤t2i−1 ≤B
P
i−1
>α
t2l−1 ·2l−1 + l−1
i=1 (t2i−1 −1)2
l−1
X
(t2i−1 − 1)2i−1
i=1
Notice that the right term in the equation resembles a binary representation. If we limit t2i−1 to the set {1, 2},
this term can represent any number in the set {0, . . . , 2l−1 − 1}, and by choosing t2l−1 = 1, the complete term
can represent any number in the set {2l−1 , . . . , 2l − 1}, and specifically, for 2l−1 ≤ α < 2l − 1, there exists an
(2l−1,α)
= α+1.
assignment for t2i−1 ∈ {1, 2} for i ∈ [l−1] such that this terms equal (α+1), and thus TR
With the above general properties for the given network, we can simplify the expression for the lower bound
given in theorem 1:
Claim
7.
If
the
l-th
B × B
layer
in
the
given
network
satisfies
(2l−1)
TR
(R(1) , S (1) , . . . , R(2l−1) , S (2l−1) ) > H/2, then the lower bound given in theorem 1 equals to
2L−2l+1
M2
Proof. From the description of the network and the previous claims it holds that D = M , H = 2L ,
(2l−1)
(2l−1,bH/2c)
TS
= 2l−1 , and TR
= 2L−1 + 1. Substituting all the above in eq. 4 results in:
(Eq. 4) = M
j
=M
= M2
L
· 2
l−1
2L −2L−1 −1
+1
2l−1
2
k
1
2L−l +1− l−1
·2L−l+1
2
2L−2l+1
With all of the above claims in place, we our ready to prove proposition 2:
22
Published as a conference paper at ICLR 2018
Proof. (of proposition 2) From claim 5, we can infer which is the first B × B layer such that its receptive field
is greater than H/2:
(2B − 1) · 2l−1 − B + 1 > 2L−1
2L + 2B − 2
2B − 1
2L + 2B − 2
⇒ l = 1 + log2
2B − 1
⇒ l > log2
Combining the above with claim 7, results in:
M
22L−2l+1
=M
2
≥ M2
(2L ≡ H) ⇒ = M
=M
The limits and the special case for B ≤
C.5
H
5
L
2L−1−2 log2 2 +2B−2
2B−1
L
2L−1−2 log2 2 +2B−2
2B−1
H2
2
H−1 −2
(1+ 2B−1
)
(2B−1)2
2
−2
·(1+ 2B−2
H )
+ 1 are both direct corollaries of the above expressions.
P ROOF OF THEOREM 2
We begin by proving an analogue of claim 3, where we show that for any given matricization of the grid tensor
A(h), induced by the overlapping network realizing the function h, the matricization rank is exponential. The
motivation behind the construction, for when parameters are “unshared”, is that we can utilized the fact that
there are separate sets of kernels, with local receptive fields the size of the input, for each spatial location.
Thus, each kernel can “connect” the index (of the grid tensor) matching its spatial location, with almost any
other index, and specifically such that the two indices come from different sets of the matricization I ∪· J of
A(h). For the “shared” case, we simply use polynomially more output channels to simulate the “unshared”
case.
2
Claim 8. For an arbitrary even partition (I, J) of {(1, 1), . . . , (H, H)}, such that |I| = |J| = H2 , there
exists an assignment to the parameters of the network given in theorem 2, for either the “unshared” or “shared”
settings, such that rank (JA(h)K) = M
H2
2
.
2
Proof. Let (I, J) be an arbitrary even partition of {(1, 1), . . . , (H, H)}, such that |I| = |J| = H2 , I =
2
{i1 , . . . , i|I| }, and J = {j1 , . . . , j|J| }, where for k ∈ [ H2 ] it holds that ik < ik+1 and jk < jk+1 (using
H2
2
such that qk = ik
lexical ordering), and we assume w.l.o.g. that i1 = (1, 1). We define the set {(qk , pk )}k=1
and pk = jk if ik < jk , and otherwise qk = jk and pk = ik .
We prove the t“unshared” case first, where the parameters of the first GC layers are given by
{(w(c,u,v) , b(c,u,v) )}D,H,H
c=1,u=1,v=1 , for which we choose the following assignment:
(c,u,v)
wm,j,i
(c,u,v)
bj,i
=
=
(
(F −1 )m,c
c ≤ M and ∃k ∈
0
(
0
Otherwise
1[c≤M ]
c ≤ M and ∃k ∈
h
h
H2
2
H2
2
i
i
, qk = (u, v) and (j, i) ∈ {(1, 1), pk − qk + (1, 1)}
, qk = (u, v) and (j, i) ∈ {(1, 1), pk − qk + (1, 1)}
Otherwise
where for u, v ∈ [H] such that qk = (u, v) it holds that pk − qk + (1, 1) ∈ {(1, 1), . . . , (H, H)} because
qk ≤ pk . Similar to the proof of claim 3, we set all layers following the first GC layer such that the following
equality holds:
(A(h))d(1,1) ,...,d(H,H) =
H X
D Y
H
Y
(c,u,v)
bji
u,s=1 c=1 j,i=1
+
M
X
m=1
23
!
(c,u,v)
wmji Om,u+j−1,v+i−1
Published as a conference paper at ICLR 2018
Which under the assignment to parameters we chose earlier, it results in:
!
H 2/2 D
M
YX X
−1
(A(h))d(1,1) ,...,d(H,H) =
(F )m,c Fdqk ,m
k=1 c=1
m=1
M
X
!
(F
−1
)m,c Fdpk ,m
m=1
H 2/2
=
D
YX
(F · F −1 )dqk ,c · (F · F −1 )dpk ,c
k=1 c=1
H 2/2
=
D
YX
1[dqk =c] · 1[dpk =c]
k=1 c=1
H 2/2
=
Y
1[dqk =dpk ]
k=1
which means JA(h)K(I,J) equals to the Kronecker product of
H2
2
M ×M -identity matrices, up to permutations
of its rows and columns which do not affect its matrix rank. Thus, rank (JA(h)K) = M
H2
2
.
For the “shared” setting, we denote the parameters of the first GC layer by {(w(d) , b(d) )}D
c=1 , and set them as:
−1
2
∃c ∈[M ]∃u,
(F )m,c
h v2∈
i [H], d = cH + uH + v
(d)
H
and ∃k ∈ 2 , qk = (u, v) and (j, i) ∈ {(1, 1), pk − qk + (1, 1)}
wm,j,i =
0
Otherwise
2
∃c ∈[M ]∃u,
0
h v2∈
i [H], d = cH + uH + v
(d)
and ∃k ∈ H2 , qk = (u, v) and (j, i) ∈ {(1, 1), pk − qk + (1, 1)}
bj,i =
1[d≤M H 2 ] Otherwise
the parameters of the other layers are set as in the “unshared” case, and the proof follows similarly.
In the preceding claim, we have found a separate example for each matricization, such that the matricization
rank is exponential. In the following proof of the theorem 2, we leverage basic properties from measure theory
to show that almost everywhere the induced grid tensor has an exponential matricization rank, under every
possible even matricization – without explicitly constructing such an example.
Proof. (of theorem 2) For any even partition (I, J) of {(1, 1), . . . , (H, H)}, according to claim 8 there exist
parameters for which rank (JA(h)K)(I,J) = M
H2
2
, and thus according to lemma 3 the set of parameters for
H2
2
which rank (JA(h)K) < M
is of measure zero. Since the finite union of sets of measure zero is also
of measure zero, then almost everywhere (with respect to the Lebesgue measure) the parameters results in
networks such that for all even partitions rank (JA(h)K)(I,J) = M
24
H2
2
.
| 9cs.NE
|
A systematic study of the class imbalance problem
in convolutional neural networks
arXiv:1710.05381v1 [cs.CV] 15 Oct 2017
Mateusz Buda1, 2
Atsuto Maki2
Maciej A. Mazurowski1, 3
1
Department of Radiology, Duke University School of Medicine, Durham, NC, USA
2
Royal Institute of Technology (KTH), Stockholm, Sweden
3
Department of Electrical and Computer Engineering, Duke University, Durham, NC, USA
{buda, atsuto}@kth.se,
[email protected]
Abstract
In this study, we systematically investigate the impact of class imbalance on classification performance of convolutional neural networks (CNNs) and compare frequently used methods to address the issue. Class imbalance is a common problem
that has been comprehensively studied in classical machine learning, yet very limited systematic research is available in the context of deep learning. In our study,
we use three benchmark datasets of increasing complexity, MNIST, CIFAR-10 and
ImageNet, to investigate the effects of imbalance on classification and perform an
extensive comparison of several methods to address the issue: oversampling, undersampling, two-phase training, and thresholding that compensates for prior class
probabilities. Our main evaluation metric is area under the receiver operating characteristic curve (ROC AUC) adjusted to multi-class tasks since overall accuracy metric
is associated with notable difficulties in the context of imbalanced data. Based on
results from our experiments we conclude that (i) the effect of class imbalance on
classification performance is detrimental; (ii) the method of addressing class imbalance that emerged as dominant in almost all analyzed scenarios was oversampling;
(iii) oversampling should be applied to the level that totally eliminates the imbalance,
whereas undersampling can perform better when the imbalance is only removed to
some extent; (iv) as opposed to some classical machine learning models, oversampling
does not necessarily cause overfitting of CNNs; (v) thresholding should be applied
to compensate for prior class probabilities when overall number of properly classified
cases is of interest.
Keywords: Class Imbalance, Convolutional Neural Networks, Deep Learning, Image
Classification
1
Introduction
Convolutional neural networks (CNNs) are gaining significance in a number of machine
learning application domains and are currently contributing to the state of the art in the
field of computer vision, which includes tasks such as object detection, image classification,
and segmentation. They are also widely used in natural language processing or speech
1
recognition where they are replacing or improving classical machine learning models [1].
CNNs integrate automatic feature extraction and discriminative classifier in one model,
which is the main difference between them and traditional machine learning techniques.
This property allows CNNs to learn hierarchical representations [2]. The standard CNN
is built with fully connected layers and a number of blocks consisting of convolutions,
activation function layer and max pooling [3, 4, 5]. The complex nature of CNNs requires
a significant computational power for training and evaluation of the networks, which is
addressed with the help of modern graphical processing units (GPUs).
A common problem in real life applications of deep learning based classifiers is that
some classes have a significantly higher number of examples in the training set than
other classes. This difference is referred to as class imbalance. There are plenty of
examples in domains like computer vision [6, 7, 8, 9, 10], medical diagnosis [11, 12],
fraud detection [13] and others [14, 15, 16] where this issue is highly significant and the
frequency of one class (e.g., cancer) can be 1000 times less than another class (e.g., healthy
patient). It has been established that class imbalance can have significant detrimental
effect on training of traditional classifiers [17] including multi-layer perceptrons [18]. It
affects both convergence during the training phase and generalization of a model on the
test set. While the issue very likely also affects deep learning, no systematic study on the
topic is available.
Methods of dealing with imbalance are well studied for classical machine learning
models [19, 17, 20, 18]. The most straightforward and common approach is the use of
sampling methods. Those methods operate on the data itself (rather than the model)
to increase its balance. Widely used and proven to be robust is oversampling [21]. Another option is undersampling. Naı̈ve version, called random majority undersampling,
simply removes a random portion of examples from majority classes [17]. The issue of
class imbalance can be also tackled on the level of the classifier. In such case, the learning algorithms are modified, e.g. by introducing different weights to misclassification of
examples from different classes [22] or explicitly adjusting prior class probabilities [23].
Some previous studies showed results on cost sensitive learning of deep neural networks [24, 25, 26]. New kinds of loss function for neural networks training were also
developed [27]. Recently, a new method for CNNs was introduced that trains the network in two-phases in which the network is trained on the balanced data first and then
the output layers are fine-tuned [28]. While little systematic analysis of imbalance and
methods to deal with it is available for deep learning, researchers employ some methods that might be addressing the problem likely based on intuition, some internal tests,
and systematic results available for traditional machine learning. Based on our review of
literature, the method most commonly applied in deep learning is oversampling.
The rest of this paper is organized as follows. Section 2 gives an overview of methods
to address the problem of imbalance. In Section 3 we describe the experimental setup.
It provides details about compared methods, datasets and models used for evaluation.
Then, in Section 4 we present the results from our experiments and compare methods.
Finally, Section 5 concludes the paper.
2
2
Methods for addressing imbalance
Methods for addressing class imbalance can be divided into two main categories [29]. The
first category is data level methods that operate on training set and change its class distribution. They aim to alter dataset in order to make standard training algorithms work.
The other category covers classifier (algorithmic) level methods. These methods keep the
training dataset unchanged and adjust training or inference algorithms. Moreover, methods that combine the two categories are available. In this section we give an overview
of commonly used approaches in both classical machine learning models and deep neural
networks.
2.1
Data level methods
Oversampling. One of the most commonly used method in deep learning [16, 30,
31, 32]. The basic version of it is called random minority oversampling, which simply
replicates randomly selected samples from minority classes. It has been shown that oversampling is effective, yet it can lead to overfitting [33, 34]. More advanced sampling
method that aims to overcome this issue is SMOTE [33]. It augments artificial examples
created by interpolating neighboring data points. Some extensions of this technique were
proposed, for example focusing only on examples near the boundary between classes [35].
Another type of oversampling approach uses data preprocessing to perform more informed
oversampling. Cluster-based oversampling first clusters the dataset and then oversamples each cluster separately [36]. This way it reduces both between-class and within-class
imbalance. DataBoost-IM, on the other hand, identifies difficult examples with boosting
preprocessing and uses them to generate synthetic data [37]. An oversampling approach
specific to neural networks optimized with stochastic gradient descent is class-aware sampling [38]. The main idea is to ensure uniform class distribution of each mini-batch and
control the selection of examples from each class.
Undersampling. Another popular method [16] that results in having the same number
of examples in each class. However, as opposed to oversampling, examples are removed
randomly from majority classes until all classes have the same number of examples. While
it might not appear intuitive, there is some evidence that in some situations undersampling
can be preferable to oversampling [39]. A significant disadvantage of this method is that
it discards a portion of available data. To overcome this shortcoming, some modifications
were introduced that more carefully select examples to be removed. E.g. one-sided
selection identifies redundant examples close to the boundary between classes [40]. More
general approach than undersampling is data decontamination that can involve relabeling
of some examples [41, 42].
2.2
Classifier level methods
Thresholding. Also known as threshold moving or post scaling, adjusts the decision
threshold of a classifier. It is applied in the test phase and involves changing the output
class probabilities. There are many ways in which the network outputs can be adjusted.
3
In general, the threshold can be set to minimize arbitrary criterion using an optimization algorithm [23]. However, the most basic version simply compensates for prior class
probabilities [43]. These are estimated for each class by its frequency in the imbalanced
dataset before sampling is applied. It was shown that neural networks estimate Bayesian
a posteriori probabilities [43]. That is, for a given datapoint x, their output for class i
implicitly corresponds to
yi (x) = p(i|x) =
p(i) · p(x|i)
.
p(x)
Therefore, correct class probabilities can be obtainedPby dividing the network output for
k |k|
each class by its estimated prior probability p(i) = |i|
, where |i| denotes the number
of unique examples in class i.
Cost sensitive learning. This method assigns different cost to misclassification of examples from different classes [44]. With respect to neural networks it can be implemented
in various ways. One approach is threshold moving [22] or post scaling [23] that are applied in the inference phase after the classifier is already trained. Similar strategy is to
adapt the output of the network and also use it in the backward pass of backpropagation
algorithm [45]. Another adaptation of neural network to be cost sensitive is to modify the
learning rate such that higher cost examples contribute more to the update of weights.
And finally we can train the network by minimizing the misclassification cost instead
of standard loss function [45]. The results of this approach are equivalent to oversampling [22, 26] described above and therefore this method will not be implemented in our
study.
One-class classification. In the context of neural networks it is usually called novelty detection. This is a concept learning technique that recognizes positive instances
rather than discriminating between two classes. Autoencoders used for this purpose are
trained to perform autoassociative mapping, i.e. identity function. Then, the classification of a new example is made based on a reconstruction error between the input and
output patterns, e.g. absolute error, squared sum of errors, Euclidean or Mahalanobis
distance [46, 47, 48]. This method has proved to work well for extremely high imbalance
when classification problem turns info anomaly detection [49].
Hybrid of methods. This is an approach that combines multiple techniques from one
or both abovementioned categories. Widely used example is ensembling. It can be viewed
as a wrapper to other methods. EasyEnsemble and BalanceCascade are methods that
train a committee of classifiers on undersampled subsets [50]. SMOTEBoost, on the other
hand, is a combination of boosting and SMOTE oversampling [51]. Recently introduced
and successfully applied to CNN training for brain tumor segmentation, is two-phase
training [28]. Even though the task was image segmentation, it was approached as a pixel
level classification. The method involves network pre-training on balanced dataset and
then fine-tuning the last output layer before softmax on the original, imbalanced data.
4
3
3.1
Experiments
Forms of imbalance
Class imbalance can take many forms particularly in the context of multiclass classification, which is typical in CNNs. In some problems only one class might be underrepresented or overrepresented and in other every class will have a different number of
examples. In this study we define and investigate two types of imbalance that we believe
are representative of most of the real-world cases.
The first type is step imbalance. In step imbalance, the number of examples is equal
within minority classes and equal within majority classes but differs between the majority
and minority classes. This type of imbalance is characterized by two parameters. One is
the fraction of minority classes defined by
µ=
|{i ∈ {1, . . . , N } : Ci is minority}|
,
N
(1)
where Ci is a set of examples in class i and N is the total number of classes. The other
parameter is a ratio between the number of examples in majority classes and the number
of examples in minority classes defined as follows.
ρ=
maxi {|Ci |}
mini {|Ci |}
(2)
An example of this type of imbalance is the situation when among the total of 10 classes,
5 of them have 500 training examples and another 5 have 5 000. In this case ρ = 10 and
µ = 0.5, as shown in Figure 1a. A dataset with the same number of examples in total
that has smaller imbalance ratio, corresponding to parameter ρ = 2, but more classes
being minority, µ = 0.9, is presented in Figure 1b.
(a) ρ = 10, µ = 0.5
(b) ρ = 2, µ = 0.9
(c) ρ = 10
Figure 1: Example distributions of imbalanced set together with corresponding values of parameters ρ and µ for step imbalance (a - b) and ρ for linear imbalance (c).
The second type of imbalance we call linear imbalance. We define it with one parameter that is a ratio between the maximum and minimum number of examples among all
classes, as in Equation 2 for imbalance ratio in step imbalance. However, the number of
examples in the remaining classes is interpolated linearly such that the difference between
consecutive pairs of classes is constant. An example of linear imbalance distribution with
ρ = 10 is shown in Figure 1c.
5
3.2
Methods of addressing imbalance compared in this study
In total, we examine seven methods to handle CNN training on a dataset with class
imbalance which cover most of the commonly used approaches in the context of deep
learning:
1. Random minority oversampling
2. Random majority undersampling
3. Two-phase training with pre-training on randomly oversampled dataset
4. Two-phase training with pre-training on randomly undersampled dataset
5. Thresholding with prior class probabilities
6. Oversampling with thresholding
7. Undersampling with thresholding
We examine two variants of two-phase training method. One on oversampled and the
other on undersampled dataset. For the second phase, we keep the same hyperparameters
and learning rate decay policy as in the first phase. Only the base learning rate from
the first phase is multiplied by the factor of 10−1 . Regarding thresholding, this method
originally uses the imbalanced training set to train a neural network. We, in addition,
combine it with oversampling and undersampling.
Selected methods are representative of the available approaches. Sampling can be
used to explicitly incorporate cost of the examples by their appearance. It makes them
one of many implementations of cost-sensitive learning [22]. Thresholding is another way
of applying cost-sensitiveness by moving the output threshold such that higher cost examples are harder to misclassify. Ensemble methods require training of multiple classifiers.
Because of considerable time needed to train deep models, it is often not practical and
may be even infeasible to train multiple deep neural networks. One-class methods have a
very limited application to datasets with extremely high imbalance. Moreover, they are
applied to anomaly detection problem that is beyond the scope of our study.
Importantly, we focused on methods that are widely used and relatively straightforward to implement as our aim is to draw conclusions that will be practical and serve as
a guidance to a large number of deep learning researchers and engineers.
3.3
Datasets and models
In our study, we used three benchmark datasets: MNIST [52], CIFAR-10 [53] and ImageNet Large Scale Visual Recognition Challenge (ILSVRC) 2012 [54]. All of them are
provided with a split on training and test set that are both labelled. For each dataset we
choose different model with a set of hyperparameters used for its training that is known to
perform well based on the literature. Datasets together with their corresponding models
are of increasing complexity. This allows us to draw some conclusions on simple task and
then verify how they scale to more complex ones.
All networks for the same dataset were trained with equal number of iterations. It
means that the number of epochs differs between the imbalanced versions of dataset. This
6
way we keep the number of weights’ updates constant. Also, all networks were trained
from a random initialization of weights and no pretraining was applied. An overview of
some information about the datasets and their corresponding models is given in Table 1.
All experiments were implemented in the deep learning framework Caffe [55].
Image dimensions
Images per class
Dataset
Width
Height
Depth
No. classes
Training
Test
CNN model
MNIST
CIFAR-10
ILSVRC-2012
28
32
≥ 256
28
32
≥ 256
1
3
3
10
10
1 000
5 000
5 000
1 000
1 000
1 000
50
LeNet-5
All-CNN
ResNet-10
Table 1: Summary of the used datasets. The number of images per class refers to the
perfectly balanced subsets used for experiments. Provided image dimensions for ImageNet
are given after rescaling.
3.3.1
MNIST
MNIST is considered simple and solved problem that involves digits’ images classification. The dataset consists of grayscale images of size 28 × 28. There are ten classes
corresponding to digits from 0 to 9. The number of examples per class in the original
training dataset ranges from 5421 in class 5 to 6742 in class 1. In artificially imbalanced
versions we uniformly at random subsample each class to contain no more than 5 000
examples.
Data dimensions
Layer
Width
Height
Depth
Kernel size
Stride
Input
Convolution
Max Pooling
Convolution
Max Pooling
Fully Connected
ReLU
Fully Connected
Softmax
28
24
12
8
4
1
1
1
1
28
24
12
8
4
1
1
1
1
1
20
20
50
50
500
500
10
10
5
2
5
2
-
1
2
1
2
-
Table 2: Architecture of LeNet-5 CNN used in MNIST experiments.
The CNN model that we use for MNIST is the modern version of LeNet-5 [52].
The network architecture is presented in Table 2. All networks for this dataset were
trained for 10 000 iterations. Optimization algorithm is stochastic gradient descent (SGD)
with momentum value of µ = 0.9 [56]. The learning rate decay policy is defined as
7
ηt = η0 · (1 + γ · t)−α , where η0 = 0.01 is a base learning rate, γ = 0.0001 and α = 0.75
are decay parameters and t is the current iteration. Furthermore, we used a batch size
of 64 and a weight decay value of λ = 0.0005. Network weights were initialized randomly
with uniform distribution and Xavier variance [57] whereas the biases were initialized
with zero. No data augmentation was used. Test error of the model trained as described
above on the original MNIST dataset was below 1%.
Experiments on MNIST dataset are performed on the following imbalance parameters
space. For linear imbalance we test values of ρ ∈ {10, 25, 50, 100, 250, 500, 1 000, 2 500, 5 000}.
For step imbalance the set of ρ values is the same and for each we use all possible number of
minority classes from 1 to 9, which corresponds to µ ∈ {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9}.
The experiment for each combination of parameters is repeated 50 times. Every time the
subset of minority classes is randomized. This way, we have created 4 050 artificially
imbalanced training sets for step imbalance and 450 for linear imbalance. As we have
evaluated four methods that require training a model, the total number of trained networks, including baseline, is 22 500.
3.3.2
CIFAR-10
CIFAR-10 is a significantly more complex image classification problem than MNIST. It
contains 32 × 32 color images with ten classes of natural objects. It does not have any
natural imbalance at all. There are exactly 5 000 training and 1 000 test examples in each
class. We do not use any data augmentation but follow standard preprocessing comprising
global contrast normalization and ZCA whitening [58].
Data dimensions
Layer
Width
Height
Depth
Kernel size
Stride
Padding
Input
Dropout (0.2)
2×(Convolution + ReLU)
Convolution + ReLU
Dropout (0.5)
2×(Convolution + ReLU)
Convolution + ReLU
Dropout (0.5)
Convolution + ReLU
Convolution + ReLU
Convolution + ReLU
Average Pooling
Softmax
32
32
32
16
16
16
8
8
6
6
6
1
1
32
32
32
16
16
16
8
8
6
6
6
1
1
3
3
96
96
96
192
192
192
192
192
10
10
10
3
3
3
3
3
1
1
6
-
1
2
1
2
1
1
1
-
1
1
1
1
0
0
0
-
Table 3: Architecture of All-CNN used in CIFAR-10 experiments.
For CIFAR-10 experiments we use one of the best performing type of CNN model on
this dataset, i.e. All-CNN [59]. The network architecture is presented in Table 3. The
8
networks were trained for 70 000 iterations using SGD with momentum µ = 0.9. The base
learning rate was multiplied by a fixed multiplier of 0.1 after 40 000, 50 000 and 60 000
iterations. The number of examples in a batch was 256 and a weight decay value was
λ = 0.001. Network weights were initialized with Xavier procedure and the biases set
to zero. Test error of the model trained as described above on the original CIFAR-10
dataset was 9.75%.
We have found the network training to be quite sensitive to initialization and the
choice of base learning rate. Sometimes the network gets stuck in a very poor local
minimum. Also, for more imbalanced datasets the training required lower base learning
rate to train at all. Therefore, for each case we were searching for the best one from the
fixed set η0 ∈ {0.05, 0.005, 0.0005, 0.00005}. Similar procedure was used by the authors
of the model architecture [59]. Moreover, each training was repeated twice on the same
dataset. For a particular method and imbalanced dataset, we pick the model with the
best score on the test set over all eight runs.
The network architecture does not have any fully connected layers. Therefore, during
the fine-tuning in two-phase training method we update the weights of two last convolutional layers with kernels of size 1.
The imbalance parameters space used in CIFAR-10 experiments is considerably sparser
than the one used for MNIST due to the significantly longer time required to train one
network. The set of tested values was narrowed to make the experiment run in a reasonable time. For linear and step imbalance, we test values of ρ ∈ {2, 10, 20, 50}. In step
imbalance, for each value of ρ, the set of values of parameter µ was µ ∈ {0.2, 0.5, 0.8},
which corresponds to having two, five and eight minority classes respectively. And for all
the cases, the classes chosen to be minority were the ones with the lowest label value. It
means that for a fixed number of minority classes the same classes were always picked as
minority. Also, all of them were included in a larger set of minority classes. In total we
trained 480 networks on this dataset.
3.3.3
ImageNet
For evaluation we use a ILSVRC-2012 competition subset of ImageNet, widely used as
a benchmark to compare classifiers’ performance. The number of examples in majority
classes was reduced from 1 200 to 1 000. Classes with less than 1 000 cases were always
chosen as a minority ones for imbalanced subsets. The only data preprocessing applied
is resizing such that the smaller dimension is 256 pixels long and the aspect ratio is
preserved. During training, as input we use a randomly cropped 224 × 224 pixel square
patch and a single centered crop in a test phase. Moreover, during training we randomly
mirror images, but there is no color, scale or aspect ratio augmentation.
A model architecture employed for this dataset is ResNet-10 [60], i.e. a residual
network [61] with batch normalization layers that are known to accelerate deep networks
training [62]. It consists of four residual blocks that give us nine convolutional layers
and one fully connected. The first residual block outputs data tensor of depth 64 and
then each one increases it by a factor of two. Fully connected layer outputs 1 000 values
to softmax that transforms them to class probabilities. The architecture of one residual
block is presented in Figure 2.
9
Figure 2: Architecture of a single residual block in ResNet used in ILSVRC-2012 experiments.
The networks were trained for 320 000 iterations using SGD with momentum µ = 0.9.
The base learning rate is set to η0 = 0.1 and decays linearly to 0 in the last iteration. The
number of examples in a batch was 256 and a weight decay λ = 0.0001. Network weights
were initialized with Kaiming (also known as MSRA) initialization procedure [63]. Top-1
test error of the model trained as described above on the original ILSVRC-2012 dataset
was 62.56% and 99.50 multi-class ROC AUC for a single centered crop.
We have chosen relatively small ResNet for the sake of faster training but without
loss of generality 1 . We test only one case of small and two cases of large step imbalance and run it on the baseline, undersampling and oversampling methods. Specifically,
all three step imbalanced subsets are defined with µ = 0.1, ρ = 10, µ = 0.8, ρ = 50 and
µ = 0.9, ρ = 100. They correspond to 100 minority classes with imbalance ratio of 10,
800 minority classes with imbalance of 50, and 900 minority classes with imbalance ratio
of 100 respectively. Moreover, for the highest imbalance, we train three networks for each
method with randomized selection of minority classes and subsampled set of examples in
each class. This is done in order to estimate variability in performance of methods. In
total, this gives us 15 ResNet-10 networks trained on five artificially imbalanced subsets
of ILSVRC-2012.
3.4
Evaluation metrics and testing
The metric that is most widely used to evaluate a classifier performance in the context of multiclass classification with CNNs is overall accuracy which is the proportion of
test examples that were correctly classified. However, it has some significant and long
acknowledged limitations, particularly in the context of imbalanced datasets [19]. Specifically, when the test set is imbalanced, accuracy will favor classes that are overrepresented
in some cases leading to highly misleading assessment. An example of this is a situation
when the majority class represents 99% of all cases and the classifier assigns the label
of the majority class to all test cases. A misleading accuracy of 99% will be assigned
to a classifier that has a very limited use. Another issue might arise when the test set
is balanced and a training set is imbalanced. This might result in a situation when a
decision threshold is moved to reflect the estimated class prior probabilities and cause a
low accuracy measure in the test set while the true discriminative power of the classifier
does not change.
A measure that addresses these issues is area under the receiver operating characteristic curve (ROC AUC) [64]. It is a well-studied and sound measure of discrimination [65]
1
It takes five days to train one ResNet-10 network on Nvidia GTX 1070 GPU.
10
and has been widely used as an evaluation metric for classifiers. ROC has also been used
to compare performance of classifiers trained on imbalanced datasets[20, 18]. Since the
basic version of ROC is only suitable for binary classification, we use a multi-class modification of it by averaging ROC AUCs for the binary classification task of distinguishing
a given class vs. all the other classes [66].
Test set of all used datasets has equal number of examples in each class. Usually, it is
assumed that the class distribution of a test set follows the one of a training set. We do
not change a test set to match artificially imbalanced training set. The reason is that the
score achieved by each classifier on the same test set is more comparable and the largest
number of cases in each of the classes provides the most accurate performance estimation.
4
4.1
Results
Effects of class imbalance on classification performance and comparison of methods to address imbalance
The results showing the impact of class imbalance on classification performance and
comparison of methods for addressing imbalance are shown in Figures 3 and 4. Figure 3
shows the results with respect to multi-class ROC AUC for a fixed number of minority
classes on MNIST and CIFAR-10. Figure 4 presents the result from the perspective of
fixed ratio of imbalance, i.e. parameter ρ, for the same two datasets.
(a) 2 minority classes
(b) 5 minority classes
(c) 8 minority classes
(d) 2 minority classes
(e) 5 minority classes
(f) 8 minority classes
Figure 3: Comparison of methods with respect to multi-class ROC AUC on MNIST (a - c) and
CIFAR-10 (d - f) for step imbalance with fixed number of minority classes.
11
Regarding the effect of class imbalance on classification performance, we observed the
following. First, the deterioration of performance due to class imbalance is substantial.
As expected, the increasing ratio of examples between majority and minority classes
as well as the number of minority classes had a negative effect on performance of the
resulting classifiers. Furthermore, by comparing the results from MNIST and CIFAR-10
we observed that the effect of imbalance is significantly stronger for the task with higher
complexity. A similar drop in performance for MNIST and CIFAR-10 corresponded to
approximately 100 times stronger level of imbalance in the MNIST dataset.
Regarding performance of different methods for addressing imbalance, in almost all
of the situations oversampling emerged as the best method. It also showed notable improvement of performance over the baseline (i.e. do-nothing strategy) in majority of the
situations and never showed a considerable decrease in performance for the two datasets
analyzed in this section making it a clear recommendation for tasks similar to MNIST
and CIFAR-10.
Undersampling showed a generally poor performance. In a large number of analyzed
scenarios undersampling showed decrease in performance as compared to the baseline.
In scenarios with a large proportion of minority classes undersampling showed some improvement over the baseline but never a notable advantage over oversampling (Figure 3).
(a) Imbalance ratio of 100
(b) Imbalance ratio of 500
(c) Imbalance ratio of 1 000
(d) Imbalance ratio of 10
(e) Imbalance ratio of 20
(f) Imbalance ratio of 50
Figure 4: Comparison of methods with respect to multi-class ROC AUC on MNIST (a - c) and
CIFAR-10 (d - f) for step imbalance with fixed imbalance ratio.
For a fixed imbalance ratio undersampling is always trained on the subset of equal
size. As a result, its performance does not change with the number of minority classes.
12
For both datasets and each case of imbalance ratio, the gap between undersampling and
oversampling is the biggest for smaller number of minority classes and decreases with the
number of minority classes, as shown in Figure 4. This is expected since with all classes
being minority these two methods become equivalent.
Two-phase training methods with both undersampling and oversampling tend to perform between the baseline and their corresponding method (undersampling or oversampling). If the baseline is better than one of these methods, fine-tuning improves the
original method. Otherwise, performance deteriorates. However, if the baseline is better,
there is still no gain from using two-phase training method. As oversampling is almost
always better than the baseline, fine-tuning always gives lower score.
(a) MNIST
(b) CIFAR-10
Figure 5: Comparison of methods with respect to multi-class ROC AUC for linear imbalance.
In Figure 5 we show the results for linear imbalance on MNIST and CIFAR-10
datasets. The highest possible linear imbalance ratio for MNIST dataset is 5 000, which
means only one example in the most underrepresented class. However, even in this case
the decrease in performance according to multi-class ROC AUC score for the baseline
model is not significant, as shown in Figure 5a. Nevertheless, oversampling improves the
score on both datasets and for all tested values of ρ, whereas the score for undersampling
decreases approximately linearly with imbalance ratio.
4.2
Results on ImageNet dataset
The results from experiments performed on ImageNet (ILSVRC-2012) dataset confirm
the impact of imbalance on classifier’s performance. Table 4 compares methods with
respect to multi-class ROC AUC. The drop in performance for the largest tested imbalance was from 99 to 90, in terms of multi-class ROC AUC. The results confirm that the
oversampling approach performs consistently better than undersampling approach across
all scenarios. A small decrease in performance as compared to baseline was observed for
oversampling for extreme imbalances. Please note, however, that these results should be
treated with caution and not as strong evidence that oversampling is inferior for highly
complex tasks with extreme imbalance. The absolute difference in performance between
three runs with respect to multi-class ROC AUC was even higher than 4 (for undersampling). Therefore, differences of 1 - 2 might be due to variability of results between
different runs of neural networks. Moreover, the highest tested imbalanced training set
13
was only about 10% of the original ILSVRC-2012 introducing confounding issues such as
the optimal training hyperparameters for this significantly changed dataset. Therefore,
while these results indicate that caution should be taken when any sampling technique
is applied to highly complex tasks with extreme imbalances, it needs a more extensive
study devoted to this specific issue.
Method
µ = 0.1, ρ = 10
µ = 0.8, ρ = 50
Baseline
Oversampling
Undersampling
99.41
99.35
96.85
96.31
95.06
94.98
µ = 0.9, ρ = 100
90.74
88.38
88.35
90.46
88.39
84.08
90.05
88.17
83.74
Table 4: Comparison of results on ImageNet with respect to multi-class ROC AUC.
4.3
Separation of effects from reduced number of examples and class
imbalance
An important question that needs to be considered in the context of our study is whether
the decrease in performance for imbalanced datasets is merely caused by the fact that
our imbalanced datasets simply had fewer training examples or is it truly caused by the
fact that the datasets are imbalanced.
First, we notice that oversampling method uses the same amount of data as the
baseline. It only eliminates the imbalance which is enough to improve the performance
in almost all the cases. Still, it does not reach the performance of a classifier trained on
the original dataset. This is an indication that the effect of imbalance is not trivial.
Second, for some cases undersampling, which reduces the total number of cases performs better than the baseline (see Figures 3c and 3f). Moreover, there are even cases
when undersampling can perform on a par with oversampling. It means that, between two
sampling methods that eliminate imbalance, even using fewer data can be comparable.
In addition, for the same value of parameter ρ we have equal number of examples in
the training set for linear imbalance and step imbalance with µ = 0.5, which corresponds
to half of the classes being minority. The drop in performance is much higher for step
imbalance. This additionally demonstrates that not only the total number of examples
matters but also its distribution between classes.
4.4
Improving accuracy score with multi-class thresholding
While our focus is on ROC AUC, we also provide the evaluation of the methods based on
overall accuracy measure with results on step imbalance shown in Figure 6. As explained
in Section 3.4, accuracy has some known limitations and in some scenarios does not reflect
the discriminative power of a classifier but rather the prevalence of classes in the training
or test set. Nevertheless, it is still commonly used evaluation score [16] and therefore we
provide some results according to this metric.
Our results show that thresholding is an appropriate approach to take to offset the
prior probabilities of different classes learned by a network based on imbalanced datasets
14
and provided an improvement in overall accuracy. In general, thresholding worked particularly well when applied jointly with oversampling.
(a) 2 minority classes
(b) 5 minority classes
(c) 8 minority classes
(d) 2 minority classes
(e) 5 minority classes
(f) 8 minority classes
Figure 6: Comparison of methods with respect to accuracy on MNIST (a - c) and CIFAR-10
(d - f) for step imbalance with fixed number of minority classes.
Please note that thresholding does not have an actual effect on the ability of the
classifier to discriminate between a given class from another but rather helps to find a
threshold on the network output that guarantees a large number of correctly classified
cases. In terms of ROC, multiplying a decision variable by any positive number does not
change the area under the ROC curve. However, finding an optimal operating point on
the ROC curve is important when the overall number of correctly classified cases is of
interest.
4.5
Undersampling and oversampling to smaller imbalance ratio
The default version of oversampling is to increase the number of cases in the minority
classes so that the number matches the majority classes. Similarly, the default of undersampling is to decrease the number of cases in the majority classes to match the minority
classes. However, a more moderate version of these algorithms could be applied. For
the case of MNIST with imbalance ratio of 1 000 we have tried to gradually decrease the
imbalance with oversampling and undersampling. The results are shown in Figure 7.
The results show that the default version of oversampling was always the best. Any
reduction of imbalance improves the score regardless of the number of minority classes,
15
as shown in Figure 7a. For undersampling, in some cases of moderate number of minority
classes, intermediate levels of undersampling performed better than both full undersampling and the baseline.
(a) Oversampling
(b) Undersampling
Figure 7: Comparison of oversampling and undersampling to reduced imbalance ratios on MNIST
with original imbalance of 1 000.
Moreover, comparing undersampling and oversampling to reduced level of imbalance,
we can notice that for each case of oversampling there is a level to which we can apply
undersampling and achieve equivalent performance. However, that level is not known a
priori rendering oversampling still the method of choice.
4.6
Generalization of sampling methods
In some cases undersampling and oversampling perform similarly. In those cases, one
would probably prefer the model that generalizes better. For classical machine learning
models it was shown that oversampling can cause overfitting, especially for minority
classes [33]. As we repeat small number of examples multiple times, the trained model
fits them too well. Thus, according to this prior knowledge undersampling would be
a better choice. The results from our experiments do not confirm this conclusion for
convolutional neural networks.
(a) Baseline
(b) Oversampling
(c) Undersampling
Figure 8: Comparison of networks convergence between baseline and sampling methods. Training
on CIFAR-10 step imbalanced with 5 minority classes and imbalance ratio of 50.
16
In Figure 8 we compare the convergence of baseline and sampling methods for CIFAR-10 experiments with respect to accuracy. Both methods helped to train a better
classifier in terms of performance and generalization. They also made training more stable. However, in this case it is not true that oversampling leads to overfitting. Accuracy
does not decrease as the training progresses and the gap between accuracy on the training
and test sets is smaller for oversampling, Figure 8b, than undersampling, Figure 8c. This
observation also holds for MNIST and ImageNet datasets and other cases of imbalance.
5
Conclusions
In this study, we examined the impact of class imbalance on classification performance
of convolutional neural networks and investigated the effectiveness of different methods
of addressing the issue. We defined and parametrized two representative types of imbalance, i.e. step and linear. Then we subsampled MNIST, CIFAR-10 and ImageNet
(ILSVRC-2012) datasets to make them artificially imbalanced. We have compared common sampling methods, basic thresholding, and two-phase training.
The conclusions from our experiments related to the class imbalance are as follows.
• The effect of class imbalance on classification performance is detrimental.
• The influence of imbalance on classification performance increases with the scale of
a task.
• The impact of imbalance cannot be explained simply by the lower total number of
training cases and depends on the distribution of examples among classes.
Regarding the choice of a method to handle CNN training on imbalanced dataset we
conclude the following.
• The method that in most of the cases outperforms all others with respect to multiclass ROC AUC was oversampling.
• For extreme ratio of imbalance and large portion of classes being minority, undersampling performs on a par with oversampling.
• To achieve the best accuracy, one should apply thresholding to compensate for prior
class probabilities. A combination of thresholding with baseline and oversampling
is the most preferable, whereas it should not be combined with undersampling.
• Oversampling should be applied to the level that totally eliminates the imbalance,
whereas undersampling can perform better when the imbalance is only removed to
some extent.
• As opposed to some classical machine learning models, oversampling does not necessarily cause overfitting of convolutional neural networks.
17
References
[1] Jiuxiang Gu, Zhenhua Wang, Jason Kuen, Lianyang Ma, Amir Shahroudy, Bing
Shuai, Ting Liu, Xingxing Wang, and Gang Wang. Recent advances in convolutional
neural networks. arXiv preprint arXiv:1512.07108, 2015.
[2] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional
networks. In European conference on computer vision, pages 818–833. Springer,
2014.
[3] Yann LeCun, Bernhard Boser, John S Denker, Donnie Henderson, Richard E Howard,
Wayne Hubbard, and Lawrence D Jackel. Backpropagation applied to handwritten
zip code recognition. Neural computation, 1(4):541–551, 1989.
[4] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with
deep convolutional neural networks. In Advances in neural information processing
systems, pages 1097–1105, 2012.
[5] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for largescale image recognition. arXiv preprint arXiv:1409.1556, 2014.
[6] Grant Van Horn, Oisin Mac Aodha, Yang Song, Alex Shepard, Hartwig Adam, Pietro
Perona, and Serge Belongie. The inaturalist challenge 2017 dataset. arXiv preprint
arXiv:1707.06642, 2017.
[7] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba.
Sun database: Large-scale scene recognition from abbey to zoo. In Computer vision
and pattern recognition (CVPR), 2010 IEEE conference on, pages 3485–3492. IEEE,
2010.
[8] Brian Alan Johnson, Ryutaro Tateishi, and Nguyen Thanh Hoan. A hybrid pansharpening approach and multiscale object-based image analysis for mapping diseased pine
and oak trees. International journal of remote sensing, 34(20):6969–6982, 2013.
[9] Miroslav Kubat, Robert C Holte, and Stan Matwin. Machine learning for the detection of oil spills in satellite radar images. Machine learning, 30(2-3):195–215,
1998.
[10] Oscar Beijbom, Peter J Edmunds, David I Kline, B Greg Mitchell, and David Kriegman. Automated annotation of coral reef survey images. In Computer Vision and
Pattern Recognition (CVPR), 2012 IEEE Conference on, pages 1170–1177. IEEE,
2012.
[11] Jerzy W Grzymala-Busse, Linda K Goodwin, Witold J Grzymala-Busse, and Xinqun
Zheng. An approach to imbalanced data sets based on changing rule strength. In
Rough-Neural Computing, pages 543–553. Springer, 2004.
18
[12] Brian Mac Namee, Padraig Cunningham, Stephen Byrne, and Owen I Corrigan. The
problem of bias in training data in regression problems in medical decision support.
Artificial intelligence in medicine, 24(1):51–70, 2002.
[13] Philip K Chan and Salvatore J Stolfo. Toward scalable learning with non-uniform
class and cost distributions: A case study in credit card fraud detection. In KDD,
volume 1998, pages 164–168, 1998.
[14] Predrag Radivojac, Nitesh V Chawla, A Keith Dunker, and Zoran Obradovic. Classification and knowledge discovery in protein databases. Journal of Biomedical Informatics, 37(4):224–239, 2004.
[15] Claire Cardie and Nicholas Howe. Improving minority class prediction using casespecific feature weights. In ICML, pages 57–65, 1997.
[16] Guo Haixiang, Li Yijing, Jennifer Shang, Gu Mingyun, Huang Yuanyue, and Gong
Bing. Learning from class-imbalanced data: Review of methods and applications.
Expert Systems with Applications, 2016.
[17] Nathalie Japkowicz and Shaju Stephen. The class imbalance problem: A systematic
study. Intelligent data analysis, 6(5):429–449, 2002.
[18] Maciej A Mazurowski, Piotr A Habas, Jacek M Zurada, Joseph Y Lo, Jay A Baker,
and Georgia D Tourassi. Training neural network classifiers for medical decision
making: The effects of imbalanced datasets on classification performance. Neural
networks, 21(2):427–436, 2008.
[19] Nitesh V Chawla. Data mining for imbalanced datasets: An overview. In Data
mining and knowledge discovery handbook, pages 853–867. Springer, 2005.
[20] Marcus A Maloof. Learning when data sets are imbalanced and when costs are
unequal and unknown. In ICML-2003 workshop on learning from imbalanced data
sets II, volume 2, pages 2–1, 2003.
[21] Charles X Ling and Chenghui Li. Data mining for direct marketing: Problems and
solutions. In KDD, volume 98, pages 73–79, 1998.
[22] Zhi-Hua Zhou and Xu-Ying Liu. Training cost-sensitive neural networks with methods addressing the class imbalance problem. IEEE Transactions on Knowledge and
Data Engineering, 18(1):63–77, 2006.
[23] Steve Lawrence, Ian Burns, Andrew Back, Ah Chung Tsoi, and C Lee Giles. Neural
network classification and prior class probabilities. In Neural networks: tricks of the
trade, pages 299–313. Springer, 1998.
[24] Salman H Khan, Mohammed Bennamoun, Ferdous Sohel, and Roberto Togneri.
Cost sensitive learning of deep feature representations from imbalanced data. arXiv
preprint arXiv:1508.03422, 2015.
19
[25] Vidwath Raj, Sven Magg, and Stefan Wermter. Towards effective classification of
imbalanced data with convolutional neural networks. In IAPR Workshop on Artificial
Neural Networks in Pattern Recognition, pages 150–162. Springer, 2016.
[26] Yu-An Chung, Hsuan-Tien Lin, and Shao-Wen Yang. Cost-aware pre-training for
multiclass cost-sensitive deep learning. arXiv preprint arXiv:1511.09337, 2015.
[27] Shoujin Wang, Wei Liu, Jia Wu, Longbing Cao, Qinxue Meng, and Paul J Kennedy.
Training deep neural networks on imbalanced data sets. In Neural Networks
(IJCNN), 2016 International Joint Conference on, pages 4368–4374. IEEE, 2016.
[28] Mohammad Havaei, Axel Davy, David Warde-Farley, Antoine Biard, Aaron
Courville, Yoshua Bengio, Chris Pal, Pierre-Marc Jodoin, and Hugo Larochelle. Brain
tumor segmentation with deep neural networks. Medical image analysis, 35:18–31,
2017.
[29] Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering, 21(9):1263–1284, 2009.
[30] Gil Levi and Tal Hassner. Age and gender classification using convolutional neural
networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition Workshops, pages 34–42, 2015.
[31] Andrew Janowczyk and Anant Madabhushi. Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases. Journal of pathology
informatics, 7, 2016.
[32] Nicolas Jaccard, Thomas W Rogers, Edward J Morton, and Lewis D Griffin. Detection of concealed cars in complex cargo x-ray imagery using deep learning. Journal
of X-Ray Science and Technology, pages 1–17, 2016.
[33] Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer.
Smote: synthetic minority over-sampling technique. Journal of artificial intelligence
research, 16:321–357, 2002.
[34] Kung-Jeng Wang, Bunjira Makond, Kun-Huang Chen, and Kung-Min Wang. A
hybrid classifier combining smote with pso to estimate 5-year survivability of breast
cancer patients. Applied Soft Computing, 20:15–24, 2014.
[35] Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderline-smote: a new oversampling method in imbalanced data sets learning. Advances in intelligent computing,
pages 878–887, 2005.
[36] Taeho Jo and Nathalie Japkowicz. Class imbalances versus small disjuncts. ACM
Sigkdd Explorations Newsletter, 6(1):40–49, 2004.
[37] Hongyu Guo and Herna L Viktor. Learning from imbalanced data sets with boosting and data generation: the databoost-im approach. ACM Sigkdd Explorations
Newsletter, 6(1):30–39, 2004.
20
[38] Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective
learning of deep convolutional neural networks. In European Conference on Computer
Vision, pages 467–482. Springer, 2016.
[39] Chris Drummond, Robert C Holte, et al. C4.5, class imbalance, and cost sensitivity:
why under-sampling beats over-sampling. In Workshop on learning from imbalanced
datasets II, volume 11, pages 1–8, 2003.
[40] Miroslav Kubat, Stan Matwin, et al. Addressing the curse of imbalanced training
sets: one-sided selection. In ICML, volume 97, pages 179–186. Nashville, USA, 1997.
[41] Jack Koplowitz and Thomas A Brown. On the relation of performance to editing in
nearest neighbor rules. Pattern Recognition, 13(3):251–255, 1981.
[42] Ricardo Barandela, E Rangel, José Salvador Sánchez, and Francesc J Ferri. Restricted decontamination for the imbalanced training sample problem. In Iberoamerican Congress on Pattern Recognition, pages 424–431. Springer, 2003.
[43] Michael D Richard and Richard P Lippmann. Neural network classifiers estimate
bayesian a posteriori probabilities. Neural computation, 3(4):461–483, 1991.
[44] Charles Elkan. The foundations of cost-sensitive learning. In International joint
conference on artificial intelligence, volume 17, pages 973–978. Lawrence Erlbaum
Associates Ltd, 2001.
[45] Matjaz Kukar, Igor Kononenko, et al. Cost-sensitive learning with neural networks.
In ECAI, pages 445–449, 1998.
[46] Nathalie Japkowicz, Catherine Myers, Mark Gluck, et al. A novelty detection approach to classification. In IJCAI, volume 1, pages 518–523, 1995.
[47] Nathalie Japkowicz, Stephen Jose Hanson, and Mark A Gluck. Nonlinear autoassociation is not equivalent to pca. Neural computation, 12(3):531–545, 2000.
[48] Hoon Sohn, Keith Worden, and Charles R Farrar. Novelty detection using autoassociative neural network. In Symposium on Identification of Mechanical Systems:
international mechanical engineering congress and exposition, New York, NY, pages
573–580, 2001.
[49] Hyoung-joo Lee and Sungzoon Cho. The novelty detection approach for different
degrees of class imbalance. In Neural Information Processing, pages 21–30. Springer,
2006.
[50] Xu-Ying Liu, Jianxin Wu, and Zhi-Hua Zhou. Exploratory undersampling for classimbalance learning. IEEE Transactions on Systems, Man, and Cybernetics, Part B
(Cybernetics), 39(2):539–550, 2009.
21
[51] Nitesh V Chawla, Aleksandar Lazarevic, Lawrence O Hall, and Kevin W Bowyer.
Smoteboost: Improving prediction of the minority class in boosting. In European
Conference on Principles of Data Mining and Knowledge Discovery, pages 107–119.
Springer, 2003.
[52] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based
learning applied to document recognition. Proceedings of the IEEE, 1998.
[53] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny
images. 2009.
[54] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma,
Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet
large scale visual recognition challenge. International Journal of Computer Vision,
115(3):211–252, 2015.
[55] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross
Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture
for fast feature embedding. In Proceedings of the 22nd ACM international conference
on Multimedia, pages 675–678. ACM, 2014.
[56] Ning Qian. On the momentum term in gradient descent learning algorithms. Neural
networks, 12(1):145–151, 1999.
[57] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep
feedforward neural networks. In Aistats, volume 9, pages 249–256, 2010.
[58] Ian J Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron C Courville, and Yoshua
Bengio. Maxout networks. ICML (3), 28:1319–1327, 2013.
[59] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller.
Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806,
2014.
[60] Marcel Simon, Erik Rodner, and Joachim Denzler. Imagenet pre-trained models with
batch normalization. arXiv preprint arXiv:1612.01452, 2016.
[61] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning
for image recognition. In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 770–778, 2016.
[62] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network
training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.
[63] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings
of the IEEE international conference on computer vision, pages 1026–1034, 2015.
22
[64] Andrew P Bradley. The use of the area under the roc curve in the evaluation of
machine learning algorithms. Pattern recognition, 30(7):1145–1159, 1997.
[65] Charles X Ling, Jin Huang, and Harry Zhang. Auc: a statistically consistent and
more discriminating measure than accuracy. In IJCAI, volume 3, pages 519–524,
2003.
[66] Foster Provost and Pedro Domingos. Tree induction for probability-based ranking.
Machine learning, 52(3):199–215, 2003.
23
| 9cs.NE
|
A GORENSTEIN CRITERION FOR STRONGLY F -REGULAR AND
LOG TERMINAL SINGULARITIES
arXiv:1603.08422v2 [math.AC] 4 Sep 2016
ANURAG K. SINGH, SHUNSUKE TAKAGI, AND MATTEO VARBARO
Dedicated to Professor Craig Huneke on the occasion of his sixty-fifth birthday.
Abstract. A conjecture of Hirose, Watanabe, and Yoshida offers a characterization of when
a standard graded strongly F -regular ring is Gorenstein, in terms of an F -pure threshold.
We prove this conjecture under the additional hypothesis that the anti-canonical cover of
the ring is Noetherian. Moreover, under this hypothesis on the anti-canonical cover, we
give a similar criterion for when a normal F -pure (resp. log canonical) singularity is quasiGorenstein, in terms of an F -pure (resp. log canonical) threshold.
1. Introduction
Let R be an F -pure domain of positive characteristic, and a a nonzero proper ideal. The
F -pure threshold fpt(a) was defined by Watanabe and the second author of this paper [37];
it may be viewed as a positive characteristic analogue of the log canonical threshold, and is
an important measure of the singularities of the pair (Spec R, V (a)). For example, a local
ring (R, m) is regular if and only if fpt(m) > dim R − 1.
We say that R is strongly F -regular if fpt(a) > 0 for each nonzero proper ideal a of R.
It is well-known that each strongly F -regular ring is Cohen-Macaulay and normal; it is then
natural to ask: when is a strongly F -regular ring Gorenstein? Toward answering this in the
graded context, Hirose, Watanabe, and Yoshida proposed the following:
Conjecture 1.1. [20, Conjecture 1.1 (2)] Let R be a standard graded strongly F -regular ring,
with R0 an F -finite field of characteristic p > 0. Let m be the unique homogeneous maximal
ideal of R. Then fpt(m) = −a(R) if and only if R is Gorenstein.
We prove that the conjecture holds for many classes of (not necessarily strongly F -regular)
F -pure normal standard graded rings:
Theorem A (Corollaries 3.17, 4.12). Let R be an F -pure, normal, standard graded ring,
with R0 an F -finite field of characteristic p > 0. Let m denote the homogeneous
maximal
L
ideal of R. Set X = Spec R, and suppose that the anti-canonical cover
O
(−nK
X)
n>0 X
of X is a Noetherian ring. Then fpt(m) = −a(R) if and only if R is quasi-Gorenstein.
Note that a ring is Gorenstein if and only if it is Cohen-Macaulay
and quasi-Gorenstein.
L
Under the hypotheses of Theorem A, the anti-canonical cover n>0 OX (−nKX ) is known to
be Noetherian in each of the following cases:
(1) R is Q-Gorenstein,
A.K.S. was partially supported by NSF grants DMS 1162585 and DMS 1500613, S.T. by JSPS KAKENHI
26400039, and M.V. by Geometria delle Varietà Algebriche PRIN 2010S47ARA 003. The third author worked
on part of this project while he was visiting the Max Planck Institute for Mathematics of Bonn.
The authors are grateful to Yoshinori Gongyo, Alessandro De Stefani, Kei-ichi Watanabe, Ken-ichi Yoshida,
and the referees for helpful comments.
1
2
A. K. SINGH, S. TAKAGI and M. VARBARO
(2) R is a semigroup ring,
(3) R is a determinantal ring,
(4) R is a strongly F -regular ring of dimension at most three,
(5) R is a four-dimensional strongly F -regular ring, of characteristic p > 5.
We give two proofs of Theorem A: the first has the advantage that it can be adapted to
obtain results in the local setting, e.g., Theorem 3.14. The second proof, while limited to
the graded context, provides a technique for computing the numerical invariants at hand; see
Proposition 4.3 for the case of determinantal rings. We describe the two techniques, after
recalling some definitions:
Recall that a ring R of prime characteristic p is called F -finite if the Frobenius map
F : R −→ F∗ R is a finite map. Let R be a local or standard graded F -finite domain of
characteristic p > 0, and suppose that R is F -pure, i.e., the e-th iterated Frobenius map
e
F e : R −→ F∗e R with x 7−→ F∗e xp splits as an R-linear map for each e > 1. Given a nonzero
ideal a ( R, and an integer e > 1, set νe (a) to be the largest integer r > 0 such that there
exists a nonzero element c in ar for which the composite map
Fe
×F e c
R −−→ F∗e R −−−∗→ F∗e R,
where
e
e
x 7−→ F∗e xp 7−→ F∗e (cxp ),
splits as an R-module homomorphism. Then, fpt(a) is defined to be lime−→∞ νe (a)/pe .
The first proof of Theorem A uses an invariant c(a) that was originally introduced in [32]:
Given a nonzero ideal a ( R, this invariant is defined in terms of the Grothendieck trace of
the iterated Frobenius map Tre : F∗e ωR −→ ωR , where ωR is the canonical module of R. For
an F -pure normal graded ring (R, m), one has
fpt(m) 6 c(m) 6 −a(R),
with equality holding when R is a quasi-Gorenstein standard graded ring; see Propositions 3.5
and 3.6. Thus, it suffices to show that if fpt(m) = c(m), then R is quasi-Gorenstein. Generalizing the argument of [37, Theorem 2.7], we are indeed able to prove this when the
anti-canonical cover of R is Noetherian. We also use the invariant c(m) in answering another
question of Hirose, Watanabe, and Yoshida, [20, Question 6.7]; see Corollary 3.18.
Our second proof uses the so-called Fedder-type criterion: Writing the standard graded
ring R as S/I, for S a polynomial ring and I a homogeneous ideal, we characterize νe (m) in
e
terms of the ideal I [p ] :S I, and use this to show that −νe (m) equals the degree of a minimal
generator of the (1 − pe )-th symbolic power of ωR , see Theorem 4.1. Using this, we give
explicit computations of fpt(m) in many situations, e.g., for determinant rings and for QGorenstein rings, see Propositions 4.3 and 4.5. We also prove that if (R, m) is a Q-Gorenstein
normal domain, with index coprime to p, then the pair (R, mfpt(m) ) is sharply F -pure; see
Proposition 4.13.
Thus far we have discussed singularities in positive characteristic; we also prove an analogous result in characteristic zero. de Fernex-Hacon [10] extended the definition of log terminal
and log canonical singularities to the non-Q-Gorenstein setting, which can be regarded as the
characteristic zero counterparts of strongly F -regular and F -pure rings. Using their definition, we formulate a characteristic zero analogue of Theorem A as follows:
Theorem B (Corollary 5.13). Let R be a standard graded normal ring, with R0 an algebraically closed field of characteristic zero. Set m to be the homogeneous maximal ideal of R.
Assume that X := Spec R has log canonical singularities in the sense of de Fernex-Hacon; set
lct(m) = sup{t > 0 | (X, mt ) is log canonical in the sense of de Fernex-Hacon}.
A GORENSTEIN CRITERION FOR STRONGLY F -REGULAR RINGS
3
(1) Then lct(m) 6 −a(R).
L
(2) Suppose, in addition, that the anti-canonical cover
n>0 OX (−nKX ) is Noetherian.
Then lct(m) = −a(R) if and only if R is quasi-Gorenstein.
We remark that in the situation of Theorem B, the anti-canonical cover is Noetherian
whenever X has log terminal singularities in the sense of de Fernex-Hacon, or if R is QGorenstein. Thus, Theorem B gives an affirmative answer to a conjecture of De StefaniNúñez-Betancourt [11, Conjecture 6.9].
In order to prove Theorem B, we introduce a new invariant d(a) for an ideal a of a normal
variety X with Du Bois singularities, in terms of a variant of multiplier modules, see Definition 5.4. We are then able to employ the same strategy as in the first proof of Theorem A,
using d(a) in place of c(a).
Throughout this paper, all rings are assumed to be Noetherian (except possibly for anticanonical covers),
commutative, with unity. By a standard graded ring, we mean an N-graded
L
ring R = n>0 Rn , with R0 a field, such that R is generated as an R0 -algebra by finitely
many elements of R1 .
2. Preliminaries on F -singularities
In this section, we briefly review the theory of F -singularities. In order to state the
definitions, we first introduce the following notation:
Let R be a ring of prime characteristic p > 0. We denote by R◦ the set of elements of R
that are not in any minimal prime ideal. Given an R-module M and e ∈ N, the R-module
F∗e M is defined by the following two conditions: (i) F∗e M = M as an abelian group, and
e
(ii) the R-module structure of F∗e M is given by r · x := r p x for r ∈ R and x ∈ F∗e M . We
e
e
write elements of F∗ M in the form F∗ x with x ∈ M . The e-th iterated Frobenius map is the
e
R-linear map F e : R −→ F∗e R sending x to F∗e xp . We say that R is F -finite if the Frobenius
map is finite, that is, F∗1 R is a finitely generated R-module. When (R, m) is local, the e-th
iterated Frobenius map F e : R −→ F∗e R induces a map FHe i (R) : Hmi (R) −→ Hmi (R) for each i.
m
We recall the definition of classical F -singularities:
Definition 2.1. Let R be an F -finite reduced ring of prime characteristic p > 0.
(1) We say that R is F -pure if the Frobenius map R −→ F∗ R splits as an R-linear map.
(2) We say that R is strongly F -regular if for every c ∈ R◦ , there exists a power q = pe of
p such that the R-linear map R −→ F∗e R sending 1 to F∗e c splits.
(3) When (R, m) is local, we say that R is F -injective if FHmi (R) : Hmi (R) −→ Hmi (R) is
injective for each i. In general, we say that R is F -injective if the localization Rm is
F -injective for each maximal ideal m of R.
(4) When (R, m) is local, we say that R is F -rational if R is Cohen-Macaulay and if for
every c ∈ R◦ , there exists e ∈ N such that cFHe d (R) : Hmd (R) −→ Hmd (R) sending z to
m
cFHe d (R) (z) is injective. In general, we say that R is F -rational if the localization Rm
m
is F -rational for each maximal ideal m of R.
Next we generalize these to the pair setting, see [17, 30, 37]:
Definition 2.2. Let a be an ideal of an F -finite reduced ring R of prime characteristic p
such that a ∩ R◦ 6= ∅.
(1) Suppose that R is local. For a real number t > 0, the pair (R, at ) is sharply F -pure if
there exist q = pe and c ∈ a⌈t(q−1)⌉ such that the R-linear map R −→ F∗e R sending 1
4
A. K. SINGH, S. TAKAGI and M. VARBARO
to F∗e c splits. The pair (R, at ) is weakly F -pure if there exist infinitely many e ∈ N and
e
associated elements ce ∈ a⌊t(p −1)⌋ such that each R-linear map R −→ F∗e R sending 1
to F∗e ce splits.
When R is not local, (R, at ) is said to be sharply F -pure (resp. weakly F -pure) if the
localization (Rm , atm ) at m is sharply F -pure (resp. weakly F -pure) for every maximal
ideal m of R.
(2) Suppose that R is F -pure. Then the F -pure threshold fpt(a) of a is defined as
fpt(a) = sup{t ∈ R>0 | (R, at ) is weakly F -pure}.
(3) Suppose that R is a normal local domain and ∆ is an effective Q-divisor on X :=
Spec R. For a real number t > 0, the pair ((R, ∆); at ) is sharply F -pure if there exists
q = pe and c ∈ a⌈t(q−1)⌉ such that the R-linear map
R −→ F∗e OX (⌈(q − 1)∆⌉)
with 1 7−→ F∗e c
splits. The pair ((R, ∆); at ) is weakly F -pure if there exist infinitely many e ∈ N and
e
associated elements ce ∈ a⌊t(p −1)⌋ such that each R-linear map
R −→ F∗e OX (⌊(pe − 1)∆⌋)
with 1 7−→ F∗e ce
splits. If, in addition, a = OX , then we simply say that (X, ∆) is F -pure.
If (R, ∆) is F -pure, then the F -pure threshold fpt(∆; a) of a with respect to the pair
(R, ∆) is defined by
fpt(∆; a) = sup{t ∈ R>0 | ((R, ∆); at ) is weakly F -pure}.
Remark 2.3. (1) Sharp F -purity implies weak F -purity. When a = R, the sharp F -purity
and the weak F -purity of (R, at ) are equivalent to the F -purity of R. Suppose that R is
F -pure. It is easy to check that if (R, at ) is weakly F -pure with t > 0, then (R, at−ε ) is
sharply F -pure for every t > ε > 0 (cf. [30, Lemma 5.2]). Thus,
fpt(a) = sup{t ∈ R>0 | (R, at ) is sharply F -pure}.
(2) Our definition of the F -purity of (R, ∆) coincides with the one in [17, Definition 2.1].
The following is a standard application of Matlis duality, which we will use in Section 3.
Lemma 2.4 (cf. [17, Proposition 2.4]). Let (R, m) be a d-dimensional F -finite normal local
ring of characteristic p > 0, ∆ be an effective Q-divisor on X = Spec R and a be a nonzero
ideal of R. For any real number t > 0, the pair ((R, ∆); at ) is weakly F -pure if and only if
e
there exist infinitely many e ∈ N and associated elements ce ∈ a⌊t(p −1)⌋ such that
e
FX,∆
×c
e
e
ce FX,∆
: Hmd (ωX ) −−−→ Hmd (OX (⌊pe KX + (pe − 1)∆⌋)) −−→
Hmd (OX (⌊pe KX + (pe − 1)∆⌋))
e
is injective, where FX,∆
is the map induced by the R-linear map R −→ F∗e OX (⌊(pe − 1)∆⌋)
sending 1 to F∗e 1.
The following is a reformulation of the so-called “Fedder-type criterion,” that we will use
in Section 4.
Proposition 2.5. Let S = k[x1 , . . . , xn ] be a polynomial ring over an F -finite field k of
characteristic p > 0 and I be a homogeneous ideal of S. Suppose that R := S/I is F -pure.
Given an e ∈ N and a homogeneous ideal a ⊂ S containing I such that aR ∩ R◦ 6= ∅, we
define the integer νe (a) by
e
e
e
νe (a) := max{r > 0 | ar (I [p ] : I) 6⊂ (xp1 , . . . , xpn )}.
A GORENSTEIN CRITERION FOR STRONGLY F -REGULAR RINGS
5
(1) For a real number t > 0, the pair (R, (aR)t ) is sharply (resp. weakly) F -pure if and
only if νe (a) > ⌈(pe − 1)t⌉ for some e (resp. νe (a) > ⌊(pe − 1)t⌋ for infinitely many e).
(2) fpt(aR) = lime−→∞ νe (a)/pe .
Proof. It follows from [36, Lemma 3.9] (where the criterion for F -purity is stated in the local
setting, but the same argument works in the graded setting).
In order to generalize the definition of F -rational and F -injective rings to the pair setting,
we use the notion of at -tight closure and at -sharp Frobenius closure.
Definition 2.6. Let a be an ideal of a reduced ring R of prime characteristic p > 0 such
that a ∩ R◦ 6= ∅, and t > 0 be a real number.
t
(1) ([18, Definition 6.1]) For an ideal I ⊆ R, the at -tight closure I ∗a of I is defined to be
the ideal of R consisting of all elements x ∈ R for which there exists c ∈ R◦ such that
ca⌈t(q−1)⌉ xq ⊆ I [q] for all large q = pe .
t
(2) ([30, Definition 3.10]) For an ideal I ⊂ R, the at -sharp Frobenius closure I F ♯a of I is
defined to be the ideal of R consisting of all elements x ∈ R such that a⌈t(q−1)⌉ xq ⊆ I [q]
for all large q = pe .
t
(3) Suppose that (R, m) is local. The at -sharp Frobenius closure 0FH♯a
i (R) of the zero subm
module in Hmi (R) is defined to be the submodule of Hmi (R) consisting of all elements
z ∈ Hmi (R) such that a⌈t(q−1)⌉ FHe i (R) (z) = 0 in Hmi (R) for all large q = pe .
m
The following technical remark is useful for the study of the invariant c(a), which will be
introduced in Section 3.
Remark 2.7. Let (R, m) be an F -finite reduced local ring of characteristic p > 0. Let
Tre : F∗e ωR −→ ωR be the e-th iteration of the trace map on R, that is, the ωR -dual of
the e-th iterated Frobenius map F e : R −→ F∗e R. It then follows from an argument similar
t
to the proof of [16, Lemma 2.1] that 0FH♯a
d (R) = 0 if and only if
m
X
e
e
Tr (F∗e (a⌈t(p −1)⌉ ωR )) = ωR
e>e0
for every integer e0 > 0.
Definition 2.8. Let R be an F -finite Cohen-Macaulay reduced ring of prime characteristic
p > 0, a be an ideal of R such that a ∩ R◦ 6= ∅, and t > 0 be a real number.
t
(1) ([32, Definition 6.1]) When R is local, (R, at ) is said to be F -rational if J ∗a = J for
every ideal J generated by a full system of parameters for R.
t
(2) When R is local, (R, at ) is said to be sharply F -injective if J F ♯a = J for every ideal J
generated by a full system of parameters for R.
When R is not local, the pair (R, at ) is said to be F -rational (resp. sharply F -injective)
if the localization (Rm , atm ) at m is F -rational (resp. sharply F -injective) for every maximal
ideal m of R. When a = R, this definition coincides with the one in Definition 2.1.
We review basic properties of sharply F -injective pairs and F -rational pairs.
Lemma 2.9. Let R be an F -finite reduced ring of prime characteristic p > 0, a be an ideal
of R such that a ∩ R◦ 6= ∅, and t > 0 be a real number. Set d := dim R.
(1) Suppose that (R, m) is Cohen-Macaulay. Then the following are equivalent:
(a) (R, at ) is sharply F -injective (resp. F -rational).
6
A. K. SINGH, S. TAKAGI and M. VARBARO
t
t
(b) J F ♯a = J (resp. J ∗a = J) for an ideal J generated by a full system of parameters.
t
t
(c) 0FH♯a
= 0 (resp. 0∗a
d (R) = 0).
d
Hm
m (R)
(2) Suppose that R is F -rational.
(a) There exists a rational number t0 > 0 such that (R, at0 ) is sharply F -injective.
(b) If (R, at ) is sharply F -injective with t > 0, then (R, at−ε ) is F -rational for every
t > ε > 0.
F ♯at
(3) Suppose that (R, m) is local. If (R, at ) is sharply F -pure, then 0H
i (R) = 0 for every i.
m
t
F ♯a
When R is quasi-Gorenstein,
is sharply F -pure if and only if 0H
= 0.
d
m (R)
(4) Let (R, m) ֒−→ (S, n) be a flat local homomorphism of F -finite reduced local rings of
characteristic p > 0. Suppose that S/mS is a field which is a separable algebraic extension
t
F ♯(aS)t
of R/m. Then 0FH♯a
d (R) = 0 if and only if 0H d (S) = 0.
(R, at )
n
m
Proof. We may assume throughout that (R, m) is local. Let J be an ideal generated by a full
system of parameters for R.
(1) The F -rational case follows from [32, Lemma 6.3] and the sharp F -injective case follows
from an analogous argument.
(2) First we will show (a). Fix a nonzero element f ∈ a. Since R is F -rational, there exists
e0 ∈ N such that f FHe0d (R) : Hmd (R) −→ Hmd (R) is injective. Then for each n ∈ N, the map
m
f 1+p
e0 +···+p(n−1)e0
FHned0(R) : Hmd (R) −→ Hmd (R)
m
t0
F ♯a
is also injective. Set t0 = 1/(pe0 − 1) and let z ∈ 0H
d (R) . Since
m
f
1+pe0 +···+p(n−1)e0
FHned0(R) (z)
m
∈ at0 (p
ne0 −1)
FHned0(R) (z) = 0
m
for sufficiently large n, one has z = 0 by the injectivity of f 1+p
e0 +···+p(n−1)e0
FHned0(R) . It follows
m
t0
F ♯a
that 0H
d (R) = 0.
m
t−ε
Next we will show (b). Let x ∈ J ∗a . Since 1 is a parameter at−ε -test element by
[32, Lemma 6.8] (see [32, Definition 6.6] for the definition of parameter at−ε -test elements),
a⌈t(q−1)⌉ xq ⊆ a⌈(t−ε)q⌉ xq ⊆ J [q] for all sufficiently large q = pe . Then the sharp F -injectivity
t−ε
of (R, at ) implies that x ∈ J, that is, J ∗a = J.
t
t
e
(3) Let z ∈ 0FH♯a
i (R) . Since (R, a ) is sharply F -pure, there exist a sufficiently large q = p
e
m
and c ∈ a⌈t(p −1)⌉ such that the R-linear map R −→ F∗e R sending 1 to F∗e c splits, and
in particular, cFHe i (R) : Hmi (R) −→ Hmi (R) is injective. Then z has to be zero, because
m
e
t
= 0.
cFHe i (R) (z) ∈ a⌈t(p −1)⌉ FHe i (R) (z) = 0. That is, 0FH♯a
i
m
m
m (R)
For the latter assertion, suppose that R is quasi-Gorenstein. Then by [30, Theorem 4.1],
(R, at ) is sharply F -pure if and only if for infinitely many q = pe , there exists c ∈ a⌈t(q−1)⌉
such that cFHe d (R) : Hmd (R) −→ Hmd (R) is injective. Looking at the socle of Hmd (R), we see
m
t
that this condition is equivalent to saying that 0FH♯a
d (R) = 0.
m
(4) Since Hnd (S) does not change by passing to the completion of S, we may assume that
S is complete. Let TreR : F∗e ωR −→ ωR (resp. TreS : F∗e ωS −→ ωS ) denote the e-th iteration
of the trace map on R (resp. S). It then follows from the proof of [35, Lemma 1.5 (2)] that
TreR ⊗R S : F∗e ωR ⊗R S −→ ωR ⊗R S is isomorphic to TreS for each e ∈ N. By Remark 2.7,
A GORENSTEIN CRITERION FOR STRONGLY F -REGULAR RINGS
7
t
0FH♯a
d (R) = 0 if and only if
m
M
e>e0
TreR :
M
F∗e (a⌈t(p
e −1)⌉
ωR ) −→ ωR
e>e0
is surjective for every integer e0 > 0. Tensoring with S, we see that this condition is equivalent
to the surjectivity of
M
M
e
TreS :
F∗e (a⌈t(p −1)⌉ ωS ) −→ ωS
e>0
e>e0
F ♯(aS)t
for every e0 > 0, which holds by Remark 2.7 again if and only if 0H d (S) = 0.
n
3. Positive characteristic case I
We introduce a new invariant of singularities in positive characteristic, and study its basic
properties. Using this, we give a partial answer to Conjecture 1.1.
Definition 3.1. Let (R, m) be a d-dimensional F -finite F -injective local ring of characteristic
p > 0, a be an ideal of R such that a ∩ R◦ 6= ∅. For each integer i, the threshold ci (a) is
defined by
t
ci (a) = sup{t ∈ R>0 0FH♯a
i (R) = 0}.
m
Note that ci (a) = ∞ when Hmi (R) = 0. Also, we simply denote cd (a) by c(a).
Remark 3.2. Let R be an N-graded ring with R0 an F -finite field of characteristic p > 0,
and m the homogeneous maximal ideal of R. Then we can define ci (m) similarly, that is,
t
F ♯m
ci (m) = sup{t ∈ R>0 0H
i (R) = 0}.
m
Since
Hmi (R)
i
∼
(Rm ), we have the equality ci (m) = ci (mRm ).
= HmR
m
Lemma 3.3. Let the notation be the same as in Definition 3.1.
(1) If R is Cohen-Macaulay, then
c(a) = sup{t ∈ R>0 | (R, at ) is sharply F -injective}.
(2) If R is F -rational, then
c(a) = sup{t ∈ R>0 | (R, at ) is F -rational}.
(3) Suppose that R is F -pure. Then ci (a) > fpt(a) for each i. In addition, if R is quasiGorenstein, then c(a) = fpt(a).
(4) c(a) is less than or equal to the height ht a of a.
(5) Suppose that R is Cohen-Macaulay and the residue field R/m is infinite. If J ⊂ R is
a minimal reduction of m, then md+1−⌈c(m)⌉ ⊆ J.
(6) Suppose that R is Cohen-Macaulay. If c(m) > d− 1, then R is regular and in particular
c(m) = d.
Proof. (1) (resp. (2), (3)) follows from Lemma 2.9 (1) (resp. (2), (3)).
(4) Since the trace map commutes with localization, by Remark 2.7, c(a) 6 c(aRp ) for
every prime ideal p containing a. Localizing at a minimal prime of a, we may assume that
ht a = d. We can also assume by Lemma 2.9 (4) that the residue field R/m is infinite. Let
t
J be a minimal reduction of a, and we will show that 0FH♯a
d (R) 6= 0 for every
d (R) ⊇ (0 : J)Hm
m
8
A. K. SINGH, S. TAKAGI and M. VARBARO
t > d. Let z ∈ (0 : J)Hmd (R) . Note that if t > d, then J [q] ⊇ a⌈t(q−1)⌉ for all large q = pe
′
(because a⌈t(q−1)⌉ = J ⌈tq (q−1)⌉ an for some fixed n ∈ N, where t′q = t − n/(q − 1)). Then
a⌈t(q−1)⌉ FHe d (R) (z) ⊆ J [q] FHe d (R) (z) = FHe d (R) (Jz) = 0
m
m
m
F ♯at
for such q = pe , which implies that z ∈ 0H d (R) .
Pm
e
(5) It follows from Remark 2.7 that e>e0 Tre (F∗e (m⌈(c(m)−ε)(p −1)⌉ ωR )) = ωR for all e0 ∈
Z>0 and for all c(m) > ε > 0 (when c(m) = 0, we put ε = 0). Multiplying by md+1−⌈c(m)⌉ on
both sides, one has
X
e
md+1−⌈c(m)⌉ ωR = md+1−⌈c(m)⌉
Tre F∗e (m⌈(c(m)−ε)(p −1)⌉ ωR )
⊆
X
e>e0
⊆
X
e>e0
⊆
X
e>e0
e>e0
e
e
Tre F∗e (m⌈(c(m)−ε)(p −1)⌉+(d+1−⌈c(m)⌉)p ωR )
e
Tre F∗e (mdp ωR )
e
Tre F∗e (J [p ] ωR )
⊆ JωR
for sufficiently large e0 and for sufficiently small ε > 0. Since R/J is the Matlis dual of
ωR /JωR , this means that md+1−⌈c(m)⌉ ⊆ J.
(6) Let J be a minimal reduction of m; we may assume by Lemma 2.9 (4) that the residue
field R/m is infinite. It then follows from (5) that m = J, which means that m is generated
by at most d elements, that is, R is regular. If R is regular, then c(m) = fpt(m) = d by (3)
and [37, Theorem 2.7 (1)].
Example 3.4. Let S be the n-dimensional polynomial ring k[x1 , . . . , xn ] over an F -finite
field k. Let R = S (r) be the r-th Veronese subring of S and mR be the homogeneous maximal
ideal of R. Then fpt(mR ) = n/r and c(mR ) = ⌈n/r⌉.
When R is an N-graded ring, the i-th a-invariant ai (R) is defined by
ai (R) = max{n ∈ Z | [Hmi (R)]n 6= 0}
for each i. The following proposition can be viewed as an extension of [11, Theorem 4.3].
Proposition 3.5. Let R be an F -injective N-graded ring, with R0 an F -finite field of characteristic p > 0. Let m be the homogeneous maximal ideal of R. Then ci (m) 6 −ai (R)
for each i. In particular, if R is F -pure, then by Lemma 3.3 (3), one has the inequality
fpt(m) 6 −ai (R) for every integer i.
Proof. We may assume that Hmi (R) 6= 0. We will then show that ci (m) 6 −ai (R) + ε, that
F ♯m−ai (R)+ε
6= 0, for every ε > 0. Note that ai (R) 6 0, because R is F -injective. Let
is, 0H
i (R)
m
z ∈ [Hmi (R)]ai (R) be a nonzero element. Since ⌈(−ai (R) + ε)(q − 1)⌉ + ai (R)q > 0 for all
sufficiently large q = pe , one has
m⌈(−ai (R)+ε)(q−1)⌉ FHe i (R) (z) ⊆ [Hmi (R)]>0 = 0
m
F ♯m−ai (R)+ε
for such q, which means that z ∈ 0H i (R)
m
.
A GORENSTEIN CRITERION FOR STRONGLY F -REGULAR RINGS
9
We record two cases in which the above result can be strengthened to an equality:
Proposition 3.6. Let R be an F -injective standard graded ring with R0 an F -finite field of
characteristic p > 0. Let m denote the homogeneous maximal ideal of R. Suppose that one
of the following conditions is satisfied:
(1) R is Cohen-Macaulay,
(2) R is normal and quasi-Gorenstein.
Then c(m) = −a(R).
Proof. Let d be the dimension of R. First, we assume that the condition (1) holds and we will
prove that c(m) > −a(R)−ε for every ε > 0. We may assume by Lemma 2.9 (4) that R0 is an
infinite field. Let J be a minimal reduction of m. As R is a Cohen-Macaulay standard graded
ring, J is generated by a homogeneous regular sequence of degree one. Then md+a(R)+1 is
contained in J but md+a(R) is not.
−a(R)−ε
= J by Lemma 2.9 (1) and Lemma 3.3 (1). Let
It is enough to show that J F ♯m
−a(R)−ε
F
♯m
, and we may assume that the degree of x is less than or equal to d + a(R).
x∈J
By definition, m⌈(−a(R)−ε)(q−1)⌉ xq ⊆ J [q] for all sufficiently large q = pe . Thus,
xq ∈ (J [q] : J ⌈(−a(R)−ε)(q−1)⌉ ) ⊆ J [q] + J dq−⌈(−a(R)−ε)(q−1)⌉−d+1 .
The degree of xq is less than or equal to (d + a(R))q, but dq − ⌈(−a(R) − ε)(q − 1)⌉ − d + 1
is greater than (d + a(R))q for sufficiently large q, so xq has to lie in J [q] for such q. It then
−a(R)−ε
= J.
follows from the F -injectivity of R that x ∈ J, that is, J F ♯m
Next, we assume that the condition (2) holds and we will show that c(m) > −a(R). It
is enough to show by Lemma 3.3 (3) that fpt(m) > −a(R). Let X = Proj R. Since R is a
quasi-Gorenstein
normal standard graded ring, there exists a very ample divisor H on X such
L
0 (X, O (nH)) and K
that R =
H
X
X ∼ a(R)H. Note that X is globally F -split and
n>0
a(R) 6 0, because R is F -pure. It then follows from an argument similar to the proof of [31,
Theorem 4.3] that there exists an effective Cartier divisor D on X such that D ∼ (1−p)a(R)H
and that the composite map
OX −→ F∗ OX −→ F∗ OX (D) x 7−→ F∗ xp 7−→ F∗ (sxp )
splits as an OX -module homomorphism, where s is a defining section for D. This map induces
the R-linear map R −→ F∗ R with 1 7−→ F∗ s, which also splits. Since s belongs to m(1−p)a(R)
by the definition of D, the pair (R, m−a(R) ) is sharply F -pure. Thus, fpt(m) > −a(R).
Motivated by Conjecture 1.1, we propose the following conjecture.
Conjecture 3.7. Let (R, m) be an F -finite F -pure normal local ring of characteristic p > 0.
Then R is quasi-Gorenstein if and only if fpt(m) = c(m).
Remark 3.8. Conjecture 3.7 can fail if R is not normal. Indeed, [11, Example 5.3] and
Proposition 3.6 give a counterexample.
Conjecture 3.7 implies an extension of Conjecture 1.1.
Proposition 3.9. Let R be an F -pure normal standard graded ring, with R0 an F -finite field
of characteristic p > 0. Let m denote the homogeneous maximal ideal of R. Suppose that
Conjecture 3.7 holds for the localization Rm of R at m. Then R is quasi-Gorenstein if and
only if fpt(m) = −a(R).
10
A. K. SINGH, S. TAKAGI and M. VARBARO
Proof. The “only if” part immediately follows from Lemma 3.3 (3) and Proposition 3.6. We
will show the “if” part. Suppose that fpt(m) = −a(R). Then by Remark 3.2, Lemma 3.3 (3)
and Proposition 3.6,
−a(R) = fpt(m) 6 fpt(mRm ) 6 c(mRm ) = c(m) = −a(R),
which implies that fpt(mRm ) = c(mRm ). It then follows from Conjecture 3.7 that Rm is
quasi-Gorenstein, which is equivalent to saying that R is quasi-Gorenstein.
Theorem 3.10. Let (R, m) be an F -finite normal local ring of characteristic p > 0 and ∆ be
an effective Q-divisor on X := Spec R such that (X, ∆) is F -pure and KX + ∆ is Q-Cartier
of index r. If R is not quasi-Gorenstein, then
1
fpt(∆; m) + 6 c(m).
r
Proof. Let d be the dimension of R. For every fpt(∆; m) > ε > 0 (when fpt(∆; m) = 0,
put ε = 0), by the definition of fpt(∆; m) and Lemma 2.4, there exist q0 = pe0 and c
in m⌊(fpt(∆;m)−ε)(q0 −1)⌋ such that
e0
cFX,∆
: Hmd (ωX ) −→ Hmd (OX (⌊q0 KX + (q0 − 1)∆⌋))
is injective. We consider the following commutative diagram:
/ H d (ωX )
m
ωX × Hmd (R)
e
0
cFX,∆
/ H d (OX (⌊q0 KX + (q0 − 1)∆⌋)),
m
OX (⌊q0 KX + (q0 − 1)∆⌋) × Hmd (R)
where the left vertical map sends (x, z) to (cxq0 , FHe0d (R) (z)).
m
For each 1/r >
′
ε′
> 0, we will show that
′
F ♯mfpt(∆;m)+1/r−ε−ε
assertion. Let ξ ∈ 0H d (R)
m
F ♯mfpt(∆;m)+1/r−ε−ε
0H
d
m (R)
= 0, which implies the
, that is, there exists q1 ∈ N such that
′
m⌈(fpt(∆;m)+1/r−ε−ε )(q−1)⌉ FHe d (R) (ξ) = 0
m
for all q = pe > q1 . By the definition of weak F -purity, we may assume that q0 is sufficiently
large so that q0 > q1 and ε′ (q0 −1) > 2. Since R is not quasi-Gorenstein, r > 2 or ∆ is strictly
r = O (K )r is contained in the fractional
effective. In either case, the fractional ideal ωX
X
X
ideal mOX (r(KX + ∆)). Therefore, for all x ∈ ωX , one has
q0
cxq0 ∈ m⌊(fpt(∆;m)−ε)(q0 −1)⌋ ωX
⊆ m⌊(fpt(∆;m)−ε)(q0 −1)⌋+⌊(q0 −1)/r⌋ OX (⌊q0 KX + (q0 − 1)∆⌋)
′
⊆ m⌈(fpt(∆;m)+1/r−ε−ε )(q0 −1)⌉ OX (⌊q0 KX + (q0 − 1)∆⌋)
by the choice of q0 . Since q0 > q1 ,
cxq0 FHe0d (R) (ξ) ∈ m⌈(fpt(∆;m)+1/r−ε−ε )(q0 −1)⌉ OX (⌊q0 KX + (q0 − 1)∆⌋)FHe0d (R) (ξ)
′
m
m
= 0 in
Hmd (OX (⌊q0 KX
+ (q0 − 1)∆⌋)),
e0
and it then follows from the commutativity of the above diagram that cFX,∆
(xξ) = 0. The
e0
injectivity of the map cFX,∆
implies that xξ = 0 for all x ∈ ωX . This forces ξ to be zero,
fpt(∆;m)+1/r−ε−ε′
F ♯m
because ωX × Hmd (R) −→ Hmd (ωX ) is the duality pairing. Thus, 0H
d (R)
m
= 0.
A GORENSTEIN CRITERION FOR STRONGLY F -REGULAR RINGS
11
We give an example of a standard graded Cohen-Macaulay ring R that is F -pure, with
a(R) = −1 and fpt(m) = 0; this is based on [34]. The ring R is Q-Gorenstein, with index 2.
Example 3.11. Let k be a field of characteristic p ≡ 1 mod 4, and set
S = k[w, x, y, z]/(w4 + x4 + y 4 + z 4 ) .
By the characteristic assumption, the ring S is F -pure. Set R to be the k-subalgebra generated
by the monomials
w4 , w3 x, w2 x2 , wx3 , x4 ,
y 4 , y 3 z, y 2 z 2 , yz 3 , z 4 .
Then R is a direct summand of S as an R-module: one way to see this is to use the Z/4×Z/4grading on S under which deg w = (1, 0) = deg x, and deg y = (0, 1) = deg z, in which case R
is the subring of S generated by elements of degree (0, 0). It follows that R is F -pure, normal,
as well as Cohen-Macaulay.
The ring R has a standard grading under which each of the monomials displayed is assigned
degree one. Computing the socle modulo the system of parameters x4 , y 4 , z 4 , it follows that
a(R) = −1. By Proposition 3.6, we have c(m) = 1.
The fractional ideal
1
ωR = 2 2 (w3 x, w2 x2 , wx3 )(y 3 z, y 2 z 2 , yz 3 )
w x
is, up to isomorphism, the graded canonical module of R; its second symbolic power is
y2z2
R,
w 2 x2
so the ring R is Q-Gorenstein. Using Theorem 4.1, one checks that that νe (m) = 0 for each
e > 1. It follows that fpt(m) = 0.
(2)
ωR =
Corollary 3.12. Let (R, m) be an F -finite F -pure normal local ring of characteristic p > 0.
(1) Suppose that there exists an effective Q-divisor ∆ on X = Spec R such that KX + ∆
is Q-Cartier, (X, ∆) is F -pure and fpt(∆; m) = fpt(m). Then Conjecture 3.7 holds for
this R.
(2) If c(m) = 0, then R is quasi-Gorenstein.
Proof. (1) immediately follows from Theorem 3.10. We will show (2). Since R is F -pure,
then by [31, Theorem 4.3 (ii)], there exists an effective Q-divisor ∆ on X such that (R, ∆) is
sharply F -pure with KX + ∆ Q-Cartier. Then
0 6 fpt(∆; m) 6 fpt(m) 6 c(m) = 0,
and the assertion follows from (1).
When is the assumption of Corollary 3.12 (1) satisfied? If the pair (R, mfpt(m) ) is sharply
F -pure, then by a similar argument to the proof of [31, Theorem 4.3 (ii)], there exists an
effective Q-divisor ∆ on X such that ((R, ∆); mfpt(m) ) is sharply F -pure with KX + ∆ QCartier. Then fpt(∆; m) = fpt(m), that is, the assumption of Corollary 3.12 (1) is satisfied.
Question 3.13 (cf. [20, Question 3.6]). Let (R, m) be an F -finite F -pure normal local ring of
characteristic p > 0. When is the pair (R, mfpt(m) ) sharply or weakly F -pure?
We will show in Proposition 4.13 that if (R, m) is an F -pure Q-Gorenstein normal standard
graded ring over an F -finite field of characteristic p > 0 with Gorenstein index not divisible
by p, then (R, mfpt(m) ) is sharply F -pure.
We now prove the main result of this section:
12
A. K. SINGH, S. TAKAGI and M. VARBARO
Theorem 3.14. Let (R, m) be an F -finite
F -pure normal local ring of characteristic p > 0.
L
Suppose that the anti-canonical cover n>0 OX (−nKX ) of X := Spec R is Noetherian. Then
fpt(m) = c(m) if and only if R is quasi-Gorenstein.
L
Proof. Since
n>0 OX (−nKX ) is Noetherian, one can find an integer r > 1 satisfying the
following: for every q = pe , if we write q − 1 = rne + je with ne > 0 and r − 1 > je > 0, then
OX ((1 − q)KX ) = OX (−rKX )ne OX (−je KX ).
Suppose that R is not quasi-Gorenstein, and we will show that fpt(m) + 1r 6 c(m). Let
ϕ1 , . . . , ϕl be a system of generators for OX (−rKX ).
For every fpt(m) > ε > 0, there exist a sufficiently large q = pe and c ∈ m⌊(fpt(m)−ε)(q−1)⌋
such that the R-linear map R −→ F∗e R sending 1 to F∗e c splits (when fpt(m) = 0, put ε = 0
and c = 1). That is, there exists an R-linear map ϕ : F∗e R −→ R sending F∗e c to 1. It follows
from Grothendieck duality that there exists an isomorphism
Φ : HomR (F∗e R, R) ∼
= F∗e OX ((1 − q)KX ) = F∗e (OX (−rKX )ne OX (−je KX )) .
P
ml
1
We write Φ(ϕ) = m F∗e (ϕm
1 · · · ϕl ψm ) with ψm ∈ OX (−je KX ), where m runs through all
elements of {(m1 , . . . , ml ) ∈ Zl>0 | m1 + · · · + ml = ne }. Then
!
X
m
e
l
1
F∗e (ϕm
1 = ϕ(F∗e c) = Φ−1
1 · · · ϕl ψm ) (F∗ c)
=
X
m
Φ
m
−1
e
ml
1
F∗e (ϕm
1 · · · ϕl ψm ) (F∗ c).
Pl
Therefore, there exists m = (m1 , . . . , ml ) ∈ Zl>0 with
i=1 mi = ne such that
ml
e c) is a unit. Replacing c by a unit multiple, we may assume
1
ψ
·
·
·
ϕ
Φ−1 (F∗e (ϕm
))(F
m
∗
1
l
ml
1
that Φ(ϕ) = F∗e (ϕm
1 · · · ϕl ψm ).
Since each ϕi determines an effective divisor Di which is linearly equivalent to −rKX , the
ml
e
1
section F∗e (ϕm
1 · · · ϕl ψm ) lies in F∗ OX ((1 − q)KX − m1 D1 − · · · − ml Dl ) and we have the
following commutative diagram:
F∗e OX ((1 − q)KX − m1 D1 − · · · − ml Dl )
O
/ F e OX ((1 − q)KX )
∗
O
Φ
HomR (F∗e OX (m1 D1 + · · · + ml Dl ), R)
/ HomR (F e R, R),
∗
where the vertical maps are isomorphisms. Therefore, ϕ induces an R-linear map
F∗e OX (m1 D1 + · · · + ml Dl ) −→ R
sending F∗e c to 1. Then its Matlis dual
cF e : Hmd (ωR ) −→ Hmd (OX (qKX + m1 D1 + · · · + ml Dl ))
is injective. On the other hand, since R is not quasi-Gorenstein and rKX + Di ∼ 0, the
r = O (K )r is contained in mO (rK + D ) for each i = 1, . . . , l. Hence,
fractional ideal ωX
X
X
X
X
i
q
ωX
is contained in m⌊(q−1)/r⌋ OX (qKX +m1 D1 +· · ·+ml Dl ). It then follows from an analogous
argument to the proof of Theorem 3.10 that fpt(m) + 1r 6 c(m).
L
Remark 3.15. In the setting of Theorem 3.14, it is well-known that
n>0 OX (−nKX ) is
Noetherian if R is Q-Gorenstein, RL
is a normal semigroup ring or R is a determinantal ring.
We briefly explain the reason why n>0 OX (−nKX ) is Noetherian in the latter case.
A GORENSTEIN CRITERION FOR STRONGLY F -REGULAR RINGS
13
Let R be the determinantal ring k[T ]/I, where T is an m × n matrix of indeterminates
with m 6 n, and I is the ideal generated by the size t minors of T where 1 6 t 6 m. Then
the anti-canonical class of R is the class of the (n − m)-th symbolic power of the prime ideal p
generated by the size t − 1 minors of the first t − 1 rows of T by [7, Theorem 8.8]. Moreover,
the symbolic powers of p coincide with its ordinary powers by [7, Corollary 7.10], so the
anti-canonical cover is the Rees algebra of pn−m . In particular, it is Noetherian.
L
Another case where n>0 OX (−nKX ) is Noetherian is the following:
Corollary 3.16. Let X be a three-dimensional strongly F -regular variety over an algebraically closed field of characteristic p > 5 and x be a closed point of X. Then fpt(mx ) =
c(mx ) if and only if X is Gorenstein at x.
Proof. We may assume that X is affine. By [31, Theorem 4.3] and [17, Theorem 3.3], there
exists an effective Q-divisor ∆ on X such that KX + ∆ is Q-Cartier and (X, ∆) is strongly F regular and in particular is klt. Since the minimal model program
holds for three-dimensional
L
klt pairs in characteristic p > 5, the anti-canonical cover
n>0 OX (−nKX ) is Noetherian
(see for example [9, Theorem 2.28]). Thus, the assertion follows from Theorem 3.14.
A combination of Proposition 3.9, Theorem 3.14, Remark 3.15 and Corollary 3.16 gives an
extension of [20, Theorem 1.2 (2)]:
Corollary 3.17. Let R be an F -pure normal standard graded ring, with R0 an F -finite field
of characteristic p >L
0. Let m be the homogeneous maximal ideal of R. Suppose that the
anti-canonical cover
n>0 OX (−nKX ) of X := Spec R is Noetherian. This assumption is
satisfied, for example, in each of the following cases:
(1) R is Q-Gorenstein,
(2) R is a semigroup ring,
(3) R is a determinantal ring,
(4) R is a strongly F -regular ring of dimension at most three,
(5) R is a four-dimensional strongly F -regular ring and p > 5.
Then fpt(m) = −a(R) if and only if R is quasi-Gorenstein.
L
Proof. If
then the assertion follows from Proposition 3.9
n>0 OX (−nKX ) is Noetherian,
L
and Theorem 3.14. By Remark
3.15,
O
n>0 X (−nKX ) is Noetherian in the case of (1), (2)
L
and (3). We will explain why n>0 OX (−nKX ) is Noetherian in the case of (4) and (5).
Since two-dimensional strongly F -regular rings are Q-Gorenstein, we may assume that
dim X > 3. Also, since strong F -regularity is preserved under flat base change by [1, Theorem 3.6], we may assume that R0L
is an algebraically closed field. Let D be a very ample
divisor on Y := Proj R so that R = m>0 H 0 (Y, OY (mD)). It follows from [31, Theorem 1.1]
that Y is a normal projective variety of Fano type of dimension at most three. It is known
that the minimal model program holds for klt surfaces and also for three-dimensional klt
pairs in characteristic p > 5 (see [15, 2, 4]). Thus, applying essentially the same argument as
the proof of [3, Corollary 1.1.9], we can see that
M
MM
OX (−nKX ) ∼
H 0 (Y, OY (mD − nKY ))
=
n>0
is Noetherian.
m∈Z n>0
We also give an answer to [20, Question 6.7]. Before stating the result, we fix some notation.
Let M = Zd , N = HomZ (M, Z), and denote the duality pairing between MR := M ⊗Z R
14
A. K. SINGH, S. TAKAGI and M. VARBARO
and NR := N ⊗Z R by h−, −i : MR × NR −→ R. Let σ ⊂ NR be a strongly convex rational
polyhedral cone and denote its dual cone by σ ∨ . Let R = k[σ ∨ ∩ M ] be the affine semigroup
ring over a field k defined by σ and m be the unique monomial maximal ideal of R. The
Newton Polyhedron P (m) ⊆ MR of m is defined as the convex hull of the set of exponents
m ∈ M of monomials xm ∈ m. We define the function λm by
λm : σ ∨ −→ R
u 7−→ sup{λ ∈ R>0 | u ∈ λP (m)},
where we set λP (m) = σ ∨ if λ = 0, and denote
aσ (R) := − min{λm (u) | u ∈ Int(σ ∨ ) ∩ M }.
Note that aσ (R) coincides with the a-invariant a(R) if R is standard graded.
Corollary 3.18. We use the above notation. Let R = k[σ ∨ ∩ M ] be a (not necessarily
standard graded) affine semigroup ring over an F -finite field k of characteristic p > 0 defined
by σ.
(1) Then c(m) = −aσ (R).
(2) fpt(m) = −aσ (R) if and only if R is Gorenstein.
Proof. Since (2) follows from (1) and Theorem 3.14, we will show only (1). Let v1 , . . . , vs be
the primitive generators for σ, that is, the first lattice points on the edges of σ. Note that
the graded canonical module ωR consists of the monomials xm such that hm, vi i > 1 for all
i = 1, . . . , s. Hence, its k-dual Hmd (R) is written as
M
Hmd (R) =
kxm ,
m∈S
where S = {m ∈ M | hm, vi i 6 −1 for all i = 1, . . . , s}. It follows from the fact that 1 is
an mt -test element by [18, Theorem 6.4] (see [18, Definition 6.3] for the definition of mt test elements) that the pair (R, mt ) is F -rational if and only if for each m ∈ S, one has
e
e
m⌈tp ⌉ xp m 6= 0 in Hmd (R), or equivalently,
(pe m + ⌈tpe ⌉P (m)) ∩ S 6= ∅,
for infinitely many e. We can rephrase this condition as saying that −m ∈ Int(tP (m)), using
an argument similar to the proof of [5, Theorem 3]. By Lemma 3.3 (2),
c(m) = sup{t ∈ R>0 | (R, mt ) is F -rational}
= sup{t ∈ R>0 | −m ∈ Int(tP (m)) for all m ∈ S}
= min λm (u).
u∈−S
Since −S = Int(σ ∨ ) ∩ M , one has the equality c(m) = −aσ (R).
4. Positive characteristic case II
In this section we give a different interpretation of the function νe (m), where m is the
homogeneous maximal ideal of an F -pure normal standard graded domain R over an F -finite
field (Theorem 4.1). Combining it with the Fedder-type criteria (Proposition 2.5), we give
explicit computations of fpt(m) in many situations (e.g. Propositions 4.3 and 4.5), eventually
yielding Corollary 3.17 as a consequence (see Corollary 4.12).
A GORENSTEIN CRITERION FOR STRONGLY F -REGULAR RINGS
15
Theorem 4.1. Let S be an n-dimensional standard graded polynomial ring over an F -finite
field of characteristic p > 0. Let I be a homogeneous ideal such that R := S/I is an F -pure
normal domain. Let ωR denote the graded canonical module of R. Then, for each q = pe ,
one has a graded isomorphism
I [q] :S I ∼
(1−q)
.
= (ωR (n))
[q]
I
In particular, if m is the homogeneous maximal ideal of R, then −νe (m) equals the degree of
(1−q)
(See Proposition 2.5 for the definition of νe (m)).
a minimal generator of ωR
Proof. After taking a flat base change, we may assume that S = k[x1 , . . . , xn ], where k
is a perfect field. It then follows that S is a free S q -module with basis xi11 · · · xinn where
0 6 ij 6 q − 1 for each j. Consider the homomorphism ϕ ∈ HomS q (S, S q ) that maps the
basis element (x1 · · · xn )q−1 to 1, and every other basis element to 0. It is readily seen that ϕ
generates HomS q (S, S q ) as an S-module.
Let J be the ideal of S q consisting of q-th powers of elements of I; note that JS = I [q] .
Then
[q]
∼ I :S I ϕ,
HomS q (S/I, S q /J) =
I [q]
see [12, page 465]. Next, note that one has graded isomorphisms
HomS q (S/I, S q /J) ∼
= HomRq (R, Rq )
∼
= HomRq (R, HomRq (ωRq , ωRq ))
∼
= HomRq (R ⊗Rq ωRq , ωRq )
(q)
∼
= HomRq (ωR , ωRq )
(q)
∼
= HomRq (ωR ⊗R R, ωRq )
(q)
∼
= HomR (ωR , HomRq (R, ωRq ))
(q)
∼
= HomR (ωR , ωR )
(1−q)
∼
= ωR .
Since the homomorphism ϕ has degree n − nq, the desired isomorphism follows.
(1−q)
is generated in degrees −d1 < · · · < −dr , then (I [q] :S I)/I [q] is generated
Suppose ωR
in degrees n(q − 1) − d1 < · · · < n(q − 1) − dr . Hence the least degree of a homogeneous
element of I [q] :S I that is not in m[q] belongs to the set
{n(q − 1) − d1 , . . . , n(q − 1) − dr },
and
I [q]
:S I ⊆
mn(q−1)−d1 .
Since the definition of νe (m) translates as
νe (m) = max{r ∈ N | (I [q] :S I) 6⊆ m[q] + mn(q−1)+1−r },
it follows that νe (m) ∈ {d1 , . . . , dr }.
As an immediate consequence we get:
Corollary 4.2. If R is an F -pure quasi-Gorenstein standard graded normal domain, over
an F -finite field, with homogeneous maximal ideal m, then
fpt(m) = −a(R).
In particular, by Proposition 3.5, in this case a(R) > ai (R) for all i = 0, . . . , dim R.
16
A. K. SINGH, S. TAKAGI and M. VARBARO
Proposition 4.3. Let R be the determinantal ring k[T ]/I, where k is an F -finite field, the
matrix of indeterminates T has size m × n with m 6 n, and I is the ideal generated by the
size t minors of T where 1 6 t 6 m. Let m be the homogeneous maximal ideal of R. Then
fpt(m) = m(t − 1).
Remark 4.4. In the notation of the proposition, the ring R has a-invariant −n(t−1). It follows
that fpt(m) = −a(R) precisely when m = n or t = 1, i.e., if and only if R is Gorenstein.
In the case t = 2, the F -pure threshold has been calculated previously, see [8, Corollary 1]
or [20, Example 6.2]. Since I is a homogeneous ideal of k[T ], which is F -pure, one can also
ask for fpt(I). This threshold has been computed in [28] (see [19] for various generalizations):
(m − l)(n − l)
fpt(I) = min
l = 0, . . . , t − 1 .
t−l
Proof of Proposition 4.3. The graded canonical module of R is computed in [6, Corollary 1.6],
namely, it equals qn−m (m − mt), where q is the prime ideal generated by the size t − 1 minors
of the first t − 1 columns of the matrix T . The divisor class group of R is described by
(1−q)
is generated by elements of degree
[7, Corollary 7.10], from which it follows that ωR
−m(q − 1)(t − 1). Theorem 4.1 now gives
νe (m) = m(q − 1)(t − 1),
from which the result follows.
Proposition 4.5. Let R be an F -pure Q-Gorenstein standard graded normal domain, over
an F -finite field, with homogeneous maximal ideal m. If c is the order of ωR in the divisor
class group and ω (c) is generated in degree D, then:
fpt(m) = D/c.
In particular, fpt(m) = −a(R) if and only if R is quasi-Gorenstein.
Proof. For q = pe let us write 1 − q = a(q)c + b(q), with 0 6 b(q) < c. By the assumptions
we have:
a(q)
(b(q))
(1−q)
= ω (c)
ωR .
ωR
(1−q)
In particular, ωR
is generated in degrees d satisfying:
a(q)D + A 6 d 6 a(q)D + B,
(b(q))
where the minimal generators of ωR
have degrees between A and B. Therefore
−a(q)D − B 6 νe (m) 6 −a(q)D − A,
and fpt(m) = limq−→∞ νe (m)/q = D/c.
(c)
For the last part of the statement, simply notice that, if ωR is principal but ωR is not,
(c)
(c)
c.
then the generator of ωR must have degree less than −a(R)c, since ωR ⊆ ωR
Remark 4.6. In the above notation, if c = p notice that a(q) = −q/p and b(q) = 1. Furthermore A can be chosen to be the negative of the a-invariant of R, so:
νe (m) 6 (q/p)D + a(R) = (q − 1)D/p + D/p + a(R) = (q − 1)fpt(m) + fpt(m) + a(R).
Given a finitely generated graded R-module M , we denote by δ(M ) the least integer d
such that Md 6= 0. In the case in which R is a normal domain, the canonical module ωR is
isomorphic (as a graded module) to a divisorial ideal of R.
A GORENSTEIN CRITERION FOR STRONGLY F -REGULAR RINGS
17
Lemma 4.7. If R is a normal standard graded domain, and a a graded divisorial ideal, then:
(1) δ(a(−1) ) > −δ(a).
(2) δ(a(−1) ) = −δ(a) if and only if a is principal.
Proof. Let a = (a1 , . . . , ar ), where the ai are homogeneous elements of the quotient field of
R of degree di ∈ Z, where δ(a) = d1 6 d2 6 . . . 6 dr . If b ∈ a(−1) is a homogeneous nonzero
element of degree l, then l + d1 > 0 since a1 b is a homogeneous nonzero element of R. This
shows (1).
Concerning point (2), if b is a homogeneous nonzero element of a(−1) of degree −d1 , then
ba1 = u is a unit of R. Because bai = fi ∈ R for all i = 1, . . . , r, we have ai = u−1 fi a1 for all
i = 1, . . . , r, so that a1 generates a as an R-module.
Proposition 4.8. Let R be an F -pure standard graded normal domain, over an F -finite
field, with homogeneous maximal ideal m. Then:
νe (m) 6 a(R)(1 − q)
∀ q = pe .
In particular fpt(m) 6 −a(R). Further, if R is not quasi-Gorenstein, then:
∀ q = pe .
(q−1)
q−1
= δ(ωR )(q−1) =
6 δ ωR
Proof. For the first part of the statement, notice that δ ωR
(1−q)
> a(R)(q − 1) by Lemma 4.7, and νe (m) 6 a(R)(1 − q) by
−a(R)(q − 1), so δ ωR
Theorem 4.1.
(q−1)
6
For the second part, assume that νe (m) = a(R)(1 − q) for some q = pe . Since δ ωR
νe (m) < a(R)(1 − q)
(q−1)
a(R)(1 − q), by putting together Theorem 4.1 and Lemma 4.7, ωR
generated
Therefore
mustbe principal and
q−1
= a(R)(1 − q).
ωR
(q−1)
q−1
in degree −a(R)(q − 1). Notice that ωR
⊇ ωR
and δ
(q−1)
q−1
= ωR
, so that ωR must
the only possibility is that ωR
be principal itself.
Proposition 4.9. Let R be an F -pure standard graded normal domain, over an F -finite
L
(−k)
of R is
field, with homogeneous maximal ideal m. If the anti-canonical cover
k>0 ωR
Noetherian, then fpt(m) = −a(R) if and only if R is quasi-Gorenstein.
Proof. If the anti-canonical cover of R is Noetherian, then there exists a positive integer c
such that, if we write 1 − q = −a(q)c − b(q) with a(q) positive and 0 6 b(q) < c:
a(q)
(−b(q))
(1−q)
.
ωR
= ω (−c)
ωR
Let us say that ω (−c) is generated in degrees −d1 < · · · < −dr . Furthermore, let −e1 < · · · <
−es be the degrees of the minimal generators of R, ω (−1) , . . . , ω (−c+1) . We have that
νe (m) ∈ {a(q)di + ej | i = 1, . . . , r and j = 1, . . . , s} ∀ q = pe .
By choosing i ∈ {1, . . . , r} such that νe (m) = a(q)di + ej (for some j) for infinitely many q,
then
fpt(m) = lim νe (m)/q = di /c.
q−→∞
For the second part of the statement, simply note that if R is not quasi-Gorenstein, with the
above notation we have −di > −d1 > a(R)c by (the same argument of) Proposition 4.8.
Remark 4.10. The above argument shows also that fpt(m) is a rational number whenever the
assumptions of the corollary are satisfied; this was already known by [9].
18
A. K. SINGH, S. TAKAGI and M. VARBARO
Proposition 4.11. Let R be an F -pure standard graded normal domain, over an F -finite
field, with
homogeneous maximal ideal m. If there exists a positive integer c such that
δ ω (c) < −a(R)c, then fpt(m) < −a(R).
Proof. Let δ ω (c) = D. With the same notation of the proof above q − 1 = a(q)c + b(q), so:
(b(q))
(q−1)
.
δ(ωR ) 6 a(q)D + δ ωR
Then, using Theorem 4.1 together with Lemma 4.7, for such q:
νe (m) 6 a(q)D.
Thus, fpt(m) = limq−→∞ νe (m)/q 6 D/c < −a(R).
The following provides strong evidence for the conjecture of Hirose-Watanabe-Yoshida 1.1
and, more generally, for the standard graded case of Conjecture 3.7.
Corollary 4.12. Let R be an F -pure standard graded normal domain, over an F -finite field,
with homogeneous maximal ideal m. Suppose that one of the following is satisfied:
L
(−k)
(1) The anti-canonical cover k>0 ωR of R is noetherian.
(c)
(2) For some positive integer c, there is a nonzero element of ωR of degree < −a(R)c.
Then fpt(m) = −a(R) if and only if R is quasi-Gorenstein.
The following gives an extension of [20, Proposition 3.4].
Proposition 4.13. Let R be an F -pure standard graded normal domain, over an F -finite
field, with homogeneous maximal ideal m. Suppose that the c-th Veronese subring of the
anti-canonical cover of R is standard graded and c is not a multiple of p. Then νe (m) =
(pe − 1)fpt(m) for infinitely many positive integers e. In particular, (R, mfpt(m) ) is sharply
F -pure.
Proof. Since p does not divide c, there is an infinite subset A ⊆ {pe | e ∈ N} such that
q − 1 = a(q)c for all q ∈ A, with a(q) ∈ N. For such q
a(q)
(1−q)
= ω (−c)
.
ωR
So if ω (−c) is generated in degrees −d1 < · · · < −dr , then
νe (m) ∈ {a(q)di | i = 1, . . . , r}
∀ q ∈ A.
So there exists i such that fpt(m) = di /c, and for all but finitely many q ∈ A
νe (m) = (q − 1)fpt(m).
5. Characteristic zero case
Throughout this section, let X be a normal variety over an algebraically closed field of
characteristic zero and a be a nonzero coherent ideal sheaf on X.
We prove a characteristic zero analogue of Conjecture 1.1. First, we define a variant of
multiplier submodules:
A GORENSTEIN CRITERION FOR STRONGLY F -REGULAR RINGS
19
Definition 5.1. Let π : Y −→ X be a log resolution of (X, a), that is, π is a proper birational
morphism from a smooth variety Y such that aOY = OY (−F ) is invertible and Exc(π) ∪
Supp(F ) is a simple normal crossing divisor. Let E be the reduced divisor supported on
Exc(π). For a real number t > 0, the multiplier submodule J (ωX , at ) is defined by
J (ωX , at ) := π∗ ωY (⌈−tF ⌉) ⊆ ωX .
This submodule of ωX is independent of the choice of π, see, for example, the proof of [32,
Proposition 3.4]. When a = OX or t = 0, we simply denote J (ωX , at ) by J (ωX ).
As a variant of J (ωX , at ), we define the submodule I(ωX , at ) of ωX by
π∗ ωY (⌈εE − (t − ε)F ⌉) if t > 0
I(ωX , at ) :=
π∗ ωY (E)
if t = 0
for sufficiently small ε > 0. It is easy to see that I(ωX , at ) is independent of the choice of ε
if ε > 0 is sufficiently small. When a = OX or t = 0, we simply denote I(ωX , at ) by I(ωX ).
Lemma 5.2. I(ωX , at ) is independent of the choice of the resolution.
Proof. Although it immediately follows from [14, Lemma 13.3 and Corollary 13.7], we give a
more direct proof here.
We consider the case where t > 0; the case t = 0 follows from a similar argument. Let
f : Y −→ X be a log resolution of (X, a) such that aOX = OX (−F ) is invertible, and
let EY be the reduced divisor supported on Exc(f ). Let g : Z −→ Y be a log resolution
of (Y, EY + F ), and let EZ be the reduced divisor supported on Exc(g). Then it is enough
to show that
ωY (⌈εEY − (t − ε)F ⌉) = g∗ ωZ (⌈ε′ (g∗−1 EY + EZ ) − (t − ε′ )g∗ F ⌉)
for sufficiently small real numbers ε,Sε′ > 0, since two log resolutions of (X, a) can be dominated by a third log resolution. Let i Ei be the irreducible decomposition of Supp(EY + F ).
For sufficiently small ε > 0, we can write
X
⌈KY + εEY − (t − ε)F ⌉ = KY − tF +
ai Ei ,
i
P
where
1
>
a
>
0
for
all
i.
Since
i
i Ei is a simple normal crossing divisor on Y , the pair
P
(Y, i ai Ei ) is log canonical. By the definition of log canonical pairs, we have
G :=⌈KZ + ε′ (g∗−1 EY + EZ ) − (t − ε′ )g ∗ F ⌉ − g∗ ⌈KY + εEY − (t − ε)F ⌉
'
&
X
′
−1
∗
∗
ai Ei > 0.
= KZ/Y + ε g∗ EY + EZ + g F − g
i
Note that G is a g-exceptional divisor for sufficiently small ε′ > 0. Therefore,
g∗ ωZ (⌈ε′ (g∗−1 EY + EZ ) − (t − ε′ )g∗ F ⌉) =g∗ (g∗ (ωY (⌈εEY − (t − ε)F ⌉)) ⊗ OZ (G))
=ωY (⌈εEY − (t − ε)F ⌉) ⊗ g∗ OZ (G)
=ωY (⌈εEY − (t − ε)F ⌉).
Remark 5.3. (1) X has only rational singularities if and only if X is Cohen-Macaulay and
J (ωX ) = ωX , see [25, Theorem 5.10].
(2) If X has only Du Bois singularities, then I(ωX ) = ωX , see [26]. In case X is CohenMacaulay, the converse holds as well. The reader is referred to [29] for the definition
and a simple characterization of Du Bois singularities.
20
A. K. SINGH, S. TAKAGI and M. VARBARO
Using I(ωX , at ), we define a new invariant of singularities in characteristic zero:
Definition 5.4. Suppose that X has only Du Bois singularities. Then the threshold d(a) is
defined by
d(a) := sup{t > 0 | I(ωX , at ) = ωX }.
If x is a closed point of X, then the threshold dx (a) is defined by
dx (a) := sup{t > 0 | I(ωX , at )x = ωX,x }.
Remark 5.5. Let R be an N-graded ring with R0 an algebraically closed field k of characteristic
zero, and let m be the homogeneous maximal ideal of R. Let X = Spec R and x ∈ X be the
closed point corresponding to m. By considering a k∗ -equivariant log resolution of (X, m), we
see that I(ωX , mt ) is a graded submodule of the graded canonical module ωR . This implies
that d(m) = dx (m).
In [10] de Fernex-Hacon extended the notion of log canonical thresholds to the non-QGorenstein setting; we recall their definition:
Definition 5.6 ([10, Proposition 7.2]). Suppose that t > 0 is a real number.
(1) The pair (X, at ) is said to be klt (resp. log canonical) in the sense of de Fernex-Hacon
if there exists an effective Q-divisor ∆ on X such that KX + ∆ is Q-Cartier and
((X, ∆); at ) is klt (resp. log canonical) in the classical sense. That is, if π : Y −→ X is
a log resolution of (X, ∆, a) such that aOY = OY (−F ) is invertible and if we write
X
KY − π ∗ (KX + ∆) − tF =
ai Ei ,
i
where the Ei are prime divisors on Y and the ai are real numbers, then ai > −1 (resp.
ai > −1) for all i. The log canonical threshold lct(a) of a is defined by
lct(a) := sup{t > 0 | (X, at ) is log canonical in the sense of de Fernex-Hacon}.
(2) Let x be a closed point of X. Then (X, at ) is klt (resp. log canonical) at x in the sense
of de Fernex-Hacon if there exists a open neighborhood U of x such that (U, (a|U )t )
is klt (resp. log canonical) in the sense of de Fernex-Hacon. If, in addition, a = OX ,
then we say that (X, x) is a log terminal (resp. log canonical) singularity in the sense
of de Fernex-Hacon. The log canonical threshold lctx (a) of a at x is defined by
lctx (a) := sup{t > 0 | (X, at ) is log canonical at x in the sense of de Fernex-Hacon}.
If ∆ is an effective Q-divisor on X such that KX + ∆ is Q-Cartier and (X, ∆) is log
canonical at x (in the classical sense), then the log canonical threshold lctx (∆; a) is
defined by
lctx (∆; a) := sup{t > 0 | ((X, ∆); at ) is log canonical at x (in the classical sense)}.
We prove some basic properties of dx (a).
Lemma 5.7. Let (X, x) be a d-dimensional normal singularity.
(1) If (X, x) is a rational singularity, then
dx (a) = sup{t > 0 | J (ωX , at )x = ωX,x }.
(2) Suppose that (X, x) is a log canonical singularity in the sense of de Fernex-Hacon.
Then lctx (a) 6 dx (a). In addition, if X is quasi-Gorenstein at x, then lctx (a) = dx (a).
(3) Suppose that (X, x) is a Cohen-Macaulay Du Bois singularity. Then dx (mx ) 6 d. If
d+1−⌈dx (mx )⌉
J ⊆ OX,x is a minimal reduction of the maximal ideal mx , then mx
⊆ J.
A GORENSTEIN CRITERION FOR STRONGLY F -REGULAR RINGS
21
(4) Suppose that X is Cohen-Macaulay at x. If dx (mx ) > d − 1, then X is nonsingular at
x and in particular dx (mx ) = d.
Proof. (1) Shrinking X if necessary, we may assume that X has only rational singularities.
First we will check that dx (a) > 0. Let π : Y −→ X be a log resolution of (X, a) such that
aOY = OY (−F ) is invertible and let E be the reduced divisor supported on Exc(π). For
sufficiently small t > ε > 0,
KY + E > ⌈KY + εE − (t − ε)F ⌉ > KY .
Taking the pushforward π∗ , we obtain inclusions ωX ⊃ I(ωX , at ) ⊃ J (ωX ) = ωX by Remark 5.3 (1). That is, dx (a) > t > 0.
Now we will show the assertion. Since J (ωX , at ) ⊆ I(ωX , at ), the inequality
dx (a) > sup{t > 0 | J (ωX , at )x = ωX,x }
is obvious. We will prove the reverse inequality. It is enough to show that if I(ωX , at )x = ωX,x
with t > 0, then J (ωX , at−ε )x = ωX,x for all t > ε > 0. Fix a real number t > ε > 0.
Shrinking X again if necessary, we may assume that X is affine and that I(ωX , at ) = ωX .
This means that for sufficiently small (1/2)ε > ε′ > 0,
ordEi (KY + divY (f ) + ε′ E − (t − ε′ )F ) > −1
for every prime divisor Ei on Y and for every f ∈ ωX . If Ei is an irreducible component of
Supp F , then
ordEi (KY + divY (f ) − (t − ε)F ) > ordEi (KY + divY (f ) + ε′ E − (t − ε′ )F ) > −1.
On the other hand, since X has only rational singularities, KY + divY (f ) > 0 by Remark 5.3 (1). Therefore, if Ei is not a component of Supp F , then
ordEi (KY + divY (f ) − (t − ε)F ) = ordEi (KY + divY (f )) > 0.
Summing up above, we conclude that ⌈KY + divY (f ) − (t − ε)F ⌉ > 0 for every f ∈ ωX , that
is, J (ωX , at−ε ) = ωX .
(2) For the former assertion, it is enough to show that if (X, at ) is log canonical at x in
the sense of de Fernex-Hacon, then I(ωX , at )x = ωX,x . Shrinking X if necessary, we may
assume that X is affine and there exists an effective Q-divisor ∆ on X such that ((X, ∆); at )
is log canonical with KX + ∆ Q-Cartier of index r. Let π : Y −→ X be a log resolution of
(X, ∆, a) such that aOY = OY (−F ) is invertible and let E be the reduced divisor supported
on Exc(π). By the definition of log canonical pairs,
⌈KY − π ∗ (KX + ∆) + ε1 E − (t − ε2 )F ⌉ > 0
for every ε1 > 0 and t > ε2 > 0 (when t = 0, put ε2 = 0). Let f ∈ ωX . Since the fractional
r is contained in O (r(K + ∆)), one has div (f ) + π ∗ (K + ∆) > 0. It follows from
ideal ωX
X
X
Y
X
these two inequalities that
⌈KY + ε1 E − (t − ε2 )F ⌉ + divY (f ) > 0,
which implies that f ∈ I(ωX , at ).
Now we will show the latter assertion. Shrinking X again if necessary, we may assume that
ωX ∼
= OX . Then I(ωX , at ) can be identified with the maximal non-lc ideal J ′ (X, at ) under
this isomorphism (see [14, Definition 7.4] for the definition of J ′ (X, at )). Since J ′ (X, at ) =
OX if and only if (X, at ) is log canonical by the definition of J ′ (X, at ), one has the equality
that lctx (a) = dx (a).
22
A. K. SINGH, S. TAKAGI and M. VARBARO
(3) The proof is essentially the same as that of [33, Theorem 5.2.5]. Let f : Y −→ Spec OX,x
e −→ Spec OX,x
be the blow-up at mx with exceptional divisor Fx . Take a log resolution π : X
P
e
of mx . Then there exists a morphism g : X −→ Y such that π = f ◦ g. Let E = si=1 Ei
be the reduced divisor supported on Exc(π). We may assume that E1 , . . . , E
r are all the
P
′
components of E dominating an irreducible component of Fx , and put E := si=r+1 Ei . If
t > d, then
I(ωX , mtx )x = π∗ OXe (⌈KXe + εE − (t − ε)g ∗ Fx ⌉)
⊆ π∗ OXe (KXe + E ′ − dg ∗ Fx )
= f∗ g∗ OXe (KXe + E ′ ) ⊗ OY (−dFx )
⊆ f∗ ωY (−dFx )
= f∗ mdx ωY
for sufficiently small ε > 0. It follows from [21, Theorem 3.7] and [22, Lemma 5.1.6] that
I(ωX , mtx )x :OX,x ωX,x ⊆ f∗ mdx ωY :OX,x ωX,x = core(m),
where core(m) is the intersection of all reductions of m. In particular, I(ωX , mtx )x ( ωX,x ,
that is, dx (mx ) < t. Thus, dx (mx ) 6 d.
d (m )−ε
Since I(ωX , mxx x )x = ωX,x for every ε > 0 (we put ε = 0 when dx (mx ) = 0), by the
same argument as above,
mxd+1−⌈dx (mx )⌉ ωX,x = mxd+1−⌈dx (mx )⌉ I(ωX , mdxx (mx )−ε )x ⊆ f∗ mdx ωY
for sufficiently small ε > 0. It then follows from [21, Theorem 3.7] and [22, Lemma 5.1.6]
again that
mxd+1−⌈dx (mx )⌉ ⊆ f∗ mdx ωY :OX,x ωX,x = core(m) ⊆ J.
(4) Let J be a minimal reduction of mx . It then follows from (3) that mx = J, which
means that mx is generated by at most d elements, that is, X is nonsingular at x. If X is
nonsingular at x, then by (2), we see that dx (mx ) = lctx (mx ) = d.
We can compute the log canonical threshold of the maximal ideal of an affine determinantal
variety using F -pure thresholds:
Proposition 5.8. Let D := Spec k[T ]/I be the affine determinantal variety over an algebraically closed field k of characteristic zero, where T is an m × n matrix of indeterminates
with m 6 n, and I is the ideal generated by the size t minors of T where 1 6 t 6 m. Let m
be the homogeneous maximal ideal of k[T ]/I, corresponding to the origin 0 in D. Then
lct(m) = m(t − 1).
Proof. For each prime integer p, let Rp := Fp [T ]/Ip be the modulo p reduction of k[T ]/I,
and mp the homogeneous maximal ideal of Rp . It then follows from [9, Theorem 6.4] and
Proposition 4.3 that
lct(m) = lim fpt(mp ) = m(t − 1).
p−→∞
Proposition 5.9. Let x be a closed point of X and ∆ be an effective Q-divisor on X such
that (X, ∆) is log canonical at x with KX + ∆ being Q-Cartier of index r. If X is not
quasi-Gorenstein at x, then
1
lctx (∆; mx ) + 6 dx (mx ).
r
In particular, if dx (mx ) = 0, then X is quasi-Gorenstein at x.
A GORENSTEIN CRITERION FOR STRONGLY F -REGULAR RINGS
23
Proof. Shrinking X if necessary, we may assume that X is affine, OX (r(KX + ∆)) ∼
= OX and
lct (∆;mx )
((X, ∆); mx x
) is log canonical. Let π : Y −→ X be a log resolution of (X, ∆, mx ) such
that mx OX = OX (−Fx ) is invertible, and let E be the reduced divisor supported on Exc(π).
Putting t = lctx (∆; mx ), one has the inequality
⌈KY − π ∗ (KX + ∆) − (t − ε)Fx + εE⌉ > 0
for every ε > 0. On the other hand, since X is not quasi-Gorenstein at x, the fractional ideal
r is contained in m O (r(K + ∆)). Hence, for each f ∈ ω , one has the inequality
ωX
x X
X
X
rdivY (f ) + rπ ∗ (KX + ∆) − Fx > 0.
It follows from these two inequalities that
0 6⌈KY − π ∗ (KX + ∆) − (t − ε)Fx + εE⌉
1
1
= KY + εE − t + − ε Fx − π ∗ (KX + ∆) + Fx
r
r
1
6KY + εE − t + − ε Fx + divY (f )
r
lctx (∆;mx )+1/r
for all ε > 0 and all f ∈ ωX . This means that I(ωX , mx
lctx (∆; mx ) + 1/r.
) = ωX , that is, dx (mx ) >
The following theorem is the main result of this section; this is a characteristic zero analogue
of Theorem 3.14.
Theorem 5.10. Suppose that (X, x) is a log canonical singularity
in the sense of de FernexL
Hacon. Assume in addition that the anti-canonical cover n>0 OX (−nKX )x is Noetherian.
Then lctx (mx ) = dx (mx ) if and only if (X, x) is quasi-Gorenstein.
Proof. Since the “if” part immediately follows from Lemma 5.7 (2), we will show the “only
if” part. Shrinking X if necessary,
we may assume that X is log canonical in the sense of
L
de Fernex-Hacon and that
O
n>0 X (−nKX ) is Noetherian. Then one can find an integer
r > 1 such that OX (−rmKX ) = OX (−rKX )m for every integer m > 1. Fix a real number
lct(m )−ε
ε with min{lct(mx ), 1/r} > ε > 0; when lct(mx ) = 0, put ε = 0. Since (X, mx x ) is log
canonical in the sense of de Fernex-Hacon, there exists an integer m0 > 1 such that the m-th
lct(m )−ε
limiting log discrepancy am,F (X, mx x ) is nonnegative for every prime divisor F over X
and for every positive multiple m of m0 by [10, Definition 7.1] (see loc. cit. for the definition
of the m-th limiting log discrepancy of a pair). By the choice of r, one has
ar,F (X, mxlct(mx )−ε ) = arm0 ,F (X, mxlct(mx )−ε ) > 0.
It follows from an argument similar to the proof of [10, Proposition 7.2] that there exists an
lct(m )−ε
effective Q-divisor ∆ on X such that KX + ∆ is Q-Cartier of index r and ((X, ∆); mx x )
is log canonical. If X is not quasi-Gorenstein at x, then by Proposition 5.9,
1
1
6 lctx (∆; mx ) + 6 dx (mx ).
r
r
This contradicts the assumption that lctx (mx ) = dx (mx ).
lctx (mx ) < lctx (mx ) − ε +
Corollary 5.11. Suppose that (X, x) is a log terminal singularity in the sense of de FernexHacon. Then lctx (mx ) = dx (mx ) if and only if (X, x) is Gorenstein.
24
A. K. SINGH, S. TAKAGI and M. VARBARO
Proof. Since (X, x) is log terminal,L
using the minimal model program for klt pairs, one can
show that the anti-canonical cover n>0 OX (−nKX )x is Noetherian (see [24, Theorem 92]).
Thus, the assertion follows from Theorem 5.10.
Proposition 5.12. Let R be a normal standard graded ring, with R0 an algebraically closed
field of characteristic zero, and let m denote the homogeneous maximal ideal of R. Suppose
that Spec R has only Du Bois singularities. Then d(m) 6 −a(R).
Proof. Put X = Spec R. Since X has only Du Bois singularities, a(R) 6 0 by [27, Theorem 4.4]. Suppose that I(ωX , mt ) = ωX with t > 0. Let ϕ : Y −→ X be the blow-up of X
e −→ Y of (Y, E)
at m and E = Proj R be its exceptional divisor. Take a log resolution ψ : X
e such that
e be the strict transform of E on X.
e We fix a canonical divisor K e on X
and let E
X
t
ψ∗ KXe = KY . Since I(ωX , m ) = ωX ,
e − (t − ε)ψ ∗ E⌉) > 0
ordEe (⌈KXe + divXe (f ) + εE
for all f ∈ ωX and all sufficiently small ε > 0. Taking the direct image by ψ, we see that
ordE (⌈KY + divY (f ) + εE − (t − ε)E⌉) > 0, that is, ϕ∗ ωY (⌈ε − t⌉E) = ωX for sufficiently
small ε > 0. On the other hand, it is easy to see by the definition of ϕ (see, for example, [22,
Proposition 6.2.1]) that
ϕ∗ ωY (⌈ε − t⌉E) = [ωX ]>⌊t−ε⌋+1 .
Thus, t 6 −a(R), that is, d(m) 6 −a(R).
As a consequence, we can prove a characteristic zero analogue of Conjecture 1.1, which
gives an affirmative answer to [11, Conjecture 6.9].
Corollary 5.13. Let R be a normal standard graded ring, with R0 an algebraically closed
field of characteristic zero. Let m denote the homogeneous maximal ideal of R. Assume that
X := Spec R has log canonical singularities in the sense of de Fernex-Hacon.
(1) Then lct(m) 6 −a(R).
L
(2) Suppose in addition that the anti-canonical cover n>0 OX (−nKX ) is Noetherian (this
assumption is satisfied, for example, if X has log terminal singularities in the sense of
de Fernex-Hacon or if R is Q-Gorenstein). Then lct(m) = −a(R) if and only if R is
quasi-Gorenstein.
Proof. Since (1) follows from Remark 5.5, Lemma 5.7 (2) and Proposition 5.12, we will show
(2). Let x ∈ X be the closed point corresponding to m. If lct(m) = −a(R), then lctx (m) has
to be equal to dx (m) by Remark 5.5, Lemma 5.7 (2) and Proposition 5.12 again. It follows
from Theorem 5.10 that X is quasi-Gorenstein at x, which is equivalent to saying that R is
quasi-Gorenstein.
Next we will show the “if” part of (2). Suppose that R is quasi-Gorenstein. Let ϕ : Y −→ X
be the blow-up of X = Spec R at m and E = Proj R be its exceptional divisor. Note that
Y is normal and quasi-Gorenstein. It is easy to see that KY /X = −(1 + a(R))E, see, for
e −→ Y of (Y, E), and
example, the proof of [31, Proposition 5.4]. Take a log resolution ψ : X
then
− ψ ∗ E.
+ ψ ∗ (KY /X + a(R)E) = KX/Y
+ a(R)ψ ∗ E = KX/Y
KX/X
e
e
e
Since X has only log canonical singularities, E has also only log canonical singularities. It
follows from inversion of adjunction for log canonical pairs [23] that (Y, E) is log canonical,
− ψ ∗ E are greater than or equal
which implies that all the coefficients of the divisor KX/Y
e
to −1. Thus, (X, m−a(R) ) is log canonical, that is, lct(m) > −a(R).
A GORENSTEIN CRITERION FOR STRONGLY F -REGULAR RINGS
25
Remark 5.14. Let (R, m) be the same as in Corollary 5.13. If X = Spec R is Q-Gorenstein,
then we can show that lct(m) 6 −ai (R) for all i (see the paragraph preceding Proposition 3.5
for the definition of ai (R)). The proof is as follows.
We may assume that i > 2. Let ϕ : Y −→ X be the blow-up of X at m and Z = Proj R
be its exceptional divisor. Since R isLa normal standard graded ring, there exists a very
ample divisor H on Z such that R = n>0 H 0 (Z, OZ (nH)) and rKZ ∼ aH for some a ∈ Z,
where r is the Gorenstein index of R. We see by the same argument as the proof of [31,
Proposition 5.4] that KY /X = −(1 + a/r)Z, which implies that lct(m) has to be less than
or equal to −a/r. Therefore, in order to prove the inequality lct(m) 6 −ai (R), it suffices
to show that −a/r 6 −ai (R). This condition is equivalent to saying that if ℓ is an integer
greater than a/r, then H i−1 (Z, OZ (ℓH)) = 0, because
ai (R) = max{ℓ ∈ Z | H i−1 (Z, OZ (ℓH)) 6= 0}.
However, since ℓH − KZ ∼Q (ℓ − a/r)H is ample, this is immediate from [13, Theorem 1.7].
When the ring is toric, we have a similar characterization in the non-standard graded case:
Corollary 5.15. Let the notation be the same as in Corollary 3.18. Let R = k[σ ∨ ∩ M ] be
an affine semigroup ring over a field k of characteristic zero, defined by a strongly convex
rational polyhedral cone σ. Let m be the unique monomial maximal ideal of R.
(1) Then d(m) = −aσ (R).
(2) lct(m) = −aσ (R) if and only if R is Gorenstein.
Proof. It follows from the existence of a toric log resolution of m that J ′ (mt ) and I(ωX , mt )
are torus-invariant, and so lct(m) and d(m) are preserved under base field extension. Thus,
we may assume that k is algebraically closed.
Since (2) follows from (1), Remark 5.5, and Corollary 5.11, it remains to justify (1). For
this, use the same strategy as the proof of Corollary 3.18, in which case the assertion follows
from [5, Theorem 2] and Lemma 5.7 (1).
References
[1] I. M. Aberbach, Extension of weakly and strongly F -regular rings by flat maps, J. Algebra 241 (2001),
799–807.
[2] C. Birkar, Existence of flips and minimal models for 3-folds in char p, Ann. Sci. Éc. Norm. Supér. (4) 49
(2016), 169–212.
[3] C. Birkar, P. Cascini, C. D. Hacon, and J. Mc Kernan, Existence of minimal models for varieties of log
general type, J. Amer. Math. Soc. 23 (2010), 405–468.
[4] C. Birkar and J. Waldron, Existence of Mori fibre spaces for 3-folds in char p, arXiv:1410.4511, preprint.
[5] M. Blickle, Multiplier ideals and modules on toric varieties, Math. Z. 248 (2004), 113–121.
[6] W. Bruns and J. Herzog, On the computation of a-invariants, Manuscripta Math. 77 (1992), 201–213.
[7] W. Bruns and U. Vetter, Determinantal rings, Lecture Notes in Math. 1327, Springer-Verlag, Berlin,
1988.
[8] T. Chiba and K. Matsuda, Diagonal F -thresholds and F -pure thresholds of Hibi rings, Comm. Algebra 43
(2015), 2830–2851.
[9] A. Chiecchio, F. Enescu, L. E. Miller, and K. Schwede, Test ideals in rings with finitely generated anticanonical algebras, arXiv:1412.6453, preprint.
[10] T. de Fernex and C. Hacon, Singularities on normal varieties, Compos. Math. 145 (2009), 393–414.
[11] A. De Stefani and L. Núñez-Betancourt, F -thresholds of graded rings, arXiv:1507.05459, preprint.
[12] R. Fedder, F -purity and rational singularity, Trans. Amer. Math. Soc. 278 (1983), 461–480.
[13] O. Fujino, Fundamental theorems for semi log canonical pairs, Algebr. Geom. 1 (2014), 194–228.
[14] O. Fujino, K. Schwede, and S. Takagi, Supplements to non-lc ideal sheaves, in: Higher dimensional
algebraic geometry, RIMS Kôkyûroku Bessatsu B24, pp. 1–46, Res. Inst. Math. Sci. (RIMS), Kyoto, 2011.
26
A. K. SINGH, S. TAKAGI and M. VARBARO
[15] C. D. Hacon and C. Xu, On the three dimensional minimal model program in positive characteristic, J.
Amer. Math. Soc. 28 (2015), 711–744.
[16] N. Hara and S. Takagi, On a generalization of test ideals, Nagoya Math. J. 175 (2004), 59–74.
[17] N. Hara and K.-i. Watanabe, F -regular and F -pure rings vs. log terminal and log canonical singularities,
J. Algebraic Geom. 11 (2002), 363–392.
[18] N. Hara and K. Yoshida, A generalization of tight closure and multiplier ideals, Trans. Amer. Math.
Soc. 355 (2003), 3143–3174.
[19] I. B. Henriques and M. Varbaro, Test, multiplier and invariant ideals, Adv. Math. 287 (2016), 704–732.
[20] D. Hirose, K.-i. Watanabe, and K. Yoshida, F -thresholds versus a-invariants for standard graded toric
rings, Comm. Algebra 42 (2014), 2704–2720.
[21] C. Huneke and N. V. Trung, On the core of ideals, Compositio Math. 141 (2005), 1–18.
[22] E. Hyry and K. E. Smith, On a non-vanishing conjecture of Kawamata and the core of an ideal, Amer.
J. Math. 125 (2003), 1349–1410.
[23] M. Kawakita, Inversion of adjunction on log canonicity, Invent. Math. 167 (2007), 129–133.
[24] J. Kollár, Exercises in the birational geometry of algebraic varieties, arXiv:0809.2579, preprint.
[25] J. Kollár and S. Mori, Birational geometry of algebraic varieties, Cambridge Tracts in Mathematics 134,
Cambridge University Press, Cambridge, 1998.
[26] S. J. Kovács, K. Schwede, and K. E. Smith, The canonical sheaf of Du Bois singularities, Adv. Math. 224
(2010), 1618–1640.
[27] L. Ma, F -injectivity and Buchsbaum singularities, Math. Ann 362 (2015), 25–42.
[28] L. E. Miller, A. K. Singh, and M. Varbaro, The F -pure threshold of a determinantal ideal, Bull. Braz.
Math. Soc. 45 (2014), 767–775.
[29] K. Schwede, A simple characterization of Du Bois singularities. Compos. Math. 143 (2007), 813–828.
[30] K. Schwede, Generalized test ideals, sharp F -purity, and sharp test elements, Math. Res. Lett. 15 (2008),
1251–1261.
[31] K. Schwede and K. E. Smith, Globally F -regular and log Fano varieties, Adv. Math. 224 (2010), 863–894.
[32] K. Schwede and S. Takagi, Rational singularities associated to pairs, Michigan Math. J. 57 (2008), 625–
658.
[33] K. Shibata, Rational singularities, ω-multiplier ideals and cores of ideals, Ph.D. thesis, University of
Tokyo, 2016.
[34] Q. Song, Questions in local cohomology and tight closure, Ph.D. thesis, University of Utah, 2007.
[35] V. Srinivas and S. Takagi, Nilpotence of Frobenius action and the Hodge filtration on local cohomology,
arXiv:1503.08772, preprint.
[36] S. Takagi, F -singularities of pairs and inversion of adjunction of arbitrary codimension, Invent. Math. 157
(2004), 123–146.
[37] S. Takagi and K.-i. Watanabe, On F -pure thresholds, J. Algebra 282 (2004), 278–297.
Department of Mathematics, University of Utah, 155 South 1400 East, Salt Lake City,
UT 84112, USA
E-mail address: [email protected]
Graduate School of Mathematical Sciences, University of Tokyo, 3-8-1 Komaba, Meguro-ku,
Tokyo 153-8914, Japan
E-mail address: [email protected]
Dipartimento di Matematica, Università di Genova, Via Dodecaneso 35, I-16146 Genova, Italy
E-mail address: [email protected]
| 0math.AC
|
Unified Statistical Theory of Spectral Graph Analysis
Subhadeep Mukhopadhyay
Department of Statistical Science, Temple University
Philadelphia, Pennsylvania, 19122, U.S.A.
Dedicated to the beloved memory of Emanuel (Manny) Parzen.
arXiv:1602.03861v4 [math.ST] 20 Sep 2016
Abstract
The goal of this paper is to show that there exists a simple, yet universal1 statistical
logic of spectral graph analysis by recasting it into a nonparametric function estimation
problem. The prescribed viewpoint appears to be good enough to accommodate most
of the existing spectral graph techniques as a consequence of just one single formalism
and algorithm.
Keywords and phrases: Nonparametric spectral graph analysis; Graph correlation density field (GraField); Empirical and smoothed spectral graph analysis; High-dimensional
discrete data smoothing.
Contents
1 Introduction
1.1 Spectral Graph Analysis: A Practitioner’s Guide
1.2 Previous Theoretical Treatments . . . . . . . . .
1.3 A Set of Questions . . . . . . . . . . . . . . . . .
1.4 Unified Construction Principle . . . . . . . . . .
.
.
.
.
2
2
3
3
4
2 Fundamentals of Statistical Spectral Graph Analysis
2.1 Graph Correlation Density Field . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Karhunen-Loéve Representation of Graph . . . . . . . . . . . . . . . . . . . . . . .
2.3 Nonparametric Spectral Approximation Theory . . . . . . . . . . . . . . . . . . . .
5
5
7
9
3 Empirical Spectral Graph Analysis
3.1 Laplacian Spectral Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Diffusion Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Modularity Spectral Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
11
13
14
4 Smoothed Spectral Graph Analysis
4.1 High-dimensional Undersampled Regime
4.2 Spectral Smoothing . . . . . . . . . . . .
4.3 Type-I Regularized Graph Laplacian . .
4.4 Type-II Regularized Graph Laplacian .
4.5 Google’s PageRank Method . . . . . . .
4.6 Other Generalizations . . . . . . . . . .
15
15
15
16
17
17
18
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5 Application to Graph Regression
19
6 Concluding Remarks
22
1
By universal, I mean the theory should not be specific to any particular spectral graph analysis technique
but be meaningfully applicable to a class of models.
1
1
Introduction
Spectral graph data analysis is undoubtedly the most favored technique for graph data
analysis, both in theory and practice. It has been extraordinarily successful in many applications including data mining, web search, quantum computing, computer vision, image
segmentation, and others. Over the last several decades, extensive efforts have been made
by researchers and practitioners to develop a suit of spectral graph analysis techniques (e.g.,
Laplacian, Modularity, Diffusion map, regularized Laplacian, Google PageRank model etc.)
with increasing sophistication and specialization. However, no single algorithm can be regarded as a panacea for dealing with the evolving complexities of modern graphs. Therefore,
the most important and pressing question for the field today appears to be whether we can
develop a unifying language to establish “bridges” between a wide variety of spectral graph
analysis techniques, and thus providing logically connected means for reaching different
ends. Undoubtedly, any such formulation would be of great theoretical significance and
practical value, that will ultimately provide the applied data scientists clear guidance and a
systematic strategy for selecting the proper spectral tools to arrive at a confident conclusion.
To that end, this work attempts to unify the theories of spectral graph analysis by purely
statistical means. That is, we seek a constructive (rather than confirmatory) theoretical
framework integrating classical and modern spectral graph analysis algorithms, which to
date have been viewed as distinct methods.
1.1
Spectral Graph Analysis: A Practitioner’s Guide
The way spectral graph analysis is currently taught and practiced can be summarized as
follows (also known as spectral heuristics):
1. Let G = (V, E) denotes a (possibly weighted) undirected graph with a finite set of
vertices |V | = n, and a set of edges E. Represent the graph using weighted adjacency
matrix A where A(x, y; G) = wxy if the nodes x and y are connected by an edge and
0 otherwise; weights are non-negative and symmetric.
2. Define “suitable” spectral graph matrix (also known as graph “shift” operator). Most
popular and successful ones are listed below:
• L = D−1/2 AD−1/2 ;
Chung (1997)
• B = A − N −1 ddT ;
Newman (2006)
• T = D−1 A;
Coifman and Lafon (2006)
−1/2
• Type-I Reg. Lτ = Dτ
−1/2
• Type-II Reg. Lτ = Dτ
−1/2
A Dτ
;
Chaudhuri et al. (2012)
−1/2
Aτ Dτ
; Amini et al. (2013)
• Google’s PageRank Tα = αD−1 A + (1 − α)F ; (Brin and Page, 1999)
D = diag(d1 , . . . , dn ) ∈ Rn×n , di denotes the degree of a node, τ > 0 regularization
P
parameter, F ∈ Rn×n with all entries 1/n, and N = 2|E| = x,y A(x, y).
2
3. Perform spectral decomposition of the matrix selected at step 2, whose eigenvectors
form an orthogonal basis of Rn . Spectral graph theory seeks to understand the interesting properties and structure of a graph by using leading nontrivial eigenvectors and
eigenvalues, first recognized by Fiedler (1973), which provide the basis for performing
harmonic analysis on graphs.
4. Compute graph Fourier transform (GFT) by expanding functions (or signals) defined
over the vertices of a graph as a linear combination of the selected eigenbasis (from step
3) and carry out learning tasks such as regression, clustering, classification, smoothing,
kriging, etc.
1.2
Previous Theoretical Treatments
There has been a flood of publications on statistical spectral graph theory, most of which
have focused on asymptotic justifications of different techniques (mentioned in the previous
section). See, in particular, those based on Laplacian (Von Luxburg et al., 2008, Belkin
and Niyogi, 2001), modularity (Davis and Sethuraman, 2016), diffusion map (Coifman and
Lafon, 2006, Coifman et al., 2005), and regularized Laplacian (Chaudhuri et al., 2012, Le
et al., 2015) methods. More refined theoretical results in the context of a specific parametric
graph model (such as stochastic block model or its variants) are discussed in Joseph and Yu
(2016), Rohe et al. (2011), Sarkar et al. (2015) and Qin and Rohe (2013).
What is this paper not about? It is important to draw a distinction between the theory
that purports to provide the constructions of different spectral graph techniques and, the
theory that does not. This article focuses on the former. Instead of statistical confirmation,
here our main interest lies in understanding the statistical origins of various spectral graph
techniques. This is mostly ill-understood and neglected territory. The current theoretical
treatments do not give us any clues on this issue.
1.3
A Set of Questions
Despite many advances, the theory and practice of spectral graph analysis are still very
much compartmentalized and unsystematic. The graph signal processing engineers are often
left bewildered by the vastness of the existing literature and huge diversity of methods
(developed by the machine learning community, applied harmonic analysts, physicists, and
statisticians). This has led to a need to develop a broader perspective on this topic, lest
we be guilty of not seeing the forest for the trees. The question how different spectral
graph techniques can naturally originate from some underlying basic principle, plays the
key role in this regard by clarifying the mystery of why and when to use them. The reality
is we still lack general tools and techniques to attack this question in a statistical way.
Given the very different character of the existing spectral models, it seems that new kinds
of abstractions need to be formulated to address this fundamental challenge. What would
these abstractions look like? How can we discover a statistical path that naturally leads
to these different spectral methods? One such promising theory is discussed in this paper,
3
taking inspiration from celebrated history on nonparametric spectral analysis of time series.
The prescribed modern viewpoint shows an exciting confluence of nonparametric function
estimation, quantile domain methods, and Fourier harmonic analysis to unveil a more simple
conceptual structure of this subject, avoiding undue complexities.
1.4
Unified Construction Principle
We start by giving a glance to the unified spectral representation scheme, whose (nonparametric) statistical interpretation and justification is deferred until the next section, as it
requires a more technical background. It turns out that all known spectral graph techniques
are just different manifestations of this single general algorithm in some way.
Probability Notations and Definitions. We will stick to the following probabilistic notations.
P
Define network probability mass function by P (x, y; G) = A(x, y; G)/ x,y A(x, y; G); Vertex
P
P
probability mass function by p(x; G) = y P (x, y; G) and p(y; G) = x P (x, y; G) with the
associated quantile functions Q(u; X, G) and Q(v; Y, G). Finally, define the important Graph
Interaction Function by GIF(x, y; G) = P (x, y; G)/p(x; G)p(y; G), x, y ∈ V (G).
Algorithm 1. Nonparametric Spectral Graph Analysis: A Unified Algorithm
Step 1. For given discrete graph G of size n, construct GraField kernel function C :
[0, 1]2 → R+ ∪ {0} defined a.e by
P Q(u; X), Q(v; Y ); G
,
(1.1)
C (u, v; G) = GIF Q(u; X), Q(v; Y ); G =
p Q(u; X) p Q(v; Y )
where u = F (x; G), v = F (y; G) for x, y ∈ {1, 2, . . . , n}.
Step 2. Let {ξj }j≥0 be a complete orthonormal system on [0, 1]. Construct discrete transform basis ξj (x; F (X; G)) := ξj (F (x; G)), evaluated at the vertex-rank-transforms. They
satisfy the following orthonormality conditions (degree-weighted):
X
ξj (x; F (X)) ξk (x; F (X)) p(x; G) = δjk ,
x
thus constitutes an orthonormal basis for L2 (F ; G).
Step 3. Transform coding of graphs. Construct generalized spectral graph matrix M(G, ξ) ∈
Rn×n with respect to ηj (u) := ξj (Q(u; X); F ):
D
Z1
M[j, k; G, ξ] = ηj ,
(C − 1)ηk
E
L2 [0,1]
0
=
X
ξj (`; F )ξk (m; F )P (`, m; G).
(1.2)
`,m
Mξ can be viewed as a transform coefficient matrix of the orthogonal series expansion of
C (u, v; G) with respect to the product bases {ηj ηk }1≤j,k≤n .
4
P
Step 4. Perform the singular value decomposition (SVD) of Mξ = U ΛU T = k uk µk uTk ,
where where uij are the elements of the singular vector of moment matrix U = (u1 , . . . , un ),
and Λ = diag(µ1 , . . . , µn ), µ1 ≥ · · · µn ≥ 0.
Step 5. Obtain approximate Karhunen-Loéve (KL) representation basis (which acts as a
graph Fourier basis) of the graph G by
φek =
n
X
ujk ξj , for k = 1, . . . , n − 1,
j=1
which can be directly used for subsequent signal processing on graphs.
Organization. The next section is devoted to the fundamental principles that underlie this
algorithm. The unifying power will be shown in Sections 3 and 4. By doing so, we will
introduce empirical and smooth spectral graph analysis techniques. We also address the
open problem of obtaining a formal interpretation of “spectral regularization”. A deep
connection between high-dimensional discrete data smoothing and spectral regularization
is discovered. This new perspective provides, for the first time, the theoretical motivation
and fundamental justification for using regularized spectral methods, which were previously
considered to be empirical guesswork-based ad hoc solutions. Important application towards
spatial graph regression is discussed in Section 5. In Section 6, we end with concluding
remarks. Additional applications are available as supplementary materials.
2
2.1
Fundamentals of Statistical Spectral Graph Analysis
Graph Correlation Density Field
Wiener’s generalized harmonic analysis formulation on spectral representation theory of
time series starts by defining the autocorrelation function (ACF) of a signal. In particular,
Wiener–Khintchine theorem asserts ACF and the spectral density are Fourier duals of each
other. Analogous to Wiener’s correlation technique, we describe a new promising starting
point, from which we can develop the whole spectral graph theory systematically by bringing
nonparametric function estimation and harmonic analysis perspectives. Both statistical and
probabilistic motivations will be given.
Definition 1. For given discrete graph G of size n, the piecewise-constant bivariate kernel
function C : [0, 1]2 → R+ ∪ {0} is defined almost everywhere through
p Q(u; X), Q(v; Y ); Gn
, 0 < u, v < 1. (2.1)
C (u, v; Gn ) = GIF Q(u; X), Q(v; Y ); Gn =
p Q(u; X) p Q(v; Y )
Theorem 1. GraField defined in Eq. (2.1) is a positive piecewise-constant kernel satisfying
ZZ
ZZ
X
C (u, v; Gn ) du dv =
C (u, v; Gn ) du dv = 1,
[0,1]2
(i,j)∈{1,...,n}2 I
ij
5
where
(
Iij (u, v) =
1,
0,
if (u, v) ∈ (F (i; X), F (i + 1; X)] × (F (j; Y ), F (j + 1; Y )]
elsewhere.
Note 1. The bivariate step-like shape of the GraField kernel is governed by the (piecewiseconstant left continuous) quantile functions Q(u; X, Gn ) and Q(v; Y, Gn ) of the discrete vertex
probability measures. As a result, in the continuum limit (let (Gn )n≥0 be a sequence of graphs
whose number of vertices tends to infinity n → ∞), the shape of the piecewise-constant discrete Cn approaches a “continuous field” over unit interval – a self-adjoint compact operator
on square integrable functions (defined on the graph) with respect to vertex probability
measure p(x; G) (see Section 2.2 for more details).
Motivation 1. We start with a diffusion based probabilistic interpretation of the GraField
kernel. The crucial point to note is that the “slices” of the C (2.1) can be expressed
as p(y|x; G)/p(y; G) in the vertex domain. This alternative conditional probability-based
viewpoint suggests a connection with the random walk on the graph. Interpret p(y|x; G) as
transition probability from vertex x to vertex y in one time step. Also note that p(y; G) is
the stationary probability distribution on the graph, as we have limt→∞ p(t, y|x; G) = p(y; G)
regardless of the initial starting point x (moreover, for connected graphs the stationary
distribution is unique). Here p(t, y|x; G) denotes the probability distribution of a random
walk landing at location y at time t, starting at the vertex x. See Lovász (1993) for an
excellent survey on the theory of random walks on graphs.
The graph affinity function measures how the transition probability p(y|x; G) is different from
the “baseline” stationary distribution (long-run stable behavior) p(y; G). That comparison
ratio is the fundamental interaction function for graphs which we denote by GIF(x, y; G).
This probabilistic interpretation along with Theorem 2 and 5 will allow us to integrate the
diffusion map (Coifman and Lafon, 2006) technique into our general statistical framework
in Section 3.2.
Motivation 2. GraField compactly represents the affinity or strength of ties (or interactions) between every pair of vertices in the graph. To make this clear, let us consider the
following adjacency matrix of a social network representing 4 employees of an organization
A=
0
2
0
0
2
0
3
3
0
3
0
3
0
3
,
3
0
where the weights reflect numbers of communication (say email messages or coappearances in social events etc.). Our interest lies in understanding the strength of association between the employees i.e., Strength(x, y) for all pairs of vertices. Looking at the
matrix A (or equivalently based on the histogram network estimator p(x, y; G) = A/N
P
with N =
x,y A(x, y) = 22) one might be tempted to conclude that the link between
6
employee 1 and 2 is the weakest, as they have communicated only twice, whereas employees 2, 3, and 4 constitute strong ties, as they have interacted more frequently. Now, the
surprising fact is that (i) Strength(1, 2) is twice that of Strength(2, 3) and Strength(2, 4);
also (ii) Strength(1, 2) is 1.5 times of Strength(3, 4)! To understand this paradox, compute the vertex-domain empirical GraField kernel matrix (Definition 1) with (x, y)th entry
N ·A(x, y; G)/d(x)d(y)
0
22/8
0
0
22/8
0
22/16 22/16
.
Cn =
0 22/16
0
22/12
0
22/16 22/12
0
This toy example is in fact a small portion (with members 1, 9, 31 and 33) of the famous
Zachary’s karate club data, where the first two members were from Mr. Hi’s group and the
remaining two were from John’s group. The purpose of this illustrative example is not to
completely dismiss the adjacency or empirical graphon (Lovász and Szegedy, 2006) based
analysis but to caution the practitioners so as not to confuse the terminology “strength of
association” with “weights” of the adjacency matrix – the two are very different objects.
Existing literature uses them interchangeably without paying much attention. The crux of
the matter is: association does not depend on the raw edge-density, it is a “comparison
edge-density” that is captured by the GraField; see Section 3.2 for its intriguing connection
with diffusion based graph-distance.
Motivation 3. GraField can also be viewed as properly “renormalized Graphon,” which
is reminiscent of Wassily Hoeffding’s “standardized distributions” idea (Hoeffding, 1940).
Thus, it can be interpreted as a discrete analogue of copula (the Latin word copula means
“a link, tie, bond”) density for random graphs that captures the underlying correlation field.
We study the structure of graphs in the spectral domain via this fundamental graph kernel
C that characterizes the implicit connectedness or tie-strength between pairs of vertices.
Fourier-type spectral expansion results of the density matrix C are discussed in the ensuing
section, which is at the heart of our approach. We will demonstrate that this correlation
density operator-based formalism provides a useful perspective for spectral analysis of graphs
that allows unification.
2.2
Karhunen-Loéve Representation of Graph
We define the Karhunen-Loéve (KL) representation of a graph G based on the spectral
expansion of its GraField function C (u, v; G). Schmidt decomposition (Schmidt, 1907) of
C yields the following spectral representation theorem of the graph.
Theorem 2. The square integrable graph correlation density kernel C : [0, 1]2 → R+ ∪ {0}
of two-variables admits the following canonical representation
C (u, v; Gn ) = 1 +
n−1
X
k=1
7
λk φk (u)φk (v),
(2.2)
where the non-negative λ1 ≥ λ2 ≥ · · · λn−1 ≥ 0 are singular values and {φk }k≥1 are the
orthonormal singular functions hφj , φk iL 2 [0,1] = δjk , for j, k = 1, . . . , n − 1, which can be
evaluated as the solution of the following integral equation relation
Z
[C (u, v; G) − 1] φk (v) dv = λk φk (u),
k = 1, 2, . . . , n − 1.
(2.3)
[0,1]
Remark 1. By virtue of the properties of Karhunen-Loéve (KL) expansion (Loève, 1955),
the eigenfunction basis φk satisfying (2.3) provides the optimal low-rank representation of a
graph in the mean square error sense. In other words, {φk } bases capture the graph topology in the smallest embedding dimension and thus carries practical significance for graph
compression. Hence, we can call those functions the optimal coordinate functions or Fourier
representation bases. Accordingly, the fundamental statistical modeling problem hinges on
finding approximate solutions to the optimal graph coordinate system {φ1 , . . . , φn−1 } satisfying the integral equation (2.3).
Definition 2. Any function or signal y ∈ Rn defined on the vertices of the graph y : V 7→ R
P
such that kyk2 = x∈V (G) |y(x)|2 p(x; G) < ∞, can be represented as a linear combination of
the Schmidt bases of the GraField density matrix C . Define the generalized graph Fourier
transform of y
n
X
yb(λk ) := hy, φk i =
y(x)φk [F (x; G)].
x=1
This spectral or frequency domain representation of a signal, belonging to the square integrable Hilbert space L 2 (F, G) equipped with the inner product
hy, ziL 2 (F ;G) =
X
y(x)z(x)p(x; G),
x∈V (G)
allows us to construct efficient graph learning algorithms. As {φk }’s are KL spectral bases,
the vector of projections onto this basis function decay rapidly, and hence may be truncated
aggressively to capture the structure in a small number of bits.
Definition 3. The entropy (or energy) of a discrete graph G, is defined using the Parseval
relation of the canonical representation
ZZ
X
Entropy(G) =
(C − 1)2 du dv =
|λk |2 .
k
[0,1]2
This quantity, which captures the departure of uniformity of the C , can be interpreted as
a measure of ‘structure’ or the ‘compressibility’ of the graph. Entropy measure can also be
used to (i) define graph homogeneity and (ii) design fast algorithms for graph isomorphism.
For homogeneous graphs, the shape of the correlation density field is uniform over the unit
square. The power at each harmonic components, as a function of frequency, is called
the power spectrum of the graph. Flat spectrum is equivalent to the analogous notion of
white-noise process.
8
2.3
Nonparametric Spectral Approximation Theory
We view the spectral graph learning algorithm as a method of approximating (λk , φk )k≥1 that
satisfies the integral equation (2.3), corresponding to the graph kernel C (u, v; G). In practice,
often the most important features of a graph can be well characterized and approximated by
a few top (dominating) singular-pairs. The statistical estimation problem can be summarized
as follows:
n
o
b1 , φb1 , . . . , λ
bn−1 , φbn−1
An×n 7→ C 7→
λ
that satisfies Eq. (2.3).
2.3.1
From Continuous to Discrete Basis and Back
The Fourier-type nonparametric spectral approximation method starts by choosing an expansion basis. Let ξj denote an orthonormal basis of L 2 [0, 1]. Construct the discrete transform basis of Rn by evaluating ξj at the vertex-rank-transforms ξj (F (x; G)) := ξj (x; F, G).
Verify that ξj (x; F, G) are orthogonal with respect to the measure p(x; G) satisfying
X
ξj (x; F, G) ξk (x; F, G) p(x; G) = 0, for j 6= k.
x
Thus form an orthogonal basis of the Hilbert space L 2 (F ; G). Define η-functions (quantile
domain unit bases), generated from mother ξj by
ηj (u; G) = ξj Q(u; X); F, G , 0 < u < 1,
as piecewise-constant (left-continuous) functions over the irregular grid {0, p(1), p(1)+p(2), . . . ,
Pn
j=1 p(j) = 1} satisfying hηj , ηk iL 2 [0,1] = 0, if j 6= k. They will provide a useful tool to recast conventional matrix calculus-based approaches as a functional statistical problem. One
unique aspect of our construction is that in the continuum limit (as the size of the graph
n → ∞) the discrete η-basis of Rn approaches the mother ξ-function, a basis of L 2 [0, 1].
We call it asymptotic “reproducing” property.
2.3.2
Projection Methods for Eigenvector Approximation
We are interested in the nonparametric estimation of eigenpairs {λk , φk }k≥0 . Approximate
the unknown eigenvectors by the projection, Pn φk , on the span{ηj , j = 1, . . . , n} defined by
φk (u) ≈ Pn φk =
n
X
θjk ηj (u),
0<u<1
(2.4)
j=1
where θjk are the unknown coefficients to be estimated.
Definition 4 (Orthogonal Discrete Graph Transform). We introduce a generalized concept
of matrices associated with graphs called the G-matrix. Define discrete graph transform
with respect to an orthonormal system η as
M[j, k; η, G] =
D
Z1
(C − 1)ηk
ηj ,
0
9
E
L 2 [0,1]
for j, k = 1, . . . , n.
(2.5)
Equivalently, we can define the discrete graph transform to be the coefficient matrix of the
orthogonal series expansion of the GraField kernel C (u, v; G) with respect to the product
bases {ηj ηk }1≤j,k≤n . As a practical significance, this generalization provides a systematic
recipe for converting the graph problem into a “suitable” matrix problem:
Gn (V, E)
−→
An×n
C (u, v; Gn )
−→
{η1 ,...,ηn }
−−−−−→
Eq. (2.5)
M(η, G) ∈ Rn×n .
Theorem 3. The G-matrix (2.5) can also be interpreted as a “covariance” operator for a
discrete graph by recognizing the following equivalent representation for x, y ∈ {1, 2, . . . . , n}
X
M[j, k; ξ, G] = EP ξj F (X; G) ξk F (Y ; G)
=
P (x, y; G) ξj F (x; G) ξj F (y; G) ,
x,y
This can be proved using the basic quantile mechanics fact that Q(F (X)) = X holds with
probability 1 (see Parzen (1979)). Next, we present a general approximation scheme that
provides an effective method of discrete graph analysis in the frequency domain.
Theorem 4 (Nonparametric spectral approximation). The Fourier coefficients {θjk } of the
projection estimators (2.4) of the GraField eigenfunctions (eigenvalues and eigenvectors),
satisfying the integral equation (2.3), can be obtained by solving the following generalized
matrix eigenvalue problem
MΘ = SΘ∆,
(2.6)
R1
where Mjk = ηj , 0 (Cn − 1)ηk L 2 [0,1] , and Sjk = ηj , ηj L 2 [0,1] .
To prove define the residual of the governing equation (2.3) by expanding φk as series
expansion (2.4),
R(u) ≡
X
j
θjk
h Z1
i
C (u, v; G) − 1 ηj (v) dv − λk ηj (u) = 0.
(2.7)
0
Now for complete and orthonormal {ηj } requiring the error R(u) to be zero is equivalent to
the statement that R(u) is orthogonal to each of the basis functions
R(u), ηk (u)
L 2 [0,1]
= 0,
k = 1, . . . , n.
(2.8)
This leads to the following set of equations:
X
j
θjk
h ZZ
1
i
X Z
C (u, v; Gn )−1 ηj (v)ηk (u) dv du − λk
θjk
ηj (u)ηk (u) du = 0. (2.9)
j
[0,1]2
0
Theorem 4 plays a key role in our statistical reformulation. In particular, we will show how
the fundamental equation (2.6) provides the desired unity among different spectral graph
techniques by systematically constructing a “suitable” class of coordinate functions.
10
Note 2. The fundamental idea behind the Rietz-Galerkin (Galerkin, 1915) style approximation scheme for solving variational problems in Hilbert space played a pivotal inspiring
role to formalize the statistical basis of the proposed computational approach.
Note 3. Our nonparametric spectral approximation theory based on eigenanalysis of G-Matrix,
remains unchanged for any choice η-function (following the recipe for construction given in
Sec 2.3.1), which paves the way for the generalized harmonic analysis of graphs.
The next two sections investigate this general scheme under various choices of η-functions,
and nonparametric estimation methods of Cn . By doing so, many “mysterious similarities”
among different spectral graph algorithms are discovered which were not known before.
3
Empirical Spectral Graph Analysis
Three popular traditional spectral graph analysis models will be synthesized in this section.
3.1
Laplacian Spectral Analysis
Laplacian is probably the most heavily used spectral graph technique in practice. Here we
will demonstrate for the first time how the Laplacian of a graph naturally originates by
purely statistical reasoning, totally free from the classical combinatorial based logic.
Degree-Adaptive Block-pulse Basis Functions. One of the fundamental, yet universally valid (for any graph) choice for {ηj }1≤j≤n is the indicator top hat functions (also known
as block-pulse basis functions, or in short BPFs). However, instead of defining the BPFs
on a uniform grid (which is the usual practice) here (following Sec 2.3.1) we define them on
P
the non-uniform mesh 0 = u0 < u1 · · · < un = 1 over [0,1], where uj = x≤j p(x; X) with
local support
(
p−1/2 (j) for uj−1 < u ≤ uj ;
(3.1)
ηj (u) =
0
elsewhere.
They are disjoint, orthogonal, and a complete set of functions satisfying
Z1
0
Z1
Z1
p
2
ηj (u) du = 1, and
ηj (u)ηk (u) du = δjk .
ηj (u) du = p(j),
0
0
Note 4. The shape (amplitudes and block lengths) of our specially designed BPFs depend
on the specific graph structure via p(x; G) as shown in Fig 1. In order to obtain the spectral
domain representation of the graph, it is required to estimate the spectra of GraField kernel
φk , by representing them as block pulse series. The next result describes the required
computational scheme for estimating the unknown expansion coefficients {θjk }.
Theorem 5. Let φ1 , . . . , φn the canonical Schmidt bases of L 2 graph kernel C (u, v; G),
satisfying the integral equation (2.3). Then the empirical solution of (2.3) for block-pulse
11
2
2
1
3
1
4
3
0.0
0.0
0.6
1.0
1.2
2.0
4
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.0
0.0
0.0
1.0
1.0
2.0
2.0
0.0
1.0
1.0
2.0
2.0
0.0
1.0
1.0
2.0
2.0
0.0
Figure 1: Two graphs and their corresponding degree-adaptive block-pulse functions. The
amplitudes and the block length of the indicator basis functions (3.1) depends on the degree
distribution of that graph.
orthogonal series approximated (3.1) Fourier coefficients {θjk } can equivalently be written
down in closed form as the following matrix eigen-value problem
L∗ [θ] = λθ,
(3.2)
1/2
where L∗ = L − uuT , L is the Laplacian matrix, u = Dp 1n , and Dp = diag(p1 , . . . , pn ).
Note that the discrete GraField kernel takes value GIF(j, k; Gn ) = P (j, k; Gn )/p(j; Gn )p(k; Gn )
and the tensor–product bases ηj (u)ηk (v) take value p−1/2 (j; Gn )p−1/2 (k; Gn ) over the rectangle Ijk (u, v) for 0 < u, v < 1. This observation reduces the master equation (2.9) to the
following system of linear algebraic equations expressed in the vertex domain:
#
"
X
p
p
p(j, k)
θjk p
− p(j) p(k) − λk δjk = 0.
(3.3)
p(j)p(k)
j
12
Empirical plugin nonparametric estimator. The estimating equation (3.3) contains
unknown network and vertex probability mass functions which need to be estimated from
the data. The most basic nonparametric estimates are Pe(j, k; Gn ) = A[j, k; Gn ]/N and
pe(j; Gn ) = d(j; Gn )/N . By plugging these empirical estimators into Eq. (3.3), said equation
can be rewritten as the following compact matrix form:
h
√ √ Ti
(3.4)
L − N −1 d d θ̃ = λ̃θ̃,
where L = D−1/2 AD−1/2 is the graph Laplacian matrix.
Significance 1. Theorem 5 allows us to interpret the graph Laplacian as the empirical
f Gn ) under degree-adaptive indicator basis choice for the η shape function.
G-matrix M(η,
Our technique provides a completely nonparametric statistical derivation of an algorithek versus k can be considered as a “raw
mic spectral graph analysis tool. The plot of λ
periodogram” analogue for graph data analysis.
3.2
Diffusion Map
We provide a statistical derivation of Coifman’s diffusion map (Coifman and Lafon, 2006)
algorithm, which hinges upon the following key result.
Theorem 6. The empirical GraField admits the following vertex-domain spectral diffusion
decomposition at any finite time t
X
pe(t, y|x; Gn )
λ̃tk φ̃k (x; F )φ̃k (y; F ),
= 1+
pe(y; Gn )
k
(3.5)
−1/2
where φek = Dp̃ uk , (uk is the kth eigenvector of the Laplacian matrix L), (φk ◦ F )(·) is
abbreviated as φk (·; F ), pe(y|x; G) = T (x, y), and T = D−1 A is the transition matrix of a
random walk on G with stationary distribution p̃(y; Gn ) = d(y; Gn )/N .
−1/2
Replacing the estimated coefficients from Theorem 5 into (2.4) yields φek = Dp uk , where
uk is the kth eigenvector of the Laplacian matrix L, immediately leads to the following
vertex-domain spectral decomposition result of the empirical GraField:
X
pe(y|x; G)
= 1+
λ̃k φ̃k (x)φ̃k (y),
pe(y; G)
k
(3.6)
For an alternative proof of the expansion (3.6) see Appendix section of Coifman and Lafon
(2006) by noting φek are the right eigenvectors of random walk Laplacian T .
Significance 2. In light of Theorem 6, define diffusion map coordinates at time t as the
mapping from x to the vector
x 7−→ λt1 φ1 (x; F ), . . . , λtk φk (x; F ) , x ∈ {1, 2, . . . , n},
13
which can be viewed as an approximate ‘optimal’ Karhunen-Loéve representation basis and
thus can be used for non-linear embedding of graphs. Define diffusion distance, a measure
of similarity between two nodes of a graph, as the Euclidean distance in the diffusion map
space
X
2
0
(3.7)
λ2t
Dt2 (x, x0 ) =
j φj (x; F ) − φj (x ; F ) .
j≥1
This procedure is known as diffusion map, which has been extremely successful tool for
manifold learning2 . Our approach provides an additional insight and justification for the
diffusion coordinates by interpreting it as the strength of connectivity profile for each vertex,
thus establishing a close connection with empirical GraField.
3.3
Modularity Spectral Analysis
Theorem 7. To approximate the KL graph basis φk =
u ≤ uj ) to be the characteristic function satisfying
Z1
Z1
ηj (u) du =
0
P
j
θjk ηj , choose ηj (u) = I(uj−1 <
ηj2 (u) du = p(j; G).
0
Then the corresponding empirically estimated spectral graph equation (2.9) can equivalently
be reduced to the following generalized eigenvalue equation in terms of the matrix B =
A − N −1 ddT
Bα = λDα.
(3.8)
Significance 3. The matrix B, known as modularity matrix, was introduced by Newman
(2006) from an entirely different motivation. Our analysis reveals that the Laplacian and
Modularity based spectral graph analyses are equivalent in the sense that they inherently use
the same underlying basis expansion (one is a rescaled version of the other) to approximate
the optimal graph bases.
Remark 2. Solutions (eigenfunctions) of the GraField estimating equation based on the
G-matrix under the proposed specialized nonparametric approximation scheme provides a
systematic and unified framework for spectral graph analysis. As an application of this
general formulation, we have shown how one can synthesize the well-known Laplacian, diffusion map, and modularity spectral algorithms and view them as “empirical” spectral graph
analysis methods. It is one of those rare occasions where one can witness the convergence
of statistical, algorithmic and geometry-motivated computational models.
2
Manifold learning: Data-driven learning of the “appropriate” coordinates to identify the intrinsic nonlinear structure of high-dimensional data. We claim the concept of GraField allows decoupling of the
geometrical aspect from the probability distribution on the manifold.
14
4
Smoothed Spectral Graph Analysis
Smoothness is a universal requirement for constructing credible nonparametric estimators.
Spectral graph analysis is also no exception. An improved smooth version of raw-empirical
spectral graph techniques will be discussed, revealing a simple and straightforward statistical
explanation of the origin of regularized Laplacian techniques.
4.1
High-dimensional Undersampled Regime
Recall from Theorem 4 that the generalized matrix eigenvalue equation (2.6) depends on the
unknown network and vertex probability mass functions. This leads us to the question of
estimating the unknown distribution P = (p1 , p2 , . . . , pn ) (support size = size of the graph
P
= n) based on N sample, where N = ni=1 di = 2|E|. Previously (Theorems 5-7) we have
used the unsmoothed maximum likelihood estimate (MLE) pe(x; G) to construct our empirical spectral approximation algorithms, which is the unique minimum variance unbiased
estimator. Under standard asymptotic setup, where the dimension of the parameter space n
is fixed and the sample size N tends to infinity, the law of large numbers ensures optimality
of pe. As a consequence, empirical spectral analysis techniques are expected to work quite
well for dense graphs.
Estimation of probabilities from sparse data. However, the raw pe is known to be strictly suboptimal (Witten and Bell, 1991) and unreliable in the high-dimensional sparse-regime where
N/n = O(1) (i.e., when parameter dimension and the sample size are comparably large).
This situation can easily arise for modern day large sparse graphs where the ratio N/n is
small and there are many nodes with low degree, as is the case of degree sparsity. The naive
MLE estimator can become unacceptably noisy (high variability) due to the huge size and
sparse nature of the distribution. In order to reduce the fluctuations of “spiky” empirical
estimates, some form of “smoothing” is necessary. The question remains: How to tackle
this high-dimensional discrete probability estimation problem, as this directly impacts the
quality of our nonparametric spectral approximation.
My main purpose in the next section is to describe one such promising technique for smoothing raw-empirical probability estimates, which is flexible and in principle can be applied to
any sparse data.
4.2
Spectral Smoothing
We seek a practical solution for circumventing this problem that lends itself to fast computation. The solution, that is both the simplest and remarkably serviceable, is the Laplace/Additive
smoothing (Laplace, 1951) and its variants, which excel in sparse regimes (Fienberg and
Holland, 1973, Witten and Bell, 1991). The MLE and Laplace estimates of the discrete
distribution p(j; Gn ) are respectively given by
Raw-empirical MLE estimates:
pe(j; Gn ) =
15
dj
;
N
Smooth Laplace estimates:
pbτ (j; Gn ) =
dj + τ
N + nτ
(j = 1, . . . , n).
Note that the smoothed distribution pbτ can be expressed as a convex combination of the
empirical distribution pe and the discrete uninform distribution 1/n
N
nτ
1
pbτ (j; Gn ) =
pe(j; Gn ) +
,
(4.1)
N + nτ
N + nτ n
which provides a Stein-type shrinkage estimator of the unknown probability mass function
p. The shrinkage significantly reduces the variance, at the expense of slightly increasing the
bias.
Choice of τ . The next issue is how to select the “flattening constant” τ . The following
choices of τ are most popular in the literature3 :
1
Laplace estimator;
1/2
Krichevsky–Trofimov estimator;
τ=
1/n
Perks estimator;
√
N /n Minimax estimator (under L2 loss).
Note 5. Under increasing-dimension asymptotics, this class of estimator is often difficult
to improve without imposing additional smoothness constraints on the vertex probabilities;
see Bishop et al. (2007, Chapter 12). The latter may not be a valid assumption as nodes of
a graph offer no natural order in general.
With this understanding, smooth generalizations of empirical spectral graph techniques will
be discussed, which have a close connection with recently proposed spectral regularized
techniques.
4.3
Type-I Regularized Graph Laplacian
Construct τ -regularized smoothed empirical ηj;τ basis function by replacing the amplitude
−1/2
p−1/2 (j) by p̂τ (j) following (4.1). Incorporating this regularized trial basis, we have the
following modified G-matrix based linear algebraic estimating equation (2.6):
"
#
X
p
p
pe(j, k)
θjk p
(4.2)
− p̂τ (j) p̂τ (k) − λk δjk = 0.
p̂
(j)p̂
(k)
τ
τ
j
Theorem 8. The τ -regularized block-pulse series based spectral approximation scheme is
equivalent to representing or embedding discrete graphs in the continuous eigenspace of
Type-I Regularized Laplacian = Dτ−1/2 A Dτ−1/2 ,
(4.3)
where Dτ is a diagonal matrix with i-th entry di + τ .
Note 6. It is interesting to note that this exact regularized Laplacian formula was proposed
by Chaudhuri et al. (2012) and Qin and Rohe (2013), albeit from a very different motivation.
3
For more details on selection of τ see Fienberg and Holland (1973) and references therein.
16
4.4
Type-II Regularized Graph Laplacian
Theorem 9. Estimate the joint probability p(j, k; G) by extending the formula given in (4.1)
for the two-dimensional case as follows:
nτ
1
N
pe(j, k; G) +
,
(4.4)
pbτ (j, k; G) =
N + nτ
N + nτ n2
which is equivalent to replacing the original adjacency matrix by Aτ = A + (τ /n)11T . This
modification via smoothing in the estimating equation (4.2) leads to the following spectral
graph matrix
Type-II Regularized Laplacian = Dτ−1/2 Aτ Dτ−1/2 .
(4.5)
Note 7. Exactly the same form of regularization of Laplacian graph matrix (4.5) was proposed by Amini et al. (2013) as a fine-tuned empirical solution.
Significance 4. Recently there has been a lot of discussion on how to choose the spectral
regularization parameter τ . Using Davis–Kahan’s theorem, Joseph and Yu (2016) have
proposed a data-driven technique called DKest. Their method involves repeated eigendecomposition of regularized Laplacian matrix (which is a dense matrix) over a grid of
values of τ . There are two major concerns for DKest: heavy computational burden and
theoretical validity (which only holds for stochastic block models and its extensions). In
contrast, our analysis provides a strikingly simple recommendation without any additional
computational overhead (by connecting it to large sparse distribution smoothing), and thus
is much easier to construct, implement, and use in practice. Similar to Joseph and Yu
(2016), the critical parameter of our theory is the ratio N/n–the average degree of the
nodes, contrary to the previous results (Chaudhuri et al., 2012, Qin and Rohe, 2013) that
relied on the minimum degree.
4.5
Google’s PageRank Method
Smoothing of network and vertex probability distributions appearing in generalized matrix
equation (2.6) resulted in Type-I and Type-II regularized Laplacian methods. The third
possibility is to directly smooth the conditional or transitional probability matrix to develop a regularized version of random walk Laplacian (which we call Type-III regularized
Laplacian) method for large sparse graphs.
Smoothing conditional probability function. Consider a random walk on G with transition
probability T (i, j; G) = Pr(Xt+1 = j|Xt = i) ≥ 0. Note that the smoothing (4.4) can
equivalently be represented as
A(i, j; G) + τ /n
1
Tτ (i, j; G) =
= (1 − ατ )T (i, j; G) + ατ
,
(4.6)
di + τ
n
17
where the degree-adaptive regularization parameter ατ = τ (di + τ )−1 . One can construct
an equivalent and more simplified (non-adaptive) estimator by directly smoothing each row
of T (as it is a row-stochastic or row-Markov matrix) via Stein-like shrinkage
1/n 1/n . . . 1/n
.
..
..
..
Tα = (1 − α)T + α ..
(4.7)
.
.
. , 0 < α < 1.
1/n 1/n . . . 1/n
Spectral analysis can now proceed on this regularized transition probability matrix (either
(4.6) or (4.7)) by substituting empirical transition matrix Te = D−1 A.
Significance 5. The non-adaptive regularized random-walk Laplacian (transition matrix)
Tα was introduced by Larry Page and Sergey Brin in 1996 (Brin and Page, 1999) and is
called the Google’s PageRank matrix. What seems natural enough from our nonparametric
smoothing perspective, is in fact known to be highly surprising adjustment–the “Teleportation trick”.
‘Teleporting is the essential distinguishing feature of the PageRank random walk that had
not appeared in the literature before’ – Vigna (2005).
Note 8. The Google’s PageRank matrix (which is different from the empirical random
walk Laplacian in an important way) is probably the most famous, earliest, and spectacular
example of spectral regularization that was originally introduced to counter the problem of
dangling nodes (nodes with no outgoing edges) by “connecting” the graph4 .
Web is a huge heterogeneous sparse graph where dangling nodes (dead-ends) and disconnected components are quite common. As a result random walk can get stuck, and may
cycle around an isolated set of pages. Smoothing allows the random walk to teleport to a
web page uniformly at random (or adaptively based on degree-weights) whenever it hits a
dead end. The steady state vector describes the long term visit rate–the PageRank score,
computed via eigen-decomposition of Tα .
4.6
Other Generalizations
The beauty of our statistical argument is that it immediately opens up several possibilities
to construct new types of spectral regularization schemes, which are otherwise hard to guess
using previous understanding. Two such promising techniques are discussed here. The first
one deals with Stein smoothing with data-driven shrinkage parameter.
Theorem 10. Under the sparse asymptotic setup where the domain size n → ∞ at the same
rate as the sample size N , the risk (expected squared-error loss) of smooth add-b
τ probability
estimate pbτb (4.1) with the following data-driven choice
P
N 2 − ni=1 d2i
(4.8)
τb = Pn 2
n i=1 di − N 2
4
From a statistical estimation viewpoint, this can be thought of as a way to escape from the “zerofrequency problem” for discrete probability estimation.
18
is uniformly smaller than the risk of the unsmoothed MLE estimator pe = N −1 (d1 , . . . , dn ).
In addition, the risk of pbτb is uniformly smaller than the risk of the estimator formed by
choosing τ = 1/2.
For proof one can use the technique described in Fienberg and Holland (1973). As a consequence of this result, we can further improve the spectral regularization algorithm by
selecting τ in a data-driven way, with no extra computation.
While the Laplace or additive smoothing performs well in general, there are situations where
they perform poorly (Gale and Church, 1994). The Good-Turing estimator (Good, 1953) is
often the next best choice which is given by
pbGT (i; G) =
$di +1 di + 1
,
·
$di
N
(4.9)
where $k denotes the number of nodes with degree k. An excellent discussion on this topic
can be found in Orlitsky et al. (2003). One can plug in the estimate (4.9) into the equation
(4.2) to generate new spectral graph regularization technique.
Significance 6. We wish to emphasize that our novelty lies in addressing the open problem
of obtaining a rigorous interpretation and extension of spectral regularization. To the best of
our knowledge this is the first work that provides a more formal and intuitive understanding
of the origin of spectral regularization. We have shown how the regularization naturally
arises as a consequence of high-dimensional discrete data smoothing. In addition, this
point of view allows us to select appropriate regularization parameter τ with no additional
computation.
5
Application to Graph Regression
We study the problem of graph regression as an interesting application of the proposed
nonparametric spectral analysis algorithm. Unlike traditional regression settings, here, one
is given n observations of the response and predictor variables over the graph. The goal is
to estimate the regression function by properly taking into account the underlying graphstructured information along with the set of covariates.
Meuse Data Modeling. We apply our frequency domain graph analysis to address the
spatial prediction problem. Fig 2 describes the Meuse data set, a well known geostatistical
dataset. There is a considerable spatial pattern one can see from Fig 2. We seek to estimate
a smooth regression function of the dependent variable Y (zinc concentration in the soil)
via generalized spectral regression that can exploit this spatial dependency. The graph was
formed according to the geographic distance between points based on the spatial locations of
the observations. We convert spatial data into signal supported on the graph by connecting
two vertices if the distance between two stations is smaller than a given coverage radius. The
maximum of the first nearest neighbor distances is used as a coverage radius to ensure at least
19
Zinc concentrations (ppm)
Flooding frequency classes
Distance to river
Soil type classes
1000
●
● ●
1800
● ● ● ●
●
● ● ●●
●
●
●
● ●
●
●
●
●
●
●●
●
● ● ●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
● ● ●
●
● ●
● ●
●
●
●
●
● ●
●
● ● ●
●
● ●
●
●
●
●
●● ●
●
●
●
●
●
●
● ● ●
●
●
●
●
●
● ●
● ● ●
●
●
●
●
●
●
●
●● ●
●
●● ●
● ● ●
● ●
●● ●
● ●
●
●●
●
●●
●
●
●
●
●
● ●
●
●
●
● ●
●●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
● ●
●
1600
1400
1200
1000
800
600
400
● ● ● ●
●
● ● ●●
●
●
●
● ●
●
●
●
●
●
●●
●
● ● ●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
● ● ●
●
● ●
● ●
●
●
●
●
● ●
●
● ● ●
●
● ●
●
●
●
●
●● ●
●
●
●
●
●
●
● ● ●
●
●
●
●
●
● ●
● ● ●
●
●
●
●
●
●
●
●● ●
●
●● ●
● ● ●
● ●
●● ●
● ●
●
●●
●
●●
●
●
●
●
●
● ●
●
●
●
● ●
●●
●
●
●
● ●
●
●
●
●
●
200
●
●
●
●
●
● ●
●
●
●
●
1
2
3
● ● ● ●
●
● ● ●●
●
●
●
● ●
●
●
●
●
●
●●
●
● ● ●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
● ● ●
●
● ●
● ●
●
●
●
●
● ●
●
● ● ●
●
● ●
●
●
●
●
●● ●
●
●
●
●
●
●
● ● ●
●
●
●
●
●
● ●
● ● ●
●
●
●
●
●
●
●
●● ●
●
●● ●
● ● ●
● ●
●● ●
● ●
●
●●
●
●●
●
●
●
●
●
● ●
●
●
●
● ●
●●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
800
600
400
200
●
●
●
● ● ● ●
●
● ● ●●
●
●
●
● ●
●
●
●
●
●
●●
●
● ● ●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
● ● ●
●
● ●
● ●
●
●
●
●
● ●
●
● ● ●
●
● ●
●
●
●
●
●● ●
●
●
●
●
●
●
● ● ●
●
●
●
●
●
● ●
● ● ●
●
●
●
●
●
●
●
●● ●
●
●● ●
● ● ●
● ●
●● ●
● ●
●
●●
●
●●
●
●
●
●
●
● ●
●
●
●
● ●
●●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Figure 2: The Meuse dataset consists of n = 155 observations taken on a support of 15 × 15
m from the top 0−20 cm of alluvial soils in a 5×2 km part of the right bank of the floodplain
of the the river Meuse, near Stein in Limburg Province (NL). The dependent variable Y
is the zinc concentration in the soil (in mg kg−1 ), shown in the leftmost figure. The other
three variables, flooding frequency class (1 = once in two years; 2 = once in ten years; 3 =
one in 50 years), distance to river Meuse (in metres), and soil type (1= light sandy clay; 2
= heavy sandy clay; 3 =silty light clay), are explanatory variables.
one neighbor for each node. Fig 3(C) shows the sizes of neighbours for each node ranging
from 1 to 22. Three nodes (# 82, 148, and 155) have only 1 neighbour; additionally one
can see a very weakly connected small cluster of three nodes, which is completely detached
from the bulk of the other nodes. The reason behind this heterogeneous degree distribution
(as shown in Fig 3) is the irregular spatial pattern of the Meuse data.
We model the relationship between Y and spatial graph topology by incorporating nonparametrically learned spectral representation basis. Thus, we expand Y in the eigenbasis of C
and the covariates for the purpose of smoothing, which effortlessly integrates the tools from
harmonic analysis on graphs and conventional regression analysis. The model (5.1) described
in the following algorithm simultaneously promotes spatial smoothness and sparsity.
Algorithm 2. Nonparametric Spectral Graph Regression
Step 1. Input: We observe {yi ; xi1 , . . . , xip } at vertex i of the graph G = (V, A) with size
|V | = n. The regularization parameters τ .
−1/2
Step 2. Construct τ -regularized block-pulse shape functions ηj;τ = p̂j;τ I(uj−1 < u ≤ uj )
for j = 1, . . . , n.
Step 3. Compute the smoothed G-Matrix with respect to trial bases {ηk;τ }1≤k≤n
Z1
D
E
Mτ [j, k; η, G] = ηj;τ , (Cn − 1)ηk;τ
for j, k = 1, . . . , n.
L2 [0,1]
0
20
1
2
3
●
● ●● ●
●
●●
●
● ● ●●
●
●
●
● ●●
●
● ●
●●
●●
●●
●●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●●●●●
●
● ● ●
●
●●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●● ●●
●
●●
●
●
●●● ●
●
●
●
●
●
●
●
●●● ●
● ●●●
●●●
●●
●●
●●●●● ● ● ● ●●
●
●
●
● ●● ●● ●●●
●
● ● ● ●●●
●
●
●● ●● ●
●
●● ● ● ●
●
●
102
101
69
103
108
137
94
154
111
68
136
112
113
104
67131
85
114
135
86 115
107
65
116
64
63
97
142 143
96
100
98 99
74 76138 140
71
77 91 152
70 72 87 88 75 139
81
73
80
89
153
79
78
90
83
66
95
144
141 145
151
149
150
146
15
110
147
93
92
148
105
106
51
61
Frequency
62
117
50
12260 121
49
128 134
11948 127
132
126
47 44
125
12358
55
54
59
53
46
124
43 42
52 45
40
41
20
13
16
14 8
17
9
129
18
15
84
130
12
133
11 10
29
24
28 27
26
30
25
19
21
22
23
33 32
155
1
2
39
38
37
36
35 34
3
7
4
5
0
57
56
5
120
10
109
6
5
31
118
10
15
20
degree of the spatial Meuse data graph
82
Figure 3: From spatial to graph representation of the Meuse river data set. For each node
we observe {yi ; x1i , x2i , x3i }ni=1 . We address the problem of approximating the regression
function by simultaneously incorporating the effect of explanatory variables X and the
underlying spatial graph dependency.
Step 4. Construct smooth graph Fourier basis
φbk;τ (u) =
n
X
ujk ηj;τ , for k = 1, . . . , n − 1,
j=1
P
T
where M = U ΛU T =
k uk µk uk , and uij are the elements of the singular vector U =
(u1 , . . . , un ). Construct spectral basis matrix Φ = φb1;τ , . . . , φbk;τ ∈ Rn×k for the graph G.
Step 5. Construct combined graph regressor matrix XG = Φ; X , where X = [X1 , . . . Xp ] ∈
Rn×p is the matrix of predictor variables.
Step 6. Solve for βG = (βΦ , βX )T
βbG = argmin ky − XG βG k22 + λkβG k1 .
(5.1)
βG ∈Rk+p
Algorithm 2 extends traditional regression to data sets represented over a graph. The proposed frequency domain smoothing algorithm efficiently captures the spatial graph topology
via the spectral coefficients βbΦ ∈ Rk and can be interpreted as covariate-adjusted discrete
graph Fourier transform of the response variable Y . The `1 sparsity penalty automatically
selects the coefficients with largest magnitudes thus provides compression.
The following table shows that incorporating the spatial correlation in the baseline unstructured regression model using spectral orthogonal basis functions (which are estimated from
the spatial graph with k = 25) boosts the model fitting from 62.78% to 80.47%, which is an
improvement of approximately 18%.
21
X
e0 + X
Φ
e Iτ =1 + X
Φ
e Iτ =.5 + X
Φ
e Iτ =.28 + X
Φ
e II
Φ
τ =1 + X
e II
Φ
τ =.5 + X
e II
Φ
τ =.28 + X
62.78
74.80
78.16
80.47
80.43
80.37
80.45
80.43
e 0, Φ
e I and Φ
e II are the graph-adaptive piecewise constant orthogonal basis functions
Here Φ
τ
τ
derived respectively from the ordinary, Type-I regularized, and Type-II regularized Laplacian matrix. Using other forms of regularization the performance does not change much,
and thus we do not show it here. Our spectral method can also be interpreted as a kernel
smoother, where the spatial dependency is captured by the discrete GraField C . We finally
conclude that extension from traditional regression to graph-structured spectral regression
significantly improves the model accuracy.
6
Concluding Remarks
What is the “common core” behind all existing spectral graph techniques? Despite half a
century of research, it remains one of the most formidable open issues, if not the core problem
in modern spectral graph theory. The need for a unified theory become particularly urgent
and pressing in the era of “algorithm deluge” in order to systematize the practice of graph
data analysis.
As an attempt to shed some light on this previously unasked question, I have introduced
a framework for statistical reasoning that appears to offer a complete and autonomous description (totally free from the classical combinatorial or algorithmic linear algebra-based
languages) with the intention of leading the reader to a broad understanding of how different
techniques interrelate. The prescribed approach brings a fresh perspective by appropriately
transforming the spectral problem into a nonparametric approximation and smoothing problem to arrive at new algorithms, as well as both practical and theoretical results. It was a
great surprise for me to be able to deduce existing techniques from some underlying basic
principles in a self-consistent way. At the end, there is always the hope that a comprehensive understanding gained through unification will inspire new tools, to attack some of the
important open problems in this area.
Acknowledgments: The author benefited greatly from many fruitful discussions and exchanges of ideas he had at the John W. Tukey 100th Birthday Celebration meeting, Princeton University September 18, 2015. The author is particularly grateful to Professor Ronald
Coifman for suggesting the potential connection between our formulation and Diffusion map
(random walk on graph). We also thank Isaac Pesenson for helpful questions and discussions.
References
Amini, A. A., Chen, A., Bickel, P. J., Levina, E. (2013). Pseudo-likelihood methods
for community detection in large sparse networks. The Annals of Statistics, 41 2097–2122.
22
Belkin, M. and Niyogi, P. (2001). Laplacian eigenmaps and spectral techniques for
embedding and clustering. NIPS, 14 585–591.
Bishop, Y. M., Fienberg, S. E. and Holland, P. W. (2007). Discrete multivariate
analysis: theory and practice. Springer Science & Business Media.
Brin, S. and Page, L. (1999). Reprint of: The anatomy of a large-scale hypertextual web
search engine. Computer networks, 56 3825–3833.
Chaudhuri, K., Graham, F. C. and Tsiatas, A. (2012). Spectral clustering of graphs
with general degrees in the extended planted partition model. Journal of Machine Learning Research Workshop and Conference Proceedings, 35 1–23.
Chung, F. R. (1997). Spectral graph theory, vol. 92. American Mathematical Soc.
Coifman, R. R. and Lafon, S. (2006). Diffusion maps. Applied and computational
harmonic analysis, 21 5–30.
Coifman, R. R., Lafon, S., Lee, A. B., Maggioni, M., Nadler, B., Warner, F.
and Zucker, S. W. (2005). Geometric diffusions as a tool for harmonic analysis and
structure definition of data: Diffusion maps. Proceedings of the National Academy of
Sciences of the United States of America, 102 7426–7431.
Davis, E. and Sethuraman, S. (2016). Consistency of modularity clustering on random
geometric graphs. arXiv preprint arXiv:1604.03993.
Fiedler, M. (1973). Algebraic connectivity of graphs. Czechoslovak Mathematical Journal,
23 298–305.
Fienberg, S. E. and Holland, P. W. (1973). Simultaneous estimation of multinomial
cell probabilities. Journal of the American Statistical Association, 68 683–691.
Gale, W. and Church, K. (1994). What is wrong with adding one. Corpus-based research
into language, 1 189–198.
Galerkin, B. (1915). Series development for some cases of equilibrium of plates and beams
(in Russian). Wjestnik Ingenerow Petrograd, 19 897–908.
Good, I. J. (1953). The population frequencies of species and the estimation of population
parameters. Biometrika, 40 237–264.
Hoeffding, W. (1940). Massstabinvariante korrelationstheorie. Schriften des Mathematischen Seminars und des Instituts für Angewandte Mathematik der Universität Berlin, 5
179–233.
Joseph, A. and Yu, B. (2016). Impact of regularization on spectral clustering. Ann.
Statist., 44 1765–1791.
23
Laplace, P. S. (1951). A philosophical essay on probabilities, translated from the 6th
French edition by Frederick Wilson Truscott and Frederick Lincoln Emory. Dover Publications (New York, 1951).
Le, C. M., Levina, E. and Vershynin, R. (2015). Sparse random graphs: regularization
and concentration of the laplacian. arXiv:1502.03049.
Loève, M. (1955). Probability Theory; Foundations, Random Sequences. New York: D.
Van Nostrand Company.
Lovász, L. (1993). Random walks on graphs. Combinatorics, Paul erdos is eighty, 2 1–46.
Lovász, L. and Szegedy, B. (2006). Limits of dense graph sequences. Journal of Combinatorial Theory, Series B, 96 933–957.
Newman, M. E. (2006). Finding community structure in networks using the eigenvectors
of matrices. Physical review E, 74 036104.
Orlitsky, A., Santhanam, N. P. and Zhang, J. (2003). Always good turing: Asymptotically optimal probability estimation. Science, 302 427–431.
Parzen, E. (1979). Nonparametric statistical data modeling (with discussion). Journal of
the American Statistical Association, 74 105–131.
Perks, W. (1947). Some observations on inverse probability including a new indifference
rule. Journal of the Institute of Actuaries, 73 285–334.
Qin, T. and Rohe, K. (2013). Regularized spectral clustering under the degree-corrected
stochastic blockmodel. In Advances in Neural Information Processing Systems. 3120–3128.
Rohe, K., Chatterjee, S. and Yu, B. (2011). Spectral clustering and the highdimensional stochastic blockmodel. The Annals of Statistics 1878–1915.
Sarkar, P., Bickel, P. J. (2015). Role of normalization in spectral clustering for stochastic blockmodels. The Annals of Statistics, 43 962–990.
Schmidt, E. (1907). Zur Theorie der linearen und nicht linearen Integralgleichungen Zweite
Abhandlung. Mathematische Annalen, 64 433–476.
Vigna, S. (2005). Trurank: taking pagerank to the limit. 14th international conference on
World Wide Web 976–977.
Von Luxburg, U., Belkin, M. and Bousquet, O. (2008). Consistency of spectral
clustering. The Annals of Statistics 555–586.
Witten, I. H. and Bell, T. (1991). The zero-frequency problem: estimating the probabilities of novel events in adaptive text compression. Information Theory, IEEE Transactions
on, 37 1085–1094.
24
Supporting Online Material
Unified Statistical Theory of Spectral Graph Analysis
Subhadeep Mukhopadhyay
Department of Statistical Science, Temple University
Philadelphia, Pennsylvania, 19122, U.S.A.
[email protected]
To further demonstrate the potential application of smooth spectral graph algorithms, in
this supplementary section we discuss the community detection problem that seeks to divides nodes into into k groups (clusters), with larger proportion of edges inside the group
(homogeneous) and comparatively sparser connections between groups to understand the
large-scale structure of network. Discovering community structure is of great importance in
many fields such as LSI design, parallel computing, computer vision, social networks, and
image segmentation.
Graph Clustering
In mathematical terms, the goal is to recover the graph signals (class labels) y : V 7→
{1, 2, . . . , k} based on the connectivity pattern or the relationship between the nodes. Representation of graph in the spectral or frequency domain via the nonlinear mapping Φ :
G(V, E) 7→ Rm using discrete KL basis of density matrix C as the co-ordinate is the most
important learning step in the community detection. This automatically generates spectral
features {φ1i , . . . , φmi }1≤i≤n for each vertex that can be used for building the distance or
similarity matrix to apply k-means or hierarchical clustering methods. In our examples, we
will apply k-means algorithm in the spectral domain, which seeks to minimizing the withincluster sum of squares. In practice, often the most stable spectral clustering algorithms
determine k by spectral gap: k = argmaxj |λj − λj+1 | + 1.
Algorithm 3. Nonparametric Spectral Graph Partitioning
1. Input: The adjacency matrix A ∈ Rn×n . Number of clusters k. The regularization
parameter τ .
2. Estimate the top k − 1 spectral connectivity profile for each node {φe1i;τ , . . . , φe(k−1)i;τ }
using Algorithm 2. Store it in Φ ∈ Rn×k−1 .
3. Apply k-means clustering by treating each row of Φ as a point in Rk−1 .
4. Output: The cluster assignments of n vertices of the graph C1 , . . . , Ck .
S–1
In addition to graph partitioning, the spectral ensemble {λk , φk }1≤k≤m of C contain a wealth
e1 (G; η) for the choice
of information on the graph structure. For example, the quantity 1 − λ
of {ηk } to be normalized top hat basis (3.1), is referred to as the algebraic connectivity,
whose magnitude reflects how well connected the overall graph is. The kmeans clustering after spectral embedding {φej (G; η)}1≤j≤k−1 finds approximate solution to the NP-hard
combinatorial optimization problem based on the normalized cut (Shi and Malik, 2000)
by relaxing the discreteness constraints into one that is continuous (sometimes known as
spectral relaxation).
Data and Results. We investigate four well-studied real-world networks for community
structure detection based on 7 variants of spectral clustering methods.
Example A [Political Blog data, Adamic and Glance (2005)] The data, which contains
1222 nodes and 16, 714 edges, were collected over a period of two months preceding the U.S.
Presidential Election of 2004 to study how often the political blogs refer to one another. The
linking structure of the political blogosphere was constructed by identifying whether a URL
present on the page of one blog references another political blog (extracted from blogrolls).
Each blog was manually labeled as liberal or conservative by Adamic and Glance (2005),
which we take as ground truth. The goal is to discover the community structure based on
these blog citations, which will shed light on the polarization in political blogs.
Table 1 shows the result of applying the spectral graph clustering algorithm on this political
web-blog data. The un-regularized Laplacian performs very poorly, whereas as both typeI/II regularized versions give significantly better results. The misclassification error drops
from 47.95% to 4.7% because of regularization. To better understand why regularization
plays a vital role, consider the degree distribution of the web-blog network as shown in
the bottom panel of Figure 2. It clearly shows the presence of a large number of lowdegree nodes, which necessitates the smoothing of high-dimensional discrete probability
p1 , . . . , p1222 . Thus, we perform the kmeans clustering after projecting the graph in the
Euclidean space spanned by Laplace smooth KL spectral basis {φek;τ }. Regularized spectral
methods correctly identify two dense clusters: liberal and conservative blogs, which rarely
links to a blog of a different political leaning, as shown in the middle panel of Fig 5.
Example B [US College Football, Grivan and Newman (2002)] The American football
network (with 115 vertices, 615 edges) depicts the schedule of football games between NCAA
Division IA colleges during the regular season of Fall 2000. Each node represents a college
team (identified by their college names) in the division, and two teams are linked if they have
played each other that season. The teams were divided into 11 “conferences” containing
around 8 to 12 teams each, which formed actual communities. The teams in the same
conference played more often compared to the other conferences, as shown in the middle
panel of Fig 5. A team played on average 7 intra- and 4 inter-conference games in the season.
Inter-conference play is not uniformly distributed; teams that are geographically close to one
another but belong to different conferences are more likely to play one another than teams
S–2
Table 1: We report % of misclassification error. We compare following seven different
Laplacian variants. K denotes the number of communities.
Type-I Reg. Laplacian
τ = 1/2 τ =
√
N /n
Type-II Reg. Laplacian
τ =1
τ = 1/2 τ =
√
Data
K
Laplacian
τ =1
N /n
PolBlogs
2
47.95%
4.9%
4.8%
5.4%
4.8%
4.7%
5.4%
Football
11
11.3%
7.83%
6.96%
6.96%
6.96%
7.83%
7.83%
MexicoPol
2
17.14%
14.2%
14.2%
14.2%
14.2%
14.2%
14.2%
Adjnoun
2
13.4%
12.5%
12.5%
12.5%
12.5%
12.5%
12.5%
separated by large geographic distances. As the communities are well defined, the American
football network provides an excellent real-world benchmark for testing community detection
algorithms.
Table 1 shows the performance of spectral community detection algorithms to identify the 11
clusters in the American football network data. The regularization boosts the performance
√
by 3-4%. In particular, τ = 1/2 and N /n produces the best result for Type-I regularized
Laplacian, while τ = 1 exhibits the best performance for Type-II regularized Laplacian.
Example C [The Political Network in Mexico, Gil-Mendieta and Schmidt (1996)] The data
(with 35 vertices and 117 edges) represents the complex relationship between politicians in
Mexico (including presidents and their close associates). The edge between two politicians
indicates a significant tie, which can either be political, business, or friendship. A classification of the politicians according to their professional background (1 - military force, 2 civilians: they fought each other for power) is given. We use this information to compare
our 7 spectral community detection algorithms.
Although this is a “small” network, challenges arise from the fact that the two communities cannot be separated easily due to the presence of a substantial number of betweencommunity edges, as depicted in Figs 4 and 5. The degree-sparsity is also evident from Fig 5
(bottom-panel). Table 1 compares seven spectral graph clustering methods. Regularization
yields 3% fewer misclassified nodes. Both the type-I and II regularized Laplacian methods
for all the choices of τ produce the same result.
Example D [Word Adjacencies, Newman (2006)] This is a adjacency network (with 112
vertices and 425 edges) of common adjectives and nouns in the novel David Copperfield
by English 19th century writer Charles Dickens. The graph was constructed by Newman
S–3
●
5
13
24
●
1
●
2
14
30
22
●
3
●
6
●
18
4
●
7
●
21
9
●
23
19
10
●
8
●
12
17
27
34
31
28
25
35
33
20
15
26
32
16
●
11
29
Figure 4:
Mexican political network. Two different colors (golden and blue) denotes the
two communities (military and civilians).
(2006). Nodes represent the 60 most commonly occurring adjectives and nouns, and an edge
connects any two words that appear adjacent to one another at any point in the book. Eight
of the words never appear adjacent to any of the others and are excluded from the network,
leaving a total of 112 vertices. The goal is to identify which words are adjectives and nouns
from the given adjacency network.
Note that typically adjectives occur next to nouns in English. Although it is possible for
adjectives to occur next to other adjectives (e.g., “united nonparametric statistics”) or for
nouns to occur next to other nouns (e.g., “machine learning”), these juxtapositions are less
common. As expected, Fig 5 (middle panel) shows an approximately bipartite connection
pattern among the nouns and adjectives.
A degree-sparse (skewed) distribution is evident from the bottom right of Fig 5. We apply
the seven spectral methods and the result is shown in Table 1. The traditional Laplacian
yields a 13.4% misclassification error for this dataset. We get better performance (although
the margin is not that significant) after spectral regularization via Laplace smoothing.
References
Adamic, L. A. and Glance, N. (2005). The political blogosphere and the 2004 us election:
divided they blog. In Proceedings of the 3rd international workshop on Link discovery.
ACM, 36–43.
Gil-Mendieta, J. and Schmidt, S. (1996). The political network in mexico. Social
Networks, 18 355–381.
S–4
Frequency
20
30
10
8
6
Frequency
10
4
20
30
Frequency
40
50
40
12
60
500
400
50
100
150
200
250
degree of the web−blog network
300
350
0
0
2
10
0
Frequency
300
200
100
0
0
7
8
9
10
11
12
13
degree of the US college football network
5
10
15
degree of the political network in Mexico
0
10
20
30
40
50
degree of the word adjacencies network
Figure 5: The columns denote the 4 datasets corresponding to the political blog, US college
football, politicians in Mexico, and word adjacencies networks. The first two rows display the
un-ordered and ordered adjacency matrix, and the final row depicts the degree distributions.
Grivan, M. and Newman, M. E. J. (2002). Community structure in social and biological
networks. Proc. Natl. Acad. Sci 7821–7826.
Newman, M. E. (2006). Finding community structure in networks using the eigenvectors
of matrices. Physical review E, 74 036104.
Shi, J. and Malik, J. (2000). Normalized cuts and image segmentation. Pattern Analysis
and Machine Intelligence, IEEE Transactions on, 22 888–905.
S–5
| 10math.ST
|
Beyond Good and Evil
Formalizing the Security Guarantees of Compartmentalizing Compilation
Yannis Juglaret1,2
arXiv:1602.04503v6 [cs.CR] 15 Apr 2017
1
Inria Paris
2
Cătălin Hriţcu1
Arthur Azevedo de Amorim4
Université Paris Diderot (Paris 7)
Abstract—Compartmentalization is good security-engineering
practice. By breaking a large software system into mutually
distrustful components that run with minimal privileges, restricting their interactions to conform to well-defined interfaces,
we can limit the damage caused by low-level attacks such as
control-flow hijacking. When used to defend against such attacks,
compartmentalization is often implemented cooperatively by a
compiler and a low-level compartmentalization mechanism. However, the formal guarantees provided by such compartmentalizing
compilation have seen surprisingly little investigation.
We propose a new security property, secure compartmentalizing
compilation (SCC), that formally characterizes the guarantees
provided by compartmentalizing compilation and clarifies its
attacker model. We reconstruct our property by starting from
the well-established notion of fully abstract compilation, then
identifying and lifting three important limitations that make
standard full abstraction unsuitable for compartmentalization.
The connection to full abstraction allows us to prove SCC by
adapting established proof techniques; we illustrate this with
a compiler from a simple unsafe imperative language with
procedures to a compartmentalized abstract machine.
1
Introduction
Computer systems are distressingly insecure. Visiting a website, opening an email, or serving a client request is often all
it takes to be subjected to a control-hijacking attack. These
devastating low-level attacks typically exploit memory-safety
vulnerabilities such as buffer overflows, use-after-frees, or
double frees, which are abundant in large software systems.
Various techniques have been proposed for guaranteeing memory safety [12], [20], [25], [44], [51]–[55], but the challenges
of efficiency [54], [55], precision [62], scalability [74], backwards compatibility [16], and effective deployment [12], [20],
[24], [25], [44], [51]–[53] have hampered their widespread
adoption.
Meanwhile, new mitigation techniques have been proposed to deal with the most onerous consequences of memory unsafety—for instance, techniques aimed at preventing
control-flow hijacking even in unsafe settings [2], [3], [15],
[26], [68]. Unfortunately, these defenses often underestimate
the power of the attackers they may face [18], [26], [28],
[29], [32], [67]—if, indeed, they have any clear model at
all of what they are protecting against. Clarifying the precise
security properties and attacker models of practical mitigation
techniques is thus an important research problem—and a
challenging one, since a good model has to capture not only
the defense mechanism itself but also the essential features of
the complex world in which low-level attacks occur.
3
Boris Eng1,3
Université Paris 8
4
Benjamin C. Pierce4
University of Pennsylvania
In this paper we focus on the use of compartmentalization [13], [33], [70] as a strong, practical defense mechanism
against low-level attacks exploiting memory unsafety. The key
idea is to break up a large software system into mutually distrustful components that run with minimal privileges and can
interact only via well-defined interfaces. This is not only good
software engineering; it also gives strong security benefits.
In particular, control-hijacking attacks can compromise only
specific components with exploitable vulnerabilities, and thus
only give the attacker direct control over the privileges held
by these components. Also, because compartmentalization can
be enforced by more coarse-grained mechanisms, acceptable
efficiency and backwards compatibility are generally easier
to achieve than for techniques enforcing full-blown memory
safety.
When used as a defense mechanism against memory unsafety, compartmentalization is often achieved via cooperation
between a compiler and a low-level compartmentalization
mechanism [16], [33], [39], [42], [61], [70], [72]. In this paper
we use compartmentalizing compilation to refer to cooperative
implementations of this sort. The compiler might, for instance,
insert dynamic checks and cleanup code when switching
components and provide information about components and
their interfaces to the low-level compartmentalizing mechanism, which generally provides at least basic isolation. Two
such low-level compartmentalization technologies are already
widely deployed: process-level privilege separation [13], [33],
[40] (used, e.g., by OpenSSH [63] and for sandboxing plugins
and tabs in modern web browsers [64]) and software fault
isolation [69] (provided, e.g., by Google Native Client [71]);
many more are on the drawing boards [12], [16], [34], [61],
[70].
So what security guarantees does compartmentalizing compilation provide, and what, exactly, is its attacker model?
A good starting point for addressing these questions is the
familiar notion of fully abstract compilation [1], [4]–[6], [8],
[10], [11], [31], [35], [56], [58]. A fully abstract compiler
toolchain (compiler, linker, loader, and underlying architecture
with its security mechanisms) protects the interactions between
a compiled program and its low-level environment, allowing
programmers to reason soundly about the behavior of their
code when it is placed in an arbitrary target-language context,
by considering only its behavior in arbitrary source-language
contexts. In particular, if we link the code produced by
such a compiler against arbitrary low-level libraries—perhaps
compiled from an unsafe language or even written directly in
assembly—the resulting execution will not be any less secure
than if we had restricted ourselves to library code written in
the same high-level language as the calling program.
(Why is it useful to restrict attention to attackers written
in a high-level language? First, because reasoning about what
attackers might do—in particular, what privileges they might
exercise—is easier in a high-level language. And second, because by phrasing the property in terms of low- and high-level
programs rather than directly in terms of attacker behaviors,
specific notions of privilege, etc., we can re-use the same
property for many different languages.)
Since full abstraction works by partitioning the world into
a program and its context, one might expect it to apply to
compartmentalized programs as well: some set of components
that are assumed to be subject to control-hijacking attacks
could be grouped into the “low-level context,” while some
others that are assumed to be immune to such attacks would
constitute the “high-level program.” Full abstraction would
then allow us to reason about the possible behaviors of
the whole system using the simplifying assumption that the
attacker’s injected behavior for the compromised components
can be expressed in the same high-level language as the
good components. Sadly, this intuition does not withstand
closer examination. Full abstraction, as previously formulated
in the literature, suffers from three important limitations that
make it unsuitable for characterizing the security guarantees
of compartmentalizing compilation.
First, fully abstract compilation assumes that the source
language itself is secure, so that it makes sense to define
target-level security with respect to the semantics of the
source language. However, compartmentalization is often applied to languages like C and C++, which do not have
a secure semantics—the C and C++ standards leave most
of the security burden to the programmer by calling out a
large number of undefined behaviors, including memory-safety
violations, that are assumed never to occur. Valid compilers
for these languages are allowed to generate code that does
literally anything—in particular, anything a remote attacker
may want—when applied to inputs that lead to undefined
behavior. There is no way to tell, statically, whether or not a
program may have undefined behavior, and compilers do not
check for this situation. (Indeed, not only do they not check:
they aggressively exploit the assumption of no undefined
behaviors to produce the fastest possible code for well-defined
programs, often leading to easily exploitable behaviors when
this assumption is broken.) The point of compartmentalizing
compilation is to ensure that the potential effects of undefined
behavior are limited to the compromise of the component in
which it occurs: other components can only be influenced
by compromised ones via controlled interactions respecting
specified interfaces.
To characterize the security of compartmentalizing compilation, we therefore need a formal property that can meaningfully accommodate source languages in which components
can be compromised via undefined behavior. Full abstraction
as conventionally formulated does not fit the bill, because,
in order to preserve equivalences of programs with undefined
behavior, compilers must abandon the aggressive optimizations
that are the reason for allowing undefined behaviors in the
first place. To see this, consider C expressions buf[42] and
buf[43] that read at different positions outside the bounds of
a buffer buf. These two programs are equivalent at the source
level: they both lead to arbitrary behavior. However, a real C
compiler would never compile these expressions to equivalent
code, since this would require runtime checks that many C
programmers would deem too expensive.
Second, fully abstract compilation makes an open world
assumption about the attacker context. While the context is
normally required to be compatible with the protected program, for instance by respecting the program’s typed interface,
the structure and privilege of the context are unrestricted (the
full abstraction definition quantifies over arbitrary low-level
contexts). This comes in direct contradiction with the idea
of least privilege, which is crucial to compartmentalization,
and which relies on the fact that even if a component is
compromised, it does not immediately get more privilege.
Compromised components cannot change the basic rules of
the compartmentalization game. For instance, in this paper
we consider a static compartmentalization setting, in which
the breakup of the application into components is fixed in
advance, as are the privileges of each component. A security
property suitable for this setting needs to be restricted to
contexts that conform to a fixed breakup into components with
static privileges.1
Third, because the definition of full abstraction involves applying the compiler only to a program and not to the untrusted
context in which it runs, a fully abstract compiler may choose
to achieve its protection goals by introducing just a single
barrier around the trusted part to protect it from the untrusted
part [8], [45], [57], [58], [60]. Such compilation schemes force
the programmer to commit in advance to a single compromise
scenario, i.e., to a single static split of their application into a
“good” trusted program and an “evil” untrusted context from
which this program has to be protected. This is not realistic
in the setting of compartmentalizing compilation, where we
generally cannot predict which components may be vulnerable
to compromise by control hijacking attacks, and instead must
simultaneously guard against multiple compromise scenarios.
Compartmentalizing compilers allow us to build more secure
applications that go beyond the blunt trusted/untrusted distinction made by some fully abstract compilers. To describe
their guarantees accurately, we thus need a new property that
captures the protection obtained by breaking up applications
into multiple mutually distrustful components, each running
with least privilege, and that permits reasoning about multiple
scenarios in which different subsets of these components are
1 In a setting where new components can be dynamically created and
privileges can be exchanged dynamically between components, the details of
this story will be more complicated; still, we expect any secure compartmentalizing compilation property to limit the ability of low-level attacker contexts
to “forge” the privileges of existing components.
2
compromised.
Our main contribution is the definition of such a property,
which we call secure compartmentalizing compilation (SCC)
(§2). While similar in many respects to full abstraction,
our property escapes the three limitations discussed above.
First, it applies to unsafe source languages with undefined
behaviors by introducing a new notion of fully defined sets of
components. While undefined behavior is a property of whole
programs, full definedness is compositional. Intuitively, a set
of components is fully defined if they cannot be blamed [30]
for undefined behavior in any context satisfying fixed interfaces. Second, SCC makes a closed-world assumption about
compromised components, enforcing the basic rules of the
compartmentalization game like the fixed division into components and the fixed privileges of each component, including
for instance with which other components it is allowed to
interact. Third, SCC ensures protection for multiple, mutually distrustful components; it does not assume we know in
advance which components are going to be compromised (i.e.,
in the C setting, which components may contain exploitable
undefined behaviors), but instead explicitly quantifies over all
possible compromise scenarios.
Our second contribution is relating SCC to standard formulations of full abstraction both intuitively and formally
(§3). We start from full abstraction and show how the three
limitations that make it unsuitable in our setting can be
lifted one by one. This results in two properties we call
structured full abstraction and separate compilation, which
can be combined and instantiated to obtain SCC. While our
property directly captures the intuition of our attacker model,
reducing it to structured full abstraction is a useful technical
step, since the latter is easier to establish for specific examples
using a variant of existing proof techniques. Moreover, arriving
at the same property by two different paths increases our
confidence that we found the right property.
Our third contribution is establishing the SCC property
for a simple unsafe imperative language with components
interacting via procedure calls and returns, compiling to an
abstract machine with protected compartments (§4). Despite
the simplicity of the setting, this result gives useful insights.
First, the source language and compilation strategy enable
interesting attacks on components with potential buffer overflows, similar to those found in C. Second, we illustrate
how SCC can be achieved by the cooperation of a compiler
(cleaning and restoring registers) and a low-level protection
mechanism (totally isolating compartments and providing a
secure interaction mechanism using calls and returns). Third,
our SCC proof adapts a standard technique called trace semantics [37], [59], via the reduction to structured full abstraction.
The closed-world assumption about the context made by
structured full abstraction requires some nontrivial changes to
the trace semantics proof technique.
The remainder of the paper describes each of our three
contributions in detail (§2–§4) and closes by discussing related work (§5) and future directions (§6). The supplemental materials associated with this paper includes: (a) a Coq
proof for Theorem 3.4; (b) technical details and proofs for
the SCC instance from §4 (while most of these proofs are
done only on paper, the main structured full abstraction
result, Theorem 4.7, is also proved in Coq); and (c) a trace
mapping algorithm in OCaml using property-based testing to
support Assumption 4.9. These materials can be found at:
https://github.com/secure-compilation/beyond-good-and-evil
2
Secure Compartmentalizing Compilation
We start with an intuitive explanation of compartmentalizing
compilation, its attacker model, and its security benefits, and
then introduce secure compartmentalizing compilation (SCC).
We consider compartmentalization mechanisms provided by
the compiler and runtime system for an unsafe programming
language with some notion of components.2 In §4 we will
present a simple example in detail, but for the present discussion it suffices to think informally of C or C++ enriched
with some compartmentalization mechanism. This mechanism
allows security-conscious developers to break large applications into mutually distrustful components running with least
privilege and interacting only via well-defined interfaces. We
assume that the interface of each component also gives a
precise description of its privilege. Our notion of interface
here is quite generic: interfaces might include any information
that can be dynamically enforced on components, including
module signatures, lists of allowed system calls, or more detailed access control specifications describing legal parameters
to inter-component calls (e.g., ACLs for files). We assume
that the division of the application into components and the
interfaces of those components are statically determined and
fixed throughout execution. In §4, we instantiate this picture
with a rather simple and rigid notion of components and
interfaces, where components don’t directly share any state
and where the only thing one component can do to another
one is to call the procedures allowed by the interfaces of both
components.
We do not fix a specific compartmentalizing compilation
mechanism; we just assume that whatever mechanism is chosen can guarantee that, even if one component is compromised
(e.g., by a control-hijacking attack), it will still be forced
to adhere to its specified interface in its interactions with
other components. What a compromised component can do
in this model is use its access to other components, as
allowed by its interface, to trick them into misusing their
own privileges (confused deputy attacks) and/or attempt to
mount further control-hijacking attacks on other components
by communicating with them via defined interfaces.
We do not assume we know in advance which components
will be compromised: the compartmentalizing compilation
mechanism has to protect each component from all the others.
This allows developers to reason informally about various
compromise scenarios and their impact on the security of
2 We use the term “runtime system” loosely to include operating system
mechanisms [13], [33], [40], [63], [64] and/or hardware protections [12], [34],
[61], [70] that may be used by the compiler.
3
the whole application [33], relying on conditional reasoning
of the form: “If these components get taken over and these
do not, then this might happen (while that cannot), whereas
if these other components get taken over, then this other
thing might happen...” If the practical consequences of some
plausible compromise scenario are too serious, developers can
further reduce or separate privilege by narrowing interfaces
or splitting components, or they can make components more
defensive by dynamically validating the inputs they receive
from other components.
For instance, developers of a compartmentalized web
browser [64] might reason about situations in which some
subset of plugins and tabs gets compromised and how this
might impact the browser kernel and the remaining plugins and
tabs. A possible outcome of this exercise might be noticing
that, if the browser kernel itself is compromised, then all
bets are off for all the components and the application as a
whole, so the developers should put extra energy in defending
the kernel against attacks from compromised plugins or tabs.
On the other hand, if interfaces between tabs and plugins
are appropriately limited, then compromise of one should not
disrupt the rest.
Our goal is to articulate a security property that supports
reasoning about multiple compromise scenarios and clarifies
the associated attacker model. At the same time, our property
is intended to serve as a benchmark for developers of compartmentalizing compilation mechanisms who want to argue
formally that their mechanisms are secure. In the rest of this
section we explain the technical ideas behind the SCC property
and then give its formal definition.
An application is a set Cs of components, with corresponding interfaces CIs. These components are separately
compiled (individually compiling each component in the set
Cs is written Cs↓) and linked together (written ./(Cs↓)) to
form an executable binary for the application.
SCC quantifies over all compromise scenarios—i.e., over all
ways of partitioning the components into a set of compromised
ones and a set of uncompromised ones. In order to ensure that
the set of compromised components doesn’t expand during
evaluation, we require that the uncompromised components be
fully defined with respect to the interfaces of the compromised
components. That is, the uncompromised components must
not exhibit undefined behaviors even if we replace the compromised components with arbitrary code (obeying the same
interfaces).
The full definedness condition is a necessary part of the
static compromise model considered in this paper. Intuitively,
if an uncompromised component can be tricked into an undefined behavior by interface-respecting communication with
other components, then we need to conservatively assume that
the already compromised components will succeed in compromising this component dynamically, so it belongs in the set
of compromised components from the start. This static model
is much simpler to reason about than a model of dynamic
compromise, in which one could perhaps provide guarantees
to not-fully-defined components up to the point at which
they exhibit undefined behavior, but which could, however,
invalidate standard compiler optimizations that involve code
motion. Moreover, it seems highly nontrivial to define our
property for such a more complex model.
Figure 1 illustrates one way to partition five components
C1 , . . . , C5 with interfaces i1 , . . . , i5 , representing the scenario
where C2 , C4 , and C5 are compromised and C1 and C3 are
not. In order for this compromise scenario to be considered by
our property, C1 and C3 need to be fully defined with respect
to interfaces i2 , i4 , and i5 , which means C1 and C3 cannot
cause undefined behaviors when linked with any components
B2 , B4 , B5 satisfying interfaces i2 , i4 , i5 .
Formally, full definedness is a language-specific parameter
to our definition of SCC, just as the program equivalence
relations are language-specific parameters to both SCC and
vanilla full abstraction. For instance, in the simple imperative
language in §4, we will say that components Cs are fully
defined with respect to a set of adversary interfaces BIs if,
for all components Bs satisfying BIs, the complete program
./(Cs ↓∪ Bs ↓) cannot reduce to a stuck non-final state
(corresponding to undefined behavior) where the currently
executing component is one of the ones in Cs (i.e., no component in Cs can be “blamed” [30] for undefined behavior).
Full definedness might well be defined differently for another
language; for instance, in a concurrent language undefined
behaviors cannot be as easily modeled by stuckness since
normally other threads can proceed even if one of the threads
is stuck. One last thing to note is that full definedness of a
set of components is generally a much weaker property than
the full definedness of each individual component in the set.
Since the interfaces of the adversary components BIs can (and
in §4 do) restrict not only the operations they export but also
the operations they import from Cs, the components in the
set can export dangerous operations just to other components
in the set; the components actually in the set might then all
use these operations properly, whereas arbitrary components
with the same interfaces could abuse them to trigger undefined
behaviors.
SCC states that, in all such compromise scenarios, the
compiled compromised components must not be able to cause
more harm to the compiled uncompromised components via
low-level attacks than can be caused by some high-level
components written in the source language. Basically this
means that any low-level attack can be mapped back to a highlevel attack by compromised components satisfying the given
interfaces. The property additionally ensures that the high-level
components produced by this “mapping back” are fully defined
with respect to the interfaces of the uncompromised components. So with SCC, instead of having to reason about the lowlevel consequences of undefined behavior in the compromised
components, we can reason in the source language and simply
replace the compromised components by equivalent ones that
are guaranteed to cause no undefined behavior.
Formally, SCC is stated by quantifying over multiple distinguishability games, one for each compromise scenario, where
the individual games are reminiscent of full abstraction. The
4
produce different observable outcomes when executed. For
this compromise scenario, SCC specifies that, if compiled
compromised components C2↓, C4↓, C5↓ can distinguish the
C1 ↓, C3 ↓ and D1 ↓, D3 ↓ variants at the low level, then there
must exist some (fully defined) components A2 , A4 , A5 that
distinguish C1 , C3 and D1 , D3 at the high level.
With all this in mind, the SCC property is formally expressed as follows:
Compartmentalized application
i1
i2
i3
i4
i5
C1
C2
C3
C4
C5
i4
i5
Any partition
i1
C1
i2
i3
C2
C3
↯
C4
↯
Definition 2.1 (SCC).
• For any complete compartmentalized program and for all
ways of partitioning this program into a set of uncompromised components Cs and their interfaces CIs, and a
set of compromised components Bs and their interfaces
BIs, so that Cs is fully defined with respect to BIs, and
• for all ways of replacing the uncompromised components
with components Ds that satisfy the same interfaces CIs
and are fully defined with respect to BIs,
• if ./(Cs↓∪ Bs↓) 6∼L ./(Ds↓∪ Bs↓),
• then there exist components As satisfying interfaces BIs
and fully defined with respect to CIs such that
./(Cs ∪ As) 6∼H ./(Ds ∪ As).
C5
↯
With one condition: C1 and C3 fully defined
i.e. no B2, B4, and B5 can compromise C1 and C3 by
tricking them into triggering undefined behavior
i1
C1
i2
i3
B2
C3
↯
i4
B4
↯
Figure 1. Compromise scenarios
i5
B5
↯
i1
i2
i3
i4
i5
C1
A2
C3
A4
A5
i3
i4
i5
D3
A4
A5
≁
As suggested before, our property applies to any fully
defined sets of components Cs and Ds (which cannot be dynamically compromised by some components with interfaces
BIs). We conjecture that this full definedness precondition
is strictly required in the static corruption model we are
assuming. It is worth noting that we are not proposing any
method for proving that programs are fully defined; this comes
with the territory when dealing with C-like languages. What
we are after is bringing formal foundations to conditional
reasoning of the form “if these Cs are fully defined and the
remaining components Bs get compromised, then...”
Note that the Bs in our SCC definition need not be fully
defined—i.e., the property allows the compromised components to contain undefined behaviors (this may well be why
they are compromised!) and promises that, even if they do, we
can find some other components As that are able to distinguish
between Cs and Ds in the source language without causing any
undefined behaviors. Indeed, for those compromise scenarios
in which Bs are already fully defined, our SCC property
trivially follows from correct compilation (Assumption 4.2)
since in that case we can always pick As = Bs.
This generic property is parameterized over a source and a
target language with a notion of component for each, sourceand target-level notions of linking sets of components (./),
source- and target-level notions of distinguishability (6∼), a
compiler mapping source components to target components
(↓), a source-level notion of interface and an interface satisfaction relation (lifted to sets of components and interfaces),
and a notion of a set of components Cs being fully defined
with respect to a set of adversary interfaces BIs.
H
i1
i2
D1
A2
∀ low-level attack from compromised C2↓, C4↓, C5↓
∃ high-level attack from some fully defined A2, A4, A5
i1
C1↓
i1
D1↓
i2
i3
C2↓
C3↓
↯
i2
C2↓
↯
i4
≁
L
i3
D3↓
C4↓
i5
C5↓
↯
↯
↯
↯
i4
C4↓
i5
C5↓
Figure 2. SCC distinguishability game, for one of the compromise scenarios
goal of the attacker in each game is to distinguish between
two variants of the uncompromised components. Figure 2
illustrates these two variants as C1 , C3 and D1 , D3 , where we
use 6∼H and 6∼L to indicate that the behaviors of two (highor low-level) complete programs are distinguishable, i.e., they
3
From Full Abstraction to SCC
§2 presented SCC by directly characterizing the attacker model
against which it defends. In this section we step back and show
5
how SCC can instead be obtained by starting from the wellestablished notion of full abstraction and removing each of the
three limitations that make it unsuitable in our setting. This results in two properties, structured full abstraction and separate
compilation, which we then combine and instantiate to obtain
SCC. This reduction is not only theoretically interesting, but
also practically useful, since structured full abstraction can
more easily be proved by adapting existing proof techniques,
as we will see in §4.
To adapt full abstraction to a source language with undefined behaviors, we need to restrict attention only to defined
complete programs in the source language. And even with this
restriction, defining full abstraction still requires a little care.
For instance, the following variant is wrong (formally, defined
is another parameter to this property):
(∀A. A[P ] and A[Q] defined ⇒ A[P ] ∼H A[Q])
(∀a. a[P↓] ∼L a[Q↓])
⇒
Any P and Q that both trigger undefined behavior as soon as
they get control would be considered equivalent in the highlevel language because there is no context that can make these
programs defined while observing some difference between
them. All such programs would thus need to be equivalent
at the low level, which is clearly not the case (since their
nondeterminism can be resolved in different ways by the
compiler). The problem here is that if P and Q trigger
undefined behavior then the context often cannot make up for
that and make the program defined in order be able to cause
an observation that distinguishes P and Q.
Full abstraction A fully abstract compiler protects compiled
programs from their interaction with unsafe low-level code
and thus allows sound reasoning about security (and other
aspects of program behavior) in terms of the source language.
Fully abstract compilation [1] intuitively states that no lowlevel attacker can do more harm to a compiled program
than some program in the source language already could.
This strong property requires enforcing all high-level language
abstractions against arbitrary low-level attackers.
Formally, full abstraction is phrased as a distinguishability
game requiring that low-level attackers have no more distinguishing power than high-level ones.
Solution 1: Full abstraction for unsafe languages The
responsibility of keeping A[P ] defined should be thus shared
between A and P . For this we assume a compositional notion
of fully defined behavior for programs and contexts as two
parameters to Definition 3.2 below. We require that these
parameters satisfy the following properties: (1) a program
is fully defined if it does not cause undefined behavior in
any fully defined context, and (2) a context is fully defined
if it does not cause undefined behavior when we plug any
fully defined program into it. Note that properties (1) and (2)
are circular and therefore cannot be used as the definition
of full definedness. For specific languages (e.g., the one in
§4) we can break this circularity and define full definedness
using blame [30]: intuitively we call a partial program fully
defined when it cannot be blamed for undefined behavior in
any context whatsoever. Similarly, we call a context fully
defined when it cannot be blamed for undefined behavior for
any program that we plug into it. We expect such a blamebased definition to satisfy the properties (1) and (2) above.
Full definedness allows us to introduce a new variant of
full abstraction that applies to unsafe source languages with
undefined behavior:
Definition 3.1. We call a compilation function (written ↓) fully
abstract if, for all P and Q,
(∀A. A[P ] ∼H A[Q]) ⇒ (∀a. a[P↓] ∼L a[Q↓]).
Here, P and Q are partial programs, A is a high-level context
whose job is to try to distinguish P from Q, and a is a lowlevel “attacker context” that tries to distinguish P↓ from Q↓.
The relations ∼L and ∼H are parameters to the definition,
representing behavioral equivalence at the two levels. To be
useful, they should allow the context to produce an observable action every time it has control, letting it convert its
knowledge into observable behaviors. For instance, a common
choice for behavioral equivalence is based on termination: two
deterministic programs are behaviorally equivalent if they both
terminate or both diverge.
When stated this way (as an implication rather than an
equivalence), full abstraction is largely orthogonal to compiler
correctness [43], [46]. While compiler correctness is about
preserving behaviors when compiling from the source to the
target, proving full abstraction requires some way to map each
distinguishing context target to a sourge-level one, which goes
in the opposite direction. This is easiest to see by looking at
the contrapositive:
Definition 3.2 (Full abstraction for unsafe languages).
We call a compiler ↓ for an unsafe language fully abstract if
for all fully defined partial programs P and Q
∀a. a[P↓] 6∼L a[Q↓] ⇒ ∃A. A[P ] 6∼H A[Q]
(∀A. A fully defined ⇒ A[P ] ∼H A[Q]) ⇒
(∀a. a[P↓] ∼L a[Q↓]).
Problem 1: Undefined behavior The first limitation of
full abstraction is that it cannot realistically be applied to
compiling from an unsafe language with undefined behaviors. Undefined behaviors are (arbitrarily!) nondeterministic,
and no realistic compiler can preserve this nondeterminism
in the target as required by full abstraction. (Removing it
from the source language would negate the performance and
optimization benefits that are the reason for allowing undefined
behaviors in the first place.)
Requiring that P , Q, and A are fully defined means that we
can safely apply ∼H to A[P ] and A[Q], because neither the
programs nor the context can cause undefined behavior. This
property is incomparable with the original definition of full
abstraction. Looking at the contrapositive,
∀P, Q fully defined.
(∃a. a[P↓] 6∼L a[Q↓])
⇒ (∃A. A fully defined ∧ A[P ] 6∼H A[Q]),
6
fully defined with respect to contexts of shape ◦s,
∀A ∈◦ s. A fully defined wrt. programs of shape •s
⇒ A[P ] ∼H A[Q]
the P, Q fully defined pre-condition makes this weaker than
full abstraction, while the A fully defined post-condition makes
it stronger. The post-condition greatly simplifies reasoning
about programs by allowing us to replace reasoning about
low-level contexts with reasoning about high-level contexts
that cannot cause undefined behavior.
⇒ (∀a ∈◦ s. a[P↓] ∼L a[Q↓]).
This property universally quantifies over any complete program shape s and requires that P ∈• s (read “program P
has shape s”), Q ∈• s, and A ∈◦ s (“context A matches
programs of shape s”). Moreover, the property only requires
programs that are fully defined with respect to contexts of the
right shape, and dually it only considers contexts that are fully
defined with respect to programs of the right shape.
One might wonder whether the P, Q fully defined precondition is too restrictive, since full definedness is a rather
strong property, requiring each component to be very defensive
about validating inputs it receives from others. In the static
compromise model inherent to full abstraction and without
additional restrictions on the program’s context, we must
be conservative and assume that, if any context can cause
undefined behavior in a program, it can compromise it in
a way that the compiler can provide no guarantees for this
program. The structured full abstraction definition below will
in fact restrict the context and thus use a weaker notion of full
definedness. Moreover, separate compilation will allow us to
quantify over all splits of a program into a fully defined partial
program and a compromised context, which also makes the
presence of the full definedness pre-condition more palatable.
Recovering secure compartmentalizing compilation SCC
can be recovered in a natural way as an instance of structured
full abstraction (Definition 3.3). For both source and target
languages, we take partial programs and contexts be sets of
components and context application be set union. Compilation
of sets of components works pointwise. To obtain an instance
of structured full abstraction we additionally take shapes to be
sets of component interfaces, where each interface is marked
as either compromised or uncompromised.
Theorem 3.4. For any deterministic target language and any
source language that is deterministic for defined programs,
structured full abstraction instantiated to components as described above implies SCC.
Problem 2: Open-world assumption about contexts While
full abstraction normally requires the contexts to be compatible
with the partial program, for instance by respecting the partial
program’s typed interface, these restrictions are minimal and
do not restrict the shape, size, exported interface, or privilege
of the contexts in any way. This open world assumption about
contexts does not fit with our compartmentalization setting, in
which the breakup of the application into components is fixed
in advance, as are the interfaces (and thus privileges) of all the
components. In our setting, the definition of full abstraction
needs to be refined to track and respect such structural
constraints; otherwise a low-level context with 2 components
might be mapped back to a high-level context with, say, 3
components that have completely different interfaces, and thus
privileges. In particular, the high-level components’ interfaces
could give them more privileges than the low-level components
had, increasing their distinguishing power.
Proof. Straightforward, though tedious. A machine-checked
Coq proof can be found in the auxiliary materials.
Problem 3: Statically known trusted/untrusted split
While SCC can deal with multiple compromise scenarios, not
all instances of structured full abstraction can. In general, if
a compiler satisfies (structured) full abstraction, how can we
know whether it can deal with multiple compromise scenarios,
and what does that even mean? While we can instantiate full
abstraction to a particular compromise scenario by letting the
partial program P contain the uncompromised components
and the low-level context a contain the compromised ones,
a fully abstract compiler (together with its linker, loader,
runtime etc.) might exploit this static split and introduce only
one single barrier protecting the uncompromised components
from the compromised ones. When presented with a different
compromise scenario for the same program, the compiler could
adapt and produce a different output.
The source of confusion here is that a fully abstract compiler
does not need to compile contexts—only programs. In fact,
even the types of contexts and of partial programs might well
be completely different (e.g., the types of lambda calculus
contexts and terms are different; a compiler for one cannot
compile the other). Even when the types do match so that
we can apply the same compiler to the context, the low-level
context-application operation A↓[P ↓] can freely exploit the
fact that its first argument is a compiled untrusted context and
its second argument is a compiled trusted program that should
be protected from the context. So if we start with a complete
high-level program C and look at two different compromise
Solution 2: Structured full abstraction We therefore introduce a structured variant of full abstraction, in which partial
programs (indicated by • below) and contexts (◦) are assigned
dual parts of predefined complete program shapes. A shape
might be anything, from a division into components with their
interfaces (as in Theorem 3.4 below), to, e.g., the maximum
size of a component’s code after compilation (which might
expose component sizes in a setting where it’s too costly to
hide them by padding to a fixed maximum size [61]).
Definition 3.3 (Structured full abstraction).
We say that a compiler ↓ for an unsafe language satisfies
structured full abstraction if, for all program shapes s and
partial programs P ∈• s and Q ∈• s so that P and Q are
7
scenarios C = A1 [P1 ] and C = A2 [P2 ], compiling each of the
parts and combining the results using context application does
not necessarily yield the same result (i.e., it could well be that
A1↓[P1↓] 6= A2↓[P2↓]) or indeed even behaviorally equivalent
results (i.e., it could well be that A1↓[P1↓] 6∼L A2↓[P2↓]). This
means that the user of a fully abstract compiler may need to
commit in advance to a single compromise scenario.
This weakness significantly limits the applicability of full
abstraction. After all, uncertainty about sources of vulnerability is precisely the motivation for compartmentalizing
compilation: if we knew which components were safe and
which were not, there would be no reason to distinguish more
than two levels of privilege, and we could merge each group
into a single mega-component. Even in rare cases where we are
certain that some code cannot be compromised—for instance
because we have proved it safe—protecting only the verified
code from all the rest using a fully abstract compiler [7] is
still suboptimal in terms of protection, since it provides no
guarantees for all the code that is not verified.
Moreover, this weakness is not hypothetical: several fully
abstract compilers proposed in the literature are only capable of protecting a single trusted module from its untrusted
context [8], [45], [57], [58], [60] (recently proposed extensions [61] do aim at lifting this restriction in some cases).
While this setup is appropriate when all one wants to achieve
is protecting trusted (e.g., verified) code from its untrusted
context [7], it is not suitable for our compartmentalizing
compilation setting where we do not know in advance which
components will be dynamically compromised and which ones
not, and thus we want to simultaneously protect against all
possible compromise scenarios.
can distinguish P from Q when compiled to low-level code,
then there exists a fully defined context A that can distinguish
them at the high-level.
In a language without undefined behavior, this property
would trivially follow just from (whole program) correct
compilation (see Assumption 4.2 below) by picking A = B.
However, it is nontrivial for a language in which context B
might cause undefined behavior, since then correct compilation
does not apply for B[P ] and B[Q]. In our setting, this
property allows us to avoid reasoning about the implications of
undefined behavior in a low-level context and instead consider
just fully defined high-level contexts.
It is trivial to check that our instance of structured full abstraction from Theorem 3.4 does satisfy separate compilation.
It should also be easy to show that many previous fully abstract
compilers [8], [45], [57], [58], [60] do not satisfy separate
compilation, since they were not designed to support a setting
of mutual distrust.
4
A Simple Instance of SCC
In this section, we illustrate the main ideas behind SCC
with a proof-of-concept compiler from an unsafe language
with components to an abstract machine with compartments.
We discuss key design decisions for providing secure compartmentalization, such as cleaning register values to prevent
unintended communication between compartments. We also
explain how a compiler optimization for component-local calls
makes unwary compiled components vulnerable to realistic
control-flow hijacking attacks. Finally, we show how to adapt
a standard full abstraction proof technique called trace semantics [58] to prove SCC.
The results in this section have been proved on paper
under the assumptions explicitly mentioned in the text. In the
following, an assumption denotes a property that we believe
is true and rely on, but that we haven’t proved. Lemmas,
theorems and corollaries denote properties that we have proved
on paper, possibly relying on some of the stated assumptions.
The proof of the structured full abstraction theorem (Theorem 4.7) has also been formalized in Coq assuming most
other results in this section as axioms, with the exception of
Theorem 4.1 and Corollary 4.3 which are also mechanized.
While not constituting a complete machine-checked proof, this
mechanization further validates the high-level proof structure
described in this section.
Solution 3: Separate compilation We can address this by
requiring that the compiler toolchain have one additional
property:
Definition 3.5. We say that the compiler toolchain (i.e., the
compiler −↓, the linker −[−], and the runtime system embodied in the low-level behavioral equivalence) satisfies separate
compilation if
1) the type of contexts and programs is the same (so that
the compiler can also compile contexts), and
2) (A[P ])↓ ∼L A↓[P↓] for all A and P .
Requiring that context application and compilation commute
(condition 2) implies that, if some complete program C can be
written as both C = A1 [P1 ] and C = A2 [P2 ], then separately
compiling each of these splits yields behaviorally equivalent
results: (A1 [P1 ])↓ ∼L (A2 [P2 ])↓. With separate compilation,
full abstraction for an unsafe language (Definition 3.2) can be
instantiated as follows:
∀B. ∀P, Q fully defined.
((B[P ])↓ 6∼L (B[Q])↓)
⇒ (∃A. A fully defined ∧ A[P ] 6∼H A[Q])
Source Language We work with an unsafe source language
with components, procedures, and buffers. A program in
this language is a set of components communicating via
procedure calls. Buffer overflows have undefined behavior and
may open the door to low-level attacks after compilation.
However, thanks to the cooperation between the low-level
compartmentalization mechanism and the compiler, the effects
of these attacks will be limited to the offending component.
Components have statically checked interfaces that specify which procedures they import and export. To satisfy an
interface, a component can only call external procedures
One compelling reading of this is that, for all compromise
scenarios (ways to break a complete program into a compromised context B and an uncompromised program P ), and for
all programs Q that we can substitute for P , if the context B
8
that it imports explicitly, and it must define all procedures
exported by its interface. Thus, interfaces define privileges by
preventing components from calling non-imported procedures,
and enable components to define private procedures (that are
not exported in their interfaces). We will use the same notion
of interfaces in our target abstract machine.
The syntax of expressions, given below, is that of a standard imperative language with mutable buffers and mutually
recursive procedures. Each component C has local procedures
“C.P ” and private local buffers b. Loops are encoded using
recursive calls, sequencing is encoded as a binary operation,
and variables are encoded using buffers. Procedures take a
single argument, which by convention is always passed in the
first cell of the first buffer of the callee component. The only
first class values are integers i; these can be passed across component boundaries using procedure calls and returns. Buffers
and procedures are second class.
e
callee compartment is updated with the call argument, while
information about the caller’s state when performing the call
gets pushed on the call stack σ. A call completes once an
empty continuation is reached and the current expression is
a value, as is the case in the left-hand side of the second
rule. In this case, the caller’s state is restored from the call
stack, and execution resumes with the call result i as the
current expression. The intermediate steps between the start
and the end of a call reduce the procedure body to a value, as
the last rule illustrates: Whenever e is not a value, reduction
deconstructs e into a subexpression e0 and a flat evaluation
context E such that e = E[e0 ], where E[e0 ] means filling the
hole in E with e0 . This expression e0 becomes the new
currently reduced expression, while E gets appended on top
of the current continuation K. Finally, when e is a value i and
the call has not completed (K 6= []), the next step is chosen
based on the flat evaluation context found on top of K, which
gets removed from K. In the left-hand side of the first rule, for
example, this flat evaluation context is C 0 .P 0 (), for which
the next chosen step, as shown on the right-hand side, is to
start a procedure call to C 0 .P 0 , using i as the call argument.
Since undefined behaviors are allowed to take the machine
to an arbitrary low-level state, it wouldn’t make much sense
to try to make the source-language semantics describe what
can happen if an undefined point is reached. We therefore
model them at the source level simply as stuckness (as done
for instance in CompCert [47]). In particular, reduction gets
stuck when trying to access or update a buffer out of bounds,
and the type safety theorem says that well-formed programs
can only go wrong (get stuck) by reducing to an out-ofbounds operation on a buffer. A program is well-formed if
all the used buffers are defined, all imported components are
defined, all imported external procedures are public, and if the
names of all components are unique. Well-formedness extends
straightforwardly to configurations.
::= i | e1 ⊗ e2 | if e then e1 else e2 | b[e] |
b[e1 ] := e2 | C.P (e) | exit
where ⊗ ∈ {; , +, −, ×, =, ≤, . . .}.
We define a standard continuation-based small-step semantics that reduces configurations cfg. It is deterministic for
programs that don’t cause undefined behavior.
cfg ::= (C, s, σ, K, e)
E
K ::= [] | E::K
::= ⊗ e2 | i1 ⊗ | if then e1 else e2 |
b[] := e2 | b[i1 ] := | C.P ()
A configuration (C, s, σ, K, e) represents a call in progress
within component C, in which e is the expression being
reduced and K is the continuation for this expression, up to
the latest procedure call. Continuations are evaluation contexts,
here represented as lists of flat evaluation contexts E. We
denote by s a global state recording the values of the local
buffers for each component. Continuations for pending calls
are stored on a call stack σ, together with their call arguments’
values and the names of the compartments they execute in. We
omit the obvious definitions for call stacks σ and states s.
Evaluation starts as a call to a fixed procedure of a fixed
main component, and completes once this call completes,
or whenever the current expression e is exit. We illustrate
the small-step semantics with the three rules that deal with
procedure call evaluation. In these rules, ∆ is a mapping from
procedure identifiers to procedure bodies.
Theorem 4.1 (Partial type safety).
For any well-formed configuration cfg = (C, s, σ, K, e), one
of the following holds:
(1) cfg is a final configuration (either e is exit or else it is
a value and K and σ are both empty);
(2) cfg reduces in one step to a well-formed configuration;
(3) cfg is stuck and has one of the following forms:
(a) (C, s, σ, b[] :: K, i) where s[C, b, i] is undefined;
(b) (C, s, σ, b[i]:=::K, i0 ) where s[C, b, i] is undefined.
s0 = s[C 0 , 0, 0 7→ i]
σ 0 = (C, s[C, 0, 0], K)::σ
∆ ` (C, s, σ, C 0 .P 0 ()::K, i) → (C 0 , s0 , σ 0 , [], ∆[C 0 , P 0 ])
In the following, we use the term undefined behavior
configurations for the configurations described in (3), and we
say that a well-formed program is defined if reducing it never
reaches an undefined behavior configuration.
s0 = s[C 0 , 0, 0 7→ i0 ]
∆ ` (C, s, (C 0 , i0 , K)::σ, [], i) → (C 0 , s0 , σ, K, i)
Target Our compiler targets a RISC-based abstract machine
extended with a compartmentalization mechanism, inspired
by a similar design featured in previous work [12]. Each
compartment in this target has its own private memory, which
cannot be directly accessed by others via loads, stores, or
jumps. Instead, compartments must communicate using special
∆ ` (C, s, σ, K, C 0 .P 0 (e)) → (C, s, σ, C 0 .P 0 ()::K, e)
As shown on the right-hand side of the first rule, a call starts
with an empty continuation and the procedure body ∆[C 0 , P 0 ]
as the current expression. The first cell in the first buffer of the
9
The Call rule checks that the call is valid with respect to
the current compartment’s interface—i.e., the target procedure
is imported by the current compartment—which ensures that
even if a compiled component is compromised it cannot
exceed its static privilege level. Then it puts the calling compartment’s name and program counter on the global protected
call stack σ. Finally, it redirects control to the entry point of the
called procedure. The Return instruction retrieves the caller’s
compartment and return address from the protected call stack
and resumes execution there.
call and return instructions, which, as explained below, include
checks to ensure that the communication respects compartment
interfaces. (Note that this scheme requires a protected call
stack, which, in a real system, could be implemented e.g.,
using a shadow call stack [3], [27] or return capabilities [39].)
Because resource exhaustion and integer overflow issues are
orthogonal to our present concerns, we assume that words are
unbounded and memory is infinite.
The instruction set for our machine is mostly standard.
instr
::= Nop | Const i → rd | Mov rs → rd
| Load ∗ rp → rd | Store ∗ rp ← rs
| Jump r | Jal r | Call C P | Return
| Binop r1 ⊗ r2 → rd | Bnz r i | Halt
Compiler We next define a simple compiler that produces
one low-level memory compartment for each high-level component. Each compartment is internally split into buffers, the
code of procedures, and a local stack that can grow infinitely.
The local stack is used to store both intermediate results and
return addresses.
In standard calling conventions, the callee is generally
expected to restore the register values of the caller, if it has
modified them, before returning from a call. Here, however,
compiled components cannot assume that other components
will necessarily follow an agreed calling convention, so they
must save any register that may be needed later. This means,
for instance, that we save the value of the current call argument
on the local stack and write the local stack pointer to a fixed
location in the current compartment’s memory before any
cross-compartment call instruction is performed, so that the
compartment can restore them when it gets control back.
The compiler must also prevent a compromised compartment from reading intermediate states from code in other
compartments that may be in the middle of a call to this one.
Intuitively, a secure compiler must prevent compromised compartments from distinguishing compiled components based on
low-level information that (fully defined) high-level attackers
don’t get. In the source language, only a single argument or
return value is communicated at call and return points. Hence,
besides preserving their values for later, the compiler ensures
that all3 registers are cleaned before transferring control to
other compartments.
The compiler implements a simple optimization for local
calls. Since all procedures of a component live in the same
address space and local calls don’t need instrumentation, these
calls can be implemented more efficiently using Jal and Jump
instructions. We therefore use different procedure entry points
for component-local and cross-component calls, and we skip,
Const i → rd puts an immediate value i into register rd .
Mov rs → rd copies the value in rs into rd . Load ∗ rp → rd
and Store ∗ rp ← rs operate on the memory location whose
address is stored in rp (the ∗ in the syntax of Load and Store
indicates that a pointer dereference is taking place), either
copying the value found at this location to rd or overwriting
the location with the content of rs . Jump r redirects control
flow to the address stored in r. Jal r (jump-and-link) does the
same but also communicates a return address in a dedicated
register rra , so that the target code can later resume execution
at the location that followed the Jal instruction. Call C P transfers control to compartment C at the entry point for procedure
“C.P ”. Return C P transfers control back to the compartment
that called the current compartment. Binop r1 ⊗ r2 → rd
performs the mathematical operation ⊗ on the values in r1 and
r2 and writes the result to rd . Finally, Bnz r i (branch-nonzero) is a conditional branch to an offset i, which is relative
to the current program counter. If r holds anything but value
zero, the branching happens, otherwise execution simply flows
to the next instruciton.
While Jal is traditionally used for procedure calls and
Jump for returns, in this machine they can only target the
current compartment’s memory. They are nonetheless useful
for optimizing compartment-local calls, which need no instrumentation; in a realistic setting, the instrumented primitives
Call and Return would likely come with monitoring overhead.
In the low-level semantics, we represent machine states
state as (C, σ, mem, reg, pc) where C is the currently executing
compartment, mem is a partitioned memory, reg is a register
file, pc is the program counter, and σ is a global protected
call stack. We assume a partial function decode from words
to instructions. We write ψ; E ` state → state’ to mean that
state reduces to state’ in an environment where component
interfaces are given by ψ and component entry points by E.
Here are the reduction rules for Call and Return:
mem[C, pc] = i
decode i = Call C 0 P 0
pc0 = E[C 0 ][P 0 ]
0
0
0
0
C = C ∨ C .P ∈ ψ[C].import
σ = (C, pc+1) :: σ
ψ; E ` (C, σ, mem, reg, pc) → (C 0 , σ 0 , mem, reg, pc0 )
mem[C, pc] = i
decode i = Return
σ = (C 0 , pc0 ) :: σ 0
0
0
ψ; E ` (C, σ, mem, reg, pc) → (C , σ , mem, reg, pc0 )
10
3 Technically speaking, we believe that, in our very simple setting, the
compiler could choose not to clean unused registers and still be secure.
However, our proof relies on compiled components cleaning all registers
except the one that holds the call argument or return value. Indeed, not
cleaning unused registers makes things harder because it can provide a covert
channel for two compromised compartments between which interfaces would
forbid any direct communication. These compartments could now exchange
values through uncleared registers by interacting with the same unsuspecting
uncompromised component. We conjecture that this possible cooperation
between compromised components doesn’t yield more attacker power in
our case. However, in a setting where registers could be used to transmit
capabilities, this would give more power to the attacker, so our compiler
clears all registers but one, which also simplifies our proof.
for local calls, the steps that store and restore register values
and clean registers.
Because we do not check bounds when compiling buffer
read and write operations, buffer overflows can corrupt a
compartment’s memory in arbitrary ways. Consequently, many
buffer overflow attacks can be reproduced even in our simple
setting, including, due to the local-call optimization, returnoriented programming attacks [14], [66]. In return-oriented
programming, an attacker overwrites return addresses on the
local stack to produce an unexpected sequence of instructions
of his choice by reusing parts of the code of componentlocal procedures. In our setting, buffer overflow attacks thus
enable compiled components to shoot themselves in the foot
by storing beyond the end of a buffer and into the local call
stack.
We assume compiler correctness as stated below for our
compiler. Note that, in the presence of partial type safety,
(Theorem 4.1), proving either (1) or (2) below is enough to
get the other.
conditions, adapted for compromised components instead of
uncompromised ones.
We say that a high-level partial program P ∈• s is
fully defined with respect to contexts of shape ◦s when it
cannot be blamed for undefined behavior when interacting
with such contexts: for every A ∈◦ s, either reducing A[P ]
never reaches an undefined behavior configuration, or else the
current component in this undefined behavior configuration
belongs to A. Similarly, a high-level context A ∈• s is fully
defined with respect to programs of shape ◦s when it cannot
be blamed for undefined behavior when interacting with such
programs.
Because we perform a point-wise compilation of highlevel programs, separate compilation (Definition 3.5) trivially
holds for our compiler. Combining it with whole-program
compiler correctness (Assumption 4.2) immediately leads to
the following corollary:
Corollary 4.3 (Separate compilation correctness).
∀P. P defined ⇒
(1) P terminates ⇐⇒ P↓ terminates ∧
(2) P diverges ⇐⇒ P↓ diverges
∀s, A ∈◦ s, P ∈• s.
P fully defined wrt. contexts of shape ◦s ⇒
A fully defined wrt. programs of shape •s ⇒
(1) A[P ] terminates ⇐⇒ A↓ [P↓] terminates ∧
(2) A[P ] diverges ⇐⇒ A↓ [P↓] diverges
Instantiating structured full abstraction We define program shapes, partial programs, and contexts in a similar way
to Theorem 3.4, as detailed below. More precisely, we use
isomorphic definitions so that we can later apply this theorem.
A program shape s is the pairing of a mapping from
component names to component interfaces and a set that
indicates uncompromised components. In the rest of the paper,
we implicitly restrict our attention to well-formed shapes. A
shape is well-formed when (1) all component interfaces in the
shape only import procedures from components that are part
of the shape, and (2) these procedures are exported according
to the shape.
High-level partial programs P and contexts A are defined
as mappings from component names to component definitions.
A high-level partial program P has shape •s when it defines
exactly the components that are marked as uncompromised in
s, with definitions that satisfy the corresponding interfaces, and
when it moreover satisfies the simple well-formedness condition that all the local buffers it uses are defined. A high-level
context A has shape ◦s under the same conditions, adapted for
compromised components instead of uncompromised ones.
A low-level partial program p or context a is formed by
pairing a partitioned memory with a mapping from procedure identifiers to entry points. This choice is isomorphic
to having sets of named compartment memories with entry
point annotations. A low-level partial program p has shape
•s when the partitioned memory has partitions under exactly
the component names that are marked as uncompromised in s,
and the entry point mapping provides addresses for exactly the
procedures that are exported by these components according
to s. A low-level context a has shape ◦s under the same
Proof technique for structured full abstraction Trace
semantics were initially proposed by Jeffrey and Rathke [36],
[37] to define fully abstract models for high-level languages.
Patrignani et al. later showed how to use trace semantics [59]
to prove full abstraction for a compiler targeting machine
code [58].
This proof technique is well suited for deterministic target languages such as machine code and proceeds in two
steps. First, we devise a trace semantics for low-level partial
programs and contexts and relate it to the target machine’s
operational semantics (e.g., by proving it fully abstract [59]).
This trace semantics will provide a set of traces for every
partial program, describing all the execution paths that this
program can reach by interacting with an arbitrary context.
Second, we use the trace semantics to characterize the interaction between an arbitrary low-level context and, separately, two
compiled programs that this context distinguishes, resulting in
two traces with a common prefix followed by different actions.
We can then use these traces to construct a high-level attacker,
proving that this attacker distinguishes between the two source
programs.
As our proof demonstrates, proving the trace semantics fully
abstract is not a mandatory first step in the technique. Instead,
we relate our trace semantics to the operational one using
two weaker trace composition and decomposition conditions
(Lemma 4.5 and Lemma 4.6), adapted from the key lemmas
that Jeffrey and Rathke used to prove their trace semantics
fully abstract [36], [37]. This reduces proof effort, since
proving a trace semantics fully abstract typically requires
proving a third lemma with a trace-mapping argument of its
own [36], [37], [59].
Assumption 4.2 (Whole-program compiler correctness).
11
Adapting the technique to undefined behavior is straightforward, essentially amounting to proving standard full abstraction for the safe subset of the language. Then one simply
proves that the context produced by the mapping is fully
defined, thus safe. Adapting to a closed world, however, takes
more work.
The trace semantics that have been used previously to
prove full abstraction characterize the interaction between a
partial program and arbitrary contexts. The context’s shape
is typically constructed as reduction goes, based on the steps
that the context takes in the generated trace. For instance, if
the trace said that the context performs a call, then the target
procedure would be appended to the context’s interface so that
this call becomes possible. For structured full abstraction, we
want a finer-grained trace semantics that enables reasoning
about the interaction with contexts of a specific shape. We
achieve this by making the shape a parameter to the reduction
relation underlying our trace semantics. To make sure that
traces are compatible with this shape, we also keep track
of the current compartment during reduction. This allows
us to generate only context steps that adhere to the current
compartment’s interface, and hence to the context’s shape.
In particular, the context will only be able to call program
procedures for which (1) there is a context compartment whose
interface explicitly imports the target procedure, thus granting
the privilege to call that procedure, and (2) this other context
compartment is reachable from the current compartment via a
chain of cross-compartment calls or returns within the context.
Moving to a closed world also makes the trace mapping
argument harder. The one from Patrignani et al. [58], for
instance, relies on changes in the context’s shape, e.g., adding
a helper component to the context that is not present in the low
level. This is no longer possible for structured full abstraction,
where the context shape is fixed.
game). It corresponds to the use of an instruction that makes
execution stuck, such as Halt.
At any point where it has control, a player can take internal
actions (any instruction that neither terminates execution nor
transfers control to the opponent); these are not reflected in
the trace. In particular, cross-compartment communication is
considered an internal action when it transfers control to
a compartment that belongs to the current player. Besides
halting, a player can also end the game by triggering an infinite
sequence of internal actions, making execution diverge. In the
trace, this will correspond to not making any move: the trace
observed thus far will be a maximal trace for the interaction
between the program and context involved, i.e., any extension
of this trace will not be shared by both the program and the
context.
Intuitively, a program p ∈• s has trace t if it answers with
the program actions described in t when facing a context a ∈◦
s that produces the context actions described in t. Similarly,
a program a ∈◦ s has trace t if it answers with the context
actions described in t when facing a program p ∈• s that
produces the program actions described in t. We define Tr◦s (p)
to be the set of traces of a partial program p with respect to
contexts of shape ◦s, and Tr•s (a) to be the set of traces of a
context a with respect to programs of shape •s.
The player that starts the game is the one that owns the
main component according to s. For each player, the trace
semantics is deterministic with respect to its own actions
and nondeterministic with respect to the opponent’s actions.
All possible actions an actual opponent could take have a
corresponding nondeterministic choice, which is formalized
by a property we call trace extensibility.
Lemma 4.4 (Trace extensibility).
∀t, s, p ∈• s, a ∈◦ s.
(t ∈ Tr◦s (p) ∧ t.γ? ∈ Tr•s (a) ⇒ t.γ? ∈ Tr◦s (p)) ∧
(t ∈ Tr•s (a) ∧ t.γ! ∈ Tr◦s (p) ⇒ t.γ! ∈ Tr•s (a))
Trace semantics for the low-level language We define
a trace semantics in which traces are finite words over an
alphabet Eα of external actions, alternating between program
external actions “γ!” and context external actions “γ?”. We
treat external actions as moves in a two-player game, viewing
the context and the partial program as the players. The trace
semantics is parameterized by a shape s, which the two players
have. External actions either transfer control to the other player
or end the game.
Eα ::= γ! | γ?
Nondeterminism disappears once we choose a particular
opponent for a player, as the two key lemmas below illustrate.
Lemma 4.5 (Trace decomposition).
∀s, p ∈• s, a ∈◦ s. a[p] terminates ⇒
∃t. t ends with X ∧ t ∈ Tr◦s (p) ∩ Tr•s (a)
Trace decomposition is stated for terminating programs. It
extracts the interaction between a program p and a context a
with dual shapes by looking at how a[p] reduces, synthesizing
that interaction into a trace t. Because execution terminates,
this trace ends with a termination marker.
γ ::= Callreg C P | Returnreg | X
Traces (Eα∗ ) track the external actions (γ) performed by the
context and the program. The first kind of external action
is cross-boundary communication, which corresponds to the
use of instrumented call instructions Call C P and Return
when they transfer control to a compartment that belongs to
the opponent. For these external actions, traces keep track of
the instruction used together with reg, the values held by all
registers when the instruction is issued. The second kind of
external action is program termination, which we denote with
a tick X and which the opponent cannot answer (X ends the
Lemma 4.6 (Trace composition).
∀t, s, p ∈• s, a ∈◦ s. t ∈ Tr◦s (p) ∩ Tr•s (a) ⇒
(∀Eα. (t.Eα) 6∈ Tr◦s (p) ∩ Tr•s (a)) ⇒
(a[p] terminates ⇐⇒ t ends with X)
Trace composition is the opposite of trace decomposition,
reconstructing a sequence of reductions based on synthesized
interaction information. It considers a program and a context
12
with dual shapes, that share a common trace t. The condition
on the second line states that the game has ended: trace t
cannot be extended by any action Eα such that the two
players share trace “t.Eα”. Under these assumptions, trace
composition tells us that one of the following holds: either (1)
the trace ends with a termination marker X and putting p in
context a will produce a terminating program, or (2) the trace
does not end in X and putting p in context a will produce
a diverging program. Intuitively, if the game has ended but
there is no termination marker, it must be because one of the
players went into an infinite sequence of internal actions and
will neither give control back nor terminate.
While the statement of these lemmas is quite close to that
used in an open world setting [36], [37], the trace semantics
itself has to be adapted in order to prove them in the presence
of our closed world assumption. To this end, we incorporate
internal actions within the trace semantics, thus adding more
options to the nondeterministic choice of the next context
action, which allows us to track at any point the currently
executing compartment. When in control, a player can only
perform communicating actions allowed by the interface of
the current compartment. This restricts external actions as
required, while also making it possible to internally switch the
current compartment through allowed internal actions. Using
our semantics, we thus end up with finer-grained traces that
include internal communication, which can be directly mapped
to high-level attackers (Assumption 4.9). The traces we use
otherwise are obtained by erasing internal actions from the
finer-grained traces.
traces gives enough distinguishing power to the high-level
context. The reason is that, because they can’t trust other
compartments, compiled fully defined components never read
values transmitted in registers with the exception of the one
used for communication. As a consequence, these components
cannot distinguish context external actions based on the content of these unread registers, which are exactly the ones a
compiled fully defined context cleans. Fully defined programs
thus perform the exact same actions when facing a trace t or
its canonicalization ζ◦ (t), as formalized by Lemma 4.8. This
means that having the high-level attacker reproduce canonical
traces instead of the original traces of the low-level context
will be enough to lead compiled programs into reproducing
the actions they took when facing the low-level context.
Lemma 4.8 (Canonicalization).
∀t, s, P ∈• s.
P fully defined wrt. contexts of shape ◦s ⇒
t ∈ Tr◦s (P↓) ⇐⇒ ζ◦ (t) ∈ Tr◦s (P↓)
The definability assumption below gives a characterization
of our mapping from a canonical trace t and an action γ1 to a
compiled fully defined context A↓ that reproduces the context
actions in t and, depending on the next action γ the program
takes, ends the game with either termination (if ζ(γ) = ζ(γ1 ))
or divergence (if ζ(γ) 6= ζ(γ1 )). The context A ↓ will thus
distinguish a program p producing trace “t.γ1 !” from any
program producing “t.γ!” with ζ(γ) 6= ζ(γ1 ).
Assumption 4.9 (Definability).
∀t, γ1 , s. t = ζ◦ (t) ∧ (∃p ∈• s. (t.γ1 !) ∈ Tr◦s (p)) ⇒
∃A ∈◦ s. A fully defined wrt. programs of shape •s ∧
(1) t ∈ Tr•s (A↓) ∧
(2) (γ1 6= X ⇒ (t.γ1 !.X?) ∈ Tr•s (A↓)) ∧
(3) ∀γ. if ζ(γ) 6= ζ(γ1 ) then ∀γ 0 . (t.γ!.γ 0 ?) 6∈ Tr•s (A↓)
Proof of SCC We prove our instance of structured full
abstraction, which implies SCC by Theorem 3.4 since we have
isomorphic definitions to the ones in §3.
Theorem 4.7 (Structured full abstraction).
Our compiler satisfies structured full abstraction.
The definability assumption gives us a fully defined context
that follows trace t (part 1) and that, if given control afterwards
via action “γ!” such that γ 6= X, acts as follows: if γ = γ1 the
context terminates (2) and if the context can distinguish γ from
γ1 , it will make execution diverge by not issuing any action
γ 0 (3). Since it is a compiled fully defined context, A↓ can
only access values transmitted using register rcom , the register
that holds the call argument or return value. So A↓ can only
distinguish between γ and γ1 when they differ in rcom , which
is captured formally by the ζ(γ) 6= ζ(γ1 ) condition.
Proving this assumption (even on paper) would be quite
tedious, so we settled for testing its correctness using
QuickCheck [17]. We built an algorithm (in OCaml) that
constructs A out of t. More precisely, the algorithm inputs a
trace with internal actions (the finer-grained trace that erases
to t) and builds a context A that reproduces context internal
and external actions as prescribed by that trace. Execution
will afterwards resume at a different point in A depending
on the next action taken by the program. At each such
point, A will either terminate execution or make it diverge
depending on whether the program action is distinguishable
Recall that the basic idea behind the proof technique is to
extract two traces that characterize the interaction between a
low-level context and two compiled fully defined high-level
programs, and then to map these two traces to a fully defined
high-level context. The high-level context should reproduce
the context actions described in the traces when facing the
same programs as the low-level context.
Unfortunately, a compiled fully defined context cannot
reproduce any arbitrary low-level trace, because the values
transmitted in registers are part of external communication
actions in low-level traces: As enforced by the compiler,
these contexts always clear all registers but the one used for
communication before giving control to the program. They can
thus only produce traces in which registers are cleared in all
context actions, which we call canonical traces. We denote by
ζ(γ) the operation that rewrites action γ so that all registers
but one are clear. A canonical trace ζ◦ (t) can be obtained from
an arbitrary trace t by replacing all context actions “γ?” by
“ζ(γ)?”. We call this operation trace canonicalization.
As we will see, being able to reproduce arbitrary canonical
13
from action γ1 . Because the trace taken as input already
includes internal actions, we do not have to reconstruct them,
hence our algorithm is not more difficult to devise than in an
open-world setting [58]. In the following, we assume that the
algorithm is correct, i.e., that Assumption 4.9 holds. We can
now turn to the main theorem.
After following trace tc , which Q↓ has from (a), Q↓ cannot
perform a terminating action: otherwise using (b) and trace
extensibility for a and tp , we could apply trace composition
to trace “tp .X” and get that a[Q ↓] terminates, which is a
contradiction. Q↓ cannot perform action γ1 either, since (c)
would then violate the fact that tp is the longest prefix of ti
in Tr◦s (Q↓). So Q↓ only has two options left. The first is to
perform no external action by going into an infinite sequence
of internal transitions. In this case, using (1), we can apply
trace composition to get that A↓ [Q↓] diverges. The second
option is to give control back to the context using an external
action γ so that X 6= γ 6= γ1 . Because fully defined compiled
programs clean registers, they only yield canonical actions, i.e.
γ = ζ(γ) ∧ γ1 = ζ(γ1 ). Combined with (3), this entails that
if A↓ produced an action γ 0 , we would have γ = γ1 , which
is false. Hence, A↓ doesn’t produce any action: it goes into
an infinite sequence of local transitions. We can again apply
trace composition to get that A↓ [Q↓] diverges.
We finally apply separate compiler correctness (Corollary 4.3) to conclude the proof.
Detailed proof of structured full abstraction. Consider a lowlevel attacker a ∈◦ s distinguishing two fully defined partial
programs P, Q ∈• s after compilation. Suppose without loss
of generality that a[P ↓] terminates and a[Q↓] diverges. We
build a high-level attacker A ∈◦ s that is fully defined with
respect to programs of shape •s and can distinguish between
P and Q.
We can first apply trace decomposition (Lemma 4.5) to a
and P ↓ to get a trace ti ∈ Tr◦s (P ↓) that ends with X, such
that ti ∈ Tr•s (a). Call tp the longest prefix of ti such that tp ∈
Tr◦s (Q↓). Because trace sets are prefix-closed by construction,
we know that tp ∈ Tr◦s (P↓) ∩ Tr•s (a).
Moreover, tp is necessarily a strict prefix of ti : otherwise,
we could apply trace composition (Lemma 4.6) and get that
a[Q↓] terminates, a contradiction. So there exists an external
action Eα such that trace “tp .Eα” is a prefix of ti . Now
Eα cannot be a context action, or else trace extensibility
(Lemma 4.4) would imply that “tp .Eα” is a trace of Tr◦s (Q↓),
which is incompatible with tp being the longest prefix of ti in
Tr◦s (Q↓). Therefore, Eα is a program action, i.e., there exists
γ1 such that “Eα = γ1 !”. Intuitively, P↓ and Q↓ take the same
external actions until the end of tp , where P ↓ takes external
action “γ1 !” and Q↓ does not (it takes either a different action
γ 6= γ1 or no external action at all).
Now, let tc be the canonicalization of trace tp , i.e.,
tc = ζ◦ (tp ). By canonicalization (Lemma 4.8), “tc .γ1 !” =
ζ◦ (tp .γ1 !) is a trace of P↓. We can thus use apply definability
(Assumption 4.9) to trace tc and action γ1 , using P ↓∈• s
as a witness having trace “tc .γ1 !”. This yields a fully defined
context A ∈◦ s such that:
5
Related Work
Fully abstract compilation Fully abstract compilation was
introduced in the seminal work of Martı́n Abadi [1] and later
investigated by the academic community. (Much before this,
the concept of full abstraction was coined by Milner [48].)
For instance, Ahmed et al. [9]–[11], [56] proved the full
abstraction of type-preserving compiler passes for functional
languages and devised proof techniques for typed target languages. Abadi and Plotkin [6] and Jagadeesan et al. [35]
expressed the protection provided by a mitigation technique
called address space layout randomization as a probabilistic
variant of full abstraction. Fournet et al. [31] devised a fully
abstract compiler from a subset of ML to JavaScript.
Patrignani et al. [45], [58] were recently the first to study
fully abstract compilation to machine code, starting from
single modules written in simple, idealized object-oriented
and functional languages and targeting hardware architectures
featuring a new coarse-grained isolation mechanism. Until
recently, they studied fully abstract compilers that by design
violate our separate compilation property, so they cannot be
applied to our compartmentalizing compilation setting.
In recent parallel work, Patrignani et al. [61] proposed an
extension of their compilation scheme to protecting multiple
components from each other. The attacker model they consider is different, especially since their source language does
not have undefined behavior. Still, if their source language
were extended with unsafe features, our SCC property could
possibly hold for their compiler.
Patrignani, Devriese et al. [21], [59] also proposed proof
techniques for full abstraction that work for untyped target
languages, and more recently proved full abstraction by approximate back-translation for a compiler between the simply
typed and the untyped lambda calculus and fully formalized
this proof in Coq [22]. As opposed to our Coq proof for
the instance from §4, which is done under well-specified
(1) tc ∈ Tr•s (A↓),
(2) γ1 =
6 X ⇒ (tc .γ1 !.X?) ∈ Tr•s (A↓),
(3) ∀γ, γ 0 . (tc .γ!.γ 0 ?) ∈ Tr•s (A↓) ⇒ ζ(γ) = ζ(γ1 ).
We now show that these conditions imply that A ↓ [P ↓]
terminates while A↓ [Q↓] diverges.
First, we look at P↓. Consider the case where γ1 = X. In
this case, by applying trace extensibility to A↓ in (1), we get
that “tc .X!” is a trace of A↓, so trace composition allows us
to conclude that A↓ [P↓] terminates. Now if γ1 6= X then this
action gives back control to the context, which, given (2), will
perform action “X?”. Applying trace extensibility to P↓, P↓
has trace “tc .γ1 !.X?”, so we can apply trace composition and
deduce that A↓ [P↓] terminates in this case as well.
Now, regarding Q ↓, we first obtain the following by
applying canonicalization to tp , “tp .X!”, and “tp .γ1 !”:
(a) tc = ζ◦ (tp ) ∈ Tr◦s (Q↓),
(b) (tc .X!) = ζ◦ (tp .X!) ∈ Tr◦s (Q↓) ⇒ (tp .X!) ∈ Tr◦s (Q↓),
(c) (tc .γ1 !) = ζ◦ (tp .γ1 !) ∈ Tr◦s (Q↓) ⇒ (tp .γ1 !) ∈ Tr◦s (Q↓).
14
assumptions but not complete, the proof of Devriese et al.
is assumption free.
including: process-level sandboxing [13], [33], [40], [63], software fault isolation (SFI) [71], capability machines [70], and
multi-PMA systems [61]. As we target lower-level machines,
new problems will appear: for instance we need to deal with
the fact that memory is finite and resource exhaustion errors
cannot be hidden from the attacker, which will require slightly
weakening the security property. Finally, we would like to
study more interesting compartmentalization models including
dynamic component creation and nested components, and the
way these extensions influence the security property.
Formal reasoning about compartmentalized code SCC is
orthogonal to formal techniques for reasoning about compartmentalized software: SCC allows transferring security guarantees for compartmentalized code written in a source language
to machine code via compartmentalizing compilation, but SCC
itself does not provide effective reasoning principles to obtain
those security guarantees in the first place. The literature
contains interesting work on formally characterizing the security benefits of compartmentalization. Promising approaches
include Jia et al.’s work on System M [38], and Devriese et
al.’s work on logical relations for a core calculus based on
JavaScript [23], both of which allow bounding the behavior
of a program fragment based on the interface or capabilities
it has access to. One significant challenge we attack in this
paper is languages with undefined behaviors, while in these
other works illegal actions such as accessing a buffer out of
bounds must be detected and make the program halt.
Acknowledgments We thank André DeHon, Deepak Garg,
and Andrew Tolmach for helpful discussions and thoughtful
feedback on earlier drafts. Yannis Juglaret is supported by a
PhD grant from the French Department of Defense (DGA) and
Inria. Arthur Azevedo de Amorim and Benjamin C. Pierce
are supported by NSF award 1513854, Micro-Policies: A
Framework for Tag-Based Security Monitors.
References
[1] M. Abadi. Protection in programming-language translations. Research
Report 154, SRC, 1998.
[2] M. Abadi, M. Budiu, Ú. Erlingsson, and J. Ligatti. Control-flow integrity.
ACM CCS. 2005.
[3] M. Abadi, M. Budiu, Ú. Erlingsson, and J. Ligatti. Control-flow integrity
principles, implementations, and applications. TISSEC, 13(1), 2009.
[4] M. Abadi and J. Planul. On layout randomization for arrays and functions.
POST. 2013.
[5] M. Abadi, J. Planul, and G. D. Plotkin. Layout randomization and
nondeterminism. MFPS, 298:29–50, 2013.
[6] M. Abadi and G. D. Plotkin. On protection by layout randomization.
TISSEC, 15(2):8, 2012.
[7] P. Agten, B. Jacobs, and F. Piessens. Sound modular verification of C
code executing in an unverified context. POPL. 2015.
[8] P. Agten, R. Strackx, B. Jacobs, and F. Piessens. Secure compilation to
modern processors. CSF. 2012.
Verifying correct low-level compartmentalization Recent
work focused on formally verifying the correctness of lowlevel compartmentalization mechanisms based on software
fault isolation [42], [49], [73] or tagged hardware [12]. That
work, however, only considers the correctness of the lowlevel compartmentalization mechanism, not the compiler and
not high-level security properties and reasoning principles for
code written in a programming language with components.
Communication between low-level compartments is generally
done by jumping to a specified set of entry points, while
the model we consider in §4 is more structured and enforces
correct calls and returns.
Finally, seL4 is a verified operating system microkernel [41], that uses a capability system to separate user level
threads and for which correct access control [65] and noninterference properties [50] were proved formally.
6
[9] A. Ahmed. Verified compilers for a multi-language world. SNAPL. 2015.
[10] A. Ahmed and M. Blume. Typed closure conversion preserves observational equivalence. ICFP. 2008.
[11] A. Ahmed and M. Blume. An equivalence-preserving CPS translation
via multi-language semantics. ICFP. 2011.
[12] A. Azevedo de Amorim, M. Dénès, N. Giannarakis, C. Hriţcu, B. C.
Pierce, A. Spector-Zabusky, and A. Tolmach. Micro-policies: Formally
verified, tag-based security monitors. In 36th IEEE Symposium on
Security and Privacy (Oakland S&P). 2015.
[13] A. Bittau, P. Marchenko, M. Handley, and B. Karp. Wedge: Splitting
applications into reduced-privilege compartments. USENIX NSDI, 2008.
Conclusion and Future Work
We have introduced a new secure compartmentalizing compilation property, related it to the established notion of full
abstraction, and applied our property in a carefully simplified
setting: a small imperative language with procedures compiling to a compartmentalized abstract machine. This lays the
formal foundations for studying the secure compilation of
mutually distrustful components written in unsafe languages.
In the future we plan to build on this groundwork to study
more realistic source and target languages, compilers, and enforcement mechanisms. In the long run, we would like to apply
this to the C language by devising a secure compartmentalizing
variant of CompCert that targets a tag-based reference monitor [12] running on a real RISC processor [19]. We have in fact
started working towards this long term goal [39], but this will
take time to achieve. Beyond tagged hardware, we would also
like to implement the abstract compartmentalization machine
from §4 in terms of various enforcement other mechanisms,
[14] E. Buchanan, R. Roemer, H. Shacham, and S. Savage. When Good Instructions Go Bad: Generalizing Return-Oriented Programming to RISC.
In Proc. ACM CCS, 2008.
[15] N. Burow, S. A. Carr, S. Brunthaler, M. Payer, J. Nash, P. Larsen, and
M. Franz. Control-flow integrity: Precision, security, and performance.
CoRR, abs/1602.04056, 2016.
[16] D. Chisnall, C. Rothwell, R. N. M. Watson, J. Woodruff, M. Vadera,
S. W. Moore, M. Roe, B. Davis, and P. G. Neumann. Beyond the PDP-11:
Architectural support for a memory-safe C abstract machine. ASPLOS.
2015.
[17] K. Claessen and J. Hughes. QuickCheck: a lightweight tool for random
testing of Haskell programs. ICFP. 2000.
[18] L. Davi, A. Sadeghi, D. Lehmann, and F. Monrose. Stitching the gadgets:
On the ineffectiveness of coarse-grained control-flow integrity protection.
USENIX Security, 2014.
15
[19] A. DeHon, E. Boling, R. Nikhil, D. Rad, J. Schwarz, N. Sharma, J. Stoy,
G. Sullivan, and A. Sutherland. DOVER: A Metadata-Extended RISCV. In RISC-V Workshop, 2016. Accompanying talk at http://youtu.be/
r5dIS1kDars.
Low-fat pointers: compact encoding and efficient gate-level implementation of fat pointers for spatial safety and capability-based security. CCS.
2013.
[45] A. Larmuseau, M. Patrignani, and D. Clarke. A secure compiler for ML
modules. APLAS, 2015.
[46] X. Leroy. Formal verification of a realistic compiler. CACM, 52(7):107–
115, 2009.
[47] X. Leroy. A formally verified compiler back-end. JAR, 43(4):363–446,
2009.
[48] R. Milner. Processes: a mathematical model of computing agents, pages
157–174. North-Holland, 1975.
[49] G. Morrisett, G. Tan, J. Tassarotti, J.-B. Tristan, and E. Gan. RockSalt:
better, faster, stronger SFI for the x86. PLDI. 2012.
[50] T. C. Murray, D. Matichuk, M. Brassil, P. Gammie, T. Bourke,
S. Seefried, C. Lewis, X. Gao, and G. Klein. seL4: from general purpose
to a proof of information flow enforcement. IEEE S&P. 2013.
[51] S. Nagarakatte, M. M. K. Martin, and S. Zdancewic. Hardware-Enforced
Comprehensive Memory Safety. IEEE Micro, 33(3):38–47, 2013.
[52] S. Nagarakatte, M. M. K. Martin, and S. Zdancewic. WatchdogLite:
Hardware-accelerated compiler-based pointer checking. CGO. 2014.
[53] S. Nagarakatte, M. M. K. Martin, and S. Zdancewic. Everything you
want to know about pointer-based checking. SNAPL. 2015.
[20] J. Devietti, C. Blundell, M. M. K. Martin, and S. Zdancewic. HardBound: Architectural support for spatial safety of the C programming
language. ASPLOS, 2008.
[21] D. Devriese, M. Patrignani, and F. Piessens. Fully-abstract compilation
by approximate back-translation. POPL, 2016.
[22] D. Devriese, M. Patrignani, F. Piessens, and S. Keuchel. Modular, fullyabstract compilation by approximate back-translation. arXiv:1703.09988,
2017.
[23] D. Devriese, F. Piessens, and L. Birkedal. Reasoning about object
capabilities with logical relations and effect parametricity. EuroS&P,
2016.
[24] U. Dhawan, C. Hriţcu, R. Rubin, N. Vasilakis, S. Chiricescu, J. M. Smith,
T. F. Knight, Jr., B. C. Pierce, and A. DeHon. Architectural support for
software-defined metadata processing. ASPLOS, 2015.
[25] U. Dhawan, A. Kwon, E. Kadric, C. Hriţcu, B. C. Pierce, J. M. Smith,
A. DeHon, G. Malecha, G. Morrisett, T. F. Knight, Jr., A. Sutherland,
T. Hawkins, A. Zyxnfryx, D. Wittenberg, P. Trei, S. Ray, and G. Sullivan.
Hardware support for safety interlocks and introspection. AHNS, 2012.
[26] Ú. Erlingsson. Low-level software security: Attacks and defenses. In
Foundations of Security Analysis and Design IV, FOSAD 2006/2007
Tutorial Lectures, 2007.
[27] Ú. Erlingsson, M. Abadi, M. Vrable, M. Budiu, and G. C. Necula. XFI:
Software guards for system address spaces. OSDI. 2006.
[54] S. Nagarakatte, J. Zhao, M. M. K. Martin, and S. Zdancewic. SoftBound:
highly compatible and complete spatial memory safety for C. PLDI. 2009.
[55] S. Nagarakatte, J. Zhao, M. M. K. Martin, and S. Zdancewic. CETS:
compiler enforced temporal safety for C. ISMM. 2010.
[28] I. Evans, S. Fingeret, J. Gonzalez, U. Otgonbaatar, T. Tang, H. Shrobe,
S. Sidiroglou-Douskos, M. Rinard, and H. Okhravi. Missing the point(er):
On the effectiveness of code pointer integrity. IEEE S&P, 2015.
[56] M. New, W. J. Bowman, and A. Ahmed. Fully abstract compilation
via universal embedding. To appear in 21st International Conference on
Functional Programming (ICFP 2016).
[29] I. Evans, F. Long, U. Otgonbaatar, H. Shrobe, M. C. Rinard, H. Okhravi,
and S. Sidiroglou-Douskos. Control jujutsu: On the weaknesses of finegrained control flow integrity. CCS. 2015.
[30] R. B. Findler and M. Felleisen. Contracts for higher-order functions.
ICFP. 2002.
[31] C. Fournet, N. Swamy, J. Chen, P.-É. Dagand, P.-Y. Strub, and
B. Livshits. Fully abstract compilation to JavaScript. POPL. 2013.
[57] M. Patrignani. The Tome of Secure Compilation: Fully Abstract
Compilation to Protected Modules Architectures. PhD thesis, KU Leuven,
Leuven, Belgium, 2015.
[58] M. Patrignani, P. Agten, R. Strackx, B. Jacobs, D. Clarke, and
F. Piessens. Secure compilation to protected module architectures.
TOPLAS, 2015.
[59] M. Patrignani and D. Clarke. Fully abstract trace semantics for protected
module architectures. Computer Languages, Systems & Structures, 42:22–
45, 2015.
[60] M. Patrignani, D. Clarke, and F. Piessens. Secure compilation of objectoriented components to protected module architectures. APLAS. 2013.
[61] M. Patrignani, D. Devriese, and F. Piessens. On modular and fullyabstract compilation. CSF, 2016.
[62] R. Prakash. Using Silicon Secured Memory to detect memory access
errors at hardware speeds. Oracle Community blog, 2016.
[63] N. Provos, M. Friedl, and P. Honeyman. Preventing privilege escalation.
In 12th USENIX Security Symposium. 2003.
[64] C. Reis and S. D. Gribble. Isolating web programs in modern browser
architectures. EuroSys. 2009.
[65] T. Sewell, S. Winwood, P. Gammie, T. C. Murray, J. Andronick, and
G. Klein. seL4 enforces integrity. ITP. 2011.
[32] E. Göktaş, E. Athanasopoulos, H. Bos, and G. Portokalidis. Out of
control: Overcoming control-flow integrity. IEEE S&P, 2014.
[33] K. Gudka, R. N. M. Watson, J. Anderson, D. Chisnall, B. Davis,
B. Laurie, I. Marinos, P. G. Neumann, and A. Richardson. Clean
application compartmentalization with SOAAP. CCS. 2015.
[34] Intel software guard extensions (Intel SGX) programming reference,
2014.
[35] R. Jagadeesan, C. Pitcher, J. Rathke, and J. Riely. Local memory via
layout randomization. CSF. 2011.
[36] A. Jeffrey and J. Rathke. A fully abstract may testing semantics for
concurrent objects. Theor. Comput. Sci., 338(1-3):17–63, 2005.
[37] A. Jeffrey and J. Rathke. Java Jr: Fully abstract trace semantics for a
core Java language. ESOP. 2005.
[38] L. Jia, S. Sen, D. Garg, and A. Datta. A logic of programs with interfaceconfined code. CSF. 2015.
[39] Y. Juglaret, C. Hritcu, A. A. de Amorim, B. C. Pierce, A. SpectorZabusky, and A. Tolmach. Towards a fully abstract compiler using micropolicies: Secure compilation for mutually distrustful components. CoRR,
abs/1510.00697, 2015.
[40] D. Kilpatrick. Privman: A library for partitioning applications. USENIX
FREENIX. 2003.
[66] H. Shacham. The geometry of innocent flesh on the bone: return-intolibc without function calls (on the x86). ACM CCS. 2007.
[67] K. Z. Snow, F. Monrose, L. Davi, A. Dmitrienko, C. Liebchen, and A.R. Sadeghi. Just-in-time code reuse: On the effectiveness of fine-grained
address space layout randomization. IEEE S&P. 2013.
[68] C. Tice, T. Roeder, P. Collingbourne, S. Checkoway, Ú. Erlingsson,
L. Lozano, and G. Pike. Enforcing forward-edge control-flow integrity
in GCC & LLVM. USENIX Security. 2014.
[69] R. Wahbe, S. Lucco, T. E. Anderson, and S. L. Graham. Efficient
software-based fault isolation. SOSP, 1993.
[70] R. N. M. Watson, J. Woodruff, P. G. Neumann, S. W. Moore, J. Anderson, D. Chisnall, N. H. Dave, B. Davis, K. Gudka, B. Laurie, S. J.
Murdoch, R. Norton, M. Roe, S. Son, and M. Vadera. CHERI: A hybrid
capability-system architecture for scalable software compartmentalization. IEEE S&P, 2015.
[41] G. Klein, K. Elphinstone, G. Heiser, J. Andronick, D. Cock, P. Derrin,
D. Elkaduwe, K. Engelhardt, R. Kolanski, M. Norrish, T. Sewell, H. Tuch,
and S. Winwood. seL4: Formal verification of an OS kernel. SOSP. 2009.
[42] J. Kroll, G. Stewart, and A. Appel. Portable software fault isolation.
CSF. 2014.
[43] R. Kumar, M. O. Myreen, M. Norrish, and S. Owens. CakeML: a
verified implementation of ML. POPL. 2014.
[44] A. Kwon, U. Dhawan, J. M. Smith, T. F. Knight, Jr., and A. DeHon.
16
[71] B. Yee, D. Sehr, G. Dardyk, J. B. Chen, R. Muth, T. Ormandy,
S. Okasaka, N. Narula, and N. Fullagar. Native Client: a sandbox for
portable, untrusted x86 native code. CACM, 53(1):91–99, 2010.
[72] B. Zeng, G. Tan, and Ú. Erlingsson. Strato: A retargetable framework
for low-level inlined-reference monitors. USENIX Security. 2013.
[73] L. Zhao, G. Li, B. D. Sutter, and J. Regehr. ARMor: fully verified
software fault isolation. EMSOFT. 2011.
[74] M. Zitser, R. Lippmann, and T. Leek. Testing static analysis tools using
exploitable buffer overflows from open source code. FSE. 2004.
17
| 6cs.PL
|
BORDER BASES FOR POLYNOMIAL RINGS OVER
NOETHERIAN RINGS
arXiv:1405.0472v5 [cs.SC] 2 Feb 2017
AMBEDKAR DUKKIPATI, NITHISH PAI AND MARIA FRANCIS
Abstract. The theory of border bases for zero-dimensional ideals has attracted
several researchers in symbolic computation due to their numerical stability and
mathematical elegance. As shown in (Francis & Dukkipati, J. Symb. Comp., 2014),
one can extend the concept of border bases over Noetherian rings whenever the corresponding residue class ring is finitely generated and free. In this paper we address
the following problem: Can the concept of border basis over Noetherian rings exists for ideals when the corresponding residue class rings are finitely generated but
need not necessarily be free modules? We present a border division algorithm and
prove the termination of the algorithm for a special class of border bases. We show
the existence of such border bases over Noetherian rings and present some characterizations in this regard. We also show that certain reduced Gröbner bases over
Noetherian rings are contained in this class of border bases.
1. Introduction
Gröbner bases theory for polynomial rings over fields gives an algorithmic technique
to determine a vector space basis of the residue class ring modulo a zero dimensional
ideal (Buchberger, 1965). The theory of Gröbner bases was extended to polynomial
rings over Noetherian commutative rings with unity a few decades ago (e.g. Trinks,
1978; Möller, 1988; Zacharias, 1978). For a good exposition on Gröbner bases over
rings one can refer to (Adams & Loustaunau, 1994).
Certain recent developments in cryptography and other fields have led to renewed
interest in polynomial rings over rings (Greuel et al., 2011). For instance, free residue
class rings over Z[x] called ideal lattices (Lyubashevsky & Micciancio, 2006) have been
shown to be isomorphic to integer lattices, an important cryptographic primitive (Ajtai, 1996) and certain cyclic lattices in Z[x] have been used in NTRU cryptographic
schemes (Hoffstein et al., 1998). Boolean polynomial rings over a boolean ring is
another important example of a polynomial ring over a ring since it can be used to
solve combinatorial puzzles like Sudoku (Sato et al., 2011). Another example is the
polynomial rings over Z/2k . They are used to prove the correctness of data paths in
system-on-chip design (Greuel et al., 2011).
Border bases, an alternative to Gröbner bases, are well studied for polynomial
rings over fields (Kehrein et al., 2005; Kehrein & Kreuzer, 2005). Though border
bases are restricted to zero dimensional ideals, the motivation for border bases comes
from the numerical stability of border bases over Gröbner bases (Stetter, 2004).
There has been considerable interest in the theory of border bases, from characterization (Kehrein & Kreuzer, 2005) to methods of computation (Kehrein & Kreuzer,
1
2
AMBEDKAR DUKKIPATI, NITHISH PAI AND MARIA FRANCIS
2006) to computational hardness (Ananth & Dukkipati, 2012). The concept of border bases can be easily extended to polynomial rings over rings if the corresponding
residue class ring has a free A-module representation w.r.t. some monomial order
and is finitely generated as an A-module (Francis & Dukkipati, 2014). In this paper,
we study border bases for ideals in polynomial rings over Noetherian commutative
rings in a more general set up.
2. Background & Preliminaries
2.1. Notations. A polynomial ring in indeterminates x1 , . . . , xn over a Noetherian,
commutative ring A is denoted by A[x1 , . . . , xn ]. Throughout this paper, the rings
we study are rings with unity. The set of all ideals of A is denoted by Id(A). When
A = k, where k is a field, we write it as k[x1 , . . . , xn ]. A monomial in indeterminates
x1 , . . . , xn is denoted by xα , where α ∈ Nn , and the set of all monomials is denoted
by Mn . By ‘term’ we mean cxα , where c ∈ A and c 6= 0. We will denote all the
terms in A[x1 , . . . , xn ] by Tn . Let T ⊆ Tn be a set of terms, possibly infinite. We
define the monomial part of T , denoted by Mon(T ), as Mon(T ) = {xα ∈ Mn : axα ∈
T, for some nonzero a ∈ A}.
P
A polynomial f is denoted by f = α∈Λ aα xα , where aα ∈ A, α ∈ Nn and Λ ⊆ Nn
is a finite set. Λ is called the support of the polynomial f , denoted by supp(f ). The
set of monomials appearing with nonzero coefficients in f is denoted by Mon(f ).
α
The set of all terms appearing in f is denoted by Ter(f ),
S i.e. Ter(f ) = {aα x |
α ∈ Λ}. If F is a set of polynomials then Mon(F ) = f ∈F Mon(f ). Similarly,
S
Ter(F ) = f ∈F Ter(f ). Given a set of terms T and an ideal a in A[x1 , . . . , xn ], the
set of residue class elements of T modulo a is denoted by T + a. That is T + a =
{axα + a : axαP
∈ T }. Given a set of polynomials F the span of F over A is given by
SpanA (F ) = { si=1 ai fi : ai ∈ A, fi ∈ F, s ∈ N}.
With respect to a monomial order ≺, the leading monomial, leading coefficient
and leading term of a polynomial f ∈ A[x1 , . . . , xn ] are denoted by Lm(f ), Lc(f ) and
Lt(f ) respectively. That is, we have Lt(f ) = Lc(f ) Lm(f ). Given a set of polynomials
S, possibly infinite, Lt(S) denotes the set of all leading terms of polynomials in S.
The leading term ideal (or initial ideal) of S is denoted by hLt(S)i and is given by
hLt(S)i = h{Lt≺ (f ) | f ∈ S}i. Similarly, the leading monomial ideal and the leading
coefficient ideal of S are denoted by hLm(S)i and hLc(S)i respectively. Note that the
leading coefficient ideal is an ideal in the coefficient ring, A. The ideal generated by
the polynomials f1 , . . . , fs is denoted by hf1 , . . . , fs i. Also a set of polynomials S is
said to be monic w.r.t. a monomial order if the leading coefficient of each polynomial
in S is 1.
2.2. Border bases over a field. Here we briefly recall definitions related to border
bases.
Definition 2.1. A finite set of monomials O ⊆ Mn is said to be an order ideal if it is
closed under forming divisors i.e., for xα ∈ Mn , if xβ ∈ O and xα |xβ , then xα ∈ O.
BORDER BASES FOR POLYNOMIAL RINGS OVER NOETHERIAN RINGS
3
Definition 2.2. Let O be an order ideal. The border of O is the set ∂O = (x1 O ∪
. . . ∪ xn O) \ O. The first border closure of O is defined as the set O ∪ ∂O and it is
denoted by ∂O.
Note that ∂O is also an order ideal. By convention, if O = ∅, then we set ∂O = {1}.
Definition 2.3. Let O = {xα1 , . . . , xαs } be an order ideal, and let ∂O = {xβ1 , . . . , xβt }
be its border. A set of polynomials B = {b1 , . . . , bt } ⊆ k[x1 , . . . , xn ] is called an Oborder prebasis if the polynomials have the form
s
X
βj
bj = x −
cij xαi ,
i=1
where cij ∈ k for 1 ≤ i ≤ s and 1 ≤ j ≤ t.
Definition 2.4. Let O = {xα1 , . . . , xαs } be an order ideal and B = {b1 , . . . , bt } be
an O-border prebasis consisting of polynomials in a ⊆ k[x1 , . . . , xn ]. We say that the
set B is an O-border basis of a if the residue classes of xα1 , . . . , xαs form a k-vector
space basis of k[x1 , . . . , xn ]/a.
The existence and uniqueness of border bases are established in (Kehrein & Kreuzer,
2005). Given an ideal one can show that there exists a border bases that do not correspond to Gröbner bases for any term ordering. An example of such a border basis
is given in (Kehrein & Kreuzer, 2006).
2.3. The A-module A[x1 , . . . , xn ]/a. From now on, unless otherwise specified we
deal with polynomials over a Noetherian, commutative ring A. Given an ideal a in
A[x1 , . . . , xn ], using Gröbner basis methods one can give an A-module representation
of residue class ring A[x1 , . . . , xn ]/a if it is finitely generated (Francis & Dukkipati,
2014). We describe this briefly below and for more details one can refer to (Francis
& Dukkipati, 2014). The notation is mostly borrowed from (Adams & Loustaunau,
1994).
Let G = {gi : i = 1, . . . , t} be a Gröbner basis for a. For each monomial, xα , let
Jxα = {i : lm(gi ) | xα , gi ∈ G} and IJxα = h{lc(gi ) : i ∈ Jxα }i. Note that IJxα is
an ideal in A. We refer to IJxα as the leading coefficient ideal w.r.t. G. Let CJxα
represent a set of coset representatives of the equivalence classes in A/IJxα . Given a
m
P
polynomial, f ∈ A[x1 , . . . , xn ], let f =
ai xαi mod hGi, where ai ∈ A, i = 1, . . . , m.
i=1
If A[x1 , . . . , xn ]/hGi is an A-module generated by m elements, then corresponding to
the coset representatives, CJxα1 , . . . , CJxαm , there exists an A-module isomorphism,
φ : A[x1 , . . . , xn ]/hGi −→ A/IJxα1 × · · · × A/IJxαm
m
X
ai xαi + hGi 7−→ (c1 + IJxα1 , · · · , cm + IJxαm ),
(1)
i=1
where ci = ai mod IJxαi and ci ∈ CJxαi . We refer to A/IJxα1 ×· · ·×A/IJxαm as the Amodule representation of A[x1 , . . . , xn ]/a w.r.t. G. If IJxαi = {0}, we have CJxαi = A,
for all i = 1, . . . , m. This implies A[x1 , . . . , xn ]/a ∼
= Am , i.e. A[x1 , . . . , xn ]/a has an
4
AMBEDKAR DUKKIPATI, NITHISH PAI AND MARIA FRANCIS
A-module basis and it is free. In this case, we define A[x1 , . . . , xn ]/a to have a “free
A-module representation w.r.t. G”. The necessary and sufficient condition for an
A-module A[x1 , . . . , xn ]/a to have a free A-module representation is given in (Francis
& Dukkipati, 2014). It makes use of the the concept of ‘short reduced Gröbner basis’
introduced in (Francis & Dukkipati, 2014) which we briefly describe below.
Definition 2.5. Let a be an ideal in A[x1 , . . . , xn ] and let G be its reduced Gröbner
basis as described in (Pauer, 2007). Consider the isomorphism in (1). G is called a
short reduced Gröbner basis if the size of the generating set of the leading coefficient
ideal, IJxα , of each leading monomial, xα in G, is minimal.
One can define reduced Gröbner bases over rings exactly as that of fields but
it may not exist in all the cases. The definition of reduced Gröbner basis given
by (Pauer, 2007) is a generalization of the concept over fields to rings that also
ensures the existence of such a basis for every ideal in the polynomial ring. The
short reduced Gröbner basis of an ideal is not to be confused with strong Gröbner
basis (Adams & Loustaunau, 1994, Definition 4.5.6.). Strong Gröbner basis exists
only if the coefficient ring is a PID. In a PID, strong Gröbner basis coincides with
the short reduced Gröbner basis.
Proposition 2.6. (Francis & Dukkipati, 2014, Proposition 3.12) Let a ⊆ A[x1 , . . . , xn ]
be a nonzero ideal such that A[x1 , . . . , xn ]/a is finitely generated. Let G be a short reduced Gröbner basis for a w.r.t. some monomial ordering, ≺. Then, A[x1 , . . . , xn ]/a
has a free A-module representation w.r.t. G if and only if G is monic.
Example 2.7. Let G = {3x, 5x, y} be the Gröbner basis of an ideal a in Z[x, y].
The short reduced Gröbner basis of a is given by Gred = {x, y}. It is monic and a
Z-module basis of Z[x, y]/a is given by {1 + a}.
With the above result the concept of border bases can be extended to ideals in
polynomial rings over rings, in the cases where the corresponding residue class rings
are finitely generated and have a free A-module representation w.r.t. some monomial
order (Francis & Dukkipati, 2014, Section 6). One can show that all the characterizations in (Kehrein & Kreuzer, 2005) hold true when the residue class ring is free. For
the sake of completeness, we state the definition of border bases in this case below.
Definition 2.8. Let O = {xα1 , . . . , xαs } be an order ideal and B = {b1 , . . . , bt } ⊆
A[x1 , . . . , xn ] be an O-border prebasis. Let a ⊆ A[x1 , . . . , xn ] be an ideal such that
A[x1 , . . . , xn ]/a is finitely generated and is a free A-module. Then B is said to be an
O-border basis if B ⊆ a and O forms an A-module basis for A[x1 , . . . , xn ]/a.
3. Order Functions and Border Prebasis Division Algorithm
The following notion that we introduce in this paper is crucial to the theory of border
bases that we develop here.
Definition 3.1 (Order Function). Let Id(A) be the set of all ideals in the ring,
A. A mapping I : Mn → Id(A) is said to be an order function if xβ | xα implies
I(xβ ) ⊆ I(xα ), for all xα , xβ ∈ Mn . From now on we denote I(xα ) by Ixα .
BORDER BASES FOR POLYNOMIAL RINGS OVER NOETHERIAN RINGS
5
Clearly, the Gröbner basis of an ideal ‘fixes’ an order function. Consider the
leading coefficient ideal, IJxα that we constructed in Section 2.3 w.r.t. G. Since Jxα
is a saturated set, the mapping xα 7→ IJxα is an order function, which is denoted by
I (G) .
Definition 3.2. An order function I : Mn → Id(A) is said to be proper if it maps
only finitely many monomials to proper ideals in A.
Example 3.3. Consider the mapping I : N2 −→ Id(Z). Let I(1) = {0}, I(x1 ) = h3i
and every other monomial be mapped to h1i. We see that for any xα ∈ N2 such that
x1 | xα , h3i ⊆ I(xα ). Therefore, I is a proper order function.
We define order ideal with respect to a proper order function I.
Definition 3.4. For each xα ∈ Mn and a proper order function I, fix Cxα ⊆ A, a
set of coset representatives of A/Ixα . A set of terms OI ⊆ Tn is said to be an order
ideal w.r.t. I if for all xα ∈ Mn , cxα ∈ OI if and only if c ∈ Cxα .
Note that for each monomial xα , one can choose any set of coset representatives
of A/Ixα and with each choice we have a different order ideal.
Example 3.5. Consider the polynomial ring Z[x, y] and let the mapping I be such
that I1 = {0}, Ix = h4i, Iy = h3i, Ix2 = h2i and the rest of the monomials map to
h1i. I is clearly a proper order function. Let the set of coset representatives be the
following, C1 = Z, Cx = {0, 1, 2, 3}, Cy = {0, 1, 2}, Cx2 = {0, 1} and for all the other
monomials, xα , Cxα = {0}. Then, OI = {a1 , a2 x, a3 y, a4 x2 | a1 ∈ C1 , a2 ∈ Cx , a3 ∈
Cy , a4 ∈ Cx2 } is an order ideal corresponding to I.
Example 3.6. Now consider a polynomial ring
k[u1 , u2 ][x, y, z] and let I be an order function defined by I1 = {0}, Ix = {0},
Iy = hu21 i, Iz = hu22 −3u1 i, Ix2 = h0i, Ixy = hu21 , u22 −1i, Ixz = hu1 , u2 i and rest of the
monomials mapping to h1i. I is a proper order function. Let C1 , Cx , Cy , Cz , Cx2 , Cxy , Cxz
represent the nonzero set of coset representatives. Then O0 I = {a1 , a2 x, a3 y, a4 z, a5 x2 ,
a6 xy, a7 xz | a1 ∈ C1 , a2 ∈ Cx , a3 ∈ Cy , a4 ∈ Cz , a5 ∈ Cx2 , a6 ∈ Cxy , a7 ∈ Cxz } is
an order ideal corresponding to I.
In sequel, we write order ideal OI as O and its dependence on the order function
is implicitly assumed. It is important to note that unlike in the case of fields, the
order ideal in the case of polynomial rings over rings have both monic and nonmonic
monomials.
Given an order ideal O we introduce two types of borders: a monomial border
∂Om and a scalar border ∂Os .
Definition 3.7. Given an order ideal O the monomial border of O is defined as
∂Om = {x1 · Mon(O) ∪ . . . ∪ xn · Mon(O)} \ Mon(O).
Definition 3.8. Let O be an order ideal with respect to a proper order function I. For
each xα such that Ixα 6= h1i define ∂Oxα = {c1 xα , . . . , cs xα }, where Ixα = hc1 , . . . , cs i
6
AMBEDKAR DUKKIPATI, NITHISH PAI AND MARIA FRANCIS
for some c1 , . . . , cs ∈ A. The scalar border of an order ideal is defined as
[
∂Os =
∂Oxα .
xα ∈Mn
Ixα 6=h1i
Definition 3.9. The border of the order ideal O, denoted as ∂O, is defined as ∂O =
∂Om ∪ ∂Os .
Example 3.10. Consider Example 3.5. The set of monic border terms that form
the monomial border is ∂Om = {xy, y 2 , x3 , x2 y} and the scalar border is ∂Os =
{4x, 3y, 2x2 }. Hence, the border of the order ideal is ∂O = {xy, y 2 , x3 , x2 y, 4x, 3y, 2x2 }.
Example 3.11. Consider Example 3.6. The monomials that form the monomial
border is ∂O0 m = {x3 , x2 y, x2 z, xy 2 , xyz, xz 2 , y 2 , yz, z 2 } and the scalar border is the
set ∂O0 s = {u21 y, (u22 − 3u1 )z, u21 xy, (u22 − 1)xy, u1 xz, u2 xz}. Hence border of the
order ideal is ∂O0 = {x3 , x2 y, x2 z, xy 2 , xyz, xz 2 , y 2 , yz, z 2 , u21 y, (u22 −3u1 )z, u21 xy, (u22 −
1)xy, u1 xz, u2 xz}.
We define ∂ 0 O = O and 0th border closure as ∂ 0 O = {axα | cxα ∈ O, for some c ∈
Cxα , c 6= 0 and a ∈ c+Ixα }. Note that in the case of fields these quantities are defined
as ∂ 0 O = ∂ 0 O = O (Kehrein & Kreuzer, 2006).
The definitions of first and higher order border closures are given below.
Definition 3.12. The first border closure ∂O of an order ideal O is defined as
∂O ={axα | ∃c ∈ A such that cxα ∈ O ∪ ∂Os ,
a ∈ A or xα ∈ ∂Om }.
Proposition 3.13. The first border closure, ∂O, of an order ideal, O is an order
ideal.
Proof. We fix Ixα = {0} for all xα ∈ ∂O. By Definition 3.12 three cases arises:
xα ∈ ∂Om or there exists c ∈ A such that cxα ∈ O or cxα ∈ ∂Os . Let xα ∈ ∂Om .
Suppose xβ |xα for some xβ ∈ Mn , then clearly, xβ ∈ Mon(O)∪∂Om . If xβ ∈ Mon(O),
then there exists some d ∈ Cxβ such that dxβ ∈ O. Therefore, in either case xβ ∈ ∂O.
In the second case, let cxα ∈ O. Suppose xβ |xα for some xβ ∈ Mn . By the closure
property of O, dxβ ∈ O for some d ∈ Cxβ . Therefore, xβ ∈ ∂O. In the third case,
let cxα ∈ ∂Os . Suppose xβ |xα for some xβ ∈ Mn . This implies that xβ ∈ Mon(O).
Thus, xβ ∈ ∂O.
The monomial part of the first border closure defined as the set of monomials in
∂O, is a finite set and it is represented as Mon(∂O). It is interesting to see that since
Ixα = {0} for all xα ∈ Mon(∂O), the scalar border for k ≥ 2 is an empty set and one
needs to consider only the monomial border.
Definition 3.14. The k th border of an order ideal O for k
1 is defined as
∂ k O = {x1 · Mon(∂ k−1 O) ∪ . . . ∪ xn · Mon(∂ k−1 O)} \ Mon(∂ k−1 O),
where Mon(∂ k−1 O) is the monomial part of the (k − 1)th border closure.
BORDER BASES FOR POLYNOMIAL RINGS OVER NOETHERIAN RINGS
7
Definition 3.15. For k ≥ 2, the k th border closure of an order ideal is defined as
∂ k O = {axα | a ∈ A, xα ∈ ∂ k O ∪ ∂ k−1 Om }.
Example 3.16. Consider Example 3.5. The set of monic border terms that form
the monomial border is ∂Om = {xy, y 2 , x3 , x2 y} and the scalar border is ∂Os =
{4x, 3y, 2x2 }. The second border of the order ideal, O, is the set, ∂ 2 O = {xy 2 , y 3 ,
x4 , x3 y, x2 y 2 }.
Example 3.17. Consider Example 3.6. The second border of the order ideal, O0 is
the set, ∂ 2 O0 = {y 3 , y 2 z, yz 2 , z 3 , x4 , x3 y, x3 z, x2 y 2 , x2 yz, x2 z 2 , xy 3 , xyz 2 , xy 2 z, xz 3 }.
Remark 3.18. The k th border closure is an infinite set of terms for k ≥ 0. Further,
for k ≥ 1, ∂ k O is closed under division and hence the set of monomials corresponding
to it, Mon(∂ k O), mimics the case of fields.
The following example explains the borders and border closures of an order ideal.
Example 3.19. Let the order function I : M2 → Id(Z) be defined as follows: I1 = 0,
Ix = 0, Iy = h4i, Ixy = h2i and for other monomials I is mapped to h1i. The order
ideal corresponding to I, OI = {a, bx, cy, dxy : a, b ∈ Z, c ∈ {0, 1, 2, 3}, d ∈ {0, 1}}.
The 0th border closure is given by {a, bx, cy, dxy : a, b ∈ Z\{0}, c ∈ Z\h4i, d ∈ Z\h2i}.
Figure 1 shows a few terms from the 0th border closure.
Figure 1. 0th border closure
The first border is hx2 , 2xy, y 2 , x2 y, xy 2 , 4yi. The first border closure is given by
hax2 , bxy, cy 2 , dx2 y, exy 2 , f y : a, c, d, e ∈ Z, b ∈ h2i, f ∈ h4i. Figure 2 is an illustration of the same.
For k ≥ 2 the k th borders are exactly as that of fields. Figure 3 shows the borders for
k = 2, 3. Note that the figure depicts the borders and not the border closures.
We give below certain properties of order ideals, their borders and border closures.
These properties are analogous to the case of polynomial rings over fields.
Proposition 3.20. Let O be an order ideal and ∂Om be its monomial border. Then
(1) For k ≥ 1, the k th monomial border closure of O, Mon(∂ k O) is the following
disjoint union, Mon(O) ∪ ∂Om ∪ (∪ki=2 ∂ k O).
8
AMBEDKAR DUKKIPATI, NITHISH PAI AND MARIA FRANCIS
Figure 2. First border closure
Figure 3. 2nd , 3rd border closure
(2) For k ≥ 1, ∂ k Om = Mnk · Mon(O) \ Mn<k . Mon(O), where ∂ k Om = ∂ k O for
k ≥ 2.
(3) A monomial, xα ∈ Mn is divisible by xβ ∈ ∂Om if and only if xα ∈ Mn \
Mon(O).
Proof. (1) The proof is by induction on k. For k = 1, clearly the monomials in
the first border closure are elements of the set Om ∪ ∂Om . From the definition of
monomial border of O we have that Om and ∂Om are disjoint. Suppose that the claim
is true for the k th monomial border closure. For k + 1, ∂ k+1 Om = ∂ k Om ∪ ∂ k+1 O. It
is easy to verify that the sets ∂ k Om and ∂ k+1 O are disjoint.
(2) The claim follows from the observation that ∂ k O = ∂ k Om \ ∂ k−1 Om .
(3) We have xβ ∈ ∂Om . This implies that there exists xγ ∈ Om and an indeterminate
xi0 such that xβ = xi0 xγ . We have xi0 xγ |xα . If xα ∈ Om then xi0 xγ ∈ Om which is a
contradiction. Now consider a monomial xα ∈ Tn \ O. Then, xα ∈ ∂Om or xα ∈ ∂ k O
for some k ≥ 2. If xα ∈ ∂ k O then it implies that there exists a monomial xγ of degree
k − 1 and a xβ ∈ ∂Om such that xα = xγ xβ . The claim follows.
Now we introduce some concepts that are essential for the division algorithm.
Definition 3.21. The index of a term cxα w.r.t. an order ideal, O is defined as
indO (cxα ) = min{k ∈ N | cxα ∈ ∂ k O}.
BORDER BASES FOR POLYNOMIAL RINGS OVER NOETHERIAN RINGS
9
Definition 3.22. Let f ∈ A[x1 , . . . , xn ] be any nonzero polynomial with support,
supp(f ), then the index of f w.r.t. an order ideal, O is defined as
indO (f ) =
max indO (cα xα ).
α∈supp(f )
Example 3.23. Consider Example 3.5. The set of monic border terms that form
the monomial border is ∂Om = {xy, y 2 , x3 , x2 y} and the scalar border is ∂Os =
{4x, 3y, 2x2 }. Then indO (3x) = 0, indO (xy) = 1 and indO (xy 2 + 8x + 7y) = 2.
Example 3.24. Consider Example 3.6. The monomials that form the monomial
border is ∂O0 m = {x3 , x2 y, x2 z, xy 2 , xyz, xz 2 , y 2 , yz, z 2 } and the scalar border is the
set ∂O0 s = {u21 y, (u22 − 3u1 )z, u21 xy, (u22 − 1)xy, u1 xz, u2 xz}. Then
indO0 ((3u2 + u2 )x) = 0, indO0 (u1 x2 y) = indO0 ((u21 + u2 )xz) = 1 and indO0 (xy 3 + 8x) =
2.
For any polynomial, the terms of highest index are grouped together to form a
border form analogous to the leading term in Gröbner bases theory. We define this
below.
Definition 3.25. Let f ∈ A[x1 , . . . , xn ] be a nonzero polynomial such that the
indO (f ) = i0 . The border form of f w.r.t. O is defined as
X
BFO (f ) =
cα x α ,
α∈supp(f ),cα ∈A
indO (cα xα )=i0
a polynomial in A[x1 , . . . , xn ].
Note that unlike leading term of a polynomial in Gröbner bases theory that is
always a monomial, border form can be a polynomial. The concept of leading term
ideal has an analogous form in border bases theory called the border form ideal.
Definition 3.26. The border form ideal of an ideal a in A[x1 , . . . , xn ] w.r.t. an order
ideal O is defined as
BFO (a) = hBFO (f ) | f ∈ ai.
Example 3.27. Consider Example 3.5. Let f = xy 2 + 2x2 y 2 + xy + 3x + 2. Then
the index of f w.r.t. the order ideal O is equal to 2. The border form of f is the
polynomial, BFO (f ) = xy 2 + 2x2 y 2 .
Example 3.28. Consider Example 3.6. Let f = 2xy 2 + (u1 − u2 )xz + 3x + 2. Then
the index of f w.r.t. the order ideal O0 is equal to 2. The border form of f is the
polynomial, BFO0 (f ) = 2xy 2 + (u1 − u2 )xz.
We now give the definition of border prebasis for an order ideal, O.
Definition 3.29. Let O be an order ideal, and ∂O = {c1 xα1 , . . . , cs xαs } be its border.
Let Cxαi be the set of coset representatives of A/Ixαi . A finite set of polynomials G
= {g1 , . . . , gs } ⊆ A[x1 , . . . , xn ] is said to be an O-border prebasis if gi = ci xαi − hi ,
where hi ∈ A[x1 , . . . , xn ] satisfying Ter(hi ) ⊆ O \ {axαi | a ∈ Cxαi }, i = 1, . . . , s.
10
AMBEDKAR DUKKIPATI, NITHISH PAI AND MARIA FRANCIS
Example 3.30. We consider Example 3.5. The set G = {g1 , . . . , g7 }, where g1 =
xy − x, g2 = y 2 − y, g3 = x3 − 2y, g4 = x2 y − x2 + 10, g5 = 4x − 2y, g6 = 3y − 3x and
g7 = 2x2 − x + 5 is an O-border prebasis but it is not acyclic. Let G0 = {g10 , . . . , g70 }
where g10 = xy − x, g20 = y 2 − y, g30 = x3 − x2 + 6, g40 = x2 y − y + 5, g50 = 4x − 7,
g60 = 3y − x and g70 = 2x2 − 2y − 3x which is also a O-border prebasis but it is acyclic
since the permutation of G0 , {g10 , g20 , g30 , g40 , g70 , g60 , g50 } satisfies the acyclicity condition.
Example 3.31. Consider Example 3.6. The set G =
{g1 , . . . , g15 }, where g1 = x3 − 3, g2 = x2 y − 3u1 y, g3 = x2 z − 2z, g4 = xy 2 − x + 10,
g5 = xyz − 11xy, g6 = xz 2 − u2 u21 x2 , g7 = y 2 − x + u1 u2 , g8 = yz − 3y + 2,
g9 = z 2 + 5xz + 11u1 x, g10 = u21 y + u2 x + 3, g11 = (u22 − 3u1 )z − u22 y, g12 =
u21 xy + 3u1 x − 2z, g13 = (u22 − 1)xy + 2x2 , g14 = u1 xz + 3u1 x2 , g15 = u2 xz +
2u1 xy + 4x2 − 4z − 10u1 y + 14 is an acyclic O-border prebasis since the following
permutation of G, {g1 , g2 , g3 , g4 , g5 , g6 , g7 , g8 , g9 , g14 , g15 , g13 , g12 , g11 , g10 } satisfies the
acyclicity condition.
Note that unlike in fields, for a monomial in the border of O, we can have more
than one polynomial in the O-border prebasis but only one polynomial corresponding
to a term in the border. With the definition of O-border prebasis, we now give a
procedure for division of any polynomial in A[x1 , . . . , xn ] with the O-border prebasis.
Algorithm 3.32. Let O be an order ideal. Let Mon(O) = {xα1 , . . . , xαt } be its
monomial part. Let ∂Om = {xβ1 , . . . , xβs0 } and ∂Os = {cs0 +1 xβs0 +1 , . . . , cs xβs } be its
monomial border and scalar border respectively. Let G = {g1 , . . . , gs } ⊆ A[x1 , . . . , xn ]
be an O-border prebasis. For f ∈ A[x1 , . . . , xn ] we perform the following steps.
(1) Initialize f1 = . . . = fs = 0, a1 = . . . = at = 0 and h = f .
(2) If h = 0 return (f1 , . . . , fs , a1 , . . . , at ).
(3) If indO (h) = 0 then find b1 , . . . , bt ∈ A such that h = b1 xα1 + . . . + bt xαt . Set
ai = bi for each 1 ≤ i ≤ t. Return (f1 , . . . , fs , a1 , . . . , at ).
(4) If indO (h) = 1 and h contains a term dxβ such that xβ ∈ ∂Om then goto
Step 5. Else, let h = d1 xγ1 + . . . + dj xγj such that indO (h) = indO (d1 xγ1 ) and
indO (d1 xγ1 ) ≥ . . . ≥ indO (dj xγj ). Find bµ+1 , . . . , bs ∈ A such that d1 xγ1 =
bs0 +1 (cs0 +1 xβs0 +1 ) + . . . + bs (cs xβs ). Subtract bs0 +1 gs0 +1 + . . . + bs gs from h, add
bi to fi for s0 + 1 ≤ i ≤ s and return to Step 2.
(5) Else, let h = d1 xγ1 +. . .+dj xγj such that indO (h) = indO (xγ1 ) and indO (d1 xγ1 ) ≥
. . . ≥ indO (dj xγj ). Determine xβi ∈ ∂Om with the smallest i such that
xγi = xµ xβi and deg(xµ ) = indO (h) − 1. Subtract d1 xµ gi from h, add d1 xµ to
fi and return to Step 2.
This procedure over rings differs from the case of fields only in Step 4. The termination of the above method is not assured because of the possibility that for a given
polynomial, f , a monomial in its support identified with index 0 in Step 3 may again
have an index 1 after Step 4. Therefore, we cannot assume the reduction in index
values at every step of the procedure.
BORDER BASES FOR POLYNOMIAL RINGS OVER NOETHERIAN RINGS
11
4. Acyclic Border Prebases and Termination of Border Division
Algorithm
Here, we identify a special class of O-border prebases called acyclic O-border prebases
for which the termination of the border division algorithm can be established.
Definition 4.1. A O-border prebasis G = {g1 , . . . , gs } is said to be acyclic if there
k,
exists a permutation of G, {gi1 , . . . , gis } such that for any gij , gik , where j
exactly one of the following conditions are satisfied
(1) cj BFO (gij ) = ck BFO (gik ) for any cj , ck ∈ A or
(2) dj xαj ∈ ∂O and dj xαj ∈ supp(gij ) implies ck xαj ∈
/ supp(gik ) for some ck , dj ∈
A.
The ordered set of acyclic O-border prebasis that satisfies the permutation given
above is called a ‘well ordered’ acyclic O-border prebasis. We now show the correctness and termination of Algorithm 3.32 when the O-border prebasis is acyclic.
Proposition 4.2. (Border Division Algorithm) Consider a polynomial f ∈
A[x1 , . . . , xn ]. If the O-border prebasis G = {g1 , . . . , gs } is acyclic, then Algorithm
3.32 terminates for f and returns a tuple,
(f1 , . . . , fs , a1 , . . . , at ) ∈ (A[x1 , . . . , xn ])s × At
such that
f = f1 g1 + . . . + fs gs + a1 xα1 + . . . + at xαt ,
and deg(fi ) ≤ indO (f ), for i = 1, . . . , s with fi gi 6= 0.
Proof. We first describe the execution of the algorithm. In Step 4, indO (d1 xγ1 ) = 1
and d1 xγ1 ∈ SpanA (h∂OiA ). This implies that d1 ∈ Ixγ1 , where Ixγ1 is an ideal
generated by hu1 , . . . , uk i, ui xγ1 ∈ ∂Os , 1 ≤ i ≤ k. Thus, there exists l1 , . . . , lk ∈ A
P
P
such that d1 = ki=1 li ui . Hence, d1 xγ1 = ss0 +1 bi (ci xβi ), where ci xβi ∈ ∂Os and
bi = lj when ci xβi = uj xγ1 for some j ∈ {1, . . . , k}, and bi = 0, otherwise. The other
steps, due to the absence of scalar border terms, mimics the border basis division in
fields (Kehrein & Kreuzer, 2005, Proposition 3). We prove that the representation,
f = f1 g1 + . . . + fs gs + a1 xα1 + . . . + at xαt + h,
computed by the algorithm is valid in every step. Clearly, it is satisfied in Step 1.
In Step 4 we subtract (bs0 +1 gs0 +1 + . . . + bs gs ) from h. These bi s are then added to
fi s, i.e. fi = fi + bi , s0 + 1 ≤ i ≤ s. Similarly in Step 5, from h we subtract d1 xµ gi
and we add d1 xµ to fi . The constants a1 , . . . , at are modified only in Step 3. The
representation of f is valid because indO (h) = 0. If the algorithm terminates, h = 0
and we have a valid representation.
Now we prove that deg(fi ) ≤ indO (f ) for all i = 1, . . . , s. In Step 5 of the
algorithm, where we divide using the monomial border, our choice of the term d1 xµ
is such that
deg(dxµ ) = indO (h) − 1. In Step 4, where we divide using the scalar border, the
index of the intermediate polynomial, h is 1. The bi , i = 1, . . . , s are constants
and the degree of fi , i = 1, . . . , s are therefore zero. All the other steps in the
12
AMBEDKAR DUKKIPATI, NITHISH PAI AND MARIA FRANCIS
algorithm do not affect fi , i = 1, . . . , s. Also, in the algorithm the index of the
intermediate polynomial, h never increases. From the above steps, the inequality
deg(fi ) ≤ indO (f ) − 1 for all i = 1, . . . , s follows.
Next, we prove that the algorithm terminates on all inputs. In Step 4, indO (h) = 1
and Ter(h) ⊆ SpanA (Mon(O))∪SpanA (∂Os ) = SpanA (Mon(O)). We claim that Step
4 terminates after a finite number of steps for an acyclic O-border prebasis. Let h =
d1 xα1 +. . .+dt xαt . For simplicity, let us assume that the acyclic O-border prebasis, G
is well ordered. It can easily be seen that g1 will be used atmost once in Step 4, while
g2
G\{g1 ,g2 }
g1
G\{g1 ,g2 }
g2
g2 will be used at most twice (h −
→ h1 −−−−−→+ h2 −
→ h3 −−−−−→+ h4 −
→ h5 ).
2
Similarly, any gi will be used atmost O i ) times. For the set G, therefore Step
4 is executed at most O s3 ) times. All the other steps of the division correspond
to either order ideal, monomial border or the k th order border, k 1 and therefore
mimic the border division in fields. Hence, the termination is guaranteed by (Kehrein
& Kreuzer, 2005, Proposition 3).
The border division algorithm gives us the remainder upon division by an acyclic
O-border prebasis as a part of its output. We now give a formal definition for
remainder.
Definition 4.3. Let O be an order ideal and Mon(O) = {xα1 , . . . , xαt }, its monomial part. Let G = {g1 , . . . , gs } be the O-border prebasis. The O-remainder of a
polynomial f w.r.t. G, if it exists, is given as
remO,G (f ) = a1 xα1 + . . . + at xαt ,
where f = f1 g1 + . . . + fs gs + a1 xα1 + . . . + at xαt and ai ∈ A for all i = 1, . . . , t is
a representation computed by the border division algorithm whenever the algorithm
terminates.
5. Order Span and Acyclic Border Bases
Consider the case when A[x1 , . . . , xn ]/a is finitely generated. Using the order function
we define a generating set for A[x1 , . . . , xn ]/a that also satisfies a weaker form of the
linear independence property.
Definition 5.1 (Order span). Let a be an ideal and let I be a proper order function
and B = {xα1 , . . . , xαm } be a finite set of monomials of size m such that xα ∈
/ B if
α
n
and only if Ixα = h1i, where x ∈ M . Let CIxα be the coset representatives of the
equivalence classes of A/Ixα . Then we say the set of residue classes of B forms an
order span for A[x1 , . . . , xn ]/a w.r.t. I if it satisfies the following properties.
(i) B + a generates A[x1 , . . . , xn ]/a as an A-module and
A[x1 , . . . , xn ]/a = {
m
X
ai xαi + a | ai ∈ CIxαi , xαi ∈ B}.
i=1
Pm
αi
(ii) If i=1 ai x + a = 0, where m ∈ N, xαi ∈ B and ai 6= 0 for all i = 1, . . . , m,
then for some j ∈ {1, . . . , m}, aj ∈ Ixαj .
BORDER BASES FOR POLYNOMIAL RINGS OVER NOETHERIAN RINGS
0
13
0
(iii) If there exists an order function I such that Ixα ⊆ Ixα for some xα ∈ Mn and
0
0
0
0
B = {xα : Ixα 6= h1i} satisfies (i) and (ii) w.r.t. I , then I = I.
P
Remark 5.2. The linear independence property requires that if ki=1 ai xαi + a = 0
then for all i ∈ {1, . . . , k}, ai ∈ Ixαi . Therefore, the second condition in Definition 5.1
is a weaker form of the linear independence property. In fact, in the case of fields
and residue class rings with a free A-module representation, the second condition
automatically implies the linear independence property.
Remark 5.3. The third condition in Definition 5.1 can be interpreted as a minimality
condition on the spanning set. Over fields, linear independence of the spanning set
ensures minimality but over rings it has to be specified separately. Consider the
0
ideal a = h4x1 , x1 2 , x2 i in Z[x1 , x2 ]. Let I and I be two order functions. Define
0
I1 = 0, Ix1 = h2i and for all the other monomials xα , Ixα = h1i. Similarly define I1 =
0
0
0
0, Ix1 = h4i and for all the other xα ∈ M2 , Ixα = h1i. Both I and I satisfy the first
0
two conditions of the order span. However, since Ix1 (= h4i) ( Ix1 (= h2i), it is w.r.t.
0
the second order function, I that we define the order span of Z[x1 , x2 ]/h4x1 , x1 2 , x2 i.
Corollary 5.4. Let a be an ideal such that A[x1 , . . . , xn ]/a is finitely generated. Let
G = {g1 , . . . , gt } be a Gröbner basis of an ideal a and I (G) the order function fixed
by G. Then,
(1) The order function I (G) is proper,
(2) The finite order span of A[x1 , . . . , xn ]/a w.r.t. I (G) is given by
B = {xα : lt(gi ) - axα , for some nonzero a ∈ A, gi ∈ G}
We now provide a better interpretation of the mapping described by (1) in terms
of order span.
Proposition 5.5. Let a be an ideal in A[x1 , . . . , xn ]. Let I be a proper order function
/ B,
and B = {xα1 , . . . , xαm } be a finite set of monomials of size m such that if xα ∈
α
n
Ixα = h1i, where x ∈ M . Consider a mapping,
φ : A[x1 , . . . , xn ]/a −→ A/Ixα1 × · · · × A/Ixαm
f + a 7−→ (c1 + Ixα1 , · · · , cm + Ixαm ),
where ci ∈ CIxαi for all i ∈ {1, . . . , m}. Then, φ is an isomorphism if B forms an
order span for A[x1 , . . . , xn ]/a.
Proof. Let B form an order span for A[x1 , . . . , xn ]/a. We first show that the mapping
φ is well defined. Consider a polynomial f ∈ A[x1 , . . . , xn ]. Suppose
φ(f + a) = (c1 + Ixα1 , · · · , cm + Ixαm ) and
φ(f + a) = (c01 + Ixα1 , · · · , c0m + Ixαm ),
where ci , c0i ∈ CIxαi , for all i = 1, . . . , m. This implies (ci − c0i ) ∈ Ixαi , i = 1, . . . , m.
Since the difference of two different coset representatives cannot give the zero coset,
we have ci = c0i for all i = 1, . . . , m. Thus φ is well defined.
14
AMBEDKAR DUKKIPATI, NITHISH PAI AND MARIA FRANCIS
Clearly, φ is a surjective map by construction. We now have to prove that φ is an
injective mapping. Consider a polynomial f ∈ A[x1 , . . . , xn ] such that φ(f + a) =
(0, . . . , 0). Let us assume that f ∈
/ a. Since B forms an order span for A[x1 , . . . , xn ]/a,
m
P
we can obtain ci ∈ CIxαi , i = 1, . . . , m such that f =
ci xαi mod a. Further, atleast
i=1
one of the ci , i ∈ {1, . . . , m}, is nonzero. This implies that φ(f + a) also maps to
(c1 , . . . , cm ). Therefore φ is not a well defined mapping. This is a contradiction and
f ∈ a. Thus, the kernel of φ, ker(φ) = {0 + a}. This implies that φ is an injective
mapping. Hence, it follows that φ is an isomorphism.
Definition 5.6. Let a be an ideal such that A[x1 , . . . , xn ]/a is a finitely generated Amodule. Let O be an order ideal and G = {g1 , . . . , gs } be an acyclic O-border prebasis
consisting of polynomials in a. G is an acyclic O-border basis of a if Mon(O) is an
order span of A[x1 , . . . , xn ]/a.
The next proposition shows that these polynomials indeed generate the ideal in a.
Proposition 5.7. Let a be an ideal such that A[x1 , . . . , xn ]/a is finitely generated as
an A-module. Let O be an order ideal and let G = {g1 , . . . , gs } be an acyclic O-border
basis for a. Then a is generated by G.
Proof. Let Mon(O) = {xα1 , . . . , xαt } be the monomial part of O and G = {g1 , . . . , gs }
be an acyclic O-border basis of a. Consider, f ∈ a. By Algorithm 4.2, we have
f1 , . . . , fs ∈ A[x1 , . . . , xn ] and a1 , . . . , at ∈ A such that
f = f1 g1 + . . . + fs gs + a1 xα1 + . . . + at xαt .
(2)
P
Pt
Ps
t
Now, f − i=1 fi gi ∈ a. This implies, i=1 ai xαi ∈ a. Let h = i=1 ai xαi . Suppose
h 6= 0, then ai ∈
/ Ixαi \ {0}, for all i = 1, . . . , t. If ai ∈ Ixαi , for all i = 1, . . . , t, then
indO (h) = 1. Then, (2) is not a valid output of Algorithm 4.2. But we are also
given that G is P
an acyclic border basis of a. The order span property of border basis
implies that if ti=1 ai xαi ∈ a for ai 6= 0 for all i = 1, . . . , t, then ai ∈PIxαi for some
i ∈ {1, . . . , t}. This is a contradiction. Hence, h = 0. We have, f = si=1 fi gi . The
other inclusion follows from the fact that G ⊆ a.
We need to verify if an acyclic O-border basis exists for every ideal, a in A[x1 , . . . , xn ].
We first address, below, whether an acylic O-border basis for a given an order ideal,
O, exists. We also prove the uniqueness of the acyclic O-border basis.
Theorem 5.8. Let O be an order ideal, and let a be an ideal such that A[x1 , . . . , xn ]/a
is a finitely generated A-module. If Mon(O) is an order span then there exists a
unique acyclic O-border basis for a.
Proof. Let Mon(O) = {xα1 , . . . , xαt } be the monomial part of O, and let ∂O =
{c1 xβ1 , . . . , cs xβs } be the border of O. We now prove that an O-border basis exists
for a. Since Mon(O) is an order span basis, for each ci xβi ∈ ∂O one can find
P
(i)
(i)
aj xαi s ∈ O such that ci xβi = tj=1 aj xαj mod a. We define G as
βi
G = {ci x −
t
X
j=1
(i)
aj xαj | 1 ≤ i ≤ s}.
BORDER BASES FOR POLYNOMIAL RINGS OVER NOETHERIAN RINGS
15
Clearly,
G is an acyclic O-border prebasis. Now, G ⊆ a and A[x1 , . . . , xn ]/a =
P
{ ai xαi + a | xαi ∈ Mon(O) and ai ∈ A}. Hence, G is an O-border basis of a.
To prove the uniqueness of O-border basis, consider two acyclic O-border bases
for a. Let G = {g1 , . . . , gt } and G = {g10 , . . . , gt0 } such that
βi
gi = ci x −
t
X
(i)
(i)
0(i)
0(i)
aj xαj , where each aj xαj ∈ O and,
j=1
gi0 = ci xβi −
t
X
aj xαj , where each aj xαj ∈ O.
j=1
We have,
gi −
gi0
=
t
X
(i)
aj x α j
j=1
−
t
X
0(i)
aj xαj ∈ a.
j=1
This implies,
t
X
0(i)
(i)
(aj − aj )xαj ∈ a.
j=1
(i)
0(i)
Since aj and aj are coset representatives of A/Ixαj and the difference of two dif0(i)
(i)
ferent cosets cannot be a zero coset, we have aj = aj . Therefore, gi = gi0 . Hence,
the acyclic O-border basis of a is unique.
Thus, the question of existence of a border basis for an ideal reduces to the following
questions. Given an ideal a, (i) does there always exist a proper order function, I
such that the monomial part of the order ideal, Mon(O) constructed from I forms
an order span for A[x1 , . . . , xn ]/a and (ii) will the corresponding O-border basis be
acyclic. We use the theory of Gröbner bases to establish the result.
Theorem 5.9. Given an ideal a such that A[x1 , . . . , xn ]/a is finitely generated as
an A-module, there always exists an acyclic border basis of a corresponding to some
order ideal, O.
Proof. Let ≺ be a monomial order on A[x1 , . . . , xn ]. Let G0 = {g10 , . . . , gt0 } be a
Gröbner basis of a. Consider the order function fixed by G, I (G) . Since A[x1 , . . . , xn ]/a
is finitely generated, the mapping is proper. Let O≺ be the order ideal corresponding to I (G) . It follows from Corollary 5.4 that Mon(O≺ ) forms an order span. Let
∂O≺ = {c1 xβ1 , . . . , cs xβs } be the border of O≺ . Let G be the O≺ -border prebasis
constructed along the same lines as in the proof of Theorem 5.8. Therefore, each
polynomial gi in G is of the form,
gi = ci xβi −
t
X
j=1
(i)
where ci xβi ∈ ∂O and each aj xαj ∈ O.
(i)
aj xαj , i = 1, . . . , s,
16
AMBEDKAR DUKKIPATI, NITHISH PAI AND MARIA FRANCIS
(i)
For any gi ∈ G, monomial ordering imposes that for every nonzero aj xαj , xαj ≺
xβi . Also, for two distinct border terms ci xβi and cj xβj such that xβi 6= xβj , either
xβi ≺ xβj or xβj ≺ xβi . The acyclic property of G follows from these two observations.
Theorem 5.8 implies that G forms a unique acyclic O≺ -border basis for a.
For any polynomial f ∈ A[x1 , . . . , xn ], given an acylic O-border prebasis, G and
an order ideal, O, O-remainder of f is denoted by remO,G (f ) (Definition 4.3).
Proposition 5.10. Let a ⊆ A[x1 , . . . , xn ] be an ideal such that A[x1 , . . . , xn ]/a is
finitely generated as an A-module and G ⊆ a be an acyclic O-border basis of a. For
any f ∈ A[x1 , . . . , xn ], f ∈ a if and only if remO,G (f ) = 0.
Proof. Let Mon(O) = {xα1 , . . . , xαt } be the monomial part of O and G = {g1 , . . . , gs }
be an acyclic O-border basis of a. By Algorithm 4.2, we have f1 , . . . , fs ∈ A[x1 , . . . , xn ]
and a1 , . . . , at ∈ A such that
f=
s
X
fi gi + remO,G (f ),
(3)
i=1
P
aj x . If remO,G (f ) = 0 then f = si=1 fi gi . Hence f ∈ a.
P
Now let f ∈ a. Then f − si=1 fi gi ∈ a. This implies, remO,G (f ) ∈ a. Suppose
remO,G (f ) is not equal to zero. The proof proceeds along the same lines as the
proof of Proposition 5.7, where we take h = remO,G (f ) and arrive at a contradiction.
Hence, when f ∈ a the remainder of f w.r.t. G is zero.
where remO,G (f ) =
Pt
j=1
αj
The above proposition enables us to solve the ideal membership problem provided
the acyclic border basis of the ideal is known. However, it must be noted that the
remainder on division by an acyclic O-border basis for any f ∈ A[x1 , . . . , xn ] is not
unique.
Below we define the normal form of a polynomial w.r.t. an acyclic border basis.
Definition 5.11. Let a be an ideal such that A[x1 , . . . , xn ]/a is finitely generated
as an A-module. Let G ⊆ a be an acyclic O-border basis for a. Let Mon(O) =
{xα1 , . . . , xαt } be the monomial part of O, CIxαi be the set of coset representatives of
the equivalence classes A/Ixαi and f be any polynomial in A[x1 , . . . , xn ]. Let r be a
polynomial given by r = a1 xα1 + · · · + at xαt , where ai ∈ CIxαi , 1 ≤ i ≤ t. Then r is
said to be the normal form of f if f = r mod a.
The normal form of a polynomial is denoted by NFO,G (f ). We now prove that
every polynomial f in A[x1 , . . . , xn ] has a unique normal form.
Proposition 5.12. Let a ⊆ A[x1 , . . . , xn ] be an ideal such that A[x1 , . . . , xn ]/a is
finitely generated as an A-module. Let G ⊆ a be an acyclic O-border basis for a. For
any polynomial f in A[x1 , . . . , xn ], the normal form of f is unique.
Proof. Let Mon(O) = {xα1 , . . . , xαt } be the monomial part of O. Let s0 ≤ t be the
number of monomials in the scalar border, ∂Os and G = {g1 , . . . , gs } be an acyclic
BORDER BASES FOR POLYNOMIAL RINGS OVER NOETHERIAN RINGS
17
O-border basis of a. The existence of a normal form, NFO,G (f ) for any polynomial
f ∈ A[x1 , . . . , xn ] is a consequence of the following equality,
A[x1 , . . . , xn ]/a = {
k
X
ai xαi + a | ai ∈ CIxαi }.
i=1
Now, we prove the uniqueness of the normal form of f . Let r1 and r2 be two different
normal forms for f . Then f = r1 mod a and
P f = r2 mod a. This
P implies, r1 =
r2 mod a. Therefore, r1 − r2 ∈ a. Let r1 = ti=1 bi xαi and r2 = ti=1 b0i xαi , where
P
bi and b0i are coset representatives in CIxαi . Then, r1 − r2 = ti=1 (bi − b0i )xαi . If
r1 6= r2 , then there is atleast one i such that bi 6= b0i . This implies that bi − b0i 6= 0.
Since bi and b0i are coset representatives of distinct cosets, bi − b0i ∈
/ Ixαi . Therefore,
(r1 − r2 ) ∈
/ a. Hence a contradiction. Thus r1 = r2 and the normal form of a
polynomial is unique.
The below result states that if we can associate a monomial order to an order ideal
O, then the reduced Gröbner basis of a w.r.t. to that monomial order is a subset of
the acyclic border basis associated with O.
Proposition 5.13. Let a be an ideal. Let ≺ be a monomial order. Let O≺ be an order
ideal corresponding to ≺ such that Mon(O) forms an order span for A[x1 , . . . , xn ]/a.
Then Pauer reduced Gröbner basis (Pauer, 2007) of a w.r.t. ≺ is a subset of the
acyclic O-border basis of a.
Proof. Let Mon(O≺ ) = {xα1 , . . . , xαt } be the monomial part of O≺ and ∂O≺ =
{c1 xβ1 , . . . , cs xβs } be the border of O≺ . Let G = {g1 , . . . , gs } be the acyclic O≺ -border
basis for a. The acyclic O≺ -border basis is constructed as in the proof of Theorem 5.9.
Since G is an acyclic O≺ -border basis we have from Proposition 5.10 that for any
f ∈ a, f reduces to zero. This implies that G is a Gröbner basis of a. Further,
hLt(a)i is generated by ∂O≺ . Recall that, hLc(α, a)i = hLc(f ) : f ∈ a, Lm(f ) = xα i
and hLc(< α, a)i = hLc(f ) : f ∈ a, α ∈ deg(f ) + Mn , Lm(f ) 6= xα i. Clearly, for
each monomial xα , in the monomial part, Mon(O≺ ), hLc(α, a)i = Ixα . From the
definition of order ideal hLc(α, a)i 6= h1i. For each monomial xα in the monomial
border, ∂O≺m , hLc(α, a)i = h1i. Also, for each monomial xα ∈ Lm(G), we have
Gen(α, a) = {η(a, hLc(< α, a)i) : a ∈ Gen(hLc(α, a)i)} \ {0},
where η(a, hLc(< α, a)i) maps to an element in the coset a + hLc(< α, a)i. Consider
the set, ∂O≺red = {cxα ∈ ∂O≺ : c ∈
/ hLc(< α, a)i}. This set contains all the terms of
α
the form cx in the border ∂O≺ such that c cannot be expressed as a combination
of leading coefficients of those monomials that properly divide xα . Clearly, ∂O≺red ⊆
∂O≺ . Let Gred ⊆ G consist of polynomials in G with the border term in ∂O≺red . It
can easily be seen that hLt(a)i = h∂O≺red i. Therefore, Gred is a Gröbner basis for a.
Also, it is clear from the construction of ∂O≺red that Gen(α, a) = {c : cxα ∈ ∂O≺red }.
We now prove that G satisfies the two properties of Pauer’s reduced Gröbner basis.
18
AMBEDKAR DUKKIPATI, NITHISH PAI AND MARIA FRANCIS
The bijectivity of the map,
φ : {g ∈ Gred : deg(g) = α} −→ Gen(α, a)
φ(g) 7−→ Lc(g)
follows from the observation that corresponding to each border term cxβ ∈ ∂O≺ ,
there is exactly one polynomial g ∈ G such that the border term in g is cxβ . If we
had not considered the reduced border ∂O≺red , then for all g ∈ G \ Gred , φ(g) will
map to zero. Also, each polynomial gi ∈ Gred is of the form,
ci x β i −
t
X
(i)
(i)
aj xαj where ci xβi ∈ ∂O≺ and aj xαj ∈ O≺ .
j=1
(i)
(i)
Since for each aj xαj , aj ∈ A/Ixαj we have that
(i)
(i)
η(aj , hLc(αj , a)i) = aj . Hence Gred satisfies the second condition of Pauer’s reduced
Gröbner basis. Therefore, Gred is a reduced Gröbner basis for a.
Theorem 5.14. Let a be a nonzero ideal in A[x1 , . . . , xn ], O be an order ideal and
∂O = {c1 xβ1 , . . . , cs xβs } be its border. Let Mon(O) = {xα1 , . . . , xαt } be the monomial
part of O and G = {g1 , . . . , gs } be an acyclic O-border prebasis. Then the following
statements are equivalent.
G is an acyclic O-border basis for a.
G
f ∈ a if and only if f −
→+ 0.
fP∈ a if and only if there exists f1 , . . . , fs ∈ A[x1 , , . . . , xn ] such that f =
s
i=1 fi gi and max{deg(fi ) | fi gi 6= 0, i = 1, . . . , s} = indO (f ) − 1.
(iv) The border form of a, BFO (a) = hc1 xβ1 , . . . , cs xβs i.
(i)
(ii)
(iii)
Proof. (i) ⇒ (ii). The claim follows from the proof of Proposition 5.10.
(ii) ⇒ (iii). Let f ∈ a. By the border division algorithm, there exists
Ps f 1 , . . . , f s ∈
A[x1 , , . . . , xn ], deg(fi ) ≤ indO (f ) − 1, 1 ≤ i ≤ s such that f =
i=1 fi gi . Assume that max{deg(fi )} indO (f ) − 1. It can easily be verified that indO (fi gi ) ≤
indO (gi ) + deg(fi ). The definition of O-border prebasis implies that indO (gi ) = 1.
Thus indO (fi gi ) ≤ deg(fi ) + 1 indO (f ) − 1 + 1. Thus indO (fi gi ) indO (f ). Also
it can be seen that, indO (f + g) ≤ max{indO (f ), indO (g)}, when either indO (f ) ≥ 1
or indP
O (g) ≥ 1. Thus,
indO ( si=1 fi gi ) indO (f ). This is a contradiction. Hence max{deg(fi )} = indO (f )−
1.
(iii) ⇒ (iv) Since each gi ∈ a and BFO (gi ) = ci xβi , we have hc1 xβ1 , . . . , cs xβs i ⊆
BFO (a). Consider a polynomial f ∈ a. Suppose indO (f ) ≥ 1, then by Definitions 3.22
and 3.25 each term in BFO (f ) is divisible by cxβ ∈ ∂O. Hence, it follows that
BFO (f ) ∈ hc1 xβ1 , . . . , cs xβs i.
Let I be the proper order function associated with the order ideal O. Now let us
assume
there exists a polynomial f ∈ a \ {0} such that indO (f ) = 0 i.e.,
Pt that αthe
i
f = i=1P
ci x where ci ∈
/ Ixαi . Then by hypothesis, there exist fi s, 1 ≤ i ≤ s, such
s
that f = i=1 fi gi and max{deg(fi ) | fi gi 6= 0, i = 1, . . . , s} = 0 − 1 = −1. This is
BORDER BASES FOR POLYNOMIAL RINGS OVER NOETHERIAN RINGS
19
not possible since deg(f ) ≥ 0 for all f ∈ A[x1 , . . . , xn ]. This implies that f = 0 which
is a contradiction. Therefore, indO (f ) ≥ 1. Thus, BFO (a) ⊆ hc1 xβ1 , . . . , cs xβs i. The
claim follows.
(iv) ⇒ (i). Consider a polynomial f ∈ A[x1 , . . . , xn ]. By the border division algorithm, we have f1 , . . . , fs ∈ A[x1 , . . . , xn ] and a1 , . . . , at ∈ A such that
f = f1 g1 + · · · + fs gs + a1 xα1 + · · · + at xαt .
P
P
Since si=1 fi gi ∈ a, f = h mod a, where h = tj=1 aj xαj . We are given that the
O-border prebasis, G is acyclic. Without loss of generality, let us assume that G is
well ordered. Find the smallest i for which the border term in gi belongs to ∂Os and
assume that the monomial in the border term is xα1 . Let G1 ⊆ G represent all the
polynomials for which the border monomial is xα1 and let |G1 | = s1 .
Let the ideal Ixα1 be generated by {c1 , . . . , cs1 } and let b1 ∈ CIxα1 be the coset
representative of a1 + Ixα1 . Then there exist d1 , . . . , ds1 ∈ A such that
(a1 − b1 ) = d1 c1 + · · · + ds1 cs1 .
Let h1 = h − d1 g1 + · · · + ds1 gs1 . Therefore we have,
h1 = b1 xα1 + b02 xα2 + · · · + b0t xαt ,
where b0i ∈ A, i ∈ {2, . . . , t}. Further, h1 = h mod a. Repeating the above process
for the remaining monomials in ∂Os in the same sequence as the well ordered basis,
we get
hs0 = b1 xα1 + . . . + bs0 xαs0 + bs0 +1 xαs+1 + · · · + bt xαt ,
where each bi is a coset representative in CIxαi . Note that for xαi , s0 + 1 ≤ i ≤ t,
Ixαi = {0} and bi ∈ A. The acyclicity of the basis ensures that at each stage, i, all
bj , j i will not be modified. Further, at every stage i, the intermediate polynomial,
hi = hi−1 mod a. Therefore, hs0 = h mod a which implies that f = hs0 mod a.
Further, each bi is a coset representative in CIxαi where xαi ∈ Mon(O). Hence,
P
A[x1 , . . . , xn ]/a = { ti=1 ai xαi + a | ai ∈ CIxαi }.
To prove the secondPcondition of the order span definition (Definition 5.1), consider
a polynomial f = ti=1 ai xαi ∈ a. Then, there exists an i ∈ {1, . . . , t} such that
ai xαi ∈ BFO (f ). By hypothesis we have, BFO (a) = hc1 xβ1 , . . . , cs xβs i. Since BFO (a)
is an ideal generated by terms, we have ai xαi ∈ BFO (a). Thus, there exists terms
di xγi , 1 ≤ i ≤ s such that
s
X
ai x α i =
(di xγi )(ci xβi ).
i=1
β
Since for all x |x
αi
Ixβ ⊆ Ixαi , we have ai ∈ Ixαi .
0
Consider a proper order function, I such that for any xα ∈ Mon(A[x1 , . . . , xn ]),
0
0
0
0
either Ixα ( Ixα or Ixα = Ixα . Let O be the order ideal associated with I and
0
0
0
0
∂O = {c1 xβ1 , . . . , cl xβl } be the corresponding border. Assume that Mon(O ) satisfies
0
0
(i) and (ii). Consider the ideal, A = hc1 xβ1 , . . . , cl xβl i. We have, A ( BFO (a)
0
since Ixα ( Ixα for some xα . Consider cω xω ∈ BFO (a) \ A. Let h denote the
remO,G (cω xω ). Then each non zero term in h is of the form dγ xγ such that dγ ∈ A/Ixγ
20
AMBEDKAR DUKKIPATI, NITHISH PAI AND MARIA FRANCIS
0
0
and xγ ∈ Mon(O). Since dγ ∈
/ Ixγ , it is not an element of Ixγ . Also, cω ∈
/ Ixω .
0
Therefore, Mon(O ) fails to satisfy Condition (ii) of the definition of order span for
cω xω −h and therefore we have a contradiction. Thus G = {g1 , . . . , gs } is an O-border
basis for a.
6. A Full Example
In this section, we illustrate the concepts given in this paper with an example.
Example 6.1. Let us consider the polynomial ring, Z[x, y]. Let I be an order function such that I1 = {0}, Ix = {0}, Iy = {0}, Ixy = {0}, Iy2 = {0}, Ix2 = h2i,
Ix2 y = h2i and the rest of the monomials map to h1i. Therefore, CI1 = CIx = CIy
= CIxy = CIy2 = Z and CIx2 = CIx2 y = {0, 1}. The set,
O ={a1 , a2 x, a3 y, a4 xy, a5 y 2 , a6 x2 , a7 x2 y |
a1 , a2 , a3 , a4 , a5 ∈ Z, a6 , a7 ∈ {0, 1}}
is an order ideal corresponding to I. The monomial part of O is the set Mon(O) =
{1, x, y, x2 , y 2 , xy, x2 y}. The scalar border of the order ideal is ∂Os = {2x2 , 2x2 y}
and the monomial border is ∂Om = {x3 , y 3 , xy 2 , x2 y 2 , x3 y}. Thus the border of O is
the union of the scalar border and the monomial border, i.e.,
∂O = ∂Om ∪ ∂Os = {x3 , y 3 , xy 2 , x2 y 2 , x3 y, 2x2 , 2x2 y}.
Consider the set G = {g1 , g2 , g3 , g4 , g5 , g6 , g7 }, where g1 = x3 − x, g2 = y 3 − y, g3 =
xy 2 −xy, g4 = x2 y 2 −x2 y, g5 = x3 y−xy, g6 = 2x2 y−y 2 −y and g7 = 2x2 +2xy−y 2 −
2x − y. The set G is an O-border prebasis for the ideal a = hg1 , g2 , g3 , g4 , g5 , g6 , g7 i.
It is also clear that the O-border prebasis satisfies the properties of acyclicity. Hence
G is an acyclic O-border prebasis.
The set G is a Gröbner basis for a with deglex order with x > y. The proof of
Theorem 5.9 implies that the set G is an acyclic O-border basis for a. Hence the
border form ideal of a is generated by the border terms i.e.,
BFO (a) = hx3 , y 3 , xy 2 , x2 y 2 , x3 y, 2x2 , 2x2 yi.
Now we demonstrate the border division algorithm (Algorithm 4.2) for a polynomial,
f = x4 + 2x3 y 2 + x2 + 4xy + 15 w.r.t. G. We have Mon(O) = {xα1 , . . . , xα7 },
where xα1 = 1, xα2 = x, xα3 = y, xα4 = x2 , xα5 = y 2 , xα6 = xy and xα7 = x2 y.
The monomial border is ∂Om = {b1 , . . . , b5 } and the scalar border is ∂Os = {b6 , b7 },
where b1 = x3 , b2 = y 3 , b3 = xy 2 , b4 = x2 y 2 , b5 = x3 y, b6 = 2x2 y and b7 = 2xy.
(1) Initialize f1 = f2 = f3 = f4 = f5 = f6 = f7 = 0, l1 = l2 = l3 = l4 = l5 = l6 =
l7 = 0 and h = x4 + 2x3 y 2 + x2 + 4xy + 15.
(2) Since indO (h) = 2, Step 5 of the algorithm is executed. We have x4 = xb1
and deg(x) = indO (h) − 1. Hence, we have f1 = f1 + x and h = x4 + 2x3 y 2 +
x2 + 4xy + 15 − x(x3 − x). Thus, h = 2x3 y 2 + 2x2 + 4xy + 15. Return to Step
2.
BORDER BASES FOR POLYNOMIAL RINGS OVER NOETHERIAN RINGS
21
(3) Again, indO (h) = 2 and we return to Step 5 of the algorithm. We have x3 y 2 =
xb4 and deg(x) = indO (h) − 1. After the reduction step we have f4 = f4 + 2x
and h = 2x3 y 2 + 2x2 + 4xy + 15 − 2x(x2 y 2 − x2 y) = 2x3 y + 2x2 + 4xy + 15.
We return to Step 2.
(4) In this step, indO (h) = 1 and the polynomial h has the monomial x3 y in its
support. Since x3 y ∈ ∂Om , we go to Step 5. We have x3 y = 1 · b5 and
deg(1) = indO (h) − 1. We perform the operation f5 = f5 + 2 and h =
2x3 y + 2x2 + 4xy + 15 − 2(x3 y − xy). Hence, h = 2x2 + 6xy + 15. We return
to Step 2.
(5) We have indO (h) = 1 and none of the terms in h are in the monomial border,
∂Om . Hence, we perform Step 4 of the algorithm. We have 2x2 = 1·b1 . Thus,
we have f7 = f7 + 1 and h = 2x2 + 6xy + 15 − 1(2x2 + 2xy − y 2 − 2x − y).
Hence, h = 4xy + y 2 + 2x + y + 15.
(6) We have indO (h) = 0 and Step 3 of the algorithm is executed. We have
l1 = l1 + 15, l2 = l2 + 2, l3 = l3 + 1, l4 = l4 + 0, l5 = l5 + 1, l6 = l6 + 4 and
l7 = l7 + 0. The algorithm terminates and returns (f1 , . . . , f7 , l1 , . . . , l7 ).
Thus we have the following representation for f ,
f =xg1 + 0g2 + 0g3 + 1g4 + 2g5 + 0g6 + 1g7 +
15 + 2x + y + 0x2 + 1y 2 + 4xy + 0x2 y.
The O-remainder of f is remO,G (f ) = 15 + 2x + y + y 2 + 4xy. Since the remainder
is not equal to zero, by Proposition 5.10, f ∈
/ a. For this example, the normal form
of f is equal to the O-remainder,
NFO,G (f ) = remO,G (f ) = 15 + 2x + y + y 2 + 4xy.
7. Concluding Remarks
The theory of border bases is popular in the symbolic computation community owing
to its numerical stability and mathematical elegance. In (Francis & Dukkipati, 2014)
it has been shown that border bases can be extended to polynomial rings over rings if
the corresponding residue class ring is free and if one can find its free representation
with respect to a monomial order by using Gröbner bases. In this paper we attempted
to extend the theory to a much more general case.
A theory was built for a special class of generators of the ideal called acyclic border
bases. We proved that the set of acyclic border bases contains all the reduced Gröbner
bases associated with all term orders. A future direction in this work is to determine
an algorithmic characterization for border bases in this case.
References
Adams, W. W. & Loustaunau, P. (1994). An Introduction to Gröbner Bases, vol. 3 of Graduate
Studies in Mathematics. American Mathematical Society.
Ajtai, M. (1996). Generating hard instances of lattice problems (Extended Abstract). In: Proceedings of the 1996 ACM Symposium on the Theory of Computing, STOC (Miller, G. L., ed.).
ACM.
22
AMBEDKAR DUKKIPATI, NITHISH PAI AND MARIA FRANCIS
Ananth, P. V. & Dukkipati, A. (2012). Complexity of Gröbner basis detection and border basis
detection. Theoretical Computer Science 459, 1–15.
Buchberger, B. (1965). An Algorithm for Finding a Basis for the Residue Class Ring of a ZeroDimensional Polynomial Ideal (in German). Ph.D. thesis, University of Innsbruck, Austria.
(reprinted in ?).
Francis, M. & Dukkipati, A. (2014). On reduced Gröbner bases and Macaulay-Buchberger basis
theorem over noetherian rings. Journal of Symbolic Computation 65, 1–14.
Greuel, G.-M., Seelisch, F. & Wienand, O. (2011). The Gröbner basis of the ideal of vanishing
polynomials. Journal of Symbolic Computation 46(5), 561–570.
Hoffstein, J., Pipher, J. & Silverman, J. (1998). NTRU: A ring-based public key cryptosystem. In: Algorithmic Number Theory (Buhler, J., ed.), vol. 1423 of Lecture Notes in Computer
Science. Springer.
Kehrein, A. & Kreuzer, M. (2005). Characterizations of border bases. Journal of Pure and
Applied Algebra 196(2-3), 251–270.
Kehrein, A. & Kreuzer, M. (2006). Computing border bases. Journal of Pure and Applied
Algebra 205(2), 279–295.
Kehrein, A., Kreuzer, M. & Robbiano, L. (2005). An algebraist’s view on border bases. In:
Solving Polynomial Equations (Dickenstein, A. & Emiris, I. Z., eds.), vol. 14 of Algorithms
and Computation in Mathematics. Springer, pp. 169–202.
Lyubashevsky, V. & Micciancio, D. (2006). Generalized compact knapsacks are collision resistant. In: ICALP (2).
Möller, H. M. (1988). On the construction of Gröbner bases using syzygies. Journal of Symbolic
Computation 6(2-3), 345–359.
Pauer, F. (2007). Gröbner bases with coefficients in rings. Journal of Symbolic Computation
42(11-12), 1003–1011.
Sato, Y., Inoue, S., Suzuki, A., Nabeshima, K. & Sakai, K. (2011). Boolean Gröbner bases.
Journal of Symbolic Computation 46(5), 622–632.
Stetter, H. J. (2004). Numerical Polynomial Algebra. Society for Industrial and Applied Mathematics.
Trinks, W. (1978). Über B. Buchbergers Verfahren, Systeme algebraischer Gleichungen zu lösen.
Journal of Number Theory 10(4), 475–488.
Zacharias, G. (1978). Generalized Gröbner Bases in Commutative Polynomial Rings. Master’s
thesis, MIT,Cambridge,MA.
Example 7.1. We consider Example 3.5. The set G = {g1 , . . . , g7 }, where g1 =
xy − x, g2 = y 2 − y, g3 = x3 − 2y, g4 = x2 y − x2 + 10, g5 = 4x − 2y, g6 = 3y − 3x and
g7 = 2x2 − x + 5 is an O-border prebasis but it is not acyclic. Let G0 = {g10 , . . . , g70 }
where g10 = xy − x, g20 = y 2 − y, g30 = x3 − x2 + 6, g40 = x2 y − y + 5, g50 = 4x − 7,
g60 = 3y − x and g70 = 2x2 − 2y − 3x which is also a O-border prebasis but it is acyclic
since the permutation of G0 , {g10 , g20 , g30 , g40 , g70 , g60 , g50 } satisfies the acyclicity condition.
Example 7.2. Consider Example 3.6. The set G =
{g1 , . . . , g15 }, where g1 = x3 − 3, g2 = x2 y − 3u1 y, g3 = x2 z − 2z, g4 = xy 2 − x + 10,
g5 = xyz − 11xy, g6 = xz 2 − u2 u21 x2 , g7 = y 2 − x + u1 u2 , g8 = yz − 3y + 2,
g9 = z 2 + 5xz + 11u1 x, g10 = u21 y + u2 x + 3, g11 = (u22 − 3u1 )z − u22 y, g12 =
u21 xy + 3u1 x − 2z, g13 = (u22 − 1)xy + 2x2 , g14 = u1 xz + 3u1 x2 , g15 = u2 xz +
2u1 xy + 4x2 − 4z − 10u1 y + 14 is an acyclic O-border prebasis since the following
permutation of G, {g1 , g2 , g3 , g4 , g5 , g6 , g7 , g8 , g9 , g14 , g15 , g13 , g12 , g11 , g10 } satisfies the
acyclicity condition.
E-mail address: [email protected], [email protected], [email protected]
BORDER BASES FOR POLYNOMIAL RINGS OVER NOETHERIAN RINGS
23
Dept. of Computer Science & Automation, Indian Institute of Science, Bangalore
- 560012
| 0math.AC
|
Under review as a conference paper at ICLR 2016
A D EEP M EMORY- BASED A RCHITECTURE
TO -S EQUENCE L EARNING
FOR
S EQUENCE -
arXiv:1506.06442v4 [cs.CL] 7 Jan 2016
Fandong Meng1∗ Zhengdong Lu2 Zhaopeng Tu2 Hang Li2 and Qun Liu1
1
Institute of Computing Technology, Chinese Academy of Sciences
{mengfandong, liuqun}@ict.ac.cn
2
Noah’s Ark Lab, Huawei Technologies
{Lu.Zhengdong, Tu.Zhaopeng, HangLi.HL}@huawei.com
A BSTRACT
We propose D EEP M EMORY, a novel deep architecture for sequence-to-sequence learning,
which performs the task through a series of nonlinear transformations from the representation
of the input sequence (e.g., a Chinese sentence) to the final output sequence (e.g., translation
to English). Inspired by the recently proposed Neural Turing Machine (Graves et al., 2014),
we store the intermediate representations in stacked layers of memories, and use read-write
operations on the memories to realize the nonlinear transformations between the representations. The types of transformations are designed in advance but the parameters are learned
from data. Through layer-by-layer transformations, D EEP M EMORY can model complicated
relations between sequences necessary for applications such as machine translation between
distant languages. The architecture can be trained with normal back-propagation on sequenceto-sequence data, and the learning can be easily scaled up to a large corpus. D EEP M EMORY
is broad enough to subsume the state-of-the-art neural translation model in (Bahdanau et al.,
2015) as its special case, while significantly improving upon the model with its deeper architecture. Remarkably, D EEP M EMORY, being purely neural network-based, can achieve
performance comparable to the traditional phrase-based machine translation system Moses
with a small vocabulary and a modest parameter size.
1
I NTRODUCTION
Sequence-to-sequence learning is a fundamental problem in natural language processing, with many important
applications such as machine translation (Bahdanau et al., 2015; Sutskever et al., 2014), part-of-speech tagging
(Collobert et al., 2011; Vinyals et al., 2014) and dependency parsing (Chen & Manning, 2014). Recently, there
has been significant progress in development of technologies for the task using purely neural network-based
models. Without loss of generality, we consider machine translation in this paper. Previous efforts on neural
machine translation generally fall into two categories:
• Encoder-Decoder: As illustrated in left panel of Figure 1, models of this type first summarize the
source sentence into a fixed-length vector by the encoder, typically implemented with a recurrent neural network (RNN) or a convolutional neural network (CNN), and then unfold the vector into the target
sentence by the decoder, typically implemented with a RNN (Auli et al., 2013; Kalchbrenner & Blunsom, 2013; Cho et al., 2014; Sutskever et al., 2014);
• Attention-Model: with RNNsearch (Bahdanau et al., 2015; Luong et al., 2015) as representative, it
represents the source sentence as a sequence of vectors after a RNN (e.g., a bi-directional RNN (Schuster & Paliwal, 1997)), and then simultaneously conducts dynamic alignment with a gating neural network and generation of the target sentence with another RNN, as illustrated in right panel of Figure 1.
Empirical comparison between the approaches indicates that the attention-model is more efficient than the
encoder-decoder approach: it can achieve comparable results with far less parameters and training instances (Jean et al., 2015). This superiority in efficiency comes mainly from the mechanism of dynamic alignment, which avoids the need to represent the entire source sentence with a fixed-length vector (Sutskever et al.,
2014).
∗
The work was done when the first author worked as intern at Noah’s Ark Lab, Huawei Technologies.
1
Under review as a conference paper at ICLR 2016
Encoder-Decoder
Attention model
Figure 1: Two types of neural machine translators. Note the pictorial illustrations may deviate from individual
models e.g., Sutskever et al. (2014), on modeling details.
1.1
D EEP M EMORY- BASED A RCHITECTURE
Both encoder-decoders and attention models can be reformalized in the language of Neural Turing Machines
(NTM) (Graves et al., 2014), by replacing different forms of representations as content in memories and the
operations on them as basic neural net-controlled read-write actions, as illustrated in the left panel of Figure 2.
This is clear after realizing that the attention mechanism (Bahdanau et al., 2015) is essentially a special case of
reading (in particular with content-based addressing ) in NTM on the memory that contains the representation
of source sentence. More importantly, under this new view, the whole process becomes transforming the source
sentence and putting it into memory (vector or array of vectors), and reading from this memory to further
transform it into the target sentence. This architecture is intrinsically shallow in terms of the transformations
on the sequence as an object, with essentially one hidden “layer”, as illustrated in the left panel of Figure 2.
Note that although RNN (as encoding/decoding or equivalently as controller in NTM) can be infinitely deep,
this depth is merely for dealing with the temporal structure within the sequence. On the other hand, many
sequence-to-sequence tasks, e.g, translation, are intrinsically complex and calls for more complex and powerful
transformation mechanism than that in encoder-decoder and attention models.
Figure 2: Space holder for NTM view of things
For this reason, we propose a novel deep memory-based architecture, named D EEP M EMORY, for sequenceto-sequence learning. As shown in the right panel of Figure 2, D EEP M EMORY carries out the task through a
series of non-linear transformations from the input sequence, to different levels of intermediate memory-based
representations, and eventually to the final output sequence. D EEP M EMORY is essentially a customized and
deep version of NTM with multiple stages of operations controlled by a program, where the choices of “layers”
and types of read/write operations between layers are tailored for a particular task.
Through layer-by-layer stacking of transformations on memory-based representations, D EEP M EMORY generalizes the notion of inter-layer nonlinear mapping in neural networks, and therefore introduces a powerful new
deep architecture for sequence-to-sequence learning. The aim of D EEP M EMORY is to learn the representation
of sequence better suited to the task (e.g., machine translation) through layer-by-layer transformations. Just as in
deep neural network (DNN), we expect that stacking relatively simple transformations can greatly enhance the
expressing power and the efficiency of D EEP M EMORY, especially in handling translation between languages
with vastly different nature (e.g., Chinese and English) and sentences with complicated structures. D EEP M EM ORY naturally subsumes current neural machine translation models (Bahdanau et al., 2015; Sutskever et al.,
2014) as special cases, but more importantly it accommodates many deeper alternatives with more modeling
power, which are empirically superior to the current shallow architectures on machine translation tasks.
Although D EEP M EMORY is initially proposed for machine translation, it can be adapted for other tasks that
require substantial transformations of sequences, including paraphrasing, reasoning (Peng et al., 2015), and
semantic parsing (Yin et al., 2015). Also, in defining the layer-by-layer transformations, we can go beyond
the read-write operations proposed in (Graves et al., 2014) and design differentiable operations for the specific
structures of the task (e.g., in (Yin et al., 2015)).
2
Under review as a conference paper at ICLR 2016
RoadMap We will first discuss in Section 2 the read-write operations as a new form of non-linear transformation, as the building block of D EEP M EMORY. Then in Section 3, we stack the transformations together to
get the full D EEP M EMORY architecture, and discuss several architectural variations of it. In Section 4 we report
our empirical study of D EEP M EMORY on a Chinese-English translation task.
2
R EAD -W RITE AS A N ONLINEAR T RANSFORMATION
We start with discussing read-write operations between two pieces of memory as a generalized form of nonlinear transformation. As illustrated in Figure 3 (left panel), this transformation is between two non-overlapping
memories, namely R-memory and W -memory, with W -memory being initially blank. A controller operates the
read-heads to get the values from R-memory (“reading”), which are then sent to the write-head for modifying
the values at specific locations in W -memory (“writing”). After those operations are completed, the content
in R-memory is considered transformed and written to W -memory. These operations therefore define a transformation from one representation (in R-memory) to another (in W -memory), which is pictorially noted in the
right panel of Figure 3.
Figure 3: Read-write as an nonlinear transformation.
These basic components are more formally defined below, following those in a generic NTM (Graves et al.,
2014), with however important modifications for the nesting architecture, implementation efficiency and description simplicity.
Memory: a memory is generally defined as a matrix with potentially infinite size, while here we limit ourselves to pre-determined (pre-claimed) N × d matrix, with N memory locations and d values in each location.
In our implementation of D EEP M EMORY, N is always instance-dependent and is pre-determined by the algorithm1 . Memories of different layers generally have different d. Now suppose for one particular instance (index
omitted for notational simplicity), the system reads from the R-memory (MR , with NR units) and writes to
W -memory (denoted MW , with NW units)
R
R-memory: MR = {x1R , x2R , · · · , xN
},
R
W
W -memory: MW = {x1W , x2W , · · · , xN
},
W
with xnR ∈ RdR and xnW ∈ RdW .
Read/write heads: a read-head gets the values from the corresponding memory, following the instructions of
the controller, which also influences the controller in feeding the state-machine. D EEP M EMORY allows multiple
read-heads for one controller, with potentially different addressing strategies (see Section 2.1 for more details).
A write-head simply takes the instruction from controller and modifies the values at specific locations.
Controller: The core to the controller is a state machine, implemented as a RNN Long Short-Term Memory
(LSTM) (Hochreiter & Schmidhuber, 1997), with state at time t denoted as st (as illustrated in Figure 3). With
st , the controller determines the reading and writing at time t, while the return of reading in turn takes part in
updating the state. For simplicity, only one reading and writing is allowed at one time step, but more than one
read-heads are allowed. The main equations for controllers are then
R EAD VECTOR :
W RITE VECTOR :
S TATE UPDATE :
rt = FR (MR , st ; ΘR )
vt = FW (st ; ΘW )
st+1 = FD (st , rt ; ΘD )
1
It is possible to let the controller learn to determine the length of the memory, but that does not yield better performance
on our tasks and is therefore omitted here.
3
Under review as a conference paper at ICLR 2016
where FD (·), FR (·) and FW (·) are respectively the operators for dynamics, reading and writing2 , parameterized
by ΘD , ΘR , and ΘW . In D EEP M EMORY, 1) it is only allowed to read from memory of lower layer and write to
memory of higher layer, and 2) reading memory can only be performed after finishing the writing to it.
The above read-write operations transform the representation in R-memory to the new representation in W memory, while the design choice of read-write specifies the inner structure of the memories in W -memory.
The transformation is therefore jointly specified by the read-write strategies (e.g., the addressing described
in Section 2.1) and the parameters learned in a supervised fashion (described later in Section 3.2). Memory
with the designed inner structure, in this particular case vector array with instance specific length, offers more
flexibility than a fixed-length vector in representing sequences. This representational flexibility is particularly
advantageous when combined with proper reading-writing strategy in defining nonlinear transformations for
sequences, which will serve as the building block of the deep architecture.
2.1
A DDRESSING
2.1.1 A DDRESSING FOR R EADING
Location-based Addressing With location-based addressing (L-addressing), the reading is simply rt = xtR .
Notice that with L-addressing, the state machine automatically runs on a clock determined by the spatial structure of R-memory. Following this clock, the write-head operates the same number of times. One important
variant, as suggested in (Bahdanau et al., 2015; Sutskever et al., 2014), is to go through R-memory backwards
after the forward reading pass, where the controller RNN has the same structure but is parameterized differently.
Content-based Addressing With a content-based addressing (C-addressing), the return at t is
rt = FR (MR , st ; ΘR ) =
NR
X
g(st , xnR ; ΘR )
g̃(st , xnR ; ΘR )xnR , and g̃(st , xnR ; ΘR ) = PNR
,
R
n=1
n0 =1 g(st , xn0 ; ΘR )
where g(st , xnR ; ΘR ), implemented as a DNN, gives an un-normalized “affliiation” score for unit xnR in Rmemory. Clearly it is related to the attention mechanism introduced in (Bahdanau et al., 2015) for machine
translation and general attention models discussed in (Gregor et al., 2015) for computer vision. Content-based
addressing offers the following two advantages in representation learning:
1. it can focus on the right segment of the representation, as demonstrated by the automatic alignment
observed in (Bahdanau et al., 2015), therefore better preserving the information in lower layers;
2. it provides a way to alter the spatial structure of the sequence representation on a large scale, for which
the re-ordering in machine translation is an intuitive example.
Hybrid Addressing: With hybrid addressing (H-addressing) for reading, we essentially use two read-heads
(can be easily extended to more), one with L-addressing and the other with C-addressing. At each time t, the
controller simply concatenates the return of two individual read-heads as the final return:
NR
X
rt = [xtR ,
g(st , xnR ; ΘR )xnR ].
n=1
It is worth noting that with H-addressing, the tempo of the state machine will be determined by the L-addressing
read-head, and therefore creates W -memory of the same number of locations in writing. As shown later, Haddressing can be readily extended to allow read-heads to work on different memories.
2.1.2
A DDRESSING FOR W RITING
Location-based Addressing With L-addressing, the writing is simple. At any time t, only the tth location in
def
W -memory is updated: xtW = vt = FW (st ; ΘW ), which will be kept unchanged afterwards. For both locationand content-based addressing, FW (st ; ΘW ) is implemented as a DNN with weights ΘW .
Content-based Addressing In a way similar to C-addressing for reading, the units to write is determined
W
through a gating network g(st , xn,t
; ΘW ), where the values in W -memory at time t is given by
W
W
W
∆n,t = g̃(st , xn,t
; ΘW )FW (st ; ΘW ), xn,t
= (1 − αt )xn,t−1
+ αt ∆n,t , n = 1, 2, · · · , NW ,
2
Note that our definition of writing is slightly different from that in (Graves et al., 2014).
4
Under review as a conference paper at ICLR 2016
W
where xn,t
stands for the values of the nth location in W -memory at time t, α is the forgetting factor (similarly
defined as in (Graves et al., 2014)), g̃ is the normalized weight (with unnormalized score implemented also with
a DNN) given to the nth location at time t.
2.2
T YPES OF N ONLINEAR T RANSFORMATIONS
As the most “conventional” special case, if we use L-addressing
for both reading and writing, we actually get the familiar structure in units found in RNN with stacked layers (Pascanu et al.,
2014). Indeed, as illustrated in the figure right to the text, this
read-write strategy will invoke a relatively local dependency based on the original spatial order in R-memory.
It is not hard to show that we can recover some deep RNN model in (Pascanu et al., 2014) after stacking layers
of read-write operations like this. This deep architecture actually partially accounts for the great performance
of the Google neural machine translation model (Sutskever et al., 2014).
The C-addressing, however, be it for reading and writing, offers a means of major reordering on the units,
while H-addressing can add into it the spatial structure of the lower layer memory. In this paper, we consider
four types of transformations induced by combinations of the read and write addressing strategies, listed pictorially in Figure 4. Notice that 1) we only include one combination with C-addressing for writing since it is
computationally expensive to optimize when combined with a C-addressing reading (see Section 3.2 for some
analysis) , and 2) for one particular read-write strategy there are still a fair amount of implementation details to
be specified, which are omitted due to the space limit. One can easily design different read/write strategies, for
example a particular way of H-addressing for writing.
Figure 4: Examples of read-write strategies.
3
D EEP M EMORY : S TACKING T HEM T OGETHER
As illustrated in Figure 5 (left panel), the stacking is straightforward: we can just apply a transformation on top
of another, with the W -memory in lower layer being the R-memory of upper layer. The entire deep architecture
of D EEP M EMORY, with diagram in Figure 5 (right panel), can be therefore defined accordingly. Basically, it
starts with a symbol sequence (Layer-0), then moves to the sequence of word embeddings (Layer-1), through
layers of transformation to reach the final intermeidate layer (Layer-L), which will be read by the output layer.
The operations in output layer, relying on another LSTM to generate the target sequence, are similar to a memory
read-write, with the following two differences:
• it predicts the symbols for the target sequence, and takes the “guess” as part of the input to update the
state of the generating LSTM, while in a memory read-write, there is no information flow from higher
layers to the controller;
• since the target sequence in general has different length as the top-layer memory, it takes only pure
C-addressing reading and relies on the built-in mechanism of the generating LSTM to stop (i.e., after
generating a End-of-Sentence token).
Memory of different layers could be equipped with different read-write strategies, and even for the same strategy,
the configurations and learned parameters are in general different. This is in contrast to DNNs, for which the
transformations of different layers are more homogeneous (mostly linear transforms with nonlinear activation
function). A sensible architecture design in combining the nonlinear transformations can greatly affect the
performance of the model, on which however little is known and future research is needed.
5
Under review as a conference paper at ICLR 2016
Figure 5: Illustration of stacked layers of memory (left) and the overall diagram of D EEP M EMORY (right).
3.1
C ROSS -L AYER R EADING
In addition to the generic read-write strategies in Section 2.2, we also introduce the cross-layer reading into
D EEP M EMORY for more modeling flexibility. In other words, for writing in any Layer-`, D EEP M EMORY
allows reading from more than one layers lower than `, instead of just Layer-` − 1. More specifically, we
consider the following two cases.
M EMORY-B UNDLE : A M EMORY-B UNDLE, as shown
in Figure 6 (left panel), concatenates the units of two
aligned memories in reading, regardless of the addressing strategy. Formally, the nth location in the bundle
(`0 +`00 )
=
of memory Layer-`0 and Layer-`00 would be xn
`0 >
`00 > >
[(xn ) , (xn ) ] . Since it requires strict alignment between the memories to put together, M EMORY-B UNDLE
is usually on layers created with spatial structure of same
origin (see Section 3.3 for examples).
Figure 6: Cross-layer reading.
S HORT-C UT Unlike M EMORY-B UNDLE, S HORT-C UT allows reading from layers with potentially different
inner structures by using multiple read-heads, as shown in Figure 6 (right panel). For example, one can use
a C-addressing read-head on memory Layer-`0 and a L-addressing read-head on Layer-`00 for the writing to
memory Layer-` with `0 , `00 < `.
3.2
O PTIMIZATION
For any designed architecture, the parameters to be optimized include {ΘD , ΘR , ΘW } for each controller, the
parameters for the LSTM in the output layer, and the word-embeddings. Since the reading from each memory
can only be done after the writing on it completes, the “feed-forward” process can be described in two scales: 1)
the flow from memory of lower layer to memory of higher layer, and 2) the forming of a memory at each layer
controlled by the corresponding state machine. Accordingly in optimization, the flow of “correction signal”
also propagates at two scales:
• On the “cross-layer” scale: the signal starts with the output layer and propagates from higher layers to
lower layers, until Layer-1 for the tuning of word embedding;
• On the “within-layer” scale: the signal back-propagates through time (BPTT) controlled by the corresponding state-machine (LSTM). In optimization, there is a correction for each reading or writing
on each location in a memory, making the C-addressing more expensive than L-addressing for it in
general involves all locations in the memory at each time t.
The optimization can be done via the standard back-propagation (BP) aiming to maximize the likelihood of the
target sequence. In practice, we use the standard stochastic gradient descent (SGD) and mini-batch (size 80)
with learning rate controlled by AdaDelta (Zeiler, 2012).
6
Under review as a conference paper at ICLR 2016
3.3
A RCHITECTURAL VARIATIONS OF D EEP M EMORY
We discuss four representative special cases of D EEP M EMORY: A RC -I, II, III and IV, as novel deep architectures for machine translation. We also show that current neural machine translation models like RNNsearch can
be described in the framework of D EEP M EMORY as a relatively shallow case.
A RC -I The first proposal, including two variants (A RC I HYB and A RC -I LOC ), is designed to demonstrate the effect
of C-addressing reading between intermediate memory
layers, with diagram shown in the figure right to the text.
Both variants employ a L-addressing reading from memory Layer-1 (the embedding layer) and L-addressing writing to Layer-2. After that, A RC -I HYB writes to Layer-3 (Laddressing) based on its H-addressing reading (two readheads) on Layer-2, while A RC -I LOC uses L-addressing to
read from Layer-2. Once Layer-3 is formed, it is then put
together with Layer-2 for a M EMORY-B UNDLE, from which the output layer reads (C-addressing) for predicting the target sequence. M EMORY-B UNDLE, with its empirical advantage over single layers (see Section 4.2),
is also used in other three architectures for generating the target sequence or forming intermediate layers.
A RC -II As an architecture similar to A RC -I HYB , A RC -II is designed to investigate the effect of H-addressing reading from different layers of memory
(or S HORT-C UT in Section 3.1). It uses the same strategy as A RC -I HYB in
generating memory Layer-1 and 2, but differs in generating Layer-3, where
A RC -II uses C-addressing reading on Layer-2 but L-addressing reading on
Layer-1. Once Layer-3 is formed, it is then put together with Layer-2 as a
M EMORY-B UNDLE, which is then read by the output layer for predicting the
target sequence.
A RC -III We intend to use this design to study a deeper architecture and more
complicated addressing strategy. A RC -III follows the same way as A RC -II
to generate Layer-1, Layer-2 and Layer-3. After that it uses two read-heads
combined with a L-addressing write to generate Layer-4, where the two readheads consist of a L-addressing read-head on Layer-1 and a C-addressing readhead on the memory bundle of Layer-2 and Layer-3. After the generation of
Layer-4, it puts Layer-2, 3 and 4 together for a bigger M EMORY-B UNDLE to
the output layer. A RC -III, with 4 intermediate layers, is the deepest among the
four special cases.
Arc-IV This proposal is designed to study the efficacy of C-addressing writing in forming intermediate representation. It employs a L-addressing reading
from memory Layer-1 and L-addressing writing to Layer-2. After that, it uses
a L-addressing reading on Layer-2 to write to Layer-3 with C-addressing. For
C-addressing writing to Layer-3, all locations in Layer-3 are randomly initialized. Once Layer-3 is formed, it is then bundled with Layer-2 for the reading
(C-addressing) of the output layer.
3.3.1
R ELATION TO OTHER NEURAL MACHINE TRANSLATORS
As pointed out earlier, RNNsearch (Bahdanau et al., 2015) with its automatic alignment, is a special case
of D EEP M EMORY with shallow architecture. As pictorially illustrated in Figure 7, it employs L-addressing
reading on memory Layer-1 (the embedding layer), and L-addressing writing to Layer-2, which then is read
(C-addressing) by the output layer to generate the target sequence. As shown in Figure 7, Layer-2 is the only
intermediate layer created by nontrivial read-write operations.
On the other hand, the connection between D EEP M EMORY and encoder-decoder architectures is less obvious
since they usually require the reading from only the last cell (i.e., for a fixed-length vector representation)
between certain layers. More specifically, Sutskever et al. (2014) can be viewed as DeepMemory with stacking
layers of L-addressing read-write (described in Section 2.2) for both the encoder and decoder part, while the
two are actually connected through last hidden states of the LSTMs of the corresponding layers.
7
Under review as a conference paper at ICLR 2016
Attention-model diagram
D EEP M EMORY diagram
Figure 7: RNNsearch as a special case of D EEP M EMORY.
4
E XPERIMENTS
We report our empirical study on applying D EEP M EMORY to Chinese-to-English translation. Our training data
consist of 1.25M sentence pairs extracted from LDC corpora, with 27.9M Chinese words and 34.5M English
words respectively. We choose NIST 2002 (MT02) dataset as our development set, and the NIST 2003 (MT03),
2004 (MT04) and 2005 (MT05) datasets as our test sets. We use the case-insensitive 4-gram NIST BLEU score
as our evaluation metric, and sign-test (Collins et al., 2005) as statistical significance test. In training of the
neural networks, we limit the source and target vocabularies to the most frequent 16K words in Chinese and
English, covering approximately 95.8% and 98.3% of the two corpora respectively.
We compare our method with two state-of-the-art SMT and NMT3 models:
• Moses (Koehn et al., 2007): an open source phrase-based translation system with default configuration
and a 4-gram language model trained on the target portion of training data with (Stolcke et al., 2002);
• RNNsearch (Bahdanau et al., 2015): an attention-based NMT model with default setting
(RNNsearchDEFAULT ), as well as an optimal re-scaling of the model (on sizes of both embedding and
hidden layers, with about 50% more parameters) (RNNsearchBEST ).
For a fair comparison, 1) the output layer in eachD EEP M EMORY variant is implemented as Gated Recurrent
Units (GRU) in (Bahdanau et al., 2015), and 2) all the D EEP M EMORY architectures are designed to have the
same embedding size as in RNNsearchDEFAULT with parameter size less or comparable to RNNsearchBEST .
4.1 R ESULTS
The main results of different models are given in Table 1. RNNsearch (best) is about 1.7 points behind Moses
in BLEU on average, which is consistent with the observations made by other authors on different machine
translation tasks (Bahdanau et al., 2015; Jean et al., 2015). Remarkably, some sensible designs of D EEP M EM ORY (e.g., A RC -II) can already achieve performance comparable to Moses, with only 42M parameters, while
RNNsearchBEST has 46M parameters.
Clearly all D EEP M EMORY architectures yield performance significantly better than (A RC -I HYB , A RC -II &
A RC -III) or comparable (A RC -I LOC & A RC -IV) to the NMT baselines. Among them, A RC -II outperforms
the best setting of NMT baseline (RNNsearchBEST ), by about 1.5 BLEU on average with less parameters.
S YSTEMS
RNNsearchDEFAULT
RNNsearchBEST
A RC -I LOC
A RC -I HYB
A RC -II
A RC -III
A RC -IV
Moses
MT03
29.02
30.28
28.98
30.14
31.27*
30.15
29.88
31.61
MT04
31.25
31.72
32.02
32.70*
33.02*
33.46*
32.00
33.48
MT05
28.32
28.52
29.53*
29.40*
30.63*
29.49*
28.76
30.75
AVERAGE
29.53
30.17
30.18
30.75
31.64
31.03
30.21
31.95
PARAMETERS #
31M
46M
54M
54M
42M
53M
48M
–
Table 1: BLEU-4 scores (%) of NMT baselines: RNNsearchDEFAULT and RNNsearchBEST , D EEP M EMORY
architectures (A RC -I, II, III and IV), and phrase-based SMT system (Moses). The “*” indicates that the results
are significantly (p<0.05) better than those of the RNNsearchBEST .
3
There are recent progress on aggregating multiple models or enlarging the vocabulary(e.g., in (Jean et al., 2015)), but
here we focus on the generic models.
8
Under review as a conference paper at ICLR 2016
D ISCUSSION
About Depth: A more detailed comparison between RNNsearch (two layers), A RC -II (three layers) and A RC -III) (four layers), both quantitatively
and qualitatively, suggests that with deep architectures are essential to the superior performance of
D EEP M EMORY. Although the deepest architecture A RC -III is about 0.6 BLEU behind the A RC II, its performance on long sentences is significant better. Figure 8 shows the BLEU scores of
generated translations on the test sets with respect
to the length of the source sentences. In particular, we test the BLEU scores on sentences longer
than {0, 10, 20, 30, 40, 50, 60} in the merged
test set of MT03, MT04 and MT05. Clearly, on
sentences with length >30, A RC -III yields consistently higher BLEU scores than A RC -II. This
observation is further confirmed by our observations of translation quality (see Appendix) and is
consistent with our intuition that D EEP M EMORY
with its multiple layers of transformation, is especially good at modeling the transformations of representations essential to machine translation on relatively complicated sentences.
33
30
27
BLEU(%)
4.2
24
21
18
RNNsearch-best
ARC-II
ARC-III
15
12
0
10
20
30
40
50
sentence length (Merge)
60
Figure 8: The BLEU scores of generated translations on
the merged three test sets with respect to the lengths of
source sentences. The numbers on X-axis of the figure
stand for sentences longer than the corresponding length,
e.g., 30 for source sentences with > 30 words.
About C-addressing Read: Further comparison between A RC -I HYB and A RC -I LOC (similar parameter sizes)
suggests that C-addressing reading plays an important role in learning a powerful transformation between intermediate representations, necessary for translation between language pairs with vastly different syntactical
structures. This conjecture is further verified by the good performances of A RC -II and A RC -III, both of which
have C-addressing read-heads in their intermediate memory layers. However, if memory Layer-`+1 is formed
with only C-addressing read from memory Layer-`, and serves as the only going to later stages, the performance is usually less satisfying. Comparison of this design with H-addressing (results omitted here) suggests
that another read-head with L-addressing can prevent the transformation from going astray by adding the tempo
from a memory with a clearer temporal structure.
About C-addressing Write: The BLEU scores of A RC -IV are lower than that of A RC -II but comparable
to that of RNNsearchBEST , suggesting that writing with C-addressing alone yields reasonable representation. A
closer look shows that although A RC -IV performs poorly on very long sentences (e.g., source sentences with
over 60 words), it does fairly well on sentences with normal length. More specifically, on source sentences
with length ≤ 40, it outperforms RNNsearchBEST with 0.79 BLEU points. One possible explanation is that our
particular implementation of C-addressing for writing in A RC -IV (Section 3.3) relies heavily on the randomly
initialized content and is hard to optimize, especially when the structure of the sentence is complex, which
might need to be “guided” by another write-head or some smart initialization.
About Cross-layer Read: As another observation, cross-layer reading almost always helps. The performances of A RC -I, II, III and IV unanimously drop after removing the M EMORY-B UNDLE and S HORT-C UT
(results omitted here), even after the broadening of memory units to keep the parameter size unchanged. It might
be due to the flexibility gained in mixing different addressing modes and representations of different stages.
5
C ONCLUSION
We propose D EEP M EMORY, a novel architecture for sequence-to-sequence learning, which is stimulated by
the recent work of Neural Turing Machine (Graves et al., 2014) and Neural Machine Translation (Bahdanau
et al., 2015). D EEP M EMORY builds its deep architecture for processing sequence data on the basis of a series of
transformations induced by the read-write operations on a stack of memories. This new architecture significantly
improves the expressive power of models in sequence-to-sequence learning, which is verified by our empirical
study on a benchmark machine translation task.
9
Under review as a conference paper at ICLR 2016
R EFERENCES
Auli, Michael, Galley, Michel, Quirk, Chris, and Zweig, Geoffrey. Joint language and translation modeling
with recurrent neural networks. In Proceedings of EMNLP, pp. 1044–1054, 2013.
Bahdanau, Dzmitry, Cho, Kyunghyun, and Bengio, Yoshua. Neural machine translation by jointly learning to
align and translate. In Proceedings of ICLR, 2015.
Chen, Danqi and Manning, Christopher D. A fast and accurate dependency parser using neural networks. In
Proceedings of EMNLP, pp. 740–750, 2014.
Cho, Kyunghyun, van Merrienboer, Bart, Gulcehre, Caglar, Bougares, Fethi, Schwenk, Holger, and Bengio,
Yoshua. Learning phrase representations using rnn encoder-decoder for statistical machine translation. In
Proceedings of EMNLP, pp. 1724–1734, 2014.
Collins, Michael, Koehn, Philipp, and Kučerová, Ivona. Clause restructuring for statistical machine translation.
In Proceedings of ACL, pp. 531–540, 2005.
Collobert, Ronan, Weston, Jason, Bottou, Léon, Karlen, Michael, Kavukcuoglu, Koray, and Kuksa, Pavel.
Natural language processing (almost) from scratch. The Journal of Machine Learning Research, 12:2493–
2537, 2011.
Graves, Alex, Wayne, Greg, and Danihelka, Ivo. Neural turing machines. arXiv preprint arXiv:1410.5401,
2014.
Gregor, Karol, Danihelka, Ivo, Graves, Alex, and Wierstra, Daan. DRAW: A recurrent neural network for image
generation. arXiv preprint arXiv:1502.04623, 2015.
Hochreiter, Sepp and Schmidhuber, Jürgen. Long short-term memory. Neural computation, 9(8):1735–1780,
1997.
Jean, Sébastien, Cho, Kyunghyun, Memisevic, Roland, and Bengio, Yoshua. On using very large target vocabulary for neural machine translation. In ACL-IJNLP, 2015.
Kalchbrenner, Nal and Blunsom, Phil. Recurrent continuous translation models. In Proceedings of EMNLP,
pp. 1700–1709, 2013.
Koehn, Philipp, Hoang, Hieu, Birch, Alexandra, Callison-Burch, Chris, Federico, Marcello, Bertoldi, Nicola,
Cowan, Brooke, Shen, Wade, Moran, Christine, Zens, Richard, Dyer, Chris, Bojar, Ondrej, Constantin,
Alexandra, and Herbst, Evan. Moses: Open source toolkit for statistical machine translation. In Proceedings
of ACL on interactive poster and demonstration sessions, pp. 177–180, Prague, Czech Republic, June 2007.
Luong, Thang, Pham, Hieu, and Manning, Christopher D. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pp. 1412–1421, 2015.
Pascanu, Razvan, Gulcehre, Caglar, Cho, Kyunghyun, and Bengio, Yoshua. How to construct deep recurrent
neural networks. In Proceedings of ICLR, 2014.
Peng, Baolin, Lu, Zhengdong, Li, Hang, and Wong, Kam-Fai. Towards neural network-based reasoning. arXiv
preprint arXiv:1508.05508, 2015.
Schuster, Mike and Paliwal, Kuldip K. Bidirectional recurrent neural networks. Signal Processing, IEEE
Transactions on, 45(11):2673–2681, 1997.
Stolcke, Andreas et al. SRILM-an extensible language modeling toolkit. In Proceedings of ICSLP, volume 2,
pp. 901–904, 2002.
Sutskever, Ilya, Vinyals, Oriol, and Le, Quoc VV. Sequence to sequence learning with neural networks. In
Advances in Neural Information Processing Systems, pp. 3104–3112, 2014.
Vinyals, Oriol, Kaiser, Lukasz, Koo, Terry, Petrov, Slav, Sutskever, Ilya, and Hinton, Geoffrey. Grammar as a
foreign language. arXiv preprint arXiv:1412.7449, 2014.
10
Under review as a conference paper at ICLR 2016
Yin, Pengcheng, Lu, Zhengdong, Li, Hang, and Ben, Kao. Neural enquirer: Learning to query tables. arXiv
preprint arXiv:1512.00965, 2015.
Zeiler, Matthew D. Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701, 2012.
11
Under review as a conference paper at ICLR 2016
APPENDIX: ACTUAL T RANSLATION E XAMPLES
In appendix we give some example translations from D EEP M EMORY, more specically, A RC -II and A RC III, and compare them against the reference and the translation given by RNNsearch. We will focus on long
sentences with relatively complicated structures.
E XAMPLE T RANSLATION OF A RC -II
Source
反对派 国会 议员 昨天 拒绝 就 政府 提出 的 宪法 修正案 进行 表决
Reference
yesterday opposition parliament members refused to vote on the constitutional amendment
bill proposed by the government
RNNsearch
the members of the opposition parliament have refused to accept the decision on the
amendment of the constitution
DeepMemory (Arc-II)
the opposition parliament members yesterday refused to vote on the constitutional
amendment proposed by the government
Source
谢瓦尔德纳泽 在 讲话 中 强调 了 在 此 问题 上 与 俄 美 两 国 进行 合作 和 交换 情
报 的 重要性 。
Reference
in his address, shevardnadze stressed the importance of the exchange of intelligence
with russia and the united states on this matter.
RNNsearch
in his speech , UNK stressed the importance of cooperation and exchange between the
two countries on this question .
DeepMemory (Arc-II)
in his speech , UNK stressed the importance of cooperation and exchange of intelligence
between russia and the united states on this issue .
Source
博尔久扎 指出 , 独联体 集体 安全 条约 组织 快速 反应 部队 的 建立 有助于 遏制 中
亚 地区 的 极端 组织 势力
Reference
bordyuzha noted that the creation of a fast reaction force by the cis collective security
treaty organization would help contain the strength of extremist groups in central asia
RNNsearch
UNK pointed out that the establishment of the organization of the commonwealth of
independent states ( cis ) is conducive to the maintenance of extremist organizations
in the central asian region
DeepMemory (Arc-II)
UNK noted that the establishment of a rapid reaction force in the organization of
collective security of the commonwealth of independent states is conducive to curbing
the extremist forces in the central asian region
Source
英格兰 说 , 五角大楼 是 在 一 、 两 天 前 确认 这 名 囚犯 身份 的 , 国务院 将
就 这 一 决定 通报 囚犯 所 属 的 国家 并 安排 有关 释放 事宜 。
Reference
england said the pentagon verified the identity of this prisoner one or two days ago
and that the state department will notify the home country of the prisoner of this
decision and make arrangements for his release.
RNNsearch
england said that the pentagon was established in a and two days ago that the state
council would inform the state of the decision and arrangements for the release of the
detainees .
DeepMemory (Arc-II)
england said that the pentagon had confirmed the identity of the prisoners before one
or two days ago , and the state council would notify the countries involved in the decision
and arrange the release of the prisoners .
12
Under review as a conference paper at ICLR 2016
E XAMPLE T RANSLATION OF A RC -III
Source
美国 副国务卿 阿米塔吉 今天 对 叙利亚 发出 警告 , 指称 大马士革 与 华府 未来 关系
的 发展 , 端 看 叙利亚 是否 愿 在 伊拉克 及 黎巴嫩 问题 上 与 美国 合作 。
Reference
us deputy secretary of state armitage issued warnings to syria today, saying the
development of future relations between damascus and washington depends on whether syria
wants to cooperate with the united states on the issues of iraq and lebanon.
RNNsearch
us deputy secretary of state richard armitage issued a warning to syria today that the
development of damascus , washington , and washington could see if syria is willing
to cooperate with the united states in iraq and lebanon .
DeepMemory (Arc-III)
us deputy secretary of state richard armitage issued a warning today that the development
of the damascus - washington relations in the future will depend on whether syria is
willing to cooperate with the united states on iraq and lebanon issues .
Source
俄罗斯 国营 石油 公司 正 与 俄罗斯 国营 瓦斯 公司 进行 合并 , 加上 尤干斯克 石油
天然气 公司 后 , 三 者 合计 将 掌控 俄罗斯 石油 总 生产 的 五分之一 以及 全部 天
然气 生产 , 一举 成为 全球 规模 最 大 能源 集团 。
Reference
the russian state petroleum company is merging with the russian state gas company, and
with the inclusion of yuganskneftegaz, the combined three will account for a fifth of
russia's oil output and all its natural gas production, making it the largest global
energy group.
RNNsearch
the russian state - run oil company is merging with the state - owned gas company ,
and the state - owned gas corp .
DeepMemory (Arc-III)
russia 's state - owned petroleum company is merging with the russian state - owned
gas company , and the third country will have to control the one - fifth of russian
oil production and to become the biggest energy group in the world .
Source
希 土 两 国 同 为 北约 成员国 , 但 双方 在 爱琴海 领海 、 领空 以及 岛屿 大陆架 的
划分 和 塞浦路斯 等 问题 上 存在 利益 冲突 和 矛盾 , 也 影响 了 两 国 关系 的 正
常化 。
Reference
both greece and turkey are nato members, but the two have conflicts of interest and
clashes over maritime and air territories and continental shelf boundary disputes in
the aegean sea and the issue of cyprus, which have affected the normalization of their
bilateral relations.
RNNsearch
the two countries have maintained conflicts and contradictions in the fields of
territorial waters , airspace , and islands , as well as contradictions and conflicts .
DeepMemory (Arc-III)
greece and turkey are the member countries of the nato , but the conflict of interests
and contradictions between the two countries on the demarcation of the territorial
waters , airspace , and the islands of the islands also adversely affect the
normalization of relations between the two countries .
13
| 9cs.NE
|
Visualizing and Understanding Curriculum Learning
for Long Short-Term Memory Networks
Volkan Cirik, Eduard Hovy, Louis-Philippe Morency
arXiv:1611.06204v1 [cs.CL] 18 Nov 2016
Department of Computer Science
Language Technologies Institute
Carnegie Mellon University
Pittsburgh, PA 15213
{vcirik,hovy,morency}@cs.cmu.edu
Abstract
Curriculum Learning emphasizes the order of training instances in a computational learning setup. The core hypothesis is that simpler instances should be learned early as building blocks to learn more complex ones. Despite its usefulness,
it is still unknown how exactly the internal representation of
models are affected by curriculum learning. In this paper, we
study the effect of curriculum learning on Long Short-Term
Memory (LSTM) networks, which have shown strong competency in many Natural Language Processing (NLP) problems.
Our experiments on sentiment analysis task and a synthetic
task similar to sequence prediction tasks in NLP show that
curriculum learning has a positive effect on the LSTM’s internal states by biasing the model towards building constructive
representations i.e. the internal representation at the previous
timesteps are used as building blocks for the final prediction. We also find that smaller models significantly improves
when they are trained with curriculum learning. Lastly, we
show that curriculum learning helps more when the amount
of training data is limited.
Introduction
Inspired by the human learning process, Curriculum Learning (Elman, 1993; Bengio et al., 2009) is an algorithm that
emphasizes the order of training instances in a computational learning setup. The main idea is that learning easy instances first could be helpful for learning more complex ones
later in the training. The first algorithm proposed by Bengio
et al. (2009), which we refer as one-pass curriculum, creates
disjoint sets of training examples ordered by the complexity
and used separately during training. The second algorithm
called baby step curriculum uses an incremental approach
where groups of more complex examples are incrementally
added to the training set (Spitkovsky, Alshawi, and Jurafsky,
2010). These curriculum learning regimens were shown to
improve performance in some Natural Language Processing
and Computer Vision tasks (Pentina, Sharmanska, and Lampert, 2015; Spitkovsky, Alshawi, and Jurafsky, 2010).
Despite its usefulness, it is still unknown how exactly
computational models are affected internally by curriculum learning. An example of computational model particuCopyright c 2017, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
larly relevant to Natural Language Processing is Long ShortTerm Memory (LSTM) network (Hochreiter and Schmidhuber, 1997). LSTM networks have shown competitive performance in several domains such as handwriting recognition
(Graves et al., 2009) and parsing (Vinyals et al., 2015). Surprisingly, curriculum learning has not been studied in the
context of LSTM networks to our knowledge. Detailed visualizations and analyses of curriculum learning regimens
with LSTM will allow us to better understand how models
are affected and provides us insights when to use these regimens. Knowing how curriculum learning works, we can design new extensions and understand the nature of tasks most
suited for these learning regimens.
In this paper, we study the effect of curriculum learning on
LSTM networks. We created experiments to directly compare two curriculum learning regimens, one-pass and baby
step, with two baseline approaches that include the conventional technique of randomly ordering the training samples.
We use two benchmarks for our analyses. First, a synthetic
task is designed which is similar to several Natural Language Processing tasks where a sequence of symbols are
observed and a particular function (e.g. analogous to a linguistic or a semantic phenomenon) is aimed to be learned.
Second, we use sentiment analysis where the polarity of subjective opinions is classified – a fundamental task in Natural
Language Processing. As mentioned previously, this is the
first work studying LSTM networks on sentiment analysis
with curriculum learning to our knowledge.
Our visualizations and analyses on these two sequence
tasks are designed to study three main factors. First, we
compare the four learning regimens on how the LSTM network’s internal representations change as the final prediction
is computed. To this end, we simply decode the representations at intermediate steps. This analysis helps us understand
how a model handles the task with the help of curriculum
learning. Second, we investigate how the performance of
models with different complexities are affected by curriculum learning. Smaller yet accurate models are crucial in limited resource settings. Third, we study how the performance
of curriculum learning changes in low-resource setups. This
analysis provides us a valuable information considering lowresource scenarios are common in several data-driven learning domains such as Natural Language Processing.
Related Work
We review a list of topics related to our work in the context
of curriculum learning (CL), analysis of neural networks and
sentiment analysis with neural networks.
Curriculum Learning. Motivated by children’s language
learning, Elman (1993) studies the effect of learning regimen on a synthetic grammar task. He shows that a Recurrent Neural Network (RNN) is able to learn a grammar
when training data is presented from simple to complex order and fails to do so when the order is random. Bengio et
al. (2009) investigate CL from an optimization perspective.
Their experiments on synthetic vision and word representation learning tasks show that CL results in better generalization and faster learning. Spitkovsky, Alshawi, and Jurafsky
(2010) apply a CL strategy to learn an unsupervised parser
for sentences of length k and initialize the next parser for
sentences of length k + 1 with the previously learned one.
They show that learning a hybrid model using the parsers
learned for each sentence lengths achieves a significant improvement over a baseline model. Pentina, Sharmanska, and
Lampert (2015) investigate the CL in a multi-task learning setup and propose a model to learn the order of multiple tasks. Their experiments on a set of vision tasks show
that learning tasks sequentially is better than learning them
jointly. Jiang et al. (2015) provide a general framework for
CL and Self-Paced Learning where model picks which instances to train based on a simplicity metric. The proposed
framework is able to combine prior knowledge of curriculum with Self-Paced Learning in the learning objective.
Long Short-Term Memory Networks. LSTM networks
are a variant of RNNs (Elman, 1990) capable of storing information and propagating loss over long distance. Using a
gating mechanism by controlling the information flow into
the internal representation, it is possible to avoid the problems of training RNNs (Bengio, Simard, and Frasconi, 1994;
Pascanu, Mikolov, and Bengio, 2012). Several architectural
variants have been proposed to improve the basic model
(Cho et al., 2014; Chung et al., 2015; Yao et al., 2015; Kalchbrenner, Danihelka, and Graves, 2015; Dyer et al., 2015;
Grefenstette et al., 2015).
tionality and negation using sentiment analysis and sequence
auto-encoding.
Synthetic Tasks. Since the early days of neural networks,
synthetic tasks were used to test the capabilities of the models (Fahlman, 1991) and often serve as unit tests for machine learning models (Weston et al., 2015). Similar to the
first work on LSTM (Hochreiter and Schmidhuber, 1997),
many of the contemporary neural network models (Graves,
Wayne, and Danihelka, 2014; Kurach, Andrychowicz, and
Sutskever, 2015; Sukhbaatar et al., 2015; Vinyals, Fortunato,
and Jaitly, 2015) use synthetic tasks to compare and contrast
several architectures. Inspired by these studies, we also use
a synthetic task as one of our tasks to understand the effect
of CL on LSTMs.
Sentiment Analysis with Neural Networks. Several approaches have been proposed to solve sentiment analysis using neural networks. Socher et al. (2013) propose Recursive
Neural Networks to exploit the syntactic structure of a sentence. A number of extensions of this model have been proposed in the context of sentiment analysis (Irsoy and Cardie,
2014; Tai, Socher, and Manning, 2015). Other proposed approaches use CNN (Kalchbrenner, Grefenstette, and Blunsom, 2014; Kim, 2014) and the averaging of word vector
models (Iyyer et al., 2015; Le and Mikolov, 2014).
To our knowledge, this work is the first to study how
the internal representation of LSTM change in a curriculum
learning setup.
Curriculum Learning Regimens
Curriculum learning emphasizes the order of training instances, prioritizing simpler instances before the more complex ones. In this section, we describe two curriculum
learning regimens: one-pass curriculum originally proposed
by Bengio et al. (2009) and baby step curriculum from
Spitkovsky, Alshawi, and Jurafsky (2010). For both regimens, we develop the curriculum C using the same strategy
proposed by Spitkovsky, Alshawi, and Jurafsky (2010) who
assume that shorter sequences are easier to learn.
The following subsections are describing the two curriculum learning regimens as well as two baseline learning regimens.
One-Pass Curriculum
Visualization of Neural Networks. Although many of the
neural network studies provide quantitative analysis, there
are few qualitative analyses of neural networks. Zeiler and
Fergus (2014) visualize the feature maps of a Convolutional
Neural Network (CNN) (LeCun et al., 1998). They show that
feature maps at different layers show sensitivity to different
shapes, textures, and objects. Similarly, Karpathy, Johnson,
and Li (2015) analyze LSTM on character level language
modeling. Their analysis shows that deeper models with gating mechanisms achieve better results. They show that some
cells in LSTM learn to detect patterns and how RNNs learn
to generalize to longer sequences. More recently, Li et al.
(2016) use visualization to show how neural network models handle several linguistics phenomena such as composi-
Bengio et al. (2009) propose to use a dataset with simpler instances in the first phase of the training. After some number
of iterations, they switch to harder target dataset. The intuition is that after some training on simpler data, the model
M is ready to handle the harder target data. Here, we name
this regimen One-Pass curriculum (see Algorithm 1). The
training data D is sorted by a curriculum C and distributed
into k number of buckets. The training starts with the easiest
bucket. Unlike the previous work (Bengio et al., 2009), we
use early stopping – training stops for the bucket when the
loss or task’s accuracy criteria on held-out set do not get any
better for p number of epochs. Afterward, the next bucket
is being used and trained in the same way. The whole training is stopped after all buckets are used. Note that the model
uses each bucket only one time for the training, hence the
name.
Algorithm 1 One-Pass Curriculum
1: procedure OP-C URRICULUM(M ,D, C)
2:
D0 = sort(D, C)
3:
{D1 , D2 , ..., Dk } = D0 where C(da ) < C(db ) da ∈
Di , db ∈ Dj , ∀i < j
4:
for s = 1...k do
5:
while not converged for p epochs do
6:
train(M , Ds )
7:
end while
8:
end for
9: end procedure
Baby Steps Curriculum
The intuition behind Baby Steps curriculum (Bengio et al.,
2009; Spitkovsky, Alshawi, and Jurafsky, 2010) is that simpler instances in the training data should not be discarded,
instead, the complexity of the training data should be increased. After distributing data into buckets based on a curriculum, training starts with the easiest bucket. When the
loss or task’s accuracy criteria on a held-out set do not get
any better for p number of epochs, the next bucket and
the current data bucket are merged. The whole training is
stopped after all buckets are used (see Algorithm 2).
Algorithm 2 Baby Steps Curriculum
1: procedure BS-C URRICULUM(M ,D, C)
2:
D0 = sort(D, C)
3:
{D1 , D2 , ..., Dk } = D0 where C(da ) < C(db ) da ∈
Di , db ∈ Dj , ∀i < j
Dtrain = Ø
for s = 1...k do
Dtrain = Dtrain ∪ Ds
while not converged for p epochs do
train(M , Dtrain )
end while
end for
end procedure
4:
5:
6:
7:
8:
9:
10:
11:
Baseline Regimens
The first baseline, named No-CL, is the common practice
of shuffling the training data. For a neural network like our
LSTM models, this means that training is performed as usual
where one epoch sees all the training set in random order.
For all experiments described in the following section, models learned with the No-CL regiment are trained 10 times1 ,
to get a proper average performance.
The second baseline, named Sorted, also sees all the data
at each epoch but the ordering of the training instances is
based on the curriculum C. This is a simplification of the
two CL regimens presented in the previous subsections since
1
Note that this favors No-CL due to lower variance in results.
we are not partitioning the data based on its complexity
(i.e., based on the curriculum). We are simply reordering the
training set. A comparison between the No-CL and Sorted
baselines will allow us to study the importance of training
instance ordering.
Experiments
The main goal of our experiments is to better understand
how a computational model, specifically LSTM networks,
are affected internally by CL. We aim to observe (1) the effect of CL on the internal model representations, (2) how the
number of model parameters affect the performance of CL,
and (3) how the amount of data size change the contribution
of CL.
The following subsections present LSTM network and
our experimental probing methodology to analyze the
LSTMs internal representations at different stages in the sequence modeling process.
LSTM
We now describe LSTM networks. Let x1 , .., xT be a sequence of one-hot coded sequence of symbols of length T .
At each time step the LSTM updates its cells as follows:
sigm
i
f sigm t xt W e
o = sigm W
ht−1
tanh
m
ct = f
ht = o
ct−1 + i
tanh(ct )
m
(1)
(2)
(3)
In above equations, W t is a [4n × 2n] matrix to calculate
gate weights and new memory information g. W e is an embedding matrix for symbols. At time t, the sigmoid (sigm)
and tanh (tanh) non-linearities are applied element-wise to
the embedding representation of input xt W e and the output
of the network from the previous time step ht−1 . Vectors
i, f, o ∈ Rn are binary gates controlling the input, forget
and output gates respectively. The vector m ∈ Rn additively
modifies the cell ct .
We use the final hidden representation hT of the LSTM
to prediction with a projection matrix W p . In the case of
regression, we use relu(W p hT ) where W p is [1 × n] and
relu is rectified non-linearity. For classification, we predict
one of k class labels using softmax(relu(W p hT ))) where
W p is [k × n] and softmax is the softmax function.
Probing Internal States of LSTM
We aim to observe how the use of the internal representation
of LSTM at intermediate steps changes depending on the
learning regimen.
Each internal representation ht is probed using the
relu(W p ht ) functions learned for regression or the
softmax(relu(W p ht )) function learned for classification. By
moving these probes along the sequences, we can study the
intermediate representation at each time t.
Table 1: Probing of the LSTM model at intermediate timesteps for the Digit Sum dataset. Left column is the input and underlined
digit emphasizes the last input digit. Ground Truth is the running sum up to that point. Predicted values by the LSTM models
are in prediction column. The number in parantheses are the standard deviation. The intermediate representation of Baby Step
curriculum model is closer to running sum of the input sequence.
Input Sequence
1
10
109
1091
10917
109173
1091735
10917356
109173567
1091735670
10917356706
109173567064
1091735670642
10917356706428
109173567064286
1091735670642861
10917356706428614
109173567064286145
1091735670642861451
10917356706428614516
Ground Truth
1
1
10
11
18
21
26
32
39
39
45
49
51
59
65
66
70
75
76
82
Baby Steps Curriculum
Prediction
0.23
1.03
10.27
11.40
18.62
21.54
26.77
32.82
40.56
40.53
46.73
50.96
52.82
61.01
67.78
69.27
72.32
76.77
78.57
83.05
Digit Sum
We aim to simulate a low-resource sequence regression
problem considering many of the NLP tasks only have a few
thousand annotated samples. To this end, the Digit Sum task
is posed as follows. Given a sequence of symbols of digits,
the model is expected to predict the sum of digits. For instance given a sequence ”5 0 2 4 6” the expected output is
17.
Digit Sum task has similarities with our second sequence
task, sentiment analysis, where digits are analogous to the
word tokens in the natural language text and the summation is analogous to the subjective position of a sequence
on a topic. Our two evaluation tasks also have some interesting differences which allow evaluating a broader range
of sequence learning tasks. In sentiment analysis, the order
of words makes a difference whereas, in the Digit Sum, the
order of digits does not change the expected answer. Secondly, the learning setup is a classification of polarity levels
for sentiment analysis whereas it is a regression for the Digit
Sum.
Dataset Details. We define the evaluation task in the Digit
Sum dataset as the summation of 20 digits, a typical length
of sentences in natural language. Both the validation and
testing sets contain 200 sequences of 20 digits randomly
generated. The training set consist of 1000 sequences each
from length 2 to 20, allowing to develop the curriculum
automatically following Spitkovsky, Alshawi, and Jurafsky
(2010) procedure. This results in a dataset of size 19K instances2 .
Experimental Details. We used LSTM with hidden units
of 2, 4, 8, ..., 512 without peephole connections. For all configurations the size of digit embeddings and hidden units
2
We experimented with 10x smaller dataset size and observed
very similar results. We do not report these due to limited space.
One-Pass Curriculum
Prediction
0.00
0.00
0.00
0.00
0.00
0.00
4.29
14.19
29.28
32.20
52.86
67.70
74.59
83.31
87.42
83.43
82.60
83.34
80.42
82.52
Sorted
Prediction
0.86
1.32
10.59
11.29
19.30
22.99
29.34
37.21
46.38
48.01
55.50
60.58
63.03
70.91
76.54
76.46
78.88
81.56
80.44
83.68
No-CL
Prediction
0.88 (0.20)
1.10 (0.28)
10.20 (0.78)
10.88 (0.95)
17.89 (1.49)
20.84 (1.94)
25.90 (2.39)
32.00 (2.87)
39.12 (3.12)
39.26 (3.00)
45.95 (3.14)
50.32 (3.09)
52.6 (2.94)
61.11 (2.87)
67.97 (2.77)
69.33 (2.57)
73.05 (2.36)
77.67 (2.27)
78.07 (2.19)
83.36 (2.03)
are the same. We use RMSprop (Dauphin et al., 2015) with
learning rate 0.001 and decay rate of 0.9 with minibatches
of size 128. The patience parameter p for early stopping is
10. We use Dropout (Srivastava et al., 2014) of rate in range
{0,0.25,0.5} as suggested for LSTMs by Gal (2015).
Results
Probing Internal Model Representations. We analyze the
behavior of the model by using the intermediate representations during processing of a sequence. As we discussed previously, we feed the hidden representations of each digit to
the regression node to predict at each timestep of the input.
Table 1 shows the input sequence, ground truth, and predictions of the best models for each learning regimen based
on validation loss. The prediction of the model trained with
One-Pass curriculum and Sorted Baseline shows no correlation with the running sum of the digits. The Baby Step curriculum model is able to predict similar values to running
sum.
A sequence model can learn to solve this task in numerous
ways such as memorizing sequences due to overfitting, using a count table of the digits, or doing a running sum at each
time step. To analyze this, we report the average differences
between successive predictions (we call it ∆) and the last input digit(see Figure 1). At each timestep, the model trained
with Baby Step curriculum updates the hidden representation such that it correlates with the sum of digits observed
up to that point. It is also interesting to observe that Baby
Step curriculum shows better variance than the average of 10
random starts (No-CL). We emphasize that models are provided with the same sequences for training, yet, the learning
regiments results in different models.
Effects on Models With Different Complexities. Our
next experiment studies the effect of learning regimens on
models with different complexities. Figure 3 shows the
9
8
7
6
5
4
3
2
1
0
-10
Baby Step
9
9
8
7
6
5
4
3
2
1
0
-10
∆
9
9
8
7
6
5
4
3
2
1
0
-10
1
2
3
4
5
Input Digit
6
7
8
No-CL
∆
∆
∆
9
8
7
6
5
4
3
2
1
0
-10
1
2
3
4
5
Input Digit
6
7
8
One-Pass
1
2
3
4
5
6
7
8
9
4
5
6
7
8
9
Input Digit
Sorted
1
2
3
Input Digit
Figure 1: Visualization of Input Digit (input digit at time t) vs ∆ (the average difference between predictions at t and t − 1).
Shaded areas represent the variance. We expect ∆ to follow the input digit. Model trained with Baby Step curriculum shows
the least variance the input digit. Note that for No-CL, we plot all 10 runs (with different seeds).
Baby Step
One-Pass
Sorted
No-CL
2500
MSE
2000
Table 2: Classification Accuracies of Training Regimens on
Sentiment Analysis Task. The numbers in parantheses are
standard deviations. Model gets better at conjunctions if it is
trained with Baby-Step curriculum.
1500
1000
500
0
2
4
8
16
32
64
128
Number of Hidden Units
256
512
Figure 3: Mean Squared Error vs the number of units of
LSTM. With much smaller model, Baby Step curriculum
achieves the best results. The other model requires the right
complexity to achieve comparable results.
Mean Squared Loss (MSE) results for the Digit Sum with
LSTMs with varying hidden unit sizes. Baby Step curriculum achieves consistently better results even if the model has
much fewer parameters. Other regimens require the model to
have the right complexity. Efficient training of small models
is particularly important if we do not have large annotated
datasets to train big models. In addition, from practical perspective, to obtain smaller yet accurate models enables deploying fast and accurate models to a limited resource setting (Bucilu, Caruana, and Niculescu-Mizil, 2006; Hinton,
Vinyals, and Dean, 2015).
Sentiment Analysis
Sentiment analysis is an application of NLP to identify the
polarity of subjective information in given source (Pang and
Regimen
All
Conjunctions
No-CL from (Tai, Socher, and Manning, 2015)
No-CL (our implementation)
Curriculum Sorted
One-Pass Curriculum
Baby Steps Curriculum
46.4 (1.1)
46.83 (1.1)
47.42
45.74
47.37
43.88 (1.9)
42.88
43.09
46.07
Lee, 2008). We use the Stanford Sentiment Treebank (SST)
(Socher et al., 2013) an extension of a dataset (Pang and Lee,
2005) which has labels for 215,154 phrases in the parse trees
of 11,855 sentences sampled from movie reviews. Realvalued sentiment labels are converted to an integer ordinal label in 0,..,4 by simple thresholding for five classes:
very negative, negative, neutral, positive, and very positive.
Therefore the task is posed as a 5-class classification problem.
Dataset Details. We use the standard train/dev/test splits
of 8544/1101/2210 for the 5-class classification problem.
We flatten the annotated tree structure into sequences of
phrases to use finer grained annotations. We treat the words
within the span of an inner phrase as a sequence and use the
phrase’s annotation as label. This results in a bigger training
set of 155019 instances.
Experimental Details. We follow the previous work (Tai,
Socher, and Manning, 2015) for the empirical setup. We use
a single layer LSTM with 168 units for the 5-class classification task. We initialized the word embeddings using 300-
No-CL
Sorted
One-Pass
Baby-Step
--
-
Neutral
+
++
0.5
there
are
slow
and repetitive parts
,
but
it
has
just enough spice
to
keep
it interesting .
there
are
slow
and repetitive parts
,
but
it
has
just enough spice
to
keep
it interesting .
there
are
slow
and repetitive parts
,
but
it
has
just enough spice
to
keep
it interesting .
there
are
slow
and repetitive parts
,
but
it
has
just enough spice
to
keep
it interesting .
0.5
0.5
0.5
Figure 2: Predictions at intermediate tokens. Colors represent the polarity and heights represent the prediction probability. Baby
Step curriculum shows a consistent behavior : only after observing a positive sub-phrase flips the prediction.
0.50
0.45
Accuracy
0.40
Baby Step
One-Pass
Sorted
No-CL
0.35
0.30
0.25
0.20
3K
6K
12K
24K
49K
Number of Training Instances
155K
Figure 4: The effect of regimen vs the amount of training
data. One-Pass and Baby Steps curriculum regimens gets
better results when the training data is limited. They converge to similar points when the amount of data increases.
dimensional Glove vectors (Pennington, Socher, and Manning, 2014) and fine-tuned them during training. For optimization, we used RMSprop (Dauphin et al., 2015) with
learning rate 0.001 and decay rate of 0.9 with mini-batches
of size 128. The patience parameter p for early stopping is
10.
Results As the first step to our more detailed analysis, Table 2 reports the overall performance of the four learning
regimens and the original results stated by Tai, Socher, and
Manning (2015). The advantage of CL is most prominent
when predicting sentiment for sentences with conjunctions
(last column in Table 2). For conjunctions where a span of
text contradicts or supports overall sentiment polarity, Baby
Step model achieves significantly better results than others.
We take a closer look at the LSTM modeling process using
a similar probing technique used for the Digit Sum dataset.
Probing Intermediate Representations. In Figure 2, we
qualitatively show how different models process a sentence
with a contrastive conjunction originally demonstrated by
Socher et al. (2013). For each model, we plot the sentiment
polarity and the probability of prediction for that polarity
after observing a word token. Unlike the others, Baby Step
model changes the sentiment at the appropriate time; after
observing “spice” which constructs a positive statement with
the sub-phrase “but it has just enough spice”. Handling contrastive conjunctions requires a model to merge two conflicting signals (i.e. positive and negative) coming from two directions (i.e. left phrase and right phrase) in an accurate way
Socher et al. (2013). Considering LSTM’s limited capacity
due to using only signal coming from previous timesteps (i.e
processing the sentence from left to right), this result is particularly interesting because Baby Step CL boosts LSTM’s
performance.
Effect of Training Data Size. To investigate the role of the
amount of training data, we use a varying fraction of training data with learning regimens. Figure 4 shows the results.
CL regimens help when training data is limited. When the
amount of training data increases, the difference between
the regimens gets lower. This result suggests that in lowresource setups, like many of the NLP problems, CL could
be useful to improve a model’s performance.
Conclusion
We examined curriculum learning on two sequence prediction tasks. Our analyses showed that curriculum learning
regimens based on shorter-first approach, help LSTM construct a partial representation of the sequence in a more intuitive way. We demonstrated that curriculum learning helps
smaller models improve performance, contributes more in a
low resource setup. Using a quantitative and qualitative analysis on sentiment analysis, we showed that a model trained
with Baby Step curriculum significantly improves for sentences with conjunctions suggesting that curriculum learning helps LSTM learn longer sequences and functional role
of the conjunctions.
References
Bengio, Y.; Louradour, J.; Collobert, R.; and Weston, J. 2009. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, 41–48. ACM.
Bengio, Y.; Simard, P.; and Frasconi, P. 1994. Learning longterm dependencies with gradient descent is difficult. Neural Networks, IEEE Transactions on 5(2):157–166.
Bucilu, C.; Caruana, R.; and Niculescu-Mizil, A. 2006. Model
compression. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining,
535–541. ACM.
Cho, K.; van Merriënboer, B.; Bahdanau, D.; and Bengio, Y.
2014. On the properties of neural machine translation: Encoderdecoder approaches. arXiv preprint arXiv:1409.1259.
Chung, J.; Gulcehre, C.; Cho, K.; and Bengio, Y.
2015.
Gated feedback recurrent neural networks. arXiv preprint
arXiv:1502.02367.
Dauphin, Y. N.; de Vries, H.; Chung, J.; and Bengio, Y. 2015. Rmsprop and equilibrated adaptive learning rates for non-convex
optimization. arXiv preprint arXiv:1502.04390.
Dyer, C.; Ballesteros, M.; Ling, W.; Matthews, A.; and Smith,
N. A. 2015. Transition-based dependency parsing with stack
long short-term memory. arXiv preprint arXiv:1505.08075.
Elman, J. L. 1990. Finding structure in time. Cognitive science
14(2):179–211.
Elman, J. L. 1993. Learning and development in neural networks:
The importance of starting small. Cognition 48(1):71–99.
Fahlman, S. E. 1991. The recurrent cascade-correlation architecture. In Advances in Neural Information Processing Systems,
190–196.
Gal, Y. 2015. A theoretically grounded application of dropout in
recurrent neural networks. arXiv preprint arXiv:1512.05287.
Graves, A.; Liwicki, M.; Fernández, S.; Bertolami, R.; Bunke, H.;
and Schmidhuber, J. 2009. A novel connectionist system for
unconstrained handwriting recognition. Pattern Analysis and
Machine Intelligence, IEEE Transactions on 31(5):855–868.
Graves, A.; Wayne, G.; and Danihelka, I. 2014. Neural turing
machines. arXiv preprint arXiv:1410.5401.
Grefenstette, E.; Hermann, K. M.; Suleyman, M.; and Blunsom, P.
2015. Learning to transduce with unbounded memory. In Advances in Neural Information Processing Systems, 1819–1827.
Hinton, G.; Vinyals, O.; and Dean, J. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531.
Hochreiter, S., and Schmidhuber, J. 1997. Long short-term memory. Neural computation 9(8):1735–1780.
Irsoy, O., and Cardie, C. 2014. Deep recursive neural networks for
compositionality in language. In Advances in Neural Information Processing Systems, 2096–2104.
Iyyer, M.; Manjunatha, V.; Boyd-Graber, J.; and III, H. D. 2015.
Deep unordered composition rivals syntactic methods for text
classification. In Proceedings of the 53rd Annual Meeting of
the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing,
volume 1, 1681–1691.
Jiang, L.; Meng, D.; Zhao, Q.; Shan, S.; and Hauptmann, A. G.
2015. Self-paced curriculum learning. In AAAI, volume 2, 6.
Kalchbrenner, N.; Danihelka, I.; and Graves, A. 2015. Grid long
short-term memory. arXiv preprint arXiv:1507.01526.
Kalchbrenner, N.; Grefenstette, E.; and Blunsom, P. 2014. A
convolutional neural network for modelling sentences. arXiv
preprint arXiv:1404.2188.
Karpathy, A.; Johnson, J.; and Li, F.-F.
2015.
Visualizing and understanding recurrent networks. arXiv preprint
arXiv:1506.02078.
Kim, Y. 2014. Convolutional neural networks for sentence classification. arXiv preprint arXiv:1408.5882.
Kurach, K.; Andrychowicz, M.; and Sutskever, I. 2015. Neural
random-access machines. arXiv preprint arXiv:1511.06392.
Le, Q. V., and Mikolov, T. 2014. Distributed representations of
sentences and documents. arXiv preprint arXiv:1405.4053.
LeCun, Y.; Bottou, L.; Bengio, Y.; and Haffner, P. 1998. Gradientbased learning applied to document recognition. Proceedings of
the IEEE 86(11):2278–2324.
Li, J.; Chen, X.; Hovy, E.; and Jurafsky, D. 2016. Visualizing and
understanding neural models in nlp. In Proceedings of the 2016
Conference of the North American Chapter of the Association
for Computational Linguistics: Human Language Technologies.
Pang, B., and Lee, L. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales.
In Proceedings of the 43rd Annual Meeting on Association for
Computational Linguistics, 115–124. Association for Computational Linguistics.
Pang, B., and Lee, L. 2008. Opinion mining and sentiment analysis. Foundations and trends in information retrieval 2(1-2):1–
135.
Pascanu, R.; Mikolov, T.; and Bengio, Y. 2012. On the difficulty of training recurrent neural networks. arXiv preprint
arXiv:1211.5063.
Pennington, J.; Socher, R.; and Manning, C. D. 2014. Glove:
Global vectors for word representation. In EMNLP, volume 14,
1532–1543.
Pentina, A.; Sharmanska, V.; and Lampert, C. H. 2015. Curriculum
learning of multiple tasks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 5492–5500.
Socher, R.; Perelygin, A.; Wu, J. Y.; Chuang, J.; Manning, C. D.;
Ng, A. Y.; and Potts, C. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the conference on empirical methods in natural language
processing (EMNLP), volume 1631, 1642. Citeseer.
Spitkovsky, V. I.; Alshawi, H.; and Jurafsky, D. 2010. From baby
steps to leapfrog: How less is more in unsupervised dependency
parsing. In Human Language Technologies: The 2010 Annual
Conference of the North American Chapter of the Association
for Computational Linguistics, 751–759. Association for Computational Linguistics.
Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; and
Salakhutdinov, R. 2014. Dropout: A simple way to prevent neural networks from overfitting. The Journal of Machine Learning
Research 15(1):1929–1958.
Sukhbaatar, S.; Weston, J.; Fergus, R.; et al. 2015. End-to-end
memory networks. In Advances in Neural Information Processing Systems, 2431–2439.
Tai, K. S.; Socher, R.; and Manning, C. D. 2015. Improved semantic representations from tree-structured long short-term memory
networks. arXiv preprint arXiv:1503.00075.
Vinyals, O.; Kaiser, Ł.; Koo, T.; Petrov, S.; Sutskever, I.; and Hinton, G. 2015. Grammar as a foreign language. In Advances in
Neural Information Processing Systems, 2755–2763.
Vinyals, O.; Fortunato, M.; and Jaitly, N. 2015. Pointer networks.
In Advances in Neural Information Processing Systems, 2674–
2682.
Weston, J.; Bordes, A.; Chopra, S.; and Mikolov, T. 2015. Towards
ai-complete question answering: a set of prerequisite toy tasks.
arXiv preprint arXiv:1502.05698.
Yao, K.; Cohn, T.; Vylomova, K.; Duh, K.; and Dyer, C.
2015. Depth-gated recurrent neural networks. arXiv preprint
arXiv:1508.03790.
Zeiler, M. D., and Fergus, R. 2014. Visualizing and understanding convolutional networks. In Computer vision–ECCV 2014.
Springer. 818–833.
| 9cs.NE
|
ALMOST ENGEL COMPACT GROUPS
arXiv:1610.02079v2 [math.GR] 12 May 2017
E. I. KHUKHRO AND P. SHUMYATSKY
To Efim Zelmanov on occasion of his 60th birthday
Abstract. We say that a group G is almost Engel if for every g ∈ G there
is a finite set E (g) such that for every x ∈ G all sufficiently long commutators
[...[[x, g], g], . . . , g] belong to E (g), that is, for every x ∈ G there is a positive
integer n(x, g) such that [...[[x, g], g], . . . , g] ∈ E (g) if g is repeated at least n(x, g)
times. (Thus, Engel groups are precisely the almost Engel groups for which we
can choose E (g) = {1} for all g ∈ G.)
We prove that if a compact (Hausdorff) group G is almost Engel, then G has
a finite normal subgroup N such that G/N is locally nilpotent. If in addition
there is a uniform bound |E (g)| 6 m for the orders of the corresponding sets,
then the subgroup N can be chosen of order bounded in terms of m. The proofs
use the Wilson–Zelmanov theorem saying that Engel profinite groups are locally
nilpotent.
1. Introduction
A group G is called an Engel group if for every x, g ∈ G the equation [x, g, g, . . . , g] =
1 holds, where g is repeated in the commutator sufficiently many times depending
on x and g. (Throughout the paper, we use the left-normed simple commutator
notation [a1 , a2 , a3 , . . . , ar ] = [...[[a1 , a2 ], a3 ], . . . , ar ].) A group is said to be locally
nilpotent if every finite subset generates a nilpotent subgroup. Clearly, any locally
nilpotent group is an Engel group. Wilson and Zelmanov [15] proved the converse
for profinite groups: any Engel profinite group is locally nilpotent. Later Medvedev
[11] extended this result to Engel compact (Hausdorff) groups.
In this paper we consider almost Engel groups in the following precise sense.
Definition 1.1. We say that a group G is almost Engel if for every g ∈ G there
is a finite set E (g) such that for every x ∈ G all sufficiently long commutators
[x, g, g, . . . , g] belong to E (g), that is, for every x ∈ G there is a positive integer
2010 Mathematics Subject Classification. 20D25, 20E18, 20F45.
Key words and phrases. Compact groups; profinite groups; finite groups; Engel condition; locally
nilpotent groups.
1
n(x, g) such that
[x, g, g, . . . , g ] ∈ E (g)
| {z }
for all n > n(x, g).
n
Thus, Engel groups are precisely the almost Engel groups for which we can choose
E (g) = {1} for all g ∈ G.
We prove that almost Engel compact groups are finite-by-(locally nilpotent). By
a compact group we mean a compact Hausdorff topological group.
Theorem 1.2. Suppose that G is an almost Engel compact group. Then G has a
finite normal subgroup N such that G/N is locally nilpotent.
In Theorem 1.2 it also follows that there is a locally nilpotent subgroup of finite
index – just consider CG (N). The proof uses the aforementioned Wilson–Zelmanov
theorem for profinite groups. First the case of a finite group G is considered, where
obviously the result must be quantitative: namely, given a uniform bound |E (g)| 6
m for the cardinalities of the sets E (g) inTthe above definition, we prove that the
order of the nilpotent residual γ∞ (G) = i γi (G) is bounded in terms of m only.
Then Theorem 1.2 is proved for profinite groups. Finally, the result for compact
groups is derived with the use of the structure theorems for compact groups.
As in the case of finite groups, if there is a uniform bound m for the cardinalities
of the sets E (g) in Theorem 1.2, then the subgroup N in the conclusion can be
chosen to be of order bounded in terms of m (Corollary 5.5).
In an earlier paper [10] we obtained similar results about finite and profinite
groups with a stronger condition of “almost Engel” type. That condition means
that every element g of the group is “almost n-Engel” for some n = n(g) depending
on g. Note, however, that Engel groups do not necessarily satisfy that condition.
The new Definition 1.1 of almost Engel groups in the present paper imposes a weaker
and more natural “almost Engel” condition and includes Engel groups (when all the
subsets E (g) consist only of 1). Thus, the results of the present paper are stronger
than in [10] even for (pro)finite groups, and cover a wider class of compact groups.
First in § 2 we collect some elementary properties of minimal subsets E (g) in the
definition of almost Engel groups. We deal with finite groups in § 3, with profinite
groups in § 4, and consider the general case of compact groups in § 5.
Our notation and terminology is standard; for profinite groups, see, for example,
[14].
We say for short that an element g of a group G is an Engel element if for any
x ∈ G we have [x, g, g, . . . , g] = 1, where g is repeated in the commutator sufficiently
many times depending on x (such elements g are often called left Engel elements).
A subgroup (topologically) generated by a subset S is denoted by hSi. For a
group A acting by automorphisms on a group B we use the usual notation for
2
commutators [b, a] = b−1 ba and [B, A] = h[b, a] | b ∈ B, a ∈ Ai, and for centralizers
CB (A) = {b ∈ B | ba = b for all a ∈ A} and CA (B) = {a ∈ A | ba = b for all b ∈ B}.
Throughout the paper we shall write, say, “(a, b, . . . )-bounded” to abbreviate
“bounded above in terms of a, b, . . . only”.
2. Properties of Engel sinks
Throughout this section we assume that G is an almost Engel group in the sense
of Definition 1.1, so that for every g ∈ G there is a finite set E (g) such that for every
x ∈ G there is a positive integer n(x, g) such that
[x, g, g, . . . , g ] ∈ E (g)
| {z }
for any n > n(x, g).
(2.1)
n
′
If E (g) is another finite set with the same property for possibly different numbers n′ (x, g), then E (g) ∩ E ′ (g) also satisfies the same condition with the numbers
n′′ (x, g) = max{n(x, g), n′(x, g)}. Hence for every g ∈ G there is a minimal set
satisfying the definition, which we again denote by E (g) and call the Engel sink for
g, or simply g-sink for short. Henceforth we shall always use the notation E (g) to
denote the (minimal ) Engel sinks, and n(x, g) the corresponding numbers satisfying
(2.1).
For a fixed g ∈ G, consider the mapping of E (g) by the rule z → [z, g], which
maps E (g) into itself by definition. By the minimality of E (g) this mapping is a
permutation of E (g). Therefore we can speak of orbits (cycles) of this permutation
on E (g). It follows that every z ∈ E (g) can be represented in the form
z = [z, g, . . . , g ]
| {z }
for some k > 1,
(2.2)
for any positive integer j.
(2.3)
k
and therefore also as
z = [z, g, . . . , g ]
| {z }
jk
Conversely, elements satisfying (2.2) belong to E (g). We have thus proved the
following.
Lemma 2.1. For any g ∈ G the g-sink E (g) consists precisely of all elements z
such that z = [z, g, . . . , g], where g occurs at least once.
Clearly, every subgroup H of G is also an almost Engel group. Moreover, by
Lemma 2.1, for h ∈ H the h-sink constructed within H is precisely the subset
E (h) ∩ H of the h-sink E (h) in G. If N is a normal subgroup of G, then G/N is
also an almost Engel group. For ḡ = gN the ḡ-sink in G/N is the image of E (g)
in this quotient group. These properties will be used throughout the paper without
special references.
3
For any element h of the centralizer CG (g) the equation z = [z, g, . . . , g] implies
z = [z h , g h , . . . , g h ] = [z h , g, . . . , g]. Hence E (g) is invariant under conjugation by
h by Lemma 2.1. If |E (g)| = m, it follows that hm! centralizes E (g). We have thus
proved the following.
h
Lemma 2.2. If h ∈ CG (g) and |E (g)| = m, then hm! centralizes E (g).
Engel sinks have especially nice properties in metabelian groups. We denote by
M ′ the derived subgroup of a group M.
Lemma 2.3. Let M be a metabelian almost Engel group.
(a) Every sink E (g) is a normal subgroup contained in M ′ .
(b) Elements of E (g) in the same orbit under the map z → [z, g] have the same
order.
Proof. (a) By (2.2) and (2.3), every element z ∈ E (g) can be represented as z =
[z, g, . . . , g] with g repeated jk(z) times for k(z) > 1 and for every j = 1, 2, . . . . In
particular, E (g) ⊆ M ′ . For z1 , z2 ∈ E (g) choose j such that jk(z1 )k(z2 ) is larger
than n(z1 z2 , g). Then, by the standard metabelian laws,
z1 z2 = [z1 , g, . . . , g][z2 , g, . . . , g]
= [z1 z2 , g, . . . , g] ∈ E (g),
where g is repeated jk(z1 )k(z2 ) times in each commutator. Thus, the finite set
E (g) is a subgroup. For any m ∈ M, choose jk(z) larger than n([z, m], g). Then
[z, m] = [[z, g, . . . , g], m] = [[z, m], g, . . . , g] ∈ E (g), where g is repeated jk(z) times
in each commutator; this means that E (g) is a normal subgroup of M.
(b) Let z ∈ E (g). Since z ∈ M ′ , we have [z k , g] = [z, g]k for any integer k. Hence
the order of [z, g] divides the order of z. Going in this way over the orbit, we return
to z, which implies that the orders of all elements in the orbit are the same.
3. Finite almost Engel groups
Of course, any finite group G is almost Engel, and every element g ∈ G has finite
(minimal) g-sink E (g). A meaningful result must be of quantitative nature, and
this is what we prove in this section. The following theorem will also be used in the
proof of the main results on profinite and compact groups.
Theorem 3.1. Let G be a finite group, and m a positive integer. Suppose that for
every g ∈ G the cardinality of the g-sink E (g) is at most m. Then G has a normal
subgroup N of order bounded in terms of m such that G/N is nilpotent.
The conclusion of the theorem can also be stated as a bound in terms of m for the
order of the nilpotent residual subgroup γ∞ (G), the intersection of all terms of the
lower central series (which for a finite group is of course also equal to some subgroup
γn (G)).
4
First we recall or prove a few preliminary results. We shall use the following
well-known properties of coprime actions: if α is an automorphism of a finite group
G of coprime order, (|α|, |G|) = 1, then CG/N (α) = CG (α)N/N for any α-invariant
normal subgroup N, the equality [G, α] = [[G, α], α] holds, and if G is in addition
abelian, then G = [G, α] × CG (α).
Lemma 3.2. Let P be a finite p-subgroup of a group G, and g ∈ G a p′ -element
normalizing P . Then the order of [P, g] is bounded in terms of the cardinality of the
g-sink E (g).
Proof. For the abelian p-group V = [P, g]/[P, g]′ we have V = [V, g] and CV (g) = 1
because the action of g on V is coprime. Then V = {[v, g] | v ∈ V } and therefore
also
V = {[v, g, . . . , g ] | v ∈ V }
| {z }
n
for any n. Hence, V is contained in the image of E (g) ∩[P, g] in [P, g]/[P, g]′, whence
|V | 6 |E (g)|.
Since [P, g] is a nilpotent group, its order is bounded in terms of |[P, g]/[P, g]′|
and its nilpotency class. We claim that, as a crude bound, the nilpotency class of
[P, g] is at most 2|E (g)| + 1. Let γi denote the terms of the lower central series of
[P, g]. The number of factors of the lower central series of [P, g] on which g acts
nontrivially is at most |E (g)|, because for any such a factor U = γi /γi+1 we have
1 6= [U, g] = {[u, g, . . . , g ] | u ∈ U}
| {z }
n
for any n, since the action of g on U is coprime, and therefore there is an element of E (g) in γi \ γi+1 . It remains to observe that g cannot act trivially on
two consecutive nontrivial factors of the lower central series of [P, g]. Indeed, if
[γi , g] 6 γi+1 and [γi+1 , g] 6 γi+2 , then by the Three Subgroup Lemma the inclusions [γi , g, [P, g]] 6 [γi+1 , [P, g]] = γi+2 and [[P, g], γi, g] = [γi+1 , g] 6 γi+2 imply the
inclusion [g, [P, g], γi] = [[P, g], γi] = γi+1 6 γi+2 , and the last inclusion implies that
γi+1 = 1.
The following lemma already appeared in [10], but we reproduce the proof for the
benefit of the reader.
Lemma 3.3. Let V be an elementary abelian q-group, and U a q ′ -group of automorphisms of V . If |[V, u]| 6 m for every u ∈ U , then |[V, U]| is m-bounded, and
therefore |U| is also m-bounded.
Proof. First suppose that U is abelian. We consider V as an Fq U-module. Pick
u1 ∈ U such that [V, u1 ] 6= 0. By Maschke’s theorem, V = [V, u1 ] ⊕ CV (u1 ), and
both summands are U-invariant, since U is abelian. If CU ([V, u1 ]) = 1, then |U|
5
is m-bounded and [V, U] has m-bounded order being generated by [V, u], u ∈ U.
Otherwise pick 1 6= u2 ∈ CU ([V, u1 ]); then V = [V, u1 ] ⊕ [V, u2 ] ⊕ CV (hu1 , u2i). If
1 6= u3 ∈ CU ([V, u1 ] ⊕ [V, u2 ]), then V = [V, u1 ] ⊕ [V, u2 ] ⊕ [V, u3 ] ⊕ CV (hu1, u2 , u3 i),
and so on. If CU ([V, u1 ] ⊕ · · · ⊕ [V, uk ]) = 1 at some m-bounded step k, then
again [V, U] has m-bounded order. However, if there are too many steps, then
for the element w = u1 u2 · · · uk we shall have 0 6= [V, ui ] = [[V, ui ], w], so that
[V, w] = [V, u1 ] ⊕ · · · ⊕ [V, uk ] will have order greater than m, a contradiction.
We now consider the general case. Since every element u ∈ U acts faithfully on
[V, u], the exponent of U is m-bounded. If P is a Sylow p-subgroup of U, let M
be a maximal normal abelian subgroup of P . By the above, |[V, M]| is m-bounded.
Since M acts faithfully on [V, M], we obtain that |M| is m-bounded. Hence |P | is
m-bounded, since CP (M) = M and P/M embeds in the automorphism group of
M. Since |U| has only m-boundedly
many prime divisors, it follows that |U| is mP
bounded. Since [V, U] = u∈U [V, u], we obtain that |[V, U]| is also m-bounded.
Recall that the Fitting series starts with the Fitting subgroup F1 (G) = F (G),
and by induction, Fk+1 (G) is the inverse image of F (G/Fk (G)). If G is a soluble
group, then the least number h such that Fh (G) = G is the Fitting height of G. The
following lemma is well known and is easy to prove (see, for example, [9, Lemma 10]).
Q
Lemma 3.4. If G is a finite group of Fitting height 2, then γ∞ (G) = q [Fq , Gq′ ],
where Fq is a Sylow q-subgroup of F (G), and Gq′ is a Hall q ′ -subgroup of G.
We now approach the proof of Theorem 3.1 with the following lemma.
Lemma 3.5. If G is a finite group such that |E (g)| 6 m for all g ∈ G, then G/F (G)
has exponent at most m!.
Proof. Every element g ∈ G centralizes all its powers. Therefore by Lemma 2.2,
since |E (g m! )| 6 m by hypothesis, g m! centralizes E (g m! ). By the minimality of the
g m! -sink, then E (g m! ) = {1}. This means that g m! is an Engel element and therefore
belongs to the Fitting subgroup F (G) by Baer’s theorem [7, Satz III.6.15].
We are now ready to prove Theorem 3.1.
Proof of Theorem 3.1. Recall that G is a finite group such that |E (g)| 6 m for every
g ∈ G. We need to show that |γ∞ (G)| is m-bounded.
First suppose that G is soluble. Since G/F (G) has m-bounded exponent by
Lemma 3.5, the Fitting height of G is m-bounded, which follows from the Hall–
Higman theorems [5]. Hence we can use induction on the Fitting height, with trivial
base when the group is nilpotent and γ∞ (G) = 1. When the Fitting height is
at least 2, consider
the second Fitting subgroup F2 (G). By Lemma 3.4 we have
Q
γ∞ (F2 (G)) = q [Fq , Hq′ ], where Fq is a Sylow q-subgroup of F (G), and Hq′ is a
Hall q ′ -subgroup of F2 (G), the product taken over prime divisors of |F (G)|. For a
6
given q, let H̄q′ = Hq′ /CHq′ (Fq ), and let V be the Frattini quotient Fq /Φ(Fq ). Note
that H̄q′ acts faithfully on V , since the action is coprime [7, Satz III.3.18].
For every x ∈ H̄q′ the order |[V, x]| is m-bounded by Lemma 3.2. Then |H̄q′ | is
m-bounded by Lemma 3.3. As a result, |[Fq , Hq′ ]| = |[Fq , H̄q′ ]| is m-bounded, since
[Fq , H̄q′ ] is the product of m-boundedly many subgroups [Fq , h̄] for h ∈ Hq′ , each of
which has m-bounded order by Lemma 3.2.
For the same reasons, there are only m-boundedly many primes q for which
[Fq , Hq′ ] 6= 1. As a result, |γ∞ (F2 (G))| is m-bounded. Induction on the Fitting
height applied to G/γ∞ (F2 (G)) completes the proof in the case of soluble G.
Now consider the general case. Most of the following arguments follow the same
scheme as in the proof of Theorem 1.2 in [10]. First we show that the quotient
G/R(G) by the soluble radical is of m-bounded order. Let E be the socle of G/R(G).
It is known that E contains its centralizer in G/R(G), so it suffices to show that
E has m-bounded order. In the quotient by the soluble radical, E = S1 × · · · × Sk
is a direct product of non-abelian finite simple groups Si . Since the exponent of
G/F (G) is m-bounded by Lemma 3.5, the exponent of E is also m-bounded. Now
the classification of finite simple groups implies that every Si has m-bounded order,
and it remains to show that the number of factors is also m-bounded. By Shmidt’s
theorem [7, Satz III.5.1], every Si has a non-nilpotent soluble subgroup Ri , for which
γ∞ (Ri ) 6= 1. Since we already proved our theorem for soluble groups, we can apply
it to T = R1 × · · · × Rk . We obtain that |γ∞ (T )| is m-bounded, whence the number
of factors is m-bounded.
Thus, |G/R(G)| is m-bounded. Since |γ∞ (R(G)| is m-bounded by the soluble
case proved above, we can consider G/γ∞ (R(G)) and assume that R(G) = F (G)
is nilpotent. Then |G/F (G)| is m-bounded. We now use induction on |G/F (G)|.
The basis of this induction includes the trivial case G/F (G) = 1 when γ∞ (G) = 1.
But the bulk of the proof deals with the case where G/F (G) is a non-abelian simple
group.
Thus, suppose that G/F (G) is a non-abelian simple group of m-bounded order.
Let g ∈ G be an arbitrary element. The subgroup F (G)hgi is soluble, and therefore
|γ∞ (F (G)hgi)| is m-bounded by the above. Since γ∞ (F (G)hgi) is normal in F (G),
its normal closure hγ∞ (F (G)hgi)G i is a product of at most |G/F (G)| conjugates,
each normal in F (G), and therefore has m-bounded order. Choose a transversal
{t1 , . . . , tk } of G modulo F (G) and set
Y
K=
hγ∞ (F (G)hti i)G i,
i
which is a normal subgroup of G of m-bounded order. It is sufficient to obtain an
m-bounded estimate for |γ∞ (G/K)|. Hence we can assume that K = 1. We remark
that then
[F (G), g, . . . , g] = 1
for any g ∈ G,
(3.1)
7
when g is repeated sufficiently many times. Indeed, g ∈ F (G)ti for some ti , and the
subgroup F (G)hti i is nilpotent due to our assumption that K = 1.
We now claim that
[F (G), G, . . . , G] = 1
(3.2)
if G is repeated sufficiently many times. It is sufficient to prove that [Fq , G, . . . , G] =
1 for every Sylow q-subgroup Fq of F (G). For any q ′ -element h ∈ G we have
[Fq , h] = [Fq , h, h] and therefore [Fq , h] = 1 in view of (3.1). Let H be the subgroup
of G generated by all q ′ -elements. Then G = Fq H since G/F (G) is non-abelian
simple, and [Fq , H] = 1, so that
[Fq , G, . . . , G] = [Fq , Fq , . . . , Fq ] = 1
for a sufficiently long commutator.
We finally show that D := γ∞ (G) has m-bounded order. First we show that
D = [D, D]. Indeed, since G/F (G) is non-abelian simple, D is nonsoluble and we
must have
G = F (G)[D, D].
Taking repeatedly commutator with G on both sides and applying (3.2), we obtain
D = γ∞ (G) 6 [D, D], so D = [D, D].
Since F (G) ∩ D is hypercentral in D by (3.2) and [D, D] = D, it follows that
F (G) ∩ D 6 Z(D) ∩ [D, D] by the well-known Grün lemma [3, Satz 4]. Thus, D
is a central covering of the simple group D/(F (G) ∩ D) ∼
= G/F (G), and therefore
by Schur’s theorem [7, Hauptsatz V.23.5] the order of D is bounded in terms of the
m-bounded order of G/F (G). Thus, we have proved that |γ∞ (G)| is m-bounded in
the case where G/F (G) is a non-abelian simple group.
We now finish the proof of Theorem 3.1 by induction on the m-bounded order
k = |G/F (G)| proving that |γ∞ (G)| is (m, k)-bounded. The basis of this induction
is the case of G/F (G) being simple: nonabelian simple was considered above, and
simple of prime order is covered by the soluble case. Now suppose that G/F (G)
has a nontrivial proper normal subgroup with full inverse image N, so that F (G) <
N ⊳ G. Since F (N) = F (G), by induction applied to N, the order |γ∞ (N)| is
bounded in terms of m and |N/F (G)| < k. Since N/γ∞ (N) 6 F (G/γ∞ (N)), by
induction applied to G/γ∞ (N) the order |γ∞ (G/γ∞ (N))| is bounded in terms of m
and |G/N| < k. As a result, |γ∞ (G)| is (m, k)-bounded, as required.
4. Profinite almost Engel groups
In this and the next sections, unless stated otherwise, a subgroup of a topological
group will always mean a closed subgroup, all homomorphisms will be continuous,
and quotients will be by closed normal subgroups. This also applies to taking
commutator subgroups, normal closures, subgroups generated by subsets, etc. Of
8
course, any finite subgroup is automatically closed. We also say that a subgroup is
generated by a subset X if it is generated by X as a topological group.
In this section we prove Theorem 1.2 for profinite groups, while Corollary 5.5 for
profinite groups is an immediate corollary of Theorem 3.1.
Theorem 4.1. Suppose that G is an almost Engel profinite group. Then G has a
finite normal subgroup N such that G/N is locally nilpotent.
Recall that pro-(finite nilpotent) groups, that is, inverse limits of finite nilpotent
groups, are called pronilpotent groups.
Lemma 4.2. An almost Engel profinite group is pronilpotent if and only if it is
locally nilpotent.
Proof. Of course, any locally nilpotent profinite group is pronilpotent. Conversely,
suppose that G is an almost Engel pronilpotent group. We claim that all Engel
sinks are trivial: E (g) = {1} for every g ∈ G. Indeed, otherwise by Lemma 2.1 E (g)
contains a non-trivial element of the form z = [z, g, . . . , g] with g occurring at least
once. Choosing an open normal subgroup N with nilpotent quotient G/N such that
z 6∈ N, we obtain a contradiction. Thus, E (g) = {1} for every g ∈ G, which means
that all elements of G are Engel elements, that is, G is an Engel profinite group.
Then G is locally nilpotent by the Wilson–Zelmanov theorem [15, Theorem 5].
T
Recall that the pronilpotent residual of a profinite group G is γ∞ (G) = i γi (G),
where γi (G) are the terms of the lower central series; this is the smallest normal
subgroup with pronilpotent quotient. The following lemma is well known and is
easy to prove. Here, element orders are understood as Steinitz numbers. The same
results also hold in the special case of finite groups.
Lemma 4.3. (a) The pronilpotent residual γ∞ (G) of a profinite group G is equal to
the subgroup generated by all commutators [x, y], where x, y are elements of coprime
orders.
(b) For any normal subgroup N of a profinite group G we have γ∞ (G/N) =
γ∞ (G)N/N .
Proof. Part (a) follows from the characterization of pronilpotent groups as profinite
groups all of whose Sylow subgroups are normal. Part (b) follows from the fact that
for any elements x̄, ȳ of coprime orders in a quotient G/N of a profinite group G
one can find pre-images x, y ∈ G which also have coprime orders.
The following generalization of Hall’s criterion for nilpotency [4], which will be
used later, already appeared in [10], but we reproduce the proof for the benefit of
the reader. We denote the derived subgroup of a group B by B ′ .
Proposition 4.4. (a) Suppose that B is a normal subgroup of a group A such
that B is nilpotent of class c and γd (A/B ′ ) is finite of order k. Then the subgroup
9
C = CA (γd (A/B ′ )) = {a ∈ A | [γd (A), a] 6 B ′ } has finite k-bounded index and is
nilpotent of (c, d)-bounded class.
(b) Suppose that B is a normal subgroup of a profinite group A such that B is
pronilpotent and γ∞ (A/B ′ ) is finite. Then the subgroup D = CA (γ∞ (A/B ′ )) = {a ∈
A | [γ∞ (A), a] 6 B ′ } is open and pronilpotent.
Proof. (a) Since A/C embeds into Aut γd (A/B ′ ), the order of A/C is k-bounded. We
claim that C is nilpotent of (c, d)-bounded class. Indeed, using simple-commutator
notation for subgroups, we have
[C, . . . , C , C, C, . . . ] 6 [[γd (A), C], C, . . . ] 6 [[B, B], C, . . . ],
| {z }
d+1
since [γd (A), C] 6 B ′ by construction. Applying repeatedly the Three Subgroup
Lemma, we obtain
Y
[[B, B], C, . . . , C , C, . . . ] 6
[[B, C, . . . , C ], [B, C, . . . , C ], C, . . . ]
| {z }
| {z }
| {z }
2d−1
i+j=2d−1
i
j
6 [[[B, C, . . . , C ], B], C, . . . ]
| {z }
d
6 [[[B, B], B], C, . . . ].
Thus, γd+1 (C) 6 γ2 (B), then γ(d+1)+(2d−1) (C) 6 γ3 (B), then a similar calculation
gives γ(d+1)+(2d−1)+(3d−2) (C) 6 γ4 (B), and so on. An easy induction shows that
γ1+f (c,d) (C) 6 γc+1(B) = 1 for 1 + f (c, d) = 1 + dc(c + 1)/2 − c(c − 1)/2, so that C
is nilpotent of class f (c, d).
(b) As a centralizer of a normal section, D is a closed normal subgroup. Since
A/D embeds into Aut γ∞ (A/B ′ ), the subgroup D has finite index; thus, D is an
open subgroup. We now show that the image of D = CA (γ∞ (A/B ′ )) in any finite
quotient Ā of A is nilpotent. Let bars denote the images in Ā. Then γ∞ (Ā/B̄ ′ ) =
γ∞ (A/B ′ ) by Lemma 4.3(b). Therefore, D̄ 6 CĀ (γ∞ (Ā/B̄ ′ )). In a finite group,
γ∞ (Ā/B̄ ′ ) = γd (Ā/B̄ ′ ) for some positive integer d. Hence D̄ is nilpotent by part
(a).
In general, the set of Engel elements in a profinite group may not be closed. But in
an almost Engel group, Engel elements form a closed set, and moreover the following
holds.
Lemma 4.5. Let G be an almost Engel profinite group, and k a positive integer.
Then the set
Ek = {x ∈ G | |E (x)| 6 k}.
is closed in G.
10
Proof. We wish to show equivalently that the complement of Ek is an open subset
of G. Every element g ∈ (G \ Ek ) is characterized by the fact that |E (g)| > k + 1.
Let z1 , z2 , . . . , zk+1 be some k + 1 distinct elements in E (g). Using Lemma 2.1 we
can write for every i = 1, . . . , k + 1
zi = [zi , g, . . . , g],
where g is repeated ki > 1 times.
(4.1)
Let N be an open normal subgroup of G such that the images of z1 , z2 , . . . , zk+1 are
distinct elements in G/N. Then equations (4.1) show that for any u ∈ N the Engel
sink E (gu) contains an element in each of the k + 1 cosets zi N. Thus, all elements in
the coset gN are contained in G \ Ek . We have shown that every element of G \ Ek
has a neighbourhood that is also contained in G \ Ek , which is therefore an open
subset of G.
Recall that in Theorem 4.1 we need to show that an almost Engel profinite group
G has a finite normal subgroup such that the quotient is locally nilpotent. The first
step is to prove the existence of an open locally nilpotent subgroup.
Proposition 4.6. If G is an almost Engel profinite group, then it has an open
normal pronilpotent subgroup.
Of course, the subgroup in question will also be locally nilpotent by Lemma 4.2;
the result can also be stated as the openness of the largest normal pronilpotent
subgroup.
Proof. For every g ∈ G we choose an open normal subgroup Ng such that E (g)∩Ng =
1. Then g is an Engel element in Ng hgi. By Baer’s theorem [7, Satz III.6.15], in
every finite quotient of Ng hgi the image of g belongs to the Fitting subgroup. As a
result, the subgroup [Ng , g] is pronilpotent.
Let Ñg be the normal closure of [Ng , g] in G. Since [Ng , g] is normal in Ng , which
has finite index, [Ng , g] has only finitely many conjugates, so Ñg is a product of
finitely many normal subgroups of Ng , each of which is pronilpotent. Hence, so is
Ñg . Therefore all the subgroups Ñg are contained in the largest normal pronilpotent
subgroup K.
It is easy to see that G/K is an F C-group (that is, every conjugacy class is finite):
indeed, every ḡ ∈ G/K is centralized by the image of Ng , which has finite index in
G. A profinite F C-group has finite derived subgroup [13, Lemma 2.6]. Hence we
can choose an open subgroup of G/K that has trivial intersection with the finite
derived subgroup of G/K and therefore is abelian; let H be its full inverse image in
G. Thus, H is an open subgroup such that the derived subgroup H ′ is contained in
K.
We now consider the metabelian quotient M = H/K ′ , which is also an almost
Engel group, and temporarily use the symbols E (g) for the Engel g-sinks in M. For
11
every positive integer k, consider the set
Ek = {x ∈ M | |E (x)| 6 k}.
By Lemma 4.5, every set Ek is closed in M. Since M is an almost Engel group, we
have
[
M=
Ei .
i
By the Baire category theorem [8, Theorem 34], one of these sets contains an open
subset; that is, there is an open subgroup U and a coset aU such that aU ⊆ Em for
some m. In other words, |E (au)| 6 m for all u ∈ U.
We claim that |E (u)| 6 m2 for any u ∈ U. Indeed, by Lemma 2.3(a) both E (a)
and E (au) are normal subgroups of M contained in M ′ . In the quotient
M̄ = M/ E (a)E (au) ,
both M̄ ′ hāi and M̄ ′ hāūi are normal locally nilpotent subgroups. Hence their product, which contains ū, is also a locally nilpotent subgroup by the Hirsch–Plotkin
theorem [12, 12.1.2]. As a result, E (u) 6 E (a)E (au) and therefore |E (u)| 6
|E (a)| · |E (au)| 6 m2 .
Since M ′ 6 K/K ′ , it is easy to see that E (u) = E (uk) for any u ∈ U and any
k ∈ K/K ′ . Therefore, setting V = U(K/K ′ ) we have |E (v)| 6 m2 for any v ∈ V .
Thus, the Engel sinks of elements of V uniformly satisfy the inequality |E (v)| 6
2
m for all v ∈ V . The same inequality holds in every finite quotient V̄ of V , to which
we can therefore apply Theorem 3.1. As a result, |γ∞ (V̄ )| 6 n for some number
n = n(m) depending only on m. Then also |γ∞ (V )| 6 n.
Let W be the full inverse image of V , which is an open subgroup of G containing
K, and let Γ be the full inverse image of γ∞ (V ). Now let F = CW (γ∞ (V )) =
{w ∈ W | [Γ, w] 6 K ′ }. By Proposition 4.4(b), this is an open normal pronilpotent
subgroup, which completes the proof of Proposition 4.6.
Proof of Theorem 4.1. Recall that G is an almost Engel profinite group, and we need
to show that γ∞ (G) is finite. Henceforth we denote by F (L) the largest normal pronilpotent subgroup of a profinite group L. By Proposition 4.6 we already know that
G has an open normal pronilpotent subgroup, so that F (G) is also open. Further
arguments largely follow the scheme of proof of Theorem 1.1 in [10].
Since G/F (G) is finite, we can use induction on |G/F (G)|. The basis of this
induction includes the trivial case G/F (G) = 1 when γ∞ (G) = 1. But the bulk of
the proof deals with the case where G/F (G) is a finite simple group.
Thus, we assume that G/F (G) is a finite simple group (abelian or non-abelian).
Let p be a prime divisor of |G/F (G)|, and g ∈ G \ F (G) an element of order pn ,
where n is either a positive integer or ∞ (so pn is a Steinitz number). For any prime
q 6= p, the element g acts by conjugation on the Sylow q-subgroup Q of F (G) as an
automorphism of order dividing pn . The subgroup [Q, g] is a normal subgroup of
12
Q and therefore also a normal subgroup of F (G). The image of [Q, g] in any finite
quotient has order bounded in terms of |E (g)| by Lemma 3.2. It follows that [Q, g]
is finite of order bounded in terms of |E (g)|.
Since [Q, g] is normal in F (G), its normal closure h[Q, g]G i in G is a product of
finitely many conjugates and is therefore also finite. Let R be the product of these
closures h[Q, g]G i over all Sylow q-subgroups Q of F (G) for q 6= p. Since [Q, g] is
finite of order bounded in terms of |E (g)| as shown above, there are only finitely
many primes q such that [Q, g] 6= 1 for the Sylow q-subgroup Q of F (G). Therefore
R is finite, and it is sufficient to prove that γ∞ (G/R) is finite. Thus, we can assume
that R = 1. Note that then [Q, g a ] = 1 for any conjugate g a of g and any Sylow
q-subgroup Q of F (G) for q 6= p.
Choose a transversal {t1 , . . . , tk } of G modulo F (G). Let G1 = hg t1 , . . . , g tk i.
Clearly, G1 F (G)/F (G) is generated by the conjugacy class of the image of g. Since
G/F (G) is simple, we have G1 F (G) = G. By our assumption, the Cartesian product
T of all Sylow q-subgroups of F (G) for q 6= p is centralized by all elements g ti . Hence,
[G1 , T ] = 1. Let P be the Sylow p-subgroup of F (G) (possibly, trivial). Then also
[P G1 , T ] = 1, and therefore
γ∞ (G) = γ∞ (G1 F (G)) = γ∞ (P G1 ).
The image of γ∞ (P G1 ) ∩ T in G/P is contained both in the centre and in the
derived subgroup of P G1 /P and therefore is isomorphic to a subgroup of the Schur
multiplier of the finite group G/F (G). Since the Schur multiplier of a finite group
is finite [7, Hauptsatz V.23.5], we obtain that γ∞ (G) ∩ T = γ∞ (P G1 ) ∩ T is finite.
Therefore we can assume that T = 1, in other words, that F (G) is a p-group.
If |G/F (G)| = p, then G is a pro-p group, so it is pronilpotent, which means that
γ∞ (G) = 1 and the proof is complete. If G/F (G) is a non-abelian simple group, then
we choose another prime r 6= p dividing |G/F (G)| and repeat the same arguments
as above with r in place of p. As a result, we reduce the proof to the case F (G) = 1,
where the result is obvious.
We now finish the proof of Theorem 4.1 by induction on |G/F (G)|. The basis
of this induction where G/F (G) is a simple group was proved above. Now suppose
that G/F (G) has a nontrivial proper normal subgroup with full inverse image N,
so that F (G) < N ⊳ G. Since F (N) = F (G), by induction applied to N the group
γ∞ (N) is finite. Since N/γ∞ (N) 6 F (G/γ∞ (N)), by induction applied to G/γ∞ (N)
the group γ∞ (G/γ∞ (N)) is also finite. As a result, γ∞ (G) is finite, as required.
5. Compact almost Engel groups
In this section we prove the main Theorem 1.2 about compact almost Engel
groups. We use the structure theorems for compact groups and the results of the
preceding section on profinite almost Engel groups.
13
Recall that a group H is said to be divisible if for every h ∈ H and every positive
integer k there is an element x ∈ H such that xk = h.
Proposition 5.1. If H is an almost Engel divisible group, then H is in fact an
Engel group.
Proof. We need to show that E (h) = {1} for every h ∈ H. Let |E (h)| = k. Let
g ∈ H be an element such that g k! = h. Since g centralizes h, by Lemma 2.2
we obtain that h = g k! centralizes E (h). By the minimality of the h-sink, then
E (h) = {1}, as required.
We are ready to prove Theorem 1.2.
Proof of Theorem 1.2. Let G be an almost Engel compact group; we need to show
that there is a finite subgroup N such that G/N is locally nilpotent. By the wellknown structure theorems (see, for example, [6, Theorems 9.24 and 9.35]), the connected component of the identity G0 in G is a divisible group such that G0 /Z(G0)
is a Cartesian product of simple compact Lie groups, while the quotient G/G0 is
a profinite group. By Proposition 5.1, G0 is an Engel group. Compact Lie groups
are linear groups, and linear Engel groups are locally nilpotent by the Garashchuk–
Suprunenko theorem [1] (see also [2]). Hence G0 = Z(G0 ) is an abelian subgroup.
Remark 5.2. If a compact group G is an Engel group, then by the above G is an extension of an abelian subgroup G0 by a profinite group G/G0 . Being an Engel group,
G/G0 is locally nilpotent by the Wilson–Zelmanov theorem [15, Theorem 5]. It is
known that an Engel abelian-by-(locally nilpotent) group is locally nilpotent (see,
for example, [12, 12.3.3]). This gives an alternative proof of Medvedev’s theorem
[11].
We proceed with the proof of Theorem 1.2.
Lemma 5.3. For every g ∈ G we have E (g) ∩ G0 = 1, which is equivalent to the
fact that for any x ∈ G0 we have
[x, g, . . . , g] = 1
if g is repeated sufficiently many times.
Proof. Suppose the opposite and choose 1 6= z ∈ E (g) ∩ G0 . By Lemma 2.1 then
z = [z, g, . . . , g] with g occurring at least once. Hence z belongs also to the gsink within the semidirect product G0 hgi. Since this subgroup is metabelian, by
Lemma 2.3(a) the g-sink in G0 hgi is a subgroup, which is equal to E (g) ∩ G0 .
Therefore we can choose z1 in E (g) ∩ G0 of some prime order p = |z1 |. Again by
Lemma 2.1 we have z1 = [z1 , g, . . . , g] with g occurring at least once. In the divisible
k
group G0 for every k = 1, 2, . . . there is an element zk such that zkp = z1 . We
have yk = [zk , g, . . . , g] ∈ E (g) ∩ G0 when g is repeated sufficiently many times.
14
k
k
Then ykp = [zkp , g, . . . , g] = [z1 , g, . . . , g], which is an element of the orbit of z1 in
E (g)∩G0 under the mapping x → [x, g] and therefore has the same order p = |z1 | by
Lemma 2.3(b). Thus, yk is an element of E (g) of order exactly pk+1 , for k = 1, 2, . . . .
As a result, E (g) is infinite, a contradiction with the hypothesis.
Applying Theorem 4.1 to the almost Engel profinite group Ḡ = G/G0 we obtain a
finite normal subgroup D with locally nilpotent quotient. Then D contains all Engel
sinks E¯(g) of elements g ∈ Ḡ and therefore the subgroup E generated by them:
E := hE¯(g) | g ∈ Ḡi 6 D.
Clearly, E¯(g)h = E¯(g h ) for any h ∈ Ḡ; hence E is a normal finite subgroup of Ḡ.
We replace D by E in the sense that Ḡ/E is also locally nilpotent by the Wilson–
Zelmanov theorem [15, Theorem 5], since this is an Engel profinite group.
We now consider the action of Ḡ by automorphisms on G0 induced by conjugation.
Lemma 5.4. The subgroup E acts trivially on G0 .
Q
Proof. The abelian divisible group G0 is a direct product A0 × p Ap of a torsion-free
divisible group A0 and Sylow subgroups Ap over various primes p. Clearly, every
Sylow subgroup is normal in G.
First we show that E acts trivially on each Ap . It is sufficient to show that for
every g ∈ Ḡ every element z ∈ E¯(g) acts trivially on Ap . Consider the action of hz, gi
on Ap . Note that hz, gi = hz hgi ihgi, where hz hgi i is a finite g-invariant subgroup, since
it is contained in the finite subgroup E. For any a ∈ Ap the subgroup
hahgi i = ha, [a, g], [a, g, g], . . . i
is a finite p-group by Lemma 5.3, and this subgroup is g-invariant. Its images under
the action of elements of the finite group hz hgi i generate a finite p-group B, which
is hz, gi-invariant. Lemma 5.3 implies that the image of hz, gi in its action on B
must be a p-group. Indeed, otherwise this image contains a p′ element y that acts
non-trivially on the Frattini quotient V = B/Φ(B). Then V = [V, y] and CV (y) = 1,
whence [V, y] = {[v, g] | v ∈ [V, y]} and therefore also [V, y] = {[v, y, . . . , y] | v ∈
[V, y]} with y repeated n times, for any n, contrary to Lemma 5.3. But since z is an
element of E¯(g), by Lemma 2.1 we have z = [z, g, . . . , g] with at least one occurrence
of g. Since a finite p-group is nilpotent, this implies that the image of z in its action
on B must be trivial. In particular, z centralizes a. As a result E acts trivially on
Ap , for every prime p.
Q
We now show that E also acts trivially on the quotient W = G0 / p Ap of G0 by its
torsion part. Note that W can be regarded as a vector space over Q. Every element
g ∈ E has finite order and therefore by Maschke’s theorem W = [W, g] × CW (g). If
[W, g] 6= 1, then [W, g] = {[w, g, . . . , g] | w ∈ [W, g]} with g repeated n times, for
any n. This contradicts Lemma 5.3.
15
Q
Thus, E acts trivially both on W and on p Ap . Then any automorphism of G0
induced by conjugation by g ∈ E acts on every element a ∈ A0 as ag = at, where
i
t = t(a) is an element of finite order in G0 . Since ag = ati , the order of t must
divide the order of g. Assuming the action of E on G0 to be non-trivial, choose an
element g ∈ E acting on G0 as an automorphism of some prime order p. Then there
is a ∈ A0 such that ag = at, where t ∈ G0 has order p. For any k = 1, 2, . . . there
k
k
is an element ak ∈ A0 such that apk = a. Then agk = ak tk , where tpk = t. Thus
|tk | = pk+1, and therefore pk+1 divides the order of g, for every k = 1, 2, . . . . We
arrived at a contradiction since g has finite order.
Let F be the full inverse image of E in G. Recall that we have normal subgroups
G0 6 F 6 G such that G0 is divisible, F/G0 is finite, and G/F is locally nilpotent.
We aim at producing a finite normal subgroup N such that G/N is locally nilpotent.
By Lemma 5.4 the subgroup G0 is contained in the centre of the full inverse
image F of E, so that F has centre of finite index. Then the derived subgroup F ′
is finite by Schur’s theorem [7, Satz IV.2.3]. We can assume that F ′ = 1, so that
then F is abelian. In the abstract abelian group F the divisible subgroup G0 has a
complement C, which is obviously finite, F = G0 × C. Let M be the normal closure
of C in G.
Note that G/M is a locally nilpotent group. Indeed, G/F is locally nilpotent,
while F/M is G-isomorphic to G0 /(G0 ∩ M) so that G/M is an Engel group by
Lemma 5.3. Being an abelian-by-(locally nilpotent) Engel group, then G/M is
locally nilpotent.
Consider the natural semidirect product M ⋊(G/CG (M)) with the induced topology, in which the action of G/CG (M) on M is continuous. Both M and G/CG (M)
are profinite groups; therefore M ⋊ (G/CG (M)) is also a profinite group (see [14,
Lemma 1.3.6]). Note that G/CG (M) is locally nilpotent, since M 6 CG (M).
The group M ⋊ (G/CG (M)) is also almost Engel, so by Theorem 4.1 it contains a
finite normal subgroup with locally nilpotent quotient. This finite subgroup therefore contains the subgroup K generated by all Engel sinks in M ⋊ (G/CG (M));
since G/CG (M) is locally nilpotent, K 6 M. Since G/M is also locally nilpotent,
the Engel sinks in G are all contained in M and coincide with the Engel sinks in
M ⋊ (G/CG (M)). Renaming K by N as a subgroup of G we arrive at the required
result. Indeed, N is a normal subgroup because E (g)h = E (g h ) for any h ∈ G. The
group G/N is locally nilpotent being an Engel group which is an extension of an
abelian group F/N by a locally nilpotent group G/F . The proof of Theorem 1.2 is
complete.
Corollary 5.5. Let G be an almost Engel compact group such that for some positive
integer m all Engel sinks E (g) have cardinality at most m. Then G has a finite
normal subgroup N of order bounded in terms of m such that G/N is locally nilpotent.
16
Proof. By Theorem 1.2 the group G is finite-by-(locally nilpotent). Therefore every
abstract finitely generated subgroup H of G is finite-by-nilpotent and residually
finite. By Theorem 3.1, every finite quotient of H has nilpotent residual of mbounded order. Hence γ∞ (H) is also finite of m-bounded order, and H/γ∞ (H) is
nilpotent. Thus, every finitely generated subgroup of G has a normal subgroup of
m-bounded order with nilpotent quotient. By the standard inverse limit argument,
the group G has a normal subgroup of m-bounded order with locally nilpotent
quotient.
Acknowledgements
The authors thank Gunnar Traustason and John Wilson for stimulating discussions.
The first author was supported by the Russian Science Foundation, project no.
14-21-00065, and the second by the Conselho Nacional de Desenvolvimento Cientı́fico
e Tecnológico (CNPq), Brazil. The first author thanks CNPq-Brazil and the University of Brasilia for support and hospitality that he enjoyed during his visits to
Brasilia.
References
[1] M. S. Garashchuk and D. A. Suprunenko, Linear nilgroups, Dokl. Akad. Nauk BSSR 4 (1960),
407–408. (Russian)
[2] K. W. Gruenberg, The Engel structure of linear groups, J. Algebra 3 (1966), 291–303.
[3] O. Grün, Beitrge zur Gruppentheorie. I, J. Reine Angew. Math. 174 (1935), 1–14.
[4] P. Hall, Some sufficient conditions for a group to be nilpotent, Illinois J. Math. 2 (1958),
787–801.
[5] P. Hall and G. Higman, The p-length of a p-soluble group and reduction theorems for Burnside’s problem, Proc. London Math. Soc. (3) 6 (1956), 1–42.
[6] K. H. Hofmann and S. A. Morris, The structure of compact groups, De Gruyter, Berlin, 2006.
[7] B. Huppert, Endliche Gruppen. I (Springer, Berlin, 1967).
[8] J. L. Kelley, General topology, Grad. Texts in Math., vol. 27, Springer, New York, 1975.
[9] E. I. Khukhro and V. D. Mazurov, Finite groups with an automorphism of prime order whose
centralizer has small rank, J. Algebra 301 (2006) 474–492.
[10] E. I. Khukhro and P. Shumyatsky, Almost Engel finite and profinite groups, Int. J. Algebra
Comput., 26, no. 5 (2016), 973–983.
[11] Yu. Medvedev, On compact Engel groups, Israel J. Math. 185 (2003), 147–156.
[12] D. J. S. Robinson, A course in the theory of groups, Springer, New York, 1996.
[13] A. Shalev, Profinite Groups with Restricted Centralizers, Proc. Amer. Math. Soc. 122, no. 4
(1994), 1279–1284.
[14] J. S. Wilson, Profinite groups (Clarendon Press, Oxford, 1998).
[15] J. S. Wilson and E. I. Zelmanov, Identities for Lie algebras of pro-p groups, J. Pure Appl.
Algebra 81, no. 1 (1992), 103–109.
17
Sobolev Institute of Mathematics, Novosibirsk, 630090, Russia, and
University of Lincoln, U.K.
E-mail address: [email protected]
Department of Mathematics, University of Brasilia, DF 70910-900, Brazil
E-mail address: [email protected]
18
| 4math.GR
|
arXiv:1801.09117v1 [math.GR] 27 Jan 2018
Centralizer fusion systems of central
involutions in a finite group with soluble
centralizer of involutions
KIVANÇ ERSOY and İPEK TUVAY
Abstract. It is a long-standing open problem raised by Starostin
to describe all finite groups with soluble centralizers of involutions.
One can observe that if the centralizer fusion system of an involution is nilpotent, then the centralizer of that involution is soluble.
In this paper, we classify the cases when the centralizer fusion system of a central involution in a finite group whose all involutions
have soluble centralizers is a nilpotent fusion system. Indeed, we
analyse the case when the solvable radical has odd order and the
corresponding factor group is simple.
2010 Mathematics Subject Classification: 20D05, 20D20
Keywords: centralizer of involution, fusion system, finite simple
group.
1. Introduction
Since Brauer-Fowler Theorem indicates that there are finitely many
simple groups with a given structure of centralizer of an involution, the
study of such proper subgroups has been an active research area on the
way to the classification of finite simple groups. Starostin raised the
following problem:
Problem 1.1. [6, Problem 2.74] Describe the finite groups all of
whose involutions have soluble centralizers.
1
2
KIVANÇ ERSOY AND İPEK TUVAY
Here one should mention the famous N-group papers of Thompson
(see [11, 12, 13, 14, 15, 16]) where finite simple groups in which normalizers of p-subgroups are soluble, are classified. Clearly, centralizer
of an involution is exactly the normalizer of the subgroup generated by
that involution. Hence, finite simple groups in Thompson’s list occur as
natural examples of finite simple groups with soluble centralizers of involutions. Indeed, Gorenstein and Lyons classified finite groups whose
all 2-local subgroups are soluble (see [5, Main Theorem]). By the classification of finite simple groups, one can list finite simple groups all
of whose involutions have soluble centralizers (See the list in Proposition 3.1). Moreover, Syskin proved the following theorem for finite
non-soluble groups in which all involutions of the Sylow 2-subgroup P
are contained in center of P .
Theorem 1.2. [9, Theorem A] Let G be a finite non-soluble group
with soluble centralizers of involutions and P a Sylow 2-subgroup of G.
Assume that the subgroup generated by all involutions of P is contained
in the center of P . Then G/O2′ (G) has a normal subgroup isomorphic
to one of the following groups:
(1) P SL2 (q) where q = 3, 5 mod 8
(2) SL2 (2n ) where n ≥ 2
(3) P SU3 (2n ) where n ≥ 2
(4) 2 B2 (22m+1 ) where m ≥ 1.
The theory of fusion systems which was originated from the work
of Burnside and Frobenius, was formulated axiomatically by Puig (for
details see [1, 2, 7]). Let G be a finite group and P be a Sylow
p-subgroup of G. The fusion system FP (G) is the category whose
objects are subgroups of P and whose morphisms are induced by the
conjugation action by elements of G on the subgroups of P . In this
paper fusion systems are used as an invariant of the groups with soluble
centralizers of involutions having odd order soluble radicals (see Lemma
2.2).
CENTRALIZER FUSION SYSTEMS OF CENTRAL INVOLUTIONS
3
In this paper we investigate a special type of finite groups with
soluble centralizers of involutions in which the soluble radical has odd
order. We ask the following question:
Question 1.3. Let G be a finite group whose all involutions have
soluble centralizers and R its soluble radical. If R has odd order and
G/R is simple, is the centralizer of a central involution of G necessarily
an extension of an odd order group by a 2-group?
We use knowledge about fusion systems to classify the cases when
this property holds. Indeed we prove that this result just depend on
the structure of G/R. The main result of the paper is the following:
Theorem 1.4. Let G be a finite non-soluble group with soluble centralizers of involutions. Assume that the soluble radical R of G has odd
order and G/R is simple. Let x be a central involution of G. Then the
following holds.
(1) CF (x) is nilpotent iff G/R is isomorphic to one of
A5 , A6 , A7 , P SL2 (q),2 B2 (2m ) where m is odd, M11 or 2 F4 (2)′ .
(2) CF (x) is non-nilpotent iff G/R is isomorphic to one of
A8 , P SL3(2n ), P SL3 (3) or P SU3 (2n ).
Therefore, whenever the centralizer fusion system is nilpotent, the
centralizer of the central involution is isomorphic to an odd order group
extended by a 2-group.
2. Preliminaries
In this section we collect definitions and results that will be used
in our proofs.
Definition 2.1. An involution x of a finite simple group G is called
a central involution if CG (x) contains a Sylow 2-subgroup of G.
Likewise for groups, there is a notion of centralizer fusion system
in general. To avoid technicality, we will define it here for just central
4
KIVANÇ ERSOY AND İPEK TUVAY
involutions, indeed by using [2, Theorem 4.27], for a central involution
x ∈ G, the centralizer fusion system CF (x) of x in F = FP (G) can be
viewed as the fusion system FP (CG (x)).
We use the following result to see that the 2-fusion system of a finite
group with a soluble radical of odd order is isomorphic to the 2-fusion
system of the quotient.
Lemma 2.2. [7, Theorem 2.12] Let G be a finite group, P be a Sylow
2-subgroup of G. If N is a normal 2′ -subgroup of G then
FP (G) ∼
= FP N/N (G/N).
The following result characterizes nilpotent fusion systems. Recall that the Sylow p-subgroup P has a normal p-complement in G iff
Op′ (G) ⋊ P ∼
= G.
Theorem
2.3. [2,
Theorem
1.12] (Frobenius Normal p-
Complement Theorem) Let G be a finite group and P be a Sylow psubgroup of G. Then FP (G) = FP (P ) iff P has a normal p-complement
in G.
Let G be a finite group and x ∈ G be an involution in G. Let P be
a Sylow 2-subgroup of G containing x.
Proposition 2.4. Let G be a finite group, P be a Sylow 2-subgroup
of G and x be a central involution in P . Let F = FP (G). If CF (x) is
a nilpotent fusion system, then CG (x) is soluble.
Proof. Observe that if CF (x) = FP (CG (x)) is nilpotent, then by
Theorem 2.3, CG (x) has a normal subgroup K of odd order where
[CG (x) : K] is a 2-power. Indeed, CG (x) = P K. Now, K is soluble by
Odd Order Theorem, and P is a 2-group, hence CG (x) is soluble.
3. Main Results
We first need to give a list of finite simple groups with soluble
centralizers of involutions.
CENTRALIZER FUSION SYSTEMS OF CENTRAL INVOLUTIONS
5
Proposition 3.1. Let S be a finite non-abelian simple group with
soluble centralizer of involutions. Then S is isomorphic to one of the
following groups:
(1) A5 , A6 , A7 , A8 ;
(2) P SL2 (q) where q ≥ 4 is a prime power;
(3) 2 B2 (22n+1 ), n ∈ N;
(4) P SL3 (2n ) n ∈ N
(5) P SL3 (3);
(6) P SU3 (3);
(7) P SU3 (2n ), n ∈ N
(8) Matthieu group M11
(9) Tits group 2 F4 (2)′
Proof. Let G be an alternating group An . Take x = (12)(34).
∼ A8 .
One can easily observe that CG (x) is non-soluble if n ≥ 9. Let G =
There are two conjugacy class of involutions, one of which is a product
of two transpositions and the other is a product of four transpositions.
It is easy to check that in any case the centralizers are soluble. Since
A5 , A6 and A7 are subgroups of A8 , they also have soluble centralizers
of involutions.
Let G be a simple group of Lie type. Then there is a simple linear
algebraic group G of adjoint type over the algebraic closure of Fp , a
′
Frobenius map σ of G such that G = O p (Gσ ). Clearly, CG (x) =
′
O p (CG (x, σ)).
First assume that p 6= 2. Then involutions are semisimple elements
of G. Assume that CG (x) is soluble. Now, by [3, Theorem D], centralizer of an involution in G over the algebraic closure of Fq involves
a simple linear algebraic group whenever 2 ≤ Σmi , namely the highest
root. Then the sum of coefficients of the highest root is 2 which means
G is of type P SL2 (q).
Now, assume CG (x) is non-soluble where CG (x) is soluble. This
means there is P SL2 (p), which is soluble, which means p = 3. Hence,
the groups are either P SL2 (q) or P SL3 (3).
6
KIVANÇ ERSOY AND İPEK TUVAY
Secondly assume p = 2. Then involutions are unipotent elements.
If rank(G) > 2, take an involution u belonging to the root subgroup
hxr (t) : t ∈ F2 i for the positive root r. Since rank(G) > 2, there is a
positive root s such that r + s nor r − s is a not root, (indeed, if you
choose r a fundamental root, there is another fundamental root s such
that r+s is not a root, r−s is not a root in any case.) Therefore, there is
a subgroup hxs (t1 ), x−s (t2 ) : t1 , t2 ∈ F2 i, which is clearly non-soluble.
Now, assume G ∼
= C2 (2n ), where the root system of C2 is generated by
r, s with highest root 2r+s. Since 2r+2s and 2r are not roots in the root
system C2 , by Chevalley Commutator Formula, the highest root commutes with hxs (t1 ), x−s (t2 ) : t1 , t2 ∈ F2n i, which is non-soluble. Hence,
the rest only follows from the list, indeed G is one of the groups listed
′
in [5, Main Theorem], namely 2 B2 (22m+1 ), P SU3(2m ),2 F4 (2) .Similarly,
Tits group 2 F4 (2)′ is another simple group which is listed in [5, Main
Theorem].
Consider the case when G is one of the sporadic simple groups. One
can observe that M11 has a semidihedral Sylow 2-subgroup P of order
16 and Z(P ) has order 2. Moreover, any involution in P has soluble
centralizer, so M11 is the only sporadic group belonging to our list. By
[5, Main Theorem], the list is complete.
Theorem 3.2. Let G be a finite group with soluble centralizers of
involutions. Suppose the soluble radical R of G has odd order. Then
S ≤ G/R ֒→ Aut(S)
where S is the unique minimal normal subgroup of G/R.
Proof. Let G be a finite group with soluble centralizers of involutions and a soluble radical R of odd order. Then CG/R (xR) =
CG (x)R/R for any involution x ∈ G, hence G/R is a finite group with
soluble centralizers of involutions and trivial soluble radical. Take a
minimal normal subgroup M of G/R. Being a non-soluble minimal
normal subgroup, M ∼
= Σt S where S is a fixed finite non-abelian
i=1
simple group. Since any involution in S × {1} × . . . × {1} ≤ M has
CENTRALIZER FUSION SYSTEMS OF CENTRAL INVOLUTIONS
7
soluble centralizers of involutions, one has t = 1. Therefore, G/R has
a minimal normal subgroup isomorphic to S where S is a finite simple
group with soluble centralizers of involutions, namely one of the groups
listed in Proposition 3.1.
Now, assume that N is another minimal normal subgroup of G/R.
Since S ∩ N must be trivial, N ≤ CG (S) ≤ CG (x) for any involution
x ∈ S. Hence, N must be soluble. But G/R has no soluble normal
subgroups, hence S is the unique minimal normal subgroup of G/R.
In particular, CG (S) is a normal subgroup of G which commute
elementwise by S, so CG (S) has odd order, that is CG (S) = 1. Now,
G/CG/R (S) ∼
= G/R embeds naturally in AutS, while S embeds in the
former trivially.
For fusion systems, we will concentrate on the groups G where
G/R ∼
= S. Recall that a fusion system F defined on a p-group P is
called nilpotent if F = FP (P ).
First, we need to show that for any central involution x in P SL2 (q),
the centralizer fusion system CF (x) is nilpotent.
Proposition 3.3. Every subgroup of odd order in the dihedral
group D2n is normal.
Proof. Consider the dihedral group D2n = hr, ρ : r n = ρ2 =
1 and rρ = ρr n−1 i. Clearly, (ρr k )2 = 1 for any k. Hence, if y ∈ D2n
is not an involution, then y ∈ H = hri, which is normal in D2n . But
since H is cyclic, any subgroup is characteristic in H, hence normal in
D2n . Therefore, any subgroup of odd order is normal in D2n .
Lemma 3.4. Let G ∼
= P SL2 (q) where q = pn , let P be a Sylow
2-subgroup of G, F = FP (G) and x is a central involution in P . Then
CF (x) is nilpotent.
Proof. Since x was assumed to be a central involution, by definition, CF (x) = FP (CG (x)).
First assume that q is even. In this case, a Sylow 2-subgroup P of G
is an elementary abelian 2-group conjugate to unitriangular subgroup
8
KIVANÇ ERSOY AND İPEK TUVAY
of P SL2 (q) and CG (x) = P for any x ∈ P . Hence, CF (x) = FP (P ),
hence it is nilpotent.
Now assume that q is odd. Then P consists of semisimple elements.
Take the central involution x ∈ P . Since x is semisimple, it is contained
in a maximal torus T of G, which has order either
q−1
2
or
q+1
.
2
Moreover,
CG (x) is a dihedral group which contains T as a normal subgroup,
that is CG (x) ∼
= Dq−1 or Dq+1 , dihedral groups of order q − 1 or q + 1
respectively. In any case, by Proposition 3.3, Hall 2′ -subgroup of CG (x)
is normal. By Theorem 2.3, FP (CG (x)) = FP (P ) iff for any CG (x) has
a normal 2-complement. Hence, FP (CG (x)) is nilpotent.
Theorem 3.5. Let G be a simple group of Lie type over a finite
field of characteristic 2. Let P be a Sylow 2-subgroup of G, F = FP (G)
and x ∈ Z(P ) be a central involution. Then CF (x) is nilpotent iff
G∼
= SL2 (2n ),2 B2 (22m+1 ) or P SL(3, 2) and not nilpotent in P SL3 (2n )
and P SU3 (2n ).
Proof. Assume that CF (x) is nilpotent, that is, equal to FP (P ).
Then it is also a soluble fusion system by [2, Corollary 5.96].
Then, CG (x) must be soluble, since P has a normal 2-complement
in CG (x) by [2, Theorem 1.12].
Then G must be isomorphic to
P SL2 (2n ),2 B2 (22m+1 )P SL3 (2n ) or P SU3 (2n ).
G = P SL2 (2n ) ∼
= SL2 (2n ) has an abelian Sylow 2-subgroup P with
CG (x) = P where x is a central involution. Hence the centralizer fusion
system is nilpotent.
Now, assume that G ∼
= P SL3 (2n ). The centralizer of the central
involution x contains the Sylow 2-subgroup as a normal subgroup. Assume the 2′ -subgroup of CG (x) is also normal. Then CG (x) becomes a
direct product of a 2-group and a group of odd order. But
a b c
CG (x) = { 0 e f Z : a, b, c, e, f ∈ F2n }
0 0 a
CENTRALIZER FUSION SYSTEMS OF CENTRAL INVOLUTIONS
a 0 0
9
does not normalize T = { 0 e 0 Z : a, e ∈ F2n }, hence the
0 0 a
centralizer fusion system of the central involution is not nilpotent.
Let G ∼
= P SU3 (22n ), let P be a Sylow 2-subgroup of G, F = FP (G)
and x is a central involution in P . Then CF (x) is not nilpotent. Indeed,
let q = 2n . By [8, p.528], one can choose the hermitian form given by
0 0 1
A = 0 1 0 and define SU3 (q 2 ) = {X ∈ SL3 (q 2 ) : X t AX = A}
1 0 0
n
where (Aij ) = (A2ij ) and P SU3 (q 2 ) = SU3 (q 2 )/Z(SU3 (q 2 )). Observe
that the order of Z(SU3 (q 2 )) divides 3. Denote G = P SU3 (q 2 ). By
[8, p. 529], all involutions in G are conjugate and every involution is
1 0 1
central. So, choose x = 0 1 0 ∈ G, and observe by [8, p.529]
0 0 1
that
α β
γ
CG (x) = { 0 αq−1 α−1 β q Z : αq+1 = 1 and αγ q +αq γ = β q+1}.
0
α−q
0
We need to show that CG (x) does not have a normal 2-complement.
Observe that |CG (x)| = q 3 (q + 1) and
a
0
Q = { 0 aq−1
0
0
0
0 Z : aq+1 = 1}
a−q
is a subgroup of CG (x) of order q + 1. Since Q is a complement of the
Sylow 2-subgroup P of CG (x) (which is itself normal), if one shows that
Q is not normal, then CG (x) does not have a normal 2-complement and
by [2, Theorem 1.12], CF (x) is not nilpotent.
Assume that Q E CG (x). Since P E CG (x), elements of Q must
commute with elements of P , indeed, CCG (x) (P ) must contain Q. If
10
KIVANÇ ERSOY AND İPEK TUVAY
a
0
0
n+1
= a, hence we get a
0 aq−1 0 ∈ CCG (x) (P ) then a2q = a2
0 0 a−q
contradiction.
Therefore, CF (x) is not nilpotent for P SU3 (2n ). Finally, we need
to check 2 B2 (22m+1 ). Observe that for any central involution x, one
has CG (x) = P where P is the Sylow 2-subgroup containing x. Hence
the centralizer fusion system is nilpotent.
Theorem 3.6. Let G be a finite group with soluble centralizers of
involutions, P be a Sylow 2-subgroup of G and F = FP (G) and x be
a central involution of G. Assume that the soluble radical R of G has
odd order and G/R is simple. Then CF (x) is a nilpotent fusion system
except G/R is either A8 , P SL3(2n ), P SU3 (2n ) or P SL3 (3).
Proof. From [7, Theorem 2.12], FP (G) ∼
= FP R/R (G/R). Moreover since CP (x) = P , < x > is fully centralized, so from [2, Theorem
4.27] we get
CF (x) = FP (CG (x)) ∼
= FP R/R (CG/R (xR))
Thus, it is enough to calculate the centralizers of central involutions for
the finite simple groups S such that G/R ∼
= S (here we should mention
that S is one of the finite simple groups listed in Proposition 3.1).
If S = A5 , any involution in S is a product of two transpositions
and P is isomorphic to Klein 4-group. Let x be one of the involutions
in P . Since CS (x) = P , we have CF (x) = FP (P ).
If S = A6 , then for any involution in S, there exists a Sylow
2-subgroup of S whose center contains this involution. Indeed, for
any involution x = (ab)(cd) and for y = (acbd)(ef ) where {e, f } ∈
{1, 2, 3, 4, 5, 6}\{a, b, c, d} one has P =< y, z | y 4 = z 2 = 1, zyz −1 =
y −1 >∼
= D8 such that x ∈ Z(P ). Here, P is a Sylow 2-subgroup of S,
so x is a central involution. Observe that CS (x) ∼
= (C2 ≀ C2 ) × C2 , from
6
this we get |CA6 (x)| = 8. Hence CF (x) = FP (P ).
CENTRALIZER FUSION SYSTEMS OF CENTRAL INVOLUTIONS
11
If S = A7 , likewise A5 and A6 any involution is a product of two
transpositions. Since CS (x) ∼
= (C2 ≀C2 )×S3 and CA (x) = CS (x)∩A7 ,
7
7
7
we get CS (x) ∼
= P × A3 where P is a Sylow 2-subgroup of S isomorphic
to D8 . Hence CF (x) = FP (P × A3 ) = FP (P ).
Let S = A8 ∼
= GL4 (2) and P is a Sylow 2-subgroup of S. In this
case, any central involution in S is a product of four transpositions.
If x is one of these involutions, CS (x) ∼
= C2 ≀ S4 , and similar to the
8
arguments above CA8 (x) ∼
= UT4 (2), CF (x) is not
= C24 ⋊ A4 . Since P ∼
nilpotent.
If S = P SL2 (q) for some q, then by Lemma 3.4, CF (x) is nilpotent
for any involution x ∈ S.
If S = P SL3 (3) any involution is central and from [4, Lemma 5.1]
CS (x) ∼
= GL2 (3) for any involution x in S. Since GL2 (3) is not 2nilpotent, by [2, Theorem 1.12] we deduce that CF (x) is not nilpotent.
If S = M11 , observe that the centralizer of the central involution
z ∈ S, namely CS (z) has order 48 and the Sylow 3-subgroup of CS (z)
is not normal. That means, CF (x) is not nilpotent.
By Theorem 3.5, when the simple group is defined over a field of
characteristic 2, the centralizer fusion system CF (x) of the central involution is nilpotent iff G ∼
= SL2 (2n ),2 B2 (22m+1 ) or P SL(3, 2) and not
nilpotent in P SL3 (2n ) and P SU3(2n ).
This completes the proof of the main theorem.
Acknowledgements
The first and second authors were partially supported by Mimar Sinan
Fine Arts University Scientific Research Project Unit with project numbers
2017/21 and 2017/22, respectively.
References
[1] M. Aschbacher, R. Kessar, B. Oliver, Fusion Systems in Algebra and Topology, Cambridge Univ. Press, Cambridge, 2011.
12
KIVANÇ ERSOY AND İPEK TUVAY
[2] D. A. Craven, The Theory of Fusion Systems, Cambridge Univ. Press, Cambridge, 2011.
[3] B. Hartley, M. Kuzucuol̆u, “Centralizers of elements in locally finite simple
groups”, Proc. London. Math. Soc.3 (62) 1991 301-324
[4] B. Huppert, N. Blackburn, Finite Groups III, Springer-Verlag, Berlin Heidelberg, 1982.
[5] D. Gorenstein, R. Lyons, “Non-solvable finite groups with solvable 2-local
subgroups”, J. Algebra 38, 453-522 (1976).
[6] V.D. Mazurov, E.I. Khukhro, Unsolved problems in group theory, The
Kourovka Notebook 18 Novosibirsk 2014.
[7] M. Linckelmann, Introduction to fusion systems, Group Representation
Theory, EPFL Press, Lausanne, 2007, 79-113.
[8] L. Pforte, “The involution module of P SU3 (22f ), Osaka J. Math. 52 (2015),
527-544
[9] S.A. Syskin, “On finite groups with solvable centralizers of involutions”,
Algebra and Logic, May 1971, Volume 10, Issue 3, pp. 206-216.
[10] J. Tits, “Algebraic and abstract simple groups”, Ann. of Math. 80 (1964),
313-329
[11] J.G. Thompson, “Nonsolvable finite groups all of whose local subgroups
are solvable”, Bulletin of the American Mathematical Society 74, (1968)
383-437
[12] J.G. Thompson, “Nonsolvable finite groups all of whose local subgroups are
solvable. II”, Pacific Journal of Mathematics 33, (1970) 451-536
[13] J.G. Thompson, “Nonsolvable finite groups all of whose local subgroups are
solvable. III”, Pacific Journal of Mathematics 39, (1971) 483-534
[14] J.G. Thompson, ”Nonsolvable finite groups all of whose local subgroups are
solvable. IV”, Pacific Journal of Mathematics 48, (1973) 511-592
[15] J.G. Thompson, “Nonsolvable finite groups all of whose local subgroups are
solvable. V”, Pacific Journal of Mathematics 50, (1974) 215-297
[16] J.G. Thompson, “Nonsolvable finite groups all of whose local subgroups are
solvable. VI”, Pacific Journal of Mathematics 51, (1974) 573-630
CENTRALIZER FUSION SYSTEMS OF CENTRAL INVOLUTIONS
13
Department of Mathematics, Mimar Sinan Fine Arts University, Istanbul, 34427 Turkey and Institute of Mathematics, Freie Universität
Berlin, Germany
E-mail address: [email protected]
Department of Mathematics, Mimar Sinan Fine Arts University,
Istanbul, 34427 Turkey
E-mail address: [email protected]
| 4math.GR
|
1
Inexact Alternating Optimization for Phase
Retrieval In the Presence of Outliers
arXiv:1605.00973v2 [cs.IT] 24 Jul 2017
Cheng Qian, Xiao Fu, Member, IEEE, Nicholas D. Sidiropoulos, Fellow, IEEE, Lei Huang, Senior Member, IEEE,
and Junhao Xie, Senior Member, IEEE
Abstract—Phase retrieval has been mainly considered in the
presence of Gaussian noise. However, the performance of the
algorithms proposed under the Gaussian noise model severely
degrades when grossly corrupted data, i.e., outliers, exist. This
paper investigates techniques for phase retrieval in the presence
of heavy-tailed noise – which is considered a better model for
situations where outliers exist. An `p -norm (0 < p < 2) based
estimator is proposed for fending against such noise, and twoblock inexact alternating optimization is proposed as the algorithmic framework to tackle the resulting optimization problem.
Two specific algorithms are devised by exploring different local
approximations within this framework. Interestingly, the core
conditional minimization steps can be interpreted as iteratively
reweighted least squares and gradient descent. Convergence
properties of the algorithms are discussed, and the CramérRao bound (CRB) is derived. Simulations demonstrate that the
proposed algorithms approach the CRB and outperform stateof-the-art algorithms in heavy-tailed noise.
Index Terms—Phase retrieval, iterative reweighted least
squares (IRLS), gradient descent, impulsive noise, Cramér-Rao
bound (CRB).
I. I NTRODUCTION
Phase retrieval aims at recovering a signal x ∈ CN
from only the magnitude of linear measurements. This is
an old problem [2], [3] that has recently attracted renewed
and growing interest. Phase retrieval arises in many fields,
such as X-ray crystallography, coherent diffraction imaging,
and optical imaging and astronomy, where the detectors only
record the intensity information, because phase is very difficult
and expensive to measure.
The mathematical description of the phase retrieval problem
is simple: given the measuring matrix A = [a1 · · · aM ]H ∈
CM ×N and the measurement vector y ∈ RM , where am ∈
CN , and
y = |Ax|,
find x ∈ CN . Early attempts to solve the phase retrieval
problem can be traced back to the 1970s, where techniques
such as Gerchberg-Saxton (GS) [2], Fienup [3] and their variants were proposed. These algorithms have been empirically
Original manuscript submitted to IEEE Trans. on Signal Processing,
October 4, 2016; revised July 25, 2017. Conference version of part of this
work appears in Proc. EUSIPCO 2016 [1].
C. Qian, X. Fu and N. D. Sidiropoulos are with the Department of Electrical
and Computer Engineering, University of Minnesota, Minneapolis, MN 55455
USA (e-mail: [email protected], [email protected], [email protected]).
L. Huang is with the College of Information Engineering, Shenzhen
University, Shenzhen, 518060 China (e-mail: [email protected]).
J. Xie is with the Department of Electronics and Information Engineering,
Harbin Institute of Technology, Harbin, 150001 China (e-mail: [email protected]).
The work of N. Sidiropoulos was supported by NSF CIF-1525194.
shown to work well under certain conditions, although little
had been known regarding their convergence properties from a
theoretical point of view. Recently, some new algorithms along
this line of work were proposed in [4], where the convergence
issue is better studied.
In recent years, more modern optimization-based approaches for phase retrieval have been proposed. For example,
Candès proposed a semidefinite relaxation approach known as
the PhaseLift algorithm [5] and proved that exact recovery
is possible with high probability in the noiseless case. Hand
recently studied the robustness of PhaseLift and showed that
it can tolerate a constant fraction of arbitrary errors [6].
Wirtinger-Flow (WF) [7] is a more recent approach that combines a good statistical initialization with a computationally
light gradient-type refinement algorithm. The combination
works very well when A is i.i.d. Gaussian. Following [7],
conceptually similar approaches, namely, the truncated WF
(TWF) [8] and truncated amplitude flow (TAF) algorithms [9],
[10] were proposed to handle more challenging scenarios. In
the recent work in [11], [12], the authors proposed a leastsquares feasible point pursuit (LS-FPP) approach that aims to
solve the same optimization problem as LS PhaseLift in the
presence of noise. Simulations in [11] indicate that LS-FPP
approaches the Cramér-Rao bound (CRB) for the Gaussian
measurement model considered in [11].
Some phase retrieval algorithms were originally developed
under an exact measurement model, and subsequently treated
the noisy case by replacing equality constraints with relaxed
inequality constraints [5]-[13]. Most of the existing algorithms
were explicitly or implicitly developed under a Gaussian noise
model. In certain applications, a subset of the measurements
may be corrupted much more significantly than the others, and
heavy-tailed noise may be encountered as well [14]-[16]. One
representative example is high energy coherent X-ray imaging
using a charge-coupled device (CCD), where the impulsive
noise originates from X-ray radiation effects on the CCD,
and the density of impulses increases with the intensity of
X-ray radiation or CCD exposure time [16]-[17]. Under such
circumstance, modeling the noise as Gaussian is no longer
appropriate.
In recent years, robust phase retrieval algorithms, e.g.,
[14], [15], [18], have been proposed to handle outliers. The
framework in [15] considered an undersampled phase retrieval
model corrupted with Laplacian-distributed outliers, but the
approach was designed specifically for sparse x. Variations of
TWF to handle outliers were also considered. For example, the
work in [18] replaces the sample mean that is used in TWF
2
by sample median in the initialization and truncated update,
which exhibits robustness to outliers under certain conditions
[18]. However, as TWF, this approach still heavily relies on
the assumption that Gaussian measurements are employed.
Another important aspect of phase retrieval is how noise
enters the measurement model. Our previous work [11] considered the noise model y = |Ax|2 + n. Another noise model
that is frequently considered in the literature [2]-[4] is
y = |Ax| + n.
(1)
Notice the subtle but important difference between the two
models: whether noise is added to the magnitude or the
squared magnitude. The choice hinges on the experimental
setup, including the measurement apparatus; but (1) is more
widely adopted by experimentalists.
Contributions: We consider the phase retrieval problem under
the model in (1) in the presence of impulsive noise, and
focus on designing robust algorithms to handle it. To fend
against impulsive noise, we adopt the `p -fitting (0 < p < 2)
based estimator that is known to be effective in dealing with
outliers, and devise two optimization algorithms using twoblock inexact alternating optimization. Specifically, the two
algorithms both solve local majorization-based approximate
subproblems for one block, instead of exactly solving the
conditional block minimization problem to optimality. Interestingly, starting from different majorizations, the resulting solutions turn out equivalent to iteratively reweighted least squares
and gradient descent, respectively. Unlike the existing inexact
and exact alternating optimization frameworks that mostly
operate with convex constraints, the proposed algorithms work
with a unit-modulus nonconvex constraint, so convergence
analysis seems difficult. Nevertheless, we prove convergence
of the proposed algorithms to a Karush-Kuhn-Tucker (KKT)
point by exploiting the two-block structure of the problem.
We also derive computationally light implementations using
Nesterov-type and stochastic gradient updates. In order to
assist in performance analysis and experimental design, we
derive the CRB for the model in (1) and under different
parameterizations, in the presence of Laplacian and Gaussian
noise. Curiously, although related CRBs for different noisy
measurement models have been previously derived in [11],
[19]-[22], to the best of our knowledge, there is no available
CRB for the model in (1) – and our work fills this gap. The
proposed algorithms are validated by extensive simulations.
The simulations show that our approaches outperform the
state-of-the-art algorithms in the presence of impulsive noise.
A preliminary conference version of part of this work
has been submitted to EUSIPCO 2016 [1]. The conference
version includes one of the two basic algorithms, the Laplacian
CRB, and limited simulations. The second algorithm, Nesterov
acceleration and stochastic gradient-type updates, additional
CRB results, and, most importantly, proof of convergence of
the iterative algorithms are all provided only in this journal
version, which naturally also includes more comprehensive
simulation results.
Notation: Throughout the paper, we use boldface lowercase
letters for vectors and boldface uppercase letters for matrices. Superscripts (·)T , (·)∗ , (·)H , (·)−1 and (·)† represent
transpose, complex conjugate, conjugate transpose, matrix
inverse and pseudo-inverse, respectively. The Re{·} and Im{·}
denote the real part and imaginary part. E[·] is the expectation
operator, | · | is the absolute value operator, || · ||F is the
Frobenius norm, || · ||p is the vector `p -norm, whose definition
PN
is ||x||p = ( i=1 |xi |p )1/p ,
is the element-wise product,
and diag(·) is a diagonal matrix with its argument on the
diagonal. δij denotes the Kronecker delta function, and ∠(·)
takes the phase of its argument. 0m , 1m , and Im stand for
the m × 1 all-zero vector, m × 1 all-one vector, and m × m
identity matrix, respectively. Furthermore, trace(·) and ∂a/∂x
denote the trace and partial derivative operators, respectively.
II. P ROPOSED A LGORITHMS
A. AltIRLS
Let us first consider the noiseless case where y = |Ax|.
Effectively, it can also be written as
y
u = Ax
(2)
where u = ej∠(Ax) is an auxiliary vector of unknown unitmodulus variables with its mth component being um =
H
ej∠(am x) . In the presence of impulsive noise, `p -(quasi)-norm
has be recognized as an effective tool for promoting sparsity
and fending against outliers [23]-[25]. Therefore, we propose
to employ an `p -fitting based estimator instead of using the
`2 -norm, which has the form of
min
|u|=1,x
M
X
2
|ym um − aH
m x| +
p/2
(3)
m=1
where 0 < p < 2 is chosen to down-weigh noise impulses
(i.e., outliers), and > 0 is a small regularization parameter
(e.g., ∈ [10−8 , 10−6 ]) that keeps the cost function within its
differentiable domain, which will prove handy in devising an
effective algorithm later.
To handle Problem (3), we follow the rationale of alternating
optimization, i.e., we first update x fixing u, and then we do
the same for u.
Assume that after some iterations, the current solution at
iteration r is (x(r) , u(r) ). At step (r + 1), the subproblem
with respect to (w.r.t.) x is
x(r+1) = arg min
x
M
p/2
X
H 2
|ym u(r)
m − am x| +
(4)
m=1
which is still difficult to handle. Particularly, when p < 1,
the subproblem itself is still non-convex; when p ≥ 1, the
subproblem is convex but has no closed-form solution. Under
such circumstances, we propose to employ the following
lemma [26]:
Lemma 2.1: Assume 0 < p < 2, > 0, and φp (w) :=
p
p−2
2−p 2
+ w. Then, we have
2
pw
x2 +
p/2
= min wx2 + φp (w)
w≥0
and the unique minimizer is
p−2
p 2
wopt =
x + 2 .
2
(5)
(6)
3
By
PM
Lemma
2.1,
an
upper
bound
of
p/2
(r)
H 2
that is tight at the current
m=1 |ym um − am x| +
solution x(r) can be easily found:
M
X
|ym u(r)
m
−
2
aH
m x|
+
Algorithm 1 AltIRLS for phase retrieval
1:
2:
(0)
W
p/2
p
2
m=1
M
X
H
(r)
≤
ym u(r)
wm
m − am x
2
(r)
+ φp wm
(7)
m=1
3:
4:
5:
where
(r)
wm
p
:=
2
ym u(r)
m
−
(r)
aH
mx
6:
p−2
2
2
+
(8)
7:
(r)
and the equality holds if and only if x = x . Instead
of directly dealing with Problem (4), we solve a surrogate
problem using the right hand side (RHS) of (7) at each iteration
to update x. Notice that the RHS of (7) is convex w.r.t. x and
the corresponding problem can be solved in closed-form:
†
x(r+1) = W(r) A W(r) y u(r)
(9)
where
W(r) = diag
q
(r)
w1
···
q
(r)
wM .
(10)
The conditional problem w.r.t. u is
u(r+1) = arg min
|u|=1
M
X
(r+1) 2
|ym um − aH
| +
mx
p/2
.
m=1
(11)
Although the problem is non-convex, it can be easily solved to
optimality. Specifically, the first observation is that the partial
minimization w.r.t. u is insensitive to the value of p; i.e., given
a fixed x, for any p > 0, the solutions w.r.t. u are identical.
Second, for all p > 0, the solution is simply to align the angle
(r+1)
of ym um to that of aH
, which is exactly
mx
H
(r+1)
u(r+1)
= ej∠(am x
m
function A LT IRLS (y, A, x(0) )
(0)
Initialize u(0)q= exp(∠(Ax
)) and W(0)
q
),
m = 1, . . . , M.
(12)
We update x and u alternately, until some convergence criterion is met. We see that the way that we construct the upper
bound of the partial problem w.r.t. x is in fact the same as the
procedure in iteratively reweighted least squares (IRLS) [25],
[26]. The difference is that we ‘embed’ this IRLS step into
an alternating optimization algorithm. We therefore call this
algorithm alternating IRLS (AltIRLS), which is summarized
in Algorithm 1.
A relevant question regarding Algorithm 1 is whether or
not this algorithm converges to a meaningful point, e.g., a
stationary or KKT point. Note that the block variable u is
constrained to a non-convex set, and we do not compute the
optimal solution for the block variable x at each iteration
of Algorithm 1. For such a type of algorithm, there is no
existing theoretical framework that establishes convergence.
We therefore need careful custom convergence analysis. We
have the following result:
Proposition 2.1: Assume that 0 < p < 2 and > 0, and that
A has full column rank. Then, the solution sequence produced
8:
9:
10:
=
diag
(0)
(0)
w1
(0)
ym um − aH
mx
···
p−2
2
2
+
(0)
wM
=
(0)
wm
with
=
while stopping criterion has not been reached do
x(r) = (W(r−1) A)† W(r−1) (y u(r−1) ).
u(r) = exp(j∠(Ax(r) ))
p−2
2
2
(r)
(r)
(r)
wm = p2 ym um − aH
, ∀m
x
+
m
q
q
(r)
(r)
W(r) = diag
w1 · · ·
wM
r =r+1
end while
end function
by Algorithm 1 converges to a set K that consists of all the
KKT points of Problem (3).
Proof: See Appendix A.
The result in Proposition 2.1 is interesting: Although the algorithm that we propose to compute the `p fitting-based estimator
solves non-convex subproblems, it ensures convergence to a
KKT point. Such a nice property is proven by exploiting the
two-block structure of the algorithm. Note that the proof itself
does not rely on the specific form of the optimization problem
in (3), and thus can be easily extended to other two-block
alternating optimization cases, which we believe is of much
broader interest.
B. AltGD
Algorithm 1 uses a simple update strategy, but its complexity may still become an issue when the problem size grows.
Specifically, the bottleneck of the AltIRLS algorithm lies in
solving the subproblem w.r.t. x, i.e.,
min
x
W(r) (y
u) − W(r) Ax
2
(13)
2
whose closed-form solution can be computed via (9). It is
observed from (9) that the matrix inversion part requires
O(N 3 ) flops to compute and also O(N 2 ) memory to store,
both of which are not efficient for high-dimensional x – e.g.,
when x is a vectorized image, N is usually very large (more
specifically, for a 100 × 100 image, N is 10,000).
To circumvent this difficulty, one way is to employ a certain
fast least squares solver such as conjugate gradient [28] to
handle Problem (13). This is a viable option, but such solvers
still typically require many iterations to obtain a fairly accurate
solution of (13). Here, we propose to deal with problem (13)
using a simpler approach. Specifically, let us denote
f (r) (x) = W(r) (y
u(r) ) − W(r) Ax
2
2
(14)
and approximate (13) using the following function
g (r) (x) = f (r) (x(r) ) + Re{(∇f (r) (x(r) ))H (x − x(r) )}
4
µ(r)
x − x(r)
2
2
.
2
3
where µ(r) ≥ 0 is a pre-specified parameter. Instead of optimizing f (r) (x), we optimize g(r) (x). By rearranging terms,
the subproblem becomes
2
1
(16)
x(r+1) = arg min x − x(r) − (r) ∇f (x(r) )
x
µ
2
and the solution is
x(r+1) = x(r) −
1
∇f (x(r) )
µ(r)
(17)
if µ(r) ≥ λmax (AH (W(r) )2 A), and the equality holds if
and only if x = x(r) . Therefore, the algorithmic structure
of Algorithm 2 is the same as that of Algorithm 1, except
that the majorizing functions of the x-block are different –
which means that the proof of convergence of Algorithm 1
also applies here:
Corollary 2.1: If µ(r) ≥ λmax (AT (W(r) )2 A) for all r,
then the whole sequence produced by Algorithm 1 converges
to K.
Algorithm 2 AltGD for phase retrieval
5:
6:
7:
8:
9:
2
10
1
which does not require any matrix inversion operation. Also,
the N × N matrix does not
2 need to be stored, if2 we take
the order Ax(r) → W(r) Ax(r) → AH W(r) Ax(r) to
compute the update of x. Therefore, using this approach, both
memory and complexity requirements are less demanding.
We summarize the algorithm in Algorithm 2. We call this
algorithm alternating gradient descent (AltGD).
Note that when µ is chosen as the Lipschitz constant, (15)
is an upper bound of the cost function in (13). In other words,
we have
f (r) (x) ≤ g (r) (x)
2:
3:
4:
AltGD
Accelerated AltGD
10
i.e., a gradient step with step-size 1/µ(r) , where the gradient
is
2
∇f (r) (x(r) ) = AH W(r) Ax(r) − y u(r)
(18)
1:
10
(15)
|| y−|Ax| ||pp
+
function A LT GD(y, A, x(0) )
Initialize u(0) = exp(∠(Ax(0) ))
while stopping criterion has not been reached do
Choose µ(r−1) as the leading eigenvalue of
H
A (W(r−1) )2 A or the trace of (W(r−1) )2
x(r) = x(r−1) − 1/µ(r−1) ∇f (x(r−1) )
u(r) = exp(j∠(Ax(r) ))
r =r+1
end while
end function
Remark 2.1: Exactly computing λmax (AH (W(r) )2 A) may
be time consuming in practice. Many practically easier ways
can be employed, e.g., the Armijo rule-based line search
[27]. In our simulations, we use a simple heuristic: we
let µ(r) = trace((W(r) )2 ) instead of λmax (AH (W(r) )2 A).
The rationale behind is that we observe that the energy of AH (W(r) )2 A is usually dominated by W(r) ,
0
10
0
20
40
60
80
100
120
Number of iterations
140
160
180
200
Fig. 1. Cost function value versus number of iterations.
and using µ(r) = trace((W(r) )2 ) is a good approximation of trace(AH (W(r) )2 A) that is an upper bound of
λmax (AH (W(r) )2 A). We should mention that this step-size
choice is a heuristic, but works well in practice, as will be
shown in the simulations.
C. Further Reducing Complexity
1) Extrapolation: Algorithm 2 is easier to compute than
Algorithm 1 in terms of per-iteration complexity. However,
first-order methods tend to require more iterations in total.
One way to alleviate this effect is to incorporate Nesterov’s
“optimal first-order” method [29]-[31], i.e., for each update,
we set
t(r−1) − 1 (r)
(r−1)
z(r) = x(r) +
x
−
x
(19)
t(r)
p
1 + 1 + 4(t(r−1) )2
(20)
t(r) =
2
1
x(r+1) = z(r) − (r) ∇f (r) (z(r) )
(21)
µ
In practice, the above ‘extrapolation’ technique greatly expedites the whole process in various applications [23], [32].
Some numerical evidence can be seen in Fig. 1, where a
simple comparison between the plain Algorithm 2 and the
extrapolated version is presented. We choose SNR= 20 dB,
N = 16 and M = 128. Each element in the signal and measurement vectors is independently drawn from the complex
circularly symmetric Gaussian distribution with mean zero and
variance one. The noise is generated from a symmetric α stable
(SαS distribution) which will be described in detail in Section
IV. Fig. 1 shows the convergence of AltGD and accelerated
AltGD using extrapolation when p = 1.3. As we can see, the
accelerated AltGD converges after 40 iterations which is much
faster than AltGD that converges after 200 iterations.
2) Block Incremental / Stochastic Gradient: When N is
very large (thus M > N is larger), even gradient computation
is too much of a burden. In such cases, a pragmatic way
is to “break down” the problem to pieces and do (block)
5
incremental or stochastic gradient. We divide the measurement
matrix into L blocks Γl . Then it is straightforward to get the
gradient for the lth block as
(r) 2
(r)
∇fΓl (x(r) ) = AH
AΓl x(r) − yΓl uΓl . (22)
Γl WΓl
The estimate of x is updated according to
x(r+1) = x(r) −
1
(r)
µΓl
∇fΓl (x(r) )
(23)
Proof: See Appendix B.
The CRB for real x is a special case of the complex case,
which can be easily derived from Proposition 3.1:
Proposition 3.2: In Laplacian noise, the CRB of real-valued
x is
CRBL,r = trace F−1
(27)
L,r
where
FL,r =
(r)
where µΓl is chosen as the leading eigenvalue of AH
Γl AΓl or
(r) 2
trace WΓl
. The algorithm can proceed block by block
with revisits, or by randomly picking blocks, resulting in
block incremental gradient and stochastic gradient versions,
respectively.
A subtle point here is that, to maintain robustness, one
should choose a block size larger than one. The reason is
that the robustness of the algorithm is brought by treating
different ym with different weights (more specifically, by
downweighting the outliers). When using only one ym for
updating, this ability vanishes.
III. C RAM ÉR -R AO B OUND A NALYSIS
In this section, the CRB on the accuracy of retrieving x
in (1) is presented. The CRB provides a lower bound on the
MSE of unbiased estimators. In many signal processing applications, one cannot guarantee that an estimator is unbiased,
or even that an unbiased estimator exists, yet the CRB is still
surprisingly predictive of estimator performance [35], [38]. For
phase retrieval under Gaussian noise, over the past few years,
several CRBs have been derived for different models, e.g., 2-D
Fourier-based measurements [19], noise added prior to taking
the magnitude [20] and noise added after taking the magnitude
square [11], [21]. To the best of our knowledge, there is no
available CRB formula for the signal model in (1). Here, we
present the CRBs for two particular types of noise: Laplacian
and Gaussian noise. Although our main interest here lies in
evaluating performance of robust algorithms and the Laplacian
CRB serves this purpose, the Gaussian CRB is of interest in
other application contexts. Note that we use subscripts r, c,
L and G to stand for real, complex, Laplacian and Gaussian,
respectively.
To get started, we should note that our derivations are
based on the assumption that aH
m x is nonzero. That is because
the term |Ax| is non-differentiable at aH
m x = 0, the Fisher
information only exists at aH
x
=
6
0.
With
this caveat, we
m
have the following proposition:
Proposition 3.1: In Laplacian noise, the CRB is
CRBL,c = trace F†L,c
(24)
GL,r = Re{AH diag(Ax)}.
(25)
CRBG = 2CRBL .
GL,c
Re{AH diag(Ax)}
=
.
Im{AH diag(Ax)}
(26)
(30)
Proof: The proof is straightforward by calculating the
Gaussian FIM and comparing it with the Laplacian FIM in
(28). We omit it here.
In certain applications of phase retrieval, we may be more
interested in the performance bound for retrieving the phase
of x. The following proposition provides the CRB on both
the phase and amplitude of x under Laplacian noise. Note
that in the Gaussian noise case, it is straightforward to apply
Proposition 3.3 to compute the lower bound. Similar results
can also be found in [36].
Proposition 3.4: In Laplacian noise, the CRB of the amplitude of x is
CRBL,|x| =
N
X
di
(31)
i=1
and the variance of any biased estimate of the phase of x is
bounded below by
CRBL,∠(x) =
2N
X
di
(32)
i=N +1
where d = [d1 · · · d2N ] contains the main diagonal elements
of F†L which in this case is defined as
2
GL diag(|Ax|−2 ) GTL
σn2
(33)
with
GL =
with
(29)
In Appendix C, we show that when A has full column
rank, FL,c is singular with rank (2N −1) while FL,r is always
nonsingular. Thus, for complex x, we adopt its pseudo-inverse
to compute an optimistic (looser) CRB, which is still a valid
lower bound that can be used to benchmark the efficiency of
any biased estimator [33]-[35]. If x is close to zero, then the
CRB is not tight at low SNRs, and more measurements should
be used to approach the CRB.
Proposition 3.3: In the Gaussian noise case, the CRB is two
times larger than the CRB in Proposition 3.1, i.e.,
FL =
2
GL,c diag(|Ax|−2 ) GTL,c
σn2
(28)
with
where
FL,c =
2
GL,r diag(|Ax|−2 ) GTL,r
σn2
diag(|x|)−1
IN
Re diag (x∗ ) AH diag(Ax)
.
Im{diag(x∗ )AH diag(Ax)}
(34)
Proof: See [11] and Appendix B.
6
In this section, we evaluate the performance of the proposed
methods and compare them with some existing algorithms,
namely, WF [7], TWF [8], TAF [9], AltMinPhase [4], MTWF
[18] and GS [2] in terms of MSE performance, where the
MSE is computed after removing the global phase ambiguity
between the true and estimated x. In the simulations, we
consider an exponential signal x = exp(j0.16πt) comprising
16 samples, i.e., t = 1, · · · , 16. We test the algorithms
using different types of A, i.e., random matrix that is usually
employed in [4], [7]-[9], [18], and the 2D Fourier matrix that
is widely used in real devices [2], [3] (cf. subsection IV-E).
For the random measuring matrix case (used in subsections
A-D), the measurement vectors are generated from a masked
Fourier transformation, which takes the form of
T
A = (DΛ1 )T · · · (DΛK )T
where K = M/N is the number of masks, D is a N × N
discrete Fourier transform (DFT) matrix with DDH = N IN
and Λk is a N × N diagonal masking matrix with its diagonal
entries generated by b1 b2 , where b1 and b2 are independent
and distributed as [5]
1 w. prob. 0.25
(√
−1 w. prob. 0.25
2/2 w. prob. 0.8
√
b1 =
and b2 =
−j
w.
prob.
0.25
3 w. prob. 0.2.
j w. prob. 0.25
All results are obtained using a computer with 3.1 GHz Intel
Xeon E31225 and 16 GB RAM. For the random measurement
case, the algorithms are all initialized from the same starting
point
is computed by picking the principal eigenvector of
PM that
2
H
y
a
i=1 i i ai , and the stopping criterion is
2
2
− y − |Ax(r−1) |
y − |Ax(r−1) |
2
2
2
2
AltIRLS
AltGD
100
10-5
10-10
IV. S IMULATION R ESULTS
y − |Ax(r) |
105
MSE
Remark 3.1: In deriving the CRB in Proposition 3.4, we
use no additional assumptions on x. Therefore, Proposition
3.4 works for both real and complex x. Specifically, in the
real case, the phase is actually the sign of x. Here, FL is
also singular with rank deficit equal to one, so we adopt its
pseudo-inverse to compute the CRB. We omit the proof of
rank-1 deficiency of FL , since it follows the line of argument
in Appendix C.
≤ 10−7
or the number of iterations reaching 1000. Furthermore, we
use AltGD with extrapolation for the simulations.
A. Selection of p
Before we do the performance comparison, let us study how
p affects the performance of the proposed methods. In this
example, SNR is fixed at 20 dB, where the SNR is computed
via
||Ax||2
SNR = 10 log10
.
||n||2
We assume that 10% of the data are corrupted by outliers that
are generated from the Gaussian distribution with mean zero
10-15
10-20
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
p
Fig. 2. MSE versus p.
and variance 100, and the remaining elements in n are zero.
K = 8 masks are employed to generate the measurements.
Fig. 2 shows the MSE as a function of p. Note that, when
p is smaller than 1, in order to achieve convergence in a
non-convex setting, we initialize our methods in a two-step
way. Specifically, for each of the proposed methods, when
0.6 < p < 1, we use the spectrum output to initialize our
methods with p = 1.3 and 100 iterations, and then choose the
corresponding output to do another 100 iterations with p = 1
and use the output as a final starting point. When p ≤ 0.6,
we take an additional intermediate step with p = 0.7 and 100
iterations to gradually stage the initialization process for our
approaches, since in this p-regime the subproblem is strictly
non-convex. It is observed that generally, p ≤ 1 provides
better performance than p > 1, especially for AltIRLS. When
p < 0.5, AltGD suffers severe performance degradation while
AltIRLS still performs well. Our understanding is that firstorder algorithms are in general more sensitive to problem
conditioning, and a small p can lead to badly-conditioned
W(r) . Working with p < 1 usually requires much more
careful initialization. In our experience, p ≈ 1.3 strikes a
good compromise between numerical stability and estimation
accuracy.
To illustrate the performance provided by our proposed
methods with p < 1 over the state-of-art algorithms including
TWF, WF, TAF, AltMinPhase, MTWF, and GS, we present the
following example, where the parameters are the same as for
Fig. 2, except that 20% of the data are corrupted by outliers.
Note that among these competitors, TAF, AltMinPhase and GS
share the same magnitude measurement model in (1) as ours,
but WF is designed under the energy measurement model (i.e.,
2
ym = |aH
m x| + nm ). In addition, TWF and MTWF adopt a
2
2
Poisson model with λ = |aH
i x| . Hence, y is fed to WF,
TWF and MTWF, where the squaring is element-wise. We set
p = 0.4 for the proposed methods, and perform Monte-Carlo
trials to estimate MSE = 10 log10 (||x̂ − x||22 ). We show the
results of 1,000 random trials in Fig. 3. One can see that for
most trials the benchmark algorithms fail to give reasonable
7
AltIRLS
AltGD
100
600
400
50
200
0
-200
-150
-100
-50
0
0
-200
-150
MSE (dB)
WF
150
150
100
100
50
50
0
-200
-150
-100
-50
-100
-50
0
MSE (dB)
TWF
0
0
-200
MSE (dB)
GS
-150
-100
-50
0
MSE (dB)
AltMinPhase
150
400
100
200
that admit closed-form PDF expression, that is, α = 1 and
α = 2 which respectively yield the Cauchy and Gaussian
distributions. For α < 2, φ(t; α, β, γ, µ) possesses heavy tails,
and thus is considered suitable for simulating impulsive noise.
The parameter α controls the density of impulses - smaller α’s
correspond to more outliers. In the following, we set α = 0.8,
and the other parameters are β = 0, γ = 2 and µ = 0, resulting
in a symmetric α-stable (SαS) distribution with zero-shift.
3) GMM: In our simulations, we also use a two-component
GMM to generate impulsive noise, whose PDF is as follows:
2
X
|n|2
c
p i
exp
−
(35)
p(n) =
2σi2
2πσi2
i=1
where σi2 is the variance of the ith term, and 0 ≤ ci ≤ 1 is the
probability of occurrence of the ith component. Note that we
have c1 +c2 = 1, and we set c2 < c1 and σ22 > σ12 – therefore,
the second component corresponds to outliers. In this example,
we choose c1 = 0.9, c2 = 0.1, σ12 = 0.1 and σ22 = 100, which
corresponds to a situation where strong outliers are present.
50
0
-200
-150
-100
-50
0
0
-200
-150
MSE (dB)
MTWF
-100
-50
0
MSE (dB)
TAF
150
1000
100
500
0
-200
50
-150
-100
-50
0
0
-200
MSE (dB)
-150
-100
-50
0
MSE (dB)
Fig. 3. Signal recovery performance comparison.
results. However, our methods produce much lower MSEs,
which indicates that our proposed methods are efficient in
suppressing outliers.
B. Statistical Performance Comparison
We now compare the MSE performance as a function of
SNR. The results are averaged over 500 Monte-Carlo trials.
To simulate outliers, we use the Laplacian distribution, the
α-stable distribution, and a Gaussian mixture model (GMM)
to generate heavy-tailed noise n, respectively. Specifically, the
outlier generating process is summarized below:
1) Laplacian: The probability density function (PDF) of
the Laplacian distribution is given in (52).
2) α-stable: The PDF of an α-stable distribution is generally not available, but its characteristic function can be written
in closed-form as
φ(t; α, β, c, µ) = exp (jtµ − γ α |t|α (1 − jβsgn(t)Φ(α)))
where Φ(α) = tan (απ/2), 0 < α ≤ 2 is the stability
parameter, −1 ≤ β ≤ 1 is a measure of asymmetry, γ > 0 is
a scale factor which measures the width of the distribution
and µ is a shift parameter. There are two special cases
Figs. 5, 6 and 7 show the MSEs of x̂ under Laplacian, SαS
and GMM noise, respectively. In the simulations, the number
of measurements is set to be M = 8N . Note that for SαS
and GMM noise, we set p = 1.3 for AltIRLS and AltGD,
while for Laplacian noise we let p = 1 which corresponds
to the maximum likelihood estimator. In the Laplacian noise
case, we include the CRB derived in Proposition 3.2 as a
performance benchmark. It is observed from Fig. 5 that our
approaches outperform the GS algorithm and yield the MSEs
which are closest to the CRB. The performance gap between
the proposed methods and their competitors becomes larger
in the SαS and GMM noise cases since the outliers become
stronger; see Fig. 6 and 7. This indicates that the proposed
algorithms work better in more critical situations, i.e., when
more severe outliers exist. Note that the MTWF performs
slightly better than TWF in GMM noise and its performance
is inferior to our methods in the three noise scenarios.
As a reference, we also plot the MSE performance versus
SNR in Gaussian noise in Fig. 8, where we set p = 2 for
having a maximum likelihood estimator. We compare the performance of the algorithms with the CRB in Proposition 3.4.
One can see that the algorithms perform similarly when the
noise follows an i.i.d. Gaussian distribution. This suggests that
the Gaussian noise case may be an easier case - where all the
algorithms under test perform reasonably.
Fig. 4 compares the CPU time as a function of N , where
SNR is 10 dB, N increases from 8 to 128 and the other
parameters are the same as Fig. 7. It is seen that AltGD is
the fastest1 .
C. Success Rate Performance Comparison
Fig. 9 shows the success rate versus the proportion of
outliers in the measurements, where the GMM is employed
1 Note that in the simulation we use the heuristic step size as stated in
Remark 2.1; if the step size µ = λmax (AH (W(r) )2 A) is used, AltGD
may consume more time.
8
103
102
101
AltIRLS
AltGD
MTWF
TAF
TWF
WF
AltMinPhase
GS
CRB:Laplacian
10-1
MSE
CPU time (s)
100
AltIRLS
AltGD
MTWF
TAF
TWF
WF
AltMinPhase
GS
100
10-2
10-1
10-2
0
50
100
150
200
250
10-3
300
N
5
10
15
20
SNR (dB)
Fig. 4. CPU time versus number of variables N in GMM noise.
Fig. 5. MSE versus SNR in Laplacian noise.
to control the outlier fraction. We vary c2 from 0 to 0.5 and
fix σ12 = 0 and σ22 = 100. SNR is 10 dB. For the proposed
methods, we choose p = 0.4 for comparison. We declare
success when the squared Euclidean distance of the estimate
x̂ from the ground-truth x is smaller than or equal to 10−4 ;
i.e., the success rate is computed as
101
100
10-1
(36)
MSE
500
1 X
success rate =
γi
500 i=1
AltIRLS
AltGD
MTWF
TAF
TWF
WF
AltMinPhase
GS
10-2
where
(
γi =
1,
0,
if kx̂i − xk22 ≤ 10−4
otherwise.
10-3
(37)
Fig. 9 shows that our methods outperform the benchmark
methods by a large margin in terms of success rate. When
the outlier fraction is greater than 0.25, our methods can still
estimate x quite accurately, while the other six competitors
all fail. Note that the performance of AltIRLS is better than
AltGD – but the complexity of AltGD is much lower.
Fig. 10 plots the success rate versus sample complexity
(measured in terms of M/N with N = 16), where SNR is 10
dB and 20% of the data are corrupted by outliers. It is obvious
that our two algorithms have the highest success rates when
M/N ≥ 5, while the other algorithms have very low success
rates even when M/N ≥ 8.
10-4
5
10
15
Fig. 6. MSE versus SNR in SαS noise.
101
AltIRLS
AltGD
MTWF
TAF
TWF
WF
AltMinPhase
GS
100
In this subsection, we showcase the performance using real
caffeine molecule data that is often used to test phase retrieval
algorithms [7]. The caffeine molecule data with size 128 ×
128 is the projection of electron density of a 3-D Caffeine
molecule’s density map onto the xy-plane. The objective is to
reconstruct the data from the magnitude of its masked Fourier
transform. Here, we use the same measuring process as in the
previous simulations and K = 8 masks are used to generate
measurements. We compare the performance of AltGD, TAF,
MSE
10-1
D. Performance Comparison on Caffeine Molecule Image
Data
20
SNR (dB)
10-2
10-3
10-4
5
10
15
SNR (dB)
Fig. 7. MSE versus SNR in GMM noise.
20
9
AltIRLS
AltGD
MTWF
TAF
TWF
WF
AltMinPhase
GS
CRB:Gaussian
100
MSE
MTWF, and TWF. We also include the block incremental (BI)
implementation of AltGD (cf. Section II.C2), which is referred
to as BI-AltGD. The outliers are generated from GMM with
c2 = 0.3, σ12 = 0 and σ22 = 100, which indicates that there
are 30% of the data are corrupted by outliers. SNR is 0 dB
and the data is normalized with unit norm. Fig. 11 plots the
retrieved molecule’s density map using AltGD, BI-AltGD with
p = 1.3, TAF, MTWF and TWF. It is seen in Fig. 11 that
our schemes still work well, while the other competitors yield
blurred images.
10-1
E. Performance Comparison with 2D Fourier Measurements
10-2
5
10
15
20
SNR (dB)
Fig. 8. MSE versus SNR in Gaussian noise.
1
AltIRLS
AltGD
MTWF
TAF
TWF
WF
AltMinPhase
0.9
0.8
Success rate
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Outliers fraction
Fig. 9. Success rate of exact recovery versus outlier fraction in GMM noise.
1
AltIRLS
AltGD
MTWF
TAF
TWF
WF
AltMinPhase
0.9
0.8
Success rate
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
2
3
4
5
6
7
8
9
10
M/N
Fig. 10. Success rate of exact recovery versus M/N under GMM noise.
In the above simulations, the measurement vectors are
generated from coded diffraction pattern (CDP) with random
masks. The success of the existing algorithms such as the
well-known PhaseLift, WF and its variants are all based on
the randomness of measurements. However, to the best of
our knowledge, the CDP model is still considered to be impractical, as there has been no device supporting the masking
technique yet. In practice, most devices record the intensity
of the Fourier transform of the object, where the intensity
is usually proportional to the magnitude of the oversampled
2D Fourier measurements with 2× oversampling. In this case,
the measurement matrix A is a Kronecker product of two
oversampled Fourier matrices.
With 2D Fourier measurements, the most successful method
is Fienup’s hybrid-input-output (HIO) algorithm [3], which
has been observed to give empirically reliable phase retrieval
performance, especially in the noiseless case. In the noisy case,
HIO can also provide good initialization for other algorithms,
e.g., GS. Therefore, HIO and HIO-initialized GS are employed
as baselines in this subsection.
In this simulation, we compare the performance of AltGD,
HIO and GS under 2D Fourier measurements in the presence
of outliers, where GMM-type outliers are generated with
c2 = 0.1, σ12 = 0 and σ22 = 100. Each entry of the signal
matrix X ∈ R16×16 is generated from the normal distribution.
After zero-padding around X, we have X̃ ∈ R32×32 . Then the
measurements are obtained by first calculating 2D fast Fourier
transformation (FFT) on X̃ and then taking the magnitude
of 2D FFT(X̃), where the sizes of left and right Fourier
matrices are both 32 × 32. Note that SNR is 10 dB and is
computed as 10 log10 (kX̃k2F /kNk2F ) where N ∈ R32×32 is
the noise matrix. For AltGD and GS, HIO with 5000 iterations
is employed to provide an initial estimate of X, and then
an additional 5000 iterations are used in AltGD and GS,
respectively. Moreover, HIO with 10000 iterations is included
as a baseline. For AltGD, we choose p = 1.3. The MSEs
obtained from 1000 Monte-Carlo tests are plotted in Fig. 12.
It is seen that HIO does not perform very well, and is inferior
to AltGD and GS. AltGD has better performance than GS,
since most of its MSEs are concentrated around −35 dB, while
those obtained by GS are located around −20 dB.
Finally, Fig. 13 shows the recovered image of 128 × 128
Cameraman from 2D Fourier measurements under GMM
noise, where SNR is 12 dB and the noise is generated in the
same way as for Fig. 12. The number of iterations used for
10
20
20
20
40
40
40
60
60
60
80
80
80
100
100
100
120
120
20
40
60
80
100
120
120
20
40
(a) Original
60
80
100
120
20
(b) BI-AltGD
20
20
40
40
40
60
60
60
80
80
80
100
100
100
120
120
40
60
80
100
60
80
100
120
80
100
120
(c) AltGD
20
20
40
120
120
20
40
(d) TAF
60
80
100
120
(e) TWF
20
40
60
(f) MTWF
Fig. 11. Retrieving molecule image in SαS noise.
50
0
-40
AltGD
-30
-20
-10
0
10
20
30
MSE (dB)
50
0
-40
GS
-30
-20
-10
0
10
20
30
MSE (dB)
40
HIO
20
0
-40
-30
-20
-10
0
10
20
(a) Original
(b) AltGD
(c) GS
(d) HIO
30
MSE (dB)
Fig. 12. Signal recovery performance comparison with 2D Fourier measurements in GMM noise.
initialization and refinement are all 5000. For AltGD, we set
p = 0.6. It is seen that the image recovered from AltGD is
much more clear than those recovered by HIO and GS.
V. C ONCLUSION
In this paper, we considered phase retrieval in the presence
of grossly corrupted data – i.e., outliers. We formulated this
problem as an `p fitting problem, where 0 < `p < 2, and
provided an algorithmic framework that is based on two-block
Fig. 13. Retrieving Cameraman image in GMM noise with 2D Fourier
measurements.
11
inexact alternating optimization. Two algorithms, namely,
AltIRLS and AltGD, were proposed under this framework.
Although the algorithms cannot be analyzed using standard
convergence results for alternating optimization due to a
nonconvex constraint, we managed to show that the algorithms
converge to a KKT point. The tools used for convergence
analysis can also be used for analyzing convergence of other
types of algorithms that involve non-convex constraints and
inexact alternating optimization. Pertinent CRBs were derived
for the noisy measurement models considered. Simulations
showed that the proposed algorithms are promising in dealing
with outliers in the context of phase retrieval using both
random measurements and 2D Fourier measurements, for
simulated and real image data.
≥ g(x̃(r+1) , {ũ(r)
m })
≥
≥
(rj )
j)
g(x̃, {ũm
}) ≥ g(x̃(rj +1) , {ũ(r
m })
≥
≥
≥
Let us denote
=
f (x, u) =
2
|ym um − aH
m x| +
p/2
m=1
Therefore, our problem can be written as
min
f (x̃, {ũm }) =
kym ũm −
p/2
+
s.t. kũm k22 = 1, ∀m.
(r)
2
(r)
wm
kym ũ(r)
m − Ãm x̃k2 + φ(wm ) .
(39)
g(x̃, {ũ(r)
m })
= arg min
x̃
ũ(r+1)
= arg
m
f (x̃(r+1) , {ũm })
min
kũm k22 =1, ∀m
(40)
(41)
Let us assume the boundedness of x̃(r) and ũ(r) and
which will be shown later. Under this assumption, the
(r)
gradients of g(x̃, {ũm }) and f (x̃, {ũm }) exist. According to
Lemma 2.1, it is easily seen that
(r)
f (x̃, {ũ(r)
m }) ≤ g(x̃, {ũm }) ∀x̃,
f (x̃ , {ũ(r)
m })
(r)
∇x̃ f (x̃ , {ũ(r)
m })
=
=
(r)
g(x̃ , {ũ(r)
m })
(r)
∇x̃ g(x̃ , {ũ(r)
m }).
(42a)
(42b)
(42c)
Now, we have
f (x̃
(44e)
(47)
(48)
f (x̃? , {um }) ≥ f (x̃? , {u?m }).
(49)
ũ?m = 0, ∀m
= g(x̃
(r)
, {ũ(r)
m })
(43a)
(50)
where λm ∈ R2 are dual variables. Combining (47) and (50),
it follows that every limit point of the solution sequence is a
KKT point.
Now, we rigorously show the boundedness of x̃(r) and ũ(r)
and W(r) . Let us first show that W(r) is bounded given
finite x̃(r) and ũ(r) . This is relatively easy, since y and A
are always bounded and thus W(r) is also bounded when
0 < p < 2 and > 0 by definition (cf. Eq. (8)). Now, we
show that x̃(r) and ũ(r) are bounded. Since the algorithm is
essentially majorization minimization, there is no risk for x̃(r)
being unbounded, if the initialization is bounded, A has full
column rank, and y is bounded. To explain, one can see that
ũ(r) − Ãx̃(0) kpp ≥ ky
ũ(r) − Ãx̃(r) kpp
≥ |kykpp − kÃx̃(r) kpp |
, {ũ(r)
m })
(44d)
Therefore, ũ?m also satisfies the partial KKT condition
ky
(r)
(44c)
(rj )
f (x̃(rj ) , u) ≥ f (x̃(rj ) , {ũm
})
∇ũm f (x̃? , {ũ?m }) + λ?m
(r)
wm ,
(r)
(44b)
and thus
Now our algorithm can be re-expressed as
x̃
(44a)
(45)
∇x̃ f (x̃? , {ũ?m }) = 0.
m=1
(r+1)
j)
f (x̃
, {ũ(r
m })
j +1)
f (x̃(rj +1) , {ũ(r
})
m
(rj+1 )
(rj+1 )
f (x̃
, {ũm
})
(rj+1 )
(rj+1 )
g(x̃
, {ũm
}),
Similarly, by the update rule in (41), we have
(38)
Note that in the new expression (38), all the variables are realvalued. Accordingly, we may define
g(x̃, {ũ(r)
m }) =
(43d)
By (42c), we immediately have
m=1
M
X
(43c)
The inequality in (45) means that x̃? is blockwise minimizer of
g(x̃, {ũ?m }). Therefore, it satisfies the partial KKT condition
w.r.t. x̃, i.e.,
(46)
∇x̃ g(x̃? , {ũ?m }) = 0.
x̃ = [Re{xT }, Im{xT }]T ,
H
Re{aH
m }, −Im{am }
.
Ãm =
Im{aH
Re{aH
m },
m}
{ũm },x̃
(rj +1)
g(x̃, {ũ?m }) ≥ g(x̃? , u? ).
ũm = [Re{um }, Im{um }]T ,
Ãm x̃k22
(43b)
where (44d) is obtained by the fact that rj+1 ≥ rj + 1
since rj indexes a subsequence. Taking j → ∞, and by the
boundedness of W(r) and continuity of g(·), we see that
.
To simplify the analysis, let us define
M
X
f (x̃
, {ũ(r)
m })
(r+1)
f (x̃
, {ũ(r+1)
})
m
where (43a) follows (42b), (43b) is obtained because of (40),
(43c) holds since we have the property in (42a), and (43b) is
obtained by the fact that the subproblem w.r.t. u is optimally
solved, i.e., (41).
Assume that {rj } denotes the index set of a convergent sub(r )
sequence, and that {x̃(rj ) , {ũmj }} converges to (x̃? , {ũ?m })
Then, we have
A PPENDIX A
P ROOF OF P ROPOSITION 2.1
M
X
(r+1)
≥
kÃx̃(r) kpp
−
kykpp .
(51)
12
The first inequality holds because block upper bound minimization algorithm always decreases the objective [40]. The
second inequality is due to the triangle inequality for k · kpp
when 0 < p < 1 [39]; for p > 1, one can apply the same
argument to k · kp . From the above, one can see that
kÃx̃(r) kpp
≤ ∞.
Therefore, if A does not have a null space, then x̃(r) has to be
bounded. Consequently, W(r) is always bounded from above.
The next step is to show that the whole sequence converges
to K, the set of all the KKT points. This is relatively
straightforward. Let us assume that the whole sequence does
not converge to K. This means that there exists a subsequence
which does not converge to K. The boundedness of x̃(r) , ũ(r)
implies that every subsequence of {x̃(r) , ũ(r) } has a limit
point. We have just shown that every limit point is a KKT
point, and thus this is a contradiction. Therefore, the whole
sequence has to converge to K.
A PPENDIX B
CRB FOR L APLACIAN N OISE
The likelihood function for Laplacian noise is given by [36][37]
)
( √
M
Y
2
1
H
√
yi − |ai x|
(52)
exp −
p(y; x) =
σn
2σn
i=1
where the noise variance is
is
σn2 ,
and its log-likelihood function
√
ln p(y; x) = −M ln( 2σn ) −
√
M
2X
σn
yi − |aH
i x| . (53)
i=1
The vector of unknown parameters for complex-valued x is
β = [Re{x1 } · · · Re{xN }, Im{x1 } · · · Im{xN }]T . (54)
It is worth noting that σn2 is actually an unknown parameter
which should be considered as a part of β. However, since σn2
is uncorrelated with the real and imaginary parts of xi , their
mutual Fisher information is zero. It will not impact the final
CRB formula for x. For this reason, we do not include σn2 in
β. Thus, the FIM can be partitioned into four parts, i.e.,
FL,rr FL,ri
FL,c =
(55)
FL,ir FL,ii
where
[FL,c ]m,n = E
(
sgn(a) =
(56)
The partial derivative of ln p(y; x) with respective to βm is
√ M
2 X ∂ yi − |aH
∂ ln p(y; x)
i x|
=−
∂βm
σn i=1
∂βm
√ M
H
2 X yi − |aH
i x| ∂|ai x|
=
H
σn i=1 yi − |ai x| ∂βm
√ M
2X
∂|aH
i x|
=
sgn(yi − |aH
(57)
i x|)
σn i=1
∂βm
1,
a>0
−1,
a<0
(58)
and
[Re ai aH
i x ]m
, for βm = Re{xm }
H
H
|a x|
∂|ai x|
i H
=
∂βm
[Im ai ai x ]m
, for βm = Im{xm }.
|aH
i x|
(59)
Substituting (57) into (56), we have
[FL,c ]m,n =
M M
H
2 X X ∂|aH
i x| ∂|ai x|
σn2 i=1 j=1 ∂βm ∂βn
H
× E sgn(yi − |aH
i x|) · sgn(yj − |aj x|) .
(60)
Next we compute the value of E[sgn(yi − |aH
i x|)sgn(yj −
H
|aH
x|)].
For
notational
simplicity,
let
s
=
sgn(y
i
i − |ai x|).
j
It is obvious that when i = j, we have
E [si sj ] = 1.
(61)
For i 6= j, we first write
E [si sj ] = Pr(si = sj ) × (+1) + Pr(si 6= sj ) × (−1)
= 2Pr(si = sj ) − 1
(62)
where Pr standards for the probability. Then the value of
Pr(si = sj ) is computed as
Pr(si = sj ) = Pr(si = 1|sj = 1)Pr(sj = 1)
+ Pr(si = −1|sj = −1)Pr(sj = −1)
= Pr(si = 1)Pr(sj = 1)
+ Pr(si = −1)Pr(sj = −1)
= 0.5
(63)
where Pr(si = 1) = Pr(si = −1) = 0.5 and the second
equation follows by independence of si and sj when i 6= j.
Substituting (63) into (62) and using (61) yields
(
1, i = j
E [si sj ] =
(64)
0, i 6= j.
Substituting (64) into (60), we obtain
[FL,c ]m,n =
∂ ln p(y; x) ∂ ln p(y; x)
.
∂βm
∂βn
where
M
H
2 X ∂|aH
i x| ∂|ai x|
.
σn2 i=1 ∂βm ∂βn
(65)
Now using (59) and (65), the four sub-FIMs can be easily
derived as
2
FL,rr = 2 Re{AH diag(Ax)} · diag(|Ax|−2 )
σn
× Re{AH diag(Ax)}T
(66)
2
FL,ii = 2 Im{AH diag(Ax)} · diag(|Ax|−2 )
σn
× Im{AH diag(Ax)}T
(67)
2
FL,ri = 2 Re{AH diag(Ax)} · diag(|Ax|−2 )
σn
13
× Im{AH diag(Ax)}T
FL,ir =
FTL,ri .
(68)
(69)
Combining (66)-(69), we obtain
FL,c =
2
GL,c diag(|Ax|−2 ) GTL,c
σn2
(70)
Re{AH diag(Ax)}
=
.
Im{AH diag(Ax)}
(71)
where
GL,c
This completes the proof of Proposition 3.1.
A PPENDIX C
P ROOF OF R ANK P ROPERTY OF FL,c AND FL,r
We note here that FL,c is derived under the assumption
−2
of nonzero aH
) is full rank. As a
m x. Therefore, diag(|Ax|
result, computing the rank of FL,c is equivalent to computing
the rank of GL,c . To this end, define a nonzero vector v =
[ v1T v2T ]T ∈ R2N , which leads to
GTL,c v = Re{AH diag(Ax)}T v1 + Im{AH diag(Ax)}T v2 .
(72)
Now let u = v1 + jv2 , then
n
H o
GTL,c v = Re AH diag(Ax) u
= Re {(Ax)∗
(Au)}
(73)
which equals to zero if and only if
u = jx
(74)
n
o
2
GTL,c v = Re j |Ax| = 0.
(75)
i.e.,
This means that there is only one direction v =
[−Im{x}T Re{x}T ]T , which is non-zero, lies in the null space
of GL,c , thus also in the null space of FL,c .
In the real x case, similar to the proof of FL,c , it suffices to
show that there is no nonzero vector v ∈ RN making FL,r v =
0. It is easy to see that
GTL,r v = Re {(Ax)∗
(Av)} .
(76)
Since v is real-valued, it cannot be set to v = jx to make
(76) zero. Given a nontrivial A, it is impossible to find a v
such that Av = 0 except v = 0. Therefore, FL,r is full rank.
This completes the proof.
R EFERENCES
[1] C. Qian, X. Fu, N. D. Sidiropoulos, and L. Huang, “Inexact alternating optimization for phase retrieval with outliers,” Proceeding of 24th European
Signal Processing Conference (EUSIPCO), pp. 1538-1542, Budapest,
2016.
[2] R. Gerchberg and W. Saxton, “A practical algorithm for the determination
of phase from image and diffraction plane pictures,” Optik, vol. 35, pp.
237-246, 1972.
[3] J. R. Fienup, “Phase retrieval algorithms: A comparison,” Applied Optics,
vol. 21, no. 15, pp. 2758-2769, 1982.
[4] P. Netrapalli, P. Jain and S. Sanghavi, “Phase retrieval using alternating
minimization,” IEEE Trans. Signal Process., vol. 63, no. 18, pp 48144826, 2015.
[5] E. J. Candès, T. Strohmer, and V. Voroninski. “PhaseLift: Exact and stable
signal recovery from magnitude measurements via convex programming,”
Communications on Pure and Applied Mathematics, vol. 66, no. 8, pp.
1241-1274, 2013.
[6] P. Hand, “PhaseLift is robust to a constant fraction of arbitrary errors,”
Applied and Computational Harmonic Analysis, vol. 42, no. 3, pp. 550562, 2017.
[7] E. J. Candès, X. Li and M. Soltanolkotabi, “Phase retrieval via Wirtinger
Flow: Theory and algorithms,” IEEE Trans. Info. Theory, vol. 61, no. 4,
pp. 1985-2007, 2015.
[8] Y. Chen and E. Candès, “Solving random quadratic systems of equations is nearly as easy as solving linear systems,” Advances in Neural
Information Processing System, pp. 739-747, 2015.
[9] G. Wang, G. B. Giannakis, and Y. C. Eldar, “Solving systems of
random quadratic equations via truncated amplitude flow,” arXiv preprint
arXiv:1605.08285, 2016.
[10] G. Wang, G. B. Giannakis and J. Chen, “Scalable solvers of random
quadratic equations via stochastic truncated amplitude flow,” IEEE Trans.
Signal Process., vol. 65, no. 8, pp. 1961-1974, 2017.
[11] C. Qian, N. D. Sidiropoulos, K. Huang, L. Huang and H. C. So, “Phase
retrieval using feasible point pursuit: Algorithms and Cramér-Rao bound,”
IEEE Trans. Signal Process., vol. 64, no. 20, pp. 5282-5296, 2016.
[12] C. Qian, N. D. Sidiropoulos, K. Huang, L. Huang and H. C. So, “Least
squares phase retrieval using feasible point pursuit,” Proceedings of the
International Conference on Acoustics, Speech, and Signal Processing
(ICASSP 2016), Shanghai, China, pp. 4288-4292, 2016.
[13] E. J. Candès, Y. C. Eldar, T. Strohmer and V. Voroninski, “Phase retrieval
via matrix completion,” SIAM Review, vol. 57, no. 2 pp. 225-251, 2015.
[14] J. Sigl, “Nonlinear residual minimization by iteratively reweighted least
squares,” Computational Optimization and Applications, vol. 64, no. 3,
pp. 755-792, 2016.
[15] D. S. Weller, A. Pnueli, G. Divon, O. Radzyner, Y. C. Eldar and J.
A. Fessler, “Undersampled phase retrieval with outliers,” IEEE Trans.
Computational Imaging, vol. 1, no. 4, pp. 247-258, 2015.
[16] S. Chen and H. Lu, “Noise characteristic and its removal in digital radiographic system,” Proceeding of 15th World Conference on Nondestructive
Testing, Roma 2000.
[17] I. Frosio and N. A. Borghese, “Statistical based impulsive noise removal
in digital radiography,” IEEE Trans. Medical Imaging, vol. 28, no. 1, pp.
3-16, 2009.
[18] H. Zhang, Y. Chi and Y. Liang, “Provable non-convex phase retrieval
with outliers: Median truncated wirtinger flow,” Proceeding of International Conference on Machine Learning (ICML), New York, 2016.
[19] J. N. Cederquist and C. C. Wackerman, “Phase-retrieval error: A lower
bound,” Journal of the Optical Society of America A, vol. 4, no. 9 pp.
1788-1792, 1987.
[20] R. Balan, “The Fisher information matrix and the CRLB in a nonAWGN model for the phase retrieval problem,” Proceeding of 2015
Internat. Conf. on Sampl. Theory and Applications (SampTA), pp. 178182, Washington, DC, 2015.
[21] R. Balan, “Reconstruction of signals from magnitudes of redundant
representations: The complex case,” Foundations of Computational Mathematics, pp. 1-45, 2013.
[22] A. S. Bandeira, J. Cahill, D. G. Mixon and A. A. Nelson, “Saving phase:
Injectivity and stability for phase retrieval,” Applied and Computational
Harmonic Analysis, vol. 37, no. 1, pp. 106-125, 2014.
[23] X. Fu, K. Huang, B. Yang, W.-K. Ma and N.D. Sidiropoulos, “Robust
volume minimization-based matrix factorization for remote sensing and
document clustering,” IEEE Trans. Signal Process., vol. 64, no. 23, pp.
6254-6268, 2016.
[24] Y. Liu, Y. Dai and S. Ma, “Joint power and admission control: Nonconvex `q approximation and an effective polynomial time deflation
approach,” IEEE Trans. Signal Process., vol. 63, no. 14, pp. 3641-3656,
2015.
[25] R. Chartrand and W. Yin, “Iteratively reweighted algorithms for compressive sensing,” Proceeding of IEEE Internat. Conf. Acoust., Speech
and Signal Process., Las Vegas, NV, pp. 3869-3872, 2008.
[26] X. Fu, K. Huang, W.-K. Ma, N. D. Sidiropoulos, and R. Bro, “Joint
tensor factorization and outlying slab suppression with applications,”
IEEE Trans. Signal Process., vol. 63, no. 23, pp. 6315-6328, 2015.
[27] A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding algorithm for linear inverse problems,” SIAM journal on Imaging Sciences,
vol.2, no.1, pp. 183–202, 2009.
[28] Y. Dai and Y. Yuan, “A nonlinear conjugate gradient method with a
strong global convergence property,” SIAM Journal on Optimization, vol.
10, no. 1, pp. 177-182, 1999.
14
[29] Y. Nesterov, “A method for unconstrained convex minimization problem
with the rate of convergence O (1/k2 ),” Doklady an SSSR, vol. 269, no.3,
pp. 543-547, 1983.
[30] O. Fercoq and P. Richtàrik, “Accelerated, parallel, and proximal coordinate descent,” SIAM Journal on Optimization, vol. 25, no. 4, pp. 19972023, 2015.
[31] Y. Nesterov, “Efficiency of coordinate descent methods on huge-scale
optimization problems,” SIAM Journal on Optimization, vol. 22, no. 2
pp. 341-362, 2012.
[32] Y. Xu, R. Hao, W. Yin, and Z. Su, “Parallel matrix factorization for
low-rank tensor completion,” Inverse Problems and Imaging, vol. 9, no.
2, pp. 601-624, 2015.
[33] P. Stoica and T. L. Marzetta, “Parameter estimation problems with
singular information matrices,” IEEE Trans. Signal Process., vol. 49, no.
1, pp. 87-90, 2001.
[34] A. O. Hero, III, J. A. Fessler and M. Usman, “Exploring estimator
bias-variance tradeoffs using the uniform CR bound,” IEEE Trans. Signal
Process., vol. 44, pp. 2026-2041, 1996.
[35] K. Huang and N. D. Sidiropoulos, “Putting nonnegative matrix factorization to the test: A tutorial derivation of pertinent Cramér-Rao bounds and
performance benchmarking,” IEEE Signal Processing Magazine, Special
Issue on Source Separation and Applications, vol. 31, no. 3, pp. 76-86,
2014.
[36] S. A. Vorobyov, Y. Rong, N. D. Sidiropoulos and A. B. Gershman,
“Robust iterative fitting of multilinear models,” IEEE Trans. Signal
Process., vol. 53, no. 8, pp. 2678-2689, 2005.
[37] S. M. Kay, Fundamentals of statistical signal processing: Detection
theory, Upper Saddle River, NJ: Prentice-Hall, 1998.
[38] S. Basu and Y. Bresler, “The stability of nonlinear least squares problems
and the Cramer-Rao bound,” IEEE Transactions on Signal Processing,
vol. 48, no. 12, pp. 3426-3436, 2000.
[39] R. Chartrand and V. Staneva, “Restricted isometry properties and
nonconvex compressive sensing,” Inverse Problems, vol. 24, no. 3, pp.
035020, 2008.
[40] M. Razaviyayn, M. Hong, and Z.-Q. Luo, “A unified convergence
analysis of block successive minimization methods for nonsmooth optimization,” SIAM Journal on Optimization, vol. 23, no. 2, pp. 1126-1153,
2013.
| 7cs.IT
|
Semantics for probabilistic programming: higher-order
functions, continuous distributions, and soft constraints
Sam Staton Hongseok Yang
Frank Wood
Chris Heunen
Ohad Kammar
University of Edinburgh
University of Cambridge
arXiv:1601.04943v3 [cs.PL] 4 May 2016
University of Oxford
Abstract
a distribution defined by normalizing another program. This is
called a nested query, by analogy with database programming.
We study the semantic foundation of expressive probabilistic programming languages, that support higher-order functions, continuous distributions, and soft constraints (such as Anglican, Church,
and Venture). We define a metalanguage (an idealised version of
Anglican) for probabilistic computation with the above features,
develop both operational and denotational semantics, and prove
soundness, adequacy, and termination. This involves measure theory, stochastic labelled transition systems, and functor categories,
but admits intuitive computational readings, one of which views
sampled random variables as dynamically allocated read-only variables. We apply our semantics to validate nontrivial equations underlying the correctness of certain compiler optimisations and inference algorithms such as sequential Monte Carlo simulation. The
language enables defining probability distributions on higher-order
functions, and we study their properties.
Here is a simple example of a program. We write gauss (µ, σ)
for the Gaussian probability distribution whose density function is
2
density gauss(a, (µ, σ)) = σ √12π exp(− (a−µ)
).
2σ 2
1
2
3
4
norm(
let x = sample(gauss (0.0, 3.0)) in
score(density gauss(5.0, (x, 1.0));
return(x < 4.5))
(1)
Line 2 samples x from a prior Gaussian distribution. The soft constraint on Line 3 expresses the likelihood of the observed data, 5.0,
coming from a Gaussian given the prior x. Line 4 says that what we
are actually interested in is a boolean random variable over the sample space. Line 1 calculates a posterior distribution for the return
value, using the prior and the likelihood. In this example we can
precisely calculate that the posterior distribution on {true, false}
has p(true) = 0.5.
Languages like this currently lack formal exact semantics. The
aim of this paper is to provide just such a foundation as a basis
for formal reasoning. Most expressive probabilistic programming
languages are explained in terms of their Monte Carlo simulation
algorithms. The simplest such algorithm, using importance and
rejection sampling, is the de facto semantics against which other
algorithms are ‘proved approximately correct’. Such ‘semantics’
are hard to handle and extend.
We provide two styles of semantics, operational and denotational. For first-order probabilistic programs, the denotational semantics is straightforward: types are interpreted as measurable
spaces, and terms are interpreted as measurable functions (§4). Operational semantics is more complicated. For discrete distributions,
an operational semantics might be a probabilistic transition system,
but for continuous distributions, it must be a stochastic relation (labelled Markov process). We resolve this by equipping the set of
configurations with the structure of a measurable space (§5).
The advantage to the operational semantics is that it is easily
extended to higher-order programs (§7). Denotational semantics for
higher-order programs poses a problem, because measurable spaces
do not support the usual β/η theory of functions: they do not form
a Cartesian closed category (indeed, RR does not exist as a measurable space [3]). Earlier work dealt with this either by excluding higher-order functions or by considering only discrete distributions. We resolve this by moving from the category of measurable
spaces, where standard probability theory takes place, to a functor category based on it (§8). The former embeds in the latter, so
we can still interpret first-order concepts. But the functor category
does have well-behaved function spaces, so we can also interpret
higher-order concepts. Moreover, by lifting the monad of probabil-
1. Introduction
Probabilistic programming is the idea to use programs to specify
probabilistic models; probabilistic programming languages blend
programming constructs with probabilistic primitives. This helps
scientists express complicated models succinctly. Moreover, such
languages come with generic inference algorithms, relieving the
programmers of the nontrivial task of (algorithmically) answering
queries about their probabilistic models. This is useful in e.g. machine learning.
Several higher-order probabilistic programming languages have
recently attracted a substantial user base. Some languages (such
as Infer.net [22], PyMC [27], and Stan [36]) are less expressive
but provide powerful inference algorithms, while others (such as
Anglican [37], Church [13], and Venture [21]) have less efficient
inference algorithms but more expressive power. We consider the
more expressive languages, that support higher-order functions,
continuous distributions, and soft constraints. More precisely, we
consider a programming language (§3) with higher-order functions
(§6) as well as the following probabilistic primitives.
Sampling The command sample(t) draws a sample from a distribution described by t, which may range over the real numbers.
Soft constraints The command score(t) puts a score t (a positive
real number) on the current execution trace. This is typically
used to record that some particular datum was observed as being
drawn from a particular distribution; the score describes how
surprising the observation is.
Normalization The command norm(u) runs a simulation algorithm over the program fragment u. This takes the scores into
account and returns a new, normalized probability distribution.
The argument to sample might be a primitive distribution, or
1
2016/5/5
Finally, for a predicate ϕ, we use the indicator expression [ϕ] to
denote 1 if ϕ holds, and 0 otherwise.
ity distributions [12] to the functor category, we can also interpret
continuous distributions. Finally, we can interpret observations by
considering probability distributions with continuous density, irrespective of the categorical machinery (§9).
The denotational semantics is sound and adequate with respect
to the operational semantics (§5.3,8.3), which means one can use
the denotational model to directly check program equations while
respecting computational issues. For example:
3. A first-order language
This section presents a first-order language for expressing Bayesian
probabilistic models. The language forms a first-order core of a
higher-order extension in Section 6, and provides a simpler setting
to illustrate key ideas. The language includes infinitary type and
term constructors, constant terms for all measurable functions between measurable spaces, and constructs for specifying Bayesian
probabilistic models, namely, operations for sampling distributions,
scoring samples, and normalizing distributions based on scores.
This highly permissive and slightly unusual syntax is not meant
to be a useful programming language itself. Rather, its purpose is
to serve as a semantic metalanguage to which a practical programming language compiles, and to provide a mathematical setting for
studying high-level constructs for probabilistic computation.
• we demonstrate a key program equation for sequential Monte
Carlo simulation (§4.1);
• we show that every term of first-order type is equal to one
without λ-abstractions or application, and hence is interpreted
as a measurable function (Proposition 8.3).
2. Preliminaries
We recall basic definitions and facts of measure theory.
Definition 2.1. A σ-algebra on a set X is a family Σ of subsets
of X, called measurable (sub)sets, which contains X and is closed
under complements and countable unions. A measurable space is a
set with a σ-algebra.
A measure on a measurable space
S (X, Σ) is
Pa function p : Σ →
[0, ∞] such that p(∅) = 0 and p( i∈N Ui ) = i∈N p(Ui ) for each
sequence U1 , U2 , . . . of disjoint measurable sets. A probability
measure or probability distribution is a measure p with p(X) = 1.
Types
The language has types
A, B ::= R | P(A) | 1 | A × B |
X
Ai
i∈I
where I ranges over countable sets. A type A stands for a measurable space JAK. For example, R denotes the measurable space
of reals, P(A) is the space of probability measures on A, and 1
is the (discrete) measurable space on the singleton set. The other
type constructors correspond to products and sums of measurable
spaces. Notice that countable sums are allowed, enabling us to express usual ground types in programming languages via standard
encoding. For instance,
the type for booleans is 1 + 1, and that for
P
natural numbers i∈N 1.
In the paper we use a few important constructions for measurable spaces. The first is to make a set X into a measurable space by
taking the full powerset of X as Σ, yielding a discrete measurable
space. When X is countable, a probability distribution on (X, Σ)
is determined by its values on singleton
P sets, that is, by specifying
a function p : X → [0, 1] such that x∈X p(x) = 1.
The second construction is to combine a collection of measurable spaces (Xi , ΣP
i )i∈I by sum or product.
Q The underlying sets are
the disjoint union i∈I XiP
and product i∈I Xi of sets. The measurable sets in the sum are i∈I Ui for Ui ∈ Σi . The σ-algebra
of
Q
the product is the smallest one containing all the subsets i∈I Ui
where Ui ∈ Σi equals Xi but for a single index i.
For the third, the real numbers form a measurable space (R, ΣR )
under the smallest σ-algebra that contains the open intervals; the
measurable sets are called Borel sets. Restricting to any measurable
subset gives a new measurable space, such as the space R≥0 of
nonnegative reals and the unit interval [0, 1].
The fourth construction is to make the set P (X) of all probability measures on a measurable space (X, ΣX ) into a measurable
space, by letting ΣP (X) be the smallest σ-algebra containing the
sets {p ∈ P (X) | p(U ) ∈ V } for all U ∈ ΣX and V ∈ Σ[0,1] .
Terms We distinguish typing judgements: Γ ⊢d t : A for deterministic terms, and Γ ⊢p t : A for probabilistic terms (see also
e.g. [20, 26, 31]). In both, A is a type, and Γ is a list of variable/type
pairs. Variables stand for deterministic terms.
Intuitively, a probabilistic term Γ ⊢p t : A may have two kinds
of effects: during evaluation, t may sample from a probability
distribution, and it may score the current execution trace (typically
according to likelihood of data). Evaluating a deterministic term
Γ ⊢d t : A does not have any effects.
Sums and products The language includes variables, and standard constructors and destructors for sum and product types.
Γ ⊢d t : Ai
P
Γ, x : A, Γ′ ⊢d x : A
Γ ⊢d (i, t) :
i∈I Ai
P
Γ ⊢d t :
(Γ, x : Ai ⊢z ui : B)i∈I
i∈I Ai
(z ∈ {d, p})
Γ ⊢z case t of {(i, x) ⇒ ui }i∈I : B
Definition 2.2. Let (X, ΣX ), (Y, ΣY ) be measurable spaces. A
function f : X → Y is measurable if f -1 (U ) ∈ ΣX for U ∈ ΣY .
We can push forward a measure along a measurable function:
if p : ΣX → [0, 1] is a probability measure on (X, ΣX ) and
f : X → Y is a measurable function, then q(U ) = p(f -1 (U ))
is a probability measure on (Y, ΣY ).
Γ ⊢d ∗ : 1
Γ ⊢d tj : Aj for all j ∈ {0, 1}
Γ ⊢d (t0 , t1 ) : A0 × A1
Γ ⊢d t : A0 × A1
Γ ⊢d πj (t) : Aj
In the rules for sums, I may be infinite. In the last rule, j is 0 or 1.
We use some standard syntactic sugar, such as false and true for the
injections in the type bool = 1 + 1, and if for case in that instance.
Definition 2.3. A stochastic relation between measurable spaces
(X, ΣX ) and (Y, ΣY ) is a function r : X × ΣY → [0, 1] such that
r(x, −) : ΣY → [0, 1] is a probability distribution for all x ∈ X,
and r(−, V ) : X → [0, 1] is measurable for all V ∈ ΣY .
Sequencing We include the standard constructs (e.g. [20, 23]).
Γ ⊢p t : A Γ, x : A ⊢p u : B
Γ ⊢d t : A
Γ ⊢p return(t) : A
Γ ⊢p let x = t in u : B
Giving a stochastic relation from (X, ΣX ) to (Y, ΣY ) is equivalent to giving a measurable function (X, ΣX ) → (P (Y ), ΣP (Y ) ).
Stochastic relations r : X × ΣY → [0, 1] and s : Y × ΣZ → [0, 1]
compose associatively to (s ◦ r) : X × ΣZ → [0, 1] via the formula
Z
(s ◦ r)(x, W ) =
s(y, W ) r(x, dy).
Where A = 1, we write (t; u) for let x = t in u.
Language-specific constructs The language has constant terms
for all measurable functions.
Γ ⊢d t : A
(f : JAK → JBK measurable)
(2)
Γ ⊢d f (t) : B
Y
2
2016/5/5
If the model evidence is 0 or ∞, the conversion fails, and this is
tracked by the ‘+1+1’.
In particular, all the usual distributions are in the language, including the Dirac distribution dirac(x) concentrated on outcome
x, the Gaussian distribution gauss(µ, σ) with mean µ and standard deviation σ, the Bernoulli distribution bern(p) with success
probability p, the exponential distribution exp(r) with rate r, and
the Beta distribution beta(α, β) with parameters α, β.1 For example, from the measurable functions 42.0 : 1 → R, e(−) : R → R,
gauss : R × R → P (R) and < : R × R → 1 + 1 we can derive:
Γ ⊢d 42.0 : R
4. Denotational semantics
This section discusses the natural denotational semantics of the
first-order language. The basic idea can be traced back a long
way (e.g. [18]) but our treatment of score and norm appear to be
novel. As described, types A are interpreted as measurable spaces
JAK. A context Γ = (x1Q
: A1 , . . . , xn : An ) is interpreted as the
def
measurable space JΓK = n
i=1 JAi K of its valuations.
Γ ⊢d t : R
Γ ⊢d et : R
Γ ⊢d µ : R Γ ⊢d σ : R
Γ ⊢d t : R Γ ⊢d u : R
Γ ⊢d gauss (µ, σ) : P(R)
Γ ⊢d t < u : bool
The following terms form the core of our language.
Γ ⊢d t : P(A)
Γ ⊢p sample(t) : A
• Deterministic terms Γ ⊢d t : A are interpreted as measurable
functions JtK : JΓK → JAK, providing a result for each valuation
of the context.
• Probabilistic terms Γ ⊢p t : A are interpreted as measurable
Γ ⊢d t : R
Γ ⊢p score(t) : 1
functions JtK : JΓK → P (R≥0 × JAK), providing a probability
measure on (score,result) pairs for each valuation of the context.
The first term samples a value from a distribution t. The second
multiplies the score of the current trace with t. Since both of these
terms express effects, they are typed under ⊢p instead of ⊢d .
The reader may think of the score of the current execution trace
as a state, but it cannot be read, nor changed arbitrarily: it can
only be multiplied by another score. The argument t in score(t) is
usually the density of a probability distribution at an observed data
point. For instance, recall the example (1) from the Introduction:
Informally, if (Ω, p : Ω→[0, 1]) is the prior sample space of the program (specified by sample), and l : Ω→R≥0 is the likelihood (specified by score), and r : Ω → JAK is the return value, then a measure
in P (R≥0 × JAK) is found by pushing forward p along (l, r).
Sums and products The interpretation of deterministic terms follows the usual pattern of the internal language of a distributive catdef
egory (e.g. [28]). For instance, JΓ, x : A, Γ′ ⊢d x : AK(γ, a, γ ′ ) = a,
def
and JΓ ⊢d f (t) : AK(γ) = f (JtK(γ)) for measurable f : JAK → JBK.
This interpretation is actually the same as the usual set-theoretic
semantics of the calculus, as one can show by induction that the
induced functions JΓK → JAK are measurable.
norm(let x = sample(gauss(0.0, 3.0)) in
score(density gauss(5.0, (x, 1.0)); return(x < 4.5))
An execution trace is scored higher when x is closer to the datum 5.
When the argument t in score(t) is 0, this is called a hard
constraint, meaning ‘reject this trace’; otherwise it is called a soft
constraint. In the discrete setting, hard constraints are more-or-less
sufficient, but even then, soft constraints tend to be more efficient.
Sequencing For probabilistic terms, we proceed as follows.
def
Jreturn(t)K(γ)(U ) = [(1, JtK(γ)) ∈ U ],
which denotes a Dirac distribution, and Jlet x = t in uK(γ)(V ) is
Z
JtK(γ)(d(r, x)) .
JuK(γ, x) (s, b) (r · s, b) ∈ V
Normalization Two representative tasks of Bayesian inference
are to calculate the so-called posterior distribution and model evidence from a prior distribution and likelihood. Programs built from
sample and score can be thought of as setting up a prior and a likelihood. Consider the following program:
R≥0 ×JAK
As we will explain shortly, these interpretations come from treating
P (R≥0 × (−)) as a commutative monad, which essentially means
the following program equations hold.
let x = sample(bern(0.25)) in
let y = (if x then return 5.0 else return 2.0) in
score(density exp(0.0, y)); return(x)
Jlet x = return(x) in uK = JuK
Jlet x = t in return(x)K = JtK
Jlet y = (let x = t in u) in vK = Jlet x = t in let y = u in vK
Here the prior y comes from the Bernoulli distribution, and the
likelihood concerns datum 0.0 coming from an exponential distribution with rate y. Recall that density exp(0.0, y) = y. So there
are two execution traces, returning either true, with probability
0.25 and score 5.0, or false, with probability 0.75 and score 2.0.
The product of the prior and likelihood gives an unnormalized
posterior distribution on the return value: (true 7→ 0.25·5=1.25,
false 7→ 0.75·2=1.5). The normalizing constant is the average
score: (0.25·5 + 0.75·2) = 2.75, so the posterior is (true 7→
1.5
1.25
≈0.45, false 7→ 2.75
≈0.55). The average score is called the
2.75
model evidence. It is a measure of how well the model encoded by
the program matches the observation.
Note that the sample x = true matches the datum better, so the
probability of true goes up from 0.25 to 0.45 in the posterior.
In our language we have a term norm(t) that will usually convert a probabilistic term t into a deterministic value, which is its
posterior distribution together with the model evidence.
Γ ⊢p t : A
Γ ⊢d norm(t) : (R × P(A)) + 1 + 1
Jlet x = t in let y = u in (x, y)K = Jlet y = u in let x = t in (x, y)K
The last equation justifies a useful program optimisation technique [37, §5.5]. (The last two equations require assumptions about
free variables of u, v and t, to avoid variable capture.)
Language-specific constructs We use the monad:
def
Jsample(t)K(γ)(U ) = JtK(γ)({a | (1, a) ∈ U })
def
Jscore(t)K(γ)(U ) = [(max(JtK(γ), 0), ∗) ∈ U ]
Here are some program equations to illustrate the semantics so far.
Jscore(7.0); score(6.1)K = Jscore(42.7)K
{
let x = sample(gauss (0.0, 1.0))
= Jsample(bern(0.5))K
in return(x > 0.0)
s
{
let x = sample(gauss (0.0, 1.0))
= Jreturn(false)K
in return(x > x)
s
Normalization We interpret norm(t) as follows. Each probability measure p on (R≥0 × X)R induces an unnormalized posterior
measure p̄ on X: let p̄(U ) = R ×U r p(d(r, x)). As long as the
≥0
average score p̄(X) is not 0 or ∞, we can normalize p̄ to build
1 Usually, gauss(0.0, 0.0) is undefined, but we just let gauss(0.0, 0.0) =
gauss(0.0, 1.0), and so on, to avoid worrying about this sort of error.
3
2016/5/5
a posterior probability measure on X. This construction forms a
natural transformation ιX : P (R≥0 × X) → (R × P (X)) + 1 + 1
if p̄(X) = 0
(1, ∗)
def
(2,
∗)
ιX (p) =
if p̄(X) = ∞ (3)
0, p̄(X), λU. p̄(U )
otherwise
p̄(X)
the resampled value in the continuation v. Line 3 propagates the
error of 0: w is a deterministic term of the right type whose choice
does not matter. Finally, line 4 detects an infinite evidence error,
and undoes the transformation. This error does not arise in most
applications of sequential Monte Carlo simulation.
5. Operational semantics
def
Let Jnorm(t)K(γ) = ι(JtK(γ)). Here are some examples:
s
{
norm(let x = sample(gauss (0.0, 3.0)) in
score(density gauss(5.0, (x, 1.0)); return(x < 4.5))
= 0, (0.949, bern (0.5))
{
s
norm let x = sample(bern(0.25)) in
(if x then score(5.0) else score(2.0)); return(x)
5
))
= 0, (2.75, bern( 11
Jnorm let x = sample(exp(1.0)) in score(ex ) K = (2, ∗)
{ s
s
{
1
norm let x = sample(beta(1,
3)) = norm score( 1+3 );
in score(x); return(x)
sample(beta(2, 3))
In this section we develop an operational semantics for the firstorder language. There are several reasons to consider this, even
though the denotational semantics is arguably straightforward.
First, extension to higher-order functions is easier in operational
semantics than in denotational semantics. Second, operational semantics conveys computational intuitions that are obscured in the
denotational semantics. We expect these computational intuitions
to play an important role in studying approximate techniques for
performing posterior inference, such as sequential Monte Carlo, in
the future.
Sampling from probability distributions complicates operational semantics. Sampling from a discrete distribution can immediately affect control flow. For example, in the term
The third equation shows how infinite model evidence errors can
arise when working with infinite distributions. In the last equation,
the parameter x of score(x) represents the probability of true
under bern(x). The equation expresses the so called conjugateprior relationship between Beta and Bernoulli distributions, which
has been used to optimise probabilistic programs [38].
let x = sample(bern(0.5)) in if x then return(1.1) else return(8.0)
the conditional depends on the result of sampling the Bernoulli
distribution. The result is 1.1 with probability 0.5 (cf. [5, §2.3]).
Sampling a distribution on R introduces another complication.
Informally, there is a transition
Monads The interpretation of letand return given above arises
from the fact that P R≥0 × (−) is a commutative monad on
the category of measurable spaces and measurable functions (see
also [7, §2.3.1], [33, §6]). Recall that a commutative monad
(T, η, µ, σ) in general comprises an endofunctor T together with
natural transformations ηX : X → T (X), µX : T (T (X)) →
T (X), σX,Y : X ×T (Y ) → T (X ×Y ) satisfying some laws [17].
Using this structure we interpret return and let following Moggi [23]:
def
JΓ ⊢p return(t) : AK(γ) = ηJAK JtK(γ)
def
JΓ ⊢p let x = t in u : BK(γ) = µJBK T (JuK) σJΓK,JAK (γ, JtK(γ))
sample(gauss (0.0, 1.0)) −→ return(r)
for every real r, but any single transition has zero probability. We
can assign non-zero probabilities to sets of transitions; informally:
Pr sample(gauss (0.0, 1.0)) −→ {return(r) | r ≤ 0} = 0.5.
To make this precise we need a σ-algebra on the set of terms, which
can be done using configurations rather than individual terms.
A configuration is a closure: a pair ht, γi of a term t with free
variables and an environment γ giving values for those variables
as elements of a measurable space. (See also [14, §3], [6, §3].)
Sampling a distribution p on R+R exhibits both complications:
Concretely, we make P (R≥0 × (−)) into a monad by combining
the standard commutative monad structure [12] on P and the commutative monoid (R≥0 , ·, 1) with the monoid monad transformer.
(We record a subtle point about ι (3). It is not a monad morphism, and we we cannot form a commutative monad of all measures because the Fubini property does not hold in general.)
let x = sample(p) in case x of (0, r) ⇒ return(r + 1.0)
|(1, r) ⇒ return(r − 1.0)
The control flow in the case distinction depends on which summand is sampled, but there is potentially a continuous distribution
over the return values. We handle this by instantiating the choice
of summand in the syntax, but keeping the value of the summand
in the environment, so that expression (4) can make a step to the
closure
4.1 Sequential Monte Carlo simulation
The program equations above justify some simple program transformations. For a more sophisticated one, consider sequential
Monte Carlo simulation [8]. A key idea of its application to probabilistic programs is: ‘Whenever there is a score, it is good to
renormalize and resample’. This increases efficiency by avoiding
too many program executions with low scores [25, Algorithm 1].
The denotational semantics justifies the soundness of this transformation. For a term with top-level score, i.e. a term of the form
(let x = t in (score(u); v)) where u and v may have x free:
Jnorm(let x = t in (score(u); v))K
= Jnorm(case (norm(let x = t in score(u); return(x))) of
(0, (e, d)) ⇒ score(e); let x = sample(d) in v
| (1, ∗) ⇒ score(0); return(w)
| (2, ∗) ⇒ let x = t in (score(u); v))K
(4)
, y 7→ 42.0i.
h let x = return(0, y) in
case x of (0, r) ⇒ return(r + 1.0)
|(1, r) ⇒ return(r − 1.0)
A type is indecomposable if it has the form R or P(A), and a
context Γ is canonical if it only involves indecomposable types.
Configurations Let z ∈ {d, p}. A z-configuration of type A is a
triple hΓ, t, γi comprising a canonical context Γ, a term Γ ⊢z t : A,
and an element γ of the measurable space JΓK. We identify contexts
that merely rename variables, such as
1
2
3
4
h(x : R, y : P(R)), f (x, y), (x 7→ 42.0, y →
7 gauss (0.0, 1.0))i
≈h(u : R, v : P(R)), f (u, v), (u 7→ 42.0, v →
7 gauss(0.0, 1.0))i.
Let us explain the right hand side. Line 1 renormalizes the program
after the score, and in non-exceptional execution returns the model
evidence e and a new normalized distribution d. Line 2 immediately
records the evidence e as a score, and then resamples d, using
We call d-configurations deterministic configurations, and p-configurations probabilistic configurations; they differ only in typing. We
will abbreviate configurations to ht, γi when Γ is obvious.
4
2016/5/5
Values v in a canonical context Γ are well-typed deterministic
terms of the form
v, w ::= xi | ∗ | (v, w) | (i, v)
parameterized by a family of measurable ‘normalization’ functions
νA : Conp (A) → R≥0 × P (JAK) + 1 + 1
(8)
(5)
indexed by types A.
where xi is a variable in Γ. Similarly, a probabilistic term t in
context Γ is called probabilistic value or p-value if t ≡ return(v0 )
for some value v0 . Remember from Section 4 that the denotational
semantics of values is simple and straightforward.
Write Cond (A) and Conp (A) for the sets of deterministic and
probabilistic configurations of type A, and make them into measurable spaces by declaring U ⊆ Conz (A) to be measurable if the set
{γ ∈ JΓK | ht, γi ∈ U } is measurable for all judgements Γ ⊢z t : A.
X
JΓK
(6)
Conz (A) =
Reduction of deterministic terms Define a type-indexed family
of relations −→ ⊆ ConNd (A) × Cond (A) as the least one that is
closed under the following rules.
hΓ, πj (v0 , v1 ), γi −→ hΓ, vj , γi
hΓ, case (i′ , v) of {(i, x) ⇒ ti }i∈I , γi −→ hΓ, ti′ [v/x], γi
hΓ, f (w), γi −→ h(Γ, Γ′ ), v, (γ, γ ′ )i
(w a value ∧ Γ′ ⊢d v : A an ordered value ∧ f (JwK(γ)) = JvK(γ ′ ))
(Γ,t)
Γ canonical,
Γ ⊢z t : A
hΓ, norm(t), γi −→ h(Γ, x:R, y:P(B)), (0, (x, y)), γ[x7→r, y7→p]i
(A = (R×P(B) + 1 + 1) ∧ νB (hΓ, t, γi) = (0, (r, p)) ∧ x, y 6∈ Γ)
Further partition Conz (A) into ConVz (A) and ConNz (A)
based on whether a term in a configuration is a value or not:
ConVd (A) = {hΓ, t, γi ∈ Cond (A)
ConNd (A) = {hΓ, t, γi ∈ Cond (A)
ConVp (A) = {hΓ, t, γi ∈ Conp (A)
ConNp (A) = {hΓ, t, γi ∈ Conp (A)
|
|
|
|
hΓ, norm(t), γi −→ hΓ, (i, ∗), γi
(A = (R × P(B) + 1 + 1) ∧ νB (hΓ, t, γi) = (i, ∗), i ∈ {1, 2})
t is a value}
t is not a value}
t is a p-value}
t is not a p-value}
hΓ, t, γi −→ hΓ′ , t′ , γ ′ i
hΓ, C[t], γi −→ hΓ′ , C[t′ ], γ ′ i
(C[−] is not (−))
Particularly well-behaved values are the ordered values Γ ⊢d v : A,
where each variable appears exactly once, and in the same order as
in Γ.
The rule for f (w) keeps the original context Γ and the closure γ
because they might be used in the continuation, even though they
are not used in v. The rules obey the following invariant.
Lemma 5.1. Consider a canonical context Γ, a type A, an ordered
value Γ ⊢d v : A, and the induced measurable function
Lemma 5.2. If hΓ, t, γi −→ hΓ′ , t′ , γ ′ i, then Γ′ = (Γ, Γ′′ ) and
γ ′ = (γ, γ ′′ ) for some Γ′′ and γ ′′ ∈ JΓ′′ K.
JvK : JΓK → JAK.
Proof. By induction on the structure of derivations.
The collection of all such functions for given A is countable, and
forms a coproduct diagram.
This lemma allows us to confirm that our specification of a relation
−→ ⊆ ConNd (A) × Cond (A) is well-formed (‘type preservation’).
Proof. By induction on the structure of types. The key fact is that
every type is a sum of products of indecomposable ones, because
the category
of measurable spaces
P
P is distributive, i.e. the canonical
map i∈I (A × Bi ) → A × i∈I Bi is an isomorphism.
Proposition 5.3. The induced relation is a measurable function.
Proof. There are three things to show: that the relation is entire
(‘progress’); that the relation is single-valued (‘determinacy’); and
that the induced function is measurable. All three are shown by
induction on the structure of terms. The case of application of
measurable functions crucially uses Lemma 5.1.
For example, A = (R × bool) + (R × R) has 3 ordered values,
first (x : R ⊢d (0, (x, true)) : A), second (x : R ⊢d (0, (x, false)) : A),
and third (x : R, y : R ⊢d (1, (x, y)) : A), inducing a canonical measurable isomorphism R + R + R × R ∼
= JAK.
Evaluation contexts We distinguish three kinds of evaluation
contexts: C[−] is a context for a deterministic term with a hole for
deterministic terms; D[−] and E [−] are contexts for probabilistic
terms, the former with a hole for probabilistic terms, the latter with
a hole for deterministic terms.
C[−] ::= (−) | πj C[−] | (C[−], t) | (v, C[−]) | (i, C[−])
| case C[−] of {(i, x) ⇒ ti }i∈I | f (C[−])
D[−] ::= (−) | let x = D[−] in t
Reduction of probabilistic terms Next, we define the stochastic
relation −→ for probabilistic terms, combining two standard approaches: for indecomposable types, which are uncountable, use
labelled Markov processes, i.e. give a distribution on the measurable set of resulting configurations; for decomposable types (sums,
products etc.), probabilistic branching is discrete and so a transition
system labelled by probabilities suffices.
(7)
Proposition 5.4. Let (Xi )i∈I be an indexed family of measurable
spaces. Suppose we are given:
E [−] ::= D[return[−]] | D[sample[−]] | D[score[−]]
| case D[−] of {(i, x) ⇒ ti }i∈I
• a function q : I → [0, 1] that
Pis only nonzero on a countable
subset I0 ⊆ I, and such that
where t, ti are general terms and v is a value.
i∈I0
q(i) = 1;
• a probability measure qi on Xi for each i ∈ I0 .
5.1 Reduction
P
This determines a probability measure p on i∈I Xi by
P
p(U ) = i∈I q(i) qi ({a | (i, a) ∈ U })
P
for U a measurable subset of i∈I Xi ,
Using the tools developed so far, we will define a measurable
function for describing the reduction of d-configurations, and a
stochastic relation for describing reduction of p-configurations:
−→ : ConNd (A) → Cond (A),
−→ : ConNp (A) × ΣR≥0 ×Conp (A) → [0, 1],
We will use three entities to define the desired stochastic relation
−→ : ConNp (A) × ΣR≥0 ×Conp (A) → [0, 1].
5
2016/5/5
(Γ′ , t′ ) and the distribution Pr(C −→(−) (−)) indexed by such
pairs satisfy the conditions in Proposition 5.4. It follows that the
partial evaluation Pr(C −→ (−)) of the function in the statement
is a probability measure, so it suffices to establish measurability of
the other partial evaluation Pr((−) −→ U ). Recall that ConNp (A)
is defined in terms of the sum of measurable spaces, and that all
p-configurations in each summand have the same term. Finally,
use induction on the term shared by all p-configurations in the
summand to see that the restriction of Pr((−) −→ U ) to each
summand is measurable.
1. A countably supported probability distribution on the set
{(Γ, t) | Γ ⊢p t : A} for each C ∈ ConNp (A). We write
Pr(C −→ (Γ, t)) for the probability of (Γ, t).
2. A probability measure on the space JΓK for each C ∈ ConNp (A)
and (Γ, t) with Pr(C −→ (Γ, t)) 6= 0. Write Pr(C −→Γ,t U )
for the probability of a measurable subset U ⊆ JΓK.
3. A measurable function Sc : ConNp (A) → R≥0 , representing
the score of the one-step transition relation. (For one-step transitions, the score is actually deterministic.)
These three entities are defined by induction on the structure of the
syntax of A-typed p-configurations in Figure 1. We combine them
to define a stochastic relation as follows.
5.2 Termination
To see that the reduction process terminates, we first define the
transitive closure. This is subtle, as sampling can introduce countably infinite branching; although each branch will terminate, the
expected number of steps to termination can be infinite.
We use the deterministic transition relation to define an evaluation relation ⇓ ⊆ Cond (A) × ConV d (A), by setting C ⇓ D if
∃n. C ⇓n D, where
Proposition 5.5. The map ConNp (A) × ΣR≥0 ×Conp (A) → [0, 1]
that sends (C, U ) to Pr(C −→ U ), defined as
X
Pr C −→ (Γ, t) Pr C −→Γ,t {γ | (Sc(C), hΓ, t, γi) ∈ U } ,
(Γ,t)
is a stochastic relation.
0
C⇓ C
on t
Proof. For each p-configuration C = h , t, i, use induction
to see that the probability distribution Pr C −→ (−) on pairs
(C ∈ ConVd (A))
D ⇓n E
C⇓
C −→ D
n+1
E
To define evaluation for probabilistic configurations, we need substochastic relations: functions f : X × ΣY → [0, 1] that are
measurable in X, satisfy f (x, Y ) ≤ S
1 for every xP∈ X, and
are countably additive in Y , i.e. f (x, i∈N Ui ) =
i∈N f (Ui )
for a sequence U1 , U2 , . . . of disjoint measurable sets. Thus a
stochastic relation (as in Definition 2.3) is a sub-stochastic relation
with f (x, Y ) = 1. Define a sub-stochastic relation
def
Pr(hΓ, E [t], γi −→ (Γ′ , E [t′ ])) = hΓ, t, γi −→ hΓ′ , t′ , γ ′ i
def
Pr(hΓ, D[t], γi −→ (Γ′ , D[t′ ])) = Pr(hΓ, t, γi −→ (Γ′ , t′ ))
def
Pr(hΓ, let x = return(v) in t, γi −→ (Γ, t[v/x])) = 1
Pr(− ⇓ −) : Conp (A) × ΣR≥0 ×ConVp (A) → [0, 1]
def P
by Pr(C ⇓ U ) = n Pr(C ⇓n U ), where Pr(C ⇓0 U ) is given
by [(1, C) ∈ U ], and Pr(C ⇓n+1 U ) is
Z
Pr(D ⇓n {(s, E) | (r · s, E) ∈ U }) Pr(C −→ d(r, D)).
def
Pr(hΓ, case (j, v) of {(i, x) ⇒ ti }i∈I , γi −→ (Γ, tj [v/x])) = 1
def
Pr(hΓ, score(v), γi −→ (Γ, return(∗))) = 1
Pr(hΓ, sample(v), γi −→ ((Γ, Γ′ ), return(v ′ )))
(
JvK(γ)({Jv ′ K(γ ′ ) | γ ′ ∈ JΓ′ K}) if Γ′ ⊢d v ′ : A ordered
def
=
0
otherwise
R≥0 ×Conp (A)
Proposition 5.6 (Termination). Evaluation of deterministic terms
is a function: ∀C. ∃D. C ⇓ D. Evaluation of probabilistic terms is
a stochastic relation: ∀C. Pr(C ⇓ (R≥0 × ConV p (A))) = 1.
Pr(hΓ, E [t], γi −→(Γ′ ,E[t′ ]) U )
def
= [hΓ, t, γi −→ hΓ′ , t′ , γ ′ i ∧ γ ′ ∈ U ]
Proof. By induction on the structure of terms.
Termination comes from the omission of recursion in our language. We do so for now, because our semantic model does not yet
handle higher-order recursion, and probabilistic while-languages
are already well-understood (e.g. [18]). (See also the discussion
about domain theory in §8.)
def
Pr(hΓ, D[t], γi −→(Γ′ ,D[t′ ]) U ) = Pr(hΓ, t, γi −→(Γ′ ,t′ ) U )
def
Pr(hΓ, let x = return(v) in t, γi −→(Γ,t[v/x]) U ) = [γ ∈ U ]
def
Pr(hΓ, case (j, v) of {(i, x) ⇒ ti }i∈I , γi −→Γ,tj [v/x] U ) = [γ ∈ U ]
def
Pr(hΓ, score(v), γi −→(Γ,return(∗)) U ) = [γ ∈ U ]
5.3 Soundness
Pr(hΓ, sample(v), γi −→((Γ,Γ′ ),return(v′ )) U )
For soundness, extend the denotational semantics to configurations:
def
=
′
′
′
′
• define sd : Cond (A) → JAK by hΓ, t, γi 7→ JtK(γ);
′
JvK(γ)({Jv K(γ ) | γ ∈ JΓ K ∧ (γ, γ ) ∈ U })
JvK(γ)({Jv ′ K(γ ′ ) | γ ′ ∈ JΓ′ K})
def
• define sp : Conp (A) × ΣR≥0 ×JAK → [0, 1] similarly by
(hΓ, t, γi, U ) 7→ JtK(γ)(U ). We may also use this stochastic relation as a measurable function sp : Conp (A) → P (R≥0 × JAK);
def
Sc(hΓ, E [t], γi) = 1
Sc(hΓ, D[t], γi) = Sc(hΓ, t, γi)
• define sV p : ConVp (A) → JAK by hΓ, return(v), γi 7→ JvK(γ).
def
Sc(hΓ, sample(v), γi) = 1
Note that in this first-order language, sV p is a surjection which
equates two value configurations iff they are related by weakening, contraction or exchange of variables.
def
Sc(hΓ, score(v), γi) = max(JvK(γ), 0)
def
Sc(hΓ, let x = return(v) in t, γi) = 1
Assumption 5.7. Throughout this section we assume that the normalization function ν on configurations (8) is perfect, i.e. it corresponds to ι, the semantic normalization function (3):
def
Sc(hΓ, case (j, v) of {(i, x) ⇒ ti }i∈I , γi) = 1
νA (hΓ, t, γi) = ιJAK (sp (hΓ, t, γi)).
Figure 1. Entities used to define reduction of probabilistic terms
6
2016/5/5
Lemma 5.8 (Context extension). Let z ∈ {d, p}. Suppose that
hΓ, t, γi and h(Γ, Γ′ ), t, (γ, γ ′ )i are configurations in Conz (A).
Then sz (hΓ, t, γi) = sz h(Γ, Γ′ ), t, (γ, γ ′ )i.
which takes a probability measure and returns a suspended program
that will sample from it. On the other hand, to reify the normalization construction, we use a different calling convention.
Proposition 5.9 (Soundness). The following diagrams commute
(in the category of measurable functions, and stochastic relations,
respectively).
⊢d λx. norm(force(x)) : T(A) ⇒ R × P(A) + 1 + 1
ConNp (A)
ConNd (A)
one-step
reduction
Cond (A)
❯❯s❯d❯*
4 JAK
✐✐✐✐✐
sd
one-step
reduction
sp
This function takes a suspended probabilistic program and returns
the result of normalizing it.
/ R≥0 × JAK
O
Example: higher-order expectation Higher types also allow us
to consider probability distributions over programs. For an example, consider this term.
multiplication
/ R≥0 × R≥0 × JAK
R≥0 × Conp (A)
id ×s
def
Eh = λ(d, f ) : T(A) × (A ⇒ R).
case (norm(let a = force(d) in score(f (a)))) of
(0, (e, y)) ⇒ e
| (1, ∗) ⇒ 0.0 | (2, ∗) ⇒ 0.0
p
Proof. By induction on the structure of syntax. The inductive steps
with evaluation contexts use the extension Lemma 5.8, which applies by the invariant Lemma 5.2.
It has type (T(A) × (A ⇒ R)) ⇒ R. Intuitively, given a thunked
probabilistic term t and a function f that is nonnegative, Eh treats
t as a probability distribution on A, and computes the expectation
of f on this distribution. Notice that A can be a higher type, so
Eh generalises the usual notion of expectation, which has not been
defined for higher types because the category of measurable spaces
is not Cartesian closed.
Adequacy
The denotational semantics is adequate, in the sense:
JtK(∗) = P (R≥0 × sV p ) Pr(h∅, t, ∗i ⇓ (−)) for all ⊢p t : A.
That is, the denotation JtK(∗) is nothing but pushing forward the
probability measure Pr(h∅, t, ∗i ⇓ (−)) from the operational semantics along the function sV p . This adequacy condition holds because Proposition 5.9 ensures that
X
Pr(h∅, t, ∗i ⇓k {(r, C) | (r, sV p (C)) ∈ U }) ≤ JtK(∗)(U )
7. Higher-order operational semantics
In this section we consider operational semantics for the higherorder extension of the language. In an operational intuition, force(t)
forces a suspended computation t to run. For example,
k≤n
for all n and U , and Proposition 5.6 then guarantees that the lefthand side of this inequality converges to the right-hand side as n
tends to infinity.
⊢d thunk(sample(gauss(0.0, 1.0))) : T(R)
is a suspended computation that, when forced, will sample the
normal distribution.
6. A higher-order language
Assumption 7.1. From the operational perspective it is unclear
how to deal with sampling from a distribution over functions. For
this reason, in this section, we only allow the type P(A) when A is a
measurable type. We still allow probabilistic terms to have higherorder types, and we still allow T(A) where A is higher-order.
This short section extends the first-order language with functions
and thunks [19], allowing variables to stand for program fragments.
In other words, ‘programs are first-class citizens’.
Types
Extend the grammar for types with two new constructors.
X
A, B ::= R | P(A) | 1 | A × B |
Ai | A ⇒ B | T(A)
7.1 Reduction
We now extend the operational semantics from Section 5 with
higher types. Values (5) are extended as follows.
i∈I
Informally, A ⇒ B contains deterministic functions, and T(A)
contains thunked (i.e. suspended) probabilistic programs. Then
A ⇒ T(B) contains probabilistic functions. A type is measurable
if it does not involve ⇒ or T, i.e. if it is in the grammar of Section 3.
v ::= . . . | λx.t | thunk(t)
Evaluation contexts (7) are extended as follows.
C[−] ::= . . . | C[−] t | v C[−]
Terms Extend the term language with the following rules. First,
the usual abstraction and application of deterministic functions:
Γ, x : A ⊢d t : B
Γ ⊢d λx. t : A ⇒ B
There are two additional redexes: (λx.t) v and force(thunk(t)).
The deterministic transition relation is extended with this β-rule:
Γ ⊢d t : A ⇒ B Γ ⊢d u : A
Γ ⊢d t u : B
hΓ, (λx.t) v, γi −→ hΓ, t[v/x], γi.
Extend the probabilistic transition relation with the following rules.
Pr hΓ, force(thunk(t)), γi −→ (Γ, t) = 1
Pr hΓ, force(thunk(t)), γi −→(Γ,t) U = [γ ∈ U ]
Second, we have syntax for thunking and forcing (e.g. [19, 23, 26]).
Γ ⊢p t : A
Γ ⊢d thunk(t) : T(A)
E [−] ::= . . . | D[force[−]]
Γ ⊢d t : T(A)
Γ ⊢p force(t) : A
Sc(hΓ, force(thunk(t)), γi) = 1
All the rules from Section 3 are also still in force. For rule (2) to
still make sense, we only include constant terms for measurable
functions f : JAK → JBK between measurable types A and B.
7.2 Termination
The evaluation relations for deterministic and probabilistic configurations of the higher-order language are defined as in Subsection 5.2. The resulting rewriting system still terminates, even
though configurations may now include higher-order terms.
Examples One motivation for higher types is to support code
structuring. The separate function types and thunk types allow us
to be flexible about calling conventions. For example, sampling can
be reified as the ground term
Proposition 7.2 (Termination). Evaluation of deterministic terms
is a function: ∀C. ∃D. C ⇓ D. Evaluation of probabilistic terms is
a stochastic relation: ∀C. Pr(C ⇓ (R≥0 × ConV p (A))) = 1.
⊢d λx. thunk(sample(x)) : P(A) ⇒ T(A),
7
2016/5/5
Proof. We sketch an invariant of higher-order terms that implies the
termination property, formulated as unary logical relations via sets
8.1 Semantic model
Fix a category Meas of measurable spaces and measurable functions that is essentially small but large enough for the purposes of
Section 4. For example, Meas could be the category of standard
Borel spaces [4, 35]: one can show that JAK is standard Borel by induction on A, and the class of all standard Borel spaces is countable
up to measurable isomorphism.
In Section 4 we interpreted first-order types A as measurable spaces JAK. We will interpret higher-order types A as functors LAM : Measop → Set. Informally, when A is a first-order
type and Γ is a first-order context, we will have LAM(JΓK) ∼
=
Meas(JΓK, JAK) ≈ {t | Γ ⊢d t : A}. For a second order type
(A ⇒ B), we will have
∼ Meas(JΓK × JAK, JBK) ≈ {t | Γ, x : A ⊢d t : B}
LA ⇒ BM(JΓK) =
R(Γ ⊢z A) ⊆ {t | Γ ⊢z t : A},
Rv (Γ ⊢z A) ⊆ {t | Γ ⊢z t : A ∧ t a z-value},
for each canonical context Γ, type A, and z ∈ {d, p}, defined by:
R(Γ ⊢d A) = {t | ∀γ. hΓ, t, γi ⇓ hΓ′ , t′ , γ ′ i ∧ t′ ∈ Rv (Γ′ ⊢d A)}
R(Γ ⊢p A) = {t | ∀γ. Pr(hΓ, t, γi ⇓
P
(R≥0 × Γ′ Rv (Γ′ ⊢p A) × JΓ′ K)) = 1}
Rv (Γ ⊢p A) = {return(v) | v ∈ Rv (Γ ⊢d A)}
Rv (Γ ⊢d A) = {x | (x : A) ∈ Γ}
for A indecomposable
Rv (Γ ⊢d 1) = {∗}
Rv (Γ ⊢d A1 × A2 ) = {(v1 , v2 ) | ∀j. vj ∈ Rv (Γ ⊢d Aj )}
X
Rv (Γ ⊢d
Ai ) = {(i, v) | v ∈ Rv (Γ ⊢d Ai )}
so that β/η equality is built in. To put it another way, LAM(Rn ) models terms of type A having n read-only real-valued memory cells.
Lemma 8.1. For a small category C with countable sums, consider the category C of countable-product-preserving functors
Cop → Set, and natural transformations between them.
Rv (Γ ⊢d T (A)) = {thunk(t) | t ∈ R(Γ ⊢p A)}
Rv (Γ ⊢d A ⇒ B) = {λx.t | ∀Γ′ ⊇ Γ, u ∈ Rv (Γ′ ⊢d A).
• C has all colimits;
• C is Cartesian closed if C has products that distribute over
t[u/x] ∈ R(Γ′ ⊢d B)}
Induction on the structure of a term Γ, x1 : A1 , . . . xn : An ⊢z t : B
for z ∈ {d, p} now proves that vi ∈ Rv (Γ ⊢d Ai ) for i = 1, . . . , n
implies t[~v /~x] ∈ R(Γ ⊢z B).
sums;
• There is a full and faithful embedding y : C → C, given by
def
y(c) = C(−, c), which preserves limits and countable sums.
Proof. See e.g. [30, §7], or [16, Theorems 5.56 and 6.25]. The
embedding y is called the Yoneda embedding.
8. Higher-order denotational semantics
This section gives denotational semantics for the higher-order language, without using Assumption 7.1. We are to interpret the new
constructs T(A), thunk, and force. We will interpret probabilistic judgements as Kleisli morphisms JΓK → T (JAK) for a certain
def
monad T , and set JT(A)K = T (JAK), so thunk and force embody
the correspondence of maps JΓK → T (JAK) and JΓK → JT(A)K.
On which category can the monad T live? Interpreting λabstraction and application needs a natural ‘currying’ bijection between morphisms JΓK × R → R and morphisms JΓK → JR ⇒ RK.
But measurable functions cannot do this: it is known that no measurable space JR ⇒ RK can support such a bijection [3].
We resolve the problem of function spaces by embedding the
category of measurable spaces in a larger one, where currying
is possible, and that still has the structure to interpret the first
order language as before. As the larger category we will take a
category of functors Measop → Set from a category Meas
of measurable spaces and measurable functions to the category
Set of sets and functions. This idea arises from two traditions.
First, we can think of a variable of type R as a read-only memory
cell, as in the operational semantics, and functor categories have
long been used to model local memory (e.g. [24]). Second, the
standard construction for building a Cartesian closed category out
of a distributive one is based on functor categories (e.g. [30]).
For a simple example, consider the category CSet of countable
sets and functions. It has countable sums and finite products, but is
not Cartesian closed. Because every countable set is a countable
sum of singletons, the category CSet is equivalent to Set.
Our semantics for the higher-order language will take place in
the category Meas. Note that products in Meas are pointwise,
e.g. (F × G)(X) = F (X) × G(X) for all F, G ∈ Meas and all
X ∈ Meas, but sums are not pointwise, e.g. (1 + 1) ∈ Meas
is the functor that assigns a measurable space X to the set of its
measurable subsets. This is essential for y to preserve sums.
Distribution types We have to interpret distribution types P(A)
in our functor category Meas. How can we interpret a probability
distribution on the type R ⇒ R? We can answer this pragmatically,
without putting σ-algebra structure on the set of all functions. If
JR ⇒ RK were a measurable space, a random variable valued
in JR ⇒ RK would be given by a measurable space (X, ΣX ),
a probability distribution on it, and a measurable function X →
JR ⇒ RK. Despite there being no such measurable space JR ⇒ RK,
we can speak of uncurried measurable functions X ×R → R. Thus
we might define a probability distribution on JR ⇒ RK to be a triple
(X, ΣX ), f : X × R → R, p : ΣX → [0, 1]
Other models of higher-order programs Semantics of higherorder languages with discrete probability are understood well. For
terminating programs, there are set-theoretic models based on a distributions monad, and for full recursion one can use probabilistic
powerdomains [15] or coherence spaces [10]. It is also plausible
one could model continuous distributions in domain theory, since
it supports computable real analysis (e.g. [9]); this could be interesting because computability is subtle for probabilistic programming (e.g. [1]). Nonetheless, we contend it is often helpful to abstract away computability issues when studying probabilistic programming languages, to have access to standard theorems of probability theory to justify program transformations.
of a measurable space (X, ΣX ) of ‘codes’, a measurable function f
where we think of f (x, r) as ‘the function coded x evaluated at r’,
and a probability distribution p on the codes. These triples should
be considered modulo renaming the codes. This is exactly the
notion of probability distribution that arises in our functor category.
Lemma 8.2. For a small category C with countable sums:
• any functor F : C → C extends to a functor F : C → C
satisfying F ◦ y ∼
= y ◦ F , given by
X
G(a) × C b, F (a) / ∼
F (G)(b) =
a
8
2016/5/5
where the equivalence relation ∼ is the least one satisfying
(a′ , x, F g ◦ f ) ∼ (a, Gg(x), f );
• similarly, any functor F : C × C → C in two arguments
extends to a functor F : C×C → C, with F ◦(y×y) ∼
= y◦F :
P
G(a)
×
H(b)
×
C
c,
F
(a,
b)
F (G, H)(c) =
/∼
a,b
A subtle point is that configuration spaces (6) involve uncountable sums: the set of terms of a given type is uncountable, but y
only preserves countable sums. This is not really a problem because
only countably many terms are reachable from a given program.
Definition 8.4. For a type A, the binary reachability relation ∗d
on {(Γ, t) | Γ ⊢d t : A ∧ Γ canonical} is the least reflexive and transitive relation with (Γ, t) ∗d (Γ′ , u) if hΓ, t, γi −→ hΓ′ , u, γ ′ i
for γ ∈ JΓK, γ ′ ∈ JΓ′ K. Similarly, ∗p is the least reflexive
and transitive relation on {(Γ, t) | Γ ⊢p t : A∧ Γ canonical} with
(Γ, t) ∗p (Γ′ , u) if Pr hΓ, t, γi −→ (Γ′ , u) 6= 0 for γ ∈ JΓK.
→ G between functors
F, G : C → C lifts to a natural transformation α : F → G,
and similarly for functors C × C → C;
• and this is functorial, i.e. G ◦ F ∼
= G ◦ F and β ◦ α = β ◦ α.
• any natural transformation α : F
Proposition 8.5. Let z ∈ {d, p}. For any closed term ⊢z t : A, the
set of reachable terms {(Γ, u) | (∅, t) ∗z (Γ, u)} is countable.
Proof. F (G) is the left Kan extension of G along F , see e.g. [16].
Direct calculation shows F (G) preserves products if G does.
Proof. One-step reachability is countable by induction on terms.
Since all programs terminate by Proposition 7.2, the reachable
terms form a countably branching well-founded tree.
Thus the commutative monads P and T = P (R≥0 × (−)) on
Meas lift to commutative monads P and T ∼
= P (yR≥0 × (−))
on Meas. The latter monad captures the informal importancesampling semantics advocated by the designers of Anglican [37].
We may thus restrict to the configurations built from a countable
set U of terms that is closed under subterms and reachability.
Extend the denotational semantics in Meas to configurations by
defining sd : y(Cond (A)) → LAM, sp : y(Conp (A)) → T LAM, and
sV p : y(ConVp (A)) → T LAM; use the isomorphisms
X
LΓM
y(Conz (A)) ∼
=
8.2 Conservativity
We interpret the types of the higher order language as objects in
Meas using its categorical structure.
P
def P
def
def
L i∈I Ai M = i∈I LAi M
LRM = yR LP(A)M = P (LAM)
def
LA × BM = LAM × LBM
def
LA ⇒ BM = LAM ⇒ LBM
def
(Γ⊢z t:A)∈U
def
LT(A)M = T (LAM)
L1M = 1
to define sd , sp , sV p by copairing the interpretation morphisms
LΓ ⊢d t : AM : LΓM → LAM and LΓ ⊢p u : AM : LΓM → T (LAM).
We extend this interpretation to contexts.
def
Lx1 : A1 , . . . , xn : An M =
Proposition 8.6 (Soundness). The following diagrams commute.
Qn
i=1 LAi M
y(ConNd (A))
Deterministic terms Γ ⊢d t : A are interpreted as natural transformations LΓM → LAM in Meas, and probabilistic terms Γ ⊢p t : A as
natural transformations LΓM → T LAM, by induction on the structure of terms as in Section 4. Application and λ-abstraction are
interpreted as usual in Cartesian closed categories [28]. Thunk and
force are trivial from the perspective of the denotational semantics,
because LTAM = T LAM. To interpret norm(t), use Lemma 8.2 to
extend the normalization functions T (X) → R≥0 × P (X) + 1 + 1
between measurable spaces (3) to natural transformations T (F ) →
y(R≥0 ) × P (F ) + 1 + 1.
Note that all measurable types A have a natural isomorphism
LAM ∼
= yJAK. This interpretation conserves the first-order semantics of Section 4:
y(reduction)
y(Cond (A))
❱❱s❱d❱+
3 LAM
❤❤❤s❤❤
y(ConNp (A))
y(reduction)
d
T (Conp (A))
❱❱s❱p❱+
3 T LAM
❤❤❤❤
µ·T (sp )
Adequacy It follows that the higher denotational semantics remains adequate, in the sense that for all probabilistic terms ⊢p t : A,
LtM1 (∗) = (T (sV p ))1 Pr(h∅, t, ∗i ⇓ (−)) .
Adequacy is usually only stated at first-order types. At first-order
types A the function sV p does very little, since global elements
of LAM correspond bijectively with value configurations modulo
weakening, contraction and exchange in the context. At higher
types, the corollary still holds, but sV p is not so trivial because we
do not reduce under thunk or λ. (See also [29].)
Proposition 8.3. For z ∈ {d, p}, and first-order Γ and A:
9. Continuous densities
• for first-order Γ ⊢z t, u : A, JtK = JuK if and only if LtM = LuM;
• every term Γ ⊢z t : A has LtM = LuM for a first-order Γ ⊢z u : A.
In most examples, the argument t to score(t) is a density function
for a probability distribution. When scores are based on density
functions, this makes the relationship between score and sample
tighter than we have expressed so far. Our language easily extends
to accommodate such distributions. We just add a collection density
types to the syntax.
Proof. We treat z = d; the other case is similar. By induction on the
structure of terms, LtM = yJtK : LΓM → LAM. The first point follows
from faithfulness of y; the second from fullness and (2).
One interesting corollary is that the interpretation of a term of
first-order type is always a measurable function, even if the term involves thunking, λ-abstraction and application. This corollary gives
a partial answer to a question by Park et. al. on the measurability
of all λ-definable ground-type terms in probabilistic programs [26]
(partial because our language does not include recursion).
D ::= R | bool | N | 1 | D × D
A ::= · · · | D(D)
The D in this grammar denotes a measurable space that: (i) carries
a separable metrisable topology that generates the σ-algebra; and
that (ii) comes with a chosen σ-finite measure µD . An example
is R with its usual Euclidean topology and the Lebesgue measure
(which maps each interval to its size).
The type D(D) denotes a measurable
R space JD(D)K of continuous functions f : JDK → R≥0 with X f dµD = 1. The σ-algebra
of JD(D)K is the least one making {f | f (x) ≤ r} measurable for
all (x, r) ∈ JDK × R≥0 .
8.3 Soundness
The same recipe as in Section 5.3 will show that the higher-order
denotational semantics is sound and adequate with respect to the
higher-order operational semantics. This needs Assumption 7.1.
9
2016/5/5
A density type D(D) comes with two measurable functions
ev : JD(D)K × JDK → R≥0
dist : JD(D)K → JP(D)K
[7] E.-E. Doberkat. Stochastic Relations: Foundations for Markov Transition Systems. Chapman & Hall, 2007.
[8] A. Doucet, N. de Freitas, and N. Gordon. Sequential Monte Carlo
Methods in Practice. Springer, 2001.
[9] A. Edalat and M. H. Escardó. Integration in Real PCF. Inf. Comput. ,
160:128–166, 2000.
[10] T. Ehrhard, C. Tasson, and M. Pagani. Probabilistic coherence spaces
are fully abstract for probabilistic PCF. In POPL, 2014.
[11] C. E. Freer and D. M. Roy. Computable de Finetti measures. Ann.
Pure Appl. Logic, 163(5):530–546, 2012.
[12] M. Giry. A categorical approach to probability theory. Categorical
Aspects of Topology and Analysis, 915:68–85, 1982.
[13] N. Goodman, V. Mansinghka, D. M. Roy, K. Bonawitz, and J. B.
Tenenbaum. Church: a language for generative models. In UAI, 2008.
[14] C. Hur, A. V. Nori, S. K. Rajamani, and S. Samuel. A provably correct
sampler for probabilistic programs. In FSTTCS, 2015.
[15] C. Jones and G. D. Plotkin. A probabilistic powerdomain of evaluations. In LiCS, 1989.
[16] G. M. Kelly. Basic concepts of enriched category theory. CUP, 1982.
[17] A. Kock. Monads on symmetric monoidal closed categories. Archiv
der Mathematik, XXI:1–10, 1970.
[18] D. Kozen. Semantics of probablistic programs. Journal of Computer
and System Sciences, 22:328–350, 1981.
[19] P. B. Levy. Call-by-push-value: A subsuming paradigm. In TLCA’02.
[20] P. B. Levy, J. Power, and H. Thielecke. Modelling environments in
call-by-value programming languages. Inf. Comput., 185(2), 2003.
[21] V. K. Mansinghka, D. Selsam, and Y. N. Perov. Venture: a higher-order
probabilistic programming platform with programmable inference.
2014. URL http://arxiv.org/abs/1404.0099.
[22] T. Minka, J. Winn, J. Guiver, and D. Knowles. Infer.NET 2.4, Microsoft Research Cambridge, 2010.
[23] E. Moggi. Notions of computation and monads. Inf. Comput., 93(1):
55–92, 1991.
[24] F. J. Oles. Type algebras, functor categories and block structure. In
Algebraic methods in semantics. CUP, 1984.
[25] B. Paige and F. Wood. A compilation target for probabilistic programming languages. In ICML, 2014.
[26] S. Park, F. Pfenning, and S. Thrun. A probabilistic language based on
sampling functions. ACM TOPLAS, 31(1):171–182, 2008.
[27] A. Patil, D. Huard, and C. J. Fonnesbeck. PyMC: Bayesian stochastic
modelling in Python. Journal of Statistical Software, 35, 2010.
[28] A. M. Pitts. Categorical logic. In Handbook of Logic in Computer
Science, volume 5. OUP, 2000.
[29] G. D. Plotkin and J. Power. Adequacy for algebraic effects. In
FOSSACS, 2001.
[30] J. Power. Generic models for computational effects. TCS, 364(2):
254–269, 2006.
[31] N. Ramsey and A. Pfeffer. Stochastic lambda calculus and monads of
probability distributions. In POPL, 2002.
[32] D. M. Roy, V. Mansinghka, N. Goodman, and J. Tenenbaum. A
stochastic programming perspective on nonparametric Bayes. In
ICML Workshop on Nonparametric Bayesian, 2008.
[33] Adam Scibior, Zoubin Ghahramani, Andrew D. Gordon: Practical
probabilistic programming with monads. Haskell 2015.
[34] C.-C. Shan and N. Ramsey. Symbolic Bayesian inference by lazy
partial evaluation, 2016.
[35] S. M. Srivastava. A course on Borel sets. Springer, 1998.
[36] Stan Development Team. Stan: A C++ library for probability and
sampling, version 2.5.0, 2014. URL http://mc-stan.org/.
[37] F. Wood, J. W. van de Meent, and V. Mansinghka. A new approach to
probabilistic programming inference. In AISTATS, 2014.
[38] H. Yang. Program transformation for probabilistic programs, 2015.
Presentation at DALI.
ev (f, x) = f (x)
Z
dist (f )(U ) =
f dµD
U
Note that measurability of ev relies on continuity of the densities [3]. The usual way of imposing a soft constraint based
on a likelihood can be encoded in our first-order language as
score(ev (f, x)), where the datum x is observed with a probability
distribution with density f . Thus the categorical machinery used
to interpret higher-order functions is not needed for such soft constraints.
There is a limit to this use of continuity: probability measures
produced by norm(t) need not have continuous density. For example, norm(return(42.0)) produces a discontinuous Dirac measure.
Probability densities are often used by importance samplers.
The importance sampler generates samples of f ∈ JD(R)K by first
sampling from a proposal distribution g where sampling is easy,
and then normalizing those samples x from g according to their
importance weight f (x)/g(x). This works provided the support of
g is R, e.g. g = density gauss(x, (0.0, 1.0)).
Jnorm(sample(dist (f )))K =
r
z
norm(let x = sample(dist(g)) in score( ev(f,x)
); return(x))
ev(g,x)
Density types can be incorporated into the higher order language straightforwardly. The only subtlety is that denotational semantics now needs the base category to contain JD(D)K.
10. Conclusion and future work
We have defined a metalanguage for higher-order probabilistic programs with continuous distributions and soft constraints, and presented operational and denotational semantics, together with useful program equations justified by the semantics. One interesting next step is to use these tools to study other old or new language features and concepts (such as recursion, function memoisation [32], measure-zero conditioning [5], disintegration [2, 34], and
exchangeability [11, 21, 37]) that have been experimented with in
the context of probabilistic programming. Another future direction
is to formulate and prove the correctness of inference algorithms,
especially those based on Monte Carlo simulation, following [14].
Acknowledgments
We thank T. Avery, I. Garnier, T. Le, K. Sturtz, and A. Westerbaan.
This work was suppoted by the EPSRC, a Royal Society University
Fellowship, An Institute for Information & communications Technology Promotion (IITP) grant funded by the Korea government
(MSIP, No. R0190-15-2011), DARPA PPAML, and the ERC grant
‘causality and symmetry — the next-generation semantics’.
References
[1] N. L. Ackerman, C. E. Freer, and D. M. Roy. Noncomputable conditional distributions. In LiCS, 2011.
[2] N. L. Ackerman, C. E. Freer, and D. M. Roy. On computability and
disintegration, 2015. URL http://arxiv.org/abs/1509.02992.
[3] R. J. Aumann. Borel structures for function spaces. Illinois Journal of
Mathematics, 5:614–630, 1961.
[4] S. K. Berberian. Borel spaces. World Scientific, 1988.
[5] J. Borgström, A. D. Gordon, M. Greenberg, J. Margetson, and J. van
Gael. Measure transformer semantics for Bayesian machine learning.
LMCS, 9(3):11, 2013.
[6] J. Borgström, U. Dal Lago, A. D. Gordon, M. Szymczak. A
lambda-calculus foundation for universal probabilistic programming.
arxiv:1512.08990. 2015.
10
2016/5/5
| 6cs.PL
|
arXiv:1602.04484v5 [cs.LG] 19 Apr 2017
Surprising properties of dropout
in deep networks
David P. Helmbold
Department of Computer Science
University of California, Santa Cruz
Santa Cruz, CA 95064, USA
[email protected]
Philip M. Long
Google
[email protected]
April 21, 2017
Abstract
We analyze dropout in deep networks with rectified linear units and the quadratic loss. Our results
expose surprising differences between the behavior of dropout and more traditional regularizers like weight
decay. For example, on some simple data sets dropout training produces negative weights even though the
output is the sum of the inputs. This provides a counterpoint to the suggestion that dropout discourages
co-adaptation of weights. We also show that the dropout penalty can grow exponentially in the depth of
the network while the weight-decay penalty remains essentially linear, and that dropout is insensitive to
various re-scalings of the input features, outputs, and network weights. This last insensitivity implies that
there are no isolated local minima of the dropout training criterion. Our work uncovers new properties
of dropout, extends our understanding of why dropout succeeds, and lays the foundation for further
progress.
1
Introduction
The 2012 ImageNet Large Scale Visual Recognition challenge was won by the University of Toronto team by
a surprisingly large margin. In an invited talk at NIPS, Hinton [15] credited the dropout training technique
for much of their success. Dropout training is a variant of stochastic gradient descent (SGD) where, as each
example is processed, the network is temporarily perturbed by randomly “dropping out” nodes of the network.
The gradient calculation and weight updates are performed on the reduced network, and the dropped out
nodes are then restored before the next SGD iteration. Since the ImageNet competition, dropout has been
successfully applied to a variety of domains [8, 10, 9, 17, 7], and is widely used [21, 13, 23]; for example, it
is incorporated into popular packages such at Torch [25], Caffe [6] and TensorFlow [24]. It is intriguing that
crippling the network during training often leads to such dramatically improved results, and dropout has
also sparked substantial research on related methods (e.g. [12, 29]).
In this work, we examine the effect of dropout on the inductive bias of the learning algorithm. A match
between dropout’s inductive bias and some important applications could explain the success of dropout, and
its popularity also motivates the study of its unusual properties.
Weight decay training optimizes the empirical error plus an L2 regularization term, λ2 ||w||22 , so we call
λ
2
2 ||w||2 the L2 penalty of w since it is the difference between training criterion evaluated at w and the
empirical loss of w. By analogy, we define the dropout penalty of w to be the difference between the dropout
training criterion and the empirical loss of w (see Section 2). Dropout penalties measure how much dropout
discriminates against weight vectors, so they are key to understanding dropout’s inductive bias.
Even in one-layer networks, conclusions drawn from (typically quadratic) approximations of the dropout
penalty can be misleading [14]. Therefore we focus on exact formal analysis of dropout in multi-layer
1
networks. Theoretical analysis of deep networks is notoriously difficult, so we might expect that a thorough
understanding of dropout in deep networks must be achieved in stages. In this paper we further the process
by exposing some of the surprising ways that the inductive bias of dropout differs from L2 and other standard
regularizers. These include the following:
• We show that dropout training can lead to negative weights even when the output is a positive multiple
of the the inputs. Arguably, such use of negative weights constitutes co-adaptation – this adds a
counterpoint to previous analyses showing that dropout discourages co-adaptation [22, 14].
• Unlike weight decay and other p-norm regularizers, dropout training is insensitive to the rescaling of
input features, and largely insensitive to rescaling of the outputs; this may play a role in dropout’s
practical success. Dropout is also unaffected if the weights in one layer are scaled up by a constant
c, and the weights of another layer are scaled down by c; this implies that dropout training does not
have isolated local minima.
• The dropout penalty grows exponentially in the depth of the network in cases where the L2 regularizer
grows linearly. This may enable dropout to penalize the complexity of the network in a way that more
meaningfully reflects the richness of the network’s behaviors. (The exponential growth with d of the
dropout penalty is reminiscent of some regularizers for deep networks studied by Neyshabur, et al [20].)
• Dropout in deep networks has a variety of other behaviors different from standard regularizers. In particular: the dropout penalty for a set of weights can be negative; the dropout penalty of a set of weights
depends on both the training instances and the labels; and although the dropout probability intuitively
measures the strength of dropout regularization, the dropout penalties are often non-monotonic in the
dropout probability. In contrast, Wager, et al [28] show that when dropout is applied to generalized
linear models, the dropout penalty is always non-negative and does not depend on the labels.
Our analysis is for multilayer neural networks with the square loss at the output node. The hidden layers
use the popular rectified linear units [19] outputting σ(a) = max(0, a) where a is the node’s activation (the
weighted sum of its inputs). We study the minimizers of a criterion that may be viewed as the objective
function when using dropout. This abstracts away sampling and optimization issues to focus on the inductive
bias, as in [5, 30, 4, 18, 14]. See Section 2 for a complete explanation.
Related work
A number of possible explanations have been suggested for dropout’s success. Hinton, et al [16] suggest
that dropout controls network complexity by restricting the ability to co-adapt weights and illustrate how
it appears to learn simpler functions at the second layer. Others [3, 1] view dropout as an ensemble method
combining the different network topologies resulting from the random deletion of nodes. Wager, et al [27]
observe that in 1-layer networks dropout essentially forces learning on a more challenging distribution akin
to ‘altitude training’ of athletes.
Most formal analysis of the inductive bias of dropout has concentrated on the single-layer setting, where
a single neuron combines the (potentially dropped-out) inputs. Wager, et al [28] considered the case that
the distribution of label y given feature vector x is a member of the exponential family, and the log-loss
is used to evaluate models. They pointed out that, in this situation, the criterion optimized by dropout
can be decomposed into the original loss and a term that does not depend on the labels. They then gave
approximations to this dropout regularizer and discussed its relationship with other regularizers. As we have
seen, many aspects of the behavior of dropout and its relationship to other regularizers are qualitatively
different when there are hidden units.
Wager, et al [27] considered dropout for learning topics modeled by a Poisson generative process. They
exploited the conditional independence assumptions of the generative process to show that the excess risk of
dropout training due to training set variation has a term that decays more rapidly than the straightforward
empirical risk minimization, but also has a second additive term related to document length. They also
discussed situations where the model learned by dropout has small bias.
2
Baldi and Sadowski [2] analyzed dropout in linear networks, and showed how dropout can be approximated by normalized geometric means of subnetworks in the nonlinear case. Gal and Ghahramani [11]
described an interpretation of dropout as an approximation to a deep Gaussian process.
The impact of dropout (and its relative dropconnect) on generalization (roughly, how much dropout
restricts the search space of the learner) was studied in [29].
In the on-line learning with experts setting, Van Erven, et al [26] showed that applying dropout in on-line
trials leads to algorithms that automatically adapt to the input sequence without requiring doubling or other
parameter-tuning techniques.
The rest of the paper is organized as follows. Section 2 introduces our notation and formally defines
the dropout model. We prove that dropout enjoys several scaling invariances that weight-decay doesn’t in
Section 3, and that dropout requires negative weights even in very simple situations in Section 4. Section 5
uncovers various properties of the dropout penalty function. Section 6 describes some simulation experiments.
We provide some concluding remarks in Section 7.
2
Preliminaries
Throughout, we will analyze fully connected layered networks with K inputs, one output, d layers (counting
the output, but not the inputs), and n nodes in each hidden layer. We assume that n is a positive multiple
of K and that K is an even perfect square and a power of two to avoid unilluminating floor/ceiling clutter
in the analysis. We will call this the standard architecture. We use W to denote a particular setting of the
weights and biases in the network and W(x) to denote the network’s output on input x using W. The hidden
nodes are ReLUs, and the output node is linear. W can be decomposed as (W1 , b1 , ..., Wd−1 , bd−1 , w, b),
where each Wj is the matrix of weights on connections from the j − 1st into the jth hidden layer, each bj
is the vector of bias inputs into the jth hidden layer, w are the weights into the output node, and b is the
bias into the output node.
We will refer to a joint probability distribution over examples (x, y) as an example distribution. We focus
on square loss, so the loss of W on example (x, y) is (W(x) − y)2 . The risk is the expected loss with respect
def
to an example distribution P , we denote the risk of W as RP (W) = E(x,y)∼P (W(x) − y)2 . The subscript
will often be omitted when P is clear from the context.
The goal of L2 training is to find weights and biases minimizing the L2 criterion with regularization
def
strength λ: J2 (W) = R(W) + λ2 ||W||2 . Here and throughout, we use ||W||2 to denote the sum of the squares
of the weights of W. (As usual, the biases are not penalized.) We use WL2 to denote a minimizer of this
criterion. The L2 penalty, λ2 ||W||2 , is non-negative. This is useful, for example, to bound the risk of a
minimizer WL2 of J2 , since R(W) ≤ J2 (W).
Dropout training independently removes nodes in the network. In our analysis each non-output node is
dropped out with the same probability q, so p = 1 − q is the probability that a node is kept. (The output
node is always kept; dropping it out has the effect of cancelling the training iteration.) When a node is
dropped out, the node’s output is set to 0. To compensate for this reduction, the values of the kept nodes
are multiplied by 1/p. With this compensation, the dropout can be viewed as injecting zero-mean additive
noise at each non-output node [28]. 1
The dropout process is the collection of random choices, for each node in the network, of whether the
node is kept or dropped out. A realization of the dropout process is a dropout pattern, which is a boolean
vector indicating the kept nodes. For a network W, an input x, and dropout pattern R, let D(W, x, R)
be the output of W when nodes are dropped out or not following R (including the 1/p rescaling of kept
nodes’ outputs). The goal of dropout training on an example distribution P is to find weights and biases
minimizing the dropout criterion for a given dropout probability:
def
JD (W) = ER E(x,y)∼P (D(W, x, R) − y)2 .
1 Some authors use a similar adjustment where the weights are scaled down at prediction time instead of inflating the kept
nodes’ outputs at training time.
3
output
1
1
1
1
1
1
1
-1
x1
x2
Figure 1: A network where the dropout penalty is negative.
This criterion is equivalent to the expected risk of the dropout-modified network, and we use WD to denote
a minimizer of it. Since the selection of dropout pattern and example from P are independent, the order of
the two expectations can be swapped, yielding
(1)
JD (W) = E(x,y)∼P ER (D(W, x, R) − y)2 .
Equation (1) is a key property of the dropout criterion. It indicates when something is true about the dropout
criterion for a family of distributions concentrated on single examples, then (usually) the same thing will be
true for any mixture of these single-example distributions.
Consider now the example network in Figure 1. The weight parameters W1 and w are all-1’s, and all of
the biases are 0. W(1, −1) = 0 as each hidden node computes 0. Each dropout pattern indicates the subset
of the four lower nodes to be kept, and when q = p = 1/2 each subset is equally likely to be kept. If R is
the dropout pattern where input x2 is dropped and the other nodes are kept, then the network computes
D(W, (1, −1), R) = 8 (recall that when p = 1/2 the values of non-dropped out nodes are doubled). Only
three dropout patterns produce a non-zero output, so if P is concentrated on the example x = (1, −1), y = 8
the dropout criterion is:
2
13
1
(8 − 8)2 + (4 − 8)2 + (0 − 8)2 = 54.
(2)
16
16
16
As mentioned in the introduction, the dropout penalty of a weight vector for a given example distribution
and dropout probability is the amount that the dropout criterion exceeds the risk, JD (W) − R(W). Wager,
et al [28] show that for 1-layer generalized linear models, the dropout penalty is non-negative.
Since W(1, −1) = 0, we have R(W) = 64, and the dropout penalty is negative in our example. This is
because the variance in the output due to dropout causes the network to better fit the data (on average) than
the network’s non-dropout evaluation. In Section 5.2, we give a necessary condition for this variance to be
beneficial. As with the dropout criterion, the dropout penalty decomposes into an expectation of penalties
over single examples:
JD (W) − R(W) = E(x,y)∼P ER (D(W, x, R) − y)2 − (W(x) − y)2 .
JD (W) =
Definition 1 Define P(x,y) as the distribution with half of its weight on example (x, y) and half of its weight
on example (0, 0).
Unless indicated otherwise, we assume p = q = 1/2 for simplicity, although this is not crucial for our
results.
3
3.1
Scaling inputs, weights and outputs
Dropout is scale-free
Here we prove that dropout regularizes deep networks in a manner that is independent of the scale of the
input features. In other words, training under dropout regularization does not penalize the use of large
4
weights when needed to compensate for small input values.
Definition 2 For any example distribution P , define the dropout aversion with P to be the maximum, over
minimizers WD of the dropout risk JD (WD ), of RP (WD ) − inf W RP (W).
The dropout aversion of P measures the extent to which P is incompatible with the inductive bias
of dropout, measured by the risk gap between the true risk minimizer and the optimizers of the dropout
criterion.
Definition 3 For example distribution P and square matrix A, denote by A ◦ P the distribution obtained by
sampling (x, y) from P , and outputting (Ax, y).
When A is diagonal and has full rank, then A ◦ P is a rescaling of the inputs, like changing one input
from minutes to seconds and another from feet to meters.
Theorem 4 For any example distribution P , and any diagonal full-rank K × K matrix A, the dropout
aversion of P equals the dropout aversion of A ◦ P .
Proof: Choose a network W = (W1 , b1 , ..., Wd−1 , bd−1 , w, b). Let W ′ = (W1 A−1 , b1 ..., Wd−1 , bd−1 , w, b).
For any x, W(x) = W ′ (Ax), as A−1 undoes the effect of A before it gets to the rest of the network, which
is unchanged. Furthermore, for any dropout pattern R, we have D(W, x, R) = D(W ′ , Ax, R). Once again
A−1 undoes the effects of A on kept nodes (since A is diagonal), and the rest of the network W ′ is modified
by R an a manner paralleling W. Thus, there is bijection between networks W and networks W ′ with
JD (W ′ ) = JD (W) and R(W ′ ) = R(W), yielding the theorem.
Theorem 4 indicates that some common normalizations of the input features (e.g. to have unit variance)
do not affect the quality of the dropout criterion minimizers, but normalization might change the speed
of convergence and which minimizer is reached. Centering the features has slightly different properties.
Although it is easy to use the biases to define a W ′ that “undoes” the centering in the non-dropout computation, different W ′ appear to be required for different dropout patterns, breaking the bijection exploited in
Theorem 4.
As we will see in Section 3.4, weight decay does not enjoy such scale-free status.
3.2
Dropout’s invariance to parameter scaling
Next, we describe an equivalence relation among parameterizations for dropout networks of depth d ≥
2. Basically, scaling the parameters at a level creates a corresponding scaling of the output. (A similar
observation was made in a somewhat different context by Neyshabur, et al [20].)
Theorem 5 For any input x, dropout pattern R, any network W = (W1 , b1 , ..., Wd−1 , bd−1 , w, b), and any
positive c1 , ..., cd , if
d−1
d
Y
Y
W ′ = c1 W1 , c1 b1 , c2 W2 , c1 c2 b2 , ..., cd−1 Wd−1 ,
cj bd−1 , cd w,
cj b ,
(3)
j=1
then D(W ′ , x, R) =
Q
d
j=1 cj
′
D(W, x, R). In particular, if
Qd
j=1 cj
j=1
= 1, then for any example distribution
P , networks W and W have the same dropout criterion, dropout penalty, and expected loss.
Note that the re-scaling of the biases at layer j depends not only on the rescaling of the connection
weights at layer j, but also the re-scalings at lower layers.
Proof: Choose an input x and a dropout pattern R. Define W ′ as in (3). For each hidden layer j,
let (hj1 , ..., hjn ) be the jth hidden layer when applying W to x with
QR, and let (h̃j1 , ..., h̃jn ) be the jth
′
hidden layer when applying W instead. By induction, for all i, h̃ji =
ℓ≤j cℓ hji ; the key step is that the
5
pre-rectified value used to compute h̃ji has the same sign as for hji , since
Q rescaling by cj preserves
Qthe sign.
Thus the same units are zeroed out in W and W ′ and D(W, x, R) = ( j cj )D(W ′ , x, R). When j cj = 1,
this implies D(W, x, R) = D(W ′ , x, R). Since this is true for all x and R, we have JD (W) = JD (W ′ ). Since,
similarly, W(x) = W ′ (x) for all x, so R(W) = R(W ′ ), the dropout penalities for W and W ′ are also the
same.
Theorem 5 implies that the dropout criterion never has isolated minimizers, since one can continuously
up-scale the weights on one layer with a compensating down-scaling at another layer to get a contiguous
family of networks computing the same function and having the same dropout criterion. It may be possible to
exploit the parameterization equivalence of Theorem 5 in training by using canonical forms for the equivalent
networks or switching to an equivalent networks whose gradients have better properties. We leave this
question for future work.
3.3
Output scaling with dropout
Scaling the output values of an example distribution P does affect the aversion, but in a very simple and
natural way.
Theorem 6 For any example distribution P , if P ′ is obtained from P by scaling the outputs of P by a
positive constant c, the dropout aversion of P ′ is c2 times the dropout aversion of P .
Proof: If a network W minimizes the dropout criterion for P , then the network W ′ obtained by scaling up
the weights and bias for the output unit by c minimizes the dropout criterion for P ′ , and for any x, y, and
dropout pattern R, (D(W, x, R) − y)2 = (D(W ′ , x, R) − cy)2 /c2 .
3.4
Scaling properties of weight decay
Weight decay does not have the same scaling properties as dropout. Define the weight-decay aversion
analogously to the dropout aversion.
We analyze the L2 criterion for depth-2 networks in Appendix B, resulting in the following theorem. Our
proof shows that, despite the non-convexity of the L2 criterion, it is still possible to identify a closed form
for one of its optimizers.
Theorem 7 Choose an arbitrary number n of hidden nodes, and λ > 0.
λ2
).
The weight-decay aversion of P(x,1) is min( 41 , x·x
Theorem 7 shows that, unlike dropout, the weight decay aversion does depend on the scaling of the input
features.
√
Furthermore, when 2λ > x · x, the weight-decay criterion for P(x,1) has only a single isolated optimum
weight setting2 – all weights set to zero and bias 1/2 at the output node. This means that weight-decay
in 2-layer networks can completely regularize away significant signal in the sample even when λ is finite,
contrasting starkly with weight-decay’s behavior in 1-layer networks.
The “vertical” flexibility to rescale weights between layers enjoyed by dropout (Theorem 5) does not hold
for L2 : one can always drive the L2 penalty to infinity by scaling one layer up by a large enough positive
c, even while scaling another down by c. On the other hand, the proof of Theorem 7 shows that the L2
criterion has an alternative “horizontal” flexibility involving the rescaling of weights across nodes on the
hidden layer (under the theorem’s assumptions). Lemma 32 shows that at the optimizers each hidden node’s
contributions to the output are a constant (depending on the input) times their contribution to the the L2
penalty. Shifting the magnitudes of these contributions between hidden nodes leads to alternative weights
2 Since the output node puts weight 0 on each hidden node and the biases are unregularized, this optimum actually represents
a class of networks differing only in the irrelevant biases at the hidden nodes. One can easily construct other cases when weightdecay has isolated minima in this sense, for example when n = 2 and there is equal probability on x and −x, both with label
1.
6
that compute the same value and have the same weight decay penalty. This is a more general observation
than the permutation symmetry between hidden nodes because any portion of a hidden node’s contribution
can be shifted to another hidden node.
4
Negative weights for monotone functions
If the weights of a unit are non-negative, then the unit computes a monotone function, in the sense that
increasing any input while keeping the others fixed increases the output. The bias does not affect a node’s
monotonicity. A network of monotone units is also monotone.
We first present our theoretical results for many features (Section 4.1) and few features (Section 4.2),
and then discuss the implication of these results in Section 4.3.
4.1
The basic case – many features
In this section, we analyze the simple distribution P(1,1) that assigns probability 1/2 to the example
(0, ..., 0), 0, and probability 1/2 to the example (1, ..., 1), 1. This is arguably the simplest monotone function. Nevertheless, we prove that dropout uses negative weights to fit this data.
The key intuition is that optimizing the dropout criterion requires controlling the variance. Negative
weights at the hidden nodes can be used to control the variance due to dropout at the input layer. When
there are enough hidden nodes this becomes so beneficial that every minimizer of the dropout criterion uses
such negative weights.
Theorem 8 For the standard architecture, if K > 18 and n is large enough relative to K and d, every
optimizer of the dropout criterion for P(1,1) uses at least one negative weight.
To prove Theorem 8, we first calculate JD (Wneg ) for a network Wneg that uses negative weights, and
then prove a lower bound greater than this value that holds for all networks using only non-negative weights.
All of the biases in Wneg are 0.
A key building block in the definition of Wneg is a block of hidden units that we call the first-one gadget.
Each such block has K hidden nodes, and takes its input from the K input nodes. The ith hidden node
in the block takes the value 1 if the Kth input node is 1, and all inputs xi′ for i′ < i are 0. This can be
accomplished with a weight vector w with wi′ = −1 for i′ < i, with wi = 1, and with wi′ = 0 for i′ > i. The
first hidden layer of Wneg comprises n/K copies of the first-one gadget.
Informally, this construction removes most of the variance in the number of 1′ s in the input, as recorded
in the following lemma.
Lemma 9 On any input x ∈ {0, 1}n except (0, 0, ..., 0), the sum of the values on the first hidden layer of
Wneg is exactly n/K.
The weights into the remaining hidden layers of Wneg are all 1, and all the weights into the output
def
layer take a value c =
K2
,
1 d−2
2nd−1 (1+ K
n )(1+ n )
following lemma analyzes Wneg .
1
Lemma 10 JD (Wneg ) = 2 1 −
chosen to minimize the dropout criterion for the network. The
(1−2−K )
d−2
(1+ Kn )(1+ n1 )
.
d−2
1 + n1
denominator in Lemma 10 approaches 1, so
When n is large relative to K and d, the 1 + K
n
JD (Wneg ) approaches 2−K /2 in this case. Lemma 16 below gives a larger lower bound for any network with
all positive weights. In the concrete case when d = 2 and n = K 3 , then Lemma 10 implies JD (Wneg ) < 1/K 2 .
Proof (of Lemma 10): Consider a computation of Wneg (1, 1, ..., 1) under dropout and let ŷ be the
(random) output. Let k0 be the number of input nodes kept, and, for each j ≥ 2, let kj be the number of
7
nodes in the jth hidden layer kept. Call the node in each first-one gadget that computes 1 a key node, and
if no node in the gadget computes 1 because the input is all dropped, arbitrarily make the gadget’s first
hidden node the key node. This ensures there is exactly one key node per gadget, and every non-key node
computes 0. Let k1 be the number
kept key nodes on the first hidden layer. If k0 = 0, the output ŷ of the
Qof
d−1
network is 0. Otherwise, ŷ = c2d j=1 kj .
Note that k0 is zero with probability 2−K . Whenever k0 ≥ 1, k1 is distributed as B(n/K, 1/2). Each
other kj is distributed as B(n, 1/2), and k1 , k2 , ..., kd−1 are independent of one another.
E(ŷ) = Pr(k0 ≥ 1)c2d E[k1 |k0 ≥ 1]
d−1
Y
j=2
E[kj ] = (1 − 2−K )c2d
n n d−2 2c
= (1 − 2−K )nd−1 .
2K
2
K
Using the value of the second moment of the binomial, we get
2
d−1
n n
Y
kj = 4c2 (1 − 2−K )
E(ŷ 2 ) = E 1k0 ≥1 c2d
+ 1 nd−2 (n + 1)d−2
K
K
j=1
=
d−2
4c2 (1 − 2−K ) 2(d−1)
1
K
1
+
1
+
.
n
K2
n
n
Thus,
1
1 − 2E(ŷ) + E(ŷ 2 )
2
d−2 !
2
−K
4c
4c
(1
−
2
)
1
1
K
1 − (1 − 2−K )nd−1 +
n2(d−1) 1 +
=
1+
2
K
K2
n
n
!
(1 − 2−K )
1
1−
=
d−2 ,
2
1+ K 1+ 1
JD (Wneg ) =
n
since c =
K
d−2
1
2nd−1 (1+ K
n )(1+ n )
n
, completing the proof.
Next we prove a lower bound on JD for networks with nonnegative weights. Let W be an arbitrary such
network.
Our lower bound will use a property of the function computed by W that we now define.
Definition 11 A function φ : RK → R is supermodular if for all x, δ 1 , δ 2 ∈ RK where δ 1 , δ 2 ≥ 0
φ(x) + φ(x + δ 1 + δ 2 ) ≥ φ(x + δ 1 ) + φ(x + δ 2 ),
or equivalently:
φ(x + δ 1 + δ 2 ) − φ(x + δ 2 ) ≥ φ(x + δ 1 ) − φ(x)
The latter indicates that adding δ 1 to the bigger input x + δ 2 has at least as large an effect as adding it to
the smaller input x.
Since W has all non-negative weights, it computes a supermodular function of its inputs. (This fact may
be of independent interest.)
Lemma 12 If a network has non-negative weights and its activation functions σ(·) are convex, continuous,
non-decreasing, and differentiable except on a finite set, then the network computes a supermodular function
of its input x.
8
Proof: We will prove by induction over the layers that, for any unit h in the network, if h(x) is the output
of unit h when x is the input to W, then h(·) is a supermodular function of its input.
The base case holds since each input node h outputs the corresponding component of the input, and
(x + δ 1 ) − x = (x + δ 1 + δ 2 ) − (x + δ 2 ).
Now, for the inductive step, let w be the weight vector for node h, let b be its bias, and σ(·) be its
activation function. Let I(x), I(x + δ 1 ), I(x + δ 2 ), and I(x + δ 1 + δ 2 ) be the inputs to node h when the
inputs to the network are x, x + δ 1 , x + δ 2 and x + δ 1 + δ 2 respectively.
By induction, these inputs to node h (componentwise) satisfy
I(x + δ 1 ) − I(x) ≤ I(x + δ 1 + δ 2 ) − I(x + δ 2 ).
Therefore, since w, δ 1 , and δ 2 are non-negative, the interval [w · I(x + δ 2 ) + b, w · I(x + δ 1 + δ 2 ) + b] is at
least as long and starts at least as high as the interval [w · I(x) + b, w · I(x + δ 1 ) + b]. Since σ is continuous
and differentiable except on a finite set, we have
Z w·I(x+δ1 )+b
h(I(x + δ 1 )) − h(I(x)) =
σ ′ (z)dz
w·I(x)+b
≤
Z
w·I(x+δ 1 +δ 2 )+b
σ ′ (z)dz (since σ ′ is non-decreasing)
w·I(x+δ 2 )+b
= h(I(x + δ 1 + δ 2 )) − h(I(x + δ 2 )).
Definition 13 Let r0 ∈ {0, 1}K be the dropout pattern concerning the input layer, and let R′ be the dropout
pattern concerning the rest of the network, so that the dropout pattern R = (r0 , R′ ).
For each ℓ ∈ {0, ..., K}, let ψW (ℓ) be the average output of W under dropout when ℓ of the inputs are
kept: i.e.,
X
r0j = ℓ .
ψW (ℓ) = E D(W, 1K , (r0 , R′ ))
j
Lemma 14 For any ℓ ∈ {1, ..., K − 1, },
ψW (ℓ + 1) − ψW (ℓ) ≥ ψW (ℓ) − ψW (ℓ − 1).
Proof: Generate u, i and j randomly by, first choosing u uniformly at random from among bit vectors with
ℓ ones, then choosing i uniformly from the 0-components of u, and j uniformly from the 1-components of u.
By Lemma 12,
W(u + ei ) − W(u) ≥ W(u − ej + ei ) − W(u − ej )
(4)
always holds. Furthermore, u + ei is uniformly distributed among bit vectors with ℓ + 1 ones, u − ej is
uniformly distributed among bit vectors with ℓ − 1 ones, and u + ei − ej is uniformly distributed among bit
vectors with ℓ ones. This is true for W, but it is also true for any network obtained by dropping out some
of the hidden nodes of W. Thus
ψW (ℓ + 1) − ψW (ℓ)
= E(D(W, 1K , (r0 , R′ )|
= E(D(W, r0 , (1K , R′ )|
≥ E(D(W, r0 , (1K , R′ )|
= ψW (ℓ) − ψW (ℓ − 1),
X
j
X
j
X
j
r0j = ℓ + 1)) − E(D(W, 1K , (r0 , R′ )|
r0j = ℓ + 1)) − E(D(W, r0 , (1K , R′ )|
r0j = ℓ)) − E(D(W, r0 , (1K , R′ )|
9
X
j
X
r0j = ℓ))
j
X
r0j = ℓ))
j
r0j = ℓ − 1)) (by (4))
completing the proof.
We will use the following lower bound on the tail of the binomial. (Many similar lower bounds are
known.)
Lemma 15 If X is distributed according to Bin(n, 1/2), then
√
√
Pr(X < n/2 − n/4) = Pr(X > n/2 + n/4) ≥ 1/4.
√
√
Proof: Using √
the fact that, for any i, Pr(X = i) ≤ 1/ n, we get Pr(|X − n/2| < n/4) ≤ 1/2, so
Pr(X < n/2 − n/4) ≥ 1/4.
Now we are ready for the lower bound on JD (W).
1
Lemma 16 If K > 18 and the weights in W are non-negative, then JD (W) ≥ 36K
.
q
1
1
Proof: Assume to the contrary that JD (W) < 36K
. First, note that ψW (0) ≤ 18K
, or else the contribution
1
. Applying Lemma 15, we have
to JD (W) due to the (0, 0), 0 example is at least 36K
r
r
√
√
2
2
ψW (K/2 − K/4) > 1 −
and ψW (K/2 + K/4) < 1 +
9K
9K
(5)
1
for the (1, . . . , 1), 1 example.
as otherwise the contribution of one of the tails to JD (W) will be at least 36K
√
We will contradict this small variation of ψW (ℓ) around K/2. The bounds on ψW (0) and ψ(K/2 − K/4)
and Lemma 14 imply that ψW (ℓ) grows rapidly when ℓ is around K/2, in particular:
q
q
2
1
1 − 9K
− 18K
√
√
1
√
ψ(K/2 − K/4 + 1) − ψ(K/2 − K/4) ≥
> p
,
K/2 − K/4
9/32K
since K > 18. Now using Lemma 14 repeatedly shows that
√
√
ψ(K/2 + K/4) − ψ(K/2 − K/4) >
which contradicts (5), completing the proof.
√
1
K
×p
=2
2
9/32K
r
2
,
9K
Putting together Lemmas 10 and 16 immediately proves Theorem 8, since for K > 18 and large enough
n, the criterion for Wneg must be less than the criterion for any network with all non-negative weights.
4.2
The case when K = 2
Theorem 8 uses the assumption that K > 18 and n is large enough; is the lower bound on K really necessary?
Here we show that it is not, by treating the case that K = 2.
Theorem 17 For the standard architecture, if K = 2, for any fixed d and large enough n, every optimizer
of the dropout criterion for P(1,1),1 uses negative weights.
Proof: Define Wneg as in the proof of Lemma 10, except that the output layer has a bias of 1/5.
We claim that
lim JD (Wneg ) = 1/10.
n→∞
To see this, consider the joint input/label distribution under dropout:
Pr((0, 0), 0) = 1/2
Pr((0, 0), 1) = 1/8
Pr((2, 0), 1) = 1/8
Pr((0, 2), 1) = 1/8
Pr((2, 2), 1) = 1/8.
10
(6)
Due to the bias of 1/5 on the output, Wneg (0, 0) = 1/5. Thus, contribution to JD from examples with
x = (0, 0) in this joint distribution is 1/2 × (1/5)2 + 1/8 × (4/5)2 = 1/10.
Now, choose x 6= (0, 0). If, after dropout, the input is x, each node in the hidden layer closest to the
input computes 1. Arguing exactly as in the proof of Lemma 10, in such cases,
E((ŷ − 1)2 ) = 1 −
1+
2
n
1
1+
1 d−2
n
.
This proves (6).
Now, let W be an arbitrary network with non-negative weights.
For our distribution,
ER (D(W, (0, 0), R) − 0)2 + ER (D(W, (1, 1), R) − 1)2
JD (W) =
.
2
Let V00 = E(W(0, 0)), V22 = E(W(2, 2)), V20 = E(W(2, 0)), V02 = E(W(0, 2)) where the expectations are
taken with respect to the dropout patterns at the hidden nodes (with no dropout at the inputs). Since each
dropout pattern over the hidden nodes defines a particular network, and Lemma 12 holds for all of them,
the relationships also hold for the expectations, so
V22 ≥ V20 + V02 − V0 .
Using this V notation, handling the dropout at the input explicitly, and the bias-variance decomposition
keeping just the bias terms we get:
(V00 − 0)2 + (V00 − 1)2 + (V22 − 1)2 + (V20 − 1)2 + (V02 − 1)2 /4
2
8JD (W) ≥ 4(V00 − 0)2 + (V00 − 1)2 + (V22 − 1)2 + (V20 − 1)2 + (V02 − 1)2 .
JD (W) ≥
(7)
(8)
We will continue lower bounding the RHS. We can re-write V22 as V20 + V02 − V00 + ǫ where ǫ ≥ 0. This
is convex and symmetric in V02 and V20 so they both take the same value at the minimizer of the RHS, so
we proceed using V20 for this common minimizing value.
2
8JD (W) ≥ (2V20 − V00 + ǫ − 1)2 + 2(V20 − 1)2 + (V00 − 1)2 + 4V00
.
Differentiating with respect to V20 , we see that the RHS is minimized when V20 = (2 + V00 − ǫ)/3, giving
8JD (W) ≥
(V00 − ǫ − 1)2
2
+ (V00 − 1)2 + 4V00
.
3
If V00 ≥ 1, then JD ≥ 1/2 (just from the (0,0),0 example) and when V00 < 1 the RHS of the above is
minimized for non-negative ǫ when ǫ = 0. Using this substitution, the minimizing value of V00 is 1/4 giving
8JD (W) ≥ 1
JD (W) ≥ 1/8.
Combining this with (6) completes the proof.
11
4.3
More general distributions and implications
In the previous sub-section we analyzed the distribution P over the 2-feature examples (1, 1), 1 and (0, 0), 0.
However, these two examples can be embedded in a larger feature space by using any fixed vector of additional
feature values, creating, for instance, a distribution over (0, 1, 0, 0, 0, 1/2, 1), 0 and (0, 1, 1, 1, 0, 1/2, 1), 1 (with
the additional features underlined) The results of Section 4.2 still apply to distribution over these extended
examples after defining Wneg network to have zero weight on the additional features, and noticing that any
weight on the additional features in the positive-weight network W can be simulated using the biases at the
hidden nodes.
It is particularly interesting when the additional features all take the value 0, we call these zero-embeddings.
Every network W with non-negative weights has JD (W) ≥ 1/8 on each of these zero-embeddings of P . On
the other hand, a single Wneg network with n/K copies of the K-input first-one gadget has JD (Wneg ) ≈ 1/10
simultaneously for all of these zero-embeddings of P (when n >> K).
Any source distribution over {0, 1}K × {0, 1} that puts probability 1/2 on the the 0, 0 example and
distributes the other 1/2 probability over examples where exactly two inputs are one is a mixture of zeroembeddings of P , and thus JD (W) ≥ 1/8 while JD (Wneg ) ≈ 1/10 for this mixture and optimizing the
dropout criterion requires negative weights.
In our analysis the negative weights used by dropout are counterintuitive for fitting monotone behavior,
but are needed to control the variance due to dropout. This suggests that dropout may be less effective
when layers with sparse activation patterns are fed into wider layers, as dropout training can hijack part of
the expressiveness of the wide layer to control the artificial variance due to dropout rather than fitting the
underlying patterns in the data.
5
5.1
Properties of the dropout penalty
Growth of the dropout penalty as a function of d
Weight-decay penalizes large weights, while Theorem 5 shows that compensating rescaling of the weights
does not affect the dropout penalty or criterion. On the other hand, dropout can be more sensitive to the
calculation of large outputs than weight decay, and large outputs can be produced in deep networks using
only small weights. We make this observation concrete by exhibiting a family of networks where the depth
and desired output are linked while the size of individual weights remains constant. For this family, the
dropout penalty grows exponentially in the depth d (as opposed to linearly for weight-decay), suggesting
that dropout training is less willing to fit the data in this kind of situation.
Theorem 18 If x = (1, 1, ..., 1) and 0 ≤ y ≤ Knd−1 , for Px,y there are weights W for the standard
y2
,
architecture with R(W) = 0 such that (a) every weight has magnitude at most one, but (b) JD (W) ≥ K+1
whereas (c) J2 (W) ≤
λy 2/d
2 (Kn
+ n2 (d − 2) + n).
1/d
Proof: Let W be the network whose weights are all c = K 1/dyn(d−1)/d and biases are all 0, so that the L2
penalty is the number of weights times λc2 /2. It is a simple induction to show that, for these weights and
input (1, 1, ..., 1), the value computed at each hidden node on level j is cj Knj−1 , so the the network outputs
cd Knd−1 , and has zero square loss (since W(x) = cd Knd−1 = y).
Consider now dropout on this network. This is equivalent to changing all of the weights from c to 2c and,
independently with probability 1/2, replacing the value of each node with 0. For a fixed dropout pattern,
each node on a given layer has the same weights, and receives the same (kept) inputs. Thus, the value
computed at every node on the same layer is the same. For each j, let Hj be the value computed by the
units in the jth hidden layer.
If k0 is the number of input nodes kept under dropout, and, for each j ∈ {1, ..., d − 1}, kj is the number of
Qℓ−1
hidden nodes kept in layer j, a straightforward induction shows that, for all ℓ, we have Hℓ = (2c)ℓ j=0 kj ,
Qd−1
so that the output ŷ of the network is (2c)d j=0 kj .
12
Using a bias-variance decomposition, E((ŷ − y)2 ) = (E[ŷ] − y)2 + Var(ŷ). Since each kj is binomially
distributed, and k0 , ..., kd−1 are independent, we have E(ŷ) = (2c)d (K/2)(n/2)d−1 = cd Knd−1 = y, so
E((ŷ − y)2 ) = Var(ŷ). Since E(ŷ 2 ) = (2c)2d (K(K + 1)/4)(n(n + 1)/4)d−1 = y 2 (1 + 1/K)(1 + 1/n)d−1 , we
have Var(ŷ) = E(ŷ 2 ) − E(ŷ)2 = y 2 ((1 + 1/K)(1 + 1/n)d−1 − 1) ≥ y 2 /K, completing the proof.
If y = exp(Θ(d)), the dropout penalty grows exponentially in d, whereas the L2 penalty grows polynomially.
5.2
A necessary condition for negative dropout penalty
Section 2 contains an example where the dropout penalty is negative. The following theorem provides a
necessary condition.
Theorem 19 The dropout penalty can be negative. For all example distributions, a necessary condition for
this in rectified linear networks is that either a weight, input, or bias is negative.
Proof: [2] show that for networks of linear units (as opposed to the non-linear rectified linear units we
focus on) the network’s output without dropout equals the expected output over dropout patterns, so in
our notation: W(x) equals ER (D(W, x, R)). Assume for the moment that the network consists of linear
units and the example distribution is concentrated on the single example (x, y). Using the bias-variance
decomposition for square loss and this property of linear networks,
JD (W) = ER (D(W, x, R)−y)2 = (ER (D(W, x, R)−y)2 + VarR (D(W, x, R)) ≥ (W(x) − y)2
and the dropout penalty is again non-negative. Since the same calculations go through when averaging
over multiple examples, we see that the dropout penalty is always non-negative for networks of linear nodes.
When all the weights, biases and inputs in a network of rectified linear units are positive, then the rectified
linear units behave as linear units, so the dropout penalty will again be non-negative.
5.3
Multi-layer dropout penalty does depend on labels
In contrast with its behavior on a variety of linear models including logistic regression [28], the dropout
penalty can depend on the value of the response variable in deep networks with ReLUs and the quadratic
loss. Thus in a fundamental and important respect, dropout differs from traditional regularizers like weightdecay or an L1 penalty.
Theorem 20 There are joint distributions P and Q, and weights W such that, for all dropout probabilities
q ∈ (0, 1), (a) the marginals of P and Q on the input variables are equal, but (b) the dropout penalties of W
with respect to P and Q are different.
We will prove Theorem 20 by describing a general, somewhat technical, condition that implies that P
and Q are witnesses to Theorem 20.
For each input x and dropout pattern R, let H(W, x, R) be the values presented to the output node with
dropout. As before, let w ∈ Rn be those weights of W on connections directly into the output node and let
b be the bias at the output node. Let r ∈ {0, 1}n be the indicator variables for whether the various nodes
connecting to the output node are kept.
Proof (of Theorem 20): Suppose that P is concentrated on a single (x, y) pair. We will then get Q by
modifying y.
Let h be the values coming into the output node in the non-dropped out network. Therefore the output
of the non-dropout network is w · h + b while the output of the network with dropout is w · H(W, x, R) + b.
We now examine the dropout penalty, which is the expected dropout loss minus the non-dropout loss. We
13
will use δ as a shorthand for w · (H(W, x, R) − h).
dropout penalty = E (w · H(W, x, R) + b − y)2 − (w · h + b − y)2
= E (w · H(W, x, R) + b − w · h + w · h − y)2 − (w · h + b − y)2
= E δ 2 + 2(w · h + b − y)E(δ)
which depends on the label y unless E(δ) = 0.
Typically E(δ) 6= 0. To prove the theorem, consider the case where x = (1, −2) and there are two hidden
nodes with weights 1 on their connections to the output node. The value at the hidden nodes without
dropout is 0, but with dropout the hidden nodes never negative and computes positive values when only the
negative input is dropped, so the expectation of δ is positive.
6
Experiments
We did simulation experiments using Torch [25]. The code is accessible at
https://www.dropbox.com/sh/6s2lcfrq17zshmp/AAAQ06uDa4gOAuAnw2MAghEMa?dl=0
6.1
Negative Weights
We trained networks with the standard architecture with K = 5 inputs, depth 3, and width 50 on the training
data studied in Section 4: one input with x = (0, 0, ..., 0) and y = 0, and one input with x = (1, 1, ..., 1) and
y = 1. We used the nn.StochasticGradient function from Torch with a maximum of 10000 iterations, and
0.1
a learning rate of 1+0.1×t
on iteration t.
With the above parameters, we repeated the following experiment 1000 times.
• Initialize the parameters of a network WD .
• Clone this network, to produce a copy W with the same initialization.
• Train WD with dropout probability 1/2 at all nodes, and train W without dropout (without using
weight decay for either one).
• Compare the number of negative weights neg(WD ) in WD with neg(W).
Counts of the outcomes are shown in the following table.
neg(WD ) > neg(W)
584
neg(WD ) < neg(W)
338
neg(WD ) = neg(W)
78
Recall that Theorem 8 shows that a global optimizer of the dropout criterion definitely uses negative
weights when trained on data sources similar to P(1,1) , despite the fact that that they are consistent with a
monotone function. This experiment shows that even when optimization is done in a standard way, using
dropout tends to create models with more negative weights for P(1,1) , and we attribute this to the variance
reduction effects associated with Theorem 8.
The experiment also indicates that training with dropout sometimes produces fewer negative weights. Due
to the random initialization, a hidden ReLU unit with negative weights can evaluate to 0 on both inputs,
so its weights will never be updated without dropout. On the other hand, the extra variance from dropout
could cause updates to these negative weights. Another way the non-dropout training could produce more
negative weights is if a hidden node whose output is too large has many small weights that turn negative
with a standard gradient descent step. However, with dropout only about half the small weights will be
updated. Theorem 8 focusses on the effect of dropout at the global minimum, and abstracts away these
kinds of initialization and optimization effects.
14
1.1
Dropout
Weight Decay
1
Training error
0.9
0.8
0.7
0.6
0.5
0.4
0.6
0.8
1
1.2
1.4
1.6
Input Scale
Figure 2: Training error as a function of the scale of the inputs for Dropout and Weight Decay in the
experiment of Section 6.2.
6.2
Scale (in)sensitivity
Our experiments regarding the sensitivity with respect to the scale of the input used the standard architecture
with K = 5, d = 2, n = 5. We used stochastic gradient using the optim package for Torch, with learning
0.01
and momentum of 0.5, and a maximum of 100000 iterations.
rate 1+0.00001t
We performed 10 sets of training runs. In each run:
• Ten training examples were generated uniformly at random from [−1, 1]K .
Q
• Target outputs were assigned using y = i sign(xi ).
• Five training sets S1 , ..., S5 with ten examples each were obtained by rescaling the inputs by {0.5, 0.75, 1, 1.25, 1.5}
and leaving the outputs unchanged.
• The weights of a network Winit were initialized using the default initialization from Torch.
• For each Si :
– Winit was cloned three times to produce WD , W2 and Wnone with identical starting parameters.
– WD was trained with dropout probability 1/2 and no weight decay.
– W2 was trained with weight decay with λ = 1/2 and no dropout.
– Wnone was trained without any regularization.
The average training losses of WD and W2 , over the 10 runs, are shown in Figure 2. (The average training
loss of Wnone was less than 0.05 at all scales.)
The theoretical insensitivity of dropout to the scale of the inputs described in Theorem 4 is also seen
here, along with the contrast with weight decay analyzed in Theorem 7.
The scale of the inputs also affects the dynamics of stochastic gradient descent. With very small inputs,
convergence is very slow, and with very large inputs, SGD is unstable. The effects of the scale of the inputs
on inductive bias analyzed in this paper are visible at the scales where optimization can be done effectively.
7
Conclusions
The reasons behind dropout’s surprisingly good performance in training deep networks across a variety of
applications are somewhat mysterious and there is relatively little existing formal analysis. A variety of
15
explanations have been offered (e.g. [1, 2, 3, 11, 27]), including the possibility that dropout reduces the
amount of coadaptation in a network’s weights [16].
The dropout criterion is an expected loss over dropout patterns, and the variance in the output values
over dropout patterns contributes to this expected loss. Therefore dropout may co-adapt weights in order
to reduce this (artificial) variance. We prove that this happens even in very simple situations where nothing
in the training data justifies negative weights (Theorem 8). This indicates that the relationship between
dropout and co-adaption is not a simple one.
The effects of dropout in deep neural networks are rather complicated, and approximations can be
misleading since the dropout penalty is very non-convex even in 1-layer networks [14]. In Section 3 we show
that dropout does enjoy several scale-invariance properties that are not shared by weight-decay. A perhaps
surprising consequence of these invariances is that there are never isolated local minima when learning a
deep network with dropout. Further exploration of these scale invariance properties is warranted to see
if they are a contributor to dropout’s empirical success or can be exploited to facilitate training. While
contrasting dropout to weight-decay in simple situations, we found that a degenerate all-zero network results
(Theorem 7) when the L2 regularization parameter is above a threshold. This is in dramatic contrast to our
previous intuition from the 1-layer case.
In [28], dropout was viewed as a regularization method, adding a data dependent penalty to the empirical
loss of (presumably) undesirable solutions. Section 5 shows that, unlike the generalized linear models case
analyzed there, the dropout penalty in deeper networks can be negative and depends on the labels in the
training data, and thus behaves unlike most regularizers. On the other hand, the dropout penalty can grow
exponentially in the depth of the network, and thus may better reflect the complexity of the underlying
model space than L2 regularization.
This paper uncovers a number of dropout’s interesting fundamental properties using formal analysis of
simple cases. However, the effects of using dropout training in deep networks are subtle and complex, and
we hope that this paper lays a foundation to promote further formal analysis of dropout’s properties and
behavior.
Acknowledgments
We are very grateful to Peter Bartlett, Seshadhri Comandur, and anonymous reviewers for valuable communications.
References
[1] P. Bachman, O. Alsharif, and D. Precup. Learning with pseudo-ensembles. NIPS, 2014.
[2] P. Baldi and P. Sadowski. The dropout learning algorithm. Artificial intelligence, 210:78–122, 2014.
[3] Pierre Baldi and Peter J Sadowski. Understanding dropout. In Advances in Neural Information Processing
Systems, pages 2814–2822, 2013.
[4] P. L. Bartlett, M. I. Jordan, and J. D. McAuliffe. Convexity, classification, and risk bounds. Journal of the
American Statistical Association, 101(473):138–156, 2006.
[5] L. Breiman. Some infinity theory for predictor ensembles. Annals of Statistics, 32(1):1–11, 2004.
[6] Caffe, 2016. http://caffe.berkeleyvision.edu.
[7] Danqi Chen and Christopher D Manning. A fast and accurate dependency parser using neural networks. In
EMNLP, pages 740–750, 2014.
[8] G. E. Dahl. Deep learning how I did it: Merck 1st place interview, 2012. http://blog.kaggle.com.
[9] G. E. Dahl, T. N. Sainath, and G. E. Hinton. Improving deep neural networks for LVCSR using rectified linear
units and dropout. ICASSP, 2013.
[10] L. Deng, J. Li, J. Huang, K. Yao, D. Yu, F. Seide, M. L. Seltzer, G. Zweig, X. He, J. Williams, Y. Gong, and
A. Acero. Recent advances in deep learning for speech research at microsoft. ICASSP, 2013.
16
[11] Yarin Gal and Zoubin Ghahramani. Dropout as a Bayesian approximation: Representing model uncertainty in
deep learning. arXiv:1506.02142, 2015.
[12] Ian J. Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron C. Courville, and Yoshua Bengio. Maxout networks.
In ICML, pages 1319–1327, 2013.
[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level
performance on imagenet classification. In ICCV, pages 1026–1034, 2015.
[14] D. P. Helmbold and P. M. Long. On the inductive bias of dropout. JMLR, 16:3403–3454, 2015.
[15] G. E. Hinton. Dropout: a simple and effective way to improve neural networks, 2012. videolectures.net.
[16] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R. Salakhutdinov. Improving neural networks
by preventing co-adaptation of feature detectors, 2012. Arxiv, arXiv:1207.0580v1.
[17] Nal Kalchbrenner, Edward Grefenstette, and Phil Blunsom. A convolutional neural network for modelling
sentences. In ACL, pages 655–665, 2014.
[18] P. M. Long and R. A. Servedio. Random classification noise defeats all convex potential boosters. Machine
Learning, 78(3):287–304, 2010.
[19] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In ICML,
pages 807–814, 2010.
[20] Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. Norm-based capacity control in neural networks. In
COLT, pages 1376–1401, 2015.
[21] Jürgen Schmidhuber. Deep learning in neural networks: An overview. Neural Networks, 61:85–117, 2015.
[22] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: A simple way to prevent
neural networks from overfitting. JMLR, 15:1929–1958, 2014.
[23] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan,
Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. CVPR, 2015.
[24] Tensorflow, 2016. https://www.tensorflow.org.
[25] Torch, 2016. http://torch.ch.
[26] T. Van Erven, W. Kotłowski, and M. K. Warmuth. Follow the leader with dropout perturbations. COLT, pages
949–974, 2014.
[27] S. Wager, W. Fithian, S. Wang, and P. S. Liang. Altitude training: Strong bounds for single-layer dropout.
NIPS, 2014.
[28] S. Wager, S. Wang, and P. Liang. Dropout training as adaptive regularization. NIPS, 2013.
[29] L. Wan, M. Zeiler, S. Zhang, Y. Le Cun, and R. Fergus. Regularization of neural networks using dropconnect.
In ICML, pages 1058–1066, 2013.
[30] T. Zhang. Statistical behavior and consistency of classification methods based on convex risk minimization.
Annals of Statistics, 32(1):56–85, 2004.
17
A
Table of Notation
Notation
1set
(x, y)
σ(·)
W
w, v
W(x)
P
Px,y
RP (W)
q, p
R
r, s
D(W, x, R)
JD (W)
J2 (W)
λ
WD
WL2
n, d
K
B
Meaning
indicator function for “set”
an example with feature vector x and label y
the rectified linear unit computing max(0, ·)
an arbitrary weight setting for the network
specific weights, often subscripted
the output value produced by weight setting W on input x
an arrelbitrary source distribution over (x, y) pairs
the source distribution concentrated on the single example (x, y)
the risk (expected square loss) of W under source P
probabilities that a node is dropped out (q) or kept (p) by the dropout process
a dropout pattern, indicates the kept nodes
dropout patterns on subsets of the nodes
Output of dropout with network weights W, input x, and dropout pattern R
the dropout criterion
the L2 criterion
the L2 regularization strength parameter
an optimizer of the dropout criterion
an optimizer of the L2 criterion
the network width and depth
the number of input nodes
Proof of Theorem 7
Here we prove Theorem 7, showing that the weight-decay aversion depends on the values of the inputs and
the number of input nodes K. Furthermore, unlike the single-layer case, the L2 regularization strength has
a threshold where the minimizer of the L2 criterion degenerates to the all-zero network.
We will focus on the standard architecture with depth d = 2. Recall that we are analyzing the distribution
P(x,1) that assigns probability 1/2 to (x, 1) and probability 1/2 to (0, 0).
Also, recall that, for P(x,1) , the weight-decay aversion is the maximum risk incurred by a minimizer of
J2 . This motivates the following definition.
Definition 21 An aversion witness is a minimizer of J2 that is also a maximizer of the risk R, from among
minimizers of J2 .
The proof of Theorem 7 involves a series of lemmas.
We first show that there is a aversion witness WL2 with a special form, and then relate any hidden node’s
effect on the output to the regularization penalty on the weights in and out of that node. This will allow us
to treat optimizing the L2 criterion as a one-dimensional problem, whose solution yields the theorem.
For some minimizer WL2 of J2 , let v∗j denote the vector of weights into hidden node j and wj∗ denote the
weight from j to the output node. Let a∗j be the bias for hidden node j and let b∗ be the bias for the output
node. Let hj be the function computed by hidden node j.
Note that, if there is an aversion witness with a certain property, then we may assume without loss of
generality that WL2 has that property.
Lemma 22 We may assume without loss of generality that for each hidden node j, there is an input x̃ ∈
{0, x} such that hj (x̃) = 0.
Proof: Suppose neither of hj (0) or hj (x) was 0. If wj∗ = 0, then replacing both with 0 does not affect
the output of WL2 , and does not increase the penalty. If wj∗ 6= 0, then subtracting min{hj (0), hj (x)}
18
from a∗j and adding
min{hj (0),hj (x)}
wj∗
min{hj (0), hj (x)} = 0.
to b does not affect WL2 or the penalty, but, after this transformation,
Lemma 23 We may assume without loss of generality that for each hidden node j, we have |v∗j · x| =
max{hj (0), hj (x)}.
Proof: If hj (x) ≥ hj (0) = 0, then a∗j = 0, and hj (x) = v∗j · x.
If hj (0) > hj (x) = 0. Then, a∗j > 0, and v∗j · x ≤ −a∗j . If needed, the magnitude of v∗j can be decreased
to make v∗j · x = −a∗j . This decrease does not affect WL2 (x) or WL2 (0), and can only reduce the L2 penalty.
Lemma 24 We may assume without loss of generality that for each hidden node j, we have hj (0) = 0.
Proof: Suppose hj (0) > 0. Let z be this old value of hj (0). Then hj (x) = 0 and z = hj (0) = −v∗j · x. If
we negate v∗ and set aj = 0, then Lemma 23 implies that we swap the values of hj (x) and hj (0).
Then, by adding zwj∗ to b∗ and negating wj∗ , we correct for this swap at the output node and do not
affect the function computed by WL2 or the penalty.
Note that Lemma 24 implies that a∗1 = ... = a∗K = 0.
Lemma 25 For all j, v∗j · x ≥ 0.
Proof: Since a∗j = 0, if v∗j · x < 0, we could make WL2 compute the same function with a smaller penalty
by replacing v∗j with 0.
Lemma 25 implies that the optimal WL2 computes the linear function,
WL2 (x̃) = (w∗ )T V ∗ x̃ + b∗ .
(9)
Later we will call (w∗ )T V ∗ x̃ the activation at the output node.
Lemma 26 b∗ =
1−(w∗ )T V ∗ x
.
2
Proof: Minimize J2 (wrt distribution P(x,1) ) as a function of b using Calculus.
Now, we have
WL2 (x̃) =
1
+ (w∗ )T V ∗ (x̃ − x/2)
2
(10)
which immediately implies
WL2 (0) = 1 − WL2 (x).
(11)
Lemma 27 Each v∗j is a rescaling of x.
Proof: Projecting v∗j onto the span of x does not affect hj , and cannot increase the penalty.
Lemma 28 WL2 (x) ≤ 1.
Proof: By (11), if WL2 (x) > 1 then WL2 (0) < 0 and the loss and the penalty would both be reduced by
scaling down w∗ .
Lemma 29 WL2 maximizes W(x) over those weight vectors W that have the same penalty as WL2 and
compute a function of the form W(x̃) = 12 + (w)T V (x̃ − x/2) (i.e. Equation 10).
19
Proof: Let W maximize W(x) over the networks considered. If WL2 (x) < W(x) ≤ 1, then W would have
the same penalty as WL2 but smaller error, contradicting the optimality of WL2 .
f obtained by scaling down the weights in the output layer so that
If W(x) > 1, then the network W
f
W(x) = 1 has a smaller penalty than WL2 and smaller error, again contradicting WL2 ’s optimality.
Informally, Lemmas 28 and 29 engender a view of the learner straining against the yolk of the L2 penalty
to produce a large enough output on x. This motivates us to ask how large W(x) can be, for a given value
of ||W||22 (recall that Lemma 25 allows us to assume that the biases at the hidden nodes are all 0).
Definition 30 For each hidden node j, let αj be the constant such v∗j = αj x, so that hj (x) = αj x · x.
Recall that the activation at the output node on input x is the weighted sum of the hidden-node outputs,
(w∗ )T V ∗ x.
Definition 31 The contribution to the activation at the output due to hidden node j is wj∗ hj (x) = wj∗ αj x · x
and the contribution to the L2 penalty from these weights is λ2 (wj∗ )2 + α2j x · x .
We now bound the contribution to the activation in terms of the contribution to the L2 penalty. Note
that as the L2 “budget” increases, so does the the maximum possible contribution to the output node’s
activation.
Lemma 32 If B is hidden node j’s weight-decay contribution, (wj∗ )2 + α2j x · x, then hidden node k’s contriq
q
B
, where it achieves the
bution to the output node’s activation is maximized when wj∗ = B2 and αj = 2x·x
√
value B x · x/2
q
Proof: Since α2j x · x + (wj∗ )2 = B, we have wj∗ = B − α2j x · x, so the contribution to the activation
q
can be re-written as αj x · x B − α2j x · x. Taking the derivative with respect to αj , and solving, we get
q
q
B
B
αj = ± 2x·x
and we want the positive solution (otherwise the node outputs 0). When αj = 2x·x
we have
q
wj∗ = B2 and thus the node’s maximum contribution to the activation is
r
B
2
r
√
B x·x
B
(x · x) =
.
2x · x
2
Lemma 33 The minimum sum-squared weights for a network W (without biases at the hidden nodes) that
has an activation A at the output node on input x is √2A
.
x·x
√
Proof: When maximized, the contribution of each hidden node to the activation at the output is x · x/2
times the hidden node’s contribution to the sum of squared-weights. Since each weight in W is used in
exactly one hidden node’s contribution to the output node’s activation, this completes the proof.
Note that this bound is independent of n, the number of hidden units, but does depend on the input x.
Proof (of Theorem 7): Let A ≥ 0 be the activation at the output node for WL2 on input x. From
Lemma 26 we get that b∗ = 1−A
2 . Combining Lemmas 28 , 29 and 33, we can re-write the J2 criterion for
WL2 and distribution P(x,1) in terms of A as follows.
λ
1
1 ∗
2
2
(b − 0) + (WL2 (x) − 1) + ||W||22
2
2
2
!
2
2
1−A
1
1+A
2A
+
=
− 1 + λ√
.
2
2
2
x·x
J2 (WL2 ) =
20
(12)
Differentiating with respect to A, we see that the criterion is minimized when
A=1− √
2λ
x·x
2λ
≤1
x·x
2λ
>1
when √
x·x
when √
A=0
since we assumed A ≥ 0; when A = 0 then WL2 has all zero weights with a bias of 1/2 at the output.
The risk part of (12) simplifies to
λ2
(1 − A)2
=
,
4
x·x
1 λ2
so the overall the risk of an aversion witness is min
.
,
4 x·x
21
| 9cs.NE
|
To appear in IEEE/ACM Design Automation Conference 2017
Hardware-Software Codesign of Accurate, Multiplier-free
Deep Neural Networks
Hokchhay Tann, Soheil Hashemi, R. Iris Bahar, Sherief Reda
School of Engineering, Brown University, Providence RI 02912
arXiv:1705.04288v1 [cs.NE] 11 May 2017
{hokchhay tann, soheil hashemi, iris bahar, sherief reda}@brown.edu
ABSTRACT
While Deep Neural Networks (DNNs) push the state-ofthe-art in many machine learning applications, they often
require millions of expensive floating-point operations for
each input classification. This computation overhead limits the applicability of DNNs to low-power, embedded platforms and incurs high cost in data centers. This motivates
recent interests in designing low-power, low-latency DNNs
based on fixed-point, ternary, or even binary data precision. While recent works in this area offer promising results,
they often lead to large accuracy drops when compared to
the floating-point networks. We propose a novel approach
to map floating-point based DNNs to 8-bit dynamic fixedpoint networks with integer power-of-two weights with no
change in network architecture. Our dynamic fixed-point
DNNs allow different radix points between layers. During
inference, power-of-two weights allow multiplications to be
replaced with arithmetic shifts, while the 8-bit fixed-point
representation simplifies both the buffer and adder design.
In addition, we propose a hardware accelerator design to
achieve low-power, low-latency inference with insignificant
degradation in accuracy. Using our custom accelerator design with the CIFAR-10 and ImageNet datasets, we show
that our method achieves significant power and energy savings while increasing the classification accuracy.
1.
INTRODUCTION
Recent availability of high-performance computing platforms has enabled the success of deep neural networks (DNNs)
in many demanding fields, especially in the domains of machine learning and computer vision. At the same time, applications of DNNs have proliferated to platforms ranging
from data centers to embedded systems, which open up new
challenges in low-power, low-latency implementations that
can maintain state-of-the-art accuracy. While systems with
general purpose CPUs and GPUs are capable of processing
very large DNNs, they have high power requirements and
are not suitable for embedded systems, which has led to in-
ACM ISBN 978-1-4503-2138-9.
DOI: 10.1145/1235
creasing interest in the design of low-power custom hardware
accelerators.
In designing low-power hardware for DNNs, one major
challenge stems from the high precision used in the network
parameters. State-of-the-art DNNs in classification accuracy are typically implemented using single precision (32bit) floating-point, which requires large memory size for both
the network parameters as well as the intermediate computations. Complex hardware multipliers and adders are also
needed to operate on such representations.
On the other hand, the inherent resiliency of DNNs to insignificant errors, has resulted in a wide array of hardwaresoftware codesign techniques targeted for lowering the energy and memory footprint of these networks. Such techniques broadly aim either to lower the cost of each operation
by reducing the precision [9, 10, 14] or to lower the number
of required operations, for example by knowledge distillation
[7, 16, 17].
While previous studies offer low-precision DNNs with little reduction in accuracy, the smallest fixed-point solutions
proposed require 8-bits or more for both the activation and
network parameters. Furthermore, while solutions with binary and ternary precisions prove effective for smaller networks with small datasets, they often lead to unacceptable
accuracy loss on large datasets such as ImageNet [19]. In
addition, these low-precision network techniques usually require precision specific network designs and therefore cannot
readily be used on a specific network without an expensive
architecture exploration.
In this work, we aim to tackle the low-power high-accuracy
challenge for DNNs by proposing a hardware-software codesign solution to transform existing floating-point networks
to 8-bit dynamic fixed-point networks with integer powerof-two weights without changing the network topology. The
use of power-of-two weights enables a multiplier-free hardware accelerator design, which efficiently performs computation on dynamic fixed-point precision. More specifically,
our contributions in this paper are as follows:
• We propose to compress floating-point networks to 8bit dynamic fixed-point precision with integer powerof-two weights. We then propose to fine-tune the quantized network using student-teacher learning to improve classification accuracy. Our technique requires
no change to the network architecture.
• We propose a new multiplier-free hardware accelerator for DNNs and synthesize it using an industry level
library. Our custom accelerator efficiently operates using 8-bit multiplier-free dynamic fixed-point precision.
To appear in IEEE/ACM Design Automation Conference 2017
• We also propose to utilize an ensemble of dynamic
fixed-point networks, resulting in improvements in classification accuracy compared to the floating-point counterpart, while still allowing large energy savings.
• We evaluate our methodologies on two state-of-the-art
and demanding test sets, namely CIFAR-10 and ImageNet, and use well-recognized network architectures
for our experiments. We compare our solution against
a baseline floating-point accelerator and quantify the
power and energy benefits of our methodology.
The rest of our paper is organized as follows. In Section 2,
we provide a breif background on deep neural networks. In
Section 3, we summarize previous work related to ours. Section 4 describes our methodologies and accelerator design.
Next, in Section 6, we provide the results obtained from our
methodologies and our custom accelerator. Here we discuss
the performance from both hardware and accuracy perspectives. Finally, in Section 7 we conclude our work.
2.
BACKGROUND
Figure 1 shows the template structure of a deep neural
network. While a large number of layer types are available
in the literature, three layer types are more commonly used
in DNNs:
• Convolutional Layers: Each neuron in this layer is
connected to a subset of inputs with the same spatial dimensions as the kernels, which are typically 3dimensional as shown in Figure 1. The
opPconvolution
P P
eration can be formulated as y = b+ i j k (xi,j,k ·
wi,j,k ). Here, x is the input subset, w is the kernel
weight matrix, and b is a scalar bias. These layers are
used for feature extractions.
• Pooling Layers: Pooling layers are simply used to down
sample input data.
• Fully-Connected Layers: These layers are similar to
convolutional layers with differences being that inputs
and kernels are one-dimensional vectors. These layers
are often used toward the end as classifier, where the
output vector from the final layer (logits) is fed to a
logistic function.
• Non-Linearity: For each scalar input x, this layer outputs σ(x), where σ(·) is a predefined non-linear function, such as tanh(·), rectify linear unit (ReLU), etc.
DNNs typically are based on floating-point precision and
trained with backpropagation algorithm. Each training step
involves two phases: forward and backward. In the forward
phase, the network is used to perform classification on the
input. Afterward, the gradients are propagated back to each
layer in the backward phase to update the network’s parameters. The biggest portion of the computational demands
are required by the multiplier blocks utilized in the convolutional and fully connected layers.
3.
RELATED WORKS
Previous work in software and hardware implementation
of DNNs has been, for the most part, disconnected. Few
studies have tried to optimize highly accurate designs with
low power budgets. On the accuracy front, one aspect of
condensing DNNs is to train much smaller networks from the
large, cumbersome models [16, 17]. Both models are based
Forward Pass
Backward Pass (Gradient Descent)
…
Input Layer
Convolutional Layer
Pooling Layer
Fully Connected Layer
Output Layer
Figure 1: Typical Structure of Deep Neural Network
on floating-point precision. This approach proposes to train
the student (smaller model) to mimic to the outputs of the
teacher (larger model). The loss function for the training is
composed of two parts: the losses with respect to the true
labels and the outputs from the teacher model.
Alternatively, DNNs with low precision data formats have
enormous potentials for reducing hardware complexity, power
and latency. Not surprisingly, there exists a rich body of literature which studies such limited precisions. Previous work
in this area have considered a wide range of reduced precision including fixed point [5, 13, 15], ternary (-1,0,1) [12] and
binary (-1,1) [8, 14]. Furthermore, comprehensive studies of
the effects of different precision on deep neural networks
are also available. Gysel et al. [10] propose Ristretto, a
hardware-oriented tool capable of simulating a wide range of
signal precisions. While they consider dynamic fixed-point,
in their work the focus is on network accuracy so the hardware metrics are not evaluated. On the other hand, Hashemi
et al. [9] provide a broad evaluation of different precisions
and quantizations on both the hardware metrics and network accuracy. However, they do not evaluate dynamic fixed
point.
In the hardware design domain, while few work have considered different bit-width fixed-point representations in their
accelerator design [6, 9, 18], in contrast to the accuracy analysis, no evaluation of hardware designs using dynamic fixedpoint is available. We fill this gap by providing an accelerator design optimized to use dynamic fixed-point representation for intermediate computations while using power-of-two
weights.
In recent years, a few works have focused on techniques
to reduce the power demands of DNNs at the cost of small
reductions in network accuracy. For instance, Tann et al.
propose an incremental learning algorithm where the network in trained in incremental steps [11]. The idea is then
to turn off large portions of the network in order to save
energy if these portions are not needed to retain accuracy.
While this work delivers significant power and energy saving
with small network accuracy degradation, it is orthogonal to
our work and can be applied in conjunction.
Sarwar et al. propose a multiplier-less neural network
where an accurate multiplier is replaced with an alphabet
set multiplier to save power. This work however, focuses on
multi-layer perceptrons and deep neural networks are not
evaluated [4]. In contrast, we evaluate our work on both
CIFAR-10 and ImageNet and highlight that our methodology is capable of delivering significant savings in energy
while even showing improvements in accuracy.
4.
MULTIPLIER-FREE DYNAMIC FIXEDPOINT (MF-DFP) NETWORKS
In order to simplify the hardware implementation, we propose to alter the compute model by replacing multipliers
To appear in IEEE/ACM Design Automation Conference 2017
with shift blocks and reducing signal bit width to 8 bits.
We represent the signals using dynamic fixed-point format
since synaptic weights and signals in different layers can vary
greatly in range. Employing a uniform fixed-point representation across the layers would require large bit widths
to accommodate for such range. As demonstrated by others [9, 10], even with 16-bit fixed-point, significant accuracy
drop is observed when compared to floating-point representation.
Dynamic fixed-point representation, as proposed in [13],
can be represented using two variables hb, f i, where b ∈
N+ is the bit width, and f ∈ Z is the fractional length.
Each b-bitPnumber in this scheme can be interpreted as
i
(−1)s 2−f b−2
i=0 2 xi , where s, xi ∈ {0, 1} are the sign bit
th
and the i bit respectively. The term dynamic refers to the
fact that different layers in DNNs can take on different values for f depending on their ranges. In this work, we deploy
8-bit dynamic fixed-point for all of our experiments.
While we adopt our quantization process from the techniques in [10], our work differs from theirs in three aspects:
(i) we perform hardware-software analysis for power-of-two
weights and dynamic fixed-point data path, (ii) we propose
to include student-teacher learning in the fine-tuning process, and (iii) we demonstrate that an ensemble of two MFDFP networks can outperform the floating-point network
while having significant savings in energy. These aspects are
described in Algorithm 1 as three phases. Next, we describe
these phases in more details.
Algorithm 1: Floating-Point to Dynamic Fixed-Point
1
2
3
4
5
6
7
8
9
10
11
12
13
14
4.1
Network Quantization (Phase 1)
In order to construct a dynamic fixed-point network, we
take as input a fully trained floating-point network. We first
quantize on this input network by rounding its weights to
the nearest powers of two. We also round the intermediate
signals to 8-bit dynamic fixed-point using Ristretto [10] (line
2). We then perform fine-tuning on the network to recover
from accuracy loss due to quantization (lines 1–9).
DNNs are typically trained using the backpropagation algorithm with variants of gradient descent methods, which
can be ill-suited for low-precision networks. The computed
gradients and learning rates are typically very small, which
means that parameters may not be updated at all due to
their low-precision format. Intuitively, this requires high
precision in order to converge to a good minima. However, integer power-of-two weights only allow large increment jumps.
To combat this disparity, we adopt solution proposed by
Courbariaux et al. [14] to keep two sets of weights during
the training process: one in quantized precision and one in
floating-point. As shown in Algorithm 1, during forward
propagation, the floating-point weight set is stochastically
or deterministically quantized before the input data is evaluated (line 4). For our work, we found that deterministic
quantization gives better performance. The output result of
the quantized network is then used to compute the loss with
respect to the true label of the data (line 5). The gradients
with respect to this loss are then used to update the floatingpoint parameters during backward propagation (line 6), and
the process is repeated until convergence. This approach allows small gradients to accumulate over time and eventually
cause incremental updates in the quantized weights.
4.2
Additional Fine-tuning (Phase 2)
15
16
17
18
19
20
// FLnet: the input floating-point net.
// t_logits: the floating-point net’s logits.
Input : FLnet, t logits
begin Phase 1
// MF_DFPnet: dynamic fixed-point network
Output: MF DFPnet
// Quantize the weights to powers of two and
// inputs to 8-bit fixed-point
MF DFPnet = Quantize 8bit(FLnet);
// Fine-tuning MF_DFPnet until convergence
// using hard data labels as in [14].
for i = 1 to Convergence do
Forward Pass(MF DFPnet);
// Compute gradients
grads=Grad(MF DFPnet, true labels);
// Backpropagate grad and update weights:
Backward update(FLnet, grads);
MF DFPnet = Quantize 8bit(FLnet);
end
end
begin Phase 2
// Additional training using different
for j = i to Convergence do
Forward Pass(MF DFPnet);
grads=Grad(MF DFPnet, true labels);
// Also with respect to teacher’s logits
grad logit=Grad(MF DFPnet, t logits);
// gradient to update:
gradients=grads + β·grad logit;
Backward update(FLnet, gradients);
MF DFPnet = Quantize 8bit(FLnet);
end
return MF DFPnet
end
// Phase 3: If using ensemble, repeat Phase 1
// and 2 with different input FLnet.
On top of the technique from Courbariaux et al. [14], we
propose additional training with a different loss function
once training with hard labels no longer improves the performance. As shown in Algorithm 1 lines 10–20, in addition
to using hard labels, we introduce student-teacher learning,
where a student network is trained to mimic the outputs of
a teacher network [16, 17]. Both networks are floating-point
based, but the student typically has a far fewer number of
parameters. In our work, we treat the dynamic fixed-point
network as the student and the floating-point network as the
teacher.
The loss function in the student-teacher learning incorporates the knowledge learned by the teacher model [16].
Suppose S is the student network, and T is the teacher
with output logit vectors zS and zT and class probability
PS and PT respectively. The softmax regression function
is relaxed by introducing a temperature parameter τ such
exp(zT ,i /τ )
exp(zS,i /τ )
and PT,i = P exp(z
. Let
that PS,i = P exp(z
S,j /τ )
T ,j /τ )
j
j
WS be the parameters of the student network, then the loss
function for the student model is defined to be:
L(WS ) = H(Y, PS ) + β · H(PT , PS )
(1)
where β is a tunable parameter, H is the cross entropy, and
Y is the one-hot true data label. Using τ >> zS , zT , we
To appear in IEEE/ACM Design Automation Conference 2017
where N is the length
P
of of vectors zS , zT . With zero-meaned zS , zT ( j zS,j =
P
j zT,j = 0), the approximated gradient is then:
8 bit <<
>>
4 bit
in
w
<<
>>
(2)
in
w
<<
>>
Ensemble of MF-DFP Networks (Phase 3)
in
w
5.
HARDWARE ACCELERATOR DESIGN
As discussed in Section 4, while we maintain low-precision
in both network signals and parameters for efficiency, providing the network with the flexibility to change the location of the radix point from layer to layer is necessary for
minimizing the accuracy degradation. While improving the
accuracy, this scheme incurs complexities in the hardware
design as some bookkeeping in needed to keep track of the
location of the radix point in different parts of the network.
In the proposed accelerator, we enable such flexibility by
providing each set of calculations with details on the indices
of both the input feature maps as well as the output activation. More specifically, we implement this feature by adding
control signals dedicated to both the input feature, and the
output activation radix indices. Dedicated hardware is then
added to the hardware to shift the result to the correct index
as determined by the radix indices.
On the other hand, while dynamic fixed-point representation for synaptic weights and activation maps allows for
compact bit widths, during inference, we would still need
to perform fixed-point multiplications. As described in Section 4, we propose to quantize the weights to integer powerof-two, which would allow the expensive multiplications to
be replaced with arithmetic shifts. These shift operators
are far more hardware-friendly than full-scale multipliers.
In this quantization scheme, for each weight w, we represent
its quantized version using two numbers hs, ei, where s is the
sign of the weight w, and e = max[round(log2 (|w|)), −7] is
the exponent for the power of 2 (i.e., 2e ). Here, round() performs rounding to the nearest integer. Note that we bound
e ≥ −7 since our input data is limited to 8 bits. For each input x, x·w is then transformed into (s·x) e, where
represents the shift operator. In addition, we observe that
+
n
17 bit
+ 18 bit
+
<<
>>
. . .
Deploying an ensemble of DNNs has been proven to be a
simple and effective method to boost the inference accuracy
of a DNN [21]. The idea is to independently train multiple
DNNs of the same architecture and use them to evaluate
each input. The output is then chosen based on the majority of votes. Suppose the ensemble consists of M networks
producing output logit vectors zi , i ∈ [1, M ]. Then the outPM
1
put class can simply be the maximum element in M
i=1 zi .
This idea is amenable in scenarios where there exists enough
time or energy budget to justify evaluating the input on a
number of networks. In section 6.2 we highlight that, since
the reduction in energy from the proposed MF-DFP are so
dramatic, the designer may implement an ensemble of MFDFP networks in parallel and still save significantly in energy consumption. More specifically, we show that an ensemble of multiplier-free dynamic fixed-point networks can
outperform a floating-point network while still achieving significant energy saving. In order to construct such ensemble,
we run Algorithm 1 multiple times with different starting
floating-point networks on line 1.
m
16
bit
+
19 bit
+
Accumulator
&
Routing
NL
clk
(b) Proposed Hardware Accelerator Design
Memory Subsystem
Memory Interface
δL(WS )
β
≈ (PS,i − Yi ) +
· (zS,i − zT,i ).
δzS,i
N · τ2
4.3
(a) The Dynamic, Fixed-Point Structure of a Single Neuron.
in
w
8 bit
1+z
P i /τ
N + j zj /τ
8 bit
≈
20 bit
Pexp(zi /τ )
j exp(zj /τ )
. . .
have Pi =
DMA
Input
Buffer
DMA
Weights
Buffer
DMA
Output
Buffer
Neural Processing Unit (NPU)
Processing Unit #1
Processing Unit #2
Processing Unit #16
Control Circuitry
Figure 2: Proposed Hardware Accelerator Design
(a) a Single Neuron and (b) The Organization of
Neurons and Hardware Blocks.
the magnitudes of the weights is less than 1, so our rounding leads to 8 possible exponents {0, −1, . . . , −7}. Therefore
the weights can be encoded into 4-bit representation. This
observation is used to simplify our hardware architecture
significantly as discussed in 6.2.
To further improve the accuracy, we ensure that there is
no loss in intermediate values by mitigating the possibility
of overflows. In order to do so, we ensure that all intermediate signals have large enough word-width, thereby effectively increasing the width of the intermediate wires as
needed. To illustrate our idea, Figure 2(a) shows the simplified structure of a single neuron in our proposed implementation, highlighting the main feature of the accelerator design. In Figure 2(a), the dedicated hardware implementing
the dynamic fixed-point scheme is shown as “Accumulator
& Routing”. Here m and n represent the locations of the
radix points for the input features and output activations
respectively.
In order to integrate our proposed neuron architecture
into a full-scale hardware accelerator, we utilize a tile-based
implementation inspired from DianNao [3], where each cycle a small number of physical neurons is fed a new set of
data for calculation. We implement three separate memory subsystems assigned to input data, weights, and output
data, respectively. This memory subsystem ensures the isolation of memory transfers from the calculation for maximum throughput. The computation itself is performed in
neural processing units (NPUs) containing a number of processing units each implementing 16 neurons with 16 synapses.
Figure 2(b) illustrates the organization of the proposed
To appear in IEEE/ACM Design Automation Conference 2017
Table 1: Design metrics of the proposed MF-DFP
accelerator against the floating-point baseline.
6.1
EXPERIMENTAL RESULTS
Experimental Setup
In this section, we discuss our results on the CIFAR-10
and ImageNet 2012 datasets [2, 19] using the well-known
DNN architectures from [2] and [20] respectively. We remove
all local response normalization layers since they are not
amenable to our multiplier-free hardware implementation.
All of our experiments are based on Caffe [1].
For CIFAR-10, we begin by training the floating-point
networks using the benchmark architecture. For the ImageNet benchmark, we obtain the floating-point model from
Caffe Model Zoo1 . We then run the networks on their corresponding training set data to obtain the pre-softmax output
logits. From these floating-point networks, we construct our
proposed MF-DFP networks using Algorithm 1.
For our hardware evaluations, we compile our designs using Synopsys Design Compiler and a 65 nm standard cell
library in the typical processing corner. We synthesize our
hardware so that we have zero timing slack for the floatingpoint design. Therefore, we use a constant clock frequency
of 250MHz for all our experiments. While the utilization of
barrel shifters instead of multipliers provides us with timing
slacks which can be used to boost the frequency, we choose
to keep the frequency constant as changing the frequency
adds another dimension for evaluation which is out of the
scope of this work.
6.2
Results
We evaluate our proposed methodology as well as our custom hardware accelerator on CIFAR-10 and ImageNet using
a broad range of performance metrics including accuracy,
power consumption, design area, inference time, and inference accuracy. Table 1 summarizes the design area and the
power consumption of the proposed multiplier-free custom
accelerator. Values shown in parenthesis, (in,w), reflect the
number of bits required for the representation of inputs and
weights respectively. We also implement a floating-point
1
https://github.com/BVLC/caffe/wiki/Model-Zoo
Power
Cons.
(mW )
1361.61
138.96
270.27
Area
Saving
(%)
0
87.97
76.00
Power
Saving
(%)
0
89.79
80.15
0.52
Training with Data labels only
Student-Teacher Training
Floating-Point Network
0.51
0.5
0.49
0.48
0.47
0.46
0.45
0.44
Phase 1
Phase 2
0.43
0
6.
Design
Area
(mm2 )
16.52
1.99
3.96
Precision (in, w)
Floating-point(32,32)
Proposed MF-DFP(8,4)
Ens. MF-DFP(8,4)
Error Rate
hardware accelerators. Here we want to stress the benefits
of our methodologies relative to the floating-point design.
Thus, an architectural design space exploration such as altering number of hardware neurons and synapses is out of
the scope of this work.
In order to incorporate the proposed ensemble of networks, the number of processing units is increased as needed
to parallelize the computation of an ensemble of networks.
Note that the memory subsystems as well as the control
logic also need to be modified to account for the number
of processing units. In section 6.2 we evaluate our methodologies using a single processing unit, resulting in a single
multiplier-free dynamic fixed point (MF-DFP) network, and
two processing units, which form an ensemble of two networks.
We also implement and compare our hardware design with
a conventional 32-bit floating-point architecture using a single processing unit as a baseline. Compared to our proposed
design, the baseline implementation utilizes multipliers in
the first stage of the design and keeps the bitwidth constant
at 32-bits throughout the design for both the activations and
the network parameters.
10
20
30
40
50
Training Epochs
Figure 3: ImageNet 2012 validation set top-1 error
rates for quantized network trained with only data
labels, with student-teacher learning, and floatingpoint network.
version of our accelerator as a baseline design and for comparison. As shown in the table our accelerator can achieve
significant benefits in both design area and power consumption using both one processing unit and using an ensemble
of two networks. Next we report the results when using
our methodologies and hardware accelerator for our benchmarks.
Figure 3 shows the classification error rate of the baseline floating-point network as well as the fine-tuning process
of MF-DFP for the ImageNet benchmark. Here, we observe that by fine-tuning using just data labels (Phase 1),
we achieve significant performance with less than a 1% increase in error rate than the floating-point counterpart. Additional training using the student-teacher model (Phase 2)
as described in Section 4.2 on top of just data labels, allows
us to reduce the error rate even more. In this experiment,
we observed that more benefit is achieved when the studentteacher training is started from a non-global optimal point
in the data labels-only training. More specifically, the value
of i in Algorithm 1 line 11 should be close to convergence but
not the global optimal point in the training process. In either case, the student-teacher learning provides consistently
better performance than using the data labels-only training.
For this training, we set τ = 20, β = 0.2 and start with a
learning rate of 1e-03. We decrease the rate by a factor of
10 when learning levels off and stop the training when the
learning rate drops below 1e-07.
Furthermore, in Table 2, we summarize the accuracy, inference time, and the energy performance of our proposed
techniques. As shown in the table, our methodology can
achieve energy savings as high as 89% in the case of single
MF-DFP network with a maximum of 0.79% degradation in
accuracy for both benchmarks. This is especially significant
as there is absolutely no modification to network depth and
To appear in IEEE/ACM Design Automation Conference 2017
Table 2: Time, energy and top-1 accuracy for CIFAR-10 and ImageNet. In addition, for Imagenet, we also
show in parenthesis the top-5 accuracy. Ensemble accuracy is obtained by deploying two MF-DFP networks.
Precision
Floating-Point (32,32)
MF-DFP (8,4)
Ensemble MF-DFP
Classification
Accuracy (%)
81.53
80.77
82.61
CIFAR-10
Time
Energy
(us)
(uJ)
246.52
335.68
246.27
34.22
246.27
66.56
Energy
Saving (%)
0
89.81
80.17
Table 3: Comparison of memory requirements for
floating-point versus MF-DFP network parameters
for CIFAR-10 and ImageNet benchmarks.
Precision
Floating-Point
MF-DFP
Ensemble MF-DFP
CIFAR-10 (MB)
0.3417
0.0428
0.0855
ImageNet (MB)
237.95
29.75
59.50
channel size. In addition, with the extra area budget, we
can implement two processing units in our accelerator and,
for each benchmark, we deploy an ensemble of two MF-DFP
networks trained using different starting points. As shown
in Table 2, we can outperform the floating networks in both
benchmarks using this ensemble while still achieving significant energy saving.
Finally, while we design our methodology with memory
footprint in mind, we do not include the power consumption
of the main memory subsystem in our evaluations. However, as a general guideline, our methodology emphasizes on
reductions in network precisions and therefore requires 8×
less memory compared to a floating-point implementation as
shown in Table 3. For the ensemble method, the memory requirement essentially doubles from single MF-DFP, however,
they are still far lower than the floating-point networks.
7.
CONCLUSION
In this work we proposed a novel hardware-software codesign approach which enables seamless mapping of full-precision
deep neural networks to a multiplier-free dynamic fixedpoint network. In our work, no change to the network architecture is required to maintain accuracy within acceptable
bounds. We also formalized the use of student-teacher learning for accuracy improvements in low-precision networks. In
addition, we proposed a hardware design capable of incorporating the dynamic fixed point as well as the multiplier-free
design aspects. We proposed to utilize an ensemble of lower
precision ML-DFP networks to increase the accuracy even
further. We evaluated our designs using two well-recognized
and demanding datasets, namely CIFAR-10 and ImageNet
running on networks well studied in the literature. Using
one single MF-DFP network on our tastbenches, our design
achieves up to 90% energy savings with an insignificant accuracy drop of approximately 1%. Using an ensemble of two
networks, energy savings of 80% are achievable while delivering accuracy gains of more than 1% for CIFAR-10, and
0.5% for ImageNet top-1 classification accuracy.
Acknowledgment
This work is supported by NSF grant 1420864. We would
like to thank NVIDIA Corporation for their generous GPU
donation.
Classification
Accuracy (%)
56.95 (79.88)
56.16 (79.13)
57.57 (80.29)
ImageNet
Time
Energy
(us)
(uJ)
15666.45 21332.38
15666.06
2176.96
15666.06
4234.07
Energy
Saving (%)
0
89.80
80.15
References
[1] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long,
R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint
arXiv:1408.5093, 2014.
[2] A. Krizhevsky and G. Hinton. Learning multiple layers of features
from tiny images, 2009.
[3] T. Chen, Z. Du, N. Sun, J. Wang, C. Wu, Y. Chen, and O. Temam.
DianNao: A small-footprint high-throughput accelerator for ubiquitous machine-learning. In Intl. Conf. on Architectural Support
for Programming Languages and Operating Systems, ASPLOS
’14, pages 269–284, 2014.
[4] S.S. Sarwar, S. Venkataramani, A. Raghunathan, K. Roy.
Multiplier-less Artificial Neurons Exploiting Error Resiliency for
Energy-Efficient Neural Computing. In Proc. DATE, 2016.
[5] C.Z. Tang, and H.K. Kwan. Multilayer feedforward neural networks with single powers-of-two weights. In IEEE Transactions
on Signal Processing, pages 2724–2727, 1993.
[6] M. Sankaradas, J. Murugan, V. Jakkula, S. Cadambi, S. Chakradhar, I. Durdanovic, E. Cosatto, and H.P. Graf. A Massively Parallel Coprocessor for Convolutional Neural Networks. In Proc.
IEEE ASAP, 2009.
[7] A. Romero, N. Ballas, S.E. Kahou, A. Chassang, C. Gatta, and
Y. Bengio. FitNets: Hints for Thin Deep Nets. arXiv preprint
arXiv:1412.6550, 2014.
[8] D. Soudry, I. Hubara, and R. Meir. Expectation backpropagation: Parameter-free training of multilayer neural networks with
continuous or discrete weights. In Proc. NIPS, pages 963–971,
2014.
[9] S. Hashemi, N. Anthony, H. Tann, R.I. Bahar, and S. Reda. Understanding the Impact of Precision Quantization on the Accuracy
and Energy of Neural Networks. In Proc. DATE, 2017.
[10] P. Gysel, M. Motamedi, S. Ghiasi, “Hardware-Oriented Approximation of Convolutional Neural Networks,” in ICLR Workshop,
2016.
[11] H. Tann, S. Hashemi, I. Bahar, S. Reda, “Runtime Configurable Deep Neural Networks for Energy-Accuracy Trade-off,” in
Proceedings of IEEE/ACM/IFIP International Conference on
Hardware/Software Codesign and System Synthesis, p. 34. 2016.
[12] K. Hwang, and W. Sung. Fixed-point feedforward deep neural
network design using weights +1, 0, and -1. In IEEE SiPS, 2014.
[13] M. Courbariaux, Y. Bengio, and D. Jean-Pierre. Low precision
arithmetic for deep learning. In arXiv preprint arXiv:1412.7024,
2014.
[14] M. Courbariaux, I. Hubara, D. Soudry, R. El-Yaniv, and Y. Bengio. Binarized Neural Networks: Training Deep Neural Networks
with Weights and Activations Constrained to +1 or -1. In arXiv
preprint arXiv:1602.02830, 2016.
[15] S. Gupta, A. Agrawal, K. Gopalakrishnan, and P. Narayanan.
Deep Learning with Limited Numerical Precision. In arXiv
preprint arXiv:1502.02551, 2015.
[16] G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in
a neural network. In arXiv preprint arXiv:1503.02531, 2015.
[17] C. Bucilua, R. Caruana, and A. Niculescu-Mizil Model compression. In Proc. of ACM SIGKDD, 2006.
[18] C. Zhang, P. Li, G. Sun, Y. Guan, B. Xiao, and J. Cong. Optimizing FPGA-based Accelerator Design for Deep Convolutional
Neural Networks. In Proc. ACM/SIGDA FPGA, 2015.
[19] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, and A.C. Berg,
and F. Li. ImageNet Large Scale Visual Recognition Challenge.
In IJCV, pages 211–252, 2015
[20] A. Krizhevsky, I. Sutskever, and G.E. Hinton Imagenet classification with deep convolutional neural networks. In Proc. NIPS,
2012.
[21] J. Ba, and R. Caruana. Do deep nets really need to be deep? In
Proc. NIPS, pages 2654–2662, 2014.
| 9cs.NE
|
Primitivity, Uniform Minimality, and State
Complexity of Boolean Operations
arXiv:1702.00877v3 [cs.FL] 26 Mar 2018
Sylvie Davies
University of Waterloo
Department of Pure Mathematics
[email protected]
Abstract. A minimal deterministic finite automaton (DFA) is uniformly
minimal if it always remains minimal when the final state set is replaced
by a non-empty proper subset of the state set. We prove that a permutation DFA is uniformly minimal if and only if its transition monoid is a
primitive group. We use this to study boolean operations on group languages, which are recognized by direct products of permutation DFAs.
A direct product cannot be uniformly minimal, except in the trivial case
where one of the DFAs in the product is a one-state DFA. However, nontrivial direct products can satisfy a weaker condition we call uniform
boolean minimality, where only final state sets used to recognize boolean
operations are considered. We give sufficient conditions for a direct product of two DFAs to be uniformly boolean minimal, which in turn gives
sufficient conditions for pairs of group languages to have maximal state
complexity under all binary boolean operations (“maximal boolean complexity”). In the case of permutation DFAs with one final state, we give
necessary and sufficient conditions for pairs of group languages to have
maximal boolean complexity. Our results demonstrate a connection between primitive groups and automata with strong minimality properties.
1
Introduction
Formal definitions are postponed until later.
The state complexity of a regular language is the minimal number of states
needed to recognize the language with a deterministic finite automaton. It is
well-known that if Lm and L′n are regular languages over a common alphabet
Σ with state complexity m and n respectively, then the state complexity of
Lm ∪ L′n is at most mn, and this bound is tight for all m, n ≥ 2. The upper
bound follows from the standard “direct product” automaton construction for
recognizing unions of regular languages. Examples which meet the bound were
given by Maslov in 1970 [15], and independently by Yu, Zhuang and Salomaa in
1994 [21], who noted that the same bound holds for intersection.
More generally, if ◦ is a binary boolean operation on languages over Σ, then
Lm ◦L′n has state complexity at most mn, and this bound is tight for all m, n ≥ 2
if and only if ◦ is proper, that is, not a constant function (L◦L′ = ∅ or L◦L′ = Σ ∗ )
or a function that depends on only one argument (for example, L ◦ L′ = Σ ∗ \ L).
This was proved by Brzozowski in 2009 [5], who gave examples showing that mn
is a tight bound for symmetric difference, and noted that the examples for union
and symmetric difference (together with their complements) suffice to prove mn
is a tight upper bound for all proper binary boolean operations.
To prove a lower bound on the worst-case state complexity of a regular operation, it suffices to give just one family of examples that meet the bound. Such
families are called witnesses. Witnesses are known for most commonly used unary
and binary operations on regular languages. However, there are several directions
of research in state complexity which necessitate finding new witnesses for operations that have previously been studied. For example, sometimes the first
witnesses found for an operation are not optimal in terms of alphabet size, so
researchers will look for new witnesses over a smaller alphabet. When studying
the n-ary versions of binary operations, such as the union of n languages, or
more generally combined operations [13,18], such as the star of a union of languages, again new witnesses are needed. It is also interesting to consider families
of languages that are simultaneous witnesses for multiple operations; it is not
generally the case that a witness for one operation will work for others. Brzozowski found a family of languages which is a simultaneous witness for reversal,
star, concatenation and all binary boolean operations [6]. Each of the problems
just mentioned, as well as the fundamental problem of determining the worstcase state complexity of an operation, may also be studied in subclasses of the
regular languages, such as the star-free languages [9] or ideal languages [8]. Often
the known witnesses do not lie in the subclass, so new witnesses must be found.
In some cases, new witnesses can be found by making slight modifications to
known witnesses, but this is not always successful. Furthermore, this technique
does little to advance our understanding of why particular witnesses work. For
these reasons, it is desirable to have results which describe the general landscape
of witnesses for a particular operation. By this we mean results that give necessary conditions for witnesses, revealing common structural properties that all
witnesses share, or sufficient conditions allowing one to easily generate examples of witnesses or check whether a candidate family is a witness. For example,
Salomaa, Wood and Yu proved that a regular language of state complexity n is
a witness for the reversal operation if the transition monoid of its minimal DFA
has the maximal possible size nn [19]; this gives a general sufficient condition
for a language to be a witness for reversal. Ideally, collecting results of this sort
would eventually lead to a complete classification of witnesses for commonly
used operations. In reality, we suspect the problem of fully classifying witnesses
is only tractable in very special cases, but even results which take small steps in
this direction can be quite useful and enlightening.
The main inspiration for this work is a paper of Bell, Brzozowski, Moreira,
and Reis [3], which considers the following question: for which pairs of languages
(Lm , L′n ) (with state complexities m and n respectively) does Lm ◦ L′n reach
the maximal state complexity mn for every proper binary boolean operation
◦? Bell et al. give sufficient conditions for this to occur. The conditions are
based on the transition monoids of the minimal deterministic automata of Lm
2
and L′n ; essentially, if the transition monoids contain the symmetric groups Sm
and Sn , then “usually” (i.e., excluding a known class of counterexamples) the
language Lm ◦ L′n will have state complexity mn. We obtain a refinement of this
result: we prove that if the transition monoids contain 2-transitive groups, then
“usually” Lm ◦ L′n has state complexity mn (though our notion of “usually” is
more restrictive than that of Bell et al.).
We also obtain necessary and sufficient conditions for Lm ◦ L′n to have state
complexity mn in the special case where the minimal automata for Lm and L′n
have exactly one final state, and their transition monoids contain a transitive
permutation group. We can view this result as solving a particular special case
of the problem of characterizing witnesses for boolean operations.
To obtain these results, we exploit a connection between a certain class of
permutation groups called primitive groups, and the notion of uniformly minimal automata introduced by Restivo and Vaglica [16]. A minimal deterministic
finite automaton (DFA) is uniformly minimal if it always remains minimal when
the final state set is replaced by a non-empty proper subset of the state set. For
a permutation DFA (that is, a DFA whose transition monoid is a permutation
group), uniform minimality is equivalent to primitivity of the transition monoid.
Although uniform minimality played an important role in the paper of Bell et
al., this connection with primitive groups was not used in their paper. Primitive
groups are an important and well-studied class of permutation groups; there are
deep results on their structure, and large libraries of primitive groups are available in computer algebra systems such as GAP [14] and Magma [4]. Uniformly
minimal DFAs have received comparatively little study; thus this connection has
significant implications for the theory of uniformly minimal DFAs.
The paper is structured as follows. Section 2 contains background material
needed to understand the paper. Section 3 discusses the relationship between
primitive groups and uniformly minimal permutation DFAs. Section 4 contains
our main results on witnesses for the maximal state complexity of boolean operations. Section 5 concludes the paper by giving a summary of our results and
stating some open problems.
2
Definitions and Notation
For a function f : X → Y , we typically write the symbol f to the right of
its arguments. For example, if the image of x under f is y, we write xf =
y. Functions are composed from left to right, and composition is denoted by
juxtaposition: if g : Y → Z, then f g denotes the composition of f and g, and
x(f g) = (xf )g = yg is an element of Z.
Let P(X) denote the power set of X, that is, the set of all subsets of X. Given
f : X → Y we may extend f by union to obtain a function f : P(X) → Ye (where
S
Ye is the closure of Y under union) defined by Sf = x∈S xf for S ⊆ X. We
denote the extension by the same symbol as the original function. Note that for
convenience, we often make no distinction between an element of a set and the
singleton containing the element; so x∪x′ = {x}∪{x′ } = {x, x′ } and xf = {x}f .
3
2.1
Monoids, Groups and Actions
A monoid is a set M equipped with an associative binary operation · and an
identity element e such that m · e = e · m = m for all m ∈ M . Typically we
omit the symbol for the operation; so the previous equation could be written as
me = em = m. For n ≥ 1 we write mn for the n-fold product of m with itself,
and define m0 = e for all m ∈ M . If for each m ∈ M , there exists m′ ∈ M such
that mm′ = m′ m = e, then M is called a group, and m′ is called the inverse of
M and denoted m−1 . The order of an element g of a group is the least integer
n ≥ 1 such that g n = e.
A submonoid of M is a subset M ′ ⊆ M which is closed under · and contains
the identity e of M . If additionally M ′ is a group, it is called a subgroup of M ;
we write M ′ ≤ M to mean that M ′ is a subgroup of M . Note that we do not
allow submonoids or subgroups of M to have an identity element different from
that of M . If x1 , . . . , xk are elements of a group G, then hx1 , . . . , xk i denotes the
group generated by x1 , . . . , xk , the smallest subgroup of G containing x1 , . . . , xk .
Let M and M ′ be monoids with identity elements e and e′ respectively. A
homomorphism from M to M ′ is a function ϕ : M → M ′ such that (m1 m2 )ϕ =
(m1 )ϕ(m2 )ϕ for all m1 , m2 ∈ M and eϕ = e′ . A bijective homomorphism is
called an isomorphism, and two monoids are said to be isomorphic if there exists
an isomorphism from one to the other. We write M ∼
= M ′ to mean that M and
′
′
M are isomorphic. If G and G are groups and ϕ : G → G′ is a homomorphism,
the kernel of ϕ is the set ker ϕ = {g ∈ G : gϕ = e′ }, that is, the set of elements of
G that map to the identity of G′ . If G is a group, N ≤ G, and gng −1 ∈ N for all
g ∈ G and n ∈ N , we say N is a normal subgroup of G. A group G is simple if it
has no non-trivial proper normal subgroups, that is, the only normal subgroups
of G are G itself and the trivial group (containing just the identity element of
G). The kernel of a homomorphism from G to another group is always a normal
subgroup of G. We occasionally use the following elementary facts about normal
subgroups and homomorphisms:
– If ϕ : G → G′ is a homomorphism and ker ϕ is the trivial one-element subgroup of G, then ϕ is injective.
– If ϕ : G → G′ is a surjective homomorphism and N is a normal subgroup of
G, then N ϕ is a normal subgroup of G′ .
A monoid action of M on a set X is a function ψ : X × M → X such
that ((x, m)ψ, m′ )ψ = (x, mm′ )ψ and (x, e)ψ = x for all m, m′ ∈ M and x ∈ X.
Equivalently, it is a family of functions mψ : X → X such that mψ m′ψ = (mm′ )ψ
for all m, m′ ∈ M and eψ is the identity map on X. The map mψ is called the
action of m. To simplify the notation, we often omit the action symbol ψ and just
write xm instead of xmψ or (x, m)ψ. Furthermore, we typically avoid assigning
a symbol to the action at all; rather than “let ψ be a monoid action of M on X”
we write “let M be a monoid acting on X”, meaning that M has a specific but
nameless action on X associated with it. If S ⊆ M generates the monoid M , a
monoid action ψ is completely determined by its values on elements of S. If M
is a group, we use the term group action rather than monoid action.
4
Let G be a group acting on X. For x ∈ X, the stabilizer subgroup or simply
stabilizer of x is the subgroup {g ∈ G : xg = x} of G. For S ⊆ X, the setwise
stabilizer of S is the subgroup {g ∈ G : Sg = S}. Elements of the setwise
stabilizer need not fix every element of S; for example, if 1g = 2 and 2g = 1
then g is in the setwise stabilizer of {1, 2}.
Let X be a finite set. A function t : X → X is called a transformation of X.
The set of all transformations of X is a monoid under composition called the
full transformation monoid TX . A submonoid of TX is called a transformation
monoid on X. The degree of a transformation monoid on X is the size of X.
If M is a transformation monoid on X, the monoid action ψ : X × M → X
given by (x, t)ψ = xt for x ∈ X, t ∈ M is called the natural action of M . If
X = {1, . . . , n} we write Tn for TX .
A bijective transformation of X is called a permutation of X. We can describe
any particular permutation of X using cycle notation as follows. For x1 , . . . , xk ∈
X, we write (x1 , . . . , xk ) for the permutation that sends xi to xi+1 for 1 ≤ i < k,
sends xk to x1 , and fixes all other elements of X. This permutation is called
a cycle of length k, or simply a k-cycle. All permutations that are not cycles
can be expressed as a product of cycles. The identity permutation is denoted
by an empty cycle, i.e., “()”. Cycle notation conflicts with the notation we use
for ordered k-tuples, but this should not cause confusion. We mainly use cycle
notation when giving concrete examples of permutations.
The set of all permutations of X is a subgroup of TX called the symmetric
group SX . A subgroup of SX is called a permutation group on X; this a special
type of transformation monoid and we have the same notions of degree and
natural action. The alternating group AX is the subgroup of SX consisting of all
permutations that can be expressed as a product of an even number of 2-cycles.
If X = {1, . . . , n} we write Sn for SX and An for AX .
Let G be a group acting on X. We say that the action of G is transitive or
that G acts transitively on X if for all x, x′ ∈ X, there exists g ∈ G such that
xg = x′ . We say the action of G is k-transitive or G acts k-transitively on X if
for all pairs of k-tuples (x1 , . . . , xk ), (x′1 , . . . , x′k ) ∈ X k , there exists g ∈ G such
that for 1 ≤ i ≤ k we have xi g = x′i ; informally, k-transitive means “transitive
on k-tuples”.
A non-empty set B ⊆ X is called a block for G if for all g ∈ G, either
Bg ∩ B = B (equivalently, Bg = B) or Bg ∩ B = ∅. A block B is trivial if it is a
singleton or the entire set X. We say the action of G is primitive or that G acts
primitively on X if it is transitive and all of its blocks are trivial. Equivalently,
a transitive group action of G is primitive if for every set S ( X with at least
two elements, there exists g ∈ G such that ∅ ( Sg ∩ S ( S.
If G is a permutation group and the natural action of G is transitive (ktransitive, primitive), then we say G is a transitive group (k-transitive group,
primitive group). For example, the cyclic group h(1, 2, 3, 4)i ≤ S4 is a transitive
group, since its natural action on {1, 2, 3, 4} is transitive. This terminology can
cause confusion, since transitivity, k-transitivity and primitivity are properties
of actions and not groups; statements like “G is transitive” or “G is primitive”
5
are statements about a particular action of G (the natural action) rather than
the abstract group itself. In particular, these properties are not preserved under
isomorphism; for example, the group h(5, 6, 7, 8)i ≤ S8 is not transitive, but it is
isomorphic to the transitive group h(1, 2, 3, 4)i ≤ S4 .
As the notions of transitivity and primitivity are central to this paper, we
give numerous examples to illustrate them below.
Example 1. Consider the group G = h(1, 2, 3, 4, 5, 6)i ≤ S6 . This group is clearly
transitive, since its natural action on {1, 2, 3, 4, 5, 6} is transitive. However, it is
imprimitive, since {1, 3, 5} and {2, 4, 6} are non-trivial blocks. Indeed, if we let
a = (1, 2, 3, 4, 5, 6), then {1, 3, 5}a = {2, 4, 6} and {2, 4, 6}a = {3, 5, 1}. Hence
for all k ≥ 0, we either have {1, 3, 5}ak ∩ {1, 3, 5} = ∅ or {1, 3, 5}ak ∩ {1, 3, 5} =
{1, 3, 5}, and similarly for {2, 4, 6}. One may also verify that {1, 4}, {2, 5} and
{3, 6} are non-trivial blocks, and that there are no blocks of size 4 or 5.
Example 2. Consider the group G = h(1, 2, 3, 4, 5)i ≤ S5 . This group is clearly
transitive, and it is also primitive. To see this, suppose for a contradiction that B
is a non-trivial block. Let a = (1, 2, 3, 4, 5) and let k = |b − b′ |, where b and b′ are
distinct elements of B. Then Bak ∩ B 6= ∅, so we must have Bak = B since B is
a block. Thus for each i ∈ B, we have iak ∈ Bak , and thus iak ∈ B. Then since
iak ∈ B, we have ia2k ∈ Bak , and thus ia2k ∈ B. By induction it follows that
{iank : n ≥ 0} ⊆ B. We claim {iank : n ≥ 0} = {1, 2, 3, 4, 5}, which contradicts
the fact that B is a non-trivial block. Indeed, for j ∈ {1, 2, 3, 4, 5}, we have
iank = j if and only if i + nk ≡ j (mod 5). Since 5 is prime and 0 < k < 5, we
see that k is coprime with 5. Hence by elementary number theory, there exists n
such that nk ≡ j − i (mod 5) and so i + nk ≡ i + j − i ≡ j (mod 5) as required.
Hence j ∈ {iank : n ≥ 0} for all j ∈ {1, 2, 3, 4, 5}, which proves the claim. It
follows G has no non-trivial blocks, and thus G is primitive.
The above argument can be generalized to prove that a cyclic group G =
h(1, 2, . . . , n)i is primitive if and only if n is prime. If a = (1, 2, . . . , n), then for
each divisor d of n and each integer 1 ≤ i ≤ n, we see that {iamd : m ≥ 0} is a
block. In particular, when n is composite, there exists a divisor d with 1 < d < n,
giving rise to a non-trivial block.
Example 3. Consider the group G = h(1, 2, 3), (4, 5, 6)i ≤ S6 . This group is intransitive, since (for example) it does not contain a permutation mapping 1 to
4. Thus it is imprimitive by definition. Alternatively, observe that {1, 2, 3} and
{4, 5, 6} are non-trivial blocks for G.
Generally an intransitive group will always have non-trivial blocks, but there
is one exception: the trivial subgroup of S2 (containing only the identity element). The natural action of this group is clearly not transitive on {1, 2}, but
its only blocks are the trivial blocks {1}, {2} and {1, 2}. To avoid dealing with
this exception, we require primitive groups to be transitive by definition.
The next example shows that we have the following hierarchy of permutation
group properties:
(2-transitive) ⇒ (primitive) ⇒ (transitive).
6
These implications do not reverse. Cyclic groups of composite order give examples of transitive imprimitive groups, while cyclic groups of prime order p ≥ 5
give examples of primitive, non-2-transitive groups. (For example, the group
h(1, 2, 3, 4, 5)i ≤ S5 is not 2-transitive on {1, 2, 3, 4, 5} since nothing maps the
pair (1, 2) to the pair (1, 3).)
Example 4. The alternating group An is 2-transitive for n ≥ 4. Indeed, given
i, i′ , j, j ′ ∈ {1, . . . , n}, the permutation (i, i′ )(j, j ′ ) is the product of an even
number of 2-cycles, and it maps the pair (i, j) to (i′ , j ′ ). We claim An is also
primitive for n ≥ 2. To see this, first note that An is a cyclic group of prime order
for 2 ≤ n ≤ 3. For n ≥ 4, suppose for a contradiction that B is a non-trivial
block. Then B has at least two elements i and j, but B is not all of {1, . . . , n}.
Choose k ∈ {1, . . . , n} \ B. Since An is 2-transitive, there exists an element
g ∈ An which maps the pair (i, j) to (j, k). Then Bg ∩ B 6= ∅ (since Bg and B
contain j), and thus Bg ∩ B = Bg = B since B is a block. But Bg contains k
and B does not, which is a contradiction. Thus all blocks of An are trivial, and
thus An is primitive. In fact, this argument shows that all 2-transitive groups
are primitive.
The following fact is immediate from the definitions of transitivity and primitivity, and is frequently useful: if H is a subgroup of G and H is transitive
(primitive), then G is also transitive (primitive). For example, the symmetric
group Sn is primitive for n ≥ 2, since it contains the primitive group An .
So far, we have only looked at cyclic groups and the symmetric and alternating groups. For our last pair of examples, we consider two subgroups of S6 that
are a little more interesting.
Example 5. Define a = (2, 4, 6), b = (1, 5)(2, 4) and c = (1, 4, 5, 2)(3, 6), and let
G = ha, b, ci. We claim this group is transitive on {1, . . . , 6}. For g ∈ G and
g
i, j ∈ {1, . . . , 6}, we will write i −→ j to mean ig = j. Observe that
c3
a2
c
c
1 −→ 2 −→ 6 −→ 3,
c
1 −→ 4 −→ 5.
Thus for each i 6= 1, there is some group element that maps 1 to i. If g ∈ G
maps 1 to i, then g −1 maps i to 1. It follows for each i, j, there is some element
x that maps i to 1, and another element y that maps 1 to j, giving
y
x
i −→ 1 −→ j.
Thus G is transitive. It is also imprimitive, with non-trivial blocks {1, 3, 5} and
{2, 4, 6}. Indeed, we see that
a
b
c
{1, 3, 5} −→ {1, 3, 5}, {1, 3, 5} −→ {5, 3, 1}, {1, 3, 5} −→ {4, 6, 2}.
Hence these sets are non-trivial blocks.
7
Example 6. Define a = (1, 2, 3, 4, 6) and b = (1, 2)(3, 4)(5, 6) and let G = ha, bi.
It is easy to see that this group is transitive on {1, . . . , 6}: just verify that 1
can be mapped to every other element and use the argument from the previous
example. This group is also primitive. To see this, first note that the subgroup hai
acts primitively on {1, 2, 3, 4, 6}, since it is a cyclic group of prime order. Hence
a non-trivial block of G cannot be a subset of {1, 2, 3, 4, 6}, so in particular a
non-trivial block of G must contain 5. Suppose B is a non-trivial block that
contains 5; then Ba ∩ B contains 5 and hence Ba ∩ B = Ba = B. Since B
is non-trivial, it contains some element i 6= 5, and since Ba = B we have
{i, ia, ia2, . . . , ia4 } = {1, 2, 3, 4, 6} ⊆ B. This implies B = {1, 2, 3, 4, 5, 6}, and so
B is trivial, which is a contradiction. Thus all blocks of G are trivial, and thus
G is primitive.
A congruence of a monoid action of M on X is an equivalence relation on
X that is M -invariant in the following sense: if E is an equivalence class, then
for all m ∈ M , there exists an equivalence class E ′ such that Em ⊆ E ′ . In
other words, if x and x′ are equivalent, then xm and x′ m are equivalent for
all m ∈ M . The equality congruence {(x, x) : x ∈ X} in which elements are
equivalent only if they are equal, and the full congruence X × X in which all
elements are equivalent, are called trivial congruences. If M is a transformation
monoid on X, a congruence of the natural action is called an M -congruence.
The notion of congruences leads to an important alternate characterization
of primitivity. In the case of a permutation group G on X, notice that for all
S ⊆ X and g ∈ G, the set Sg has the same size as S. Hence a G-congruence
has the following property: if E is an equivalence class, then for all g ∈ G, the
set Eg is also an equivalence class. In particular, we either have E ∩ Eg = E or
E ∩ Eg = ∅ for all g ∈ G; thus the classes of G-congruences are blocks.
In fact, if G is transitive, then every G-congruence arises from the blocks
of G as follows. If B is a block for G, the block system corresponding to B
is the set {Bg : g ∈ G}. As the name implies, each set in a block system is
also a block for G. Indeed, for all g ′ ∈ G, we either have Bgg ′ ∩ Bg = ∅ or
Bgg ′ ∩ Bg 6= ∅, and in the latter case, Bgg ′ g −1 ∩ B 6= ∅. But B is a block, so this
implies Bgg ′ g −1 = B and thus Bgg ′ = Bg. Thus every set in a block system in
a block, so in particular, all distinct sets in a block system are pairwise disjoint.
Furthermore, since G is transitive, each element of X appears in at least one
block of the system. It follows that block systems are partitions of X, and thus
equivalence relations on X. It is easy to see that block systems are G-invariant,
and thus are G-congruences.
Thus every block gives rise to a block system that is a G-congruence, and every G-congruence consists of blocks; it follows block systems and G-congruences
are one and the same if G is a transitive group. If all G-congruences are trivial,
then all block systems of G consist only of trivial blocks, and vice versa. Thus
we obtain our alternate characterization of primitivity: a transitive permutation
group G on X is primitive if and only if all G-congruences are trivial.
Let us revisit some of our earlier examples of primitive and imprimitive
groups in the context of this new characterization.
8
Example 7. Consider the imprimitive cyclic group G = ha = (1, 2, 3, 4, 5, 6)i ≤
S6 of Example 1. Put an equivalence relation ∼ on X = {1, . . . , 6} by letting
i ∼ j if i and j have the same parity (odd or even). Notice that for i ∈ X, the
elements i and ia have opposite parity. Thus ∼ is a G-congruence, since if i ∼ j
then ia ∼ ja, and so if [i] is the equivalence class of i then [i]a = [ia] is also an
equivalence class. In fact, the classes of ∼ are just the blocks {1, 3, 5} and {2, 4, 6}
we found in Example 1; thus the G-congruence ∼ corresponds to the block system
{{1, 3, 5}, {2, 4, 6}}. If we define an equivalence relation by i ∼ j if i and j are
equivalent modulo 3, we obtain a non-trival G-congruence corresponding to the
block system {{1, 4}, {2, 5}, {3, 6}}. As for the trivial G-congruences, the equality
congruence corresponds to the block system {{1}, {2}, . . . , {6}} containing the
singletons, and the full congruence corresponds to the block system {{1, . . . , 6}}
that just contains the full set X.
Example 8. Consider the primitive cyclic group G = h(1, 2, 3, 4, 5)i ≤ S5 of
Example 2. With the notion of G-congruences, it is much easier to prove that
this group is primitive. Indeed, fix a G-congruence on X. By G-invariance, all
classes of the G-congruence must have the same size, say m. If the congruence
has n classes, then we have mn = |X| = 5. So m is either 1 or 5 since 5 is prime,
which means the classes are either singletons (giving the equality congruence) or
the full set X (giving the full congruence). Thus all G-congruences are trivial,
and thus G is primitive. Alternatively, we could make the same argument in
terms of block systems, using the fact that all blocks in a system have the same
size to show all blocks must be trivial. This argument actually shows that not
only are cyclic groups of prime order primitive, but all transitive groups of prime
degree are primitive (since |X| is the degree of a permutation group on X).
2.2
Languages, Automata and State Complexity
Let Σ be a finite set. The set of all finite-length sequences of elements of Σ
is called the free monoid generated by Σ, and is denoted Σ ∗ . In this context,
elements of Σ are called letters, and elements of Σ ∗ are called words over Σ. The
operation of the free monoid is concatenation of words, and the identity element
is the empty word ε of length zero. A set L ⊆ Σ ∗ is called a language over Σ,
and Σ is called the alphabet of L.
We use the convention that a language L ⊆ Σ ∗ is implicitly a pair (L, Σ), so
for example, the language {a, ab} over alphabet {a, b} and the language {a, ab}
over alphabet {a, b, c} are distinct. In particular, two words over different alphabets are necessarily distinct. This is similar to the convention which views two
functions with different codomains as necessarily distinct.
A deterministic finite automaton (DFA) is a tuple A = (Q, Σ, δ, 1, F ) where
Q and Σ are finite sets, δ : Q × Σ ∗ → Q is a monoid action, 1 ∈ Q, and F ⊆ Q.
The elements of Q are called states; the state 1 is called the initial state and the
states in F are called final states. The set Σ is the alphabet of the automaton.
The monoid action δ is called the transition function.
9
Since Σ generates Σ ∗ , we may completely specify the action δ by defining
the function aδ : Q → Q for each a ∈ Σ. If w = a1 · · · ak for a1 , . . . , ak ∈ Σ,
then wδ : Q → Q is the composition (a1 )δ · · · (ak )δ . The function εδ : Q → Q
is necessarily the identity map. The monoid M (A) = {wδ : w ∈ Σ ∗ } is called
the transition monoid of A; it is a submonoid of TQ and thus has a natural
action on Q. We call the function wδ the action of w. Under our notational
conventions, we may write δ(p, w) = q as pwδ = q or simply pw = q. We may
also extend wδ by union and apply it to subsets of the state set: for X ⊆ Q we
w
have Xw = {qw : q ∈ X}. We also sometimes write p −→ q to mean pw = q.
A state q ∈ Q is reachable from p ∈ Q if pw = q for some w. Two states
p, q ∈ Q are distinguishable by X ⊆ Q if there exists w ∈ Σ ∗ such that pw ∈
X ⇔ qw 6∈ X. We frequently use two special cases of these definitions: A state
q ∈ Q is reachable if it is reachable from the initial state 1, and states p, q ∈ Q
are distinguishable if they are distinguishable by F . We say A is accessible if
every state is reachable (from the initial state 1), and strongly connected if every
state is reachable from every other state. A state q ∈ Q is empty if qw 6∈ F for
all w ∈ Σ ∗ . In a strongly connected DFA, there exists an empty state if and only
if all states are empty.
Consider the following relation on Q: two states p, q ∈ Q are related if and
only if they are indistinguishable by X ⊆ Q, that is, for all w ∈ Σ ∗ we have
pw ∈ X ⇔ qw ∈ X. This is an equivalence relation on Q, and in fact it is an
M (A)-congruence. Indeed, if p and q are equivalent, we have pw ∈ X ⇔ qw ∈ X
for all w ∈ Σ ∗ . So in particular, if we take w = xy for some fixed x ∈ Σ ∗ , then
(px)y ∈ X ⇔ (qx)y ∈ X for all y ∈ Σ ∗ , and thus px and qx are equivalent for
all x ∈ Σ ∗ . This congruence is called the indistiguishability congruence of X.
The language recognized by A or simply language of A is the language L(A) =
{w ∈ Σ ∗ : 1w ∈ F } over Σ. A language which can be recognized by a DFA
is called a regular language. Two DFAs are equivalent if they have the same
language. Two DFAs A = (Q, Σ, δ, 1, F ) and A′ = (Q′ , Σ ′ , δ ′ , 1′ , F ′ ) with Σ = Σ ′
are isomorphic if there is a bijection f : Q → Q′ such that 1f = 1′ , F f = F ′ ,
and (qaδ )f = (qf )aδ′ for all a ∈ Σ; in other words, they are identical up to the
naming of the states. In particular, isomorphic DFAs are equivalent.
We say A is minimal if the number of states is minimal among all DFAs
equivalent to A. It is well-known that for each regular language L, all minimal
DFAs recognizing L are isomorphic and hence have the same number of states.
The number of states in a minimal DFA for L is called the state complexity of L,
and is denoted sc(L). A DFA A is minimal if and only if all states are reachable
and all pairs of states are distinguishable.
Given a binary regular operation ◦, the state complexity of the operation ◦
is the following function:
(m, n) 7→ max{sc(Lm ◦ Kn ) : sc(Lm ) = m, sc(Kn ) = n}.
That is, it is the maximal state complexity of the language resulting from the
operation, expressed as a function of the state complexities of the operation’s
arguments.
10
For f : N × N → N and g : N × N → N, we say f ≤ g if f (m, n) ≤ g(m, n) for
all (m, n) ∈ N × N; we say that f is an upper bound for the state complexity of
the operation ◦ if sc(◦) ≤ f , and a tight upper bound if sc(◦) = f .
In the definition of state complexity of operations, we assume that ◦ takes
two languages over the same alphabet as arguments. This is justified by our view
that words over different alphabets are necessarily distinct; hence, for example,
the union of two languages over different alphabets would be a set containing
a mixture of words over different alphabets, which is not a language. To perform such an operation, one must first convert the operands to languages over a
common alphabet. This convention is very common in the literature; however,
Brzozowski has recently argued this convention is unnecessary, and in fact leads
to incorrect state complexity bounds for operations on languages over different alphabets, since converting the input languages to a common alphabet can
change their state complexities [7]. Brzozowski introduces a distinction between
restricted state complexity of operations, the traditional model in which operands
must have the same alphabet, and unrestricted state complexity of operations, a
new model which produces accurate state complexity bounds for operations on
languages over different alphabets.
We use restricted state complexity in this paper for the following reasons.
First, computing unrestricted state complexity requires using DFAs which have
an empty state, and in particular are not strongly connected. In this paper, we
mainly study DFAs whose transition monoids are permutation groups, which
are always strongly connected. This group-theoretic focus is essential to most of
our results. Working with DFAs that are not strongly connected would take us
into the realm of semigroup theory, and we are unsure how much of our work
would carry over. Second, restricted state complexity has been the dominant
model of state complexity of operations for many years, while unrestricted state
complexity is a recent generalization. Although restricted state complexity gives
incorrect results when applied to languages over different alphabets, it is otherwise a correct model. We have chosen to study the simpler case of restricted
state complexity in this paper and leave the more general unrestricted case for
potential future work.
3
Primitive Groups and Uniform Minimality
A DFA A = (Q, Σ, δ, 1, F ) is called a permutation DFA if M (A) is a permutation
group on Q. In this case we call M (A) the transition group rather than the
transition monoid. The languages recognized by permutation DFAs are called
group languages.
Proposition 1. For a permutation DFA A, the following are equivalent:
1. A is accessible.
2. A is strongly connected.
3. M (A) is transitive.
11
Proof. (1) ⇒ (3): Since A is accessible, for each q ∈ Q there exists wq ∈ Σ ∗
such that 1wq = q. Since M (A) is a group, the element wq has an inverse, and
thus for all p, q ∈ Q we have p(wp )−1 wq = 1wq = q. It follows M (A) is transitive.
(3) ⇒ (2): Since M (A) is transitive, for all p, q ∈ Q there exists w ∈ Σ ∗
such that pw = q. This is precisely saying that A is strongly connected.
The last implication (2) ⇒ (1) is immediate.
⊓
⊔
Note that (2) ⇔ (3) holds for arbitrary DFAs, not only permutation DFAs.
Let A = (Q, Σ, δ, 1, F ) be a DFA and let L = L(A) be its language. For
S ⊆ Q, we write A(S) for the DFA A = (Q, Σ, δ, 1, S) obtained by replacing the
final state set of A with S. We say a regular language L′ is a cognate of L if
L′ = L(A(S)) for some S ⊆ Q. We say a DFA A′ is a cognate of A if A′ = A(S)
for some S; so a language is a cognate of L if and only if it is recognized by a
cognate of A. If S = Q or S = ∅, then A(S) is called a trivial cognate of A,
since L(A(S)) is either Σ ∗ or the empty language ∅.
We say A is uniformly minimal if all non-trivial cognates of A are minimal.
That is, we can reassign the final state set of the DFA in any non-trivial way and
the new DFA will always be minimal. Equivalently, all cognates of L = L(A)
have the same state complexity |Q|. This definition is essentially restricted to
accessible DFAs, since if A is not accessible, then not all states are reachable
and hence no cognate of A can be minimal.
Restivo and Vaglica introduced and studied uniformly minimal DFAs in [17].
Their notion of uniform minimality is almost the same as ours, except it is
restricted to strongly connected DFAs. Presumably, Restivo and Vaglica were
interested in DFAs that are minimal for every reassignment of initial and final
states; if a DFA is not strongly connected, we can reassign the initial state to
obtain a new DFA which is not accessible and hence not minimal. However, for
strongly connected DFAs, the choice of initial state has no effect on minimality
since every state is reachable from each possible choice of initial state. Hence we
lose nothing by fixing an initial state and generalizing to accessible DFAs.
Remark. Restivo and Vaglica also studied uniformly minimal DFAs in [16], but
they used different terminology. They used the term “almost uniformly minimal”
for the notion discussed above, and used “uniformly minimal” for a stronger
condition that can only be met by incomplete DFAs (which we do not discuss in
this paper).
A DFA A is called simple if all M (A)-congruences are trivial. Ésik proved
the following result for strongly connected DFAs [17, Proposition 1]. The same
proof works for accessible DFAs.
Proposition 2. An accessible DFA A is uniformly minimal if and only if it is
simple.
Proof. Suppose A is simple, that is, all M (A)-congruences are trivial. Then in
particular, for every S ⊆ Q, the indistiguishability congruence of S is trivial. If
the indistiguishability congruence for S is the equality relation, then each state
12
lies in its own class, so all pairs of states are distinguishable. Since A is accessible, all states are reachable, and hence A is minimal. If the indistinguishability
congruence for S is the full relation, then all states are indistinguishable. But
final states are always distinguishable from non-final states, so this can only
happen if all states are final (S = Q) or all states are non-final (S = ∅). Hence
if ∅ ( S ( Q, then A(S) is minimal, so it follows that A is uniformly minimal.
Conversely, suppose A is not simple, and there exists a non-trivial M (A)congruence. Then this congruence has a class E which has at least two elements,
but is not all of Q. Let E be the final state set of A and let p, q ∈ E. For all
w ∈ Σ ∗ , the states pw and qw both lie in the set Ew, which is contained in some
congruence class E ′ . If E ′ = E, then we have pw, qw ∈ E. If E ′ ∩ E = ∅, then
we have pw, qw 6∈ E. Thus for all w ∈ Σ ∗ , we have pw ∈ E ⇔ qw ∈ E, and so
p and q are not distinguishable by E. Hence A is not uniformly minimal, since
A(E) is not minimal.
⊓
⊔
In the special case of permutation DFAs, we have:
Corollary 1. An accessible permutation DFA A is uniformly minimal if and
only if M (A) is primitive.
Proof. If A is uniformly minimal, then it is simple, so all M (A)-congruences are
trivial. Now, recall that a group G is primitive if and only if all G-congruences
are trivial. Since M (A) is a group, we see that M (A) is primitive.
Conversely, if M (A) is primitive, then all M (A)-congruences are trivial.
Hence A is simple and hence uniformly minimal.
⊓
⊔
Note that both implications in Corollary 1 are vacuously true if A is not accessible: A cannot be uniformly minimal, and M (A) cannot be transitive and thus
cannot be primitive. Thus one can technically omit the accessible assumption.
It seems this relationship between primitivity and minimality has been overlooked until recently. Primitive groups have seen increasing application in automata theory over the past decade, particularly in connection with the classical
synchronization problem for DFAs; for a survey of such work see [2]. The connection between simple DFAs and primitive groups was recently noted by Almeida
and Rodaro [1]. However, primitive groups are not mentioned in Restivo and
Vaglica’s work on uniformly minimal DFAs, nor in any other work on DFA minimality that we are aware of.
The wealth of results on primitive groups makes Corollary 1 quite useful for
studying and constructing uniformly minimal DFAs. For example, we can use
this corollary to easily prove that for each n ≥ 2, there exists a uniformly minimal
DFA with n states. Restivo and Vaglica proved this using a rather complicated
construction [16, Theorem 3].
Proposition 3. For each n ≥ 2, there exists a uniformly minimal permutation
DFA with n states.
13
Proof. The symmetric group Sn is primitive for all n ≥ 2, and clearly for each
n ≥ 2 there exists an n-state DFA with transition group Sn . For example, let
{g1 , . . . , gk } be a generating set of the symmetric group and let A be a DFA
with states {1, . . . , n}, alphabet Σ = {a1 , . . . , ak }, and transition function δ
with (ai )δ = gi for 1 ≤ i ≤ k. In fact we can use a binary alphabet, since Sn has
generating sets of size two for all n ≥ 2.
⊓
⊔
This proof illustrates a technique that is very useful for producing examples of
DFAs. If we have a generating set for a transformation monoid, we can construct
a DFA which has that monoid as its transition monoid.
Example 9. Let A be the DFA with alphabet {a, b} defined as follows.
– The states are {1, 2, 3, 4}, the initial state is 1, and the final states are {3, 4}.
– The transformations are the permutations a = (2, 3, 4) and b = (1, 2)(3, 4).
More formally, we mean that the transition function δ of A is given by aδ =
(2, 3, 4) and bδ = (1, 2)(3, 4). However, we will generally be brief when describing
DFAs, as above.
The permutations (1, 2)(3, 4) and (2, 3, 4) generate the alternating group A4 .
Thus the transition group of A is A4 . We saw in Example 4 that A4 is transitive
and primitive. Hence by Proposition 1, A is strongly connected, and by Corollary
1, A is uniformly minimal.
A state diagram of A is given in Figure 1. We can see from the diagram
that A is indeed strongly connected. It is tedious, but possible to verify that
A is uniformly minimal by checking that it is minimal with respect to every
non-empty, proper subset of {1, 2, 3, 4}.
a
a
a, b
b
1
a
2
3
4
b
b
Fig. 1. Uniformly minimal DFA A of Example 9.
a
1
a
2
a
3
a
4
a
5
a
6
Fig. 2. Non-minimal DFA A of Example 10 with an imprimitive transition group.
14
Example 10. Let A be the DFA with alphabet {a}, states {1, . . . , 6}, initial state
1, final states F = {1, 3, 5} and a = (1, 2, 3, 4, 5, 6). A diagram is in Figure 2.
The transition group G of A is the cyclic group of order six, which is imprimitive. We saw in Example 1 that F = {1, 3, 5} is a block for this group.
Hence for all k, we either have F ak = F or F ak ∩ F = ∅. Thus if i, j ∈ F , then
for all k we have iak ∈ F ⇔ jak ∈ F . This means all pairs of states in F are
indistinguishable by F , and hence A is not minimal.
This argument actually shows that whenever F is a non-trivial block of G,
the DFA A is not minimal. In fact, this also holds whenever F is a union of
non-trivial blocks of G (see Lemma 1 below).
Note that if we construct a DFA from a cyclic group of prime order, we get
a uniformly minimal DFA, since cyclic groups of prime order are primitive.
There exist many infinite families of primitive groups, and hence of uniformly
minimal permutation DFAs. However, there are infinitely many positive integers
n for which the only primitive groups of degree n are Sn and An [12, pg. 66].
Hence other infinite families of primitive groups cannot be used to construct
n-state uniformly minimal DFAs for every n, unless we “fill in the gaps” with
symmetric or alternating groups.
Remark. Steinberg has extended the notion of primitivity to transformation
monoids [20]. Steinberg defines a transformation monoid M to be primitive if
there are no non-trivial M -congruences. Under this definition, an accessible DFA
A is uniformly minimal if and only if the transition monoid M (A) is primitive.
However, we have not investigated whether any of our other results that hold
for primitive groups are also true for primitive monoids.
We close this section with a technical lemma that generalizes Proposition 2.
If M is a transformation monoid on X and S ⊆ X, we say that S is saturated
by an M -congruence if it is a union of classes of the M -congruence.
Lemma 1. An accessible DFA A with ∅ ( F ( Q is minimal if and only if
there is no non-trivial M (A)-congruence that saturates F .
It follows that if all M (A)-congruences are trivial, then A is uniformly minimal.
Conversely, if there is a non-trivial M (A)-congruence, then it saturates its own
congruence classes and at least one class is a proper non-empty subset of Q, and
thus A is not uniformly minimal. Hence this indeed generalizes Proposition 2.
Proof. We prove the contrapositive: A is not minimal if and only if there exists
a non-trivial M (A)-congruence that saturates F .
Suppose A is not minimal. Then the indistinguishability congruence of F is
a non-trivial M (A)-congruence, since at least two states are indistinguishable.
Suppose there is an indistinguishability class E that is neither contained in F
nor disjoint from F . Then there exist p, q ∈ E such that p ∈ F and q 6∈ F .
But then p and q are distinguishable by F , which cannot happen since E is an
indistinguishability class. Thus for each indistinguishability class [q], we have
15
S
[q] ⊆ F or [q] ∩ F = ∅. Then we have F = f ∈F [f ], so F is saturated by its
indistinguishability congruence.
Conversely, let E1 , . . . , Ek ⊆ Q be the congruence classes of a non-trivial
M (A)-congruence that saturates F . Choose a congruence class Ei of size at
least two. Then for all w ∈ Σ ∗ we have Ei w ⊆ Ej for some j. Since F is a union
of congruence classes, either Ej ⊆ F or Ej ∩ F = ∅. Hence for p, q ∈ Ei and all
w ∈ Σ ∗ , we have pw ∈ F ⇔ Ej ⊆ F ⇔ Ei w ⊆ F ⇔ qw ∈ F . It follows that
states in Ei are indistinguishable, and thus A is not minimal.
⊓
⊔
In the special case of permutation DFAs, this has a useful consequence.
Corollary 2. Let A be a permutation DFA. If |F | = 1 or |F | = |Q| − 1, then A
is minimal if and only if it is accessible.
Proof. Recall that if G is a transitive permutation group and E and E ′ are
classes of a G-congruence, then |E| = |E ′ |. It follows that if |F | = 1, then a
non-trivial M (A)-congruence cannot saturate F since all the congruence classes
have size at least two. Furthermore, an M (A)-congruence saturates F if and only
if it saturates Q \ F , and if |F | = |Q| − 1 then a non-trivial M (A)-congruence
cannot saturate the set Q \ F of size one. Hence if A is accessible, it is minimal
by Lemma 1. On the other hand, if A is not accessible, it cannot be minimal. ⊓
⊔
4
Main Results
Throughout this section, A = (Q, Σ, δ, 1, F ) and A′ = (Q′ , Σ ′ , δ ′ , 1′ , F ′ ) are
minimal DFAs with a common alphabet Σ = Σ ′ . The languages of A and A′ are
L and L′ , and the transition monoids are M and M ′ , respectively. For w ∈ Σ ∗ ,
we write w for wδ ∈ M and w′ for wδ′ ∈ M ′ . Sometimes we will assume A and
A′ are permutation DFAs, and then we will use G and G′ for the transition
groups rather than M and M ′ .
4.1
Direct Products and Boolean Operations
The direct product of A and A′ is the DFA A×A′ with state set Q×Q′ , alphabet
a
Σ, transitions (q, q ′ ) −→ (qa, q ′ a′ ) for each (q, q ′ ) ∈ Q × Q′ and a ∈ Σ, initial
state (i, i′ ), and an unspecified set of final states. By assigning particular sets
of final states to A × A′ as described below, we can recognize the languages
resulting from arbitrary binary boolean operations on L and L′ .
Fix a function ◦ : {0, 1}2 × {0, 1}; these are called binary boolean functions.
For a set S, let χS : S → {0, 1} denote the characteristic function of S, defined
by χS (x) = 1 if x ∈ S and χS (x) = 0 otherwise. We can think of χS (x) as giving
the “truth value” of the proposition “ x ∈ S ”, where 0 is false and 1 is true. Now
for F ⊆ Q and F ′ ⊆ Q′ , define F ◦ F ′ = {(q, q ′ ) ∈ Q × Q′ : χF (q) ◦ χF ′ (q ′ ) = 1}.
Then A × A′ with final states F ◦ F ′ recognizes the language L ◦ L′ defined by
L ◦ L′ = {w ∈ Σ ∗ : χL (w) ◦ χL′ (w) = 1}.
16
For example, if ◦ : {0, 1}2 → {0, 1} is the “logical or” function, then L ◦ L′ =
L ∪ L′ , since w ∈ L ◦ L′ if w ∈ L or w ∈ L′ . Similarly, the “logical and” function
gives the intersection L ∩ L′ .
We say that a boolean function (and the associated boolean operation on
languages) is proper if its output depends on both of its arguments. For example,
u ◦ v = 1 − u (giving L ◦ L′ = Σ ∗ \ L) only depends on the first argument,
and u ◦ v = 0 (giving with L ◦ L′ = ∅) depends on neither argument, so they
are not proper. If L and L′ have state complexity m and n respectively, then
the improper binary boolean operations have state complexity 1 (if they are
constant), state complexity m (if they depend only on the first operand), or
state complexity n (if they depend only on the second operand). There are 16
binary boolean operations in total, and one may easily verify that 10 of them
are proper.
If A and A′ have m and n states respectively, A × A′ has mn states. Hence
every proper binary boolean operation has state complexity bounded by mn.
It is well-known that this bound is tight for general regular languages, and it
remains tight for regular group languages. In fact, the original witnesses for
union given by Maslov [15] and Yu et al. [21] are group languages, and we will
demonstrate later (in Example 13) that these languages are also witnesses for
all other proper binary boolean operations. We will say the pair (L, L′ ) has
maximal boolean complexity if sc(L ◦ L′ ) = sc(L) sc(L′ ) for all proper binary
boolean operations ◦.
Suppose that ∅ ( F ( Q and ∅ ( F ′ ( Q′ . We say a subset of Q × Q′
is (F, F ′ )-compatible if it is equal to F ◦ F ′ for some proper binary boolean
operation ◦. Notice that (L, L′ ) has maximal boolean complexity if and only if
A × A′ is minimal for every (F, F ′ )-compatible subset of Q × Q′ . We disallow
F = ∅ and F = Q since then A × A′ is minimal for F ◦ F ′ only if |Q| = 1
and L = ∅ or L = Σ ∗ ; these cases are uninteresting. Similarly, we exclude
F ′ = ∅ and F ′ = Q′ . We say the pair (A, A′ ) (or the direct product A × A′ ) is
uniformly boolean minimal if for every pair of sets (S, S ′ ) with ∅ ( S ( Q and
∅ ( S ′ ( Q′ and every (S, S ′ )-compatible set S ◦ S ′ , the DFA (A × A′ )(S ◦ S ′ )
is minimal. In other words, if every pair of cognates of L and L′ has maximal
boolean complexity.
We give an example of a pair of DFAs that are not uniformly boolean minimal,
as well as a pair of DFAs that are.
Example 11. Define two DFAs over alphabet Σ = {a, b, c} as follows:
– A has state set Q = {1, 2}, initial state 1, final state set F = {1}, and
transformations a = b = (1, 2), c = ().
– A′ has state set Q′ = {1, 2}, initial state 1, final state set F ′ = {1}, and
transformations a′ = c′ = (1, 2), b′ = ().
We show that (L(A), L(A′ )) does not have maximal boolean complexity, and
thus A × A′ is not uniformly boolean minimal.
17
To see this, consider the symmetric difference operator ⊕, arising from the
“exclusive or” boolean function: for u, v ∈ {0, 1}, the “exclusive or” u ⊕ v is zero
if u = v and one if u 6= v. The corresponding operation on languages over Σ is
L ⊕ L′ = {w ∈ Σ ∗ : w ∈ L or w ∈ L′ , but not both} = (L \ L′ ) ∪ (L′ \ L).
The final state set that makes A × A′ recognize L(A) ⊕ L(A′ ) is:
F ⊕ F ′ = {(q, q ′ ) ∈ Q × Q : q ∈ F or q ′ ∈ F ′ , but not both} = {(1, 2), (2, 1)}.
State diagrams of the DFAs A, A′ and (A × A′ )(F ⊕ F ′ ) are shown in Figure
3. Notice that (A × A′ )(F ⊕ F ′ ) is not minimal: the states (1, 2) and (2, 1)
cannot be distinguished. Since F ⊕ F ′ is an (F, F ′ )-compatible set, it follows
that (L(A), L(A′ )) does not have maximal boolean complexity, and that A × A′
is not uniformly boolean minimal.
c
c
1, 1
c
1, 2
a, b
1
c
2
a, b
b
b
b
a
a
a
b
a, c
1
a
b
b
2
a, c
c
2, 1
2, 2
c
Fig. 3. DFAs A, A′ and A × A′ of Example 11. The final state set of A × A′ is chosen
so that A × A′ recognizes the symmetric difference of the languages of A and A′ .
Example 12. Define two DFAs over alphabet Σ = {a, b} as follows:
– A has state set Q = {1, 2} and transformations a = (1, 2), b = ().
– A′ has state set Q′ = {1, 2, 3} and transformations a′ = (1, 2), b′ = (1, 2, 3).
The initial and final states are not important for this example.
The direct product A × A′ is shown in Figure 4. Notice that the transition
group of A′ is S3 . We will see much later (Corollary 3) that this implies A × A′
is uniformly boolean minimal.
Note that A × A′ is not uniformly minimal; for example, it is not minimal
with respect to the final state set {(1, 1), (1, 2), (1, 3)}. If |Q|, |Q′ | ≥ 2, a direct
product DFA with state set Q×Q′ can never be uniformly minimal. In particular,
it cannot be minimal for final state sets of the form S × Q′ (“unions of rows”)
or Q × S ′ (“unions of columns”). However, the definition of uniform boolean
minimality excludes these sets.
18
b
b
1, 1
1, 2
a
a
a
a
b
1, 3
a
2, 1
b
2, 2
b
a
2, 3
b
Fig. 4. Uniformly boolean minimal DFA A × A′ of Example 12.
Bell, Brzozowski, Moreira and Reis found sufficient conditions for a pair of
DFAs to be uniformly boolean minimal [3]. However, these conditions require
that the transition monoids of the DFAs contain the symmetric group, in the
sense that they contain every permutation of the DFA’s state set. In particular,
for permutation DFAs, these conditions only apply when the transition group
is the symmetric group on the state set. We obtain more general sufficient conditions for uniform boolean minimality in permutation DFAs, which apply to a
larger class of transition groups. Additionally, we show that DFAs whose transition monoids contain 2-transitive groups “usually” meet these conditions, up to
some technical assumptions we will state later.
We also obtain necessary and sufficient conditions for a pair of languages
(L, L′ ) to have maximal boolean complexity in the special case where L and L′
are recognized by permutation DFAs A and A′ with exactly one final state. In
this special case, it turns out (L, L′ ) has maximal boolean complexity if and
only if A × A′ is accessible. We give several group-theoretic conditions and a
graph-theoretic condition that are equivalent to A × A′ being accessible.
We begin with a proposition which characterizes (F, F ′ )-compatible subsets.
If Q is the state set of a DFA and S ⊆ Q, write S for Q \ S. Similarly, if L is a
language over Σ, write L for Σ ∗ \ L.
Proposition 4. Let ∅ ( F ( Q and ∅ ( F ′ ( Q′ . A subset of Q × Q′ is
(F, F ′ )-compatible if and only it is equal to one of the following sets:
(a)
(b)
(c)
(d)
(e)
(f )
F × F ′ (corresponding to L ∩ L′ ).
F × F ′ (corresponding to L ∩ L′ = L \ L′ ).
F × F ′ (corresponding to L ∩ L′ = L′ \ L).
F × F ′ (corresponding to L ∩ L′ = L ∪ L).
(F ×F ′ )∪(F ×F ′ ) (corresponding to symmetric difference (L\L′ )∪(L′ \L)).
The complement (Q × Q′ ) \ S, where S is one of the above sets.
Proof. Let ◦ be a proper binary boolean function. Let k be the number of pairs
(u, v) ∈ {0, 1} × {0, 1} such that u ◦ v = 1.
19
Case 1 (k = 1): If k = 1, then there is a unique pair (u, v) such that
u ◦ v = 1. Hence F ◦ F ′ = {(q, q ′ ) : χF (q) = u and χF ′ (q ′ ) = v}. Consider
possible values for (u, v):
–
–
–
–
If
If
If
If
(u, v) = (0, 0)
(u, v) = (0, 1)
(u, v) = (1, 0)
(u, v) = (1, 1)
then
then
then
then
F
F
F
F
◦ F′
◦ F′
◦ F′
◦ F′
=F
=F
=F
=F
× F ′.
× F ′.
× F ′.
× F ′.
Hence F ◦ F ′ is a set of type (a), (b), (c) or (d).
Case 2 (k = 2): If k = 2, there are exactly two pairs (u, v) and (u′ , v ′ )
such that u ◦ v = u′ ◦ v ′ = 1. We claim that u 6= u′ and v 6= v ′ . To see this,
suppose u = u′ . Then we must have v 6= v ′ , or else the pairs are not distinct.
Thus {v, v ′ } = {0, 1} and it follows u ◦ 0 = u ◦ 1 = 1. Hence ◦ only depends on
the value of the first argument, which contradicts the fact that ◦ is proper. By
a symmetric argument, we cannot have v = v ′ . Now, observe that (q, q ′ ) is in
F ◦ F ′ if and only if
χF (q) = u and χF ′ (q ′ ) = v or χF (q) = u′ and χF ′ (q ′ ) = v ′ .
Suppose (u, v) = (1, 0). Then we necessarily have (u′ , v ′ ) = (0, 1) and we get
F ◦ F ′ = (F × F ′ ) ∪ (F × F ′ ).
If (u, v) = (0, 1), then (u′ , v ′ ) = (1, 0) and we get the same set. If (u, v) = (1, 1)
or (u, v) = (0, 0), then we get
F ◦ F ′ = (F × F ′ ) ∪ (F × F ′ ).
But this is simply the complement of the previous set. So we either get a set of
type (e) or the complement of such a set, which is type (f).
Case 3 (k = 3): If k = 3, then there is a unique pair (u, v) such that
u ◦ v = 0. Hence F ◦ F ′ is the complement of a set of type (a), (b), (c) or (d),
that is, a set of type (f).
This proves that every (F, F ′ )-compatible set, that is, every set of the form
F ◦ F ′ where ◦ is a proper binary boolean function, has one of the given forms
(a)–(f). Conversely, if we are given sets F and F ′ and a set X ⊆ Q × Q′ with one
of the forms (a)–(f), the proof shows how to construct a proper binary boolean
function ◦ such that X = F ◦ F ′ . It follows X is (F, F ′ )-compatible if and only
if it has one of the forms (a)–(f).
⊓
⊔
4.2
Accessibility of A × A′
In this section, we consider the problem of determining when A×A′ is accessible.
This is essential for proving that A × A′ is minimal for a certain final state set,
and also an interesting question in its own right. Recall that by Proposition 1, the
DFA A × A′ is strongly connected if and only if the transition monoid of A × A′
is transitive. Furthermore, if A× A′ is a permutation DFA, then it is accessible if
and only if its transition group is transitive. The following proposition describes
the structure of the transition monoid of A × A′ .
20
Proposition 5. Let M× denote the transition monoid of A × A′ .
1. M× is isomorphic to the submonoid of M ×M ′ generated by {(a, a′ ) : a ∈ Σ}.
We often identify M× with this submonoid.
2. The projections π : M× → M and π ′ : M× → M ′ given by (w, w′ )π = w and
(w, w′ )π ′ = w′ are surjective.
3. If M and M ′ are groups, then M× is a group.
Proof. (1): Write w× for wA×A′ ∈ M× . Consider the map ϕ : M× → M × M ′
given by w× 7→ (w, w′ ). This map is clearly a monoid homomorphism. Furthermore, if (x, x′ ) = (y, y ′ ) then qx = qy and q ′ x′ = q ′ y ′ for all q ∈ Q and q ′ ∈ Q′ ,
and thus in M× we have (q, q ′ )x× = (q, q ′ )y× for all (q, q ′ ) ∈ Q × Q′ . Hence
x× = y× whenever x× ϕ = y× ϕ, and it follows that ϕ is injective.
Since ϕ is injective, (M× )ϕ is a finite monoid of the same size as M× . It
follows ϕ is bijective when viewed as homomorphism between M× and (M× )ϕ,
and thus ϕ is an isomorphism between these monoids. Since {a× : a ∈ Σ}
generates M× , we see that {(a, a′ ) : a ∈ Σ} generates (M× )ϕ. Hence we have
M× ∼
= (M× )ϕ = h(a, a′ ) : a ∈ Σi as required.
(2): Fix w ∈ M . Then for the element (w, w′ ) ∈ M× we have (w, w′ )π = w.
Hence π maps surjectively onto M . Similarly, π ′ maps surjectively onto M ′ .
(3): Since M× is a monoid, it suffices to show every element of M× has an inverse. Recall that the identity elements of M and M ′ are ε and ε′ respectively. For
(w, w′ ) ∈ M× , pick m and n such that wm = ε in M and (w′ )n = ε′ in M ′ . This
is possible since M and M ′ are finite groups. We have (wmn−1 , (w′ )mn−1 ) ∈ M× ,
and (w, w′ )(wmn−1 , (w′ )mn−1 ) = (wmn−1 , (w′ )mn−1 )(w, w′ ) = (wmn , (w′ )mn ) =
(ε, ε′ ), the identity of M× . Thus (wmn−1 , (w′ )mn−1 ) is an inverse of (w, w′ ). ⊓
⊔
Recall that for permutation DFAs A and A′ , we denote the transition group of
A by G and the transition group of A′ by G′ ; in this case we will also write
G× for the transition group of A × A′ . If A and A′ are permutation DFAs, the
transitivity of G× is a necessary and sufficient condition for all states of A × A′
to be reachable. However, the structure of G× can be difficult to understand.
Hence we will derive a simpler characterization of transitivity that depends only
on properties of G and G′ .
Suppose that A and A′ are permutation DFAs. Consider the subgroup ker π ≤
G× . It contains all (w, w′ ) ∈ G× such that w is the identity in G. View Q × Q′
as a grid, where elements of Q are “row indices” and elements of Q′ are “column indices”. Then ker π consists of the elements of G× which fix all row indices.
Hence we define R = ker π and call R the full row stabilizer. Similarly, C = ker π ′
fixes all column indices and we call it the full column stabilizer. Both of these
subgroups are normal, since they are kernels of homomorphisms.
Fix q ∈ Q and q ′ ∈ Q′ . Let (q, ∗) denote the set {(q, i′ ) ∈ Q × Q′ : i′ ∈ Q′ },
that is, the “q-th row” of Q × Q′ . Similarly, let (∗, q ′ ) = {(i, q ′ ) ∈ Q × Q′ : i ∈ Q}
denote the “q ′ -th column” of Q × Q′ . Let Rq ≤ G× be the setwise stabilizer of
(q, ∗) and let Cq′ ≤ G× be the setwise stabilizer of (∗, q ′ ). We call the subgroups
Rq the single row stabilizers and the subgroups Cq′ the single column stabilizers.
21
The full row stabilizer R is the intersection of all single row stabilizers, and hence
is a subgroup of each Rq ; the analogous fact holds for C.
We now give necessary and sufficient conditions for A × A′ to be transitive
in the case where A and A′ are permutation DFAs.
Lemma 2. Let A and A′ be permutation DFAs. The following are equivalent:
1. A × A′ is accessible.
2. G and G′ are transitive and for all q ∈ Q and q ′ ∈ Q′ , the subgroups Rq π ′ ≤
G′ and Cq′ π ≤ G are transitive.
3. G is transitive and Rq π ′ ≤ G′ is transitive for some q ∈ Q, or G′ is transitive
and Cq′ π ≤ G is transitive for some q ′ ∈ Q′ .
4. G× is transitive.
Proof. Since A × A′ is a permutation DFA, we see that (1) ⇔ (4). Also, the
implication (2) ⇒ (3) is immediate.
(3) ⇒ (4): Fix (i, i′ ), (j, j ′ ) ∈ Q × Q′ . Suppose that Rq π ′ is transitive for
some q ∈ Q; the case where some Cq′ π is transitive is symmetric.
– Since G is transitive, there exists x ∈ G such that ix = q. Let k ′ ∈ Q′ be the
x
element such that i′ x′ = k ′ . Then (i, i′ ) −→ (q, k ′ ).
– Since G is transitive, there exists y ∈ G such that qy = j in G.
– Since Rq π ′ ≤ G′ is transitive on Q′ , there exists z ′ ∈ Rq π ′ such that k ′ z ′ =
z
j ′ (y ′ )−1 . Since (z, z ′ ) ∈ Rq , we have qz = q. Hence (q, k ′ ) −→ (q, j ′ (y ′ )−1 ).
It follows that
x
y
z
(i, i′ ) −→ (q, k ′ ) −→ (q, j ′ (y ′ )−1 ) −→ (qy, j ′ ) = (j, j ′ ).
Thus G× is transitive on Q × Q′ , since for all elements (i, i′ ), (j, j ′ ) ∈ Q × Q′ ,
there exists an element of G× that maps one to the other.
(4) ⇒ (2): If G× is transitive, then for all q, i, j ∈ Q and q ′ , i′ , j ′ ∈ Q′ , there
exist x, y ∈ Σ ∗ such that
y
x
(q, i′ ) −→ (qx, i′ x′ ) = (q, j ′ ) and (i, q ′ ) −→ (iy, q ′ y ′ ) = (j, q ′ ).
Thus, we see that:
– Since qx = q, we have (x, x′ ) ∈ Rq , and since q ′ y ′ = q ′ , we have (y, y ′ ) ∈ Cq′ .
– For all i, j ∈ Q, there exists a word y ∈ Cq′ π ≤ G that maps i to j.
– For all i′ , j ′ ∈ Q′ , there exists a word x′ ∈ Rq π ′ ≤ G′ that maps i′ to j ′ .
Hence for all q ∈ Q and q ′ ∈ Q′ , we see that Rq π ′ is transitive on Q′ and Cq′ π
is transitive on Q. Since Cq′ π ≤ G and Rq π ′ ≤ G′ , it follows that G and G′ are
transitive.
This establishes a cycle of implications (2) ⇒ (3) ⇒ (4) ⇒ (2). Since we
also have (1) ⇔ (4), all the statements are equivalent.
⊓
⊔
22
This lemma reduces the problem of checking transitivity of G× to just checking the the transitivity of a row stabilizer on the column indices, or of a column
stabilizer on the row indices. The following proposition gives a graph-theoretic
interpretation of this idea, which may be easier to understand and apply. This
graph-theoretic condition for accessibility of A × A′ can easily be proved without appeal to group theory, but for illustrative purposes we will connect it with
condition (3) of Lemma 2.
Proposition 6. Let A and A′ be permutation DFAs. The following statements
are equivalent:
1. A is accessible and there exists q ∈ Q such that all states in (q, ∗) are reachable, or A′ is accessible and there exists q ′ ∈ Q′ such that all states in (∗, q ′ )
are reachable.
2. G is transitive and Rq π ′ ≤ G′ is transitive for some q ∈ Q, or G′ is transitive
and Cq′ π ≤ G is transitive for some q ′ ∈ Q′ .
Proof. (1) ⇒ (2): Suppose A is accessible, and there exists q ∈ Q such that
all states in (q, ∗) are reachable. Since A is an accessible permutation DFA, G is
transitive on Q.
To see that Rq π ′ ≤ G′ is transitive on Q′ , fix i′ , j ′ ∈ Q′ . Since all states in
(q, ∗) are reachable from the initial state (1, 1′ ) of A × A′ , there is some word
x
x ∈ G× such that (1, 1′ ) −→ (q, i′ ). Also, there is some y ∈ G× such that
y
(1, 1′ ) −→ (q, j ′ ). Since A × A′ is a permutation DFA, there is some z ∈ Σ ∗ such
that z = x−1 . Thus we have
y
z
(q, i′ ) −→ (1, 1′ ) −→ (q, j ′ ).
We see that zy maps q to itself, so z ′ y ′ ∈ Rq π ′ . Hence Rq π ′ is transitive on Q′ .
By a symmetric argument, if A′ is accessible and there exists q ′ ∈ Q′ such
that all states in (∗, q ′ ) are reachable, it follows that G′ is transitive on Q′ and
Cq′ π ≤ G is transitive on Q.
(2) ⇒ (1): Suppose G is transitive and Rq π ′ ≤ G′ is transitive for some
q ∈ Q. Since G is transitive, A is accessible. In particular, there exists w ∈ Σ ∗
w
such that 1w = q, and it follows that (1, 1′ ) −→ (q, 1′ w′ ) in A × A′ . Since Rq π ′ is
x
transitive on Q′ , for all q ′ ∈ Q there exists x ∈ Rq π ′ such that (q, 1′ w′ ) −→ (q, q ′ ).
′
Hence every state in (q, ∗) is reachable. In the case where G and some Cq′ π are
transitive, we can use a symmetric argument.
⊓
⊔
We now prove one of our main results, which gives necessary and sufficient
conditions for pairs of group languages recognized by DFAs with exactly one
final state to have maximal boolean complexity.
23
Theorem 1. Suppose |Q| ≥ 3 or |Q′ | ≥ 3. Let A and A′ be permutation DFAs
with exactly one final state. Then the following are equivalent:
1. A × A′ is accessible.
2. For all proper binary boolean operations ◦, the language L ◦ L′ has maximal
state complexity. That is, (L, L′ ) has maximal boolean complexity.
3. There exists a proper binary boolean operation ◦ such that the language L◦L′
has maximal complexity.
Determining whether A × A′ is accessible can be difficult in general. Perhaps the
easiest method is to use the graph-theoretic condition of Proposition 6, which
states that assuming A and A′ are accessible, the direct product A × A′ is
accessible if either of the following holds.
– There exists a row (q, ∗) such that all states in (q, ∗) are reachable.
– There exists a column (∗, q ′ ) such that all states in (∗, q ′ ) are reachable.
This reduces the problem to just checking reachability for a single row or column.
The group-theoretic conditions of Lemma 2 may also be used, but they are
perhaps harder to understand. Much later in the paper (Section 4.4) we will use
these to obtain simpler group-theoretic conditions for accessibility of A × A′ .
In particular, provided that A and A′ are both accessible, and also satisfy an
additional criterion called dissimilarity (which is usually easy to check), we have:
– If G or G′ is a transitive simple group, then A × A′ is accessible.
– If G or G′ is a primitive group, then A × A′ is accessible.
Proof (Theorem 1). The only difficult implication here is (1) ⇒ (2). Suppose
A × A′ is accessible; we want to show that L ◦ L′ has maximal state complexity
for every proper binary boolean operation ◦. That is, we want to show that all
pairs of states of A × A′ are distinguishable by each (F, F ′ )-compatible subset
of Q × Q′ .
Note that since A × A′ is accessible, by Lemma 2 we know that G× is transitive and that Rq π ′ ≤ G′ and Cq′ π ≤ G are transitive for all q ∈ Q and q ′ ∈ Q′ .
What this means is:
– For every pair of states (p, p′ ) and (q, q ′ ) of A × A′ , there exists a word
w
w ∈ Σ ∗ such that (p, p′ ) −→ (q, q ′ ). (Transitivity of G× )
– Fix a state q ∈ Q. For every pair of states i′ , j ′ ∈ Q′ , there exists a word
w
w ∈ Σ ∗ such that (q, i′ ) −→ (q, j ′ ). (Transitivity of Rq π ′ )
– Fix a state q ′ ∈ Q′ . For every pair of states i, j ∈ Q, there exists a word
w
w ∈ Σ ∗ such that (i, q ′ ) −→ (j, q ′ ). (Transitivity of Cq′ π)
We will use these facts repeatedly throughout the proof.
Let F = {f } and F ′ = {f ′ }, so that F × F ′ = {(f, f ′ )}. Let (p, p′ ) and (q, q ′ )
be distinct states of A×A′ that we wish to distinguish. We will show these states
are distinguishable with respect to each type of set described in Proposition 4.
24
We only need to consider types (a) through (e), since sets of type (f) are just
complements of sets of types (a) through (e), and two states are distinguishable
by a set X if and only if they are distinguishable by the complement of X.
Case 1 (States in the same row or same column): Suppose p = q, that
is, both states (p, p′ ) and (q, q ′ ) are in the same row. Then we necessarily have
p′ 6= q ′ , since the states are distinct.
w
– By transitivity of G× , for all r ∈ Q there exists w ∈ Σ ∗ such that (p, p′ ) −→
(r, f ′ ).
w
– Since p = q and p′ 6= q ′ , we have (q, q ′ ) −→ (r, s) for some s 6= f ′ . (Since w′
′
′
is a permutation, it must map p and q to different states.)
If r ∈ F , we have (r, f ′ ) ∈ F × F ′ and (r, s) ∈ F × F ′ . Hence we can distinguish
the states if the final state set is F × F ′ , F × F ′ , or (F × F ′ ) ∪ (F × F ′ ).
If r 6∈ F , we have (r, f ′ ) ∈ F × F ′ and (r, s) ∈ F × F ′ . Hence we can distinguish
the states if the final state set is F × F ′ or F × F ′ .
This covers all the possible sets of final states. If p 6= q and p′ = q ′ (that is,
the states are in the same column) we can use a symmetric argument.
Case 2 (States in different rows and different columns): Assume p 6= q
and p′ 6= q ′ . We consider each possible set of final states in turn.
F × F ′ : Here A × A′ has exactly one final state (f, f ′ ), so it is minimal by
Corollary 2.
F × F ′ : We make a few observations:
w
– By transitivity of G× , there exists w ∈ Σ ∗ such that (p, p′ ) −→ (f, f ′ ).
– Since p 6= q, p′ 6= q ′ and w is a permutation, we must have qw 6= f and
q ′ w′ 6= f ′ .
x
– Since Cf ′ π is transitive, there exists x ∈ Σ ∗ such (qw, f ′ ) −→ (f, f ′ ).
It follows that
w
x
(p, p′ ) −→ (f, f ′ ) −→ (f x, f ′ ),
w
x
(q, q ′ ) −→ (qw, q ′ w′ ) −→ (f, q ′ w′ x′ ).
– Since qwx = f , qw 6= f and x is a permutation, we have f x 6= f . It follows
that (f x, f ′ ) ∈ F × F ′ .
– Since f ′ x′ = f ′ , q ′ w′ 6= f ′ and x′ is a permutation, we have q ′ w′ x′ 6= f ′ . It
follows that (f, q ′ w′ x′ ) ∈ F × F ′ .
Hence wx maps (p, p′ ) to a non-final state and (q, q ′ ) to a final state. Thus we
have distinguished the two states.
F × F ′ : We can use a symmetric argument to the previous case.
w
F × F ′ : As in the case of F × F ′ , pick w such that (p, p′ ) −→ (f, f ′ ). Then
w
(q, q ′ ) −→ (qw, q ′ w′ ), which is in F × F ′ since qw 6= f and q ′ w′ 6= f ′ . Thus
w
w sends (q, q ′ ) to a final state. But (p, p′ ) −→ (f, f ′ ) is non-final, so we have
distinguished the states.
25
(F × F ′ ) ∪ (F × F ′ ): This is the most complicated case.
u
– By transitivity of G× , there exists u ∈ Σ ∗ such that (p, p′ ) −→ (f, r′ ), where
r′ 6= f ′ .
u
– We have (f, r′ ) ∈ F × F ′ , so u sends (p, p′ ) to a final state. If (q, q ′ ) −→
′ ′
(qu, q u ) is non-final, then u distinguishes the states, so we may assume
without loss of generality that it is final.
– We cannot have qu = f , since p 6= q and pu = f . Thus qu ∈ F . Since
(qu, q ′ u′ ) is final we therefore must have q ′ u′ ∈ F ′ , that is, q ′ u′ = f ′ .
Define r = qu; now we have reduced the problem to distinguishing two states of
the forms (f, r′ ) and (r, f ′ ), with r 6= f and r′ 6= f ′ .
Suppose |Q| ≥ 3; if we only have |Q′ | ≥ 3 we can use a symmetric argument
to the argument below.
– Since Cf ′ π is transitive and |Q| ≥ 3, there is a word v ∈ Σ ∗ such that
v
(f, f ′ ) −→ (s, f ′ ) for some s 6∈ {r, f }.
v
– It follows that (f, r′ ) −→ (s, r′ v ′ ), where r′ v ′ 6= f ′ .
v
– The state (s, r′ v ′ ) is in F × F ′ , and thus is non-final. If (r, f ′ ) −→ (rv, f ′ ) is
′
′
final, then v distinguishes (f, r ) and (r, f ). Hence we may assume without
loss of generality that (rv, f ′ ) is non-final.
– A non-final state either lies in F × F ′ or F × F ′ . Since f ′ ∈ F ′ , we must have
(rv, f ′ ) ∈ F × F ′ . But then rv = f .
Thus we have
u
v
(p, p′ ) −→ (f, r′ ) −→ (s, r′ v ′ ),
u
v
(q, q ′ ) −→ (r, f ′ ) −→ (f, f ′ ).
Now, apply v again to both states.
–
–
–
–
Since s 6= r and rv = f , we have sv 6= f .
Since f ′ v ′ = f ′ and r′ 6= f ′ , we have r′ v ′ 6= f ′ and r′ v ′ v ′ 6= f ′ .
v
It follows that (s, r′ v ′ ) −→ (sv, r′ v ′ v ′ ) ∈ F × F ′ , and thus is non-final.
v
However, recall that f v = s and s 6= f ; thus (f, f ′ ) −→ (s, f ′ ) is in F × F ′ .
Hence (p, p′ ) and (q, q ′ ) are distinguished by uv 2 .
We have shown that all pairs of states of A × A′ are distinguishable by all
(F, F ′ )-compatible sets of final states, and so this proves (1) ⇒ (2).
The implication (2) ⇒ (3) is immediate. For (3) ⇒ (1), just note that for
each proper binary boolean operation ◦, the language L ◦ L′ is recognized by
(A × A′ )(X) for some set of final states X. If (A × A′ )(X) is not accessible, then
it cannot be minimal and thus L ◦ L′ cannot have maximal state complexity. ⊓
⊔
Note that Example 11 gives a pair of languages recognized by two-state permutation DFAs which have maximal complexity for intersection, but not symmetric
difference (see also [3, Example 2]). Hence in the previous theorem, it was necessary to assume that at least one DFA has three or more states.
26
Note that Theorem 1 also holds in the following cases:
– A and A′ both have exactly one non-final state.
– A has exactly one final state and A′ has exactly one non-final state.
– A has exactly one non-final state and A′ has exactly one final state.
The same arguments we gave in Theorem 1 can be used in the above three
cases, but the role of each argument is changed. For example, consider the case
where A has one final state and A′ has one non-final state. Let F = {f } and
let F ′ = Q′ \ F ′ = {q ′ }. We can use the same arguments as in the original proof
of Theorem 1, except wherever F ′ appears we substitute F ′ . So for example, we
deal with the case of F × F ′ = {(f, q ′ )} by appealing to Corollary 2, just like
we did for F × F ′ in the original proof. This works because distinguishability
arguments are the same whether we distinguish with respect to a set of final
states or a set of non-final states.
We now apply Theorem 1 to show that the original witnesses for the maximal state complexity of union (found by Maslow and later by Yu, Zhuang and
Salomaa) are in fact witnesses for all proper binary boolean operations.
Example 13. In [15], Maslov defined two families of DFAs over alphabet {0, 1} as
follows, and claimed that the languages they recognize are witnesses for union.
The DFA A has states {S0 , . . . , Sm−1 } with S0 initial and Sm−1 final, and the
transitions are given by Si 0 = Si , Si 1 = Si+1 for i 6= m − 1, and Sm−1 1 = S0 .
The DFA B has states {P0 , . . . , Pn−1 } with P0 initial and Pn−1 final, and the
transitions are given by Pi 1 = Pi , Pi 0 = Pi+1 for i 6= n − 1, and Pn−1 0 = P0 .
It is easy to see that A × B is accessible: the state (Si , Pj ) can be reached from
(S0 , P0 ) via the word 1i 0j . Furthermore, A and B are permutation DFAs: the
symbol 0 acts as the identity permutation in A and as a cyclic permutation of
the states in B, while 1 acts as a cyclic permutation in A and the identity in
B. They also have exactly one final state. So in fact, for m, n ≥ 3, the pair of
languages (L(A), L(B)) has maximal boolean complexity by Theorem 1. That
is, Maslov’s languages are witnesses for all proper binary boolean operations, in
addition to union.
Yu, Zhuang and Salomaa gave a different family of witnesses in [21]. For a
word w ∈ Σ ∗ and a ∈ Σ, let |w|a denote the number of occurrences of the letter
a in w. Yu et al. defined languages Lm = {w ∈ {a, b}∗ : |w|a ≡ 0 (mod m)}
and Ln = {w ∈ {a, b}∗ : |w|b ≡ 0 (mod n)}, then proved that Lm ∩ Ln and
Lm ∪ Ln both have state complexity mn. In fact, (Lm , Ln ) has maximal boolean
complexity for m, n ≥ 3. Indeed, one may verify that the minimal DFA of Lm
has m states, with the initial and final states equal; the letter a acts as a cyclic
permutation of the state set, and the letter b acts as the identity. The minimal
DFA of Ln is similar, except there are n states, b is the cyclic permutation, and a
is the identity. These DFAs are almost identical to the DFAs defined by Maslov,
except with a different choice of final state. This does not change the fact that
they are permutation DFAs with one final state and an accessible direct product,
and hence meet the conditions of Theorem 1.
27
4.3
Uniform Boolean Minimality
We now give sufficient conditions for a pair of permutation DFAs (A, A′ ) to be
uniformly boolean minimal. Just as with uniform minimality, primitive groups
play an important role in our conditions for uniform boolean minimality. To
state our conditions, we need some new notation.
For p, q ∈ Q and p′ , q ′ ∈ Q′ , we write Rp,q for the setwise stabilizer of
(p, ∗) ∪ (q, ∗), and Cp′ ,q′ for the setwise stabilizer of (∗, p′ ) ∪ (∗, q ′ ). If p = q
then Rp,q = Rp = Rq , and similarly if p′ = q ′ then Cp′ ,q′ = Cp′ = Cq′ . We call
these subgroups double row stabilizers and double column stabilizers. Under this
definition, single row stabilizers are special cases of double row stabilizers, and
similarly for column stabilizers.
Note that Rp and Rq are not necessarily subgroups of Rp,q , nor the other
way around: the group Rp might contain elements that map q to some state
r 6∈ {p, q}, while the group Rp,q might contain elements that swap p and q.
However, the full row stabilizer R is a common subgroup of Rp , Rq and Rp,q .
The analogous facts hold for column stabilizers.
Lemma 3. Suppose A and A′ are permutation DFAs.
1. If |Q| ≥ 3, the group G is primitive, and Rp,q π ′ ≤ G′ is primitive for all
p, q ∈ Q, then A × A′ is uniformly boolean minimal.
2. If |Q′ | ≥ 3, the group G′ is primitive, and Cp′ ,q′ π ≤ G is primitive for all
p′ , q ′ ∈ Q′ , then A × A′ is uniformly boolean minimal.
Note that we do not require p 6= q or p′ 6= q ′ , so in case (1) both the single and
double row stabilizers must be primitive, and in case (2) both the single and
double column stabilizers must be primitive.
Proof. Suppose that the conditions of (1) hold, that is, |Q| ≥ 3, G is primitive
and for all p, q ∈ Q, the subgroup Rp,q π ′ ≤ G′ is primitive. The case where the
conditions of (2) hold is symmetric. We want to show that for every pair of sets
∅ ( S ( Q and ∅ ( S ′ ( Q′ and each (S, S ′ )-compatible subset X ⊆ Q × Q′ ,
the DFA (A × A′ )(X) is minimal.
It suffices to consider the cases where X = S × S ′ and where X = (S × S ′ ) ∪
(S ×S ′ ). It may seem that this would only cover sets of type (a) and complements
of sets of type (e) from Proposition 4. However, these two cases actually cover
all possible types of (S, S ′ )-compatible sets.
To see this, consider a set S ×S ′ of type (b). If we prove that (A×A′ )(T ×T ′ )
is minimal for all pairs of sets ∅ ( T ( Q and ∅ ( T ′ ( Q′ , then in particular
we can take T = S and T ′ = S ′ to show that (A × A′ )(S × S ′ ) is covered. A
similar argument works for sets of type (c) and (d) .
Now note that if (A × A′ )(X) is minimal, then (A × A′ )(X) is also minimal,
so we also cover sets of type (e) and (f). So all types of sets from Proposition 4
are covered by just looking at the cases X = S × S ′ and X = (S × S ′ ) ∪ (S × S ′ ).
Let (i, i′ ) and (j, j ′ ) be distinct states of A × A′ ; we will show they are
distinguishable with respect to X.
Case 1 (States in the same row): Suppose i = j, that is, the states are
in the same row. Since the states are distinct, we have i′ 6= j ′ .
28
– Since G is primitive, it is transitive, and thus there exists a word w ∈ G that
w
w
maps i = j to some element s ∈ S. Thus (i, i′ ) −→ (s, i′ w′ ) and (j, j ′ ) −→
′ ′
(s, j w ).
– Suppose w does not distinguish (i, i′ ) and (j, j ′ ) with respect to X. Then
(s, i′ w′ ) ∈ X ⇔ (s, j ′ w′ ) ∈ X.
– Since Rs π ′ is primitive, all permutation DFAs with state set Q′ and transition group Rs π ′ are uniformly minimal by Corollary 1. It follows there exists
x′ ∈ Rs π ′ that distinguishes i′ w′ and j ′ w′ with respect to S ′ .
We have:
w
x
(i, i′ ) −→ (s, i′ w′ ) −→ (s, i′ w′ x′ ),
w
x
(j, j ′ ) −→ (s, j ′ w′ ) −→ (s, j ′ w′ x′ ).
– Since x′ distinguishes i′ w′ and j ′ w′ with respect to S ′ , we see that i′ w′ x′ ∈
S ⇔ j ′ w′ x′ 6∈ S.
– Hence (s, i′ w′ x′ ) ∈ S × S ′ ⇔ (s, j ′ w′ x′ ) ∈ S × S ′ .
It follows that either w or wx distinguishes (i, i′ ) and (j, j ′ ) with respect to X,
regardless of whether we have X = S × S ′ or X = (S × S ′ ) ∪ (S × S ′ ).
Case 2 (States in the same column): Suppose i′ = j ′ , that is, the states
are in the same column but different rows. Since G is primitive, all permutation
DFAs with state set Q and transition group G are uniformly minimal. Hence
there exists x ∈ G that distinguishes i and j with respect to S. Suppose without
loss of generality that ix ∈ S and jx 6∈ S. Since Rix,jx π ′ is primitive, it is
transitive, and thus there exists y ′ ∈ Rix,jx π ′ such that i′ x′ y ′ ∈ S. If y fixes ix
and jx, then (ixy, i′ x′ y ′ ) = (ix, i′ x′ y ′ ) ∈ S ×S ′ , and (jxy, j ′ x′ y ′ ) = (jx, j ′ x′ y ′ ) ∈
S×S ′ since i′ = j ′ implies i′ x′ y ′ = j ′ x′ y ′ . If y swaps ix and jx then (ixy, i′ x′ y ′ ) =
(jx, i′ x′ y ′ ) ∈ S × S ′ and (jxy, j ′ x′ y ′ ) = (ix, j ′ x′ y ′ ) ∈ S × S ′ . In either case, it
follows that xy distinguishes (i, i′ ) and (j, j ′ ) with respect to X, regardless of
whether we have X = S × S ′ or X = (S × S ′ ) ∪ (S × S ′ ).
Case 3 (States in different rows and different columns): Suppose
i 6= i′ and j 6= j ′ . We divide this case into two subcases.
Subcase 3a (X = S × S ′ ): We may assume without loss of generality that
(i, i′ ) and (j, j ′ ) are both in X. To see this, observe that since G is transitive and
Rq π ′ ≤ G′ is transitive for each q ∈ Q, we know that G× is transitive by Lemma
2. Thus there exists w ∈ Σ ∗ that maps (i, i′ ) to a state in X. Hence either w
distinguishes the states, or w also maps (j, j ′ ) into X.
Suppose we have (i, i′ ), (j, j ′ ) ∈ X = S × S ′ . Since Ri,j π ′ is primitive, there
exists w ∈ Ri,j π ′ that distinguishes i′ and j ′ with respect to S ′ . Without loss of
generality, assume i′ ∈ S ′ and j ′ 6∈ S ′ . Then (i, i′ )(w, w′ ) = (i, i′ ) ∈ S × S ′ and
(j, j ′ )(w, w′ ) ∈ S × S ′ . Hence w distinguishes the states.
Subcase 3b (X = (S × S ′ ) ∪ (S × S ′ )): This is the final case we must
deal with, and most complicated part of the proof. We introduce a notion of
polarity to simplify the arguments. We assign a polarity of 1 or −1 to each state
in Q × Q′ as follows. First, let q ∈ Q have polarity 1 if q ∈ S and polarity −1
if q 6∈ S. Similarly, q ′ ∈ Q′ has polarity 1 if q ′ ∈ S ′ and polarity −1 if q ′ 6∈ S ′ .
Then the polarity of (q, q ′ ) ∈ Q × Q′ is the product of the polarities of q and q ′ .
29
Next, we partition Q×Q′ into four quadrants: S×S ′ , S×S ′ , S×S ′ , and S×S ′ .
Notice that in each quadrant, all states have the same polarity. Furthermore, the
set X = (S × S ′ ) ∪ (S × S ′ ) is the set of all states with positive polarity, and the
set X = (S × S ′ ) ∪ (S × S ′ ) is the set of all states with negative polarity. Hence
to show all states are distinguishable by X, we must show that for each pair of
states of equal polarity, there is a word that preserves the polarity of one state
and reverses the polarity of the other state.
We now prove two claims, which together complete the proof of this subcase.
First we show that pairs of states in the same quadrant are distinguishable, and
then we show that pairs of states in different quadrants are distinguishable.
Claim 1 (States in the same quadrant are distinguishable): Suppose
(i, i′ ) and (j, j ′ ) are in the same quadrant. This means that (i, i′ ) and (j, j ′ ) have
the same polarity; furthermore, i and j have the same polarity, and i′ and j ′
have the same polarity.
– Choose a word w′ ∈ Ri,j π ′ that distinguishes i′ and j ′ with respect to S ′
(by primitivity of Ri,j π ′ ).
– Notice that w preserves the polarity of both i and j, since it either fixes both
i and j or it swaps them, and i and j have the same polarity.
– Since (i, i′ ) and (j, j ′ ) are in the same quadrant, we either have i′ , j ′ ∈ S ′ or
i′ , j ′ ∈ S ′ .
– Since w′ distinguishes i′ and j ′ with respect to S ′ , it follows that w′ acts on
i′ and j ′ by preserving the polarity of one state and reversing the polarity
of the other.
It follows that (w, w′ ) acts on (i, i′ ) and (j, j ′ ) by preserving the polarity of one
state and reversing the polarity of the other. In other words, w distinguishes
these states.
Claim 2 (States in different quadrants are distinguishable): Suppose
that (i, i′ ) and (j, j ′ ) lie in different quadrants.
– We may assume without loss of generality that (i, i′ ) and (j, j ′ ) have the
same polarity; otherwise they are trivially distinguishable.
– We may also assume without loss of generality that (i, i′ ), (j, j ′ ) ∈ X, by the
same argument we used in Subcase 3a. Thus we must have (i, i′ ) ∈ S × S ′
and (j, j ′ ) ∈ S × S ′ , or vice versa.
– We may assume without loss of generality that (i, i′ ) ∈ S × S ′ and (j, j ′ ) ∈
S × S ′ , by swapping the names of (i, i′ ) and (j, j ′ ) if necessary.
So we have reduced to the case where one state is in quadrant S × S ′ and the
other is in quadrant S × S ′ .
– Since G is primitive, S and S are either not blocks, or they are trivial blocks.
– S and S are proper non-empty subsets of Q, so they can only be trivial
blocks if |S| = |S| = 1.
– This would imply |Q| = |S| + |S| = 2, and we are assuming |Q| ≥ 3, so
they cannot both be trivial blocks. So at least one of S or S is not a block.
Note that in the case where G′ is primitive and the groups Ci′ ,j ′ π ≤ G are
primitive, we would use |Q′ | ≥ 3 here.
30
If S is not a block, let w ∈ G be a word such that ∅ ( Sw ∩ S ( S. Otherwise,
S is not a block, so let w ∈ G be a word such that ∅ ( Sw ∩ S ( S.
Now, we partition Q into two sets:
P = {q ∈ S : qw ∈ S} ∪ {q ∈ S : qw ∈ S},
P = {q ∈ S : qw ∈ S} ∪ {q ∈ S : qw ∈ S}.
Note that P is non-empty, since if it was empty we would have Sw ∩ S = ∅ and
Sw∩S = ∅. Similarly, P is non-empty, since otherwise we would have Sw∩S = S
and Sw ∩ S = S.
Observe that if i, j ∈ P , then P is a proper subset of Q with size at least
two, so it cannot be a block for G. Hence some word in G distinguishes i and j
by P . Similarly, if i, j ∈ P , then i and j are distinguishable by P . So we may
assume that either i ∈ P and j ∈ P , or i ∈ P and j ∈ P .
Suppose that i ∈ P and j ∈ P . Recall that we have (i, i′ ) ∈ S × S ′ and
(j, j ′ ) ∈ S × S ′ . Thus since i ∈ S we have iw ∈ S, and since j ∈ S we have
jw ∈ S. Hence (iw, i′ w′ ), (jw, j ′ w′ ) ∈ S × Q′ , that is, w maps both (i, i′ ) and
(j, j ′ ) into S × Q′ . There are two possibilities:
– The states (iw, i′ w′ ) and (jw, j ′ w′ ) are in the same quadrant, and thus are
distinguishable by Claim 1.
– The states (iw, i′ w′ ) and (jw, j ′ w′ ) are in different quadrants. Since iw and
jw are both in S, one state must lie in S × S ′ and the other S × S ′ . Thus w
distinguishes (i, i′ ) and (j, j ′ ), and we are done.
So if i ∈ P and j ∈ P , we have proved the claim. If we have i ∈ P and j ∈ P ,
then we have iw ∈ S and jw ∈ S, and so a symmetric argument shows that the
states are distinguishable. This completes the proof of Claim 2. By Claim 1 and
Claim 2, we see that all pairs of states are distinguishable by sets of the form
(S × S ′ ) ∪ (S × S ′ ), completing the proof of Subcase 3b and hence Case 3.
We have shown that for every pair of sets ∅ ( S ( Q and ∅ ( S ′ ( Q′ and
each (S, S ′ )-compatible subset X ⊆ Q × Q′ , each pair of states in (A × A′ )(X)
is distinguishable by X. Thus A × A′ is uniformly boolean minimal.
⊓
⊔
While Lemma 3 gives sufficient conditions for uniform boolean minimality,
the conditions are not necessary. We will demonstrate this later, in Example 20.
In the above proof, most of the difficulty came from dealing with the case
(S ×S ′ )∪(S ×S ′ ), which corresponds to the operation of symmetric difference (or
complement of symmetric difference). In fact, if we choose to ignore the operation
of symmetric difference, we can obtain necessary and sufficient conditions for the
corresponding weaker version of uniform boolean minimality.
Proposition 7. Suppose A and A′ are permutation DFAs. The following are
equivalent:
1. Rq π ′ ≤ G′ and Cq′ π ≤ G are primitive for all q ∈ Q and q ′ ∈ Q′ .
2. For all sets ∅ ( S ( Q and ∅ ( S ′ ( Q′ , the DFA (A × A′ )(S × S ′ ) is
minimal.
31
Proof. (1) ⇒ (2): We proceed as in the proof of Lemma 3. Let (i, i′ ) and (j, j ′ )
be distinct states of A × A′ ; we will show they are distinguishable with respect
to S × S ′ .
Case 1 (States in the same row): In the proof of Lemma 3, we proved that
states in the same row are distinguishable, using the facts that G is transitive and
Rq π ′ is primitive for all q ∈ Q. Those facts still hold under our new hypotheses,
so the same argument can be used here.
Case 2 (States in the same column): The argument we used for this case
in Lemma 3 relied on the double row stabilizers being transitive, so we cannot
use it here. However, under our new hypotheses, we know that G′ is transitive
and Cq′ π is primitive for all q ′ ∈ Q. Thus we can just use a symmetric argument
to the one for states in the same row.
Case 3 (States in different rows and different columns): Suppose
i 6= j and i′ 6= j ′ . As in the proof of Lemma 3, we may assume without loss
of generality that (i, i′ ), (j, j ′ ) ∈ S × S ′ . Since Ci′ π ′ is primitive, there exists
w ∈ Ci′ π that distinguishes i and j with respect to S. We claim that we can
choose w so that iw ∈ S and jw 6∈ S:
– Suppose for a contradiction that for all words w ∈ Ci′ π which distinguish i
and j by S, we have iw 6∈ S and jw ∈ S.
– Fix x ∈ Ci′ π such that ix 6∈ S and jx ∈ S.
y
– By transitivity of Ci′ π, we can choose y ∈ Ci′ π such that (ixy, i′ ) −→ (jx, i′ ).
It follows ixy = jx, and thus ixy ∈ S.
– Suppose jxy 6∈ S. Then we have ixy ∈ S and jxy 6∈ S, where xy ∈ Ci′ π. This
contradicts our assumption that for all words w ∈ Ci′ π which distinguish i
and j by S, we must have iw 6∈ S and jw ∈ S.
– Thus we can assume jxy ∈ S. But then since ixy = jx, we have ixy 2 =
jxy ∈ S. If jxy 2 6∈ S, we get a contradiction as before, so jxy 2 ∈ S.
– In general, we have ixy k = jxy k−1 , and it follows by induction on k that
ixy k ∈ S and jxy k ∈ S for all k > 1.
– But Ci′ π is a finite group, so y k = ε for some k. This gives ixy k = ix ∈ S
and jxy k = jx ∈ S, which contradicts the fact that ix 6∈ S.
This proves the claim; there must exist w ∈ Ci′ π such that iw ∈ S and jw 6∈ S.
w
w
Then we have (i, i′ ) −→ (iw, i′ ) ∈ S × S ′ , but (j, j ′ ) −→ (jw, j ′ w′ ) 6∈ S × S ′ since
jw 6∈ S. Thus w distinguishes the states. We have now shown that all pairs of
states (i, i′ ) and (j, j ′ ) are distinguishable.
(2) ⇒ (1): Suppose that for all sets ∅ ( S ( Q and ∅ ( S ′ ( Q′ , the DFA
(A × A′ )(S × S ′ ) is minimal. Assume for a contradiction that there exists q ′ ∈ Q′
such that Cq′ π is not primitive. Then there exists a non-trivial block B ( Q for
Cq′ π. We claim that (A × A)(B × {q ′ }) is not minimal:
– Since B is a non-trivial block, it contains at least two distinct elements. Let
i, j ∈ B be distinct and consider the states (i, q ′ ) and (j, q ′ ) of A × A′ .
– If q ′ w′ 6= q ′ , then w does not distinguish (i, q ′ ) and (j, q ′ ). Indeed, if q ′ w′ 6= q ′ ,
then (iw, q ′ w′ ) and (jw, q ′ w′ ) both lie outside of B × {q ′ }.
32
– Hence if w ∈ Σ ∗ distinguishes (i, q ′ ) and (j, q ′ ), we must have q ′ w′ = q ′ , and
thus w ∈ Cq′ π.
– Since B is a block for Cq′ π, we either have iw, jw ∈ B (if Bw = B) or
{iw, jw} ∩ B = ∅ (if Bw ∩ B = ∅).
– Thus w ∈ Cq′ π cannot distinguish (i, q ′ ) and (j, q ′ ): if iw, jw ∈ B then w
maps both states to B × {q}; otherwise it maps both states to B × {q}.
– It follows that no word can distinguish (i, q ′ ) and (j, q ′ ) by B × {q ′ }.
This shows that (A × A′ )(B × {q ′ }) is not minimal, which is a contradiction.
It follows that Cq′ π must be primitive for all q ′ ∈ Q′ . A symmetric argument
shows that Rq π ′ must be primitive for all q ∈ Q.
⊓
⊔
One may wonder whether condition (1) of Proposition 7 are actually sufficient
to prove Lemma 3. We will show later (Example 19) that this is not the case.
4.4
Dissimilar DFAs
While the conditions of Lemma 3 are somewhat complicated, there are cases
where we can easily verify that they hold. We consider some of these cases next.
We will say that the DFAs A and A′ are similar if the maps π : M× → M and
′
π : M× → M ′ are injective. By Proposition 5, π and π ′ are always surjective,
so if they are also injective then they are isomorphisms. Hence if A and A′ are
similar, the map π −1 π ′ : M → M ′ given by w 7→ w′ is a well-defined monoid
isomorphism. If A and A′ are not similar, we say they are dissimilar. If π and
π ′ both fail to be injective, we say that A and A′ are strongly dissimilar.
We give some examples of dissimilar DFAs. All DFAs will be over the twoletter alphabet {a, b}, and we will not specify the initial and final states since
they do not affect whether DFAs are similar.
Example 14. Let A have states {1, 2} and transformations a = (1, 2) and b = ().
Let A′ have states {1, 2} and transformations a′ = () and b′ = (1, 2). Notice that
in the transition group of A×A′ , we have (b, b′ ) = ((), (1, 2)) and (ε, ε′ ) = ((), ()).
Thus (b, b′ )π = (ε, ε′ )π = () and it follows that π is not injective. Hence A and
A′ are dissimilar. In fact, a symmetric argument shows that π ′ is not injective,
and thus these DFAs are strongly dissimilar.
Another way to see that these DFAs are dissimilar is to consider the “map”
w 7→ w′ . This “map” is not actually well-defined, since from the fact that b 7→ b′
we must have () 7→ (1, 2), but from the fact that ε 7→ ε′ we must have () 7→ ().
(Formally, this “map” is a binary relation; we say it is “well-defined” if the
relation happens to be a function, and otherwise is not.) This means A and
A′ cannot be similar, since we know that if they are similar, then π −1 π ′ is a
well-defined isomorphism which sends w to w′ .
Alternatively, without even checking whether the “map” w 7→ w′ is well
defined, we can see that since () = b 7→ b′ = (1, 2), this “map” sends an element
of order one to an element of order two, and thus it cannot possibly be a group
isomorphism. But then A and A′ cannot be similar.
33
Usually, the easiest way to prove that a pair of DFAs is dissimilar is to
examine the “map” w 7→ w′ and show that either it is not well-defined or not
an isomorphism.
Example 15. Let A have states {1, 2} and transformations a = (1, 2), b = (1, 2).
Let A′ have states {1, 2, 3, 4} and transformations a′ = (1, 2)(3, 4), b = (1, 3)(2, 4).
The transition group of A has two elements: ε = () and a = b = (1, 2). However,
the transition group of A′ has four elements:
ε = (), a′ = (1, 2)(3, 4), b′ = (1, 3)(2, 4), a′ b′ = (1, 4)(2, 3).
Hence these DFAs must be dissimilar, since they have different transition groups.
Similar DFAs always have isomorphic transition monoids/groups.
These DFAs are not strongly dissimilar. To see this, observe that the transition group of A × A′ has four elements:
(ε, ε′ ) = ((), ()),
(a, a′ ) = ((1, 2), (1, 2)(3, 4)),
(b, b′ ) = ((1, 2), (1, 3)(2, 4)),
(ab, a′ b′ ) = ((), (1, 4)(2, 3)).
It is easy to verify that any other product of elements will be equal to one of
these four. Now note that π ′ is a bijection but π is not, and thus A and A′
are not strongly dissimilar. In this case, the “map” w 7→ w′ is not well-defined.
However, the map (π ′ )−1 π given by w′ 7→ w is well-defined, and in fact is a
group homomorphism (but not an isomorphism).
As for examples of similar DFAs, we have the following fact: isomorphic DFAs
are necessarily similar. Indeed, suppose there is an isomorphism f : Q → Q′ .
Then for all q ∈ Q and a ∈ Σ, we have (qa)f = (qf )a′ , and thus qa = q(f a′ f −1 ).
It follows that a = f a′ f −1 for all a ∈ Σ ∗ , and thus w = f w′ f −1 for all w ∈ Σ ∗ .
Hence π : M× → M is given by (w, w′ )π = (f w′ f −1 , w′ )π = f w′ f −1 , and this
map is clearly injective since if f w′ f −1 = f x′ f −1 then w′ = x′ . Similarly, we
have w′ = f −1 wf and π ′ : M× → M is injective.
This shows that DFA similarity is a generalization of DFA isomorphism.
However, the next example shows that it is possible for two DFAs with different
numbers of states to be similar, and that the monoid isomorphism w 7→ w′ does
not necessarily have to be conjugation by a permutation.
Example 16. Consider the symmetric group S10 . This group contains an intransitive subgroup that is isomorphic to S5 , given by permutations of {1, . . . , 10}
which fix every point in {6, . . . , 10}. This can be considered the “natural” embedding of S5 in S10 . However, there is also a primitive subgroup of S10 that
is isomorphic to S5 , and is not conjugate to this natural embedding. In GAP’s
library of primitive groups, this subgroup can be accessed with the command
PrimitiveGroup(10,2). We can use GAP to compute an explicit isomorphism
between S5 and this subgroup:
gap> IsomorphismGroups(SymmetricGroup(5),PrimitiveGroup(10,2));
[ (3,4), (1,2,3)(4,5) ] ->
[ (1,2)(6,8)(7,9), (2,6,4,5,3,7)(8,10,9) ]
34
The GAP output tells us that the map which sends (3, 4) to (1, 2)(6, 8)(7, 9) and
(1, 2, 3)(4, 5) to (2, 6, 4, 5, 3, 7)(8, 10, 9) can be extended multiplicatively to an
isomorphism between S5 and the aforementioned primitive subgroup of S10 . We
use this isomorphism to construct similar permutation DFAs of different sizes.
Let A have states {1, . . . , 5} and transformations a = (3, 4), b = (1, 2, 3)(4, 5).
Let A′ have states {1, . . . , 10} and transformations a′ = (1, 2)(6, 8)(7, 9) and
b′ = (2, 6, 4, 5, 3, 7)(8, 10, 9).
The transition group G of A is S5 , and the transition group G′ of A′ is
a primitive subgroup of S10 that is isomorphic to S5 . Furthermore, the map
w 7→ w′ is an isomorphism of G and G′ . Hence A and A′ are similar. This pair
of DFAs has other interesting properties; we will revisit them in Example 21.
Notice that similarity of DFAs is a very fragile property; if we simply switch
the roles of a′ and b′ in A′ , giving a = (3, 4) and a′ = (2, 6, 4, 5, 3, 7)(8, 10, 9),
then A and A′ are no longer similar. Indeed, after the switch, wee see that
w 7→ w′ sends an element of order two to an element of order six, which means
it cannot be an isomorphism of G and G′ .
Dissimilar permutation DFAs have the following nice property.
Proposition 8. If A and A′ are dissimilar permutation DFAs, then at least one
of the following statements holds:
1. Rπ ′ is a non-trivial normal subgroup of G′ .
2. Cπ is a non-trivial normal subgroup of G.
If A and A′ are strongly dissimilar, then both hold.
Proof. Recall that R = ker π and C = ker π ′ ; these are normal subgroups of G× .
Since π and π ′ are surjective, Rπ ′ is a normal subgroup of G′ and Cπ is a normal
subgroup of G.
– We have ker π = {(w, w′ ) ∈ G× : w = ε}, so Rπ ′ = {w′ ∈ G′ : w = ε} and
similarly Cπ = {w ∈ G : w′ = ε′ }.
– If Rπ ′ is trivial, then whenever w = ε we have w′ = ε′ , and so R = ker π =
{(ε, ε′ )} is trivial; hence π is injective.
– Similarly, if Cπ is trivial, then C = ker π ′ is trivial, and thus π ′ is injective.
Thus we see that if A and A′ are dissimilar, then π and π ′ cannot both be
injective, and so Rπ ′ and Cπ cannot both be trivial. Furthermore, if A and A′
are strongly dissimilar, then neither π nor π ′ can be injective, and so neither
Rπ ′ nor Cπ can be trivial.
⊓
⊔
This leads to a useful theorem:
Theorem 2. Let A and A′ be dissimilar permutation DFAs with |Q|, |Q′ | ≥ 3.
1. Suppose G and G′ are transitive. If all non-trivial normal subgroups of G
and of G′ are transitive, then A × A′ is accessible.
2. Suppose G and G′ are primitive. If all non-trivial normal subgroups of G
and of G′ are primitive, then A × A′ is uniformly Boolean minimal.
35
Proof. By Proposition 8, since A and A′ are dissimilar, one of Rπ ′ or Cπ is a
non-trivial normal subgroup. Suppose that Cπ ≤ G is non-trivial; the other case
is symmetric.
(1): Since all non-trivial normal subgroups of G are transitive, Cπ is transitive. Hence Cq′ π is transitive for all q ′ ∈ Q, since Cq′ π ≥ Cπ. Since G′ is
transitive, we see that condition (3) of Lemma 2 holds. Thus A × A′ is accessible.
(2): Since all non-trivial normal subgroups of G are primitive, Cπ is primitive. Hence Cp′ ,q′ π is primitive for all p′ , q ′ ∈ Q, since Cp′ ,q′ π ≥ Cπ. Since G′ is
primitive, we see that A and A′ meet the conditions of Lemma 3. Thus A × A′
is uniformly Boolean minimal.
⊓
⊔
The power of Theorem 2 comes from the fact that several interesting classes of
groups have the property that all non-trivial normal subgroups are transitive or
primitive. The next corollary gives examples of when Theorem 2 can be applied.
Corollary 3. Let A and A′ be dissimilar permutation DFAs with |Q|, |Q′ | ≥ 3.
1. Suppose G and G′ are transitive.
(a) If G and G′ are transitive simple groups, then A × A′ is accessible.
(b) If G and G′ are primitive groups, then A × A′ is accessible.
2. Suppose G and G′ are primitive.
(a) If G and G′ are primitive simple groups, then A×A′ is uniformly Boolean
minimal.
(b) If G is SQ or AQ , and G′ is SQ′ or AQ′ , and we have (|Q|, |Q′ |) 6∈
{(3, 4), (4, 3), (4, 4)}, then A × A′ is uniformly Boolean minimal.
(c) If G and G′ are 2-transitive groups which are not of affine type, then
A × A′ is uniformly Boolean minimal.
Before proving this, we will explain what we mean by “affine type”. The notion
of “affine type” comes from the O’Nan-Scott theorem [12, Theorem 4.1A], a
structure theorem for primitive groups. The O’Nan-Scott theorem divides the
primitive groups into different types based on their socle. The socle of a group
G is the subgroup generated by all the minimal normal subgroups of G, that is,
the normal subgroups N of G for which there does not exist a non-trivial normal
subgroup N ′ of G with N ′ ⊆ N .
A group is abelian if its binary operation is commutative. A primitive group
with an abelian socle is necessarily of affine type, which means it is a permutation
group of degree pd for p prime and d ≥ 1, and is isomorphic to a subgroup of the
affine group AGL(d, p). This is a group of permutations of the set Fdp , where Fp is
the finite field with p elements; it consists of all maps of the form (γ1 , . . . , γd ) 7→
(αγ1 + β, . . . , αγd + β), where α, β, γ1 , . . . , γd ∈ Fp . After this proof, we will show
that there exist DFAs A and A′ whose transition groups are 2-transitive groups
of affine type, such that A × A′ is not uniformly Boolean minimal. Hence in
the statement of Corollary 3 (2c), excluding 2-transitive groups of affine type
is necessary. We remark that A3 , S3 , A4 and S4 happen to be groups of affine
type; we will see in the proof that this is not true for symmetric and alternating
groups of larger degree.
36
Proof. (1a): Recall that a group is simple if it has no non-trivial proper normal
subgroups. Hence if G is a transitive simple group, then the only non-trivial
normal subgroup of G is G itself. Similarly, the only non-trivial normal subgroup
of G′ is G′ itself. Thus the conditions of Theorem 2 hold.
(1b) Suppose G and G′ are primitive. It is an easy exercise in group theory
to show that all non-trivial normal subgroups of a primitive group are transitive
(e.g., see [12, Theorem 1.6A]). Thus the conditions of Theorem 2 hold in this
case.
(2a): Since G is a primitive simple group, then the only non-trivial normal
subgroup of G is G itself, and similarly for G′ . Thus the conditions of Theorem
2 hold.
(2b): First suppose |Q| 6= 4 and |Q′ | 6= 4. It is well-known that An is simple
for n 6= 4 (e.g., see [12, Corollary 3.3A]). Thus AQ and AQ′ are primitive simple
groups. So if G = AQ and G′ = AQ′ , this case follows from (2a).
Now, consider the symmetric groups. It is well-known that for n 6= 4, the
only non-trivial normal subgroups of Sn are An and Sn .
To see this, let N be a non-trivial normal subgroup of Sn . If An ≤ N , we
claim that either N = An or N = Sn . Indeed, if N properly contains An , then
it contains some permutation p that cannot be written as a product of an even
number of 2-cycles. But every permutation can be written as a product of 2cycles, so p can be written as a product of an odd number of 2-cycles. Thus for
each 2-cycle (i, j), we have (i, j)p ∈ An ≤ N . So N contains (i, j)pp−1 = (i, j);
since N contains all 2-cycles, we must have N = Sn .
If N ≤ An , consider ghg −1 for g ∈ An and h ∈ N . Since g, h and g −1 are
all in An , each can be written a product of an even number of 2-cycles, and
thus ghg −1 can be written this way as well. So N is also a non-trivial normal
subgroup of An ; but An is simple for n 6= 4, so we must have N = An .
Thus a non-trivial normal subgroup is either Sn or An . both of which are
′
)}, Theorem 2 applies
primitive. Thus if (G, G′ ) ∈ {(AQ , SQ′ ), (SQ , AQ′ ), (SQ , SQ
and gives the result. Note we have also shown that for n ≥ 5, the alternating
group An is the unique minimal normal subgroup of Sn and An . Thus the socle
of Sn and An is non-abelian, which shows that Sn and An are not of affine type
when n ≥ 5.
Now, suppose |Q| = 4 or |Q′ | = 4 and (|Q|, |Q′ |) 6∈ {(3, 4), (4, 3), (4, 4)}. Then
|Q| ≥ 5 or |Q′ | ≥ 5. Assume without loss of generality that |Q| = 4 and |Q′ | ≥ 5;
the other case is symmetric. Consider the normal subgroup Rπ ′ of G′ . If Rπ ′ is
trivial, then as we argued in the proof of Proposition 8, the map π : G× → G
must be injective. Since G is either A4 or S4 , this means |G× | ≤ |S4 | = 24. But
the map π ′ : G× → G′ is surjective, and G is either AQ′ or SQ′ for |Q′ | ≥ 5. This
means |G× | ≥ |A5 | = 60. So we have 60 ≤ 24, which is a contradiction. Thus
Rπ ′ cannot be trivial. So Rπ ′ is a non-trivial normal subgroup of G′ , and thus
it is primitive, since G′ is SQ′ or AQ′ and |Q′ | ≥ 5. Thus the arguments in the
proof of Theorem 2 apply and A × A′ is uniformly Boolean minimal.
(2c): The results from permutation group theory that we use for this case
are somewhat more advanced. We will need the fact that a 2-transitive group
37
has a unique minimal normal subgroup; this follows from two theorems in [12]
(Theorem 4.1B and Theorem 4.3B), or alternatively is stated as a single result
in [10] (Proposition 5.2). The other fact we need is that if the socle of a 2transitive group is non-abelian, it is necessarily primitive [12, Theorem 7.2E].
Now, suppose G is 2-transitive. The socle of G is the subgroup generated by
all the minimal normal subgroups; but G has a unique minimal normal subgroup
N , so the socle of G is just equal to N . If N is abelian, then G is of affine
type. Thus we may assume the socle N is non-abelian; then it follows that N is
primitive. Since N is the unique minimal normal subgroup of G, every non-trivial
normal subgroup of G contains N , and thus is primitive. Similarly, every nontrivial normal subgroup of G′ is primitive. It follows that Theorem 2 applies. ⊓
⊔
4.5
Affine Groups
We now construct an infinite family of pairs of dissimilar permutation DFAs
which have 2-transitive transition groups of affine type, and are not uniformly
boolean minimal. The details of the construction require some knowledge of
finite fields. We will first give the construction in full generality, and then use
the construction to produce an explicit pair of 8-state DFAs.
Example 17. For k ≥ 0, let F2k denote the finite field of order 2k . For α, β, ξ ∈
F2k with α 6= 0, define tα,β : F2k → F2k to be the map ξ 7→ αξ + β. The
set of all such maps forms a group of permutations of F2k , which is called the
1-dimensional affine group on F2k and is denoted AGL(1, 2k ). Multiplication
(that is, composition of maps) in the affine group is given by the rule
tα,β tγ,ξ = tαγ,βγ+ξ .
It is an easy but somewhat tedious exercise to show that the affine group is
2-transitive. In [12, Chapter 4] it is proved that the affine group has an abelian
socle.
Recall that the multiplicative group of a finite field is cyclic. Let x be a
generator for the multiplicative group of F2k . We claim that the elements tx,0
and t1,1 generate AGL(1, 2k ). This is once again an easy but tedious exercise,
so we omit a proof.
An element of AGL(1, 2k ) of the form t1,β for some β ∈ F2k is called a
translation. The translations form a subgroup of AGL(1, 2k ), which we call T .
We claim that the subgroup of translations T is imprimitive and contains a block
B of size 2k−1 .
To see this, recall that F2k is a k-dimensional vector space over F2 . Pick
k − 1 non-zero elements of F2k and let B be the subspace spanned by them. Now
consider Bt1,β for β ∈ F2k .
– If β ∈ B, then α+β ∈ B for all α ∈ B, since B is a subspace; thus Bt1,β = B.
– If β 6∈ B, then for all α ∈ B we have α + β 6∈ B. Indeed, if α + β was in B
then (α + β) − α = β would be in B, since B is a subspace.
– Thus if β 6∈ B, then Bt1,β ∩ B = ∅.
38
And so, we see that B is indeed a block for T .
Consider the subgroup A of AGL(1, 2k ) × AGL(1, 2k ) generated by the elements a = (tx,0 , tx,0 ), b = (t1,1 , t1,0 ) and c = (t1,0 , t1,1 ). We claim that every
element (tα,β , tγ,ξ ) of A has the property that α = γ. For simplicity, we will call
elements with this property balanced.
Certainly the elements a, b and c are balanced, and so is the identity element
(t1,0 , t1,0 ). We will show that multiplying a balanced element on the right by a,
b or c results in a balanced element. Indeed, first observe that tα,β tx,0 = tαx,βx
and tα,β t1,1 = tα,β+1 . Thus if we take an arbitrary balanced element (tα,β , tα,γ ),
then we have:
(tα,β , tα,γ )a = (tα,β tx,0 , tα,γ tx,0 ) = (tαx,βx , tαx,γx ).
(tα,β , tα,γ )b = (tα,β t1,1 , tα,γ t1,0 ) = (tα,β+1 , tα,γ ).
(tα,β , tα,γ )c = (tα,β t1,0 , tα,γ t1,1 ) = (tα,β , tα,γ+1 ).
Since a, b and c and the identity are balanced, and multiplying a balanced
element by a, b or c results in a balanced element, it follows the group ha, b, ci = A
consists solely of balanced elements.
Next, let B = F2k \B and consider the set X = (B ×B)∪(B ×B) ⊆ F2k ×F2k .
Notice that (0, 0), (1, 1) ∈ F2k × F2k both lie in X. We claim that elements of A
cannot distinguish (0, 0) and (1, 1) with respect to X, that is, for all g ∈ A we
have (0, 0)g ∈ X ⇔ (1, 1)g ∈ X.
– To see this, consider an arbitrary element g = (tα,β , tα,γ ) of A.
– We have 0tα,β = β and 1tα,β = α + β. It follows that (0, 0)g = (β, γ) and
(1, 1)g = (α + β, α + γ) = (βt1,α , γt1,α ).
– Since B is a block for the subgroup of translations T , we either have Bt1,α =
B or Bt1,α ∩ B = ∅.
– But B has size 2k−1 , which is exactly half the size of F2k , so if Bt1,α ∩ B = ∅
then Bt1,α = B.
It follows that if (β, γ) is in B × B or B × B (that is, (β, γ) is in X) then
(βt1,α , γt1,α ) is also in B × B or B × B (and thus in X).
Likewise, if (β, γ) is not in X, then it is either in B × B or B × B, and it
follows (βt1,α , γt1,α ) is not in X. Thus (0, 0)g ∈ X ⇔ (1, 1)g ∈ X for all g ∈ A.
Finally, for each k ≥ 1, we construct a pair of DFAs over the alphabet {a, b, c}
with 2k states each, which both have AGL(1, 2k ) as their transition group, but
do not have a uniformly boolean minimal direct product.
We define a DFA A as follows:
– The state set is F2k , the initial state is 0, and the final state set is B.
– The transformations are a = tx,0 , b = t1,1 and c = t1,0 .
We define A′ in the same way as A, except the roles of b and c are swapped:
– The transformations are a′ = tx,0 , b′ = t1,0 and c′ = t1,1 .
39
Since tx,0 and t1,1 generate AGL(1, 2k ), it is clear that both DFAs have AGL(1, 2k )
as their transition group.
Now consider A × A′ . Let ◦ be the “complement of symmetric difference”
operation, so that B ◦ B = (B × B) ∪ (B × B). Observe that the transition group
of A × A′ is simply the group A. Thus the states (0, 0) and (1, 1) of A × A′
are not distinguishable by B ◦ B. Hence (A × A′ )(B ◦ B) is not minimal, and it
follows that A × A′ is not uniformly boolean minimal.
Example 18. We now carry out the construction of Example 17 for k = 3, 2k = 8.
First, we must construct the finite field F8 . Let F2 = {0, 1} denote the ring of
integers modulo two. We define F8 to be the quotient ring F2 [x]/hx3 + x + 1i.
This ring consists of polynomials of the form {a + bx + cx2 : a, b, c ∈ F2 }.
Addition and multiplication work as usual for polynomials, except when multiplying, any terms of degree 3 or higher are reduced by repeatedly applying
the rule x3 = x + 1. Note also that since the coefficients of the polynomials
are in F2 , we have 2f (x) = 0 for all polynomials f (x). For example, we have
(x2 + x + 1) + (x2 + 1) = x. Applying these facts, we see that:
x3 = x + 1,
x4 = x2 + x,
x5 = x3 + x2 = x2 + x + 1,
x6 = x3 + x2 + x = x2 + x + x + 1 = x2 + 1,
x7 = x3 + x = x + 1 + x = 1.
These calculations show that every non-zero element of F8 can be written as
a power of the monomial x; in particular, 1 can be written as a power of x,
and thus x is invertible. Hence F8 is indeed a field and x is a generator of the
multiplicative group of F8 .
Next, we explicitly write out the permutations tx,0 and t1,1 in cycle notation:
tx,0 = (x, x2 , x3 , x4 , x5 , x6 , x7 ) = (x, x2 , x + 1, x2 + x, x2 + x + 1, x2 + 1, 1).
t1,1 = (0, 1)(x, x+1)(x2 , x2 +1)(x2 +x, x2 +x+1) = (0, x7 )(x, x3 )(x2 , x6 )(x4 , x5 ).
Finally, we need to find a block B for the subgroup of translations of AGL(1, 8).
The construction tells us to pick two non-zero elements of F8 and let B be the
subspace spanned by them. If we take 1 and x, we get B = {0, 1, x, x + 1}.
We now have all the information we need to construct A and A′ . A state
diagram for A is shown in Figure 5, with the self-loops on c omitted. One may
verify computationally that A × A′ is not minimal when it is assigned the final
state set (B × B) ∪ (B × B).
For k = 1, we get DFAs A, A′ and A × A′ that are isomorphic to the
DFAs of Example 11. For k = 2, it happens that AGL(1, 4) is isomorphic to
the alternating group A4 . Hence the k = 2 case gives an example of dissimilar
DFAs that are not uniformly boolean minimal, and have alternating groups as
their transition groups. As Corollary 3 shows, this example does not generalize
to alternating groups of higher degree. The DFA A of Example 9 is isomorphic
to the DFA A produced by the construction of Example 17 with k = 2.
The construction of Example 17 also shows that condition (1) of Proposition
7 is not sufficient for uniform boolean minimality.
40
a
b
b
0
x7
a
a
x
x6
a
b
x2
b b
b
a
a
x3
x5
a
b
x4
a, b
Fig. 5. DFA A of Example 18. Each state also has a self-loop on letter c; these transitions are omitted from the diagram. The final state set is B = {0, x7 = 1, x, x3 = x+1},
the block of the subgroup of translations of AGL(1, 8) that was found in Example 18.
Example 19. The DFAs A and A′ constructed in Example 17 are not uniformly
boolean minimal. However, we claim the subgroups Rα π ′ ≤ G′ and Cα π ≤ G are
primitive for all α ∈ F2k , and thus A and A′ meet condition (1) of Proposition
7. In fact, the groups Rα π ′ and Cα π are equal to AGL(1, 2k ). First, we show
that R0 π ′ and C0 π are equal to AGL(1, 2k ). Consider (a, a′ ) = (tx,0 , tx,0 ).
– Since tx,0 fixes 0, it follows that (a, a′ ) ∈ R0 and (a, a′ ) ∈ C0 .
Thus a′ = tx,0 ∈ R0 π ′ and a = tx,0 ∈ C0 π.
– Since (b, b′ ) = (t1,1 , t1,0 ) and b′ = t1,0 fixes 0, we see that (b, b′ ) ∈ C0 .
Thus b = t1,1 ∈ C0 π.
– Similarly, since (c, c′ ) = (t1,0 , t1,1 ), we see that (c, c′ ) ∈ R0 .
Thus c′ = t1,1 ∈ R0 π ′ .
Since tx,0 and t1,1 generate AGL(1, 2k ), and these elements are in R0 π ′ and C0 π,
it follows R0 π ′ and C0 π are equal to AGL(1, 2k ) and thus are primitive.
To show that Rα π ′ and Cα π are primitive for all α 6= 0, we prove a general
fact about single row and column stabilizers: if Cp′ π ≤ G is primitive for some
p′ ∈ Q′ and G′ is transitive, then Cq′ π is primitive for all q ′ ∈ Q′ (and similarly
for single row stabilizers).
To see this, choose w′ ∈ G′ such that p′ w′ = q ′ .
Let B be a block for Cq′ π. We claim Bw−1 is a block for Cp′ π.
To see this, choose x ∈ Cp′ π and consider Bw−1 x ∩ Bw−1 .
If Bw−1 x ∩ Bw−1 6= ∅, then Bw−1 xw ∩ B 6= ∅.
Now, for all x ∈ Cp′ π, we have p′ x′ = p′ by definition.
It follows q ′ (w′ )−1 x′ w′ = p′ x′ w′ = p′ w′ = q ′ . Since (w′ )−1 x′ w′ fixes q ′ , we
have w−1 xw ∈ Cq′ π.
– Since B is a block for Cq′ π and Bw−1 xw ∩ B 6= ∅, we have Bw−1 xw = B.
– Hence Bw−1 x = Bw−1 , which proves Bw−1 is a block for Cp′ π.
–
–
–
–
–
–
41
It follows that if B is a block for Cq′ π, it must be a trivial block; otherwise Bw−1
is a non-trivial block for the primitive group Cp′ π, which is a contradiction.
Thus Cα π is primitive for all α ∈ F2k , and symmetrically we see that Rα π ′
is primitive for all α ∈ F2k . This shows that A and A′ satisfy condition (1) of
Proposition 7, yet A × A′ is not uniformly boolean minimal.
We can also use DFAs derived from affine groups to show that the conditions
of Lemma 3 are not necessary for uniform boolean minimality.
Example 20. As in Example 18, we define two DFAs A and A′ that have transition group AGL(1, 8). However, this time the direct product of the DFAs will
be uniformly boolean minimal. We define A as follows (leaving the final state
set unspecified).
– The state set is F8 , constructed as in Example 18, and the initial state is 0.
– The transformations are a = tx,0 and b = t1,1 .
′
Define A′ to have the same states as A and transformations a′ = t−1
x,0 , b = t1,1 .
′
Since A and A only have 8 states each, we were able to verify computationally that A × A′ is uniformly boolean minimal by a brute force approach. We
computed A × A′ , and for each pair of sets ∅ ( S ( Q and ∅ ( S ′ ( Q′ , we
checked the minimality of (A × A′ )(X) for all (S, S ′ )-compatible sets X.
We have also verified computationally that R0,1 π ′ and C0,1 π are imprimitive,
and hence A and A′ do not meet the conditions of Lemma 3. We verified this
by using the cycle notation representation we found for tx,0 and t1,1 to explicitly
construct the transition group G× of A × A′ in GAP. Then we computed the
setwise stabilizer R0,1 of {0, 1} × F8 (the rows indexed by 0 and 1), and the
setwise stabilizer C0,1 of F8 × {0, 1} (the columns indexed by 0 and 1). Next,
we computed R0,1 π ′ ≤ G′ and C0,1 π ≤ G. These groups turned out to both be
equal to T , the subgroup of translations in AGL(1, 8). We saw earlier than T is
imprimitive.
We suspect that if this construction is generalized to AGL(1, 2k ), the resulting DFAs will have the same property of being uniformly boolean minimal but
having R0,1 π ′ and C0,1 π imprimitive. However, we were unable to prove this.
4.6
Similar DFAs
To close out Section 4, we consider what happens when the DFAs A and A′
are similar. We have not investigated this case very deeply. In some ways, it
seems much more difficult than the dissimilar case. Particularly, most of our
results rely on the projections of various kinds of row and column stabilizers
being transitive or primitive. For similar DFAs, the projections of the full row
and column stabilizers Cπ and Rπ ′ are both trivial. Hence there is no guarantee
that other types of stabilizers such as Rq π ′ or Cp′ ,q′ π have useful properties, or
even that they are non-trivial, and so we cannot necessarily use these groups to
our advantage.
42
On the other hand, similarity imposes the very strong condition that the
groups G, G′ and G× are all isomorphic. It may be possible to exploit this to
prove some interesting things in the similar case.
It is not difficult to prove that if A and A′ are isomorphic as DFAs, then G×
is necessarily intransitive, so the case of isomorphic similar DFAs is uninteresting
for our purposes. We give two examples demonstrating what can happen with
non-isomorphic similar DFAs.
Example 21. Recall that in Example 16, we constructed two similar DFAs that
are of different sizes (and hence are non-isomorphic) and have primitive transition groups.
– A has states {1, . . . , 5} and transformations a = (3, 4), b = (1, 2, 3)(4, 5).
– A′ has states {1, . . . , 10} and transformations a′ = (1, 2)(6, 8)(7, 9) and b′ =
(2, 6, 4, 5, 3, 7)(8, 10, 9).
We have verified computationally that A × A′ has an intransitive transition
group, and so is not accessible. Thus even if non-isomorphic similar DFAs have
primitive transition groups, their direct product might have an intransitive transition group (compare this with Corollary 3 for dissimilar DFAs).
Note that in Example 16, we also showed that by simply swapping the roles
of a′ and b′ in A′ , the two DFAs A and A′ become dissimilar. Furthermore, A
has the symmetric group S5 as its transition group. Thus by Corollary 3, the
two DFAs actually become uniformly boolean minimal if we swap the roles of a′
and b′ .
Example 22. There is a primitive subgroup of S6 that is isomorphic to S5 . Using
GAP, we can obtain an explicit isomorphism between S5 and this subgroup, just
as we did in Example 16.
gap> IsomorphismGroups(SymmetricGroup(5),PrimitiveGroup(6,2));
[ (3,4), (1,2,3)(4,5) ] -> [ (1,2)(3,4)(5,6), (1,2,3,5,4,6) ]
We then use this isomorphism to construct similar DFAs:
– A has states {1, . . . , 5} and transformations a = (3, 4) and b = (1, 2, 3)(4, 5).
– A′ has states {1, . . . , 6} and transformations a′ = (1, 2)(3, 4)(5, 6) and b′ =
(1, 2, 3, 4, 5, 6).
Unlike the DFAs of Example 21, here we verified computationally that A ×
A′ actually has a transitive transition group. Hence a direct product of nonisomorphic similar DFAs with primitive transition groups may or may not be
accessible.
Note that A × A′ is not uniformly boolean minimal. For example, we have
verified computationally that (A×A′ )(X) is not minimal for X = {1}×{1, 3, 5}.
We have not found an example of two similar DFAs that are uniformly boolean
minimal, but we also have not proved that no such example exists.
43
5
Conclusion
We summarize the major results proved in Section 4.
Theorem 1 gives necessary and sufficient conditions for a pair of regular
languages (L, L′ ) to have maximal boolean complexity, with the requirement
that these languages are recognized by permutation DFAs A and A′ with exactly
one final state. In this special case, it turns out that (L, L′ ) is uniformly boolean
minimal if and only if A × A′ is accessible. This gives a partial characterization
of witnesses for the state complexity of proper binary boolean operations.
We have several results which may help to determine whether A × A′ is
accessible. Lemma 2 gives group-theoretic conditions for accessibility, while
Proposition 6 gives a useful graph-theoretic condition.
Theorem 2 gives a particularly useful group-theoretic condition for accessibility of A × A′ , as well as a similar condition for uniform boolean minimality.
The power of this theorem is demonstrated by Corollary 3, which gives several
classes of groups where the condition of Theorem 2 holds. If one can show that
the transition group A or of A′ lies in one of these classes, one immediately
gets useful information about A × A′ . Corollary 3 is also useful for constructing
examples of DFAs whose direct product is accessible or uniformly boolean minimal: one may pick a pair of groups from the classes mentioned in the corollary,
and use them as the transition groups of a pair of DFAs.
Lemma 3 gives some sufficient conditions for uniform boolean minimality
of permutation DFAs. The conditions are stronger than those of Theorem 2,
but more difficult to verify. Unfortunately, Example 20 shows that these conditions are not necessary. Necessary and sufficient conditions for uniform boolean
minimality are still unknown.
Proposition 7 gives a necessary and sufficient condition (1) for a property
that is slightly weaker than uniform boolean minimality to hold (in permutation
DFAs). Specifically, condition (1) of Proposition 7 does not guarantee minimality
for final state sets corresponding to the symmetric difference operation or its
complement.
Unfortunately, Example 19 shows that condition (1) of Proposition 7 is not
sufficient for uniform boolean minimality. This means a precise characterization of uniform boolean minimality lies strictly between the conditions given by
Lemma 3 and Proposition 7.
We state some unsolved problems and potential new directions of research
arising from our work in this paper.
– Find necessary and sufficient conditions for A × A′ to be uniform boolean
minimal. Even in the special case of permutation DFAs, we were unable to
resolve this.
– Find more classes of groups where the hypotheses of Theorem 2 hold, thus
extending the reach of Corollary 3.
– Find an example of two similar DFAs that are uniformly boolean minimal, or
prove that no such DFAs exist. Even pairs of similar DFAs with an accessible
direct product seem to be rare; one such pair is given in Example 22.
44
– Prove anything interesting about accessibility and/or uniform boolean minimality of direct products of non-isomorphic similar DFAs.
– Investigate uniform boolean minimality with respect to “unrestricted state
complexity” (see the last two paragraphs of Section 2).
– Investigate uniform boolean minimality for DFAs that do not contain “interesting” subgroups of permutations. That is, let us say the transition group
of a DFA is “interesting” if it satisfies the hypotheses of any of our major
results. Although we stated our results exclusively for permutation DFAs,
they hold more generally for DFAs whose transition monoids contain an “interesting” subgroup of permutations. However, many DFAs have transition
monoids which do not contain “interesting” subgroups; for example, DFAs of
star-free languages have transition monoids with no non-trivial subgroups.
In these cases, what can we say about uniform boolean minimality, or even
accessibility of direct products?
– Look for necessary and/or sufficient conditions characterizing state complexity witnesses for operations on regular languages other than boolean
operations (e.g., concatenation, star, reverse). As these problems could be
extremely difficult, it may be useful to start with the special case of group
languages or some other subclass of the regular languages.
Acknowledgements
I thank Jason Bell and Janusz Brzozowski for careful proofreading and helpful
comments. The computer algebra system GAP [14] was invaluable for this research; I cannot overstate its importance in obtaining these results. In particular,
I thank the authors of the Automata GAP package [11] and all contributors to
GAP’s library of primitive groups.
Funding: This work was supported by the Natural Sciences and Engineering
Research Council of Canada under grant No. OGP0000871.
References
1. Almeida, J., Rodaro, E.: Semisimple synchronizing automata and the WedderburnArtin theory. Int. J. Found. Comput. Sc. 27(02), 127–145 (2016)
2. Araújo, J., Cameron, P.J., Steinberg, B.: Between primitive and 2-transitive: synchronization and its friends. https://arxiv.org/abs/1511.03184 (Nov 2015)
3. Bell, J., Brzozowski, J., Moreira, N., Reis, R.: Symmetric groups and quotient
complexity of boolean operations. In: Esparza, J., et al. (eds.) ICALP 2014. LNCS,
vol. 8573, pp. 1–12. Springer (2014)
4. Bosma, W., Cannon, J., Playoust, C.: The Magma algebra system I: The user
language. J. Symbolic Comput. 24(3–4), 235–265 (1997)
5. Brzozowski, J.: Quotient complexity of regular languages. J. Autom. Lang. Comb.
15(1/2), 71–89 (2010)
6. Brzozowski, J.: In search of most complex regular languages. Int. J. Found. Comput. Sc. 24(06), 691–708 (2013)
45
7. Brzozowski, J.: Unrestricted state complexity of binary operations on regular languages. In: Câmpeanu, C., Manea, F., Shallit, J. (eds.) DCFS 2016. LNCS, vol.
9777, pp. 60–72. Springer (2016)
8. Brzozowski, J., Jirásková, G., Li, B.: Quotient complexity of ideal languages. Theoret. Comput. Sci. 470, 36–52 (2013)
9. Brzozowski, J., Liu, B.: Quotient complexity of star-free languages. Int. J. Found.
Comput. Sc. 23(06), 1261–1276 (2012)
10. Cameron, P.J.: Finite permutation groups and finite simple groups. Bull. London
Math. Soc 13(1), 1–22 (1981)
11. Delgado, M., Linton, S., Morais, J.: Automata – a GAP package, Version 1.13.
http://cmup.fc.up.pt/cmup/mdelgado/automata/ (19 November 2011)
12. Dixon, J.D., Mortimer, B.: Permutation groups. Springer (1996)
13. Ésik, Z., Gao, Y., Liu, G., Yu, S.: Estimation of state complexity of combined
operations. Theoret. Comput. Sci. 410(35), 3272–3280 (2009)
14. The GAP Group: GAP – Groups, Algorithms, and Programming, Version 4.8.6
(2016), http://www.gap-system.org
15. Maslov, A.N.: Estimates of the number of states of finite automata. Dokl. Akad.
Nauk SSSR 194, 1266–1268 (Russian) (1970), English translation: Soviet Math.
Dokl. 11(1970) 1373–1375
16. Restivo, A., Vaglica, R.: Extremal minimality conditions on automata. Theor.
Comput. Sci. 440–441, 73–84 (2012)
17. Restivo, A., Vaglica, R.: A graph theoretic approach to automata minimality. Theoret. Comput. Sci. 429, 282–291 (2012)
18. Salomaa, A., Salomaa, K., Yu, S.: State complexity of combined operations. Theoret. Comput. Sci. 383(2), 140–152 (2007)
19. Salomaa, A., Wood, D., Yu, S.: On the state complexity of reversals of regular
languages. Theoret. Comput. Sci. 320(2), 315–329 (2004)
20. Steinberg, B.: A theory of transformation monoids: combinatorics and representation theory. https://arxiv.org/abs/1004.2982 (Apr 2010)
21. Yu, S., Zhuang, Q., Salomaa, K.: The state complexities of some basic operations
on regular languages. Theor. Comput. Sci. 125(2), 315–328 (1994)
46
| 4math.GR
|
Taming Context-Sensitive Languages
with Principled Stateful Parsing
Nicolas Laurent
Kim Mens
arXiv:1609.05365v1 [cs.PL] 17 Sep 2016
Université catholique de Louvain, ICTEAM
{nicolas.laurent, kim.mens}@uclouvain.be
Abstract
1.
Historically, true context-sensitive parsing has seldom been
applied to programming languages, due to its inherent
complexity. However, many mainstream programming and
markup languages (C, Haskell, Python, XML, and more)
possess context-sensitive features. These features are traditionally handled with ad-hoc code (e.g., custom lexers),
outside of the scope of parsing theory.
Current grammar formalisms struggle to express contextsensitive features. Most solutions lack context transparency:
they make grammars hard to write, maintain and compose
by hardwiring context through the entire grammar. Instead,
we approach context-sensitive parsing through the idea that
parsers may recall previously matched input (or data derived
therefrom) in order to make parsing decisions. We make use
of mutable parse state to enable this form of recall.
We introduce principled stateful parsing as a new transactional discipline that makes state changes transparent to
parsing mechanisms such as backtracking and memoization.
To enforce this discipline, users specify parsers using formally specified primitive state manipulation operations.
Our solution is available as a parsing library named Autumn. We illustrate our solution by implementing some practical context-sensitive grammar features such as significant
whitespace handling and namespace classification.
In this section, we review the notion of context-sensitive
parsing (1.1), describe how we tackle it through recall and
stateful parsing (1.2), and situate our implementation (1.3).
In what follows, we use the term parser to refer to any
unit of functionality that can match some input text (both at
the the lexical and grammatical levels) and produce a result
based on this match. We assume that simple parsers can be
combined into increasingly complex parsers.
Categories and Subject Descriptors
Languages]: Processors—Parsing
1.1
Introduction
Context-Sensitive Parsing
True context-sensitive parsing is seldom applied to programming languages. Writing context-sensitive grammars can
prove challenging, for instance the grammar for the language
an bn cn is notoriously tricky [10]. In addition, most language
features can be adequately expressed using the much more
tractable context-free grammars. Nevertheless, many mainstream languages exhibit context-sensitive features. Here are
a few examples:
• In C, in order to determine whether the statement x*y; is
the product of x by y, or rather the declaration of a variable
y which is a pointer to type x, one must analyze the type
definitions preceding the statement.
• In Haskell and Standard ML, programmers can introduce
operators with custom precedence and associativity. The
parser needs to interpret these definitions in order to be
able to parse the remainder of the input.
D.3.4 [Programming
Keywords stateful parsing, grammars, context sensitivity,
data dependence, parsing expressions
• Since Python has significant indentation, a Python parser
Nicolas Laurent is a research fellow of the Belgian fund for scientific
research (F.R.S.-FNRS).
• In XML, opening tags must be matched with corresponding closing tags. For instance, <foo></foo> is valid while
<foo></bar> is not. As such, an XML parser must memo-
needs to detect when the indentation level increases or
decreases.
rize the names of open tags, at arbitrary levels of nesting.
• Many network protocols, including TCP, make use of
length-delimited fields whose length is not known in
advance but indicated by a length field that precedes them.
Most parsing tools cannot adequately handle these syntactic peculiarities, leading to all sorts of hacks, and sometimes
[Copyright notice will appear here once ’preprint’ option is removed.]
Taming Context-Sensitive Languages with Principled Stateful Parsing
1
2016/9/20
2.
to the rejection of parsing tools altogether. There are a few
exceptions however, which we review in Section 2. Section 3
deals with the key properties almost all these solutions lack,
namely context transparency, hence calling for a new solution.
1.2
As the problem of context-sensitive features in programming
languages is not new, it is not surprising that several solutions
have been proposed. We review these solutions in order to
better put our contributions in perspective. We do not purport
to review the entire body of work on context-sensitive parsing,
but only the approaches closest to our goal. In particular, we
left out the literature on context-sensitive lexical analysis
(e.g., [3, 30]) which by definition only handles a small subset
of all context sensitivity issues.
Recall and Stateful Parsing
While context sensitivity was first characterized by Chomsky [5], his Context-Sensitive Grammars (CSG) are only of
little help, due to the intricate coding that they require1 . A
CSG is made of rewrite rules X → Y where X and Y are
strings of mixed terminals and non-terminals. These rules
must be non-contracting: Y, as a string of symbols, must not
be shorter than X.2 As a matter of fact, these grammars were
never meant to describe programming languages, but natural
languages, where the shape of the rules make much more
sense. In particular, it is difficult to encode recall constraints:
for instance, requiring the same string of tokens to appear at
two different locations in the sentence (assuming the string is
not fixed in advance).
We propose instead to approach context sensitivity through
the notion of recall, i.e. the ability to accept sentences based
on relationships between some of their parts. This is more
easily understood in parsing terms as the capability to make
parsing decisions based on previously matched input.
We enable recall by allowing users to write parsers which
can manipulate mutable state. However, unlike solutions
deployed by existing parsing tools such as ANTLR [24]
and Rats! [9], we are principled about state use. Indeed,
parsing algorithms do not proceed linearly. When faced
with a choice, they may speculatively try an alternative, and
need to backtrack if this alternative does not succeed. When
backtracking happens, all changes made to the state during the
speculative execution need to be reversed. Parsers may also
memoize the result of a speculative execution. In a stateful
model, these results need to include the state changes incurred
by the execution. As will be explained in Section 4, we satisfy
these requirements by introducing primitive operations to
manipulate mutable state in a principled way.
2.1
Implementation
Data-dependent grammars
Jim et al. [15] proposed data-dependent grammars, a formalism which permits context sensitivity by allowing rules to
be parameterized by semantic values. A parameterized nonterminal appearing on the right-hand side of a rule acts as a
form of function call that also returns a semantic value. These
semantic values are computed by semantic actions written
in a general-purpose programming language. There are also
semantic predicates which can make grammar branches succeed or fail depending on a semantic value.
Data-dependent grammars can be compiled to a format
accepted by a target parsing tool, which must support fairly
general semantic actions. In subsequent work [14], the authors introduced a new kind of automaton that can be used
We implement our stateful parsing approach as a PEGlike [8] (top-down recursive descent) functional-style 3 parsercombinator library named Autumn, which can be used with
Java, Kotlin, and other Java-compatible languages. Implementation details are given in Section 5 and a detailed use
case is worked out in Section 6.
1 The
same holds for a large body of work on mildly context-sensitive
grammars. [16]
2 In reality, CSG rules are not required to be non-contracting, but noncontracting grammars and CSG describe the same set of languages. [6]
3 Parsers are merely functions. Moreover, custom parsers can be defined.
Taming Context-Sensitive Languages with Principled Stateful Parsing
Backtracking Semantic Actions
Parsing with backtracking semantic actions [29] is an approach that extends a (general) backtracking LR parser with
reversible semantic actions. Upon backtracking, state changes
are reversed. Two important restrictions apply: state changes
can only occur during term reduction, and the state can only
affect the parse through semantic conditions that trigger backtracking.
Compared to our approach, the difference in implementation (top-down recursive versus LR) has far-reaching consequences. Backtracking semantic actions reverse state changes
rather than making snapshots of the state. Accordingly it
becomes impossible to compare state snapshots. These capabilities are very useful in the context of PEG parsing, as they
enable the definition of custom parsers, left-recursion and
memoization; but less so within an LR system where custom
parsers cannot be defined.
We also believe our top-down recursive-descent model
to be more intuitive in the presence of state. Logically, state
changes occur as parser are invoked, from left to right and
from top to bottom. Backtracking semantic actions, on the
other hand, are executed upon term reduction. This means
that a parser may modify the state before other parsers that
matched input on its left.
Despite these caveats, we consider parsing with backtracking semantic actions [29] to be the safest and most convenient
system for context-sensitive parsing among those presented
in this section.
2.2
1.3
Related Work
2
2016/9/20
to implement parsers recognizing data-dependent grammars.
These techniques are put to work in a tool called Yakker.
Data-dependent grammars, though theoretically compelling, suffer from usability issues. The value-passing model
means that the parse state needs to be threaded throughout
the grammar. Making a rule dependent on a new semantic
value means that all rules through which this rule is reachable might need to be modified to pass this value around.
Maintainability-wise, this is far from ideal. Moreover, it
harms composability, as a rule must be aware of all states it
has to pass through.
Afroozeh and Izmaylova [1] show how advanced parser
features such as lexical disambiguation filters, operator precedence, significant indentation and conditional preprocessor directives can be translated to data-dependent grammars. Quite
clearly, the task is non-trivial and one comes away with the
feeling that dependent grammars are better suited as an elegant calculus to be targeted by parsing tool writers rather than
as a paradigm that fits the needs of tool users. The machinery
implementing the formalism is also distinctively non-trivial,
involving a multi-stage transformation into a continuation
routine or into a new kind of automaton. In contrast, our approach consists of a lightweight library that can be layered
on top of a general-purpose programming language.
Finally, we note that the much older Definite Clause
Grammars (DCGs) [25] formalism works on almost exactly
the same principle, but building upon logic programming.
Accordingly, it suffers from similar limitations.
2.3
tors generate grammar fragments at parse-time (much like a
monadic parser generates a new parser), hence the term active. While monadic parsing seems at first sight very similar
to the data-dependent grammars from Section 2.2, Atkey [2]
carefully contrasts the two approaches:
We characterise their [Jim et al.] approach as refining
context-free grammars: each Yakker grammar has an
underlying context-free grammar with regular righthand sides, and the constraints allow for sophisticated
data-dependent filtering of parses. In contrast, we
consider active right-hand sides that generate the
grammar as the input is read.
Nevertheless, monadic parsers suffer from the same pitfalls as data-dependent grammars: the state is threaded
through the grammar (or code), leading to poor maintainability and composability.
2.4
Attribute grammars [18] associate attributes to AST nodes
(assuming an AST node per matched grammar rule). The
attributes can be synthesized: their value derived from the
attributes of subnodes, or inherited: their value computed
by a parent node. The formalism supports context-sensitive
parsing through production guards predicated over attributes.
However, attribute grammars are not context-transparent.
To enable recall, they need to propagate the recalled value
from the definition site to the use site, through a chain of of
synthesized and inherited attributes. Even reference attributed
grammars [11], which allow attributes to contain references
to nodes, do not fully solve this distribution problem.
Monadic parsers
Monadic parsing [12] is a well-known way to build functionalstyle parser-combinator libraries, made popular by Haskell
libraries such as Parsec [21]. In this paradigm, the type of
a parser is a function parameterized by a result type, i.e.
with signature string → (string, result), where the parameter
string is the input text and the returned string is the input
remaining after parsing. The parser type is also a monad
instance, meaning there is a bind function whose signature,
in Haskell notation, is:
2.5
Stateful Parsing
Manipulating parse-wide state can be an effective solution
to the problem of data dependence: the data depended upon
can be written in the state when encountered and read or even
altered later on.
Broadly speaking, we can distinguish two big classes of
stateful parsing tools. First, there are parser combinator libraries that allow users to write their own subparsers. Notable
examples include Parboiled [7], Lua Peg [13] and Scala’s
parser combinators [22]. Since these custom parsers are implemented in a general-purpose programming language, they
can manipulate state, even though the libraries make no provision for this. Second, there are parsing tools that provide
very general semantic actions and semantic predicates. Notable examples include Bison [28] and ANTLR [24]. These
work much like their counterpart in Yakker (cf. Section 2.2),
except that instead of returning a value, semantic actions may
modify a global state object.
Unfortunately, most parsing tools in both categories do not
make the necessary provisions for dealing with backtracking
and memoization: if the parser backtracks over a construct
that made state changes (semantic action or custom parser),
these changes need to be undone; if the parser can memoize
Parser r1 -> (r1 -> Parser r2) -> Parser r2
where r1 and r2 are result types. This function takes a parser
as first parameter, and a function which transforms the result
of the parse into another parser as second parameter. When
invoked, the parser returned by bind will invoke the first
parser, pass its result (of type r1) to the function, then invoke
the parser this function returns, yielding a result of type r2.
The important point about monadic parsers is that they can
handle context sensitivity. Indeed, the second parameter to
bind (the function) returns a parser from a result. This means
that the behaviour of the parser returned by bind depends on
data acquired during the parse: this is a form of recall.
An in-depth analysis of this aspect was done by Atkey [2].
In particular, he formalizes monadic parsers by introducing
active right-hand sides, which are the right-hand sides of
rules that can contain monadic combinators. These combinaTaming Context-Sensitive Languages with Principled Stateful Parsing
Attribute Grammars
3
2016/9/20
the result of a construct, state changes need to be memoized
as well. In the absence of such guarantees, a construct can
only access state which it is sure has not been corrupted by
changes that should have been discarded. It must also be
sure that some state-altering construct was not skipped due to
memoization. These are tricky propositions to verify even for
medium-sized grammars, and every change to the grammar
threatens to falsify them.
One may think that solving the backtracking problem
is simply a matter of inserting a construct that reverses
state changes whenever a rule fails. However, a rule can
be backtracked over even if it succeeded. It suffices that one
of the rules through which our rule was reached fails. Hence
this scheme would entail, for each state-altering construct, the
modification of every rule through which it can be reached.
2.6
A grammatical construct is context-transparent if it
is unaware of the context shared between its ancestors
and its descendants.
Data-dependent grammars, monadic parsers, DCGs and
attribute grammars are not context-transparent because of the
need to explicitly pass values around. For instance, consider
two data-dependent grammars4 : a grammar for a Pythonlike language with significant indentation, in which the
rules for block-level constructs (statements, definitions) are
paremeterized by the indentation level; and a grammar for a
generic macro definition language (e.g., GNU M4). We want
to compose these two languages such that macro definitions
may appear anywhere where definitions can appear in our
Python-like language. Additionally, we want macro bodies to
include Python-like code.
The issue is that the rules in the macro language grammar
know nothing about indentation level, yet the indentation
level needs to be shared between the block holding the macro
definition and the Python-like code appearing inside macro
definitions. In this case, the lack of context transparency
would force us to rewrite all rules in the macro language
grammar to carry around the indentation level.
Stateful parsers also are not context-transparent, as they
must ensure that no unforeseen backtracking or memoization
takes place. For instance, if a parser a manipulates the state
and its callers do not expect it to backtrack, it cannot be
swapped for a parser c(a) (where c is some parser combinator)
without first ensuring that c(a) never backtracks over a.
Lack of context-transparency makes grammars hard to
reason about, hence hard to write and to maintain: refactoring,
extending or composing grammars becomes particularly
challenging, because each change to a rule might entail the
need to modify all rules through which it is (transitively)
reachable. In stateful parsers, such changes are liable to
introduce undesired backtracking or memoization.
We suggest a simple solution: use stateful parsing (which
does not thread context through the grammar), but undo state
changes upon backtracking and allow the memoization of
state changes. And to achieve this, we introduce a new context
sensitivity handling discipline: principled stateful parsing.
Rats!
Rats! [9] is a fully-memoizing (packrat) PEG parser. Rats! is,
to the best of our knowledge, the only stateful parsing tool
that provides some guarantees for state usage, by ensuring
that state changes are discarded if certain conditions are met.
For this purpose, Rats! introduces transactions that wrap
rules under which state changes might occur. A transaction
can either succeed, in which case its state changes are
retained, or fail, in which case the changes are discarded.
Rats! also requires that a nonterminal invoked at a given
position within a transaction must always modify the state in
the same way, no matter how that nonterminal was reached.
Combining transactions with this requirement ensures that
Rats! will never have to discard the memoization of a rule,
hence upholding the linear-time guarantee of packrat parsers.
In spite of its advantages, this scheme has two important
pitfalls. First, it requires nonterminal invocations at a given
position to always return the same result. This precludes
parsing expressions that modify the behaviour of the parsing
expression they invoke. However, this capability is valuable
in practice. For example, we use it to enable left-recursion
handling in our library (cf. Section 5.4).
Second, state changes are not memoized. If a rule succeeds
after applying a state change, but the enclosing transaction
fails, the changes are lost. If we wanted to call the rule at
the same position again, the memoized result would be used
and it does not include the state changes. This means that
a state change cannot safely be referenced by two different
transactions, and that transactions cannot be re-tried after a
state change higher up in the grammar hierarchy.
3.
4.
In Section 1, we established the relevance of context-sensitive
parsing and introduced the notion of recall as a way to
express context-sensitive features in terms of backreferences
to previously matched input. We enable recall by storing the
matched input (or data derived thereof) in a mutable data
store: the parse state. This section expounds how principled
stateful parsing is able to work with parse state while avoiding
the usual pitfalls of stateful parsing (cf. sections 2.5 and 3).
Context Transparency
As the previous section has shown, enabling the definition
of context-sensitive languages without jeopardizing maintainability, composability or even safety is no easy feat. We
put forward the notion of context transparency as the gold
standard that a context sensitive parsing mechanism needs to
meet in order to be considered sufficiently practical.
Taming Context-Sensitive Languages with Principled Stateful Parsing
Principled Stateful Parsing
4 The
same reasoning applies to monadic parsers, DCGs and attribute
grammars.
4
2016/9/20
4.1
Intuition
be made using the same notation. We also note that in Z, all
types used in the lambda calculus are sets.
Since we adopt the functional parser-combinator approach
(cf. Section 1), parsers are simply functions manipulating
parse state (Section 4.2.1) whose set-theoretic signature is
given in Section 4.2.2. Section 4.2.3 formally specifies the
primitive state-manipulation operations that were briefly introduced in Section 4.1. Finally, Section 4.2.4 gives the semantics of parser invocation by specifying the call operation,
which maps a parser (as defined in 4.2.2) to a single state
transformation.
Before diving into a formal explanation, we present the
remarkably simple intuition behind the approach.
The point of using state is to pass context around implicitly,
without the need to hardwire context in the grammar, hence
achieving context transparency (cf. Section 3).
If the execution of a parser were linear, simply reading/writing to this state would suffice. Unfortunately, parsers
must sometimes perform speculative executions that may fail
further down the line, a phenomenon called backtracking.
When backtracking occurs, all state changes in the speculative execution being backtracked over must be reversed.
Hence, we need an operation that can take a snapshot of the
state at a given point, and an operation that can restore the
state described by such a snapshot.
Given these requirements, it helps to think of the parse
state as a log of the operations applied to the state, which can
be snapshot and rolled back as required. Appropriately, this
is also how we formalize the parse state.
Additionally, it is sometimes desirable to save the result
of a speculative execution (whether it failed or not), i.e., the
state changes it induced: a delta acquired by performing a
diff between the states before and after the execution. It is
also necessary to be able to merge these changes back into
the state. The most straightforward application of the diff
and merge capabilities is the memoization of parse results.
However, other valuable use cases exist, such as longestmatch parsing and left-recursive parsing (see Section 5.4).
This motivates the need for four primitive state-manipulation
operations: snapshot, restore, diff and merge. These operations are described in section 4.2.3.
4.2.1
At the core of our approach lies the notion of parse state. The
parse state abstracts over a general mutable data store. We do
not place any constraint on the data within the store. This is
formalized as follows.
[CHANGE]
STATE = seq CHANGE
The square brackets introduce the abstract set CHANGE
of all state changes. What exactly constitutes a state change
(most likely the mutation of a memory location) is an implementation concern that is not relevant to the formalization.
STATE is the set of possible parse states: i.e., of possible
configurations of our mutable store. We represent a parse
state as a sequence of state changes. This means that a state
can be seen as a log of the operations over the mutable store
it represents, assuming some well-defined initial state.
In Z, the set of sequences of items from the set S is written
seq S and corresponds to the powerset of pairs (i, s) ∈ N × S,
or equivalently to the powerset of partial functions N
7 S. In
each sequence, the indices are unique and consecutive.
In practice, an implementation of the approach will want
to use parse state to reify important parsing notions, such
as input position. We consciously avoided making our formalism needlessly specific, hence the absence of some usual
parsing notions such as input position. This enables using our
approach to parse non-linear inputs (e.g., object graphs), or
perform computations that only bear nominal resemblance to
traditional parsing, even though this direction is outside the
scope of the current paper.
Principled stateful parsing is an approach where
parsers behave transactionally: each parser invocation either succeeds or leaves the state untouched. Additionally,
it is possible to generate and merge deltas corresponding
to state changes made by parser invocations. All this is
made possible through the use of formally specified state
manipulation operations.
4.2
Formalization
4.2.2
We formalize our approach using the Z notation [26], though
eschewing its schema calculus in favor of a purely functional presentation.5 The Z notation is a formal specification
language that builds on top of Zermelo-Frankel set theory,
first-order logic and simply typed lambda calculus. As such,
Z can be seen as a language where functions can be defined
in lambda calculus extended with predicates from first-order
logic and set theory. Formal assertions over the functions can
Parsers
A parser represents a computation over the parse state that
either succeeds or fails, and has side effects on the parse state,
in the form of state changes, as introduced in the previous
section.
TRANSFORM = STATE → STATE
PARSER = STATE → seq TRANSFORM
RESULT ::= success | failure
result : STATE → PARSER → RESULT
5 To
improve the presentation, we took some liberty with the Z layout (but
not with the notation). A machine-understandable version of the specification
is available online [20].
Taming Context-Sensitive Languages with Principled Stateful Parsing
Parse State
Formally, a parser is a function from a state — the
current state at the time of invocation — to a sequence of
5
2016/9/20
Call Of these six, call has a special status: it represents
the invocation of a parser. We will define this operation in
section 4.2.4, hence specifying the semantics of parsers given
their execution trace. Note that the signature definition of call
expands to PARSER → STATE → STATE: a parser must be
called with a state as parameter.
transformations, which move from one state to another. This
amounts to defining a parser in terms of its execution trace.
Two things seem to be missing from this definition. First,
it does not say if the parse succeeds when run over a specific
state. This property is exposed separately through the result
predicate rather than as part of the PARSER signature. This
approach is not significant: it simply makes the math look
nicer. Second, the input being parsed does not explicitly
appear in the signature. Instead, the input is assumed to be
held within the parse state.6
A parser is a recognizer of states. It accepts states for
which result state = success holds. If within the input state
one dissociates the parse input from the rest of the state (the
context), one can see that the parser recognizes — hence also
defines — different languages depending on the context.
But a parser is also a transformer of states as well: when
invoked it performs a STATE → STATE transformation. In
section 4.2.4 we explain how to derive this transformation
from a parser (recall that parsers have type PARSER defined
as STATE → seq TRANSFORM), as a means of defining the
semantics of a parser given its execution trace. We could alternatively have defined PARSER as STATE → TRANSFORM
(with the result being the composition of the transformations
in the sequence), or directly as STATE → STATE. We chose
to emphasize the execution trace — a sequence of transformations — instead, because the primitive state operations
described in the next section are suppliers of such transformations, to be composed to yield the transformation performed
by the parser.
This representation also emphasizes that the parse state
is both an input of the parser and an input of the returned
transformations. This reflects the fact that a parser is contextsensitive: it chooses which operation to perform depending
on the state. This is closely related to the notions of active
right-hand sides [2] and monadic parsing [12]. In fact, each
operation in the sequence is chosen depending on the state
obtained by running the initial state through the composition
of all preceding transformations. Abstracting over this makes
the specification much simpler, without altering its meaning.
4.2.3
Snapshot A snapshot, as the name implies, is a capture of
the state at a specific point during the execution. Naturally,
this makes SNAPSHOT, the set of all snapshots, equivalent
to STATE. Formally, the snapshot operation, which creates
such a capture, is simply the identity function.
snapshot = λ x : STATE • x
Diff The diff operation returns a DELTA object representing
the difference between a snapshot and the current state, as a
set of state changes. As a precondition, this operation requires
the snapshot it receives to be a prefix of the current state.
This is expressed with the Z built-in prefix infix operator. By
keeping the deltas append-only, we ensure that a delta can be
later merged to any state, not just the one corresponding to
the snapshot.
∀ sn : dom diff • ∀ st : dom (diff sn) •
sn prefix st
Since deltas are state suffixes, DELTA, the state of all
deltas, is equivalent to STATE.
Assuming the precondition is respected, diff can be defined as the remainder of the current state after chopping off
the prefix corresponding to the snapshot. In Z, the squash
function packs the indices (left-hand side) of a set of pairs
in N × S, where S is some set, in order to turn this set into
a proper sequence. For instance, it turns {(2, x), (5, y)} to
{(1, x), (2, y)}.
diff = λ sn : SNAPSHOT • λ st : STATE •
squash (st \ sn)
Transformations All operations except diff and snapshot
return a transformation. Recall that we defined PARSER as
STATE → seq TRANSFORM. The transformations returned
by the operations are precisely those which will be part of
a parser’s execution trace. diff and snapshot are different
because they do not modify the parse state. Instead, diff and
snapshot create new objects, which can be freely passed
through the parse state.
Primitive Operations
We now present six primitive operations (amongst which the
four announced in Section 4.1) that parsers can perform.
SNAPSHOT = seq CHANGE
DELTA = seq CHANGE
ApplyChange The applyChange operation is very simple:
given a change, it simply returns a transformation that applies
this change, by appending it to the change log. It can be
defined as follows, using the concatenation operator (a) to
append the change to the old log.
call : PARSER → TRANSFORM
snapshot : STATE → STATE
diff : SNAPSHOT → STATE → DELTA
applyChange : CHANGE → TRANSFORM
restore : SNAPSHOT → TRANSFORM
merge : DELTA → TRANSFORM
applyChange = λ c : CHANGE • λ st : STATE •
st a hci
This “operation” models the fact that parsers can perform
arbitrary state changes.
6 Nothing
precludes the input from being mutable, even though we have not
investigated the usefulness of the idea.
Taming Context-Sensitive Languages with Principled Stateful Parsing
6
2016/9/20
Restore The restore operation takes a snapshot as input and
returns a transformation that brings the state to that described
by the snapshot.
encounter them. Our approach is not language-specific and
can easily be ported to other languages.
We start by exposing the fundamentals of the Autumn API
and how it relates to our formalization (Section 5.1). We then
show the API in action on a simple example (Section 5.3).
Finally we discuss how the API enables simple left-recursion
handling (Section 5.4).
restore = λ sn : SNAPSHOT • λ st : STATE • sn
Merge The merge operation takes a delta as input and
returns a transformation that appends this delta to the input
state.
5.1
In this section, we review how our implementation relates
to our formalization of principled stateful parsing (Section
4). Figure 1 shows the key interfaces and classes in our
implementation.
merge = λ d : DELTA • λ st : STATE • st a d
4.2.4
Parser Invocation Semantics
We now look at how the transformation returned by the call
operation can be derived from the execution trace returned
by a parser. Recall that the call operation’s signature is
PARSER → TRANSFORM.
We start by defining two helper functions. composeTwo
maps sequences of transformations of length n ≥ 2 to a
sequence of length n − 1 similar to the input sequence,
but where the first two items have been replaced by their
composition (s 1 and s 2 access the first two items of s while
# is the relational composition operator). reduceN takes a
natural n and a sequence of transformations and returns the
composition of its n first items, or the identity transformation
if n = 0. This is achieved by iteratively running the sequence
through composeTwo, using the Z built-in iter operator.
interface Parser {
fun parse ( ctx : Context ): Result
}
sealed class Result {
object Success : Result ()
open class Failure ( val pos : Int , val msg : String )
: Result ()
}
class Context ( input : String ,
vararg states : State <* , * >) {
var pos : Int = 0
val text : String = input + ’\ u0000 ’
fun <T : State <* ,* > > state ( klass : Class <T >): T { ... }
fun
fun
fun
fun
composeTwo = λ s : seq TRANSFORM •
hs 1 # s 2i a tail (tail s)
reduceN = λ n : N • λ s : seq TRANSFORM •
if (n = 0) then id STATE
else iter (n − 1) composeTwo s 1
snapshot (): Snapshot { ... }
restore ( snap : Snapshot ) { ... }
diff ( snap : Snapshot ) { ... }
merge ( delta : Delta ) { ... }
...
}
class Snapshot { ... }
class Delta { ... }
interface State < Snapshot , Delta > {
fun snapshot (): Snapshot
fun restore ( snap : Snapshot )
fun diff ( snap : Snapshot ): Delta
fun merge ( delta : Delta )
}
With this in place, we define the result of call as the
composition of all transformations within the call’s execution
trace, assuming the parser invocation is successful. Otherwise,
the identity transformation is returned. The hash sign (#) is
an operator returning the cardinality of a set.
abstract class Grammar {
open val whitespace : Parser
= ZeroMore ( CharPred ( Char :: isWhitespace ))
open val root : Parser
open val requiredStates : List < State <* ,* > = emptyList ()
...
}
call = λ p : PARSER • λ st : STATE •
if (result st p = success)
then reduceN (# p st) (p st) st
else st
5.
The Autumn API
Figure 1. Key interfaces and classes in the implementation
of Autumn.
Implementation
Parser We represent a parser by an instance of the Parser
interface. Implementers must override the parse method,
which takes a Context as parameter and returns a Result:
either a Success or a Failure which holds the position
at which the failure occurred, together with a diagnostic
message. It can also hold custom diagnostic information
through subclassing of Failure. This gives a lot of control
over the error messages that will be shown to the user.
We implemented the principled stateful parsing approach in
a general-purpose parsing library called Autumn. It is freely
available online [20]. Autumn is implemented in Kotlin, an
up-and-coming JVM language that closely matches Java’s
semantics while reducing boilerplate. Kotlin possesses many
features that make it particularly well suited for writing
domain-specific languages (DSLs), an ability we exploit to
define grammars. We will introduce these features as we
Taming Context-Sensitive Languages with Principled Stateful Parsing
7
2016/9/20
Context Each parse — the invocation of a parser on a
complete piece of input text — has an associated Context
object. The role of this object is to hold the state for the
parse. The context is passed down to parsers during parser
invocation, so that all parsers may access it. Using a context
object rather than global state allows multiple parses to coexist, potentially in parallel.
The mutable store mentioned in the formalization is represented as a collection of singleton classes implementing
the State interface. Parsers can retrieve a state instance by
calling the state method with the proper class object. Note
that the state held by the context also includes the input text
and the input position, although, as a special provision, these
can be accessed directly through the text and pos properties
respectively.
the stack and can be grafted back at a later time, allowing
for granular change handling.
• MapState: represents a map as an immutable Hash Array
Map Trie (HAMT) [4]. Our implementation is based on
that of Steindorfer and Vinju [27] which ensures good
performance. The map is treated as a unit.
• InertState: represents state that does not change during
the parse, or whose change is not significant (e.g., logging
logic). All operations are implemented as no-ops.
Grammar Programmers must subclass the Grammar class in
order to define a new grammar. When doing so, they must
define the root parser by overriding root(), and provide any
required State instances by overriding requiredStates().
Grammar also defines a default whitespace parser which
consumes any number of characters matching the Java
Char::isWhitespace predicate.
Parser combinator libraries traditionally struggle with
the definition of recursive parsers: it is forbidden to write
[val A = Seq(..., A)] because A is not defined yet when
it is evaluated on the right-hand side. Using the Grammar
initialization logic, we can replace the recursive reference
to A with the [!"A"] operator-overloading syntax: [val A =
Seq(..., !"A")]. This creates a stub parser which will be
patched with a reference to A at parse-time. To achieve this,
the Grammar class maps names to parsers through reflection
over its Parser-valued fields.
State The State interface has four methods: snapshot,
restore, diff and merge, corresponding to the four key operations introduced in Section 4.1, but only locally for the State
instance itself. To get the parse-wide semantics of Section
4, we aggregate the state operations over all State instances.
This is achieved through the methods in the Context class
that mimic the State interface. These methods manipulate
the Snapshot and Delta classes (not to be confused with
the eponymous type parameters of interface State), which
aggregate State-level snapshots and deltas.
In our formalization, we represented the mutable store as
a log of all operations over the store. In practice, this might
not be a good idea, as parsers must compute using the state,
meaning the “current state” would need to be re-derived from
the whole log at least every time backtracking happens, unless
all operations were fully reversible.
The approach we took instead was to give maximum
implementation flexibility to the programmer: he can choose,
for each State instance, the most appropriate strategy to
create snapshots and deltas, as well as to restore/merge them
back in. Maintaining a log of reversible operations is only
one possibility among many.
However, having the programmer implement the State
interface for each data structure he wishes to manipulate
would be tedious and repetitive. Hence, we supply base
implementations for common use cases, such as:
5.2
Autumn enforces the principled stateful parsing guarantees,
but only if its interfaces are implemented correctly. In particular:
• Each implementation of Parser must either succeed or
undo all the state changes it incurred. This is usually
achieved through the use of snapshot() and restore().
• Each implementation of State must implement its opera-
tions according to the specification given in Section 4.
5.3
Example
As an illustration, Figure 2 shows the implementation of one
of the most fundamental parser combinators, the sequential
composition of parsers. The resulting parser calls its subparsers sequentially, succeeding if they all succeed. If one of
them fails, the parser reverts the state to its initial condition.
• CopyState: for states that are records containing just a few
fields, which we can afford to copy every time, and which
can be treated as a unit (i.e., a delta cannot represent that a
field changed while the others retained their previous value
— the value of every field is systematically captured).
5.4
Left-Recursion
When a parser invokes itself (either directly or indirectly
through intermediate parsers) without intervening state
changes, the result is an infinite loop of parser invocations.
This is a well-known problem of top-down recursive parsers,
called left-recursion. Fortunately, it can be mitigated as follows: start by running the left-recursive parser while failing
all left-recursive invocations, then re-run it, using the result of
the initial parse as the result of all left-recursive invocations.
• StackState: represents a stack as a singly linked list
(which is naturally immutable). Snapshots and deltas are
represented as nodes in the list. The list is treated as a unit.
• MonotonicStack: Similar to StackState, but adds the restriction that diff must only be called with a snapshot that
is a suffix of the current stack. Deltas are then prefixes of
Taming Context-Sensitive Languages with Principled Stateful Parsing
Contract
8
2016/9/20
• Namespace classification — The parser needs to know
class Seq ( vararg children : Parser ): Parser (* children ) {
override fun _parse_ ( ctx : Context ): Result {
val snapshot = ctx . snapshot ()
for ( child in children ) {
val r = child . parse ( ctx )
if (r is Failure ) {
ctx . restore ( snapshot )
return ctx . failure
}
}
return Success
}
}
which identifiers refer to types. In our language, this is
needed because there is an ambiguity between the syntax
of function calls — which can receive a closure parameter
as an indented block — and the syntax of anonymous
classes:
val a = myFunction ()
myFunction2 ()
val b = MyClass ()
var x : Int
fun foo () { ... }
Figure 2. Implementation of the sequential parser combinator in Autumn.
The body of a class only admits declarations, while
the block part of a function call admits all statements
(including declarations). These two constructs result in
different nodes being added to the abstract syntax tree
(AST) produced by the parse.
Repeat until as much input as possible has been matched.
Refer to our earlier paper [19] for more details.
Interestingly, this strategy can be implemented entirely
within the stateful parsing paradigm. In particular, when
we speak of result of a parse, we are really referring to a
delta of the parse state. These deltas need to be stored in
the parse state, so that they can be retrieved by left-recursive
invocations. We also need to track, within the state, which leftrecursive parsers have been invoked at which input position,
so that we may recognize left-recursive invocations.7
Recognizing left-recursive invocations is the task of a dedicated parser that must be wrapped around all left-recursive
parsers. We do so by annotating recursive parsers with the !
operator. Since left-recursion requires recursive references
(cf. Section 5.1), we can easily check that all such parsers
have been properly annotated: a missing annotation will result in an unresolved reference instead of a mystifying infinite
loop.
6.
6.1
We now explain how Examply handles significant whitespace.
The code enabling this feature is shown in Figure 3.
The usual whitespace handling strategy is to assume
that every parser consumes its trailing whitespace through
invocation of the whitespace parser. To do so, it is only
necessary to ensure that all “primitive” parsers (fullfilling
what is traditionally the role of lexical analysis: matching
identifiers, literals, keywords and operators) consume their
trailing whitespace: then all parsers will do so by transitivity.
The Grammar class provides some support for this, including
lexical analysis emulation (not demonstrated in this paper),
and the [+"lit"] syntax which evaluates to a parser matching
a literal string and any trailing whitespace. Ultimately, this
leads to an important guarantee: all parsers are invoked at an
input position where no leading whitespace is present.
In order to handle significant whitespace, we maintain two
data structures. The first one, IndentMap, maps line numbers
to two quantities: the input position at which the indentation
ends on that line, and the indentation count, which is obtained
by expanding tabs to tab stops aligned to multiples of 4. The
second data structure, IndentStack, is a stack that stores the
indentation counts for all enclosing blocks.
Context.indent and Context.istack are extension properties for the Context class and provide syntactic sugar to access
the indentation count on the current line, and the indentation
stack, respectively.
We build IndentMap at the beginning of the parse, through
the invocation of the buildIndentMap parser. This does not
require any special tricks: each parser has access to the whole
input (ctx.text). The parser does not advance the input
position (ctx.pos), so that subsequent parsers are free to
proceed.
Within the grammar, we use the indent parser to require
an indented block, and the dedent parser to test for the end
of an indented block. The implementation of these parsers is
straightforward. indent checks if the current line is indented
with respect to the indentation of the current block (the top of
Use Case
In this section, we demonstrate the stateful parsing approach
with a realistic use case, using our Autumn library. We
implement a parser for a simple, yet non-trivial, staticallytyped, object-oriented programming language, which we call
Examply. This imaginary language draws inspiration from
Java (its main unit of definition is the class), Kotlin (its postfix
type notation and closure notation) and Python (significant
indentation). Its full grammar, written with Autumn, can be
found online [20]. Examply possesses two common contextsensitive features:
• Significant whitespace — Indentation is significant: Like
Python, the language does not use curly braces or keywords to delimit blocks of code such as loop or function bodies. Instead, these constructs expect to have their
body indented with respect to their first line. Similarly,
a decrease in indentation signifies the end of the block.
Newlines are also significant, as they are used to separate
successive statements and declarations.
7 We
could track the whole state instead of the input position, but we enforce
the stronger requirement that a parser invocation has to advance the input
position not to be considered left-recursive: unlike other state changes, an
increase in input position is a strong proof of progress.
Taming Context-Sensitive Languages with Principled Stateful Parsing
Significant Whitespace
9
2016/9/20
IndentStack, initialized to 0). If so, it succeeds after pushing
the new indentation count onto the stack. dedent checks to
new) to disambiguate constructor invocation from function
invocation. Or we could perform AST-disambiguation passes
after parsing. We stress that Examply is designed to showcase
the strength of the principled stateful approach, and notably
its ability to deal with the quirks of existing languages.
To understand the logistics of namespace classification,
we must first define how our imaginary language handles
type references:
see if the indentation count of the current line is less than
that of the current block, or if we have reached the end of the
input. If so it succeeds, after popping the previous count from
the stack.
Finally, the newline parser succeeds if and only if it is
invoked at the end of the indentation on the current line or at
the end of the input.
1. Types are always referenced through a single identifier,
except within imports where they are preceded by a
package string.
data class IndentEntry ( val count : Int , val end : Int )
class IndentMap : InertState < IndentMap > {
lateinit var map : Map < Int , IndentEntry >
fun get ( ctx : Context ): IndentEntry =
map [ ctx . lineMap . lineFromOffset ( ctx . pos )]!!
}
2. A class name can only be referenced after or within its
definition.
3. A class definition can appear anywhere other declarations
can: at the top-level, within another class, or within a code
block.
class IndentStack : StackState < Int >()
val Context . indent : IndentEntry
get () = state ( IndentMap :: class ). get ( this )
val Context . istack : IndentStack
get () = state ( IndentStack :: class )
4. Class definitions are lexically scoped: a class has access
to all imported classes and to all classes defined before it
within one of its outer scopes (class body or code block).
val buildIndentMap = Parser { ctx ->
val map = HashMap < Int , IndentEntry >()
var pos = 0
ctx . text . split ( ’\n ’). forEachIndexed { i , str ->
val wspace = str . takeWhile {
it == ’ ’ || it == ’\t ’ }
val count = wspace . wspace . expandTabs (4). length
map . put (i , IndentEntry ( count , pos + wspace . length ))
pos += str . length + 1
}
ctx . state ( IndentMap :: class ). map = map
Success
}
5. A class has access to all classes defined within its superclass and other ancestors.
6. A class cannot inherit from one of its outer classes.
7. In order to avoid ambiguous type names (for instance,
both a class defined in an outer class and a class defined
in a superclass could bear the same name), Examply
features type aliases that assign an alternate name to an
existing type. Type aliases can appear anywhere a class
definition can appear, and have the same visibility as class
definitions.
val indent = Parser { ctx ->
val new = ctx . indent . count
val old = ctx . istack . peek () ?: 0
if ( new > old ) Success after { ctx . istack . push ( new ) }
else ctx . failure {
" Expecting indentation > $old positions " }
}
Figure 4 shows the code handling namespace classification.
We do not engage with its minutiae, but instead give a highlevel description of its operation. The code itself should
demonstrate that the implementation of these ideas is terse
and readable, even to those who ignore the precise semantics
of some operations. We note however that in the code,
ctx.stack refers to the stack used to construct AST nodes. We
occasionally peek in this stack in order to retrieve identifiers.
The main data structure is the TypeStack State instance.
It holds a stack of Types, which are pairs formed by a string
and a list of other Types. Intuitively, each Type instance
represents a class name alongside with a list of classes
accessible through it: its inner classes and the inner classes
of its ancestors. We call these classes the private classes of
a class: they cease being accessible once the class definition
ends. We define two helper functions over the type stack:
isType checks if an identifier refers to a type, and priv returns
the private classes of the named class, or an empty list if no
such class exists.
Each time we encounter a new type during the parse, it is
pushed onto the type stack. This is the the task of the NewType
parser, which is applied to identifiers introduced by classes
val dedent = Parser { ctx ->
val new = ctx . indent . count
val old = ctx . istack . peek () ?: 0
if ( new < old || ctx . pos == ctx . text . length - 1)
Success after { ctx . istack . pop () }
else ctx . failure {
" Expecting indentation < $old positions " }
}
val newline = Predicate {
indent . end == pos || ctx . pos == ctx . text . length - 1 }
Figure 3. Using Autumn to implement significant whitespace handling for Examply.
6.2
Namespace Classification
Examply needs to distinguish between types (i.e., class
names) and other identifiers at parse-time, in order to resolve
ambiguities. We call this process namespace classification:
we want to know whether an identifier belongs to the namespace of types or not. Such parse-time tracking is reminiscent
of the C language, and is seldom seen in modern languages.
In Examply, we could avoid it by adding a keyword (e.g.,
Taming Context-Sensitive Languages with Principled Stateful Parsing
10
2016/9/20
and type aliases. A parameter controls whether the new type
is an alias, in which case it inherits the private classes of the
aliased class. Note that classes start with an empty list of
private classes. This will be updated once the class definition
is complete.
Once a scope (a class body, or some code block) is exited,
the types introduced within it are not longer accessible. To
enforce this, we use the Scoped parser: it saves the type stack
size, invokes its child parser (corresponding to a scope), then
removes any extraneous items from the type stack.
If the scoped body was a class body, the Type representing
the class on the stack must be updated with a list of its private
classes, so that an inner class may access them. This is the role
of the ClassDef parser. It looks up the class and superclass
names on the AST stack, then looks up the list of private
classes of the superclass. If found, this list is pushed on the
type stack. All of this is done after taking a snapshot of
type stack. Subsequently, the body parser is invoked and, if
successful, a delta of the type stack is generated using the
snapshot. This delta corresponds to the private classes of
the class, including those introduced by its superclass. The
snapshot is restored and the topmost entry on the type stack
(which represents the class) is removed and replaced with a
new one that binds the class name to its private classes.
A reduced version of this process also needs to happen
for anonymous classes: they need to access their superclass’
private classes, but no Type record must be created for them.
The anonClassInherit parser fullfills this role, by reusing the
inherit function.
Finally, to resolve the ambiguity, we use the classGuard
parser: it performs a lookahead, attempting to match an
identifier, and succeeding only this identifier refers to a type.
We defer our assessment of the approach until Section 6.4.
6.3
data class Type ( val name : String , val priv : LinkList < Type >)
class TypeStack : MonotonicStack < Type >()
val Context . types : TypeStack
get () = state ( TypeStack :: class )
fun isType ( ctx : Context , iden : String ): Boolean
= ctx . types . stream (). any { it . name == iden }
fun priv ( ctx : Context , iden : String ): LinkList < Type >
= ctx . types . stream ()
. filter { it . name == iden }
. next () ?. priv ?: LinkList ()
fun NewType ( child : Parser , alias : Boolean = false )
= Parser { ctx ->
child . parse ( ctx ) andDo {
val name = ctx . stack . peek () as String
val priv = if ( alias ) priv ( ctx , name )
else LinkList ()
ctx . types . push ( Type ( name , priv ))
} }
fun Scoped ( body : Parser ) = Parser { ctx ->
val size = ctx . types . size
body . parse ( ctx ) andDo { ctx . types . truncate ( size ) }
}
fun inherit ( ctx : Context , name : String )
= priv ( ctx , name ). stream (). each { ctx . types . push ( it ) }
fun ClassDef ( body : Parser )
= Parser { ctx ->
val parent = ctx . stack . at (0) as Maybe < SimpleType >
val name = ctx . stack . at (1) as String
val snapshot = ctx . types . snapshot ()
if ( parent is Some < SimpleType >)
inherit ( ctx , parent . value . name )
body . parse ( ctx ) andDo {
val diff = ctx . types . diff ( snapshot )
ctx . types . restore ( snapshot )
ctx . types . pop ()
ctx . types . push ( Type ( name , diff ))
} }
val anonClassInherit = Perform { ctx ->
inherit ( ctx , ctx . stack . at (1) as String ) }
val classGuard = Seq ( iden , Predicate { ctx ->
isType ( ctx , ctx . stack . peek () as String ) }
). ahead
Putting it all together
Figure 4. Using Autumn to implement namespace classification for Examply.
To illustrate the use of significant whitespace and namespace
classification (as presented in sections 6.1 and 6.2), let’s
look at two short examples. First, here is how we define an
indented code block:
The parser is simply guarded with the classGuard parser,
which checks if there is an identifier at the current input
position, and whether this identifier refers to a type. The
body of the class can access the superclass’ private classes
through anonClassInherit. The Scope wrapper ensures that
this access is restricted to the class and does not spread to
the code that follows. decls refers to an indented block of
declarations.
val statements =
Seq ( indent , Scoped (!" statement " until dedent ))
. collect < Stmt >()
The block starts by an increase in indentation (indent), and
ends when a decrease in indentation is encountered (dedent),
parsing statements in the mean time (!"statement" until
dedent). The collect part instructs the parser to collect all
statement nodes pushed onto the AST stack and to aggregate
them in a list, which is itself pushed onto that stack. We also
see that all indented statements form a scope (Scoped) in the
sense of Section 6.2.
Second, here is how the parser for “block constructors”
(i.e., anonymous classes) is defined:
6.4
We have implemented two context-sensitive features in for an
imaginary but non-trivial programming language. With the
code we have shown so far, the rest of the grammar is trivially
able to define code delimited by changes in indentation, or
by newlines (Section 6.1); or to direct the parse depending
on whether an identifier refers to a type (Section 6.2).
val blockCtorBody = Scoped ( Seq ( anonClassInherit , decls ))
val blockCtor
= Seq ( classGuard , iden , paramList , blockCtorBody )
. build { CtorCall ( get () , get () , get ()) }
Taming Context-Sensitive Languages with Principled Stateful Parsing
Discussion
11
2016/9/20
All this, by itself, is no mean feat. There are few parsing
tools where this is possible to begin with (most of them are
presented in Section 2). Significant indentation handling, in
particular, is non-trivial because Autumn does not include a
lexical analysis layer out of the box.
We also underline that the presented implementations are
rather terse, less than 50 lines of code each. Some of that
does come from our choice of implementation language, but
it also shows that the principled stateful parsing approach
does not impose a significant boilerplate overhead. In particular, the ability to reuse state-handling strategies, such
as MonotonicStack means that context-transparency comes
almost for free.
The state manipulations operations from sections 4.1
and 4.2.3 are strangely discreet in our examples. Significant
whitespace handling does not use them at all, while namespace classification performs a diff in order to capture the
types introduced by a class body. But because they do not
appear in the code does not mean the operations are not used,
they are simply hidden from view. The basic contract of principled stateful parsing is that each parser either succeeds or
leaves the state untouched. You can convince yourself that
all the parsers and parser combinators we introduced satisfy
this condition, either by reusing existing combinators, or by
delegating the responsibility for this to their single subparser.
It remains that the parsers we introduce do get backtracked
over during the parse. As such, their state mostly get saved
and restored by other parsers that invoke them, directly or
indirectly. It is in fact crucial for the state to get restored when
backtracking occurs: we need to know the correct indentation
level whenever we backtrack out of a block; we also need to
know which identifiers are classes, even when backtracking
over a type definition. Granted, given that most constructs
in the grammar are guarded by specific keywords, such
backtracking occurences should be rare. However, unlike
the other, often accidentally stateful parsers (cf. Section 2),
context transparency ensures that we can evolve the grammar
as we see fit, without fear of breaking the mechanisms we
just introduced. It is also a pre-requisite for safe grammar
composition.
In fact, the scarcity of state operations is a boon: it means
that the benefits of our approach come at very little cost,
at least implementation-wise. We would also disabuse the
reader of the notion that the Autumn codebase hides some
devilish complexity in order to make up for this: the whole
library [20] is less than 2500 lines of code. All pre-defined
parsers live in a single file of less than 500 lines. This file
defines around 50 parsers: those corresponding to all basic
PEG [8] operators, as well as numerous extensions, notably
to work with error messages, AST nodes, ...
6.5
Parboiled [7] for context-free grammars. The implementation
currently incurs overhead even for context-free grammars,
which we are working to reduce. The overhead scales with
the amount of state in use, depending on the State implementation details. The costly operations are the creation of
snapshots and deltas. In general, memory allocations tend to
be the bottleneck, so increased sharing between snapshots
results in better performances. Indeed, we’ve had success
with purely functional data structures [23].
Just like PEG parsing without full memoization, parsing
has exponential complexity in the worst case. In practice
however run times are acceptable, as programming language
grammars are fairly deterministic.
7.
Conclusion
In this paper, we proposed an approach to tackle the problem
of context-sensitive parsing. Our solution, unlike existing
ones, possesses the property of context transparency: grammatical constructs are unaware of the context shared between
their ancestors and their descendants, making it easier to
write, evolve and compose context-sensitive grammars.
We proceeded in two parts. First, we allowed parsers to
manipulate a mutable data store, so as to enable contextsensitivity through recall. Second, we required parsers to
behave transactionally: a parser must either succeed, or fail
and leave the state unaltered. This transactional discipline,
which we call principled stateful parsing, prevents parsing
mechanisms such as backtracking and memoization to break
the guarantee of context transparency.
To enforce the principled stateful parsing discipline, we
supplied formally specified state manipulation operations.
The role of these operations is to snapshot and restore the
state, as well as to create and merge deltas between a snapshot
and the current state.
We implemented our approach in a parsing library called
Autumn, and showed how it can be used in practice to implement common context-sensitive grammar features such
as significant whitespace and namespace classification. We
underline the low boilerplate and conceptual overhead introduced by the approach.
We view this work as a first step towards bringing disciplined grammarware engineering (as defined by Klint, Lämmel and Verhoef [17]) to context-sensitive parsing.
Performance
Performance has not been our focal point, but preliminary testing seems to indicate that performance is within an order of
magnitude of mainstream parsing tools such as Rats! [9] and
Taming Context-Sensitive Languages with Principled Stateful Parsing
12
2016/9/20
References
[17] P. Klint, R. Lämmel, and C. Verhoef. Toward an engineering
discipline for grammarware. ACM Transactions on Software
Engineering and Methodology, 14(3):331–380, July 2005.
[1] A. Afroozeh and A. Izmaylova. One parser to rule them all. In
ACM International Symposium on New Ideas, New Paradigms,
and Reflections on Programming and Software, Onward! 2015,
pages 151–170. ACM, 2015.
[18] D. Knuth. Semantics of context-free languages. Mathematical
systems theory, 2(2):127–145, 1968.
[19] N. Laurent and K. Mens. Parsing expression grammars made
practical. In Proceedings of the ACM SIGPLAN International
Conference on Software Language Engineering, SLE 2015,
pages 167–172. ACM, 2015.
[20] N. Laurent and K. Mens. Taming context-sensitive languages
with principled stateful parsing: Artifacts. Software Language
Engineering: Artifacts Track, 2016. https://github.com/
ncellar/sle2016.
[2] R. Atkey. The semantics of parsing with semantic actions.
In Proceedings of the 27th Annual IEEE/ACM Symposium on
Logic in Computer Science, LICS 2015, pages 75–84. IEEE
Computer Society, 2012.
[3] J. Aycock and R. N. Horspool. Schrödinger’s token. Software:
Practice and Experience, 31(8):803–814, 2001.
[4] P. Bagwell. Ideal hash trees. Technical Report LAMPREPORT-2001-001, Ecole polytechnique fédérale de Lausanne, 2001.
[5] N. Chomsky. Three models for the description of language.
IRE Transactions on Information Theory, 2(3):113–124, 1956.
[21] D. Leijen and E. Meijer. Parsec: Direct style monadic parser
combinators for the real world. Technical Report UU-CS2001-35, Department of Information and Computing Sciences,
Utrecht University, 2001.
[6] N. Chomsky. Formal properties of grammar. In Handbook of
Mathematical Psychology, chapter 12, pages 360–363 and 367.
Wiley, 1963.
[22] A. Moors, F. Piessens, and M. Odersky. Parser combinators
in Scala. Technical Report CW491, Department of Computer
Science, KU Leuven, Feb. 2008.
[23] C. Okasaki. Purely Functional Data Structures. Cambridge
University Press, New York, NY, USA, 1998.
[7] M. Doenitz. The Parboiled homepage, 2015. https://github.
com/sirthias/parboiled.
[24] T. Parr, S. Harwell, and K. Fisher. Adaptive LL(*) parsing: The
power of dynamic analysis. In Proceedings of the 2014 ACM
International Conference on Object Oriented Programming
Systems Languages & Applications, OOPSLA ’14, pages
579–598. ACM, 2014.
[8] B. Ford. Parsing expression grammars: A recognition-based
syntactic foundation. SIGPLAN Notices, 39(1):111–122, Jan.
2004.
[9] R. Grimm. Better extensibility through modular syntax. SIGPLAN Notices, 41(6):38–51, June 2006.
[25] F. Pereira and D. Warren. Readings in natural language
processing. chapter Definite Clause Grammars for Language
Analysis, pages 101–124. Morgan Kaufmann, 1986.
[10] D. Grune and C. J. Jacobs. Parsing Techniques: A Practical
Guide, p. 21–23. Springer, 2nd edition, 2008.
Informatica
[26] J. M. Spivey and J. Abrial. The Z notation. Prentice Hall, 1992.
[12] G. Hutton and E. Meijer. Monadic parsing in Haskell. Journal
of Functional Programming, 8(4):437–444, July 1998.
[27] M. J. Steindorfer and J. J. Vinju. Optimizing hash-array
mapped tries for fast and lean immutable jvm collections. In
Proceedings of the 2015 ACM SIGPLAN International Conference on Object-Oriented Programming, Systems, Languages,
and Applications, OOPSLA 2015, pages 783–800. ACM, 2015.
[11] G. Hedin. Reference attributed grammars.
(Slovenia), 24(3), 2000.
[13] R. Ierusalimschy. A text pattern-matching tool based on parsing expression grammars. Software: Practice and Experience,
39(3):221–258, Mar. 2009.
[28] The Free Software Foundation. The GNU Bison homepage,
2014. http://www.gnu.org/software/bison/.
[14] T. Jim and Y. Mandelbaum. A new method for dependent
parsing. In Proceedings of the 20th European Conference
on Programming Languages and Systems, ESOP 2011, pages
378–397. Springer, 2011.
[29] A. D. Thurston and J. R. Cordy. A backtracking LR algorithm for parsing ambiguous context-dependent languages.
In Proceedings of the 2006 conference of the Centre for Advanced Studies on Collaborative Research, October 16-19,
2006, Toronto, Ontario, Canada, pages 39–53, 2006.
[15] T. Jim, Y. Mandelbaum, and D. Walker. Semantics and
algorithms for data-dependent grammars. SIGPLAN Notices,
45(1):417–430, Jan. 2010.
[30] E. Van Wyk and A. Schwerdfeger. Context-aware scanning for
parsing extensible languages. In International Conference on
Generative Programming and Component Engineering, GPCE
2007. ACM, October 2007.
[16] L. Kallmeyer. Parsing Beyond Context-Free Grammars.
Springer, 2010.
Taming Context-Sensitive Languages with Principled Stateful Parsing
13
2016/9/20
| 6cs.PL
|
Schur’s theory for partial projective representations
arXiv:1711.06739v1 [math.GR] 17 Nov 2017
Mikhailo Dokuchaev and Nicola Sambonet∗
November 21, 2017
Abstract
This article focuses on those aspects about partial actions of groups
which are related to Schur’s theory on projective representations. It
provides an exhaustive description of the partial Schur multiplier, and
this result is achieved by introducing the concept of a second partial
cohomology group relative to an ideal, together with an appropriate
analogue of a central extension. In addition, the new framework is
proved to be consistent with the earlier notion of cohomology over
partial modules.
1
Introduction
It is becoming relatively common to encounter the word partial as a prefix for some classical terminology in algebra. This phenomenon is due to
the increasing awareness about the algebraic relevance of partial symmetries which explicitly or implicitly appear in various local-global aspects of
well-known mathematical topics. It was indeed the theory of C ∗ -algebras
which motivated the introduction of the notions of partial actions and partial representations of groups [17]. In this regard, the most prominent application of the partial actions is their involvement in the construction of
a crossed product, which encloses relevant classes of algebras [7]. Partial
representations are intimately related to partial actions in various ways, in
particular, they play a fundamental role in providing partial crossed product structures on concrete algebras. This can be seen applied to a series
∗
The first author was partially supported by Fapesp of Brazil and by CNPq of Brazil.
The second author was supported by Fapesp of Brazil. Address: Instituto de Matemática
e Estatı́stica da Universidade de São Paulo, Rua do Matão 1010, 05508-090 São Paulo,
Brazil. Email: [email protected] and [email protected]
1
of important algebras, among the most recent examples being C ∗ -algebras
related to dynamical systems of type (m, n) [2], the Carlsen-Matsumoto algebras of subshifts [8] and algebras related to separated graphs [1]. In the
latter paper, a partial-global passage was remarkably employed to a problem
on paradoxical decompositions.
The general concept of a twisted partial group action, used to define
more general crossed products, involves a kind of a 2-cocycle (see [9, 16]),
raising the desire to fit it into some cohomology theory. In order to obtain
some testing material for the development of such a theory, partial projective group representations were introduced and studied in [13, 14, 15, 28].
The notion of a partial multiplier naturally appears in analogy with the
classical Schur’s theory of projective representations. Since the algebraic
structures involved in partial representations are not only groups but also
inverse semigroups, the new theory diverges from the original one and the
multiplier results in a semilattice of abelian groups, called components. That
is briefly to say, a collection composed of abelian groups and parametrized by
a meet semilattice, where the operations of group-multiplications are compatible with the partial order. Subsequently, a cohomology theory based on
partial actions was developed in [10, 11, 12]. Notice that partial cohomology
turned out to be useful to deal with the ideal separation property of (global)
reduced C ∗ -crossed products in [25], where partial 2-cocyles appear as factor sets (twists) of partial projective representations (called twisted partial
representations in [25]) naturally related to (global) C ∗ -dynamical systems.
With respect to the partial Schur multiplier, it was shown in [10, Theorem 2.14] that each component is a union of cohomology groups with values
in non-necessarily trivial partial modules. Furthermore, in the case of an
algebraically closed field K, it was proved in [15, Theorem 5.9] that each
component of the partial Schur multiplier is an epimorphic image of a direct
product of copies of K× . However, computations of the partial Schur multiplier of concrete groups show that each component is, in fact, isomorphic
to a direct product of copies of K× (see [6, 15, 26, 27, 28, 29]), raising the
conjecture that this should be true for all groups.
The present article is inspired by the aforementioned conjecture, of which
it provides an affirmative answer for finite groups (Corollary 5.3), and it actuates an essential refinement for our understanding of partial projective
group representations. To this aim, the theory is reformed to encompass a
2
broader setting, specifically to allow arbitrary coefficients and to establish
a partial analogue for Schur’s theory of central extensions (in particular,
see Theorem 3.3). Influenced by known properties of partial factor sets, the
concepts of a pre-cohomology group, of a second partial cohomology group
relative to an ideal, and of a second partial cohomology semilattice of groups
with values in an arbitrary abelian group are introduced. Then the partial Schur multiplier is shown to coincide with the second partial cohomology semilattice of groups with values in K× . This new level of abstraction
conveys standard techniques into the scene, which promptly produce new
theorems.
Most remarkably, in the finite case the pre-cohomology groups are completely determined by minor information about the underlying group (Theorem 4.3).
Also, taking the ideals into accounts, a general statement is established for
the second partial cohomology groups (Theorem 5.2). Furthermore, considering appropriate partial G-modules, one is able to conclude, in particular,
that each component of the partial Schur multiplier is isomorphic to a single
partial cohomology group (see Theorem 6.1) in the sense of [10], bringing
thus the partial Schur theory closer to the classical one.
The text is structured as follows: in section 2 some fundamental concepts in the theory of partial actions of groups and in Schur’s theory will
be recalled; in section 3 new concepts will be introduced, and the partial
analogue for Schur’s theory with arbitrary coefficients will be developed;
in section 4 and, respectively, in section 5 the pre-cohomology groups and
the second partial cohomology semilattice of groups will be studied in details; lastly, in section 6 the consistency with the previous notion of partial
cohomology will be proved.
2
2.1
Prerequisites
The universal inverse semigroup of a group
Inverse semigroups are the algebraic structures which capture the idea of
partial symmetry [21]. In an inverse semigroup S the system of equations
xyx = x ∧ yxy = y
(1)
admits a unique solution for y in terms of x, commonly denoted by y = x−1 .
This is easily recognized to be a generalization of groups, still it incorporates
3
the semilattices, which are the semigroups consisting of commuting idempotents. A partial bijection over a set X is any bijection between subsets of X.
Multiplication of partial bijection is defined as follows: if α : A → Aα and
β : B → B β , then β ◦ α is their composition on the largest subset where it
makes sense, i. e. α−1 (Aα ∩ B). Necessarily, one has to admit the existence
of the map ∞ : ∅ → ∅, and this serves as the zero element. As the reader
has noticed, we convene to denote with ∞ the zero element of a semigroup,
if this ever exists, coherently with the additive notation for abelian groups.
The set of partial bijections over X turns out to be an inverse semigroup
in the obvious way, that is by inversion of functions, which is named the
symmetric inverse semigroup I(X).
We recall that, given an ideal I of a semigroup S, the Rees congruence
is defined by setting x ∼ y when either x = y or both x and y belong to I.
The corresponding quotient is denoted by S/I. In general, if ϑ : S → T is a
homomorphism of semigroups and T admits the zero ∞, then ϑ−1 (∞) is an
ideal of S.
Inverse semigroups apply to the theory of groups through the notion
of partial actions and partial homomorphisms of groups [17]. A partial
action is a map ψ : G → I(X) satisfying the compatibility conditions that
ψ(1) = idX and the composition ψ(g) ◦ ψ(h) coincides with the restriction
of ψ(gh). In turn, this notion results to be a special case of the following.
A partial homomorphism is a map ψ from a group G into a semigroup M
satisfying the conditions
ψ(1) · ψ(g) = ψ(g) = ψ(g) · ψ(1)
(2)
ψ(g−1 ) · ψ(g) · ψ(h) = ψ(g −1 ) · ψ(gh)
(3)
ψ(g) · ψ(h) · ψ(h−1 ) = ψ(gh) · ψ(h−1 )
(4)
for any pair of elements g and h of G. These relations, which openly emerge
from the compatibility requirement of partial action, are used to introduce
the universal inverse semigroup S(G) associated with G (see [17]). Precisely,
S(G) is the abstract semigroup generated by symbols Π(g) subject to relations analogue to (2)-(4), so that the map Π : G → S(G) is an injective
partial homomorphism, and this construction is universal in sense that
S(G)
③<<
Π ③③
③
ψ̄
③③
③③ ψ
// M
G
4
(5)
every partial homomorphism ψ extends to a semigroup homomorphism ψ̄
uniquely by means of Π.
In fact, S(G) is isomorphic to the Birget-Rhodes expansion of G [24] (see
also [3, 4]). Thus, a generic element of S(G) is identified with a pair (R, g)
where R is a finite subset of G containing 1 and g, whereas multiplication
obeys to the rule
(R, g) · (S, h) = (R ∪ gS, gh) .
The result is an inverse semigroup since the system of equations (1) admits
the unique solution (R, g)−1 = (g−1 R, g −1 ), and the universal partial homomorphism Π associates an element g of the group with Π(g) = ({1, g}, g). It
is convenient to extend the definition of Π to a higher number of elements
as Π(g1 , . . . , gk ) = Π(g1 ) · · · Π(gk ). In particular, considering two elements
yields
Π(g, h) = ({1, g}, g) · ({1, h}, h) = ({1, g, gh}, gh) ,
(6)
whereas considering three elements yields
Π(x, y, z) = ({1, x}, x) · ({1, y}, y) · ({1, z}, z) = ({1, x, xy, xyz}, xyz) . (7)
One may notice how these relations confirm that Π satisfies (2)-(4).
It is important to describe the ideals of S(G) since these determine the
vanishing of a partial homomorphism. Precisely, in the notation of (5) the
equality ψ(g1 ) · · · ψ(gk ) = ∞ occurs if and only if Π(g1 , . . . , gk ) belongs to
I = ψ̄ −1 (∞). To begin with, the principal ideals in S(G), which we will
denote by h·iS , can be characterized as follows [14, Corollary 2]:
h(R, g)iS = h(S, h)iS ⇐⇒ (S, h) = (x−1 R, x−1 y) for some x, y ∈ R . (8)
For our purpose, one may just consider the following relations which are
direct consequences of the definition of a partial representation:
hΠ(g, g−1 )iS = hΠ(g)iS = hΠ(g−1 )iS = hΠ(g−1 , g)iS ,
(9)
and more generally, for any 1 ≤ j ≤ r and 1 ≤ k ≤ r,
hΠ(g1 , . . . , gr )iS
= hΠ((g1 · · · gj )−1 , g1 , . . . , gr )iS
= hΠ(g1 , . . . , gr , (gk · · · gr )−1 )iS
= hΠ((g1 · · · gj )−1 , g1 , . . . , gr , (gk · · · gr )−1 )iS .
5
(10)
In addition, with respect to (6), taking y = 1 and either x = g or x = gh in
the above characterization (8) yields
hΠ(g, h)iS = hΠ(h, h−1 g −1 )iS = hΠ(h−1 , g −1 )iS .
(11)
These identities prove that, given a partial homomorphism ψ, the subset of
G × G consisting of those pairs (g, h) for which ψ(g) · ψ(h) = ∞ is invariant
under the action of the symmetric group S3 generated by the transformations
(g, h) 7→ (h, h−1 g−1 ) , (g, h) 7→ (h−1 , g−1 ) .
(12)
This action, which has been discovered in [13], will be central in the present
essay.
Another family of ideals can be described by the simple observation that,
for any pair of elements (R, g) and (S, h) of S(G), since |R ∪ gS| ≥ |S| and
|R∪gS| ≥ |R|, left and right multiplication are “non-decreasing operations”.
Consequently, for any positive integer k, the set
Nk = {(R, g) ∈ S(G) | |R| ≥ k + 1}
is an ideal of S(G).
In the present essay the crucial case is when k = 3. Indeed, in the notation of (5), for any partial homomrphisms ψ considered hereby it will be
assumed that the ideal I = ψ̄ −1 (∞) is proper in S(G) and it contains N3 .
This assumption is motivated by the theory on partial projective representations where the semilattice
Λ = (Λ, ∪) , Λ = {I | N3 ≤ I ⊳ S(G)}
(13)
plays a prominent role (see [15]).
Taking this into account, suppose that Π(x, y, z) is not contained in I, for
I = ψ̄ −1 (∞) lying in Λ. Accordingly with (7), one has |{1, x, xy, xyz}| ≤ 3,
which can be rewritten as
1 ∈ {x, y, z, xy, yz, xyz} .
(14)
Therefore, the failure on this condition implies ψ(x) · ψ(y) · ψ(z) = ∞.
2.2
A sketch of classical Schur’s theory
The ordinary multiplier is the fundamental object in the theory of projective
representations and central extensions [22, 23, 31]. Hereby, few of the very
6
elementary concepts are mentioned in order to fix the notation and facilitate
the description of their partial analogue.
A central extension of a group M is a group E endowed with a surjective
homomorphism π : E → M such that A = ker π is contained in the center
Z(P ). This situation is summarized in the diagram
A
ι
π
// E
// M
(15)
The interest is to consider such a central extension together with a group G
and a homomorphism ψ : G → M . The archetypes are:
i. When M = G and ψ is the identity, then E is a central extension of G.
ii. When E = GL(n, C) and M = PGL(n, C) for some positive integer n,
then ψ is a projective representation of G.
Hence, the choice of a section ϕ : G → E completes the commutative diagram:
G
(16)
ϕ
A
ι
// E
~~
ψ
π
// M
One can not expect that the map ϕ is a homomorphism. However, the
failure for this to happen is encoded in the equation
ϕ(g) · ϕ(h) = ϕ(gh) · σ(g, h)
(17)
where σ is a function in AG×G . Clearly, not any function is admissible:
adopting the additive notation for group multiplication in A, associativity
of E proves that σ satisfies the equality
σ(x, y) + σ(xy, z) = σ(x, yz) + σ(y, z)
for any x, y and z in G. In connection with cohomology this fact is restated
as follows. The coboundary homomorphism
δ2 : AG
×2
→ AG
×3
is defined by means of the relation
δ2 σ(x, y, z) = σ(y, z) − σ(xy, z) + σ(x, yz) − σ(x, y) ,
7
(18)
and the group of cocycles is the kernel of this homomorphism
Z 2 (G, A) = {σ ∈ AG×G | δ2 σ = 0} .
Thus, the function σ determined by (17) is an element of Z 2 (G, A). Luckily
enough the process is reversible and any σ ∈ Z 2 (G, A) defines a central
extension (15) of the group G as follows. The underlying set is A × G and
multiplication is given by
(λ, g) · (µ, h) = (λ + µ + σ(g, h), gh) .
(19)
Therefore, there is a correspondence between the central extensions and the
cocycles of G which, however, depends on the choice of section. Nonetheless,
any change of section, as it can be seen easily, corresponds to addition of a
coboundary
δ1 ζ(g, h) = ζ(h) − ζ(gh) + ζ(g) .
(20)
Hence, it is of interest to introduce the group
B 2 (G, A) = {δ1 ζ | ζ ∈ AG } ,
and to consider the second cohomology group
H 2 (G, A) = Z 2 (G, A)/B 2 (G, A) .
The terminology Schur multiplier applies to the case when G is a finite
group and A = C× , which answers to the original problem of classifying the
projective representations. For general coefficients, the facts here discussed
provide a one-to-one correspondence between the isomorphism classes of the
central extensions of G by A and the elements of H 2 (G, A)
Of course, the above definition is a special case in the theory of group
cohomology (see [5]). Indeed, in the present manuscript it is always assumed
that the (global) group actions over the coefficient modules are trivial. In
view of this, the first cohomology group simply is
H 1 (G, A) = ker δ1 = Hom(G, A) .
3
The partial analogue notions
Also the partial counterpart of the theory originates with the study of partial projective representations in K× -cancellative monoids [13, 14, 15] (previously named K-cancellative). However, the most general situation can be
extended to semigroups and partial homomorphisms.
8
Firstly, referring to the diagram (15), in the current framework E will be
taken as an A-cancellative central extension of a semigroup M by an abelian
group A, in the following sense. A central extension is a semigroup E which
contains A as a central subgroup, and M is the quotient semigroup E/A
with respect to the following congruence: x ∼ y whenever there exist a and
b in A such that ax = by. Given a central extension E endowed with zero
∞, this extension is called A-cancellative if, for every a, b in A and every x
in E \ {∞}, the equality ax = bx implies a = b. Observe that, this condition
yields that, for every a in A, the equation ax = ∞ has no solution for x in
E other than x = ∞.
In second place, the due modification in the diagram (16) is given by an
A-cancellative central extension of M together with a group G and a partial
homomorphism ψ : G → M . Notice that, when ∞ ∈ E, then M has at least
two elements. In this case, it is convenient to assume that ψ(1) 6= ∞ since,
otherwise, ψ(g) = ∞ for all g. With respect to (3)-(4) and (17), this time
the failure for the section ϕ to be a partial homomorphism can be shown to
be encoded in the equations
ϕ(g−1 ) · ϕ(g) · ϕ(h) = ϕ(g −1 ) · ϕ(gh) · σ(g, h)
−1
ϕ(g) · ϕ(h) · ϕ(h
−1
) = ϕ(gh) · ϕ(h
) · σ(g, h)
(21)
(22)
where σ(g, h) is an element of A. Actually, it is convenient to regard σ(g, h)
as an element of the semigroup A∞ = A ∪ {∞}, obtained adjoining the
zero ∞ to the abelian group A, and set σ(g, h) = ∞ whenever the above
equations reduce to the triviality ∞ = ∞.
Similarly with the case of a partial projective representation over a field,
the fact that the same function σ appears in both (21) and (22) is not
obvious, but the arguments given in the proof of [13, Theorem 3] can be
easily adapted to this case. Moreover, also in the current situation the map
σ has not to satisfy the cocycle identity [13, page 260]. However, by means of
(10), associativity of E applied to the relation ϕ(x−1 )·ϕ(x)·ϕ(y)·ϕ(z)·ϕ(z −1 )
shows that δ2 σ(x, y, z) = 0 whenever Π(x, y, z) does not belong to the ideal
I = ψ −1 (∞). Accordingly to (14), it is of interest to introduce the following:
Definition 1. A pre-cocycle of G with values in A is a function σ in AG×G
satisfying the weak cocycle condition:
1 ∈ {x, y, z, xy, yz, xyz} ⇒ δ2 σ(x, y, z) = 0 .
9
The group of pre-cocycles is denoted by pZ 2 (G, A), and
pH 2 (G, A) = pZ 2 (G, A)/B 2 (G, A)
is the pre-cohomology group of G.
The zero ∞, which does not appear in the above definition, is reintroduced as follows. Any fixed ideal I in the semilattice Λ, introduced in (13),
defines an additive characteristic function ǫI : G × G → A∞ obeying to
(
∞ Π(g, h) ∈ I
ǫI (g, h) =
0 otherwise.
associating α 7→ α + ǫI is a homoIn these terms, the map AG×G → AG×G
∞
morphism of semigroups.
Definition 2. In the current notation, define
Z 2 (G, I; A) = pZ 2 (G, A) + ǫI , B 2 (G, I; A) = B 2 (G, A) + ǫI .
Then, the quotient
H 2 (G, I; A) = Z 2 (G, I; A)/B 2 (G, I; A) ,
is the second partial cohomology group relative to I with coefficients in A.
We shall see below (Theorem 6.1) that this definition is coherent with
the notion of partial cohomology groups given in [10]. Moreover, the fact
Z 2 (G, N3 ; A) = pZ 2 (G, A)
justifies that only ideals containing N3 are considered.
Clearly, the identity element of H 2 (G, I; A) is the class [ǫI ]. In addition,
in AG×G
one has ǫI + ǫJ = ǫI∪J . Thus, by allowing external sums, that is
∞
[σ + ǫI ] + [τ + ǫJ ] = [σ + τ + ǫI∪J ], it is possible to join all of these groups
as follows:
Definition 3. The second partial cohomology semilattice of groups of G is
a
H 2 (G, Λ; A) =
H 2 (G, I; A)
I∈Λ
The semigroup of idempotents in H 2 (G, Λ; A) consists of the classes [ǫI ]
as I varies in Λ, and it is canonically isomorphic with (Λ, ∪). Furthermore,
the definition of Z 2 (G, −; A) and B 2 (G, −; A) immediately yields:
10
Proposition 3.1. For any pair of ideals I and J in Λ, the homomorphism
H 2 (G, J; A) → H 2 (G, I ∪ J; A) mapping µ 7→ µ + [ǫI ] is surjective.
From the above discussion, it follows that a map σ defined by (21)-(22)
is an element of Z 2 (G, I; A) for I = ψ(∞)−1 . The following construction,
which is analogue to (19), provides the converse of this statement. Namely,
for any ideal I in Λ, every element of Z 2 (G, I; A) can be obtained as a map
σ arising from an A-cancellative central extension together with a partial
homomorphism, as in (21)-(22).
Lemma 3.2. Given σ in Z 2 (G, I; A), then the set E = (A × S(G) \ I)∪{∞}
endowed with multiplication
(
(λ + µ + σ(x, y), R ∪ xS, xy) if (R ∪ xS, xy) ∈
/I
(λ, R, x) · (µ, S, y) =
∞
otherwise
results in an A-cancellative central extension of S(G)/I. Moreover, E is a
monoid with identity element (−σ(1, 1), Π(1)), and the natural partial homomorphism G → S(G)/I admits the section
(
(−σ(1, 1), Π(g)) if Π(g) ∈
/I
ϕ : G → E , g 7→
∞
otherwise
which is associated with σ.
Proof. To prove the associativity of E, given X = (R, x), Y = (S, y) and
Z = (T, z) in S(G) \ I together with λ, µ and ν in A, one considers the two
ways for computing the product (λ, X)·(µ, Y )·(ν, Z). There is no loss of generality assuming that XY Z does not belong to I, since otherwise the product
results ∞. In particular, associativity corresponds to the cocycle identity
δ2 σ(x, y, z) = 0 under this condition. Write XY Z = (R ∪ xS ∪ xyT, xyz).
Since {1, x, xy, xyz} is contained in R ∪ xS ∪ xyT and N3 is contained in
I, it follows that |{1, x, xy, xyz}| ≤ 3, that is 1 ∈ {x, y, z, xy, yz, xyz}. This
proves the associativity of E since, by hypothesis, σ = σ ′ + ǫI for some
σ ′ ∈ pZ 2 (G, A). Finally, denoting by ψ the canonical partial homomorphism G → S(G)/I, sending g to Π(g) whenever the latter is not in I, and
to ∞ otherwise, and taking the section ϕ : G → E which accordingly maps
g either to (−σ(1, 1), ψ(g)) or to ∞, it is evident that the associated map is
σ = σ ′ + ǫI .
11
It is routine to check that the isomorphism class of the extension E only
depends on the partial cohomology class of σ in H 2 (G, I; A), consequently:
Theorem 3.3. For any ideal I in Λ, there is a one-to-one correspondence between the equivalence classes of A-cancellative central extensions
of S(G)/I and the elements of H 2 (G, I; A).
In particular, in the case A = K× , any partial cocycle in Z 2 (G, Λ; K× )
defines a K× -cancellative monoid together with a partial projective representation, and vice-versa. Therefore, the partial multiplier pM (G) introduced
in [13] turns out to be a partial cohomology semilattice of groups:
Proposition 3.4. Let G be a group and K a field. Then the partial multiplier pM (G) over K coincides with H 2 (G, Λ; K× ).
4
Pre-cohomology
The pre-cocycles can be characterized in several different ways:
Lemma 4.1. Let σ ∈ AG×G . The following properies are equivalent:
i. σ is a pre-cocycle.
ii. 1 ∈ {xy, xyz} ⇒ δ2 σ(x, y, z) = 0
iii. For every g, h ∈ G the following holds:
a. σ(h, h−1 g−1 ) = σ(g, h) + σ(gh, h−1 g−1 ) − σ(g, g −1 )
b. σ(h−1 , g−1 ) = −σ(h, h−1 g−1 ) + σ(h, h−1 ) + σ(1, 1)
Moreover, if the above properties are satisfied, then:
c. σ(g, 1) = σ(1, 1) = σ(1, g)
d. σ(g, g−1 ) = σ(g −1 , g)
e. σ(h−1 g−1 , g) = σ(g, h) + σ(gh, h−1 g−1 ) − σ(h, h−1 )
f. σ(gh, h−1 ) = −σ(g, h) + σ(h, h−1 ) + σ(1, 1)
g. σ(g−1 , gh) = −σ(g, h) + σ(g, g −1 ) + σ(1, 1)
Proof. Denote by i.x the property [x = 1 ⇒ δ2 σ(x, y, z) = 0, ∀y, z], and
so on, so that the property i is the conjunction of the properties: i =
i.x ∧ i.y ∧ i.z ∧ i.xy ∧ i.yz ∧ i.xyz. The proof is divided in eight steps.
12
I) i ⇒ ii. This is trivial, since ii = i.xy ∧ i.xyz.
II) i.xyz = iii.a. Indeed, iii.a = [δ2 σ(g, h, h−1 g−1 ) = 0, ∀g, h] = i.xyz.
III) i.xyz ⇒ c. Observe that g−1 ·g ·1 = 1·g ·g −1 = 1, ∀g. Thus, i.xyz ⇒
[σ(g, 1)−σ(1, 1) = δ2 σ(g−1 , g, 1) = 0]∧[σ(1, 1)−σ(1, g) = δ2 σ(1, g, g−1 ) = 0].
IV) ii ⇐⇒ iii. Write down
δ2 σ(h, h−1 , g−1 ) = σ(h−1 , g−1 ) − σ(1, g −1 ) + σ(h, h−1 g −1 ) − σ(h, h−1 ) .
Then, by c one has σ(1, g−1 ) = σ(1, 1), so that i.xy ∧ c ⇒ iii.b, as well as
iii.b ∧ c ⇒ i.xy. Therefore, the result follows by II and III.
V) i.xyz ⇒ i.x ∧ i.y ∧ i.z. It is easily seen that c ⇒ i.x ∧ i.y ∧ i.z, simply
by consideration of δ2 σ(1, y, z), δ2 σ(x, 1, z), and δ2 σ(x, y, 1). Therefore, this
statement follows by III.
VI) ii ⇒ d. Firstly, i.xy ⇒ [δ2 (g, g −1 , g) = 0, ∀g]. Thus, writing down
explicitly 0 = δ2 σ(g, g−1 , g) = σ(g −1 , g) − σ(1, g) + σ(g, 1) − σ(g, g−1 ), the
claim follows by III.
VII) iii ⇒ i. In view of IV and V, it is left to prove that iii ⇒ i.yz.
Observe that i.yz = [δ2 σ(x, y, y −1 ) = 0, ∀x, y], and explicitly write down
δ2 σ(x, y, y −1 ) = σ(y, y −1 ) − σ(xy, y −1 ) + σ(x, 1) − σ(x, y) .
(23)
Setting h = xy and g = x−1 in iii.b, yields
σ(y −1 x−1 , x) = −σ(xy, y −1 ) + σ(xy, y −1 x−1 ) + σ(1, 1) .
By the use of c, which is allowed by III, σ(1, 1) = σ(x, 1), so that
−σ(xy, y −1 ) + σ(x, 1) = σ(y −1 x−1 , x) − σ(xy, y −1 x−1 ) .
(24)
Therefore, substituting (24) in (23), one obtains
δ2 σ(x, y, y −1 ) = σ(y, y −1 ) + σ(y −1 x−1 , x) − σ(xy, y −1 x−1 ) − σ(x, y) . (25)
Now, setting g = y −1 x−1 and h = x in iii.a, yields
σ(x, y) = σ(y −1 x−1 , x) + σ(y −1 , y) − σ(y −1 x−1 , xy) ,
that is
σ(y −1 , y) + σ(y −1 x−1 , x) − σ(y −1 x−1 , xy) − σ(x, y) = 0 .
13
(26)
The use of d, which is allowed by VI, yields σ(y −1 , y) = σ(y, y −1 ), and
σ(xy, y −1 x−1 ) = σ(y −1 x−1 , xy). Thus, (26) implies that (24) vanishes, that
is δ2 σ(x, y, y −1 ) = 0.
VIII) i ⇒ e ∧ f ∧ g. Writing down the equalities δ2 σ(h−1 g−1 , g, h) = 0,
δ2 σ(g, h, h−1 ) = 0, and δ2 σ(g−1 , g, h) = 0 explicitly, which are guaranteed
by i.xyz, i.yz, and i.xy respectively, it can be easily seen that the result just
requires the use of c and d, which is allowed by III and VI.
The above description of the pre-cocycles allows one to relate these functions with the action of S3 defined in (12).
Proposition 4.2. Let G be any group. Introduce on G × G the following
equivalence relation: for any element g of G set
(g, 1) ∼ (1, 1) ∼ (1, g) , (g, g −1 ) ∼ (g −1 , g)
and, for any pair of elements g and h of G satisfying g 6= 1 6= h 6= g −1 set
(g, h) ∼ (h, h−1 g −1 ) ∼ (h−1 , g −1 ) .
Then the group of pre-cocycles pZ 2 (G, A) is isomorphic with the group of
functions AΩ over the set of equivalence classes Ω = G × G/ ∼.
Proof. Fix a choice of representatives ∆ : Ω → G × G satisfying, for convenience, ∆(1, 1)∼ = (1, 1). As for the homomorphism pZ 2 (G, A) → AΩ
simply take the composition σ 7→ σ ◦ ∆. Whereas, the homomorphism
AΩ → pZ 2 (G, A), such that f 7→ σ, is defined as follows. First, set
σ(g, 1) = σ(1, g) = f (1, 1)∼
, σ(g, g −1 ) = σ(g −1 , g) = f (g, g −1 )∼
for any g ∈ G. Then, for any other class ω of Ω, that is ∆(ω) = (g, h) where
g 6= 1 6= h 6= g −1 , set
σ(g, h) = f (ω) ,
and, in accordance with the relations iii.a, iii.b iii.e, iii.f iii.g of Lemma 4.1,
define the values of σ over the remaining elements of the orbit as follows:
σ(h, h−1 g−1 ) = f (ω) + σ(gh, h−1 g−1 ) − σ(g, g −1 )
σ(h−1 , g−1 ) = −f (ω) + σ(g, g −1 ) + σ(h, h−1 ) − σ(gh, h−1 g−1 ) + σ(1, 1)
σ(h−1 g−1 , g) = f (ω) + σ(gh, h−1 g−1 ) − σ(h, h−1 )
σ(gh, h−1 ) = −f (ω) + σ(h, h−1 ) + σ(1, 1)
σ(g−1 , gh) = −f (ω) + σ(g, g −1 ) + σ(1, 1) .
14
It is easy to check that these homomorphisms are inverse of each other.
In the case of a finite group G, the previous result proves that pZ 2 (G, Z)
is a finitely generated abelian group of functions, so that standard techniques
can be used (see [19, §97]).
Theorem 4.3. Let G be a group of order n < ∞, and denote by G(k) the
set of elements of order k in G. Then, for any abelian group A,
pH 2 (G, A) ≃ Am−n ⊕ A ⊗ G/[G, G] ,
where m = (n2 + 2|G(3) | + 3|G(2) | + 5)/6.
Proof. Since G is finite then Ω is also finite and, consequently, the group
AΩ is a finitely generated A-module. Denoting m = |Ω|, the above formula
for m in terms of n, |G(2) |, and |G(3) | can be easily achieved (a similar
computation is given in [6]). Furthermore, if Ω = {ω1 , . . . , ωm } then
AΩ = Ahω1 ⊕ · · · ⊕ Ahωm ,
where hωi denotes the characteristic function hωi : ωj 7→ δij (the Kronecker
symbol). Hence,
AΩ ≃ A ⊗ ZΩ , ZΩ = Zhω1 ⊕ · · · ⊕ Zhωm .
By Proposition 4.2, unlike the analogue case in the classical theory, it follows
that
pZ 2 (G, A) ≃ A ⊗ pZ 2 (G, Z) .
In view of this fact, one first considers the case A = Z. Hence, in
the subgroup Z 2 (G, Z) admits a complement: indeed, given any
σ ∈ ZG×G , and a positive integer q, one has δ2 (qσ) = 0 if and only if δ2 σ = 0,
so that Z 2 (G, Z) is a pure subgroup and thus, since pZ 2 (G, Z) is finitely
generated, it is complemented [18, Theorem 28.2]. Hence, denoting by r the
free-abelian rank of Z 2 (G, Z), one has pZ 2 (G, Z) = Zm−r ⊕ Z 2 (G, Z), and
so
pH 2 (G, Z) = Zm−r ⊕ H 2 (G, Z) .
pZ 2 (G, Z)
It is well known that, since G is finite, H 2 (G, Z) ≃ H 1 (G, Q/Z) ≃ G/[G, G]
(see [5, Proposition 6.1, Corollary 10.2]). Therefore, in order to prove the
theorem’s claim for A = Z it is left to show that r = n. To see this, it is
15
sufficient to find a finite index subgroup of Z 2 (G, Z) isomorphic with Zn ,
and this goal is achieved by taking B 2 (G, Z). Indeed, since |H 2 (G, Z)| <
∞, it has finite index and, since n < ∞ implies that H 1 (G, Z) = 0, the
desired isomorphism is given by the differential δ1 : C1 (G, Z) = ZG ≃ Zn →
B 2 (G, Z).
Returning to the case of generic coefficients one has to consider the
quotient pZ 2 (G, A)/B 2 (G, A) and, more precisely, to describe B 2 (G, A) as
a subgroup of pZ 2 (G, A). The only difficulty is that, denoting by hg the
characteristic function relative to g in C 1 (G, Z) = ZG , the canonical (finite)
set {δ1 hg | g ∈ G} freely generates B 2 (G, Z) as a Z-module, but in B 2 (G, A)
the corresponding generating set {1A ⊗ δ1 hg | g ∈ G} is not necessarily free.
Consequently, a more delicate analysis of pZ 2 (G, Z) is needed in order to find
a basis presenting a good behavior. To this aim, one considers the inclusion
B 2 (G, Z) ⊆ Z 2 (G, Z) of finitely generated free Z-modules, and uses Smith’s
normalization theorem to obtain simultaneous basis (see [30, Theorem 8.61]);
therefore, there exists elements f1 , . . . , fn of Z 2 (G, Z) together with nonnegative integers λ1 , . . . , λn , which are uniquely determined by imposing
the condition that λi divides λi+1 for every i < n, satisfying
Z 2 (G, Z) = Zf1 ⊕ · · · ⊕ Zfn , B 2 (G, Z) = λ1 Zf1 ⊕ · · · ⊕ λn Zfn .
As already remarked, there is an isomorphism in cohomology which allows
to compute the integral coefficients as follows:
Zλ1 ⊕ · · · ⊕ Zλn = H 2 (G, Z) ≃ G/[G, G] .
At this moment one chooses a basis {gm+1 , . . . , gn } of the complement Zm−n
to obtain a decomposition pZ 2 (G, Z) = Zf1 ⊕ · · · ⊕ Zfn ⊕ Zgn+1 ⊕ · · · ⊕ Zgm .
It is readily seen that 1A ⊗ f1 , . . . , 1A ⊗ gm freely generates pZ 2 (G, A) as
an A-module. Therefore, denoting by T = λ1 Af1 ⊕ · · · ⊕ λn Afn , it is left
to prove that B 2 (G, A) = T . Since each δ1 hg ∈ B 2 (G, Z) can be expressed
as an integral combination of the elements λ1 f1 , . . . , λn fn , it follows that
1A ⊗ δ1 hg ∈ T for every g ∈ G, so that B 2 (G, A) ≤ T . On the other
hand, since the elements {δ1 hg | g ∈ G} constitute a basis of B 2 (G, Z),
then every λi fi is an integral combination of these elements, proving that
λi 1A ⊗ fi ∈ B 2 (G, A), so that T ≤ B 2 (G, A).
16
5
Partial cohomology relative to an ideal
In this section the results about the pre-cohomology groups are generalized
to the partial cohomology groups relative to ideals I ∈ Λ.
In the above notation, let ∆ : Ω → G × G denote a choice of representatives satisfying ∆(1, 1)∼ = (1, 1), and let Ω∗ = Ω \ {(1, 1)∼ }. Then the map
Π : G × G → S(G) defined in (6) induces a map
Π∼ : Ω∗ → Λ , ω 7→ hΠ ◦ ∆(ω)iS ∪ N3 ,
which, in view of the equalities (9) and (11), does not depend on the choice
of ∆. In this regard the relative pre-cocycles are characterized as follows:
Lemma 5.1. A function σ in AG×G
lies in Z 2 (G, I; A) if and only if it
∞
satisfies
i. Π(g, h) ∈ I ⇐⇒ σ(g, h) = ∞,
ii. Π(x, y, z) ∈
/ I ⇒ δ2 σ(x, y, z) = 0.
Moreover, the group Z 2 (G, I; A) is isomorphic with the group of function
AΩI where ΩI = {(1, 1)∼ } ∪ {ω ∈ Ω∗ | Π∼ (ω) 6⊆ I}.
Proof. Clearly, any function of the form σ = σ ′ + ǫI where σ ′ ∈ pZ 2 (G, A)
satisfies i. To see that σ also satisfies ii, firstly one has to show that no
summand of δ2 σ is equal to ∞. Equivalently, that no-one among Π(x, y),
Π(xy, z), Π(x, yz), and Π(y, z) is contained in I. To this aim, the cases of
Π(x, y) and Π(y, z) are trivial; on the other hand, assuming that Π(xy, z) ∈
I yields Π(x−1 , x, y, z) = Π(x−1 , xy, z) ∈ I, but then also Π(x, y, z) =
Π(x, x−1 , x, y, z) ∈ I contradicting the hypothesis, and the case of Π(x, yz) is
similar. Therefore, when Π(x, y, z) ∈
/ I then δσ(x, y, z) = δσ ′ (x, y, z). However, since I contains N3 , Π(x, y, z) ∈
/ I implies that 1 ∈ {x, xy, xyz, yz, z},
′
so that δσ (x, y, z) = 0.
Vice versa, given σ satisfying the properties i and ii, one has to produce
a pre-cocycle σ ′ for which σ = σ ′ + ǫI . Accordingly to Proposition 4.2, one
defines f in AΩ by setting f (ω) = 0 if σ ◦ ∆(ω) = ∞, and f (ω) = σ ◦ ∆(ω)
otherwise. Then, the desired σ ′ is obtained by means of the isomorphism
AΩ → pZ 2 (G, A) as f 7→ σ ′ .
Finally, the above defined function f ∈ AΩ can be regarded as an element of AΩI simply by restriction. Conversely, any function in AΩI can be
17
extended to an element f of AΩ by imposing the value 0 over the complementary set Ω \ ΩI , resulting as above in an element σ ∈ Z 2 (G, I; A).
For the general case of partial cohomology groups relative to ideals, one
is able to come at a weaker form of Theorem 4.3:
Theorem 5.2. Let G be a finite group. Then
H 2 (G, I; Z) ≃ ZmI −nI ⊕ Z/µ1 Z ⊕ · · · ⊕ Z/µnI Z
where mI = |ΩI | ≥ nI = rkZ B 2 (G, I; Z), and µ1 , . . . , µnI are positive integers ordered by recursive division. Moreover,
H 2 (G, I; A) ≃ A ⊗ H 2 (G, I; Z) ≃ AmI −nI ⊕ A/µ1 A ⊕ · · · ⊕ A/µnI A
for any abelian group A.
Proof. The result follows in a way similar to the proof of Theorem 4.3, with
the following minor modification. In this case, using Lemma 5.1, one also
has Z 2 (G, I; A) ≃ A ⊗ Z 2 (G, I; Z). Then, applying Smith’s normalization
theorem, one obtains a free basis f1 , . . . , fmI of Z 2 (G, I; Z) together with
positive integers µ1 , . . . , µnI such that µ1 f1 , . . . , µnI fnI is a free basis of
B 2 (G, I; Z). Finally, the fact B 2 (G, I; A) = µ1 Af1 ⊕ · · · ⊕ µnI Afni is proved
considering the canonical generating set {δhg + ǫI | g ∈ G}.
In particular, when A is a divisible group, then for every ideal I in Λ
each summand A/µi A in the decomposition of H 2 (G, I; A) is trivial. This
observation confirms the conjecture mentioned in the introduction:
Corollary 5.3. If G is finite and K is an algebraically closed field, then any
component of the partial multiplier is a finite direct sum of copies of K× .
5.1
An example
It can be seen in the proof of Theorem 4.3 that the torsion part of pH 2 (G, Z)
comes from the classical cohomology group H 2 (G, Z). A similar statement
can not be formulated for a generic ideal I. Moreover, the next example
shows that the torsion part of H 2 (G, I; Z) is not necessarily a quotient of
H 2 (G, Z). The general procedure to compute H 2 (G, I; Z) explicitly, which
can be reproduced for various small groups using GAP [20], is summarized
18
as follows: one determines m = |Ω| together with a set of representatives
∆(ω1 ), . . . , ∆(ωm ) for the classes of Ω; the matrix
M = (δ1 hgi ◦ ∆(ωj ))i,j ∈ Matn×m (Z)
yields the coefficients of the natural basis δ1 hg1 , . . . , δ1 hgn of B 2 (G, Z) in
terms of the characteristic functions hω1 , . . . , hωm ; given an ideal I in Λ,
one determines Ω \ ΩI = {ωi1 , . . . , ωim−mI } to obtain, removing from A
the columns i1 , . . . , im−mI , a matrix MI ; finally, the coefficients µ1 , . . . , µnI
which describe the torsion part of H 2 (G, I; Z) in Theorem 5.2 are the nonzero diagonal entries of the Smith’s normal form of the matrix MI .
Example 4. Consider the cyclic group G = Z6 . It is directly seen that G
admits 8 = (62 + 2 · 2 + 3 · 1 + 5)/6 classes in Ω, which are represented by
the following pairs:
ω1
ω2
ω3
ω4
ω5
ω6
ω7
ω8
∆ (0, 0) (1, 5) (2, 4) (3, 3) (1, 1) (1, 2) (1, 3) (2, 2)
Thus, the coefficients of the matrix M = (δ1 hgi ◦ ∆(ωj ))i,j ∈ Mat6×8 (Z) can
be read in the following table:
ω1
δh0
δh1
δh2
δh3
δh4
δh5
ω2
ω3
ω4
ω5
ω6
ω7
ω8
1 −1 −1 −1
·
·
·
·
·
1
·
·
2
1
1
·
·
·
1
· −1
1
·
2
·
·
·
2
· −1
1
·
·
·
1
·
·
· −1 −1
·
1
·
·
·
·
·
·
It can be checked that the Smith’s normalization algorithm provides two
matrices P ∈ GL(6, Z) and Q ∈ GL(8, Z), such that the diagonal entries
of D = P M Q are λ1 = · · · = λ5 = 1 and λ6 = 6. In accordance with
Theorem 4.3, then pH 2 (Z6 , Z) ≃ Z ⊕ Z ⊕ Z6 .
Consider now I = N3 ∪ hΠ(1, 2), Π(1, 3)iS , so that Ω \ ΩI = {ω6 , ω7 }.
In particular nI = |ΩI | = 6, and Smith’s normalization for MI provides
a matrix with diagonal entries µ1 = . . . = µ4 = 1, µ5 = 2 and µ6 = 6.
Therefore, H 2 (Z6 , I; Z) ≃ Z2 ⊕ Z6 that, as anticipated, is not a quotient of
H 2 (Z6 , Z) = Z6 .
19
N3 : Z ⊕ Z❋ ⊕
❘ Z6
❋❋❘❘❘❘
❋❋ ❘❘❘
❋❋
❘
❋❋ ❘❘❘❘❘
❘❘❘
❋❋
❘❘❘
❋❋
❘❘
❋
①①
①①
①
①①
①①
①
①①
①①
ω5 : Z ⊕❋Z6
ω6 : Z ⊕❋Z❘6
ω7 : Z ⊕ ◆Z6 ω8 : Z ⊕
✼✼ Z6
◆◆◆❧❧❧
❋❋
❋❋❘❘❘❘❦❦❦❦
♠♠
✂✂
✼✼
❧
❋❋♠♠♠♠♠
❋❦❋❦❦ ❘❘❘
◆
❧
✂
❧ ◆◆
✼
◆◆◆
✂✂
❦❦❦ ❋❋❋ ❘❘❘❘❧❧❧❧❧
♠♠ ❋❋❋
❦
♠
✂
❦
♠
◆◆◆ ✼✼✼
❋❋ ❦❦❦❦
❋❋ ❧❧❧❧ ❘❘❘❘
✂
♠♠
♠
✂
♠
❋
❋
❦
❧
◆
❘
❧ ❋
❘❘❘
◆◆◆ ✼✼✼
❦❦ ❋
✂✂ ♠♠♠
❧❧❧ ❋❋
❘❘❘
❦❦❦ ❋❋
❧
◆◆
❦
❧
✂♠✂♠♠♠♠
❦
❘
❦❦
Z
⊕
Z
Z6 ❁
Z
Z ⊕ Z2
Z6 ❋
❤
2
6
6
❤
✐✐ Z6
❇❇ ❤❤❤❤❤
❋❋
①
❁❁
✂
❤❤❤❤
✐✐✐✐ ✞✞
❤❇
❤
①
✐
❇
❤
❤
✂
✐
❤
❤
❋
①
✐
❤
❤
❁❁ ✂✂
❤
❋❋ ①①
✞
❤❤
❇❇ ✐✐✐✐
❤❤❤❤
❋❋①①❤❤❤❤❤❤❤❤
❁❁✂✂
✞✞
✐✐✐❇✐❇❇
❤❤❤❤
❋
✐
❤
✞
①
❤
✐
❤
❁
❤
✂
❋
❤
✐
❤
①
✞
✂ ❁
❇❇
❤❤❤
❤❤❤① ❋
✐✐✐
✞✞
✂✂ ❁❤❁❁❤❤❤❤❤①❤①①① ❤❤❤❋❤❋❤❋❤❋❤❤
❇❇
✐✐✐✐
✐
✞
✂
✐
❤
✐
❤
❤
✞
✂ ❤❤❤
①❤❤❤❤
✐✐✐✐
ω4 : Z❇6
Z2
Z2 ❤
Z6 ❁◗❤◗◗
❤ Z2
❤
❤
❤
❤
❁❁ ◗◗◗
①①
①①
❤❇❇❤❤❤
❁❁ ◗◗◗◗
①①
①❤①❤❤❤❤❤❤ ❇❇❇
①
①
❤
①
◗◗◗
❁❁
❇❇
①①❤
❤❤❤①❤
◗◗◗ ①①①
❁❁
❇❇
❤❤❤❤①①①
①◗◗◗
❤
❤
①
❁❁
❇❇
❤
❤
①
①
❇❇
❁ ①①①① ❤❤❤❤◗❤◗❤◗❤
①
◗
①
❤
◗
❤
❤
Z3
0
Z2 ❤
ω2 : Z6 ◗
◗◗◗
◗◗◗
◗◗◗
◗◗◗
◗◗◗
◗◗◗
◗◗◗
◗
ω3 : Z2 ❙
Z2 ❁
0
Z3
❙❙❙
❁❁
⑤⑤
❙❙❙
⑤
❁❁
❙❙❙
⑤⑤
❁❁
❙❙❙
⑤⑤
❙❙❙
❁❁
⑤
❙❙❙
⑤⑤
❁❁
❙❙❙
⑤⑤
❁
❙❙❙
⑤
❙❙ ⑤
Z2 ●
●
●●
●●
●●
●●
●●
●●
●
0
✇0
✇✇
✇
✇✇
✇✇
✇
✇
✇✇
✇✇
✇
✇
Figure 1: The partial cohomology semilattice of groups H 2 (Z6 , Λ; Z). The
semilattice structure is the same of (Λ, ∪). At the vertex corresponding to
the ideal I appears H 2 (Z6 , I; Z). The principal-modulo-N3 ideals are marked
by their generator, thus “ω5 : Z⊕Z6 ” stands for H 2 (G, Π∼ (ω5 ); Z) ≃ Z⊕Z6 .
20
For another instance, consider I = N3 ∪hΠ(2, 4), Π(3, 3)iS . Since, by (9),
hΠ(2, 4)iS = hΠ(2)iS = hΠ(4)iS and hΠ(3, 3)iS = hΠ(3)iS , it immediately
follows that Ω \ ΩI = {ω3 , . . . , ω8 }. Thus, only the first and the second
column in the above table are relevant, and one sees that δ1 h5 +ǫI = hω2 +ǫI
and δ1 (h0 + h5 ) + ǫI = hω1 + ǫI . Consequently, H 2 (G, I; Z) = 0 in this case.
The full description of H 2 (Z6 , Λ; Z) is shown in Figure 1.
6
Consistency with the previous notion
It is left to show that for any abelian group of coefficients A and any ideal
I in Λ, the group H 2 (G, I; A) is a partial cohomology group according to
[10]. To this aim, one has to associate the pair (I; A) with a unital partial
G-module B (that is to say, a commutative monoid together with a unital
partial G-action).
Theorem 6.1. For any pair (I; A), there exists a unital partial G-module
B such that H 2 (G, I; A) ≃ H 2 (G, B).
Proof. The construction is similar to that of Lemma 3.2: let E(S(G) \ I)
denote the collection of the idempotents in S(G) \ I, then the set
B = (A × E(S(G) \ I)) ∪ {∞}
endowed with zero element ∞ and multiplication
(
(a + b, εη) if εη ∈
/I
(a, ε) · (b, η) =
∞
otherwise,
results in an A-cancellative central extension of the semilattice E(S(G)/I).
Moreover, B is a commutative monoid with identity element (0A , Π(1)),
and it is naturally endowed with the following unital partial action: for any
g ∈ G, one considers the idempotent
(
(0A , Π(g, g−1 )) if Π(g) ∈
/I
eg =
∞
otherwise
together with the ideal Bg = Beg of B, and the partial action ϑ is given by
the family of isomorphisms {ϑg : Bg−1 → Bg | g ∈ G}, defined by:
(
(a, Π(g)εΠ(g−1 )) if Π(g)εΠ(g−1 ) ∈
/I
ϑg (a, ε) =
∞
otherwise.
21
Thus, for any x1 , . . . , xn ∈ G, the ideal B(x1 ,...,xn ) = Bx1 Bx1 x2 · · · Bx1 x2 ···xn
is either reduced to the trivial semigroup {∞}, or it is generated by the
idempotent
ex1 ex1 x2 · · · ex1 ···xn = 0A , Π(x1 , x2 , . . . , xn , (x1 x2 · · · xn )−1 ) ,
which obviously plays the role of the identity element. In the latter case,
the subgroup of B(x1 ,...,xn ) consisting of invertible elements is
U (B(x1 ,...,xn ) ) = a, Π(x1 , x2 , . . . , xn , (x1 x2 · · · xn )−1 ) | a ∈ A ,
which is isomorphic to A by means of the projection on the first component
π : B \ {∞} → A. It is convenient to extend π to the whole B by the
assignment of π(∞) = ∞. Therefore, an n-cochain σ ∈ C n (G, B), which
is by definition a map σ : Gn → B satisfying σ(x1 , . . . , xn ) ∈ U (B(x1 ,...,xn ) )
for every x1 , . . . , xn ∈ G, can be identified with the composite function
σ̄ = π ◦ σ : Gn → A∞ . Clearly, one has that σ̄(x1 , . . . , xn ) = ∞ if and
only if Π(x1 , . . . , xn , (x1 . . . xn )−1 ) ∈ I, which indeed corresponds to the
case B(x1 ,...,xn ) = {∞}. By the above construction of the partial action of G
over B, the coboundary map δn : C n (G, B) → C n+1 (G, B) is given by the
relation
(
(δn σ̄(x1 , . . . , xn+1 ), η) if η ∈
/I
δn σ(x1 , . . . , xn+1 ) =
∞
otherwise,
where
η = Π(x1 , . . . , xn+1 , (x1 · · · xn+1 )−1 )
and δn : C n (G, A) → C n+1 (G, A) is the standard coboundary homomorphism relative to the the trivial (global) G-module A. Notice that, in
view of (10), the condition η ∈
/ I is equivalent to Π(x1 , . . . , xn+1 ) ∈
/ I.
2
Therefore, taking n = 2, Lemma 5.1 shows that σ ∈ Z (G, B) if and only
if σ̄ ∈ Z 2 (G, I; A). Finally, one checks the case n = 1 to conclude that
σ ∈ B 2 (G, B) if and only if σ̄ ∈ B 2 (G, I; A), so that the map σ → σ̄ induces
the desired isomorphism of cohomology groups.
References
[1] P. Ara and R. Exel. Dynamical systems associated to separated graphs,
graph algebras, and paradoxical decompositions. Advances in Mathematics, 252:748–804, 2014.
22
[2] P. Ara, R. Exel, and T. Katsura. Dynamical systems of type (m, n) and
their C∗ -algebras. Ergodic Theory and Dynamical Systems, 33(5):1291–
1325, 2013.
[3] J. C. Birget and J. Rhodes. Almost finite expansions of arbitrary semigroups. Journal of Pure and Applied Algebra, 32(3):239–287, 1984.
[4] J. C. Birget and J. Rhodes. Group theory via global semigroup theory.
Journal of Algebra, 120(2):284–300, 1989.
[5] K. S. Brown. Cohomology of groups, volume 87 of Graduate Texts in
Mathematics. Springer-Verlag, New York-Berlin, 1982.
[6] H. G. G. de Lima and H. Pinedo. On the total component of the partial Schur multiplier. Journal of the Australian Mathematical Society,
100(3):374–402, 2016.
[7] M. Dokuchaev. Partial actions: a survey. In Groups, algebras and
applications, volume 537 of Contemp. Math., pages 173–184. Amer.
Math. Soc., Providence, RI, 2011.
[8] M. Dokuchaev and R. Exel. Partial actions and subshifts. Journal of
Functional Analysis, 272(12):5038–5106, 2017.
[9] M. Dokuchaev, R. Exel, and J. J. Simón. Crossed products by twisted
partial actions and graded algebras. Journal of Algebra, 320(8):3278–
3310, 2008.
[10] M. Dokuchaev and M. Khrypchenko. Partial cohomology of groups.
Journal of Algebra, 427:142–182, 2015.
[11] M. Dokuchaev and M. Khrypchenko. Twisted partial actions and extensions of semilattices of groups by groups. International Journal of
Algebra and Computation, online-ready, pages 1–47, 2016.
[12] M. Dokuchaev and M. Khrypchenko. Partial cohomology of groups
and extensions of semilattices of abelian groups. Journal of Pure and
Applied Algebra, to appear, arXiv:1705.09654, 2017.
[13] M. Dokuchaev and B. Novikov. Partial projective representations and
partial actions. Journal of Pure and Applied Algebra, 214(3):251–268,
2010.
23
[14] M. Dokuchaev and B. Novikov. Partial projective representations and
partial actions II. Journal of Pure and Applied Algebra, 216(2):438–455,
2012.
[15] M. Dokuchaev, B. Novikov, and H. Pinedo. The partial Schur multiplier
of a group. Journal of Algebra, 392:199–225, 2013.
[16] R. Exel. Twisted partial actions: a classification of regular C ∗ -algebraic
bundles. Proceedings of the London Mathematical Society. Third Series,
74(2):417–443, 1997.
[17] R. Exel. Partial actions of groups and actions of inverse semigroups.
Proceedings of the American Mathematical Society, 126(12):3481–3494,
1998.
[18] L. Fuchs. Infinite abelian groups. Vol. I. Pure and Applied Mathematics, Vol. 36. Academic Press, New York-London, 1970.
[19] L. Fuchs. Infinite abelian groups. Vol. II. Academic Press, New YorkLondon, 1973. Pure and Applied Mathematics. Vol. 36-II.
[20] The GAP Group. GAP – Groups, Algorithms, and Programming, Version 4.8.8, 2017.
[21] J. M. Howie. An introduction to semigroup theory. Academic Press
[Harcourt Brace Jovanovich, Publishers], London-New York, 1976.
L.M.S. Monographs, No. 7.
[22] I. M. Isaacs. Character theory of finite groups. Academic Press [Harcourt Brace Jovanovich, Publishers], New York-London, 1976. Pure
and Applied Mathematics, No. 69.
[23] G. Karpilovsky. The Schur multiplier, volume 2 of London Mathematical Society Monographs. New Series. The Clarendon Press, Oxford
University Press, New York, 1987.
[24] J. Kellendonk and M. V. Lawson. Partial actions of groups. International Journal of Algebra and Computation, 14(1):87–114, 2004.
[25] M. Kennedy and C. Schafhauser. Noncommutative boundaries and the
ideal structure of reduced crossed products, 2017.
24
[26] B. Novikov and H. Pinedo. On components of the partial Schur multiplier. Communications in Algebra, 42(6):2484–2495, 2014.
[27] H. Pinedo. A calculation of the partial Schur multiplier of S3 . Int. J.
Math. Game Theory Algebra, 22(4):405–417, 2014.
[28] H. Pinedo. Partial projective representations and the partial Schur
multiplier: a survey. Bol. Mat., 22(2):167–175, 2015.
[29] H. Pinedo. The total component of the partial Schur multiplier of the
elementary abelian 3-group. Rev. Colombiana Mat., 50(1):75–83, 2016.
[30] J. J. Rotman. Advanced modern algebra. Part 1, volume 165 of Graduate
Studies in Mathematics. American Mathematical Society, Providence,
RI, third edition, 2015.
[31] I. Schur. Über die darstellung der endlichen gruppen durch gebrochen
lineare substitutionen. Journal für die reine und angewandte Mathematik, 127:20–50, 1904.
25
| 4math.GR
|
arXiv:1309.7919v2 [nlin.CD] 11 Mar 2014
CRITICAL TRANSITIONS IN A MODEL OF A GENETIC
REGULATORY SYSTEM
Jesse Berwald
Institute for Mathematics and Its Applications
Minneapolis, Minneapolis, MN 55455, USA
Marian Gidea
Yeshiva University, Department of Mathematical Sciences,
New York, NY 10016
and
School of Mathematics, Institute for Advanced Study
Princeton, NJ 08540, USA
(Communicated by the associate editor name)
Abstract. We consider a model for substrate-depletion oscillations in genetic
systems, based on a stochastic differential equation with a slowly evolving external signal. We show the existence of critical transitions in the system. We
apply two methods to numerically test the synthetic time series generated by
the system for early indicators of critical transitions: a detrended fluctuation
analysis method, and a novel method based on topological data analysis (persistence diagrams).
1. Introduction
Gene expression is the process by which the genetic code is used to synthesize
functional gene products (proteins, functional RNAs). The timing and the level
of gene production is specified by a wide range of mechanisms, termed gene regulation. It is believed that a significant number of genes express cyclically, with
about 10 − 15% of genes directly regulated by the circadian molecular clock. Such
gene expression oscillations allow for rapid adaptation to changes in intracellular
and environmental conditions. In general, the phase and amplitude of gene expression depend on the function of the gene and internal and external stimuli. It has
been postulated that gene expression oscillation is a basic property of all genes, not
necessarily connected with any specific gene function [27].
Gene regulatory systems are intrinsically stochastic. Stochasticity originates in
the statistical uncertainty of the chemical reactions between molecules, and is inversely proportional to the square root of the number of molecules. Thus, lower
numbers of interacting molecules yield increasingly significant statistical fluctuations. Besides intrinsic stochasticity, gene expression is also subjected to extrinsic
1991 Mathematics Subject Classification. Primary: 92D20; 34C23; 55N99. Secondary: 34F05;
57M99.
Key words and phrases. Gene regulatory networks; critical transitions; stochastic differential
equations; persistence diagrams.
The first author is supported by the MCRN. The second author is supported by NSF grant
DMS-1201357, by the National Cancer Institute grant P20 CA157069, by The Fund for Math of
the IAS, and by the MCRN.
1
2
JESSE BERWALD AND MARIAN GIDEA
stochasticity, due to environmental effects. In general, stochastic fluctuations are
seen as a source of robustness and stability, but sometimes can adversely affect a
cell function [22].
A motivation for the modeling and simulation of gene regulatory systems comes
from synthetic biology, an emerging field of research devoted to the design and
construction of biological systems, based on engineering principles. An interesting
analogy in [31] compares a molecular network to an electrical circuit, where, instead
of resistors, capacitors and transistors connected via circuits one has genes, proteins,
and metabolites connected via chemical reactions and molecular pathways. Some
milestone achievements in this direction have been reported, e.g., in [1, 10, 13]. Relatedly, one would also like to achieve a clear understanding of the functionality of
the molecular circuits and networks through measurements of various product outputs, in the same way one understands an electrical circuit through measurements
of current, voltage, and resistance. In particular, one would like to discern possible ways in which systems subjected to varying parameters and noise may switch
between different stable regimes, or more generally, between potential attractors.
It is the concept of suddenly shifts amongst stable regimes which we explore in
this paper. We investigate the occurrence of these critical transitions in a model
of a genetic regulatory system. By a critical transition we mean a sudden change
of a system from one stable regime (fixed point, limit cycle) to an unstable regime,
possibly followed by some other stable regime. We consider a simple genetic circuit
that exhibits an oscillatory regime, and we study the behavior of the system under
noise. Explicitly, we consider a two-gene model whose oscillations depend on several
parameters. We show that the system undergoes a critical transition under slow
parameter drift. We accomplish this by recording the time series generated by
this model and analyzing the critical transitions. We utilize numerical methods to
identify early warning signals indicative of critical transitions in the synthetic data.
First, we apply a statistical method, based on detrended fluctuation analysis,
to analyze these time series data. The method is described in detail in Subsection
4.1. The tests are performed using a windowed analysis of the data and reveal that
the autocorrelation of the time series increases towards 1, and the variance of the
time series distribution grows steadily prior to a critical transition. These signs are
consistent with early warning indicators of critical transitions described by others;
for instance, see results in Scheffer, et al [28].
Secondly, we apply a method from topological data analysis, based on persistence
diagrams, which we describe in more detail in Subsection 4.2. Again, we consider
windows of data from the time series. In this case, we consider these windows as
strings of data points to which we associate filtrations of Rips complexes and for
which we generate associated persistence diagrams to analyze the topology of the
data at different resolutions. The persistence diagrams reveal qualitative changes
in the topology of the strings of data points prior to the critical transitions: the
distribution of data points becomes more widespread and/or asymmetric. While
the detrended fluctuation analysis has been previously used for detection of critical
transitions, the application of persistence diagrams, a method from topological data
analysis, is novel.
We also make a comparison between the two methods. While the detrended
fluctuation analysis introduces artificial choices and possible bias (see also [2]), the
proposed topological method is inherently robust. Note that both methods can
be applied to detect or predict critical transitions in experimental data as well.
CRITICAL TRANSITIONS IN A MODEL OF A GENETIC REGULATORY SYSTEM
3
As such, the results from model testing may serve as benchmarks for testing data
measured from real world sources.
2. Background
In this section we briefly describe critical transitions in the context of fast-slow
systems. We consider both deterministic and stochastic systems. Then we present
a simple genetic regulatory model of the substrate-depletion oscillator type.
2.1. Critical transitions. A fast-slow system of ordinary differential equations is
a system of the type
(1)
x0
=
(2)
0
=
εy
m
m+n
g(x, y),
where (x, y) ∈ R × R , f : R
→ R , g : Rm+n → Rn are C r -functions with
d
. One can regard y as a fast variable.
r ≥ 3, ε > 0 is a small parameter, and 0 = dt
Rescaling the time t = ετ yields
(3)
n
f (x, y),
m
ẋ = εf (x, y),
(4)
ẏ
= g(x, y),
d
dτ .
where ˙ =
The singular limit of (1),(2) when ε → 0 gives the slow subsystem,
and the singular limit of (3),(4) gives the fast subsystem. The critical set is defined
as
C0 = {(x, y) : g(x, y) = 0}
∂g
is
and consists of equilibrium points for the fast subsystem. If the Jacobian ∂y
nonsingular on C0 , then C0 is an m-dimensional manifold, and is the graph of a
smooth function y = h0 (x). The slow subsystem is determined by x0 = f (x, h0 (x))
∂g
at a point have non-zero
and restricts to C0 . In the case that all eigenvalues of ∂y
real parts, then the point is normally hyperbolic. The set of normally hyperbolic
points forms a normally hyperbolic invariant manifold (NHIM). In particular, it has
stable and unstable manifolds. The NHIM can have attractive components, where
∂g
all the eigenvalues of ∂y
have negative real part, and repelling components, where
at least one eigenvalue has positive real part. An attractive component has an ndimensional stable manifold, while a repelling component has a non-trivial unstable
manifold.
Fenichel’s Theorem [11] implies that, for all sufficiently small ε, every compact
submanifold (with boundary) S0 of C0 can be continued to a NHIM Sε (not uniquely
defined) for the flow of (1), (2), which is the graph of a smooth function y = hε (x).
The stable and unstable manifolds of S0 continue to stable and unstable manifolds
of Sε . The flow on Sε converges to the slow flow as ε → 0. Such a manifold Sε is
referred to as a slow manifold.
We define a critical transition for this type of system following [17]. We assume
that the critical set C0 can be decomposed as C0 = S0a ∪ S0r ∪ S0b , where S0a is
an attractive NHIM, S0r a repelling NHIM, and S0b is the part of C0 that is not
normally hyperbolic (corresponding to bifurcation points). By definition, a point
p0 = (x0 , y0 ) on C0 that is not normally hyperbolic is a critical transition if there
exists a concatenation of trajectories γ0 , γ1 , where γ0 : [t−1 , t0 ] → Rm+n , γ1 :
[t0 , t1 ] → Rm+n satisfy the following properties:
(1) γ0 (t−1 , t0 ) is a trajectory of the slow subsystem, oriented from γ0 (t−1 ) to
γ0 (t0 ), contained in the attracting NHIM S0a ;
4
JESSE BERWALD AND MARIAN GIDEA
(2) γ0 (t0 ) = γ1 (t0 ) = p0 ∈ S0b is a point that is not normally hyperbolic;
(3) γ1 (t0 , t1 ) is a trajectory of the fast subsystem, oriented from γ1 (t0 ) to γ1 (t1 ).
When we consider the dynamics of the system for ε > 0 sufficiently small, a
trajectory starting near γ0 follows closely the slow dynamics around γ0 for some
time, after which it transitions to follow the fast dynamics near γ1 for a period of
time. In [17, 18] several types of bifurcations are examined to determine whether
or not they exhibit the characteristics of critical transitions, under some suitable
conditions on the smoothness, compactness, and non-degeneracy on the system. We
summarize the findings below:
(a) for m = n = 1, saddle-node (fold) bifurcation points determine critical
transitions;
(b) for m = n = 1, subcritical pitchfork bifurcation points determine critical
transitions;
(c) for m = n = 1, transcritical bifurcation points determine critical transitions;
(d) for m = 2 and n = 1, subcritical non-degenerate Hopf bifurcations determine critical transitions.
2.2. Incorporating noise. It is often important for the understanding of a physical system to incorporate stochastic effects. We consider the Langevin form of (1),
(2)
(5)
dx
=
(6)
dy
=
f (x, y) + σ1 dW1 ,
1
1
g(x, y) + √ σ2 dW2 ,
ε
ε
where σ1 , σ2 represent noise levels (depending on ε), and W1 , W2 are one-dimensional
Wiener processes (Brownian motions). Assuming σ1 , σ2 are sufficiently small, the
sample paths of the system (5), (6) stay near S0a with high probability, up to a
neighborhood of the critical transition, after which they exit the neighborhood.
In [17, 18], it is argued that the following behaviors are typical of a system prior to
a critical transition:
(i) The system recovery from small perturbations is ‘critically’ slows down;
(ii) The variance in the time series increases steadily;
(iii) The autocorrelation of the time series increases towards 1;
(iv) The distribution of the time series becomes more asymmetric.
We note that (iv) from above depends on whether or not the underlying bifurcation has symmetry. For example, a saddle-node bifurcation as in (a) from above will
typically yield asymmetric fluctuations, while a pitchfork bifurcation as in (b) from
above will typically yield symmetric fluctuations. A complementary characteristic
to (iv) is that the distribution of the time series loses its normality, for example it
changes from uni-modal to multi-modal.
Such system response characteristics can be monitored numerically and serve
as ‘early warnings’ of critical transitions in real-world systems. Examples include
Earth’s climate, ecological systems, global finance, asthma attacks or epileptic
seizures; see, e.g., [7, 28, 29, 30], and the references therein.
2.3. A simple genetic circuit. We briefly describe a model of simple genetic
circuit which generates oscillations of varying amplitude. The model consists of two
genes, one producing the protein R(t), and the other producing the protein X(t).
The protein X(t) is a substrate for the activator protein R(t) that is produced in
an autocatalytic process. As R(t) accumulates, the production of R(t) accelerates
CRITICAL TRANSITIONS IN A MODEL OF A GENETIC REGULATORY SYSTEM
5
until there is an explosive conversion of the whole of X(t) into R(t). This rapid
change corresponds to a critical transition in the underlying system. With the
substrate X(t) depleted, the autocatalytic reaction terminates, and the activator
R(t) degrades in time. This allows the level of X(t) to grow again, leading to
another cycle of explosive growth in R(t). This process is know as a substratedepletion oscillator.
An example of this mechanism is the oscillation of the M-phase-promoting factor
(MPF) activator in the frog egg, where the substrate is the phosphorylated form of
the B-cyclin-dependent kinase (note that the true mechanism involves several other
proteins and reactions) [26]. A mathematical model for the substrate-depletion
oscillator is given by the following system:
(7)
X 0 (t)
=
k1 S − [k00 + k0 EP (R(t))]X(t),
(8)
R0 (t)
=
[k00 + k0 EP (R(t))]X(t) − k2 R(t).
Here EP (R(t)) represents the level of the phosphorylated version of the protein
R(t) – involved with R(t) in a mutual activation process – given by EP (R) =
G(k3 R, k4 , J, K), where G, the Goldbeter-Koshland function, is defined by
G(u, v, J, K) =
v − u + vJ + uK +
2uK
p
.
(v − u + vJ + uK)2 − 4(v − u)uK
The Goldbeter-Koshland function represents the equilibrium concentration of the
phosphorylated form of a protein, for a phosphorylation-dephosphorylation reaction
governed by Michelis-Menten kinetics. The Goldbeter-Koshland function is responsible for creating a switch-like signal-response in the evolution of the protein R(t).
The quantities k0 , k00 , k1 , k2 , k3 , k4 , J, K, S are parameters. The parameter S is the
strength of a signal, representing the rate of synthesis of the substrate X, which we
regard as an external input to the system.
This system presents both positive and negative feedback. The positive feedback
loop creates a bistable system and the negative-feedback loop drives the system
back and forth between two stable steady states. In what follows, we will modify
the simple genetic circuit in (7) and (8) by considering the external input S as a
slowly varying parameter, in addition to including a stochastic term. We will study
critical transitions in the resulting system.
3. Model
We consider a new model of a genetic regulatory network with a slowly dependent
signal, given by (7), (8), with S being now a slowly evolving parameter, i.e.
(9)
X 0 (t)
= k1 S(t) − [k00 + k0 EP (R(t))]X(t),
(10)
R0 (t)
=
(11)
S 0 (t)
= ε,
[k00 + k0 EP (R(t))]X(t) − k2 R(t),
where ε > 0 is small. We will fix the parameter as in [31], k0 = 0.04, k00 = 0.01, k1 =
k2 = k3 = 1, k4 = 0.3, J = K = 0.05.
The fast subsystem is obtained by letting ε → 0 yielding
(12)
X 0 (t)
=
k1 S(t) − [k00 + k0 EP (R(t))]X(t),
(13)
0
R (t)
=
[k00 + k0 EP (R(t))]X(t) − k2 R(t),
(14)
S 0 (t)
=
0.
6
JESSE BERWALD AND MARIAN GIDEA
R
X,R
X(t)
R(t)
X
t
Figure 1. A sample trajectory, and the X-time series (higher
value) and R-time series (lower values).
We rescale time τ = εt, and we rewrite the corresponding system relative to the
rescaled time
(15)
εẊ
= k1 S − [k00 + k0 EP (R)]X,
(16)
εṘ
=
[k00 + k0 EP (R)]X − k2 R,
(17)
Ṡ
=
1.
Sample trajectories for this system are plotted in Fig. 1. From the above equations
the slow subsystem is obtained by letting ε → 0 yielding
(18)
0
= k1 S − [k00 + k0 EP (R)]X,
(19)
0
=
[k00 + k0 EP (R)]X − k2 R,
(20)
Ṡ
=
1.
From this we see that the critical submanifold is given by
)
(
k1
k1 S
,R = S .
C0 = (X, R, S) , X = 0
k2
k0 + k0 EP ( kk21 S)
The critical submanifold consists of equilibrium points of the fast subsystem. The
stability of the C0 is determined by the eigenvalues of the Jacobi matrix evaluated
at the equilibrium points
−[k00 + k0 EP (R)] −[k0 ∂R EP (R)X]
J=
.
k00 + k0 EP (R)
k0 ∂R EP (R)X − k2
For the values of the parameters chosen above, we find that the stability at an
equilibrium point changes at Scrit1 = 0.13326703 and Scrit2 = 0.34680193, respectively. The corresponding equilibria are Xcrit1 = 5.41285587, Rcrit1 = 0.13326703,
and Xcrit2 = 1.07370450, Rcrit2 = 0.34680193.
Precisely, for S < Scrit1 and S > Scrit2 the equilibrium point is stable. For
S ∈ (Scrit1 , Scrit2 ) the equilibrium point is unstable, and there exists a periodic
orbit that is asymptotically stable, whose existence can be established numerically,
as observed in Fig. 1. The values S = Scrit1 , S = Scrit2 yield subcritical Hopf
bifurcations, where an unstable equilibrium point is turned into a stable one and
a small unstable periodic orbit is born (or vice versa). In addition, one expects
canard-type solutions in some exponentially small neighborhoods of Scrit1 , Scrit2 ,
relative to ε (see, e.g., [19]). In Fig. 2 we plot the nullclines of the system for the
critical points Scrit1 and Scrit2 .
CRITICAL TRANSITIONS IN A MODEL OF A GENETIC REGULATORY SYSTEM
7
Figure 2. The nullclines of the fast system for S = Scrit1 and S = Scrit2 .
The specific stochastic differential equation (SDE) system associated to (15),
(16), (17) can be written
(21)
Ẋ
=
(22)
Ṙ
=
(23)
Ṡ
=
1
σ1
(k1 S − [k00 + k0 EP (R)]X) + √ dW1 ,
ε
ε
1 0
σ2
([k + k0 EP (R)]X − k2 R) + √ dW2 ,
ε 0
ε
1,
where W1 , W2 represent Brownian motions, and σ1 , σ2 are noise levels. Since the parameter values Scrit1 and Scrit2 yield subcritical Hopf bifurcations, the theory from
Subsection 2.1 allows us to conclude that the corresponding points (Xcrit1 , Rcrit1 ),
(Xcrit2 , Rcrit2 ) determine critical transitions. A typical trajectory of the stochastic
system in the phase space and its corresponding R-time series are shown in Fig. 3.
Examining the R-time series, one can see that a critical transition occurs near time
t ≈ 1000. We note that a similar analysis for activator-inhibitor oscillations has
been performed in [18].
Remark 1. As mentioned in Section 1, noise in the form of random fluctuations
arises naturally in gene regulatory networks. One typically distinguishes between
intrinsic noise, inherent in the biochemical reactions, and extrinsic noise, originating
in the random variation of the externally set control parameters. Both types of
noise can be model by augmenting the governing rate equations with additive or
multiplicative stochastic terms. We refer the interested reader to [14].
In our model we only consider the effects of additive noise, which can be thought
of as a randomly varying external field acting on the biochemical reactions. The
field enters into the governing rate equations as an additive stochastic term in
the Langevin equation. We choose to focus on additive extrinsic noise as this
could be used as a switch and/or amplifier for gene expression, which has potential
applications to gene therapy [14]. Switching mechanisms are exactly the type of
phenomena that we would like to capture via the critical transitions approach.
4. Methods
We use the model proposed in Section 3 to generate a synthetic time series given
by successive reading of one of the variables. We investigate the synthetic time series
for early warning signs of critical transitions. Below we describe two such detection
8
JESSE BERWALD AND MARIAN GIDEA
Phase plane plot
x
1.4
1.4
R
1.2
1.2
1
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0
0
t
0.2
0
1
2
3
4
5
6
0.2
0
200
400
600
800
1000
1200
1400
1600
1800
2000
Figure 3. Phase space of the model described by (21), (22), (23),
and a corresponding R-time series.
methods: a well-known detrended fluctuation analysis method, and a novel method
inspired by topological data analysis.
4.1. Detrended fluctuation analysis. Detrended fluctuation analysis (DFA) is
a technique introduced by Hurst half a century ago to analyze fluctuations in time
series. The DFA procedure has been widely used for early detection of critical
transitions [23, 28, 30]. We outline the algorithm below.
4.1.1. Algorithmic description of DFA. The DFA procedure takes as input a time
series (sk , zk ), k = 1, . . . , N , where sk is the instant of time of the k-th measurement
(not necessarily equally spaced), and zk is the k-th measurement of some observable.
To detect whether the system undergoes a critical transition, the DFA algorithms
proceeds as follows:
Interpolation: Choose an optimal step size ∆t, and interpolate the given
time series such that it is evenly spaced in time. Denote the new series
(tk , xk ), with tk = k∆t.
Detrending: One way to remove a general trend from statistical data is by
subtracting a moving average. For example, using a Gaussian kernel
1
(t − k∆t)2
Gk (t) = √
exp −
2d2
2πd
of bandwidth d, one may compute the weighted average of xk ,
PN
Gk (i∆t)xi
X(k∆t) = Pi=1
.
N
i=1 Gk (i∆t)
Subtracting the weighted average from the time series yields the detrended
series,
yk = xk − X(k∆t)
.
CRITICAL TRANSITIONS IN A MODEL OF A GENETIC REGULATORY SYSTEM
9
Remark 2. Instead of Gaussian kernel detrending, alternative detrending
methods can be used, such as linear, cubic spline, or Fourier interpolation,
depending on the nature of the data. We also explore these additional
methods in Section 5
Lag-1 autocorrelation: The final step involves fitting a first-order autoregressive (AR(1)) process
yk+1 = ck yk + σξk ,
to the detrended time series yk , where (σξk )k∈N is white noise of intensity
σ. In order to compute ck , choose a sliding window of size w and determine
the least-squares fit
yj+1 ≈ ck yj ,
for j = k, . . . , k + w − 1.
Hence, for each window we extract the value of AR(1) ck . Recall that
the lag-1 autocorrelation AR(1) is 0 for white noise and close to 1 for red
(autocorrelated) noise.
4.1.2. Detection of critical transitions via DFA. The following criterion has been
proposed for the detection of critical transition [28]:
• Given a time series measured from a system approaching a critical transition, the DFA outputs a time series (tk , yk ) for which
(1) the autocorrelation has a general trend which increases towards 1;
(2) the variance has a positive trend.
In certain special cases, this criterion has a rigorous justification [17, 18].
The DFA method is an effective tool in detecting early signs of critical transitions
in noisy data. However, the method comes with several significant drawbacks, such
as its sensitivity to the procedures and parameters used in processing the data.
For instance, the sample frequency, detrending method (e.g., the bandwidth of the
Gaussian detrending), or the size of the sliding window all have a strong effect on the
conclusions drawn from the algorithm in subsection 4.1 and hence on the power of
the DFA method to serve as a prediction tool. One concern is that the measurement
of the AR(1) values as well as the variance are strongly influenced by the fit of the
detrending method, with a poor fit being likely to signal ‘false positives’ for critical
transitions (see [2]).
4.2. Persistence diagrams. As mentioned in Section 1, we propose to use tools
from the field of topological data analysis as a new method to detect critical transitions in dynamical systems. In particular, we leverage the stability properties
of persistence diagrams to detect critical transitions. Topological persistence is
a relatively recent development that forms the core of topological data analysis
and has been widely used to extract relevant information from noisy data (see
[8] for background in persistence topology in general). There are numerous applications, including computer vision, cluster analysis, biological networks, cancer
survival analysis, and granular material (see [4, 8, 12, 20, 25] and the references
listed there).
In this section we describe the way in which we adapt this method to observe
changes in time series from systems approaching or undergoing critical transitions.
The key idea is to extract from the time series consecutive strings of data points
of a fixed length, which we regard as individual point cloud data sets. To each
such point cloud we assign a topological invariant, namely its persistence diagram.
10
JESSE BERWALD AND MARIAN GIDEA
Roughly speaking, the persistence diagram is a representation of the data set in an
abstract metric space which encodes information about topological features of the
data.
The highlight of this method is that when the system undergoes a critical transition, the topological features associated to the point cloud data sets also change
significantly. The fact that the corresponding persistence diagrams and distances
between them can be computed algorithmically enables us to describe these changes
quantitatively.
4.2.1. Description of persistence diagrams. We describe the concept of a persistence
diagram associated with point cloud data starting with an informal description.
From a high-level perspective, the data analysis pipeline works as follows:
Data =⇒ Filtration =⇒ Persistence Module =⇒ Persistence Diagram
We focus on the first two and the fourth parts of this pipeline, and only briefly detail
the algebraic aspects of the third component below. Suppose that one is provided
with a point cloud data set, X0 , that is an approximation of some geometric shape.
One would like to infer from the data the topological information on that shape.
However, a finite collection of points has only trivial topology. One way to convert
the collection of points into a non-trivial topological space is to replace the points
of the set by balls of a certain radius . One then computes the topological features
of the resulting set, X . Typical invariants resulting from this computation, which
serve to classify the set, include the number connected components along with the
number of ‘tunnels’ and number of ‘cavities’ (known as Betti numbers).
Of course, the topology of X depends on the choice of the radius of the balls in
this construction. Instead of fixing a certain radius, topological persistence considers
all possible radii, from some sufficiently small value, up to a sufficiently large radius.
This growth of the radius yields the filtration step above. As the radius is gradually
increased, new topological features will be ‘born’ and certain existing ones will ‘die’.
A schematic representation of this process is depicted in Fig. 4. The birth and death
of each topological feature at a given dimension is recorded by a persistence diagram.
This is a collection (multiset actually) of (birth,death) times in R2 . The 0- and 1dimensional diagrams for the associated filtration are shown in the bottom row of
Fig. 4.
The lifespan of a feature is easily computed by calculating (death time) - (birth
time). A topological feature with a long lifespan, measured by the range of radii
over which it ‘persists’, is likely to capture an essential topological feature of the
underlying space from which the data was sampled. On the other hand, short
lived features are likely to result from ‘noise’ in the data. However, rather than
discriminating between what is an essential feature of the topology and what is not,
the persistence diagram method provides a summary of topological features that
appear and disappear throughout the variation of the radii of the balls, as well as
a ranking of the significance of these features, expressed in terms of the lifespans.
We now continue with a concise, formal description of persistence diagrams. For
an introduction to algebraic topology and homology, see [15]; surveys of persistent
topology can be found in [8, 32] Given point cloud data X ⊂ Rd , i.e., a collection
of points in Rd , and δ > 0, we associate to them the Rips complex Rδ . This is,
by definition, the abstract simplicial complex whose 0-simplices are points xα , and
whose k-simplices are given by unordered (k+1)-tuples of points {xαj }j=0,...,k which
CRITICAL TRANSITIONS IN A MODEL OF A GENETIC REGULATORY SYSTEM
t=1
t=2
t=3
0
1
2
3
0-dimensional
birth
4
0
0
1
1
2
2
3
3
4
death
4
death
t=4
0
1
2
3
birth
4
1-dimensional
Figure 4. A collection of points in the plane, resembling a ‘noisy’
circle, is given. At each instant of time t = 1, . . . , 4, around each
point we constructs disks of radii δ1 /2, . . . , δ4 /2, respectively, with
δ1 < δ2 < δ3 < δ4 . The corresponding Rips complexes are also
constructed at each instant. The topological features of the Rips
complexes change as time increases. At time t = 1 there are 8 connected components and no 1-dimensional hole. At time t = 2 the
connected components coalesce into a single component (thus in
the 0-dimensional diagram there are actually 7 deaths represented
at (1, 2)), and a 1-dimensional hole is born. Both the single connected component and the 1-dimensional hole survive to t = 3. At
time t = 4 the 1-dimensional hole dies as it fills in (the death is represented at (2, 4)), while the single connected component continues
living (in fact, it has infinite lifespan). The ♦ at (1, 4) represents
this fact.
11
12
JESSE BERWALD AND MARIAN GIDEA
are pairwise within a distance δ. Figure 4 provides an example for X ⊂ R2 . For all
0 < a < b we have Ra ⊂ Rb . That is, the family {Rδ }δ>0 forms a filtration.
Denote by Hp (Ra ) the p-homology of Ra with Z2 coefficients. Heuristically,
the homology of a simplicial complex provides information about the topological
features of the complex, e.g., the number of connected components, tunnels and
cavities in that complex. The inclusion Ra ,→ Rb induces the homomorphisms
f∗a,b : H∗ (Ra ) → H∗ (Rb ) in all dimensions. Note that the image F a−ρ,b of f∗a−δ,a
in H∗ (Rb ) is independent of ρ for all ρ > 0 sufficiently small. We denote this image
by F∗a−,b .
A real value c > 0 is called a homological critical value if there exists q such
that the homomorphism fqc−ρ,c : Hq (Rc−ρ ) → Hq (Rc ) is not an isomorphism for all
sufficiently small ρ > 0. The image Fqc−,c of fqc−ρ,c in Hq (Rc ) is independent of ρ,
if this is small enough. The quotient group Bqc = Hq (Rc )/Fqc−,c is called the q-th
birth group at Rc , and it captures the homology classes that did not exist in Rc−ρ .
A homology class α ∈ Hq (Rc ) is born in Rc if it represents a non-trivial element in
Bqc , that is, the canonical projection of α is non-zero.
Now consider the homomorphism gqa,b : Bqa → Hq (Rb )/F a−,b , where gqa,b ([α]) =
a,b
[fq (α)], for α ∈ Hq (Rb ), where the notation [·] stands for equivalence class. We
set gqa,b = 0 for all b > 0 sufficiently large. The kernel Dqa,b of the map gqa,b is called
the death subgroup of Bqa at Rbq . A homology class α ∈ Hq (Ra ) dies entering Rb if
[α] ∈ Dqa,b but [α] 6∈ Dqa,b−ρ , for ρ > 0 sufficiently small. The degree r of the death
value b of Bqa is defined by r = rankDqa,b − rankDqa,b− . The sum of the degrees of
all death value of the birth group Bqa is clearly equal to rank(Bqa ). The birth time
of a homology class α is the value a > 0 where the α is born in Ra , and the death
time is the value b > 0 where α dies in Rb .
The q-persistence diagram of the filtration (Rδ )δ>0 is defined as a multiset Pq in
R2 consisting of points of the type zi = (a, bi ), where a is a birth value corresponding
to a non-trivial group Bpa , and bi is a death value of Bpa ; the point zi appears in
the diagram with multiplicity equal to the degree ri of the death value bi . Since
deaths occur after births, all points (a, bi ) lie above the diagonal set of R2 . By
default, the diagonal set of R2 is part of the persistence diagram, representing all
trivial homology generators that are born and die at every level. Each point on
the diagonal has infinite multiplicity. The axes of the persistence diagram are birth
values on the horizontal axis and death values on the vertical axis. Again, see Figure
4 for a schematic representation of the construction of a persistence diagram.
It is convenient to define a metric on the space of persistence diagrams. A number
of options exist. A fairly standard metric is the p-Wasserstein metric. On the set of
the q-persistence diagrams consider the p-Wasserstein metric, 1 ≤ p ≤ ∞, defined
by
1/p
X
dp (Pq1 , Pq2 ) = inf
kz − φ(z)kp∞ ,
φ
z∈Pq1
where Pq1 , Pq2 are two q-persistence diagrams, and the sum is taken over all bijections
φ : Pq1 → Pq2 . The set of bijections, {φ : Pq1 → Pq2 }, is nonempty owing to the fact
that each diagram includes the diagonal set, allowing one to match off-diagonal
elements in one diagram with diagonal elements in another when their numbers
differ.
CRITICAL TRANSITIONS IN A MODEL OF A GENETIC REGULATORY SYSTEM
13
The space of q-persistence diagrams together with the p-Wasserstein metric forms
a metric space, which is complete and separable. The Wasserstein distance takes the
‘best’ matching; that is, it minimizes the distance, relative to the Lp norm, that one
has to shift generators in Pq1 to match them with those in Pq2 . In probability theory,
and in particular cases with continuous or weighted distributions, the Wasserstein
metric is sometimes termed the ‘earth mover distance’, which refers to the operation
of transforming one distribution into another with the minimal change in mass. In
what follows we set p = 2 and drop the reference to p. For details on the Wasserstein
metric, see [6].
One of the remarkable properties of persistence diagrams is their stability, meaning that small changes in the initial point cloud data produce persistence diagrams
that are close to one another relative to Wasserstein metric. The stability results
are very general for the ‘bottleneck distance’, when p = ∞, and more restrictive for
the Wasserstein metric with p < ∞. The essence of the stability results, as shown
in [3, 6, 9], is that the persistence diagrams depend Lipschitz-continuously on point
cloud data.
In applications, the stability result ensures the robustness of the data analysis
performed via persistence diagrams, which makes them a powerful alternative to
statistical methods. This is particularly useful in context of data from stochastic
systems, since persistence diagrams turn out to be quite versatile in distinguishing
between small but relevant features in a data set and noise.
4.2.2. Detection of critical transitions via persistence diagrams. We now describe
how to apply this method to detect critical transitions in time series. Consider
a time series (tj , xj ), j = 1, . . . , J, with xj ∈ Rd . (In the case of a time series
obtained from the model discussed in Section 3, we will chose d = 1). Assume that
the time series (tj , xj ) is obtained as a time discretization of a process (t, xt ) which
is Lipschitz continuous. (This is indeed the case when a time series is obtained
from a Langevin equation, as in Subsection 3.) To each ti we associate a string of
N consecutive data points points from the time series, with N sufficiently large,
which we denote
ti 7→ Xi = (xi , xi+1 , . . . , xi+N −1 ).
(24)
We regard each Xi as a point cloud set in Rd . We compute the persistence diagrams
P∗ (Xi ) of Xi , in all dimensions, and follow the evolution of the persistence diagrams
in time.
Diagrams corresponding to nearby times will be close to one another, due to the
Lipschitz continuity of the process underlying the time series and to the robustness
of persistence diagrams. Within this context, persistence diagrams that are near to
each other in time, but relatively far from one another in the Wasserstein metric,
indicate a sudden change in the time series. Therefore, we propose the following
empirical criterion for detection of critical transitions in slow-fast systems:
•
Persistence diagrams undergo significant changes, measured using the Wasserstein metric, prior to a critical transition.
This criterion follows from the following heuristic argument. If the noise level in
the Langevin equation is small, then far from a critical transition the time series
follows closely, with high probability, a trajectory of the slow subsystem. A point
cloud associated to a data string displays significant topological features similar to
those of the slow manifold, plus less significant topological features due to noise. In
14
JESSE BERWALD AND MARIAN GIDEA
0.3
X
0.25
0.2
0.15
0.1
0.05
0
t
0.05
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
Figure 5. R-time series for the gene regulatory network model
from Section 3 cut off before the critical transition.
addition, the corresponding persistence diagrams at nearby times are close to one
another relative to the Wasserstein metric.
When the system undergoes a critical transition, the time series ceases to follow
the slow manifold, as the dynamics enters a transient regime. The topological features associated to the slow manifold are destroyed, and new topological features
appear in the point cloud structure. Furthermore, if the system moves to a different stable regime after a finite time, the point cloud will reflect the topological
features associated to that regime. Critically, for a point cloud data near a critical
transition the corresponding persistence diagrams shift away from those diagrams
corresponding to data far from the critical transition. Consequently, successive
distances between diagrams in this region exhibit a large jump prior to a critical
transition.
5. Results
We numerically solve the SDE defined in (21) – (23) using the Euler-Maruyama
procedure, with stepsize 0.01 and noise level σ1 = σ2 = 0.02. We fix the rate of
change of the parameter S to be ε = 10−4 . As output, we choose the time series
given by the R-component; a particular realization of this time series is shown in
the righthand panel of Fig. 3. Since the solution values are dense in time, we
subsample the time series by taking every 10-th data point. The time series follows
a slowly varying attractive equilibrium point, until it reaches a critical transition,
at which point it enters an oscillatory mode. To test for early signs of the critical
transition, we truncate the time series before it enters the oscillatory regime. This
truncated region is shown in Fig. 5.
5.1. DFA analysis. We conduct three experiments to detect critical transitions
using the DFA1 methodology on the time series generated by our model. The
first experiment uses a Gaussian kernel to detrend the time series; in the second
experiment we use a cubic spline interpolation; and in the third experiment, we
use Fourier interpolation. For each of the detrended time series we compute AR(1)
and the variance for a windowed time series. The results are summarized in Fig.
6. All three experiments show AR(1) increases to 1 as the system approaches the
1The DFA procedure that we use here was implemented in Matlab by Rebecca M. Jones [16]
CRITICAL TRANSITIONS IN A MODEL OF A GENETIC REGULATORY SYSTEM
15
transition, while the variance also grows steadily, both behaviors being consistent
with a critical transition.
5.2. Persistence diagram analysis. We compute the 0-dimensional persistence
diagrams for strings of N data points Xi (see Eq. (24)) as the ti approaches a critical
transition, as for the example time series in Fig. 5. We construct Rips complexes,
with an initial radius of δ = 10−4 around each data point, and then grow the radii
of the balls δ, 2δ, . . . , nδ, where n is chosen large enough so that the final complex
in the resulting filtration has a single connected component. From this filtration we
compute the 0-persistence diagrams. We choose the size of the data sets N = 300;
larger sizes make little difference in the qualitative behavior of the diagrams.
The 0-dimensional persistence diagrams are easy to interpret: they track of the
births and deaths of connected components in the Rips complex, as the radii of
the balls increase. Note that all births occur at the same time, when the radius
of balls is zero and we have N disjoint points. After this initial stage, a large
number of connected components die as the radii of the balls increase, as connected
components merge with one another.
Consider Fig. 7, where the persistence diagrams for a data string far from the
critical transition (left panel) and a data string close to the critical transition (right
panel) are shown. For data far from the critical transition, points cluster near
the attractive equilibrium. Thus, when balls are constructed in the Rips filtration
around these points, they will quickly yield a robust connected component around
the attractive equilibrium , plus a small number of scattered connected components
corresponding to points that escape for brief periods time from the equilibrium point
due to stochastic effects. The implication is that, in the corresponding persistence
diagram, the vertical spread of the death times is relatively small, and consists
of a small numbers of points away from the diagonal (accounting for the robust
connected component and a few outliers), plus many short-lived points close to the
diagonal (accounting for noise).
Conversely, when a data string originates from close to a critical transition, the
points tend to spread further away from the attractive equilibrium point, due to
changes in the potential field. Heuristically, the equilibrium loses its attractiveness.
This causes the distribution of the data points from the time series to grow. The
implication is that, in the corresponding persistence diagram, the vertical spread of
the death times is much larger, with a tendency to form multiple small clusters.
The visual inspection of persistence diagrams provides intuition, but is not a
precise way to indicate the approach to a critical transition. To quantify the above
assessment, we study the behavior of the Wasserstein distances between consecutive
diagrams which are summarized in Fig. 8. In the figure, time increases from left to
right. The figures in the top row represent persistence diagrams for data sampled far
from the critical transition, and those in the middle represent persistence diagrams
for data sampled close to the critical transition. The five time frames captured in
each column spread over a time interval of size ∆t = 0.5. We then compute the
Wasserstein distances between consecutive persistence diagrams. These changes in
the persistence diagrams are quantified in the bottom row of Fig. 8. The solid curve,
corresponding to data near the critical transition, shows a significant increase in the
distances between consecutive diagrams as the point cloud anayzed near the critical
transition. The dotted curve, corresponding to distances between diagrams far from
the critical transition, shows only small variations in the consecutive distances.
16
JESSE BERWALD AND MARIAN GIDEA
Time series and its trend
Detrended time series
0.2
Autocorrelation coefficient
0.1
x 10
0.08
4
Variance
4
0.986
3.8
0.984
0.15
0.06
3.6
0.982
0.04
3.4
0.98
0.1
0.02
3.2
0.978
0
3
0.05
0.976
0.02
2.8
0.974
0.04
2.6
0
0.972
0.06
0.05
0.08
0
5000
10000
Time series and its cubic interpolation
0.3
2.4
2.2
0.97
0
5000
10000
Detrended time series using spline method
0.15
0
2000
4000
6000
8000
Autocorrelation coefficient
0.974
0
7
x 10
2000
4000
6000
Variance
8000
2000
4000
6000
Variance
8000
4
0.972
0.25
0.1
6.5
0.97
0.2
0.05
0.968
0.15
0
6
0.966
5.5
0.1
0.05
0.05
0.1
0
0.15
0.964
0.962
5
0.96
4.5
0.958
0.05
0
5000
10000
Time series and its Fourier interpolation
0.3
0.25
0.2
0
5000
10000
Detrended time series using FFT method
0.15
0.956
0
2000
4000
6000
8000
Autocorrelation coefficient
0.975
4
7
0.1
0
x 10
4
6.5
0.97
0.2
0.05
6
0.965
0.15
0
5.5
0.1
0.05
0.96
5
0.05
0.1
0
0.15
0.955
0.05
0
5000
10000
0.2
4.5
0
5000
10000
0.95
0
2000
4000
6000
8000
4
0
2000
4000
Figure 6. DFA analysis of the time series using (top to bottom)
Gaussian kernel, cubic spline, and Fourier interpolation detrending.
6000
8000
CRITICAL TRANSITIONS IN A MODEL OF A GENETIC REGULATORY SYSTEM
17
50
20
40
death
death
15
30
10
20
5
10
0
0
0
5
10
15
20
0
birth
10
20
30
birth
40
50
Figure 7. The 0-persistence diagram on the left corresponds to a
string of data sampled far from the critical transition; the diagram
on the right corresponds to a string of data sampled near the critical
transition. In each persistence diagram, the dots mark finite deaths
times, and the diamond indicates infinite death times.
The computed distance are indicated by the symbols on each curve and are placed
between the diagrams from which they were computed.2
Note that the variance of the lifespans is related to the variance in the time
series. An increase in the vertical spread while approaching a critical transition is
consistent with the findings by the DFA method in Subsection 5.1. Also, the change
observed in the clustering of the diagram coordinates is related to asymmetric or
multimodal properties of the data, as mentioned in Subsection 4.2.
6. Conclusions
The substrate-depletion oscillator that we analyze in this paper is a realistic
model for certain types of molecular regulation circuits studied experimentally. The
methods for detecting critical transitions that we propose are suitable for the analysis of real data as well. Indeed, most of the experimental data obtained about gene
regulatory networks (e.g., data obtained from microarrays, or reverse transcriptase
polymerase chain reaction) is limited by background noise, and both the DFA and
persistence diagram methods are robust to noise in data, as long as the noise does
not overwhelm the signal. Also, in comparison with the DFA method, which necessitates a number of ‘ad-hoc’ choices of statistical parameters and procedures, the
persistence diagram method appears more robust and objective.
In current and future work, we are developing a theoretical framework for the
empirical criterion proposed in this paper. Namely, we plan to establish rigorously
that bifurcation-induced critical transitions determine large changes in the persistence diagrams, and, conversely, large changes in the persistence diagrams imply
the existence of bifurcations.
2For the computation of persistence diagrams we use the Perseus software developed by Vidit
Nanda [24]. The Wasserstein distances were computed using software written by Miroslav Kramar [21].
18
JESSE BERWALD AND MARIAN GIDEA
Figure 8. The 0-persistence diagrams on the top row correspond
to consecutive strings of data sampled far from the critical transition; the diagrams in the middle row correspond to consecutive
strings of data sampled near the critical transition. The bottom
curves describe the distances between consecutive diagrams. The
plot labels are positioned between two successive diagrams with
values on the y-axis indicating the 2-Wasserstein distances between
them.
Acknowledgement
A portion of this work has been done while M.G. was a member of the IAS, to
which he is very grateful. Also, we thank Konstantin Mischaikow, Miroslav Kramar,
Vidit Nanda, and Rebecca M. Jones for many useful discussions on this subject.
References
[1] T. Bulter, S.-G. Lee, W.-W. Wong, E. Fung, M.R. Connor, and J.C. Liao, Design of
artificial cell-cell communication using gene and metabolic networks, Proc. Natl. Acad.
Sci. USA, 101 (2004), 2299–2304.
[2] R.M. Bryce and K.B. Sprague, Revisiting detrended fluctuation analysis, Scientific
Reports 2, (2012), doi:10.1038/srep00315
[3] F. Chazal, D. Cohen-Steiner, L. J. Guibas, F. Mémoli, S. Oudot, Gromov-Hausdorff
Stable Signatures for Shapes using Persistence, Computer Graphics Forum (proc. SGP
2009) (2009), 1393–1403.
[4] F. Chazal, L.Guibas, S. Oudot, and P. Skraba, Persistence-Based Clustering in Riemannian Manifolds, Proc. 27th Annual ACM Symposium on Computational Geometry,
(2011), 97–106.
[5] F. Chazal, V. de Silva, S. Oudot, Persistence stability for geometric complexes, Geometriae Dedicata, (2013), doi: 10.1007/s10711-013-9937-z
[6] D. Cohen-Steiner, H. Edelsbrunner, J. Harer, Y. Mileyko. Lipschitz Functions Have
Lp -Stable Persistence. Foundations of Computational Mathematics. Vol. 10 (2010),
doi: 10.1007/s10208-010-9060-6
CRITICAL TRANSITIONS IN A MODEL OF A GENETIC REGULATORY SYSTEM
19
[7] P.D. Ditlevsen and S.J. Johnsen, Tipping points: Early warning and wishful thinking,
Geophysical Research Letters, Vol. 37, L19703 (2010), doi:10.1029/2010GL044486
[8] H. Edelsbrunner and J. Harer, Persistent homology — a survey, in “Surveys on Discrete
and Computational Geometry. Twenty Years Later”, 257-282, eds. J. E. Goodman, J.
Pach and R. Pollack, Contemporary Mathematics 453, Amer. Math. Soc., Providence,
Rhode Island, (2008).
[9] H. Edelsbrunner and M. Morozov, Persistent Homology: Theory and Applications,
Proceedings of the European Congress of Mathematics, (2012).
[10] M. Elowitz and S.Leibler, A Synthetic Oscillatory Network of Transcriptional Regulators, Nature, Vol. 403, (2000), 335–338.
[11] N. Fenichel, Geometric singular perturbation theory for ordinary differential equations,
Journal of Differential Equations, 31 (1979), 53–98.
[12] M. Gameiro, Y. Hiraoka, S. Izumi, M. Kramar, K. Mischaikow, and V. Nanda, A
Topological Measurement of Protein Compressibility, preprint, (2013).
[13] T.S Gardner, C.R. Cantor, and J.J. Collins, Construction of a genetic toggle switch in
Escherichia coli, Nature, Vol. 403 (2000), 339–342.
[14] J. Hasty, J. Pradines, M. Dolnik, and J.J. Collins, Noise-based switches and amplifiers
for gene expression, PNAS 97 (2000), 2075–2080.
[15] A. Hatcher,Algebraic Topology. Cambridge University Press, (2002).
[16] R.M.
Jones,
Matlab
code
for
the
DFA
procedure,
http://criticaltransitions.wikispot.org/ (2013).
[17] C. Kuehn, A mathematical framework for critical transitions: Bifurcations, fast-slow
systems and stochastic dynamics, Physica D: Nonlinear Phenomena, 240(12) (2011),
1020-1035.
[18] C. Kuehn, A Mathematical Framework for Critical Transitions: Normal Forms, Variance and Applications, Journal of Nonlinear Science, DOI 10.1007/s00332-012-9158-x
[19] M. Krupa, and P. Szmolyan, Extending geometric singular perturbation theory to nonhyperbolic points -fold and canard points in two dimensions, SIAM J. of Math. Anal.
33 (2001), 286– 314.
[20] L. Kondic, A. Goullet, C.S. O’Hern, M. Kramar, K.Mischaikow, R.P. Behringer, Topology of force networks in compressed granular media, Europhys. Lett. 97 (2012), 54001.
[21] M. Kramar, C++ code to compute the Wasserstein metric, Personal communication,
http://www.math.rutgers.edu/∼miroslav (2013).
[22] A. Leier, P.D. Kuo, W. Banzhaf, and K. Burrage, Evolving noisy oscillatory dynamics in
genetic regulatory networks, In EuroGP’06 Proceedings of the 9th European conference
on Genetic Programming, LNCS 3905 (2006), 290–299.
[23] V.N. Livina and T. M. Lenton, A modified method for detecting incipient bifurcations
in a dynamical system, Geophys. Res. Lett., 34, (2007), L03712.
[24] V. Nanda, The Perseus Software Project for Rapid Computation of Persistent Homology, http://www.math.rutgers.edu/˜vidit/perseus.html
[25] M. Nicolau, A.J. Levine, and G. Carlsson, Topology based data analysis identifies a
subgroup of breast cancers with a unique mutational profile and excellent survival,
PNAS,17 (2011)
[26] B. Novak, and J.J. Tyson, Numerical analysis of a comprehensive model of M-phase
control in Xenopus oocyte extracts and intact embryos, J. Cell. Sci., 106 (1993), 1153–
1168.
[27] A.A. Ptitsyn, S. Zvonic, and J.M. Gimble, Digital Signal Processing Reveals Circadian
Baseline Oscillation in Majority of Mammalian Genes, PLoS Comput Biol 3 (2007),
e120.
[28] M. Scheffer et al., Early-warning signals for critical transitions, Nature, 461 (2009),
53–59.
[29] M. Scheffer et al, Anticipating Critical Transitions, Science 338, (2012) 344, DOI:
10.1126/science.1225244
[30] J.M.T. Thompson, and J. Sieber, Climate tipping as a noisy bifurcation: a predictive
technique, IMA Journal of Applied Mathematics (2010), 1–20.
[31] J.J. Tyson, K.C. Chen, and B. Novak, Sniffers, buzzers, toggles and blinkers: dynamics
of regulatory and signaling pathways in the cell, Curr. Opin. Cell. Biol. 15(2), (2003),
221–231.
[32] A.J. Zomorodian, Topology for Computing, Cambridge University, (2005)
20
JESSE BERWALD AND MARIAN GIDEA
E-mail address: [email protected]
E-mail address: [email protected]
| 5cs.CE
|
Semi-Automatic Construction of a Domain Ontology for Wind
Energy Using Wikipedia Articles
Dilek Küçüka,∗, Yusuf Arslanb
arXiv:1410.8581v1 [cs.CL] 30 Oct 2014
b
a
Power Electronics Department, TÜBİTAK Energy Institute, 06531, Ankara, Turkey
Department of Computer Engineering, Middle East Technical University, 06531, Ankara, Turkey
Abstract
Domain ontologies are important information sources for knowledge-based systems. Yet,
building domain ontologies from scratch is known to be a very labor-intensive process. In
this study, we present our semi-automatic approach to building an ontology for the domain
of wind energy which is an important type of renewable energy with a growing share in
electricity generation all over the world. Related Wikipedia articles are first processed in
an automated manner to determine the basic concepts of the domain together with their
properties and next the concepts, properties, and relationships are organized to arrive at
the ultimate ontology. We also provide pointers to other engineering ontologies which could
be utilized together with the proposed wind energy ontology in addition to its prospective
application areas. The current study is significant as, to the best of our knowledge, it
proposes the first considerably wide-coverage ontology for the wind energy domain and the
ontology is built through a semi-automatic process which makes use of the related Web
resources, thereby reducing the overall cost of the ontology building process.
Keywords: wind energy, wind power, domain ontology, text mining, Web mining
Corresponding author. Tel: +90-312-2101310; fax: +90-312-2101315
Email addresses: [email protected] (Dilek Küçük), [email protected]
(Yusuf Arslan)
∗
Preprint submitted to Elsevier
November 3, 2014
1. Introduction
Renewable energy resources (such as wind, solar, and biomass, among others) have
emerged as an important engineering topic mainly due to the profound advantages of the
corresponding plants over conventional plants (such as fossil-fuelled and nuclear). Considering wind energy, the primary benefits of the wind energy plants are that (i) energy
production does not lead to environmental pollution and (ii) as wind is abundant in many
countries, the financial costs of these plants are comparably lower than that of the conventional plants [1]. Compared to the other renewable resources, wind is considerably variable
and hence it is hard to determine optimal places for wind power plant installations and to
forecast daily energy productions of the plants precisely. The latter point (the uncertainty
of the production) results in particular problems during the integration of these plants to
the electrical grid as the electricity transmission system operators should be informed of
prospective daily productions in advance to make short-term planning and to ensure that
the supply matches the demand [1]. Other directions of research on the topic include predictions for the economic viability of wind power plant projects and wind turbine design
[2]. As wind energy keeps increasing its share in the electricity generation (and annually
growing at a rate of 30% [3]), research projects targeting at these problems have started
to increase (such as [4]).
In this paper, we target at wind power applications from a knowledge-based perspective
and propose a wide-coverage domain ontology for wind energy domain1 . A domain ontology
1
In this study, we utilize the terms wind power and wind energy interchangeably to denote the electrical
power generated from wind through the wind turbines installed on the wind power plants (as they are
commonly used interchangeably and they denote wind generated power), although the terms power and
energy are quite distinct terms in physics.
2
is defined as a reusable vocabulary of concepts, relationships, and activities in the domain
along with the governing theories and principles [5]. Various domain ontologies have been
reported in the literature [6, 7, 8, 9, 10] and there is ongoing research on the topic along
with research on the Semantic Web. Although domain ontologies serve various purposes,
building them from scratch is too labor-intensive and time-consuming. Hence, we employ
a semi-automatic approach to build our wind energy ontology which comprises two main
phases: in the first phase, Wikipedia [11] articles on the topic are processed to learn the
concepts from highly frequent words and phrases (n-grams) and in the second phase, these
concepts are organized with their properties and relationships to arrive at the ultimate
ontology. Yet, the final form of the ontology is open to extensions as well, that is, other
concepts and properties covering more details about the domain can be added to the
ontology. It can also be integrated with other related engineering ontologies for better
coverage of the domain.
Ontology-based studies on wind power domain are rather limited. We come across
only three studies on the topic, the first of which [12] proposes an ontology on wind power
plant information and the resulting ontology comprises some generic concepts which are
not specific to wind power domain and moreover it basically aims to cover concepts related
to the management of wind power plant information. The second study [13] describes an
ontology for wind turbines condition monitoring and is more domain-specific compared to
the ontology proposed in [12]. Yet this study [13] also suffers from low coverage, i.e., it only
considers the semantics of wind turbines instead of the whole wind power domain. The
last related study [14] utilizes an ontology of facts to be utilized by a multi-agent system
which aims to control a photovoltaic (PV)-based microgrid. This ontology covers specific
3
facts or concepts related to agent operations which are far from representing the domain
of renewable energy or wind energy. Hence, to the best of our knowledge, in the current
study, we propose the first large-scale ontology for the wind power domain. Yet, as will
be emphasized in Section 3, this ontology can well be integrated with related engineering
ontologies and aligned with the aforementioned ontologies including [12, 13].
The rest of the paper is organized as follows: In Section 2, the ontology building process
for wind energy domain is presented in details. Section 3 is devoted to prospective integration opportunities with related engineering ontologies and plausible application areas
of the proposed ontology. Finally, Section 4 concludes the paper and provides pointers to
future work.
2. Building an Ontology for Wind Energy
Our semi-automatic ontology construction process for wind energy domain comprises
two main phases:
• Learning the concepts and properties of these concepts from related Wikipedia articles by extracting the frequent phrases as unigrams, bigrams, and trigrams from the
article texts. This phase executes in fully automated mode.
• Building the ontology by organizing these components and properties utilizing other
written information sources like related international standards and textbooks.
We provide the execution flow of the overall approach with more detailed steps in Figure
1. In the figure, the resources (input and output) are shown as rectangular shapes, while
the functional steps are presented as round rectangles (fully automated steps are shown
with darker color than the manual steps). At the end of this execution flow, we realize the
4
ontology on Protégé [15] which is an ontology editor and a knowledge-base framework. The
details of the aforementioned two phases are described in the rest of this section basically
following the functional steps in Figure 1.
Figure 1: Execution Flow of the Employed Ontology Building Approach.
5
2.1. Learning Ontology Concepts from the Web
Ontology learning is defined as the task of automatic or semi-automatic knowledge discovery from different data sources and representation of the ultimate ontology [16]. Within
the course of the current study, we consider Wikipedia [11] articles as our data source for
learning an ontology for the wind energy domain. Wikipedia is a large and valuable source
of semantic information. It has been commonly pointed out in the literature that it can
serve various tasks including information extraction, information retrieval, question answering systems, and ontology building [17, 18]. Therefore, we choose the Wikipedia article for
the term wind power [19] as our initial point and utilized this article and linked articles as
the information source to facilitate the ontology building process for wind energy domain.
The approach followed is very much like that of the Web search engines which traverse the
links in Web pages in a breadth-first manner, yet, we only consider the articles linked to
the initial article (links in the first level), i.e., we do not process the articles linked to the
linked articles of the initial article.
Table 1: Frequent N-grams in the Initial Wikipedia Article for Wind Power (with Frequencies in Parentheses).
Unigrams
wind (346)
power (149)
energy (135)
turbine (65)
capacity (54)
electricity (50)
grid (36)
Bigrams
wind power (82)
wind energy (51)
wind turbine (45)
wind farm (39)
renewable energy (22)
capacity factor (16)
offshore wind (15)
Trigrams
offshore wind power (8)
wind power capacity (4)
wind power industry (3)
meteorology and climatology (2)
wind turbine engineering (2)
electric power transmission (2)
wind generated power (2)
Our overall ontology learning procedure for wind energy domain from Wikipedia articles
is carried out as follows: first, the text of the initial Wikipedia article is obtained and
stopwords in the text are eliminated. After stopword elimination, named entities (such as
6
person, location, and organization names) as well as numeric expressions are also removed
from the text. Next, unigrams, bigrams, and trigrams are obtained with their frequencies
from the resulting text. Especially highly frequent phrases are then manually examined
to determine whether they are relevant to the target ontology or not. Relevant phrases
are utilized as concepts and concept properties of the ontology. The same procedure is
employed for the linked Wikipedia articles and again relevant phrases contribute to the
ontology. To better illustrate the process, the highly frequent sample phrases in the text
of the initial Wikipedia article are provided in Table 1 together with their frequencies in
parentheses as sorted by their frequencies.
2.2. Organizing the Ontology
Apart from the frequent phrases determined in the steps described in the previous
section, we have also examined several related textbooks of the domain [2, 1] and the set of
wind turbine standards by International Electrotechnical Commission (IEC) [20] to arrive
at a wide-coverage (yet extensible) domain ontology for wind power domain. At this stage,
the relevant phrases extracted from the Web are complemented with those extracted from
the aforementioned written resources and the resulting information is organized to form
the ultimate ontology. The taxonomy of the concepts in the final form of the proposed
domain ontology is presented in Figure 2. The ontology is built utilizing Protégé [15, 21]
as the ontology editor and the hierarchy of the ontology classes is presented in Figure 3 as
snapshots.
7
Wind Power Plant
(Wind Farm)
is-a
has
generates
causes
utilizes
measures
controls
Blade
Gearbox
Hub
Rotor
Generator
H-axis
Turbine
Wind Turbine
(Wind Turbine
Generator)
Meteorological
Tower
Data
Logger
Humidity
Sensor
Sensor
Pressure
Sensor
8
Temperature
Wind
Profiler
Solar
Radiation
Sensor
Solar
Radiation
DAQ
System
Monitoring and
Control System
Forecast
System
Turbine
Status
Wind
Vane
Propeller
Anemometer
Measured
Power
Wind
Speed
LIDAR
Sensor
Meteorological
Data
Wind
Turbulence
Wind
Direction
RADAR
Sensor
Power
Quality
Wind Power
(Wind Energy)
Sonic
Anemometer
SODAR
Sensor
Analysis
System
Numerical
Weather
Prediction
Physical
Model
Statistical
Model
Cup
Anemometer
Temperature
Sensor
Control
System
Anemometer
Humidity
Pressure
V-axis
Turbine
Tower
Nacelle
Vertical Wind
Component
Horizontal Wind
Component
Wind
Shear
U-component
V-component
Figure 2: The Domain Ontology for Wind Energy.
Forecast
Power
Figure 3: The Class Hierarchy of the Wind Energy Ontology on Protégé [15].
At the center of the ontology is the Wind Power Plant class which has the major
components represented by the classes of Meteorological Tower, Wind Turbine, Monitoring and Control System, and Forecast System. The Meteorological Tower class in turn
comprises Data Logger and Sensor classes, where the latter class has several subclasses,
namely, Humidity Sensor, Pressure Sensor, Temperature Sensor, Solar Radiation Sensor,
Wind Profiler, Wind Vane, and Anemometer. Among these classes, the latter three of
them represent devices for measuring wind properties and are therefore more significant
9
for the overall wind energy ontology. To clarify the corresponding devices, a wind profiler is a device for measuring wind speed and direction, while wind vanes only measure
wind direction and anemometers only measure wind speed. Anemometers usually come in
three basic types represented in our ontology by the classes of Cup Anemometer, Propeller
Anemometer, and Sonic Anemometer. All of the Sensor subclasses in the proposed wind
energy ontology represent devices for measuring distinct types of data on meteorological
phenomena which are represented by the Meteorological Data class. The Wind class has
the components of Wind Speed, Wind Direction, Vertical Wind Component, Horizontal
Wind Component (which in turn has U-component and V-component as components),
Wind Shear and Turbulence. The names of the first four of these classes are fairly selfexplanatory. Considering the last two classes, wind shear is the change of wind speed as
a function of height while turbulence is defined to be the standard deviation of horizontal
wind speed, vertical wind speed, and wind direction as 10-min averages [1].
The second significant component of a wind farm is a set of wind turbines (i.e., wind
turbine generators) represented with the Wind Turbine class. The turbines usually fall into
one of the two types: horizontal axis turbines (modeled with H-axis Turbine) and vertical
axis turbines (modeled with V-axis Turbine class). Main components of a generic wind
turbine are rotor, nacelle, and tower which are modeled with the corresponding classes of
Rotor, Nacelle, and Tower, respectively. The rotor of a wind turbine is usually composed of
the blades (modeled by the Blade class) and the rotor hub (represented with the Hub class)
which connects the blades to the shaft [1]. Nacelle is the part of the turbine that includes
the main parts apart from those included in the rotor system such gearbox, generator, and
brake [1] where we only include the Gearbox and the Generator classes in our ontology as
10
parts of the Nacelle class. Depending on the requirements of the application in which the
wind energy ontology will be utilized, the parts of the Nacelle class can be extended and
more classes to represent these parts can be added to the ontology. The last part of a wind
turbine is the structural part represented with the Tower class. Wind turbines generate
wind power represented with the Wind Power class in our ontology.
The last two parts of a typical wind power plant are the monitoring and control system
(modeled with Monitoring and Control System class) and the forecast system (modeled
with Forecast System class). The former system has several components including a control system, a data acquisition (DAQ) system, and an analysis system denoted by the
classes of Control System, DAQ System, and Analysis System, respectively. This system is
used to monitor and measure turbine status information, as well as generated power and
electrical power quality, which are represented by Turbine Status, Measured Power, and
Power Quality classes, respectively. The power quality is actually another related domain
which is not modeled in our ontology in details but as will be emphasized in the following
section, an ontology for this domain can be integrated into the current ontology to cover
this power quality concept and related concepts adequately. The Forecast System class represents the system which utilizes numerical weather predictions together with previously
generated physical and statistical models of the wind power plants to produce forecasts
regarding the wind generated power. Hence, related classes with self-explanatory names
include Numerical Weather Prediction, Physical Model, Statistical Model, and finally Forecast Power. Although some forecast systems may also output forecasts for wind speed and
direction, we do not consider that case here as the ultimate aim of the system is usually
power forecast.
11
Wind Power Plant, which is the central class of our wind energy ontology, has the
synonymous names of Wind Farm, Wind Plant, and Wind Energy Plant. It is also commonly abbreviated as WPP. As other concepts of the domain also have several synonyms,
we have added a synonymSet property for applicable concepts to hold the synonyms of
the actual class names corresponding to these concepts. Abbreviations are also treated as
synonyms and this property is used to hold abbreviations as well. We provide synonyms
of some of the concepts of the ontology in Table 2 for illustrative purposes. As shown in
the table, WPP is added as a synonym of the Wind Power Plant class. Similarly, WTG
is the abbreviation for Wind Turbine Generator which in turn is a synonym of the Wind
Turbine class, hence WTG is also added to the synonym set of this class.
Table 2: Synonyms of Some of the Concepts in the Wind Energy Ontology.
Ontology Class
Wind Power Plant
Wind Power
Wind Turbine
Wind Vane
Synonyms
WPP, Wind Plant, Wind Energy Plant, Wind Farm
Wind Power Generation, Energy, Generation,
Wind Energy, Wind Generation, Wind Generated Power,
Electricity Production
Turbine, Wind Turbine Generator, WTG, Generator
Weather Vane, Weather Cock
As we have previously pointed out, properties and relations are also specified within the
ultimate wind energy ontology. While properties are mostly obtained from the extracted
keywords from Wikipedia articles (where the procedure is outlined in the previous subsection), relations are manually formed considering the interrelations between the semantic
concepts (corresponding to the classes in Figure 2). Plausible properties of some of the
classes are provided in Table 3 for illustrative purposes. It should be noted that properties
may also have some synonymous names. For instance, installed capacity property of Wind
Power Plant class has rated capacity, nominal capacity, maximum effect, power capacity,
12
nameplate capacity, and wind power capacity as its synonyms. Similarly, utilisation rate is
a synonym for the capacity factor property of the same class2 . Therefore such information
is also modeled in the ontology through the synonymSet property of the applicable concept
properties. Among the properties of Wind Power Plant class in Table 3, voltage level corresponds to that level through which the plant is integrated to the electrical grid and the
names of the remaining ones are fairly self-explanatory. The properties of Wind Turbine
mostly model the technical characteristics of the turbines while the properties of the Wind
Speed class aim to cover the actual speed, the height at which the speed is measured or
forecasted, the date of measurement or forecast, and the plant where the speed is measured
or forecasted.
Table 3: Properties of Some of the Concepts of the Wind Energy Ontology.
Ontology Class
Wind Power Plant
Wind Turbine
Wind Speed
Properties
name, owner, location, license date,
number of turbines, installed capacity, capacity factor,
voltage level
model, hub height, swept area, rated power,
power curve, cut-in wind speed, cut-out wind speed,
rotor diameter, number of blades
speed, height, date, wind power plant
Considering the relations among the concepts in the ontology, there are basically seven
distinct relation types which are outlined below:
• is-a: This type denotes the subclass-class relation between the corresponding concepts in the ontology. For instance, is-a relation exists between the Anemometer and
Sensor classes to specify that an anemometer is a kind of sensor.
2
Installed capacity of a power plant is defined as the “intended technical fullload sustained output” [19]
of the corresponding plant while capacity factor is usually defined as the ratio of generated power to the
installed capacity.
13
• has: This relation corresponds to the part-of information between concepts. It exists
between the Wind Power Plant and Wind Turbine classes as wind power plants have
a number of turbines.
• generates: This relation type and the remaining types denote functional relations
with quite self-explanatory names. Hence, we only provide examples from the ontology in Figure 2 to illustrate the relations, hence, this generates relation can be
observed between Wind Turbine and Wind Power classes of the ontology.
• causes: This type of relation is illustrated between Wind and Wind Power classes
as depicted in Figure 2.
• utilizes: This relation exists between the classes of Wind Turbine and Wind as wind
turbines utilize wind to generate wind power.
• measures: This relation type can be observed between the Anemometer and Wind
Speed classes in the ontology.
• controls: This relation exists between Control System and Wind Power Plant classes
in the wind energy ontology.
It should be noted that we do not claim to cover all meteorological phenomena and
other wind related concepts in our ontology and as it will be clarified in the upcoming
section, relevant ontologies on the topic can be integrated into the wind power ontology to
increase its coverage.
The final form of the proposed ontology is made available for research purposes at
http://www.ceng.metu.edu.tr/~e120329/wont.owl as a Web Ontology Language (OWL)
file.
14
3. Prospective Integration with Related Ontologies and Application Areas
The proposed wind energy ontology has some concepts, particularly Power Quality
and Meteorological Data, corresponding to distinct but related domains which may be
represented in a more detailed manner. But, instead of extending the proposed wind
energy ontology with the concepts for these domains from scratch, ontologies for these two
domains can be integrated into the wind energy ontology. For the domain of meteorological
data, the ontology covered in [22], which is an ontology for atmospheric science, can be
integrated into the wind energy ontology. For the domain of electrical power quality,
PQONT domain ontology [9] has been proposed and it can readily be integrated into the
wind power ontology as it is publicly available as an OWL file. Another plausible option is
to extend the proposed ontology to make it a larger renewable energy ontology, considering
the semantics of other renewable energy types like solar and biomass. With this prospective
extension and aforementioned integration opportunities with the related domain ontologies,
the wind energy ontology can be turned into a wider coverage ontology and hence can serve
a wider range of applications.
There are several significant application areas of domain ontologies. Basically, they
are known to be of particular importance for semantic information extraction, information
retrieval, and question answering systems, among others. Below, we outline the plausible application areas in which the wind energy ontology that we have proposed can be
employed:
• The ontology could act as a shared vocabulary of concepts to ensure the interoperability of the various wind energy applications like monitoring and forecast systems.
The proposed ontology could also aid in reducing the software analysis and design
15
costs during the development of the aforementioned systems as it acts as a viable
source of semantic information for the underlying domain of wind energy.
• The ontology can be integrated into a domain-specific information extraction system
to automatically extract significant pieces of information on wind energy and related
topics from free natural language texts. Utilizing such an information extraction
system, the proposed ontology could be populated with relevant instances. Similar
to the previous point, the ontology could be employed as part of an information retrieval system to determine the index terms to be used during retrieval or as part of
a question answering system to determine relevant answers from relevant documents
for the specified questions. A text categorization system can also benefit from the
proposed ontology to detect documents related to wind energy (or, renewable energy)
in a large set of text documents. The ontology could also be employed in a natural
language interface over the populated ontology where queries in natural language can
be transformed into valid ontology queries making use of the concept names, properties, and their synonyms to pinpoint the queried information. This final application
is exemplified in [9] where a natural language interface, based on the PQONT domain
ontology for electrical power quality, is described where the interface basically utilizes the ontology to determine the concepts specified in the natural language query
expressions..
4. Conclusion
There is a growing interest in wind energy which is a significant type of renewable
energy with a growing share in the electricity production all over the world. Accord-
16
ingly, research on topics related to wind energy, including monitoring and forecast of wind
power, has gained considerable attention. In this study, we deal with the wind energy domain from a knowledge-based perspective and propose the first considerably wide-coverage
domain ontology for wind energy. The ontology building process is a learning-based semiautomated procedure where in the first phase plausible keywords to be utilized during the
determination of the concepts and their properties are learned from the Web. Namely, related Wikipedia articles are processed to determine the relevant keywords and next related
textbooks and international standards are examined to organize these keywords to arrive
at the ultimate ontology. Hence, we decrease the ontology building costs considerably by
employing a learning procedure and utilizing Web as a source of semantic information.
It should be noted that the proposed ontology can be extended or customized to better
satisfy the needs of prospective applications in which the ontology will be employed. The
final form of the proposed ontology is also made publicly available for research purposes.
Plausible directions of future research include the integration of the ontology with
related engineering ontologies such as those for atmospheric science and electrical power
quality. It can also be extended into a more comprehensive renewable energy ontology.
Moreover, the ontology can be employed in knowledge-based systems such as information
extractors, information retrieval, question answering, and text categorization systems in
addition to natural language interfaces, to assess its practical contribution to such systems.
References
[1] P. Jain, Wind Energy Engineering, 1st Edition, McGraw-Hill, 2011.
17
[2] T. Burton, D. Sharpe, N. Jenkins, E. Bossanyi, Wind Energy Handbook, 1st Edition,
John Wiley & Sons, 2001.
[3] Renewable energy - Wikipedia, http://en.wikipedia.org/wiki/Renewable_energy.
[4] Wind
Power
Monitoring
and
Forecast
System
(RİTM),
http://www.ritm.gov.tr/en/root/index.php.
[5] A. Gómez-Pérez, M. Fernández-López, O. Corcho, Ontological Engineering, 3rd Edition, Springer-Verlag, 2004.
[6] R. Stevens, C. Goble, I. Horrocks, S. Bechhofer, Building a bioinformatics ontology
using OIL, IEEE Transactions on Information Technology in Biomedicine 6 (2) (2002)
135–141.
[7] D. Gasević, V. Devedzić, Petri net ontology, Knowledge-Based Systems 19 (4) (2006)
220–234.
[8] J. Morbach, A. Yang, W. Marquardt, Ontocape-a large-scale ontology for chemical
process engineering, Engineering Applications of Artificial Intelligence 20 (2) (2007)
147–161.
[9] D. Küçük, O. Salor, T. İnan, I. Çadırcı, M. Ermiş, PQONT: A domain ontology for
electrical power quality, Advanced Engineering Informatics 24 (2010) 84–95.
[10] C. Zhang, C. Cao, Y. Sui, X. Wu, A chinese time ontology for the Semantic Web,
Knowledge-Based Systems 24 (7) (2011) 1057–1074.
[11] Wiki - Wikipedia, http://en.wikipedia.org/wiki/Wiki.
18
[12] Y. Zhu, X. Wang, D. Cheng, Ontology-based research on wind power plant information interaction, in: Proceedings of the Workshop on Knowledge Discovery and Data
Mining, 2008.
[13] P. Papadopoulos, L. Cipcigan, Wind turbines’ condition monitoring: an ontology
model, in: Proceedings of the International Conference on Sustainable Power Generation and Supply, 2009.
[14] M. Pipattanasomporn, H. Feroze, S. Rahman, Securing critical loads in a pv-based
microgrid with a multi-agent system, Renewable Energy 39 (1) (2012) 166–174.
[15] The
Protégé
Ontology
Editor
and
Knowledge
Acquisition
System,
http://protege.stanford.edu.
[16] L. Drumond, R. Girardi, A survey of ontology learning procedures, in: Proceedings of
the 3rd Workshop on Ontologies and Their Applications, 2008.
[17] O. Medelyan, D. Milne, C. Legg, I. H. Witten, Mining meaning from Wikipedia,
International Journal of Human-Computer Studies 67 (9) (2009) 716–754.
[18] S. Ferrández, A. Toral, O. Ferrández, A. Ferrández, R. Munoz, Exploiting Wikipedia
and EuroWordNet to solve cross-lingual question answering, Information Sciences
179 (20) (2009) 3473–3488.
[19] Wind power - Wikipedia, http://en.wikipedia.org/wiki/Wind_power.
[20] IEC 61400 - Wikipedia, http://en.wikipedia.org/wiki/IEC_61400.
[21] N. F. Noy, M. Sintek, S. Decler, M. Crubezy, R. W. Fergerson, M. A. Musen, Creating
19
Semantic Web contents with Protege-2000, IEEE Intelligent Systems 16 (2) (2001)
60–71.
[22] R. Ramachandran, S. Movva, S. Graves, S. Tanner, Ontology-based semantic search
tool for atmospheric science, in: Proceedings of the 22nd International Conference
on Interactive Information Processing Systems for Meteorology, Oceanography, and
Hydrology, 2006.
20
| 5cs.CE
|
Publications of the Astronomical Society of Australia (PASA)
c Astronomical Society of Australia 2015; published by Cambridge University Press.
doi: 10.1017/pas.2015.xxx.
arXiv:1501.05992v1 [astro-ph.IM] 24 Jan 2015
The Murchison Widefield Array Correlator
S. M. Ord4,20 , B. Crosse4 , D. Emrich4 , D. Pallot4 , R. B. Wayth4,20 , M. A. Clark23,5,22 , S. E. Tremblay4,20 ,
W. Arcus4 , D. Barnes25 , M. Bell14 , G. Bernardi21,24,5 , N. D. R. Bhat,4 , J. D. Bowman1 , F. Briggs2 , J. D. Bunton3 ,
R. J. Cappallo6 , B. E. Corey6 , A. A. Deshpande9 , L. deSouza3,14 , A. Ewell-Wice7 , L. Feng7 , R. Goeke7 ,
L. J. Greenhill5 , B. J. Hazelton16 , D. Herne4 , J. N. Hewitt7 , L. Hindson19 , H. Hurley-Walker4 , D. Jacobs1 ,
M. Johnston-Hollitt19 , D. L. Kaplan18 , J. C. Kasper13,5 , B. B. Kincaid6 , R. Koenig3 , E. Kratzenberg6 ,
N. Kudryavtseva4 , E. Lenc,14,20 C. J. Lonsdale6 , M. J. Lynch4 , B. McKinley2 , S. R. McWhirter6 , D. A. Mitchell3,20 ,
M. F. Morales16 , E. Morgan7 , D. Oberoi11 , A. Offringa,2,20 , J. Pathikulangara3 , B. Pindor12 , T. Prabu9 ,
P. Procopio12 , R. A. Remillard7 , J. Riding12 , A. E. E. Rogers6 , A. Roshi8 , J. E. Salah6 , R. J. Sault12 ,
N. Udaya Shankar9 , K. S. Srivani9 , J. Stevens3 , R. Subrahmanyan9,20 , S. J. Tingay4,20 , M. Waterson4,2 ,
R. L. Webster12,20 , A. R. Whitney6 , A. Williams4 , C. L. Williams7 , J. S. B. Wyithe12,20
1
Arizona State University, Tempe, USA
The Australian National University, Canberra, Australia
3
CSIRO Astronomy and Space Science, Australia
4
International Centre for Radio Astronomy Research (ICRAR), Curtin University, Perth, Australia
5
Harvard-Smithsonian Center for Astrophysics, Cambridge, USA
6
MIT Haystack Observatory, Westford, MA, USA
7
MIT Kavli Institute for Astrophysics and Space Research, Cambridge, USA
8
National Radio Astronomy Observatory, Charlottesville, USA
9
Raman Research Institute, Bangalore, India
10
Swinburne University of Technology, Melbourne, Australia
11
National Center for Radio Astrophysics, Pune, India
12
The University of Melbourne, Melbourne, Australia
13
University of Michigan, Ann Arbor, USA
14
University of Sydney, Sydney, Australia
15
University of Tasmania, Hobart, Australia
16
University of Washington, Seattle, USA
17
University of Western Australia, Perth, Australia
18
University of Wisconsin–Milwaukee, Milwaukee, USA
19
Victoria University of Wellington, New Zealand
20
ARC Centre of Excellence for All-sky Astrophysics (CAASTRO)
21
Square Kilometre Array South Africa (SKA SA), Cape Town, South Africa
22
California Institute of Technology, California, USA
23
NVIDIA, Santa Clara, California, USA
24
Department of Physics and Electronics, Rhodes University, Grahamstown, South Africa
25
Monash University, Melbourne, Australia
2
Abstract
The Murchison Widefield Array (MWA) is a Square Kilometre Array (SKA) Precursor. The telescope
is located at the Murchison Radio–astronomy Observatory (MRO) in Western Australia (WA). The MWA
consists of 4096 dipoles arranged into 128 dual polarisation aperture arrays forming a connected element
interferometer that cross-correlates signals from all 256 inputs. A hybrid approach to the correlation task is
employed, with some processing stages being performed by bespoke hardware, based on Field Programmable
Gate Arrays (FPGAs), and others by Graphics Processing Units (GPUs) housed in general purpose rack
mounted servers. The correlation capability required is approximately 8 TFLOPS (Tera FLoating point
Operations Per Second). The MWA has commenced operations and the correlator is generating 8.3 TB/day
of correlation products, that are subsequently transferred 700 km from the MRO to Perth (WA) in real-time
for storage and offline processing. In this paper we outline the1correlator design, signal path, and processing
elements and present the data format for the internal and external interfaces.
Keywords: instrumentation: interferometers, techniques: interferometric
2
1 Introduction
The MWA is a 128 element dual polarisation interferometer, each element is a 4x4 array of analog beam
formed dipole antennas. The antennas of each array are
arranged in a regular grid approximately 1m apart, and
these small aperture arrays are known as tiles. The science goals that have driven the MWA design and development process are discussed in the instrument description papers(Tingay et al. 2013a; Lonsdale et al. 2009),
and the MWA science paper (Bowman et al. 2013).
These are (1) the detection of redshifted 21cm neutral
hydrogen from the Epoch of Re-ionization (EoR); (2)
Galactic and extra-Galactic surveys; (3) time-domain
astrophysics; (4) solar, heliospheric and ionospheric science and space weather.
1.1 Specific MWA Correlator Requirements
The requirements and science goals have driven the
MWA into a compact configuration of 128 dualpolarisation tiles. 50 tiles are concentrated in the 100m diameter core, with 62 tiles distributed within 750-m
and the remaining 16 distributed up to 1.5-km from the
core.
The combination of the low operating frequency of
the MWA and its compact configuration allow the correlator to be greatly reduced in complexity, however
this trade-off does drive the correlator specifications.
Traditional correlators are required to compensate for
the changing geometry of the array with respect to the
source in order to permit coherent integration of the
correlated products. In the case of the MWA, no such
corrections are performed. This drives the temporal resolution specifications of the correlator and forces the
products to be rapidly generated in order to maintain
coherence. Tingay et al. (2013a) list the system parameters and these include a temporal resolution of 0.5 s.
The temporal decoherence introduced by integrating
for 0.5 s without correcting for the changing array geometry; and the requirement to image a large fraction
of the primary beam, drive the system channel resolution specification to 10 kHz. It should be noted that this
does not drive correlator complexity as it is total processed bandwidth that is the performance driver, not
the number of channels. Albeit the number of output
channels does drive the data storage and archiving specifications. The MWA correlator is required to process
the full bandwidth as presented by the current MWA
digital receivers, which is 30.72 MHz per polarisation.
In summary, in order to meet the MWA science requirements, the correlator is required to correlate 128
dual polarisation inputs, each input is 3072x10kHz in
bandwidth. The correlator must present products, integrated for no more than 0.5 seconds at this native
channel resolution, to the data archive for storage.
PASA (2015)
doi:10.1017/pas.2015.xxx
Ord et al.
1.2 Benefits of Software Correlator
Implementations
The correlation task has previously been addressed by
Application Specific Integrated Circuits (ASICs) and
Field Programmable Gate Arrays (FPGAs). However
the current generation of low frequency arrays including
the MWA (Wayth et al. 2009), LOFAR (van Haarlem
et al. 2013), PAPER (Parsons et al. 2010) and LEDA
(Kocz et al. 2014) have chosen to utilise, to varying degrees, general purpose computing assets to perform this
operation. The MWA leverages two technologies to perform the correlation task, a Fourier transformation performed by a purpose built FPGA-based solution, and a
cross-multiply and accumulate task (XMAC) utilising
the xGPU library1 described by Clark et al. (2012), also
used by the PAPER telescope and the LEDA project.
The MWA system was deployed over 2012/13 as described by Tingay et al. (2013a), and is now operational.
The MWA is unique amongst the recent low frequency imaging dipole arrays, in that it was not designed to utilise a software correlator at the outset. The
flexibility provided by the application of general purpose, GPU based, software solution has allowed a correlator to be rapidly developed and fielded. This was
required in-order to respond to a changed funding environment, which resulted in a significant design evolution from that initially proposed by Lonsdale et al.
(2009) to that described by Tingay et al. (2013a). Expeditious correlation development and deployment was
possible because General Purpose GPU (GPGPU) computing provides a compute capability, comparable to
that available from the largest FPGAs, in a form that
is much more accessible to software developers. It is
true that GPU solutions are more power intensive than
FPGA or ASIC solutions, but their compute capability is already packaged in a form factor that is readily
available and the development cycle is identical to other
software projects. Furthermore the GPU processor lifecycle is very fast and it is also generally very simple
to benefit from architecture improvements. For example, the GPU kernel used as the cross-multiply engine
in the MWA correlator will run on any GPU released
after 2010. We could directly swap out the GPU in the
current MWA cluster, replace them with cards from the
Kepler series (K20X), and realise a factor of 2.5 increase
in performance (and a threefold improvement in power
efficiency).
The organisation of this paper begins with a short
introduction to the correlation problem, followed by an
outline of the MWA correlator design and then a description of the sub-elements of the correlator following
the signal path; we finally discuss the relevance of this
correlator design to the Square Kilometre Array and
1
xGPU
available
correlators/xGPU
at:
https://github.com/GPU-
The MWA Correlator
3
there are several Appendices describing the various internal and external interfaces.
2 The Correlation Problem
A traditional telescope has a filled aperture, where a
surface or a lens is used to focus incoming radiation to
a focal point. In contrast, in an interferometric array like
the MWA the purpose of a correlator is to measure the
level of signal correlation between all antenna pairs at
different frequencies across the observing band. These
products can then be added coherently, and phased, or
focussed, to obtain a measurement of the sky brightness
distribution in any direction
The result of the correlation operation is commonly
called a visibility and is a representation of the measured signal power (brightness) from the sky on angular scales commensurate with the distance between the
constituent pair of antennas. Visibilities generated between antennas that are relatively far apart measure
power on smaller angular scales and vice versa. When
all visibilities are calculated from all pairs of antennas in
the array many spatial scales are sampled (see Thompson et al. (2001) for an extensive review of interferometry and synthesis imaging). When formed as a function
of observing frequency (ν), this visibility set forms the
cross-power spectrum. For any two voltage time-series
from any two antennas V1 and V2 this product can be
formed in two ways. First the cross correlation as a function of lag, τ , can be found, typically by using a delay
line and multipliers to form the lag correlation between
the time-series.
Z ∞
(V1 ? V2 )(τ ) =
V1 (t)V2 (t − τ )dt.
(1)
−∞
The cross power spectrum, S(ν), is then obtained by
application of a Fourier transform:
Z ∞
S12 (ν) =
(V1 ? V2 )(τ )e−2πiντ dτ.
(2)
−∞
When the tasks required to form the cross power spectrum are performed in this order (lag cross-correlation,
followed by Fourier transform) the combined operation
is considered an XF correlator. However the cross correlation analogue of the convolution theorem allows Equation 2 to be written as the product of the Fourier transform of the voltage time series from each antenna:
Z ∞
Z ∞
S12 (ν) =
V1 (t)e−2πiνt dt ×
V2 (t)e−2πiνt dt.
−∞
−∞
(3)
Implemented as described by Equation 3 the operation
is an FX correlator. For large-N telescopes the FX correlator has a large computational advantage. In an XF
correlator for an array of N inputs the cross correlation
for all baselines requires O(N 2 ) operations for every lag,
and there is a one to one correspondence between lags
PASA (2015)
doi:10.1017/pas.2015.xxx
and output channels, F, resulting in O(F N 2 ) operations
to generate the full set of lags. The Fourier transform requires a further O(F log2 F ) operations, but this can be
performed after averaging the lag spectrum and is therefore inconsequential. For the FX correlator we require
O(N F log2 F ) operations for the Fourier transform of
all input data streams, but only O(N 2 ) operations per
sample for the cross multiply (although we have F channels the sample rate is now lower by the same factor).
Therefore as long as N is greater than log2 F there is a
computational advantage in implementing an FX correlator. XF correlators have been historically favoured
by the astronomy community, at least in real-time applications, as until very recently N has been small, and
there are disadvantages to the FX implementation. The
predominant disadvantage is data growth: the precision of the output from the Fourier transform is generally larger than the input, resulting in a data rate
increase. There is also the complexity of implementing
the Fourier transform in real-time.
3 The MWA Correlator System
The tasks detailed in this section cover the full signal path after digitisation, including fine channelisation,
data distribution, correlation and output. In the MWA
correlator the F stage is performed by a dedicated channeliser, subsets of frequency channels from all antennas
are then distributed to a cluster of processing nodes via
an ethernet switch. The correlation products are then
assembled and distributed to an archiving system. An
outline of the system as a whole is shown in Figures
1 and 2. As shown in Figure 3 the correlator is conceptually composed of 4 sub-packages, the Polyphase
Filterbank (PFB), that performs the fine channelisation, the Voltage Capture System (VCS), responsible
for converting the data transport protocol into ethernet
and distributing the data, the correlator itself (XMAC),
and the output buffering that provides the final interface to the archive. Figure 2 presents the physical layout
of the system.
3.1 The System Level View
The computer system consists of 40 servers, grouped
into different tasks and connected by a 10GbE data
network and several 1GbE monitoring, command and
control networks. The correlator computer system has
been designed for ease of maintenance and reliability.
All of the servers are configured as thin-clients and
have no physical hard-disk storage that is utilised for
any critical functions - any hard-disk is used for noncritical storage and could be removed (or fail) without
limiting correlator operation. All of the servers receive
their Operating System through a process known as the
4
Ord et al.
REC 1
VCS 1
V
REC 2
32 MWA
ANTENNAS
VCS 2
V
PFB 1
REC 3
VCS 3
V
VCS 4
V
REC 4
4 BLOCKS (2 NOT SHOWN)
SWITCH
REC 13
VCS 13
V
REC 14
32 MWA
ANTENNAS
VCS 14
V
PFB 4
REC 15
VCS 15
V
REC 16
128 Antennas
16 digital
g
receivers
(Digitisation and first
g
stage
Fourier transform)
XMAC
PRO
PROCESSING
AND
A
R
ARCHIVING
10 GbE Switch for
cross connect
24 Servers
Final corner turn
bit promotion
p
GPU-based XMAC
Local and Remote
Archiving
VCS 16
V
FPGA
Polyphase
yp
p
Filter
Banks
(Second stage Fourier
transform)
16 Servers
Rocket-IO to 10 GbE
Protocol Conversion
+ Demultiplex
Figure 2. The full physical layout of the MWA digital signal processing path. The initial digitalisation and coarse channelisation is
performed in the field. The fine channelisation, data distribution and correlation is performed in the computer facility at the Murchison
Radio Observatory, the data products are finally archived at the Pawsey Centre in Perth.
PXE (Pre-eXecution Environment) boot process, and
import all of their software over NFS (Network File
System) and mount it locally in memory. The servers
are grouped by task into the VCS (voltage capture system) machines that house the FPGA capture card and
the servers that actually perform the cross-multiplyaccumulate (XMAC) and output the visibility sets for
archiving. Any machine can be replaced without any
more effort than is required to physically connect the
server box and update an entry in the IP address mapping tables.
3.2 The MWA Data Receivers: Input to the
Correlator
As described in Prabu et al. (2014) there are 16 digital
receivers deployed across the MWA site. Power and fiPASA (2015)
doi:10.1017/pas.2015.xxx
bre is reticulated to them as described by Tingay et al.
(2013a). The individual tiles are connected via coaxial
cable to the receivers, the receivers are connected via
fibre runs of varying lengths to the central processing
building of the MRO, where the rest of the signal processing is located.
3.2.1 The Polyphase Filterbank
Polyphase Filter Banks (PFBs) are commonly used in
digital signal communications as a realisation of a Discrete Fourier Transform (DFT; see Crochiere & Rabiner
(1983) for a detailed explanation). It specifically refers
to the formation of a DFT via a method of decimating input samples into a polyphase filter structure and
forming the DFT via the application of a Fast Fourier
Transform to the output of the polyphase filter (Bunton
2003).
The MWA Correlator
5
Figure 3. The decomposition of the MWA Correlator system demonstrating the relationship between the major sub-elements.
Figure 1. The MWA signal processing path, following the MWA
data receivers, in the form of a flow diagram.
As described by Prabu et al. (2014) the first or
coarse PFB is as an 8 tap, 512 point, critically sampled,
polyphase filter bank. This is implemented as; an input
filtering stage realised by 8x512 point Kaiser windowing function, the output of which is subsampled into a
PASA (2015)
doi:10.1017/pas.2015.xxx
512 point FFT. This process transforms the 327.68 MHz
input bandwidth into 256, 1.28 MHz wide, subbands.
The second stage of the F is performed by four dedicated FPGA based polyphase filter bank boards housed
in two ATCA (Advanced Telecommunications Computing Architecture) racks, which implement a 12 tap, 128
point filter response weighted DFT. Because the first
stage is critically sampled the fine channels that are
formed at the boundary of the coarse channels are corrupted by aliasing. These channels are processed by the
pipeline, but removed during post-processing, the fraction of the band excised due to aliasing is approximately
12%. The second stage is also critically sampled so there
is a similar degree of aliasing present in each 10kHz wide
channel. The aliased signal correlates, and the phase of
the correlation does not change significantly across the
narrow channel, any further loss in sensitivity is negligible.
Development of the PFB boards was initially funded
through the University of Sydney and the firmware initially developed by CSIRO. The boards were designed
to form part of the original MWA correlator which
would share technology with the Square Kilometre Array Molongolo Prototype (SKAMP) (de Souza et al.
2007). Subsequent modifications to the firmware were
undertaken at MIT-Haystack and the final boards, with
firmware specific to the MWA, were first deployed as
part of the 32-tile MWA prototype.
3.2.2 Input Format, Skew and Packet Alignment
The PFB board input data format is the Xilinx serial
protocol RocketIO, although there have been some customisations at a low level, made within the RocketIO
CUSTOM scheme. The data can be read by any multigigabit-transceiver (MGT) that can use this protocol,
6
which in practice is restricted to the Xilinx FPGA (Vertex 5 and newer). The packet format as presented to the
PFB board for the second stage channeliser is detailed
in Prabu et al. (2014).
A single PFB board processes 12 fibres that have
come from 4 different receivers. The receivers are distributed over the 3 km diameter site and there exists
the possibility of there being considerable difference in
packet arrival time between inputs that are connected to
different receivers. The PFB boards have an input buffer
that is used to align all receiver inputs on packet number 0 (which is the 1 second tick marker see Appendix
A). The buffer is of limited length (±8 time-samples,
corresponding to several hundred metres of fibre) and if
the relative delay between input lines from nearby and
distant receivers is greater than this buffer length, the
time delay between receivers will be undetermined. We
have consolidated the receivers into 4 groups with comparable distances to the central processing building and
constrained these groups to have fibre runs of 1380, 925,
515, and 270 metres. Each of the 4 groups is allocated to
a single PFB (see Figure 2). This has resulted in more
fibre being deployed than was strictly necessary, but has
guaranteed that all inputs to a PFB board will arrive
within a packet-time. The limited buffer space available
can therefore be utilised to deal with variable delays induced by environmental factors and will guarantee that
all PFB inputs will be aligned when presented to the
correlator.
This packet alignment aligns all inputs onto the same
1 second mark but does not compensate for the outward clock signal delay, which traverses the same cable
length, and produces a large cable delay between the
receiver groups, commensurate with the differing fibre
lengths. The largest cable length difference is 1110 metres, being the difference in length between the shortest
and longest fibres and is removed during calibration.
3.2.3 PFB Output Format
As the MWA PFB boards were originally purposed as
the hardware F-stage of an FPGA based FX correlator, the output of the PFB was never intended to be
processed independently. The output format is also the
Xilinx serial protocol RocketIO, and the PFB presents
eight output data lanes on two CX4 connectors. These
connectors generally house four-lane XAUI as used by
Inifiniband and 10Gb Ethernet, but these are simply
used as a physical layer. In actuality all of the pins
on the CX4 connectors are driven as transmitters, as
opposed to receive and transmit pairs, therefore only
half of the available pins in the CX4 connectors were
being used. We were subsequently able to change the
PFB firmware to force all output data onto connector
pins that are traditionally transmit pins, and leave what
would normally be considered the receive pins unused.
We then custom designed breakout cables that split the
PASA (2015)
doi:10.1017/pas.2015.xxx
Ord et al.
CX4 into 4 single lane SFP connectors (see Appendix
A).
3.3 The Voltage Capture System
In order to capture data generated by the PFBs with
off-the-shelf hardware, the custom built SFP connectors were then plugged into an off-the-shelf Xilinx based
RocketIO capture cards that are housed within linux
servers. Each card can capture 2 lanes from the PFB
cards, therefore 4 cards were required per PFB, and 16
cards required in total. A set of 16 machines are dedicated to this voltage capture system (VCS), these are
16 CISCO UCS C240 M3 servers. They each house two
Intel Xeon E5-2650 processors, 128 GB of RAM, and
2x1TB RAID5 disk arrays. These machines also mount
the Xilinx-based FPGA board, supplied by Engineering
Design Team Incorporated (EDT) that is used to capture the output from the PFB. All of the initial buffering
and packet synchronisation is performed within these
machines.
3.3.1 Raw Packet Capture
The data capture and distribution is enabled by a succession of software packages outlined in Figure 4. First,
the EDT supplied capture card transfers a raw packet
stream from device memory to CPU host memory. The
transfer is mediated by the device driver and immediately copied to a larger shared memory buffer. At this
stage the data are checked for integrity and aligned on
packet boundaries to ensure efficient routing without
extensive re-examination of the packet contents. Each
of the 16 MWA receivers in the field obtain time synchronisation via a distributed 1 pulse per second (PPS)
that is placed into the data stream in the “second tick”
field as described in Appendix A. We maintain synchronisation with this “tick” and ensure all subsequent processing within the correlator labels the UTC second correctly. As each machine operates independently the success or failure of this synchronisation method can only
be detected when packets from all the servers arrive for
correlation, at which point the system is automatically
resynchronised if an error is detected. No unsynchronised data are correlated.
3.3.2 The Demultiplex
Data distribution in a connected element interferometer
is governed by two considerations; all the inputs for the
same frequency channels for each time step must be
presented to the correlator; and the correlator must be
able to keep up with the data rate.
As there are four PFB boards (see Figure 2), each
processing one quarter of the array, each output packet
contains the antenna inputs to that PFB, for a subset
of the channels for a single time step. Individual data
streams from each antenna in time order need to be
The MWA Correlator
7
Figure 4. The decomposition of the Voltage Capture System. The diagram shows the individual elements that comprise the VCS on
a single host. There are 16 VCS hosts operating independently (but synchronously) within the correlator system as a whole.
cross-connected to satisfy the requirement that a single time step for all antennas is presented simultaneously for cross correlation. To achieve this we utilise
the fact that each PFB packet is uniquely identified by
the contents of its header (see Appendix A). We route
each packet (actually each block of 2000 packets) to an
endpoint based upon the contents of this header, each
VCS server sends 128 contiguous, 10kHz channels to
the same physical endpoint, from all of the antennas in
its allocation. This results in the same 128 frequency
channels, from all antennas in the array, arriving at the
same physical endpoint. This is repeated for 24 different endpoints – each receiving a different 128 channels.
These endpoints are the next group of servers, the cross
multiply and accumulate (XMAC) machines.
This cross-connect is facilitated by a 10 Gb ethernet
switch. In actuality this permits the channels to be aggregated on any number of XMAC servers. We use 24
as it suits the frequency topology of the MWA. Regardless of topology some level of distribution is required to
reduce the load on an individual XMAC server in a flexible manner, as compute limitations govern how many
frequency channels can be simultaneously processed in
the XMAC stage. Clark et al. (2012) demonstrates that
the XMAC engine performance is limited to processing
approximately 4096 channels per NVIDIA Fermi GPU
at the 32 antenna level and 1024 channels at the 128
antenna level. This is well within the MWA computational limitations as we are required to process 3072
channels, and have up to 48 GPU available. However,
the benefit of packet switching correlators in general is
clear: if performance were an issue we could simply add
more GPU onto the switch to reduce the load per GPU
without adding complexity. Conversely, as GPU perforPASA (2015)
doi:10.1017/pas.2015.xxx
mance improves we can reduce the number of servers
and still perform the same task.
Each of the VCS servers is required to maintain 48
concurrent TCP/IP connections, two to each endpoint,
because each VCS server captures two lanes from a PFB
board . This amounts to 48 open sockets that are being
written to sequentially. As there are many more connections than available CPU cores individual threads,
tasked to manage each connection, the operation is subject to the linux thread scheduler, which attempts to
distribute CPU resources in a round robin manner. Each
thread on each transmit line utilises a small ring buffer
to allow continuous operation despite the inevitable device contention on the single 10 Gb ethernet line. Any
such contention causes the threads without access to
the interface to block. This time spent in this wait condition allows the scheduler to redistribute resources. We
also employ some context-based thread waiting on the
receiving side of the sockets to help the scheduler in its
decision making and to even the load across the data
capture threads. We have found that this scheme results
in a thread scheduling pattern that is remarkably fair
and equitable in the allocation of CPU resources and
provided that a reasonably-sized ring buffer is maintained for each data line, all the data are transmitted
without loss.
3.3.3 VCS Data Output
To summarise, we are switching packets as they are received and they are routed based upon the contents
of their header. Therefore any VCS server can process
any PFB lane output. Each server holds a single PCIbased capture card, and captures two PFB data lanes.
The data from each lane is grouped in time contiguous
8
blocks of 2000 packets (50 milliseconds worth of samples for 16 adjacent 10 kHz frequency channels), from all
32 of the antennas connected to that PFB. The packet
header provides sufficient information to uniquely identify the PFB, channel group, and time block it is, and
the packet block is routed on that basis. Further information is given in Appendix B. A static routing table is
used to ensure that each XMAC server receives a contiguous block of frequency channels, the precise number
of which is flexible and determined by the routing table,
but it is not alterable at runtime. 2
3.3.4 Monitoring, Command, Control and the
Watchdog
Monitor and control functionality within the correlator is mediated by a watchdog process that runs independently on each server. The watchdog launches each
process, monitors activity, and checks error conditions.
It can restart the system if synchronisation is lost and
mediates all start/stop/idle functionality as dictated by
the observing schedule. All interprocess command and
control, for example propagation of HALT instructions
to child processes and time synchronisation information, is maintained in a small block of shared memory
that holds a set of key-value pairs in plain text that can
be interrogated or set by third-party tools to control
the data capture and check status.
3.3.5 Voltage Recording
The VCS system also has the capability to record the
complete output of the PFB to local disk. The properties and capabilities of the VCS system will be detailed
in a companion paper (Tremblay et al. 2014).
3.4 The Cross-Multiply and Accumulate
(XMAC)
The GPU application for performing the cross-multiply
and accumulate is running on 24 NVIDIA M2070 Fermi
GPU housed in 24 IBM iDataplex servers. These machines are housed in racks adjacent to the VCS servers
and connected via a 10Gb ethernet network. The package diagram detailing this aspect of the correlator is
presented in Figure 5 and the packet interface and data
format is described in Appendix C. The cross-multiply
servers receive data from the 32 open sockets and internally align and unpack the data into a form suitable
for the GPU. The GPU kernel is launched to process
the data allocation and internally integrates over a user
defined length of time. There are other operations possible, such as incoherent beam forming and the raw dump
of data products (or input voltages). These modes will
2 In
order to achieve optimal loop unrolling and compile-time evaluation of conditional statements, xGPU requires that the number of channels, number of stations, and minimum integration
time are compile-time parameters.
PASA (2015)
doi:10.1017/pas.2015.xxx
Ord et al.
be expanded upon during the development of other processing pipelines and will be detailed in companion papers when the modes are available to the community.
3.4.1 Input Data Reception, and Management
A hierarchical ring buffer arrangement is employed to
manage the data flow. Each TCP connection is managed by a thread that receives a 2000 packet block and
places it in a shared memory buffer. Exactly where in a
shared memory block each of the 2000 packets is placed
is a function of the time/frequency/antenna block that
it represents. A small corner-turn, or transpose, is required as the PFB mixes the order of time and frequency. The final step in this process is the promotion
of the sample from 4 to 8 bits, which is required to enable a machine instruction that performs rapid promotion from 8 to 32 bits within the GPU. This promotion
is not performed if the data are just being output to
local disk after reordering.
Each of the 32 threads is filling a shared memory
block that represents 1 second of GPU input data, and
a management thread is periodically checking this block
to determine whether it is full. Safeguards are in place
that prevent a thread getting ahead of its colleagues or
overwriting a block. The threads wait on a condition
variable when they have finished their allocation, so if a
data line is not getting sufficient CPU resources it will
soon be the only thread not waiting and is guaranteed
to complete. This mechanism helps the Linux thread
scheduler to divide the resources equitably. The use
of TCP allows the receiver to actually prioritise which
threads on the send side of the connection are blocked
to facilitate an even distribution of CPU resources.
Once the management thread judges the GPU input
block to be full, it releases it, launches the GPU kernel
and frees all the waiting threads to fill the next buffer
while the GPU is running.
3.4.2 xGPU - Correlation on a GPU
It is possible to address the resources of a GPU in
a number of ways, utilising different application programming interfaces (API), such as OpenCL, CUDA,
and openGL. The MWA correlator uses the xGPU library as is described in detail in Clark et al. (2012).
The xGPU library is a CUDA application which is specific to GPUs built by the NVIDIA corporation. There
are many references in the literature to the CUDA programming model and examples of its use (Sanders &
Kandrot 2010).
The performance improvement seen when porting an
application to a GPU is in general due to the large
aggregate FLOPS and memory bandwidth rates, but
to realise this performance requires effective use of the
large number of concurrent threads of execution that
can be supported by the architecture. This massively
parallel architecture is permitted by the large number of
The MWA Correlator
9
Figure 5. The decomposition of the cross-multiply and accumulate operation running on each of 24, GPU enabled IBM iDataplex
servers.
processing cores on modern GPUs. The TESLA M2070
are examples of NVIDIA Fermi architecture and have
448 cores, grouped into 14 streaming multiprocessors or
SM, each with 32 cores. The allocation of resources follows the following model: threads are grouped into a
thread block and a thread block is assigned to an SM;
there can be more thread blocks than SM, but only
one block will be executing at a time on any SM; the
threads within each block are then divided into groups
of 32, (one for each core of the associated SM), and this
subdivision called a warp; execution is serialised within
a warp with the same instruction being performed by all
32 threads, but on different data elements in the single
instruction multiple data paradigm.
GPGPU application performance is often limited by
memory access bandwidth. A good predictor of algorithm performance is therefore arithmetic intensity,
or the number of floating point operations per byte
transferred. A complex multiply and add for two dual
polarisation antennas at 32 bit precision requires 32
bytes of input, 32 FLOPS (16 multiply-adds), and 64
bytes of output. The arithmetic intensity of this is
32/96=0.33. Compare this to the Fermi C2070 architecture which has a peak single-precision performance
of 1030 GFLOPS and memory bandwidth of 144 GB/s;
the ratio of which (7.2) tells us that the performance of
the algorithm will be completely dominated by memPASA (2015)
doi:10.1017/pas.2015.xxx
ory traffic. The high performance kernel developed by
Clark et al. (2011) increases the arithmetic intensity in
two ways. Firstly the output memory traffic can be reduced by integrating the products for a time, I, at the
register level. Secondly, instead of considering a single
baseline, groups of (m × n) baselines denoted tiles (see
Figure 6), are constructed cooperatively by a block of
threads. In order to fill an m × n region of the final
correlation matrix, only two vectors of antennas need
be transferred from host memory. A thread block loads
two vectors of antenna samples (of length n and m)
and forms nm baselines. These two steps then alter the
arithmetic intensity calculation to
Arithmetic Intensity =
32mnI
16(m + n)I + 64mn
(4)
(5)
which implies that the arithmetic intensity can be made
arbitrarily large by increasing the tile size until the number of available registers is exhausted. In practice a balance must be struck between the available resources and
the arithmetic intensity and this balance is achieved in
the xGPU implementation by multi-level tiling and we
direct the reader to Clark et al. (2012) for a complete
discussion.
10
Ord et al.
Each tile is mapped to a
block of threads
Tx
The correlation matrix is
broken into tiles, each
tile consists of many baselines
Ty
Rx
Each thread is allocated a
a sub-tile of baselines
Ry
Figure 6. The arithmetic intensity of the correlation operation
on the GPU is increased by tiling the correlation matrix. Threads
are assigned groups of baselines instead of a single baseline.
The MWA correlator is not a delay tracking correlator. Therefore, we do not have to adjust the correlator inputs for whole, or partial sample delays. The
correlator always provides correlation output products
phased to zenith. The system parameters (integration
time, baseline length, channel width) have all been chosen with this in mind and the long operating wavelengths result in the system showing only minimal ( 1%)
decorrelation even far from zenith for typical baseline
lengths, integration times and channel widths (see Figure 7). Typical observing resolutions have been between
0.5 and 2 seconds in time and 40 kHz and 10 kHz in frequency.
3.4.3 Output Data Format
The correlation products are accumulated on the GPU
device and then copied asynchronously back to host
memory. The length of accumulation is chosen by the
operator. Each accumulation block is topped with a
FITS header (Wells et al. 1981) and transferred to an
archiving system where the products are concatenated
into larger FITS files with multiple extensions, wherein
each integration is a FITS extension (see Appendix
D). The archiving system is an instantiation of the
Next Generation Archiving System (NGAS) (Wicenec
& Knudstrup 2007; Wu et al. 2013) and was originally
developed to store data from the European Southern
Observatory (ESO) telescopes. The heart of the system is a database that links the output products with
the meta-data describing the observation and that manages the safe transportation and archiving of data. In
the MWA operating paradigm all visibilities are stored,
which at the native resolution is 33024 visibilities and
3072 channels or approximately 800 MBytes per time
resolution unit (more than 6 Gb/s for 1 second integraPASA (2015)
doi:10.1017/pas.2015.xxx
tions). Due to the short baselines and long wavelengths
used by the array it is reasonable to integrate by a factor of two in each dimension (of time and frequency,
see Figure 7 for the parameter space), but even this reduced rate amounts to 17 TBytes/day at 100% duty
cycle. The large data volume precludes handling by humans and the archiving scheme is entirely automated.
NGAS provides tools to access the data remotely and
provides the link between the data products to the observatory SQL database. As discussed in Tingay et al.
(2013a), the data are taken at the Murchison Radio Observatory in remote Western Australia and transferred
to the Pawsey HPC Centre for SKA Science in Perth,
where 15 PB of storage is allocated to the MWA over its
five-year lifetime. Subsequently the archive database is
mirrored by the NGAS system to other locations around
the world: MIT in the USA; The Victoria University of
Wellington, NZ; and the Raman Research Institute in
India. These remote users are then able to access their
data locally.
4 Instrument Verification and
Commissioning
After initial instrument commissioning with simple test
vectors and noise inputs the telescope entered a commissioning phase in September 2012. The commissioning team initially consisted of 19 scientists from 10 institutions across 3 countries and was led by Curtin University. This commissioning phase was successfully concluded in July 2013 giving way to “Early Operations”
and the MWA is now fully operational.
4.1 Verification
The development of the MWA correlator proceeded separately to the rest of the MWA systems and the correlator GPU based elements were verified against other
software correlator tools. The MWA correlator is comparatively simple, it does not track delays, performs the
correlation at 32-bit floating point precision, and does
not have to perform any Fourier transformations. The
only verification that was required was to ensure that
the cross multiply and accumulate was performed to
sufficient precision and matched results generated by a
simple CPU implementation.
Subsequent to ensuring that the actual cross multiply and accumulate operation was being performed correctly the subsequent testing and integration was concerned with ensuring that the signal and path was maintained. This was a relatively complex operation due to
the requirement that the correlator interface with legacy
hardware systems. The details of these interfaces are
presented in the Appendices. In Figure 8 a correlation
matrix is shown, the colours representing the length
of the baseline associated with that antenna pair. The
The MWA Correlator
11
Figure 7. The combined reduction in coherence due to time and bandwidth smearing 45 degrees from zenith on a 1km baseline, at an
observing frequency of 200MHz as a function of channel width and integration time, demonstrating that even this far from zenith the
decorrelation
PASA (2015) is generally less the 1%. The longest MWA baselines are 3km, and these baselines show closer to 5% decorrelation for the
same observing parameters. Decorrelation factors above 1% are not plotted for clarity.
doi:10.1017/pas.2015.xxx
12
Ord et al.
Figure 8. A correlation matrix showing the baseline length distribution. The antennas are grouped into receivers, each servicing
16 antennas and the antenna layout displays a pronounced centrally dense core. The layout is detailed in Tingay et al. (2013a).
The colours indicate baseline length and the core regions are also
indicated by arrows within the figure.
following Figure 9 shows the response of the baselines
when the telescope is pointed at the radio galaxy Centaurus A, demonstrating the response of the different
baseline lengths of the interferometer to structure in the
source. Images of this object obtained with the MWA
can be found in McKinley et al. 2013. The arrows on
Figure 8 indicate those baselines associated with the
densely-packed core (Tingay et al. 2013a).
Figure 9. Visibility amplitude vs baseline length for a 2 minute
snapshot pointing at the giant radio galaxy Centaurus A. Centre frequency 120 MHz. Only 1 polarisation is shown for clarity.
The source has structure on a wide range of scales and does not
dominate the visibilities as a point source.
4.2 Verification Experiments
The MWA Array and correlator is also the verification
platform for the Aperture Array Verification System
(AAVS) within the SKA reconstruction activities pursued by the Low Frequency Aperture Array (LFAA)
consortium(bij de Vaate et al. 2011). This group have
recently performed a successful verification experiment,
using both the MWA and the AAVS telescopes. Using
an interferometric observation of the radio galaxy Hydra A taken with the MWA system, they have obtained
a measurement of antenna sensitivity (A/T) and compared this with a full-wave electromagnetic simulation
(FEKO3 ). The measurements and simulation show good
agreement at all frequencies within the MWA observing band and are presented in detail in Colegate et al.
(2014).
3 www.feko.info
PASA (2015)
doi:10.1017/pas.2015.xxx
4.3 Commissioning Science
In order to verify the instrumental performance of the
array, correlator and archive as a whole, the MWA Science Commissioning Team have performed a 4300 deg2
survey, and have published a catalogue of flux densities
and spectral indices from 14,121 compact sources. The
survey covered approximately 21 h < Right Ascension
< 8 h, -55◦ < Declination < -10◦ over three frequency
bands centred on 119, 150 and 180 MHz. This survey
will be detailed in Hurley-Walker et al. (2014). Data
taken during and shortly after the commisioning phase
of the instrument has also been used to demonstrate the
tracking of space debris with the MWA (Tingay et al.
2013b), novel imaging and deconvolution schemes (Offringa et al. 2014), and to present multi-frequency observations of the Fornax radio galaxy (McKinley et al.
2013) and the galaxy cluster A3667 (Hindson et al.
2014).
The MWA Correlator
5 The Future of the Correlator
5.1 Upgrade Path
We do not fully utilise the GPU in the MWA correlator and we estimate that the X-Stage of the correlator
could currently support a factor of 2 increase in array
size (to 256 dual polarisation elements). As discussed
in Tingay et al. (2013a) the operational life-span of the
MWA is intended to be approximately five years. The
GPUs (NVIDIA M2070s) the MWA correlator are from
the Fermi family of NVIDIA GPU and are already almost obsolete. One huge advantage of off-the-shelf signal processing solutions is that we can easily benefit
from improvements in technology. We could swap out
the GPU in the current MWA cluster, replace them with
cards from the Kepler series (K20X) with no code alteration, and would benefit from a factor of 2.5 increase
in performance (and a threefold improvement in power
efficiency) as not only are the number of FLOPS provided by the GPUs increasing, but also their efficiency
(in FLOPS/Watt) is improving rapidly. This would permit the MWA to be scaled to 512 elements. However,
not all of our problem is with the correlation step and
increasing the array to 512 elements would probably require an upgrade to the networking capability of the
correlator.
5.2 SKA Activities
The MRO is the proposed site of the low-frequency portion of the Square Kilometre Array. Curtin University
is a recipient of grants from the Australian government
to support the design and pre-construction effort associated with the construction and verification of the
Low-Frequency Aperture Array (LFAA) and software
correlation systems for the SKA project. The MWA
correlator will therefore be used as a verification platform within the prototype for the LFAA, known as the
Aperture Array Verification System (AAVS). One of the
benefits of a software based signal processing system
is flexibility. We have been able to easily incorporate
AAVS antennas into the MWA processing chain to facilitate these verification activities, which have been of
benefit to both LFAA and the MWA (Colegate et al.
2014). Furthermore we will be developing and deploying SKA software correlator technologies throughout
the pre-construction phase of the SKA using the MWA
correlator for both testing and verification.
13
eration radio telescope. We have outlined the various
stages in the correlator signal path and detailed the
form of all internal and external interfaces.
7 Acknowledgements
We acknowledge the Wajarri Yamatji people as the traditional owners of the Observatory site. Initial correlator development was enabled by equipment supplied
by an IBM Shared University Research Grant (VUW
& Curtin University), and by an Internal Research and
Development Grant from the Smithsonian Astrophysical Observatory.
Support for the MWA comes from the U.S. National
Science Foundation (grants AST CAREER-0847753,
AST-0457585, AST-0908884 and PHY-0835713), the
Australian Research Council (LIEF grants LE0775621
and LE0882938), the U.S. Air Force Office of Scientific Research (grant FA9550-0510247), the Centre for
All-sky Astrophysics (an Australian Research Council
Centre of Excellence funded by grant CE110001020),
New Zealand Ministry of Economic Development (grant
MED-E1799), an IBM Shared University Research
Grant (via VUW & Curtin), the Smithsonian Astrophysical Observatory, the MIT School of Science, the
Raman Research Institute, the Australian National
University, the Victoria University of Wellington, the
Australian Federal government via the National Collaborative Research Infrastructure Strategy, Education
Investment Fund and the Australia India Strategic Research Fund and Astronomy Australia Limited, under contract to Curtin University, the iVEC Petabyte
Data Store, the Initiative in Innovative Computing and
NVIDIA sponsored CUDA Center for Excellence at
Harvard, and the International Centre for Radio Astronomy Research, a Joint Venture of Curtin University
and The University of Western Australia, funded by the
Western Australian State government.
A MWA F-Stage (PFB) to Voltage Capture
System (VCS)
The contents of this Appendix are extended from the original MWA ICD for the polyphase filterbank written by R.
Cappallo. We have incorporated the changes made to the
physical mappings, electrical connections and header contents to enable capture of the PFB output by general purpose computers.
6 Summary
This paper outlines the structure, interfaces, operations
and data formats of the MWA Hybrid FPGA/GPU
correlator. This system combines off-the-shelf computer
hardware with bespoke digital electronics to provide a
flexible and extensible correlator solution for a next genPASA (2015)
doi:10.1017/pas.2015.xxx
A.1 Physical
• PFB presents data on two CX-4 connectors. ALL pins
are wired with transmit drivers, there are no receivers
on any pins
14
Ord et al.
A.2 Interface Details
• Two custom made CX4 to SFP breakout cables. This
cable breaks a CX4 connector into 4 1X transmit and
receive pairs on individual SFP terminated cables.
• Data flows only in a single direction (simplex) from
PFB to CB. Each (normally Tx/Rx) pair of differential pairs is wired as two one-way data paths (Tx /Tx).
We have distributed these one-way data paths to occupy every other CX-4 pin. Therefore all cables are
wired correctly as Tx/Rx but the Rx channels are not
utilised.
• The serial protocol to be used a custom protocol as described below. The interface to the Xilinx Rocket I/O
MGTs is GT11 CUSTOM, which is a lean protocol.
The word width will be 16 bits.
• The mean data rate per 1X cable will be 2.01216 Gb/s,
carried on a high speed serial channel of burst rate
of 2.08 Gb/s formatted, or 2.6 Gb/s including the
10B/8B encoding overhead.
A.3 Data Format
• Each data stream carries all of the antenna information for 384 fine frequency channels. In the entire 32T
correlation system there are 8 such data streams going
from the PFB to the CB, each carrying the channels
making up a 3.84 MHz segment of the processed spectrum.
• Each sample in the stream is an 8-bit (4R+4I) complex
number, 2s complement with 8 denoting invalid data,
representing the voltage in a 10 kHz fine frequency
channel.
• Data are packetised and put in a strictly defined sequence, in f1t[f2a] (frequency-time-frequency-antenna)
order. The most rapidly varying index, a, is the antenna number (0, 16, 32, 48, 1, 17, 33, 49, 2,18, , 15,
31, 47, 63), followed by f2, the fine frequency channel
(n..n+3), then time (0..499), and most slowly the fine
frequency channel group index (0, 4, 8, 12, 128, 132,
136, 140, ,, 2944, 2948, 2952, 2956 for fibre 0; increment by 16n for fibre n). The brackets indicate the
portion of the stream that is contained within a single
packet. The antenna ordering is non intuitive as for
historical reasons each PFB combines its inputs into
an order originally considered conducive to its interface to a full mesh backplane and a hardware FPGA
based correlator.
• A single packet consists of all antenna data for a group
of four fine channels for a single time point. Thus there
are 500 sequential time packets for each of the 96 frequency channel groups. Over the interval of 50 ms
there are 48,000 (=96channel groups x 500 time samples) packets sent per data stream, so there are 960,000
packets/second.
• Using a packet size of 264 bytes allows a 16-bit packet
header (0x0800), two 16-bit words carrying the second
of time tick and packet counter, 256 bytes of antenna
data, finally followed by a 16-bit checksum (see Table
1).
PASA (2015)
doi:10.1017/pas.2015.xxx
• The header word is a 16-bit field with value 0x0800,
used to denote the first word in a packet. Note that
it cannot be confused with the data words, since the
value 8 is not a legal voltage sample code.
• The second tick in word two is in bit 0, and has the
value 0 at all times, except for the first packet of a new
second, when it is 1.
• The leading bits of word two, and all of word three
contains the following counter values that can be decoded to determine the precise input and channel that
a given packet belongs to.
– sec tick (1 bit): Is this the first packet for this data
lane, for this one second of data? 0=No, 1=Yes.
Only found on the very first packet of that second.
If set, then the mgt bank, mgt channel, mgt group
and mgt frame will always be 0. The preceding 3
bits are unused.
– pfb id (2 bits): Which physical PFB board generated this stream [0-3]. This defines the set of receivers and tiles the data refers to. A lane’s pfb id
will remain constant unless physically shifted to a
different PFB via a cable swap.
– mgt id (3 bits): Which 1/8 of coarse channels
this Rocket–IO lane contains. Should be masked
with 0x7, not 0xf. Contains [0-7] inclusive. A
lane’s mgt id will remain constant unless physically
shifted to a different port on the PFB via a cable
swap.
Within an individual data lane, the data packets cycle
in the following order, listed from slowest to fastest.
– mgt bank (5 bits): Which one of the twenty 50ms
time banks in the current second is this one? [0-19]
[0-0x13].
– mgt channel (5 bits): Which coarse channel [0-23]
[0-0x17] does the packet relate to?
– mgt group (2 bits): Which 40KHz wide packet of the
contiguous 160KHz does this packet contain the 4
x 10KHz samples for? [0-3]
– mgt frame (5 bits): Which time stamp within a
50ms block this packet is. [0-499] or [0-0x1F3]. Cycles fastest. Loops back to 0 after 0x1F3. There are
20 complete cycles in a second.
• The last word of the packet is a 16-bit checksum,
formed by taking the bitwise XOR of all 128 (16-bit)
words of antenna samples in the packet.
B VCS to XMAC Server
B.1 Physical
• Data presented on either fibre optic or direct attach copper cables terminated with SFP pluggable
transceivers. Note that all optical transceivers used in
the CISCO UCS servers must be supplied by CISCO,
the firmware within the 10Gb ethernet cards requires
it.
• A single 10GbE interface is sufficient to handle the
data rate
The MWA Correlator
B.2 Interface Details
• Ethernet IEEE 802.3 frames, with TCP/IPv4
• The packet format is precisely as described in the PFB
to VCS interface
• The communication is from the VCS to the XMAC is
uni-directional and mediated by a switch.
• Each XMAC requires a data packet from all VCS machines. This requires 32 open connections mediated by
TCP
• Data from all sources is stripped of header information
and assembled into a common 1 second buffer which
requires 20 blocks of 2000 packets from each of the 32
connections to be assembled.
• As the packets for four PFB are being combined the
antenna ordering is also being concatenated. The order
within each packet is maintained with the 64 inputs
concatenated into a final block of 256 inputs.
C XMAC to GPU Interface
15
Therefore the correlation products cannot be simply processed without a remapping.
The remapping is performed by a utility that has been
supplied by the builders to the data analysis teams that
simply reorders the data products into one that would be
expected if no PFB reordering had taken place. It is important to keep track of which products have been generated
but the XMAC as conjugates of the desired products, and
which need conjugating. It should further be noted that this
correction is performed to the order as presented to the PFB
and may not the order as expected by the ordering of the
receiver inputs. Care should be taken in mapping antennas
on the ground as the order of the receiver processing within
a PFB is not only function of the physical wiring but the
firmware mapping. We have carefully determined this mapping and the antenna mapping tables are held within the
same database as the observing metadata.
For historical reasons the files are transferred into an intermediate file format that was originally used during instrument commissioning before subsequent transformation to
UVFITS files (or measurement sets) by the research teams.
C.1 Physical
The interface is internal to the CPU host, being a section
of memory shared between the GPU device driver and the
host.
D XMAC Server to NGAS (Next
Generation Archiving System)
D.1 Physical
C.2 Interface details
C.2.1 GPU Input
The data ordering in the assembly buffer for input buffers,
running from slowest to fastest index is time(t), channel(f),
station(s), polarisation(p), complexity(i). The data is an 8
bit integer, promoted from 4-bit twos-complement sample
via a lookup table. The correlator is a dual polarisation correlator as implemented, but the input station/polarisation
ordering is not in a convenient order as we have concatenated
4 PFB outputs during the assembly of this buffer. Two options presented themselves. One that we re-order the input
stations and polarisation to undo the partial corner turn.
The disadvantage begin that each input packet would need
to be broken open and the stations reordered - which would
have been extremely compute intensive. Or to correlate without changing the order and remapping the output products
to the desired order. This is preferable as although there
are now N2 products instead of N stations, the products are
integrated in time and can be reordered in post-processing.
C.2.2 GPU Output
The data are formated as 32-bit complex floating point
numbers. An individual visibility set has the order (from
slowest index to fastest index): channel(f), station(s1), station(s2), polarisation(p1), polarisation(p2), complexity(i).
The full cross-correaltion matrix is Hermitian so only the
lower, or upper, triangular elements need to be calculated,
along with the on-diagonal elements which are the autocorrelation products. Therefore only a packed triangular matrix is actually transferred from the GPU to the host. The
GPU processing kernel is agnostic to the ordering of the
antenna inputs and the PFB output order is maintained.
PASA (2015)
doi:10.1017/pas.2015.xxx
The interface is internal to the CPU host, being a section
of memory shared between the XMAC application and an
NGAS application running on the same host.
D.2 Interface details
The data are handed over as a packed triangular matrix
with the same ordering as was produced by the GPU. It has
a FITS header added and is padded to the correct size as
expected by a FITS extension. The FITS header includes a
time-tag.
Once the buffer is presented to the NGAS system it is
buffered locally and presented to a central archive server,
subsequently it is added to a database and mirrored to multiple sites across the world.
REFERENCES
bij de Vaate, J. G., de Lera Acedo, E., Virone, G., Jiwani,
A., Razavi, N., Perini, F., Zarb-Adami, K., Monari, J.,
Padhi, S., Addamo, G., Peverini, O., Montebugnoli, S.,
Gunst, A., Hall, P., Faulkner, A., & van Ardenne, A.
2011, General Assembly and Scientific Symposium, 2011
XXXth URSI, 1
Bowman, J. D., Cairns, I., Kaplan, D. L., Murphy, T.,
Oberoi, D., Staveley-Smith, L., Arcus, W., Barnes, D. G.,
Bernardi, G., Briggs, F. H., Brown, S., Bunton, J. D.,
Burgasser, A. J., Cappallo, R. J., Chatterjee, S., Corey,
B. E., Coster, A., Deshpande, A., Desouza, L., Emrich, D.,
Erickson, P., Goeke, R. F., Gaensler, B. M., Greenhill, L.,
Harvey-Smith, L., Hazelton, B. J., Herne, D. E., Hewitt,
J. N., Johnston-Hollitt, M., Kasper, J. C., Kincaid, B. B.,
Koeing, R., Kratzenberg, E., Lonsdale, C. J., Lynch,
16
M. J., Matthews, L. D., McWhirter, S. R., Mitchell, D.,
Morales, M. F., Morgan, E. H., Ord, S. M., Pathikulangara, J., Prabu, T., Remillard, R. A., Robishaw, T.,
Rogers, A. E. E., Roshi, A. A., Salah, J. E., Sault, R. J.,
Shankar, N. U., Srivani, K. S., Stevens, J. B., Subrahmanyan, R., Tingay, S. J., Wayth, R. B., Waterson, M.,
Webster, R. L., Whitney, A. R., Williams, A. J., Williams,
C. L., & Wyithe, J. S. B. 2013, Publications of the Astronomical Society of Australia, 30, 31
Bunton, J. D. 2003, ALMA Memo Series, 1
Clark, M. A., La Plante, P. C., & Greenhill, L. J. 2012, International Journal of High Performance Computing Applications, 27(2), 178
Colegate, T. M., Sutinjo, A. T., Hall, P. J., Padhi, S. K.,
Wayth, R. B., de Vaate, J. G. B., Crosse, B. W., Emrich,
D., Faulkner, A. J., Hurley-Walker, N., de Lera Acedo,
E., Juswardy, B., Razavi-Ghods, N., Tingay, S. J., &
Williams, A. J. 2014, in Antenna Measurements & Applications (CAMA), 2014 IEEE Conference on (IEEE),
1–4
Crochiere, R. E. & Rabiner, L. R. 1983, Multirate Digital
Signal Processing (Prentice-Hall Processing Series), 1st
edn. (Prentice-Hall)
de Souza, L., Bunton, J. D., Campbell-Wilson, D., Cappallo,
R. J., & Kincaid, B. 2007, in FPL, ed. K. Bertels, W. A.
Najjar, A. J. van Genderen, & S. Vassiliadis (IEEE), 62–
67
Hindson, L., Johnston-Hollitt, M., Hurley-Walker, N., Buckley, K., Morgan, J., Carretti, E., Dwarakanath, K. S., Bell,
M., Bernardi, G., Bhat, N. D. R., Bowman, J. D., Briggs,
F., Cappallo, R. J., Corey, B. E., Deshpande, A. A., Emrich, D., Ewall-Wice, A., Feng, L., Gaensler, B. M., Goeke,
R., Greenhill, L. J., Hazelton, B. J., Jacobs, D., Kaplan,
D. L., Kasper, J. C., Kratzenberg, E., Kudryavtseva, N.,
Lenc, E., Lonsdale, C. J., Lynch, M. J., McWhirter, S. R.,
McKinley, B., Mitchell, D. A., Morales, M. F., Morgan, E.,
Oberoi, D., Ord, S. M., Pindor, B., Prabu, T., Procopio,
P., Offringa, A. R., Riding, J., Rogers, A. E. E., Roshi,
A., Udaya Shankar, N., Srivani, K. S., Subrahmanyan,
R., Tingay, S. J., Waterson, M., Wayth, R. B., Webster,
R. L., Whitney, A. R., Williams, A., & Williams, C. L.
2014, ArXiv e-prints
Hurley-Walker, N. 2014, submitted for publication.
Kocz, J., Greenhill, L. J., Barsdell, B. R., Bernardi, G.,
Jameson, A., Clark, M. A., Craig, J., Price, D., Taylor,
G. B., Schinzel, F., & Werthimer, D. 2014, Journal of
Astronomical Instrumentation, 3, 50002
Lonsdale, C. J., Cappallo, R. J., Morales, M. F., Briggs,
F. H., Benkevitch, L., Bowman, J. D., Bunton, J. D.,
Burns, S., Corey, B. E., Desouza, L., Doeleman, S. S.,
Derome, M., Deshpande, A., Gopala, M. R., Greenhill,
L. J., Herne, D. E., Hewitt, J. N., Kamini, P. A., Kasper,
J. C., Kincaid, B. B., Kocz, J., Kowald, E., Kratzenberg,
E., Kumar, D., Lynch, M. J., Madhavi, S., Matejek, M.,
Mitchell, D. A., Morgan, E., Oberoi, D., Ord, S., Pathikulangara, J., Prabu, T., Rogers, A., Roshi, A., Salah, J. E.,
Sault, R. J., Shankar, N. U., Srivani, K. S., Stevens, J.,
Tingay, S., Vaccarella, A., Waterson, M., Wayth, R. B.,
Webster, R. L., Whitney, A. R., Williams, A., & Williams,
C. 2009, IEEE Proceedings, 97, 1497
PASA (2015)
doi:10.1017/pas.2015.xxx
Ord et al.
McKinley, B., Briggs, F., Gaensler, B. M., Feain, I. J.,
Bernardi, G., Wayth, R. B., Johnston-Hollitt, M., Offringa, A. R., Arcus, W., Barnes, D. G., Bowman, J. D.,
Bunton, J. D., Cappallo, R. J., Corey, B. E., Deshpande,
A. A., deSouza, L., Emrich, D., Goeke, R., Greenhill,
L. J., Hazelton, B. J., Herne, D., Hewitt, J. N., Kaplan,
D. L., Kasper, J. C., Kincaid, B. B., Koenig, R., Kratzenberg, E., Lonsdale, C. J., Lynch, M. J., McWhirter, S. R.,
Mitchell, D. A., Morales, M. F., Morgan, E., Oberoi, D.,
Ord, S. M., Pathikulangara, J., Prabu, T., Remillard,
R. A., Rogers, A. E. E., Roshi, D. A., Salah, J. E., Sault,
R. J., Shankar, N. U., Srivani, K. S., Stevens, J., Subrahmanyan, R., Tingay, S. J., Waterson, M., Webster, R. L.,
Whitney, A. R., Williams, A., Williams, C. L., & Wyithe,
J. S. B. 2013, MNRAS, 436, 1286
Offringa, A. R., McKinley, B., Hurley-Walker, N., Briggs,
F. H., Wayth, R. B., Kaplan, D. L., Bell, M. E., Feng, L.,
Neben, A. R., Hughes, J. D., Rhee, J., Murphy, T., Bhat,
N. D. R., Bernardi, G., Bowman, J. D., Cappallo, R. J.,
Corey, B. E., Deshpande, A. A., Emrich, D., Ewall-Wice,
A., Gaensler, B. M., Goeke, R., Greenhill, L. J., Hazelton,
B. J., Hindson, L., Johnston-Hollitt, M., Jacobs, D. C.,
Kasper, J. C., Kratzenberg, E., Lenc, E., Lonsdale, C. J.,
Lynch, M. J., McWhirter, S. R., Mitchell, D. A., Morales,
M. F., Morgan, E., Kudryavtseva, N., Oberoi, D., Ord,
S. M., Pindor, B., Procopio, P., Prabu, T., Riding, J.,
Roshi, D. A., Udaya Shankar, N., Srivani, K. S., Subrahmanyan, R., Tingay, S. J., Waterson, M., Webster, R. L.,
Whitney, A. R., Williams, A., & Williams, C. L. 2014,
ArXiv e-prints
Parsons, A. R., Backer, D. C., Foster, G. S., Wright, M.
C. H., Bradley, R. F., Gugliucci, N. E., Parashare, C. R.,
Benoit, E. E., Aguirre, J. E., Jacobs, D. C., Carilli, C. L.,
Herne, D. E., Lynch, M. J., Manley, J. R., & Werthimer,
D. J. 2010, The Astronomical Journal, 139, 1468
Prabu, T., Srivani, K., & Anish, R. 2014, in prep.
Sanders, J. & Kandrot, E. 2010, CUDA by Example,
An Introduction to General-Purpose GPU Programming
(Addison-Wesley Professional)
Thompson, A. R., Moran, J. M., & Swenson, Jr., G. W.
2001, Interferometry and Synthesis in Radio Astronomy,
2nd Edition (John Wiley and Sons)
Tingay, S. J., Goeke, R., Bowman, J. D., Emrich, D., Ord,
S. M., Mitchell, D. A., Morales, M. F., Booler, T., Crosse,
B., Wayth, R. B., Lonsdale, C. J., Tremblay, S., Pallot,
D., Colegate, T., Wicenec, A., Kudryavtseva, N., Arcus,
W., Barnes, D., Bernardi, G., Briggs, F., Burns, S., Bunton, J. D., Cappallo, R. J., Corey, B. E., Deshpande,
A., Desouza, L., Gaensler, B. M., Greenhill, L. J., Hall,
P. J., Hazelton, B. J., Herne, D., Hewitt, J. N., JohnstonHollitt, M., Kaplan, D. L., Kasper, J. C., Kincaid, B. B.,
Koenig, R., Kratzenberg, E., Lynch, M. J., Mckinley, B.,
Mcwhirter, S. R., Morgan, E., Oberoi, D., Pathikulangara, J., Prabu, T., Remillard, R. A., Rogers, A. E. E.,
Roshi, A., Salah, J. E., Sault, R. J., Udaya-Shankar,
N., Schlagenhaufer, F., Srivani, K. S., Stevens, J., Subrahmanyan, R., Waterson, M., Webster, R. L., Whitney,
A. R., Williams, A., Williams, C. L., & Wyithe, J. S. B.
2013a, Publications of the Astronomical Society of Australia, 30, 7
The MWA Correlator
Tingay, S. J., Kaplan, D. L., McKinley, B., Briggs, F.,
Wayth, R. B., Hurley-Walker, N., Kennewell, J., Smith,
C., Zhang, K., Arcus, W., Bhat, N. D. R., Emrich, D.,
Herne, D., Kudryavtseva, N., Lynch, M., Ord, S. M., Waterson, M., Barnes, D. G., Bell, M., Gaensler, B. M., Lenc,
E., Bernardi, G., Greenhill, L. J., Kasper, J. C., Bowman,
J. D., Jacobs, D., Bunton, J. D., deSouza, L., Koenig, R.,
Pathikulangara, J., Stevens, J., Cappallo, R. J., Corey,
B. E., Kincaid, B. B., Kratzenberg, E., Lonsdale, C. J.,
McWhirter, S. R., Rogers, A. E. E., Salah, J. E., Whitney, A. R., Deshpande, A., Prabu, T., Udaya Shankar, N.,
Srivani, K. S., Subrahmanyan, R., Ewall-Wice, A., Feng,
L., Goeke, R., Morgan, E., Remillard, R. A., Williams,
C. L., Hazelton, B. J., Morales, M. F., Johnston-Hollitt,
M., Mitchell, D. A., Procopio, P., Riding, J., Webster,
R. L., Wyithe, J. S. B., Oberoi, D., Roshi, A., Sault, R. J.,
& Williams, A. 2013b, AJ, 146, 103
Tremblay, S., Ord, S., Bhat, N., Tingay, S., Crosse, B., Pallotand S. Oronsaye, D., Bernardi, G., Bowman, J. D.,
Briggs, F., Cappallo, R. J., Corey, B. E., Deshpande,
A. A., Emrich, D., Gaensler, B. M., Goeke, R., Greenhill, L. J., Hazelton, B. J., Johnston-Hollitt, M., Kaplan,
D. L., Kasper, J. C., Kratzenberg, E., Lonsdale, C. J.,
Lynch, M. J., McWhirter, S. R., Mitchell, D. A., Morales,
M. F., Morgan, E., Oberoi, D., Prabu, T., Rogers, A.
E. E., Roshi, A., Shankar, N. U., Srivani, K. S., Subrahmanyan, R., Waterson, M., Wayth, R. B., Webster, R. L.,
Whitney, A. R., Williams, A., & Williams, C. L. 2014, in
preparation.
van Haarlem, M. P., Wise, M. W., Gunst, A. W., Heald, G.,
McKean, J. P., Hessels, J. W. T., de Bruyn, A. G., Nijboer, R., Swinbank, J., Fallows, R., Brentjens, M., Nelles,
A., Beck, R., Falcke, H., Fender, R., Hörandel, J., Koopmans, L. V. E., Mann, G., Miley, G., Röttgering, H., Stappers, B. W., Wijers, R. A. M. J., Zaroubi, S., van den
Akker, M., Alexov, A., Anderson, J., Anderson, K., van
Ardenne, A., Arts, M., Asgekar, A., Avruch, I. M., Batejat, F., Bähren, L., Bell, M. E., Bell, M. R., van Bemmel, I., Bennema, P., Bentum, M. J., Bernardi, G., Best,
P., Bı̂rzan, L., Bonafede, A., Boonstra, A.-J., Braun, R.,
Bregman, J., Breitling, F., van de Brink, R. H., Broderick,
J., Broekema, P. C., Brouw, W. N., Brüggen, M., Butcher,
H. R., van Cappellen, W., Ciardi, B., Coenen, T., Conway,
J., Coolen, A., Corstanje, A., Damstra, S., Davies, O.,
Deller, A. T., Dettmar, R.-J., van Diepen, G., Dijkstra,
K., Donker, P., Doorduin, A., Dromer, J., Drost, M., van
Duin, A., Eislöffel, J., van Enst, J., Ferrari, C., Frieswijk,
W., Gankema, H., Garrett, M. A., de Gasperin, F., Gerbers, M., de Geus, E., Grießmeier, J.-M., Grit, T., Gruppen, P., Hamaker, J. P., Hassall, T., Hoeft, M., Holties,
H. A., Horneffer, A., van der Horst, A., van Houwelingen,
A., Huijgen, A., Iacobelli, M., Intema, H., Jackson, N.,
Jelic, V., de Jong, A., Juette, E., Kant, D., Karastergiou,
A., Koers, A., Kollen, H., Kondratiev, V. I., Kooistra, E.,
Koopman, Y., Koster, A., Kuniyoshi, M., Kramer, M.,
Kuper, G., Lambropoulos, P., Law, C., van Leeuwen, J.,
Lemaitre, J., Loose, M., Maat, P., Macario, G., Markoff,
S., Masters, J., McFadden, R. A., McKay-Bukowski, D.,
Meijering, H., Meulman, H., Mevius, M., Middelberg, E.,
Millenaar, R., Miller-Jones, J. C. A., Mohan, R. N., Mol,
PASA (2015)
doi:10.1017/pas.2015.xxx
17
J. D., Morawietz, J., Morganti, R., Mulcahy, D. D., Mulder, E., Munk, H., Nieuwenhuis, L., van Nieuwpoort, R.,
Noordam, J. E., Norden, M., Noutsos, A., Offringa, A. R.,
Olofsson, H., Omar, A., Orrú, E., Overeem, R., Paas, H.,
Pandey-Pommier, M., Pandey, V. N., Pizzo, R., Polatidis, A., Rafferty, D., Rawlings, S., Reich, W., de Reijer, J.-P., Reitsma, J., Renting, G. A., Riemers, P., Rol,
E., Romein, J. W., Roosjen, J., Ruiter, M., Scaife, A.,
van der Schaaf, K., Scheers, B., Schellart, P., Schoenmakers, A., Schoonderbeek, G., Serylak, M., Shulevski, A.,
Sluman, J., Smirnov, O., Sobey, C., Spreeuw, H., Steinmetz, M., Sterks, C. G. M., Stiepel, H.-J., Stuurwold, K.,
Tagger, M., Tang, Y., Tasse, C., Thomas, I., Thoudam,
S., Toribio, M. C., van der Tol, B., Usov, O., van Veelen,
M., van der Veen, A.-J., ter Veen, S., Verbiest, J. P. W.,
Vermeulen, R., Vermaas, N., Vocks, C., Vogt, C., de Vos,
M., van der Wal, E., van Weeren, R., Weggemans, H.,
Weltevrede, P., White, S., Wijnholds, S. J., Wilhelmsson,
T., Wucknitz, O., Yatawatta, S., Zarka, P., Zensus, A., &
van Zwieten, J. 2013, Astronomy & Astrophysics, 556, A2
Wayth, R. B., Greenhill, L. J., & Briggs, F. H. 2009, Publications of the Astronomical Society of Australia, 121,
857
Wells, D. C., Greisen, E. W., & Harten, R. H. 1981, Astronomy & Astrophysics Supplement, 44, 363
Wicenec, A. & Knudstrup, J. 2007, The Messenger, 129, 27
Wu, C., Wicenec, A., Pallot, D., & Checcucci, A. 2013, Experimental Astronomy, 36, 679
| 5cs.CE
|
GENERALIZED IRREDUCIBLE DIVISOR GRAPHS
CHRISTOPHER PARK MOONEY
arXiv:1312.7406v1 [math.AC] 28 Dec 2013
Abstract. In 1988, I. Beck introduced the notion of a zero-divisor graph of a commutative
rings with 1. There have been several generalizations in recent years. In particular, in 2007
J. Coykendall and J. Maney developed the irreducible divisor graph. Much work has been
done on generalized factorization, especially τ -factorization. The goal of this paper is to
synthesize the notions of τ -factorization and irreducible divisor graphs in domains. We will
define a τ -irreducible divisor graph for non-zero non-unit elements of a domain. We show
that by studying τ -irreducible divisor graphs, we find equivalent characterizations of several
finite τ -factorization properties.
2010 AMS Subject Classification: 13A05, 13E99, 13F15, 5C25
1. Introduction
In this article, D will denote an integral domain. We will always assume that all rings
have an identity which is not zero. We will use G = (V, E) to denote a graph with V the set
of vertices and E the set of edges. In 1988, I. Beck in [8], introduced for a commutative ring
R, the notion of a zero-divisor graph Γ(R). The vertices of Γ(R) are the set of zero-divisors,
and there is an edge between a, b ∈ Z(R) if ab = 0. This has been studied and developed by
many authors including, but not limited to D.D. Anderson, D.F. Anderson, A. Frazier, A.
Lauve, P.S. Livingston, and M. Naseer in [3, 4, 5].
There have been several generalizations and extensions of this concept. In this paper, we
Date: November 21, 2017.
Key words and phrases. factorization, commutative rings, zero-divisor graphs, irreducible divisor graphs.
1
2
CHRISTOPHER PARK MOONEY
focus on the notion of an irreducible divisor graph first formulated by J. Coykendall and J.
Maney in [9]. Instead of looking exclusively at divisors of zero in a ring, they restrict to a
domain D and choose any non-zero, non-unit x ∈ D and study the relationships between
the irreducible divisors of x. In [7, 9], M. Axtell, N. Baeth, J. Coykendall, J. Maney, and J.
Stickles present several nice results about factorization properties of domains based on their
associated irreducible divisor graphs. M. Axtell and J. Stickles have also studied irreducible
divisor graphs in commutative rings with zero-divisors in [6].
Unique factorization domains are well known and widely studied. Lesser known are other
finite factorization properties that domains might possess which are weaker than UFDs.
These come in the form of half factorization domains or half factorial domains (HFDs),
finite factorization domains (FFDs) and bounded factorization domains (BFDs). See [1] for
more information on the developments in the theory of factorization in integral domains.
More recently, these concepts have been further generalized by way of τ -factorization in
several papers, especially by D.D. Anderson and A. Frazier in [2] as well as the author in
[12, 13]. In this paper, we seek to take the notion of τ -factorization and apply it to irreducible
divisor graphs. We will find that many equivalent characterizations of τ -finite factorization
properties given in the aforementioned papers can be given by studying τ -irreducible divisor
graphs.
Section Two gives the necessary preliminary background information and definitions from
the study of irreducible and zero-divisor graphs as well as τ -factorization. In Section Three,
we define the τ -irreducible divisor graph of a domain D with a fixed relation τ . We give a few
examples of τ -irreducible divisor graphs, especially in comparison with the irreducible divisor
graphs of [6, 9]. In Section Four, we prove several theorems illustrating how τ -irreducible
GENERALIZED IRREDUCIBLE DIVISOR GRAPHS
3
divisor graphs give us another way to characterize various τ -finite factorization properties
domains may possess as defined in [2].
2. Preliminary Definitions
2.1. Irreducible Divisor Graph Definitions.
We begin with some definitions from J. Coykendall and J. Maney [9]. Let Irr(D) be
the set of all irreducible elements in a domain D. We will let Irr(D) be a (pre-chosen)
set of coset representatives of the collection {aU(D) | a ∈ Irr(D)}. Let x ∈ D # have a
factorization into irreducibles. The irreducible divisor graph of x ∈ D # , will be the graph
G(x) = (V, E) where V = {a ∈ Irr(D) | a | x}, i.e. the set of irreducible divisors of x up to
associate. Given a1 , a2 ∈ Irr(D), a1 a2 ∈ E if and only if a1 a2 | x. Furthermore, n − 1 loops
will be attached to a if an | x. If arbitrarily large powers of a divide x, we allow an infinite
number of loops. They define the reduced irreducible divisor graph of x to be the subgraph
of G(x) which is formed by deleting all the loops and denote it as G(x).
A clique will refer to a simple (no loops or multiple edges), complete (all vertices are pairwise adjacent) graph. A clique on n ∈ N vertices will be denoted Kn . We will call a graph G
a pseudoclique if G is a complete graph having some number of loops (possibly zero). This
means a clique is still considered a pseudoclique.
Let G be a graph, possibly with loops, and let a ∈ V (G). We have two ways of counting
the degree of this vertex. We define deg(a) := |{a1 ∈ V (G) | a1 6= a, a1 a ∈ E(G)}|, i.e. the
number of distinct vertices adjacent to a. Suppose a vertex a has n loops. We define
degl (a) := n + deg(a), the sum of the degree and the number of loops. Given a, b ∈ V (G), we
4
CHRISTOPHER PARK MOONEY
define d(a, b) to be the shortest path between a and b. If no such path exists, i.e. a and b are
in disconnected components of G, or the shortest path is infinite, then we say d(a, b) = ∞.
We define Diam(G) :=sup({d(a, b) | a, b ∈ V (G)}).
2.2. τ -Factorization Definitions.
Let D be a domain. Let D ∗ = D − {0}, let U(D) be the set of units of D, and let
D # = D ∗ − U(D) be the non-zero, non-units of D. We will say a and b are associates if
a = λb for some λ ∈ U(D), and denote this a ∼ b. Let τ be a relation on D # , that is,
τ ⊆ D # × D # . We will always assume further that τ is symmetric. Let a be a non-unit. A
factorization of the form a = λa1 · · · an is said to be a τ -factorization if ai ∈ D # , λ ∈ U(D)
and ai τ aj for all i 6= j. If n = 1, then this is said to be a trivial τ -factorization. Each ai is
said to be a τ -factor, or that ai τ -divides a, written ai |τ a.
We say that τ is multiplicative (resp. divisive) if for a, b, c ∈ D # (resp. a, b, b′ ∈ D # ), aτ b
and aτ c imply aτ bc (resp. aτ b and b′ | b imply aτ b′ ). We say τ is associate preserving if
for a, b, b′ ∈ D # with b ∼ b′ , aτ b implies aτ b′ . We define a τ -refinement of a τ -factorization
λa1 · · · an to be a factorization of the form
(λλ1 · · · λn ) · b11 · · · b1m1 · b21 · · · b2m2 · · · bn1 · · · bnmn
where ai = λi bi1 · · · bimi is a τ -factorization for each i. This is slightly different from the
original definition in [2] where no unit factor was allowed. One can see they are equivalent
when τ is associate preserving. We then say that τ is refinable if every τ -refinement of a
τ -factorization is a τ -factorization. We say τ is combinable if whenever λa1 · · · an is a τ factorization, then so is each λa1 · · · ai−1 (ai ai+1 )ai+2 · · · an .
We pause briefly to give some examples of particular relations τ .
GENERALIZED IRREDUCIBLE DIVISOR GRAPHS
5
Example 2.1. Let D be a domain and let τ = D # × D # .
This yields the usual factorizations in D and |τ is the same as the usual divides. Moreover,
τ is multiplicative and divisive (hence associate preserving).
Example 2.2. Let D be a domain and let τ = ∅.
For every a ∈ D # , there is only the trivial factorization. Furthermore, all τ -divisors of a are
associate to a. If b | τ a, then the factorization is forced to be of the form a = λb = λ(λ−1 (a))
for λ ∈ U(D). That is, a ∼ b. Again τ is both multiplicative and divisive (vacuously).
Example 2.3. Let D be a domain and let S be a non-empty subset of D # . Let τ = S × S.
Define aτ b ⇔ a, b ∈ S.
In this case, τ is multiplicative (resp. divisive) if and only if S is multiplicatively closed
(resp. closed under non-unit factors). A non-trivial τ -factorization is (up to unit factors) a
factorization into elements from S. Some examples of nice sets S might be the set of primes
or irreducibles, then a τ -factorization is a prime decomposition or an atomic factorization
respectively.
Example 2.4. Let D be a domain and let aτ b if and only if (a, b) = D
In this case we get the comaximal factorizations studied by S. McAdam and R. Swan in
[11]. More generally, as in J. Juett in [10], we could let ⋆ be a star-operation on D and define
aτ b ⇔ (a, b)⋆ = D, that is a and b are ⋆-coprime or ⋆-comaximal.
Let a ∈ D # . As in [2], we will say a is τ -irreducible or a τ -atom if factorizations of the
form a = λ(λ−1 a) are the only τ -factorizations of a. Then D is said to be τ -atomic if every
6
CHRISTOPHER PARK MOONEY
a ∈ D # has a τ -factorization a = λa1 · · · an with ai being τ -atomic for all 1 ≤ i ≤ n. We
will call such a factorization a τ -atomic-factorization. We say D satisfies τ -ascending chain
condition on principal ideals (τ -ACCP) if for every chain (a0 ) ⊆ (a1 ) ⊆ · · · ⊆ (ai ) ⊆ · · ·
with ai+1 |τ ai , there exists an N ∈ N such that (ai ) = (aN ) for all i > N.
A domain D is said to be a τ -unique factorization domain (τ -UFD) if (1) D is τ -atomic
and (2) for every a ∈ D # any two τ -atomic factorizations a = λ1 a1 · · · an = λ2 b1 · · · bm have
m = n and there is a rearrangement so that ai and bi are associate for each 1 ≤ i ≤ n.
A domain D is said to be a τ -half factorization domain (τ -HFD) if (1) D is τ -atomic and
(2) for every a ∈ D # any two τ -atomic-factorizations have the same length. A domain D
is said to be a τ -finite factorization domain (τ -FFD) if for every a ∈ D # there are only a
finite number of τ -factorizations up to rearrangement and associate. A domain D is said
to be a τ -weak finite factorization domain (τ -WFFD) if for every a ∈ D # , there are only
finitely many b ∈ D such that b is a τ -divisor of a up to associate. A domain D is said
to be a τ -irreducible divisor finite domain (idf-domain) if for every a ∈ D # , there are only
finitely many τ -atomic τ -divisors of a up to associate. A domain D is said to be a τ -bounded
factorization domain (τ -BFD) if for every a ∈ D # , there exists a natural number N(a) such
that for any τ -factorization a = λa1 · · · an , n ≤ N(a).
We have the following set of relationships between the above finite factorization properties from [2], where ∇ indicates the relationship requires τ to be refinable and associate
GENERALIZED IRREDUCIBLE DIVISOR GRAPHS
7
preserving.
2:
❧❧❧
❧❧❧
❧
❧
❧❧❧
❧❧❧
τ -UFD
❘❘❘❘
❘❘❘ ∇
❘❘❘❘
❘❘❘
$,
τ -HFD ❙
❙❙❙
❙❙❙❙∇
❙❙❙
❙❙❙❙
%-
τ -FFD
KS
∇
∇
τ -BFD
19
❦❦❦
❦❦❦
❦
❦
❦
❦❦❦
❦❦❦
+3 τ -WFFD
❧
❧
❧
❧❧❧
❧❧❧
❧
❧
qy ❧❧
∇
+3
τ -ACCP
+3
τ -atomic
τ -atomic τ -idf domain
3. The τ -Irreducible Divisor Graph and Examples
Let D be a domain and let τ be a symmetric and associate preserving relation on D # . Let
Irrτ (D) be the collection of τ -irreducible elements. We will let Irrτ (D) be fixed, pre-chosen
coset representatives of the cosets {aU(D) | a ∈ Irrτ (D)}. Given an element a ∈ D # with
a τ -atomic factorization, we define the τ -irreducible divisor graph of x to be Gτ (x) = (V, E)
with V = {a ∈ Irrτ (D) | a |τ x}, and given a1 , a2 ∈ Irrτ (D), a1 a2 ∈ E if and only if
there is a τ -factorization of the form x = λa1 a2 · · · an . Furthermore, n − 1 loops will be
attached to the vertex corresponding to a if there is a τ -atomic factorization of the form
x = λa · · · aa1 · · · an where a occurs n times. Again, if this occurs for arbitrarily large powers
of a, we allow the possibility of an infinite number of loops. We then define the reduced
τ -irreducible divisor graph of x to be the subgraph of Gτ (x) formed by removing all loops,
and denote it Gτ (x). If there are no τ -atoms in D, then D is said to be a τ -antimatter
domain. In this case, we define Gτ (x) = Gτ (x) = ∅ for all x ∈ D # .
Example 3.1. Let D be a domain. Suppose τ = ∅.
8
CHRISTOPHER PARK MOONEY
By letting τ = ∅, we have eliminated all non-trivial τ -factorizations. This has the effect of
making every non-zero, non-unit a τ -atom. This means for every x ∈ D # , Gτ (x) = ({x}, ∅).
The τ -irreducible divisor graph of x consists of the single vertex, x itself (or whichever
associate of x chosen initially in Irrτ (D)).
Example 3.2. Let D be a domain. Suppose τ = D # × D # .
In this case, we are looking at the usual factorizations in D. This means a | b if and only
if a |τ b. Every τ -factorization is a usual factorization and conversely. Moreover, x ∈ D #
is τ -atomic if and only if x is atomic. This results in the irreducible divisor graphs being
identical. Hence, we have Gτ (x) = G(x) and Gτ (x) = G(x) as defined in Coykendall and
Maney [9].
Example 3.3. (Inspired by [9, Example 2.4]) Let D = Q[x2 , x3 ] and let f (x) = x8 − x9 and
consider the relation τ defined by g(x)τ g ′ (x) if and only if deg(g(x)) = deg(g ′(x)).
The irreducible factorizations of f (x) are
f (x) = x8 − x9 = x2 · x2 · x2 · (x2 − x3 ) = (x3 − x4 ) · x2 · x3 = (x2 − x3 ) · x3 · x3 .
It is clear that only the last factorization above is a τ -factorization. It is the only factorization
in which all of the factors have the same degree. Every irreducible is certainly τ -irreducible,
but the question is: are there any additional τ -atomic elements which τ -divide f (x)? For
instance, g(x) = x5 is τ -atomic since the only non-trivial factorization up to associates and
rearrangement is g(x) = x2 · x3 is thrown out due to the factors having different degrees.
Any τ -factorization of f (x) must have τ -factors of the same degree. This amounts to finding
a proper partition of deg(f (x)) = 9 into parts of equal size. This is done only by nine degree
GENERALIZED IRREDUCIBLE DIVISOR GRAPHS
9
1 parts, or three degree 3 parts. Since all non-units in D have degree at least 2, the former is
not possible in this ring. This leaves only one possible τ -atomic factorization up to associates
and rearrangement:
f (x) = (x2 − x3 ) · x3 · x3 .
We show both G(f (x)) and Gτ (f (x)) below in Figure 1 for the sake of comparison.
x3
x3
x3-x2
x4-x3
x2
x3-x2
(a)
(b)
Figure 1. (a) G(f (x)), (b) Gτ (f (x))
This yields the following corresponding reduced irreducible divisor graphs G(f (x)) and
Gτ (f (x)) in Figure 2.
x3
x3
x3-x2
x4-x3
x2
(a)
x3-x2
(b)
Figure 2. (a) G(f (x)), (b) Gτ (f (x))
Remark. This above partition argument shows that any polynomial of prime degree p is a
τ -atom. The only way to non-trivially partition a prime integer p into equal parts is p parts
of degree 1 each. But all non-unit elements have degree at least 2. Hence there can only be
the trivial factorization. Thus we know the τ -irreducible divisor graph of any polynomial of
prime degree, p(x) consists of the single vertex, p(x) (or possibly whichever associate of p(x)
10
CHRISTOPHER PARK MOONEY
was taken in Irrτ (D)). It is perhaps worth noting that the converse is false as conveniently
demonstrated above by x4 − x3 being irreducible and hence τ -irreducible, despite having
composite degree.
4. The τ -Irreducible Divisor Graph and τ -Finite Factorization Properties
We first observe that the τ -irreducible divisor graph yields a new equivalent characterization of a τ -irreducible element of [2]. This formalizes something we observed in Example
3.1.
Theorem 4.1. Let D be a domain and τ a symmetric, associate preserving relation on D # .
If D is τ -atomic, then a non-unit x is τ -irreducible if and only if Gτ (x) ∼
= K1 , the complete
graph on a single vertex which is some associate of x.
Proof. If x is τ -irreducible, then the only τ -factorizations are of the form x = λ(λ−1 x), hence
the only τ -divisors of x are associates of x. Moreover, x is a non-zero, non-unit, so x2 does
not divide x, hence there are no loops. Thus, Gτ (x) is a single isolated vertex generated by
whichever associate of x was chosen in Irrτ (D). Conversely, suppose x ∈ R is a non-unit such
that Gτ (x) ∼
= K1 . We suppose x were not τ -irreducible. Then there is a non-trivial τ -atomic
factorization of the form x = λa1 · · · an with n ≥ 2. This yields a1 , a2 ∈ V (Gτ (x)), but there
is only one vertex and no loops in Gτ (x), contradicting the hypothesis and completing the
proof.
Theorem 4.2. Let D be a domain and τ a symmetric, refinable and associate preserving
relation on D # . If D is τ -atomic such that for all x ∈ D # , and for all a ∈ V (Gτ (x)),
degl(a) < ∞, then D satisfies τ -ACCP.
GENERALIZED IRREDUCIBLE DIVISOR GRAPHS
11
Proof. Suppose D did not satisfy τ -ACCP. Then there exists a chain of principal ideals
(x1 ) ( (x2 ) ( (x3 ) ( · · · such that xi+1 |τ xi . Say
xi = λi xi+1 · ai1 · · · aini
(1)
is a τ -factorization for each i. Because D is τ -atomic and τ is refinable and associate
preserving, we may replace each aij with a τ -atomic factorization. This allows us to assume
each factor in Equation (1) is a τ -atom. Since τ is associate preserving, we may assume
further that each aij ∈ Irrτ (D). Because τ is refinable,
(2)
x1 = λ1 x2 · a11 · · · a1n1 = (λ1 λ2 )x3 · a21 · · · a2n2 · a11 · · · a1n1 = · · ·
are all τ -factorizations with aij τ -atomic. Because xi ( xi+1 , we must have in Equation
(1) that ni ≥ 1 or else xi ∼ xi+1 . This means the factorizations in each iteration of Equation
(2) increase in length. If {aij } is infinite, then a11 has an infinite number of adjacent vertices
in Gτ (x1 ), i.e degl(a11 ) ≥ deg(a11 ) = ∞. Otherwise, if {aij } is finite, then one of the ai0 j0
for some i0 and j0 occurs an infinite number of times. Hence degl(ai0 j0 ) = ∞ in Gτ (x1 )
since arbitrarily high powers of ai0 j0 τ -divide x1 . This is a contradiction and D must satisfy
τ -ACCP as desired.
Theorem 4.3. Let D be a domain and τ a symmetric and associate preserving relation on
D # . If D is τ -atomic, then the following are equivalent.
(1) D is a τ -UFD.
(2) Gτ (x) is a pseudoclique for every x ∈ D # .
(3) Gτ (x) is a clique for every x ∈ D # .
12
CHRISTOPHER PARK MOONEY
(4) Gτ (x) is connected and Diam(Gτ (x)) = 1 for every x ∈ D # .
(5) Gτ (x) is connected and Diam(Gτ (x)) = 1 for every x ∈ D # .
(6) Gτ (x) is connected for every x ∈ D # .
(7) Gτ (x) is connected for every x ∈ D # .
Proof. (1) ⇒ (2) Let x ∈ D # . Let x = λa1 · · · am be the unique τ -atomic factorization of
x up to rearrangement and associates. Because τ is associate preserving, one may adjust
the unit factor in front if necessary, to assume without loss of generality ai ∈ Irr(D) for
each i. After rearrangement, we may take the first n τ -factors to be distinct up to associate,
and assume the last m − n τ -factors are repeated associates of a1 , . . . , an . Since these are
the only irreducible τ -divisors of x, V (Gτ (x)) = {a1 , . . . , an }. It is clear that given any
two vertices, ai , aj ∈ V (Gτ (x)), ai aj ∈ E(Gτ (x)), because ai τ aj and ai aj |τ x. If there are
repeated associates in the τ -factorization, say x = λae11 · · · aenn with n ≤ m is a factorization
with e1 + e2 + . . . + en = m. Then if ei > 1, we have ai τ ai and ai ai |τ x, so we have ei − 1
loops attached to ai . Hence Gτ (x) is a pseudoclique.
It is immediate from the definition of the reduced τ -irreducible divisor graph that (2) ⇔
(3) and (6) ⇔ (7). Furthermore, (2) ⇔ (4) and (3) ⇔ (5) respectively since a connected
graph is complete if and only if the diameter is 1.
If Gτ (x) is complete it is certainly connected, so (2) ⇒ (6). It now suffices to show that (6)
⇒ (1). The following is a modification of the proof of [6, Theorem 2.1]. Let A be the set of all
non-zero, non-units which admit at least two distinct τ -atomic factorizations. We show A =
∅. Suppose otherwise and let n := minx∈A {k | x = λa1 · · · ak is a τ -atomic factorization}.
Clearly, n ≥ 2. Let y ∈ A such that y = λa1 · · · an . Then there is a distinct τ -atomic
factorization y = µb1 · · · bt with t ≥ n. There is a path in Gτ (x) connecting b1 and a1 ,
GENERALIZED IRREDUCIBLE DIVISOR GRAPHS
13
so without loss of generality we may as well have picked at τ -atomic factorization such
that b1 and a1 are actually adjacent in Gτ (x). If ai and bj were associates for any 1 ≤
i ≤ n, 1 ≤ j ≤ t, then
y
ai
= λ′ a1 · · · abi · · · an = µ′ b1 · · · bbk · · · bt provides two distinct τ -
atomic factorizations of a non-zero, non-unit but contradicts the minimality of n. Since
we chose a1 and b1 to be adjacent, we know there is a τ -atomic factorization of the form
y = γa1 b1 c1 · · · cm = λa1 · · · an . But this again yields two distinct τ -atomic factorizations of
y
:
a1
y
= γb1 c1 · · · cm = λa2 · · · an
a1
again contradicting the minimality of n. Hence, A must be empty as desired, completing
the proof.
Theorem 4.4. Let D be a domain and τ a symmetric, refinable and associate preserving
relation on D # . If D is τ -atomic, then consider the following statements.
(1) Gτ (x) is finite for every x ∈ D # .
(2) Gτ (x) is finite for every x ∈ D # .
(3) D is a τ -irreducible divisor finite domain.
(4) D is a τ -weak finite factorization domain.
(5) D is a τ -finite factorization domain.
(6) For all x ∈ D # , degl(a) < ∞ for all a ∈ V (Gτ (x)).
(7) For all x ∈ D # , deg(a) < ∞ for all a ∈ V (Gτ (x)).
We have (1)-(5) are equivalent, (3) ⇒ (6) ⇒ (7), and if we assume further that τ is reflexive,
then (7) ⇒ (3) and all are equivalent.
Proof. We begin by showing (1)-(5) are equivalent. (1) ⇔ (2) and (3) ⇔ (1) are immediate
from the definitions. (5) ⇒ (4) is clear since every every τ -divisor up to associate must
14
CHRISTOPHER PARK MOONEY
appear as a τ -factor in one of the finitely many τ -factorizations. (4) ⇒ (3) every τ -atomic
divisor up to associate is certainly among the τ -divisors.
We need only prove (3) ⇒ (5). We modify the proof of [1, Theorem 5.1]. Let a ∈ D # .
Let a1 , a2 , . . . an be the collection of all non-associate τ -irreducible divisors of a. Because D
is τ -atomic and τ is refinable and associate preserving, given any τ -factorization, we could
τ -refine it into a τ -atomic factorization. In this way, every τ -factorization corresponds to a
τ -factorization of the form
a = λas11 · · · asnn
with 0 ≤ si . If we can show that the set of factorizations of this form is finite, then every
τ -factorization occurs as some grouping of these τ -atomic factors.
We suppose first that for each i, there is a bound Ni such that every τ -refinement of a
τ -factorization as above has 0 ≤ si ≤ Ni . Then this set is finite, with N1 · N2 · · · Nn elements
in it. So now we must have some si which is unbounded, say s1 is unbounded. Then for each
k ≥ 1, we can write
x = λk as1k1 · · · asnkn
τ -factorizations with sk1 < sk2 < sk3 < · · · . Suppose that in the set of factorizations, {ski }
is bounded for 1 < i ≤ n. Then there are only finitely many choices for sk2 , . . . , skn , we must
have sk2 = sj2 , sk3 = sj3, . . . , skn = sjn for some j > k. But then we have
s
x = λj a1j1 · · · asnjn = λk as1k1 · · · asnkn .
s
Since D is a domain, we can cancel leaving λj a1j1 = λk as1k1 with sj1 > sk1 , a contradiction.
This means some set {ski } is unbounded for a fixed i with 1 < i ≤ n. Without loss of
generality suppose it is i = 2. We continue in this manner until we get subsequences with
GENERALIZED IRREDUCIBLE DIVISOR GRAPHS
15
s11 < s21 < s31 < · · · and s21 < s22 < s32 < · · · . But this means we have
x = λ1 as111 · · · asn1n = λ2 as121 · · · asn2n
with s1i < s2i for each i, a contradiction. Thus this set must be finite, and the proof is
complete.
(5) ⇒ (6) ⇒ (7) is immediate.
We now suppose in addition that τ is reflexive and show (7) ⇒ (3). Suppose there is a
a ∈ D # such that {ai }i∈I is an infinite collection of non-associate τ -atomic divisors of a.
Suppose for each i ∈ I, that a = λi ai ai1 · · · aini are the given τ -factorizations showing ai is a
τ -atomic divisor of a. But then since τ is reflexive, we have aτ a, so
a2 = a · a = (λi ai ai1 · · · aini ) λj aj aj1 · · · ajnj
is a τ -factorization for any choice of i, j ∈ I. Because τ -is refinable and associate preserving,
we have
a2 = (λi λj )ai aj ai1 · · · aini aj1 · · · ajnj
is a τ -factorization, showing ai and aj are adjacent in Gτ (a2 ), for any choice of i, j ∈ I.
Thus when we fix i = i0 , and let j range over all possible choices of aj ∈ {ai }i∈I , we see
ai0 aj ∈ E(Gτ (a2 )) for all j ∈ I, j 6= i0 . Hence deg(ai0 ) = ∞ in Gτ (a2 ).
Corollary 4.5. Let D be a domain and τ a symmetric, refinable and associate preserving
relation on D # . If D is τ -atomic and if Gτ (x) (resp. Gτ (x)) is connected for every x ∈ D # ,
then Gτ (x) (resp. Gτ (x)) is a finite pseudoclique for every x ∈ D # . Furthermore, Gτ (x) =
Kn for some finite n.
16
CHRISTOPHER PARK MOONEY
Proof. If Gτ (x) is connected for every x ∈ D # , then by Theorem 4.3 D is a τ -UFD which
implies D is a τ -atomic τ -irreducible divisor finite ring. Theorem 4.4 shows Gτ (x) has a finite
number of vertices. Moreover, by Theorem 4.3, each of these vertices is adjacent to every
other vertex, i.e. Gτ (x) is a finite a pseudoclique. The last statement is immediate.
The following result is well known and was proven in [2], but we include it since the
τ -irreducible divisor graph results yield a nice proof.
Corollary 4.6. Let D be a domain and τ a symmetric, refinable and associate preserving
relation on D # . If D is a τ -atomic, τ -irreducible divisor finite domain, then D satisfies
τ -ACCP.
Proof. Theorem 4.4 shows for an associate preserving and refinable τ , a τ -atomic τ -idf domain has the property that for all x ∈ D # , degl(a) < ∞ for any vertex a ∈ V (Gτ (x)). By
Theorem 4.2, this shows D satisfies τ -ACCP.
Acknowledgment
I would like to thank the referee for their helpful comments and careful reading which
have improved the quality of this article. This research was conducted as a research fellow
under the supervision of Professor Daniel D. Anderson while at The University of Iowa.
References
[1] D.D. Anderson, D.F. Anderson, and M. Zafrullah, Factorization in Integral Domains, J. Pure Appl.
Algebra., 69 (1990), 1-19.
[2] D. D. Anderson and Andrea M. Frazier, On a general theory of factorization in integral domains, Rocky
Mountain J. Math., 41 (2011), 663–705.
[3] D.D. Anderson and M. Naseer, Beck’s coloring of a commutative ring, J. Algebra, 159 (1993), 500–514.
GENERALIZED IRREDUCIBLE DIVISOR GRAPHS
17
[4] D.F. Anderson, A. Frazier, A. Lauve, and P.S. Livingston, The zero-divisor graph of a commutative ring.
II, Ideal theoretic methods in commutative algebra (Columbia,MO, 1999), Lecture Notes in Pure and
Appl. Math., 220, Dekker, New York, (2001), 61–72.
[5] D.F. Anderson and P.S. Livingston, The zero-divisor graph of a commutative ring, J. Algebra, 217 (1999),
434–447.
[6] M. Axtell, N. Baeth, and J. Stickles, Irreducible divisor graphs and factorization properties of domains,
Comm. Algebra, 39 (2011), 4148–4162.
[7] M. Axtell and J. Stickles, Irreducible divisor graphs in commutative rings with zero-divisors, Comm.
Algebra, 36 (2008), 1883–1893.
[8] I. Beck, Coloring of commutative rings, J. Algebra, 116 (1988), 208–226.
[9] J. Coykendall and J. Maney, Irreducible divisor graphs, Comm. Algebra, 35 (2007), 885–895.
[10] J. Juett, Generalized comaximal factorization of ideals, J. Algebra, 352 (2012), 141–166.
[11] S. McAdam and R. Swan, Unique comaximal factorization, J. Algebra, 276 (2004), 180–192.
[12] C.P. Mooney, Generalized factorization in commutative rings with zero-divisors, Houston J. Math., to
appear.
[13] C.P. Mooney, Generalized u-factorization in commutative rings with zero-divisors, Rocky Mountain J.
Math., to appear.
Reinhart Center, Viterbo University, 900 Viterbo Drive, La Crosse, WI 54601
E-mail address: [email protected]
| 0math.AC
|
Lecture Notes on the ARV Algorithm for Sparsest Cut
Thomas Rothvoss∗
University of Washington, Seattle
arXiv:1607.00854v1 [cs.DS] 4 Jul 2016
July 5, 2016
Abstract
p
One of the landmarks in approximation algorithms is the O( log n)-approximation algorithm for
the Uniform Sparsest Cut problem by Arora, Rao and Vazirani from 2004. The algorithm is based on a
semidefinite program that finds an embedding of the nodes respecting the triangle inequality. Their
core argument shows that a random
p hyperplane approach will find two large sets of Θ(n) many nodes
each that have a distance of Θ(1/ log n) to each other if measured in terms of k · k22 .
Here we give a detailed set of lecture notes describing the algorithm. For the proof of the Structure
Theorem we use a cleaner argument based on expected maxima over k-neighborhoods that significantly simplifies the analysis.
1 Introduction
Let G = (V, E ) be a complete, undirected graph on |V | = n nodes and let c : E → R≥0 be a cost function
on the edges. For a subset S ⊆ V of nodes, let δ(S) := {{i , j } ∈ E | |{i , j } ∩ S| = 1} be the induced cut. We
P
abbreviate c(δ(S)) := e∈δ(S) c(e) as the cost of the cut. The (Uniform) Sparsest Cut problem is then to
find the cut that minimizes the cost-over-separated-pairs ratio:
½
¾
c(δ(U ))
min
| ; ⊂U ⊂ V .
|U | · |V \U |
There is also a non-uniform version of the problem where each pair i , j ∈ V has an associated demand
d (i , j ) ≥ 0 and one aims for the cut minimizing the ratio c(δ(S))/d
(δ(S)). We will now see the celebrated
p
algorithm by Arora, Rao and Vazirani [ARV04] that finds a O( log n)-approximation in polynomial time.
For the algorithm we will not try to optimize any constant. To fix some notation, we will denote any
vector in bold font, like v i ∈ Rm . If we write i ∼ V , then we mean that i is a uniform random node from V .
We denote N (0, 1) as the 1-dimensional Gaussian distribution with mean 0 and variance 1. In particular,
2
a random variable g ∼ N (0, 1) has density p1 e −x /2 . If we write g ∼ N m (0, 1), then we mean that g is
2π
an m-dimensional Gaussian. Recall that the vector g = (g 1 , . . . , g m ) can be generated by sampling each
coordinate independently with g i ∼ N (0, 1). In reverse, for any pair of orthonormal vectors u, v ∈ Rm , the
inner products 〈g , u〉 , 〈g , v 〉 are independently distributed from N (0, 1).
∗ Email: [email protected]. Supported by NSF grant 1420180 with title “Limitations of convex relaxations in combinatorial
optimization” and an Alfred P. Sloan Research Fellowship.
1
2 A semidefinite program
Sparsest Cut is an unusual problem in the sense that it minimizes the ratio of two functions. Let us
assume for the sake of simplicity that we guessed the cost C ∗ and the size S ∗ of an optimum cut, say with
S ∗ ≤ n2 . Then we define a semidefinite programming relaxation
X
X
i ∈V
V
2
{i , j }∈(
X
{i , j }∈E
)
kv i k22
kv i − v j k22
c i j kv i − v j k22
kv i − v j k22
kv i k22
kv 0 k22
= S∗
(I )
= S ∗ · (n − S ∗ )
(I I )
= C∗
(I I I )
≤ kv i − v k k22 + kv k − v j k22
≤ 1
= 0
∀i , j , k ∈ V ∪ {0}
∀i ∈ V
(I V )
(V )
(V I )
where we use an artificial index 0 with v 0 := 0, so that the triangle inequality also holds for the origin.
Lemma 1. If there is cut U ∗ ⊆ V of cost C ∗ and size S ∗ , then the above SDP has a solution.
Proof. One could choose 1-dimensional vectors by defining
(
1 if i ∈ U ∗
v i :=
0 if i ∉ U ∗ .
Then the only non-trivial case is verifying the triangle inequalities in (I V ). These are satisfied by our
choice of v i ∈ {0, 1} since if kv i − v j k22 = 1, then i and j have to be on different sides of the cut U ∗ and any
node k has to be either not on the side of i or not on the side of j .
We can solve the semi-definite program (SDP) in polynomial time [GLS93]; let {v i }i ∈V ⊆ Rm be the
solution. Due to the triangle inequalities (IV) we can define a metric d : V × V → R≥0 by setting d (i , j ) :=
kv i − v j k22 . Note that while k·k2 is always a metric, k·k22 is not a metric on all points sets. For sets of nodes
A, B ⊆ V we define d (i , A) := min{d (i , j ) : j ∈ A} and d (A, B ) := min{d (i , j ) : i ∈ A, j ∈ B }.
3 A ball rounding scheme
Given that family {v i }i ∈V of SDP vectors, there are several natural rounding procedures that would come
to mind. For example one could try the hyperplane rounding that Goemans and Williamson [GW94]
have used for MaxCut. The natural algorithm for Sparsest Cut would be to take a random Gaussian
g ∼ N m (0, 1) and set U := {i ∈ V | 〈g , v i 〉 ≥ 0}. Assume for the sake of simplicity that we are in the balanced
case of Sparsest Cut with S ∗ = Θ(n). Then an edge (i , j ) ∈ E has a contribution to the objective function
of Θ(d (i , j )/n 2 ). On the other hand, the probability that (i , j ) is separated is roughly proportional to the
Euclidean distance kv i − v j k2 and even if the hyperplane generates
p perfectly balanced cuts, the expected
contribution of an edge (i , j ) to the hyperplane cut would be Θ( d (i , j )/n 2 ). In other words, short edges
would be separated far too likely.
The second best idea that one might have, would be to select a node i ∈ V and take a random cut
U := { j ∈ V | d (i , j ) ≤ r } where r ∼ [0, 1]. Now every edge is cut with a probability not exceeding d (i , j ).
On the other hand, this argument seems to not give any guarantee on the size of U and V \U , hence the
2
objective function can be arbitrarily bad again. But a slight fix of this rounding argument can work. We
only need a large “core” of nodes so that the remaining nodes still have a decently large distance to it.
Lemma 2. Suppose we have a set of nodes A ⊆ V with |A| ≥ αn and
2
cut of the form {i ∈ V | d (i , A) ≤ r } is a αβ
-approximation.
Proof. Recall that the optimum value of the objective function is
and take U := U (r ) := {i ∈ V | d (i , A) ≤ r } as a random cut.
P
i ∈V
d (i , A) ≥ β · S ∗ . Then the best
C∗
.
S ∗ ·(n−S ∗ )
Suppose we sample r ∼ [0, 1]
U := U (r )
r
A
Then for an edge (i , j ) ∈ E , say with d (i , A) ≤ d ( j , A), we have
Pr [(i , j ) ∈ δ(U )] = Pr [d (i , A) ≤ r ≤ d ( j , A)]
r ∼[0,1]
triangle inequality
r ∼[0,1]
≤
d (i , j ).
Hence the expected cost of the cut U is
E [c(δ(U ))] ≤
r ∼[0,1]
X
(i , j )∈E
d (i , j ) = C ∗ .
Note that in any case |U | ≥ |A| ≥ αn. We know that d (i , A) ≤ 2 for all i ∈ V and hence Prr ∼[0,1] [i ∉ U ] =
Prr ∼[0,1] [r < d (i , A)] ≥ 12 d (i , A). Hence
E [|V \U |] ≥
r ∼[0,1]
1X
β
d (i , A) ≥ · S ∗ .
2 i ∈V
2
αβ
Then Er ∼[0,1] [|U | · |V \U |] ≥ 2 · S ∗ · n. In other words, the random cut seems to have the right expected
nominator and denominator to satisfy the claim. But this is not enough to argue that their ratio satisfies
c(δ(U ))
2
C∗
· S ∗ ·(n−S
Er ∼[0,1] [ |U |·|V \U | ] ≤ αβ
∗ ) . The following insight comes to rescue:
Fact. Let a, b ∈ Rm
≥0 be non-negative numbers and D be a distribution over indices in [m].
Then
½ ¾
ai
Ei ∼D [a i ]
.
≤
min
i ∈{1,...,m} b i
Ei ∼D [b i ]
Now, this fact implies that best choice of U (over all r ∈ [0, 1]) will indeed satisfy the claim and the lemma
is proven.
We should also remark that if A is given, we can find the cut U in polynomial time as we only need to
try out at most n 2 many values of r .
3
4 The case of heavy clusters
Let B (i , r ) := { j ∈ V | d (i , j ) ≤ r } be the “ball” of radius r around node i . A slight annoyance of the ARV
algorithm is that it requires a case split. If we can find a cluster center i ∗ ∈ V , then we can use Lemma 2
to get a constant factor approximation by just taking a ball around the center i ∗ .
∗
Lemma 3. Suppose there is a node i ∗ ∈ V ∪ {0} with |B (i ∗ , 18 · Sn )| ≥ n4 . Then in polynomial time one can
find a cut that gives a O(1)-approximation.
∗
Proof. We set A := B (i ∗ , 18 · Sn ). Then by assumption |A| ≥ n4 . Moreover, bounding the average distance
of pairs of nodes from above and from below gives
³
triangle ineq.
1 S ∗ S ∗ n − S ∗ (I I )
1 S∗ ´
.
·
≤
·
≤ E [d (i , j )]
≤
2 · E d (i , A) + ·
2 n
n | {z
n }
8 n
i , j ∼V
i ∼V
≥1/2
∗
This can be rearranged to Ei ∼V [d (i , A)] ≥ 81 · Sn . We obtain a 64-approximation by applying Lemma 2.
5 An algorithm for the main case
From now on we make the assumption that no cluster exists:
¯ ³
∗ ´¯¯
¯
¯B i , 1 · S ¯ < n ∀i ∈ V ∪ {0}.
¯
8 n ¯ 4
∗
We will prove that in this case, there are sets L, R ⊆ V of size |L|, |R| ≥ Ω(n) with d (L, R) ≥ ∆ · Sn for ∆ :=
p
Θ(1/ log n). Then choosing Lemma 2 with A := L will give a O( ∆1 )-approximation. Before we start prov∗
ing this, we want to further simplify the situation. Note that by (I ) we have Ei ∼V [d (i , 0)] = Sn , and hence
∗
∗
at most half the nodes can have a distance of more than 2 · Sn to 0. Moreover we have |B (0, 18 · Sn )| ≤ n4 .
∗
∗
Then we only loose a constant factor if we delete those nodes and assume that 18 · Sn ≤ d (i , 0) ≤ 2 · Sn for
p
all remaining nodes. Next we scale the vectors v i by a factor of n/S ∗ , which scales the distances d (i , j )
by a factor of Sn∗ . After this transformation it suffices to prove the following structure theorem:
Theorem 4 (ARV Structure Theorem). Given any set of |V | = n vectors {v i }i ∈V ⊆ Rm with
and |B (i , 18 )| ≤ 43 n for all i ∈ V that satisfy the triangle inequalities
1
8
≤ kv i k22 ≤ 2
kv i − v j k22 ≤ kv i − v k k22 + kv k − v j k22 ∀i , j , k ∈ V.
Then there is a polynomial time algorithm
pthat with constant probability finds sets L, R ⊆ V of size
|L|, |R| ≥ Ω(n) with d (L, R) ≥ ∆ for ∆ := Θ(1/ log n).
From now on, we complete ignore the cost function and only use the properties given in the Structure Theorem. Such sets L and R with d (L, R) ≥ ∆ are called ∆-separated. Let c ′ > 0 be a small enough
constants. The algorithm to produce the ∆-separated sets is as follows:
4
Well-separated sets algorithm
Input: Vectors {v i }i ∈V ⊆ Rm satisfying the triangle inequality with
3
4 n for all i ∈ V ∪ {0}.
Output: Either ∆-separated sets L ′ , R ′ of size |L ′ |, |R ′ | ≥
(1)
(2)
(3)
(4)
(5)
(6)
c′
2n
1
8
≤ kv i k22 ≤ 2 and |B (i , 18 )| ≤
or FAIL
m
Pick a random Gaussian g ∼ N (0, 1)
Set L := {i ∈ V | 〈v i , g 〉 ≤ −1} and R := {i ∈ V | 〈v i , g 〉 ≥ 1}
If either |L| ≤ c ′ n or |R| ≤ c ′ n then FAIL
Compute any inclusion-wise maximal matching M (g ) ⊆ {(i , j ) ∈ L × R | d (i , j ) ≤ ∆}
′
If |M (g )| > c2 n then FAIL
Return L ′ := {i ∈ L | i not covered by M (g )} and R ′ := {i ∈ V | i not covered by M (g )}
Observe that any pair i ∈ L ′ and j ∈ R ′ that remains will have d (i , j ) > ∆ as otherwise the matching M (g )
′
would not have been maximal. Also, if the algorithm reaches (6), then min{|L ′ |, |R ′ |} ≥ c2 n.
L
R
L′
R′
i
∈ M (g )
j
d (i , j ) ≤ ∆
g
The first step is to argue that with constant probability both L and R have size at least Ω(n).
Lemma 5. There is an absolute constant c ′ > 0 so that Prg ∼N m (0,1) [min{|L|, |R|} ≥ c ′ n] ≥ c ′ .
Proof. We will prove that E[|L| · |R|] ≥ Ω(n 2 ) which then implies the claim. Fix any i ∈ V and select one of
the at least 41 n nodes j with d (i , j ) ≥ 81 . Let w be the orthogonal projection of v j on v i , see figure below.
Let α ∈ [0, π2 ] be the angle between v i − v j and w and let β ∈ [0, π2 ] be the angle between v j and w . Due to
the triangle inequalities, the angle spanned by the points 0, v i and v j is non-obtuse and α + β ≤ π2 . Then
we have either α ≤ π4 and
1
kw k2 = cos(α) · kv i − v j k2 ≥
| {z } | {z } 4
p
p
≥1/ 8
≥1/ 2
or otherwise we have β ≤
π
4
and
1
kw k2 = cos(β) · kv j k2 ≥
| {z } | {z } 4
p
≥1/ 2
5
p
≥1/ 8
Either way, kw k2 ≥ 41 . Since v i ⊥ w , the inner products 〈g , v i 〉 and 〈g , w 〉 are independent random variables and we can estimate
Pr[i ∈ L and j ∈ R] ≥ Pr[−2 ≤ 〈g , v i 〉 ≤ −1 and 〈g , w 〉 ≥ 3]
h
h
vi
1
2 i
3 i
w
〉≤−
〉≥
≤ 〈g ,
= Pr −
>0
· Pr 〈g ,
kv i k2
kv i k2
kv i k2
kw k2
kw k
| {z 2}
| {z
}
∈[ p1 , p1 ]
8
≤12
2
which is some tiny, yet absolute constant. Note that in case the latter event happens, then indeed
〈v i , v j 〉
〈g , v j 〉 = 〈g , v i 〉 ·
+ 〈g , w 〉 ≥ 1.
| {z } kv i k2 | {z }
2
| {z }
≥−2
≥3
∈[0,1]
vi
w
p
2
0
α
β
vj
p1
8
This implies that with a constant probability, the algorithm does not fail in (3). The main technical
part lies in proving that Eg ∼N m (0,1) [|M (g )|] ≤ cn where we can make the constant c as small as we want,
′
′
′
at the expense of a smaller value of ∆. If we choose c := ( c2 )2 , then Pr[|M (g )| > c2 ] ≤ c2 and the success
probability of the algorithm is at least
c′
.
2
6 The proof of the Structure Theorem
p
The following geometric theorem by Arora, Rao and Vazirani is the heart of their O( log n)-approximation
−2/3
for Sparsest Cut. To be precise, the original ARV
)
p result [ARV04] only showed this theorem for ∆ = Θ((log n)
and needed a lot of extra work to get the O( log n)-approximation. The claim as it is stated here was first
proven by Lee [Lee05]. For an edge set E ′ , let β(E ′ ) be the size of the maximum matching.
p
Theorem 6 ([ARV04, Lee05]). For any constant c > 0 there is a choice of ∆ := Θc (1/ log n) so that the
following holds: Let {v i }i ∈V ⊆ Rm be a set of |V | = n vectors satisfying the triangle inequality
kv i − v j k22 ≤ kv i − v k k22 + kv k − v j k22 ∀i , j , k ∈ V.
For a vector g ∈ Rm define
©
ª
E (g ) := (i , j ) ∈ V × V | 〈v j − v i , g 〉 ≥ 2 and kv i − v j k22 ≤ ∆ .
Then Eg ∼N m (0,1) [β(E (g ))] ≤ cn .
6
Here we think of E (g ) as directed edges. Let M (g ) be a maximum matching attaining β(E (g )). We
will assume the existence of such a matching M (g ) and lead this to a contradiction. By inducing on a
subgraph and reducing the constant c one can even assume that for every node the probability of having
an outgoing edge is at least c and the same is true for ingoing edges. First, there is no harm in assuming
that M (g ) has the reverse edges of M (−g ), which implies that each node has an outgoing edge with the
same probability as it has an incoming edge.
Lemma 7. Assume that Theorem 6 is false for vectors {v i }i ∈V ⊆ Rm . Then there is a subset V ′ ⊆ V of size
|V ′ | ≥ cn and a matching M ′ (g ) ⊆ M (g ) on V ′ so that every node i ∈ V ′ has an outgoing edge in M ′ (g )
with probability at least c8 and an ingoing edge with probability at least c8 .
Proof. For each node i ∈ V define p(i ) := Prg ∼N m (0,1) [i has outgoing edge in M (g )]. If there is a node
i with p(i ) ≤ 8c , then we imagine to delete the node from the graph and remove from M (g ) any edge
containing node i . Note that this decreases the expected size of the matching by at most 2· 8c . We continue
this procedure until no such node exists anymore. Let V ′ be the remaining set of nodes with M ′ (g ) :=
M (g ) ∩ (V ′ × V ′ ). Then Eg ∼N m (0,1) [|M ′ (g )|] ≥ Eg ∼N m (0,1) [|M (g )|] − n · 4c ≥ c2 n. Then there must be at least
|V ′ | ≥ cn many nodes left.
After changing the constants and adapting the value of n, we assume to have n nodes and for every
node i ∈ V , the matching M (g ) has an outgoing and an incoming edge with probability at least c.
We call an edge (i , j ) ∆-short if d (i , j ) ≤ ∆. For a node i ∈ V , let Γ(i ) := { j ∈ V | d (i , j ) ≤ ∆} be the
neighborhood of i with respect to the graph of ∆-short edges. Moreover, let Γk (i ) := Γk−1 (Γ(i )) be the
nodes that can be reached from i via at most k many ∆-short edges.
Lemma 8. For any k ∈ Z≥0 and i ′ ∈ Γk (i ) one has kv i − v i ′ k2 ≤
p
k∆.
Proof. We have kv i − v i ′ k22 = d (i , i ′ ) ≤ k · ∆ by the SDP triangle inequality. Taking square roots gives the
claim.
At the heart of the arguments lies the fact that the value of Lipschitz functions is well concentrated.
Recall that a function F : Rm → R is called L-Lipschitz, if |F (x) − F (y )| ≤ L · kx − y k2 for all x, y ∈ Rm .
Lemma 9 (Concentration for Lipschitz Functions (Sudakov-Tsirelson, Borrell)). Let F : Rm → R be an L 1
2
Lipschitz function with Gaussian mean µ := Eg ∼N m (0,1) [F (g )]. Then Prg ∼N m (0,1) [|F (g )−µ| ≥ α] ≤ 2e − 2 ·(α/L)
for any α ≥ 0.
We define a function
F i ,k (g ) := max{〈g , v i − v j 〉 | j ∈ Γk (i )}.
In other words, F i ,k (g ) gives the maximum inner product 〈g , v i − v j 〉 over all nodes j ∈ V that are within
k many ∆-short edges of node i . Note that F i ,k (g ) ≥ 〈g , v i − v i 〉 = 0 for all g ∈ Rm as i ∈ Γk (i ).
Lemma 10. The function F i ,k : Rm → R is
p
k · ∆-Lipschitz.
Proof. Fix g , g ′ ∈ Rm and assume for the sake of symmetry that F (g ) ≥ F (g ′ ). Let j , j ′ ∈ Γk (i ) be the nodes
7
attaining F (g ) and F (g ′ ), resp. Then
Cauchy-Schwarz
|F (g ) − F (g ′ )| = 〈g , v i − v j 〉 − 〈g ′ , v i − v j ′ 〉 ≤ 〈g − g ′ , v i − v j 〉
≤
kg − g ′ k2 · kv i − v j k2 .
| {z }
p
≤ k·∆ by Lem. 8
Here we used in the first inequality that j ′ ∈ Γk (i ) maximizes 〈g ′ , v i − v j ′ 〉.
Now, let µi ,k := Eg ∼N m (0,1) [F i ,k (g )] be the expected maximum inner product over k-neighbors of node
i ∈ V . One useful argument will be a nice relation between expectations of neighbors:
Lemma 11. For any node i ∈ V and i ′ ∈ Γ(i ) and k ∈ Z≥0 one has µi ′ ,k+1 ≥ µi ,k .
Proof. We have
µi ′ ,k+1
=
Em
g ∼N (0,1)
′
Γk+1 (i )⊇Γk (i )
≥
Em
g ∼N (0,1)
h
£
n
oi
max 〈v i ′ − v i , g 〉 + 〈v i − v j , g 〉 | j ∈ Γk+1 (i ′ )
©
ª¤
max 〈v i − v j , g 〉 | j ∈ Γk (i ) +
E
[〈v i ′ − v i , g 〉] = µi ,k .
{z
}
g ∼N m (0,1)
|
=0
On the other hand, we can get the following upper bound:
p
p
Lemma 12. For any k ∈ Z≥0 and any i ∈ V one has µi ,k ≤ 10 log n · k∆.
Proof. For any j ∈ Γk (i ) we have kv i − v j k2 ≤
Pr
g ∼N m (0,1)
h
p
k∆ and generously
q
p i Z∞
〈v i − v j , g 〉 ≥ 8 log n · k∆ ≤ p
8
logn
2
1
1
.
p e −x /2 d x ≤
2n
2π
p
p
p
m (0,1) [F i ,k (g ) ≤ 8
That means
log n· k∆] ≥ 1/2. Again, F i ,k is k∆-Lipschitz, hence Prg ∼N m (0,1) [|F i ,k (g )−
p Prg ∼N
p
µi ,k | ≥ 2 k∆] ≤ 12 and 2 ≤ 2 log(n) for n ≥ 2.
6.1 Extend or expand
The core argument to get to a contradiction is the following:
1
log(1/c) · ∆1 } be parameters and U ⊆ {i ∈ V | µi ,k ≥ δ} be a set
Lemma 13. Let ∆ > 0, δ ∈ R and k ∈ {0, . . . , 100
of nodes. Then
(A) either there is a subset U ′ ⊆ Γ(U ) so that |U ′ | ≥ 4c · |U | and µi ,k+1 ≥ δ + 1 for all i ∈ U ′ .
(B) or the neighborhood U ′ := Γ(U ) satisfies |U ′ | ≥ c4 · |U | and µi ,k+1 ≥ δ for all i ∈ U ′ .
Proof. If |Γ(U )| ≥ 4c · |U |, then every node in i ∈ Γ(U ) has µi ,k+1 ≥ δ by Lemma 11 and we are in case (B).
So suppose that |Γ(U )| < c4 · |U |. Consider the random matching
n
1o
M̃(g ) := (i , j ) ∈ M (g ) | i ∈ U and F i ,k (g ) ≥ δ −
2
8
that is the restriction of M (g ) to edges that are going out
p of U and where F i ,k (g ) is large enough. Note
m
that µi ,k = Eg ∼N (0,1) [F i ,k (g )] ≥ δ for all i ∈ U and F i ,k is k∆-Lipschitz. Hence by Lemma 10 we have
³
h
1 ´ c
1i
≤ 2 exp −
<
F i ,k (g ) < δ −
(0,1)
2
8k∆
2
Pr
m
g ∼N
∀i ∈ U .
This implies that each node in U will have an outgoing edge in M̃ (g ) with probability at least 2c . Now,
2
c
define U ′ := { j ∈ Γ(U ) | Pr[ j has incoming edge from M̃ (g )] ≥ 16
}. Since M̃(g ) is a matching we have
c
c2
c
· |U | ≤ E[|M̃ (g )|] ≤
· |Γ(U ) \U ′ | +|U ′ | ≤ · |U | + |U ′ |
|
{z
}
2
16
4
≤(4/c)·|U |
which implies that |U ′ | ≥ 4c · |U |. Now fix a node j ∈ U ′ . It remains to argue that µ j ,k+1 ≥ δ + 1 for all
j ∈ U ′ . First, condition on the event that M̃ (g ) has an edge incoming to j . We denote that edge by
(i (g ), j ) ∈ M̃ (g ) with i (g ) ∈ U 1 . We know by the definition of M̃ (g ) that 〈v j − v i (g ) , g 〉 ≥ 2. Moreover, we
know that there is a node h(g ) ∈ Γk (i (g )) so that 〈v i (g ) − v h(g ) , g 〉 ≥ δ − 12 .
Γ(U )
U
Γk (i (g ))
U′
M̃ (g )
h(g )
i (g )
j
Then 〈v j − v h(g ) , g 〉 = 〈v j − v i (g ) , g 〉 + 〈v i (g ) − v h(g ) , g 〉 ≥ δ + 32 . In other words, Prg ∼N m (0,1) [F j ,k+1(g ) ≥ δ +
p
3
c2
2 ] ≥ 16 . Again, the function F j ,k+1 is (k + 1)∆-Lipschitz and
´ c2
³
h
1
1 i Lem. 9
<
≤ 2 exp −
|F j ,k+1(g ) − µ j ,k+1 | ≥
(0,1)
2
8(k + 1)∆
16
Pr
m
g ∼N
and hence µ j ,k+1 ≥ δ + 1. This shows the claim.
Now, suppose we run Lemma 13 iteratively, starting with U := V and in each iteration we replace
the current U by the set U ′ . We iterate this until the upper bound on k is reached. Note that case (B)
1
cannot happen more often than case (A) as always |Γ(U )| ≤ n. Then after being k = 100∆
log(1/c) times
in Case (A) and ℓ ∈ {0, . . . k} times in Case (B), we end up with a set U ⊆ V with |U | ≥ n · ( 4c )k−ℓ ≥ n · ( 4c )k
p
p
and µi ,2k ≥ µi ,k+ℓ ≥ k for allp
i ∈ U . On the other hand, µi ,2k ≤ 10 log n · 2k∆ by Lemma 12. Choosing
∆ := Θc ( p 1 ) and k := Θc ( log n) with proper choice of constants, then gives a contradiction.
logn
Acknowledgement. The author is very grateful to James R. Lee, Harishchandra Ramadas, Rebecca Hoberg
and Alireza Rezaei for helpful discussion and comments.
1 Here we write i (g ) to indicate that this node will depend on the choice of g .
9
References
[ARV04] Sanjeev Arora, Satish Rao, and Umesh V. Vazirani. Expander flows, geometric embeddings and
graph partitioning. In László Babai, editor, Proceedings of the 36th Annual ACM Symposium on
Theory of Computing, Chicago, IL, USA, June 13-16, 2004, pages 222–231. ACM, 2004.
[GLS93] Martin Grötschel, László Lovász, and Alexander Schrijver. Geometric algorithms and combinatorial optimization, volume 2 of Algorithms and Combinatorics. Springer-Verlag, Berlin, second
edition, 1993.
[GW94] Michel X. Goemans and David P. Williamson. .879-approximation algorithms for MAX CUT
and MAX 2sat. In Frank Thomson Leighton and Michael T. Goodrich, editors, Proceedings of
the Twenty-Sixth Annual ACM Symposium on Theory of Computing, 23-25 May 1994, Montréal,
Québec, Canada, pages 422–431. ACM, 1994.
[Lee05] James R. Lee. On distance scales, embeddings, and efficient relaxations of the cut cone. In
Proceedings of the Sixteenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2005,
Vancouver, British Columbia, Canada, January 23-25, 2005, pages 92–101. SIAM, 2005.
10
| 8cs.DS
|
arXiv:1707.07184v2 [cs.CV] 24 Nov 2017
A survey of exemplar-based texture synthesis
Lara Raad∗1 , Axel Davy†2 , Agnès Desolneux‡2 , and Jean-Michel
Morel§2
1
Dept. of Information & Communications Technologies, Universitat Pompeu Fabra
2
CMLA, École normale supérieure Paris-Saclay
November 28, 2017
Abstract
Exemplar-based texture synthesis is the process of generating, from an
input sample, new texture images of arbitrary size and which are perceptually equivalent to the sample. The two main approaches are statisticsbased methods and patch re-arrangement methods. In the first class, a
texture is characterized by a statistical signature; then, a random sampling conditioned to this signature produces genuinely different texture
images. The second class boils down to a clever “copy-paste” procedure,
which stitches together large regions of the sample. Hybrid methods try
to combine ideas from both approaches to avoid their hurdles. The recent
approaches using convolutional neural networks fit to this classification,
some being statistical and others performing patch re-arrangement in the
feature space. They produce impressive synthesis on various kinds of
textures. Nevertheless, we found that most real textures are organized at
multiple scales, with global structures revealed at coarse scales and highly
varying details at finer ones. Thus, when confronted with large natural
images of textures the results of state-of-the-art methods degrade rapidly,
and the problem of modeling them remains wide open.
1
Introduction
This paper proposes a review of exemplar-based texture theory, a topic that occupied David Mumford at the end of the last century [97, 98], and again in
∗ [email protected]
† [email protected]
‡ [email protected]
§ [email protected]
The images in this document are lossy compressed. To compare the zoomed-in images,
please refer to the uncompressed pdf, which can be found at http://desolneux.perso.math.
cnrs.fr/papers/exemplar-based-texture-synthesis-v2-full-res.pdf
1
his book on pattern theory [62]. Textures are ubiquitous in our visual environment. In the past fifty years their definition has occupied psychophysicists,
mathematicians and computer scientists who have built increasingly sophisticated models. The main progress on the elusive topic of defining textures has
come from computer graphics with the problem of reproducing other examples
of the same texture given a sample. There is so far no complete mathematical
theory that would, first, give a formal axiomatic of texture, and then prove
that some texture synthesis algorithm matches this definition. Rather, each
exemplar-based texture method formulates its own definition of texture and
sometimes (but rarely) convergence or consistency proofs. The method to work
on texture modeling still relies on a visual exploration of synthesized textures,
their defects and successes being linked to some improvement or shortcoming of
the mathematical model. All the more, texture modeling remains a valid challenge for mathematicians, as textures represent arguably the vaster and most
common class of observable functions. They indeed cover a majority of the area
of most digital images. This article accounts for the very rapid and impressive
recent apparition of new texture synthesis methods with striking results. We
shall retrace their theoretical roots. By performing objective experiments and
not hiding the failures of each method, this paper will uncover some flaws in the
current definition of exemplar-based texture modeling. This will lead us to propose a slightly different definition of the problem that seems to address better
its challenges.
The Oxford Dictionary of English defines texture as the feel, appearance, or
consistency of a surface or a substance. Focusing on visual appearance, texture
is analog to color, a perceived quality of a surface, where the RGB bands are
replaced by the output of a specific bank of filters [62, p.215]. Julesz defined
textures as classes of pictures that cannot be discriminated in preattentive vision and advanced two statistical hypotheses to characterize them [41, 38, 40].
Grenander proposed to use the term “texture” for strictly stationary stochastic
processes [31, p.398]. Giving a precise definition of textures is a slippery task;
in a sense, each model implicitly proposes one and as we will see the jury is still
out.
Exemplar-based texture synthesis is the process of generating, from an input
texture sample, new texture images of arbitrary size and which are perceptually
equivalent to the input. It is common to classify them under the two classical statistical estimation categories: parametric methods and non-parametric
methods. The parametric methods aim at characterizing a given texture sample by estimating a set of statistics which will define an underlying stochastic
process. The new images will then be samples of this stochastic process, i.e.
they will have the same statistics as the input sample. The question here is:
what would be the appropriate set of statistics to yield a correct synthesis for
the wide variety of texture images? The results of these methods are satisfying
but only on a small group of textures, and often fail when important structures
are visible in the input. The non-parametric methods reorganize local neighbourhoods from the input sample in a consistent way to create new texture
images. These methods return impressive visual results. Nevertheless, they of2
ten yield verbatim copies of large parts of the input sample. Furthermore, they
can diverge, starting to reproduce iteratively one part of the input sample and
neglecting the rest of it, thus growing what experts call “garbage”. Because
“non-parametric” methods are not completely parameter-free, and “parametric” methods can have a reduced set of parameters, in this paper we will denote
by patch re-arrangement methods the former and by statistics-based methods
the latter.
What constitutes a texture? The answer depends of course on human perception. But a mathematical formulation can be used to characterize patterns
that are perceived as textures. The statistical characterization of texture images was initiated by Béla Julesz [38, 42]. Julesz was the first to point out
that texture images could be reliably organized according to their N-th order
statistics into groups of textures that are preattentively indistinguishable by
humans [38]. (Focusing on pre-attentive vision helps to reduce the subjective
impact of high level processing.) Julesz [42] demonstrated that many texture
pairs sharing the same second-order statistics would not be discerned by human preattentive vision. This hypothesis constitutes the first Julesz axiom for
texture perception. One consequence of this axiom is that two textures sharing the same Fourier modulus but with different phase should be perceptually
equivalent. Indeed, the square Fourier modulus of an image corresponds to its
spatial auto-correlation, thus the second-order statistics. This motivates a class
of algorithms (the random phase methods) aiming at creating textures with a
given second-order statistic. An example of such algorithms is [87]. In a more
recent extension [23], a texture is generated by randomizing the Fourier phase
while maintaining the Fourier modulus. The Random Phase Noise method in
[23] correctly synthesizes textures with no salient details, namely microtexture,
which adapt well to a Gaussian distribution, but it fails for more structured
ones, macrotextures, as can be experimented in the executable paper [22]. Indeed, textures may share the same second and even third order statistics while
being visually different [43, 11]. This led Julesz [40, 39] to propose a second
theory to explain texture preattentive discrimination by introducing the notion
of textons. Textons are local conspicuous features like bars or corners. Giving a
mathematical definition for textons is far from trivial and was studied in for example [95, 17]. Julesz’ second theory states that only the first order statistics of
these textons are relevant for texture perception: images having the same texton
densities (thus, just a first order statistic) could not be discriminated. Texton
theory proposes the main axiom that texture perception is invariant to random
shifts of the textons [40]. This axiom is extensively used in the stochastic dead
leaves models [61, 77, 8].
Several models of the early visual processing in mammals are based on a
multiscale analysis with Gabor kernels, and are used in particular for modeling the perception of texture [5, 82, 59]. Wavelet analysis provided a natural
frame for these models and resulted in effective methods for texture classification and segmentation [12, 48, 84, 60]. Heeger and Bergen [33] extended Julesz’
approach to multiscale statistics. They characterized a texture sample by the
histograms of its wavelet coefficients. By enforcing the same histograms on a
3
white noise image they obtained a new multiscale exemplar-based texture synthesis method. Yet this method only measures marginal statistics. It misses
important correlations between pixels across scales and orientations which are
crucial to generate edges and conspicuous patterns. We refer to the on-line execution of this method [9] where some successes but many failures are evident, as
is also the case for RPN [22]. Within a similar range of results, the De Bonet [14]
method randomizes the initial texture image and preserves only a few statistics,
namely the dependencies across scales of a multi-resolution filter bench response.
Other methods are also based on statistics of wavelet coefficients or more involved multiscale image representations [69, 67, 72]. The Heeger-Bergen method
was extended by Portilla and Simoncelli [69] who proposed to evaluate on the
sample some 700 cross-correlations, autocorrelations and statistical moments
of the wavelet coefficients. Enforcing the same statistics on synthetic images,
starting from white noise, achieves striking results for a wide range of texture
examples. This method, which for a decade represented the state-of-the-art
for psychophysically and statistically founded algorithms is nevertheless computationally heavy, and its convergence is not guaranteed. Its results, though
generally indiscernible from the original samples in a pre-attentive examination, often present blur and phantoms. Earlier, Zhu, Wu and Mumford [98]
proposed to model texture images by inferring a probability distribution on a
set of images with the same texture appearances and then to sample from it. To
infer this probability distribution, the set of images is filtered by a pre-selected
set of filters (which capture the important features of a given texture image)
and their histograms are extracted. These are estimates of the marginals of
the probability distribution sought for. Then the maximum entropy probability
distribution is constructed matching the previous marginals. To sample from
this probability distribution the Gibbs sampler is adopted, thus generating new
texture images. The resulting model is a Markov random field. The limitation
of this method is its practical aspect. Inferring the probability distribution and
sampling from it are complex tasks. More recent work by Zhu et al. [96, 92]
advanced the Julesz ensembles texture model based on a common set of statistics; they proved that this model is equivalent to FRAME in the limit of an
infinite image grid. An efficient MCMC sampling method was also proposed.
These two texture generators have been recently revisited with neural networks.
Gatys’ texture generator [27] and DeepFrame [58] can be seen respectively as
extended versions of [69] and [98], and get significantly better results. Some new
neural network methods, based on generative neural networks, also get notable
results [36]. All these recent methods show that the Julesz program of seeking
the right statistics to characterize a texture is still well alive.
It is worth mentioning that texture models can be used to complete missing
parts of an image or texture inpainting. These methods rely on the definition
of texture images as the realization of a random field. For inpainting this boils
down to the estimation of a random texture model on the masked input image
(a set of valid pixels of the image) from which a new image is sampled conditioned to some of the known values of input image. The method presented in
[26, 25, 24] is particularly well adapted for micro-textures. A Gaussian model
4
is estimated on the masked input image; then the result is generated by a
conditional sampling from the estimated model using the kriging estimation
framework.
Patch re-arrangement methods constitute a totally different category of texture synthesis algorithms. This category started by pixel re-arrangement using
square patches as context. The initial Efros and Leung [19] method was inspired by Shannon’s Markov random field model for the English language [78].
In analogy with Shannon’s algorithm for synthesizing sentences, the texture is
constructed pixel by pixel. For each new pixel in the reconstructed image, a
patch centered in the pixel is compared to all the patches of the input sample.
The patches in the sample that are similar help predict the pixel value in the
synthetic image. Several optimizations have been proposed to accelerate this
algorithm. Among them Wei and Levoy [90] managed to fix the shape and size
of the learning patch, and Ashikhmin [3] proposed to extend existing patches
whenever possible instead of searching in the entire sample texture. Yet, as already pointed out in the original paper [19], an iterative procedure may fail by
producing “garbage” when the neighborhood’s size is too small. On the other
hand, it can lead to a trivial verbatim reproduction of big pieces of the sample
when the neighborhood is too large. This can be experimented in the online
executable paper [1]. Many extensions of [19] have been proposed that manage
to accelerate the procedure and reduce the “garbage” problem by stitching entire patches instead of pixels. Among the first methods proposing to re-arrange
whole patches, Xu et al. [32] proposed to synthesize a texture by picking random patches from the sample texture and placing them randomly in the output
texture image. A blending step is applied across the overlapping blocks to avoid
edge artifacts. In [54] the authors proposed to synthesize the image by quilting together patches that were taken from the input image among those who
best match the patch under construction. A blending step was also added to
overcome some edge artifacts. Efros and Freeman [20] proposed an extension
of the latter introducing the quilting method (a blending step) that computes
a path with minimal contrast across overlapping patches, thus mitigating the
transition effect from patch to patch.
Kwatra et al. [47] extended [20] by using a graph-cut algorithm to define the
edges of the patch to quilt in the synthesis image. Another extension of [19]
was proposed by Kwatra et al. [46] where to synthesize a texture image they
improve the quality of the synthesis image sequentially by minimizing a patchbased energy function. In the same spirit as [46], where texture optimization
is performed, the authors in [50] proposed to synthesize textures in a multiscale framework using the coordinate maps of the sample texture at different
scales. They introduced spatial randomness by applying a jitter function to the
coordinates at each level, combined to a correction step inspired by [3]. One
of the key strengths of the method is that it is a parallel synthesis algorithm
which makes it extremely fast. These patch-based approaches often present satisfactory visual results. In particular they have the ability to reproduce highly
structured textures (macrotextures). However, the risk remains of copying even
several times verbatim large parts of the input sample. For practical applica5
tions this may result in the appearance of repeated patterns in the synthesized
image. Furthermore, a fidelity to the global statistics of the initial sample is not
guaranteed, in particular when the texture sample is not stationary. We refer
to [89] for an extensive overview of the different patch re-arrangement methods.
Recent research tries to revisit the use of previous methods. Using neural
networks has seen some success, as well as combining patch re-arrangement and
statistics-based methods to overcome the drawbacks mentioned previously [66,
81]. These approaches will be called hybrid methods. Peyré [66] proposed to use
a patch-based approach where all the patches of the synthesized image are created from a sparse dictionary learnt on the input sample. Tartavel et al. [81] extended [66] by minimizing an energy that involves a sparse dictionary of patches
combined to constraints on the Fourier spectrum of the input sample in a multiscale framework. Raad et al. [70] proposed to model the self-similarities of
a given input texture with conditional multivariate Gaussian distributions in
the patch space in a multiscale framework. A new image is generated patch
by patch, where for each given patch a multivariate Gaussian model is inferred
from its nearest neighbours in the patch space of the input sample, and hereafter
sampled from this model.
The academic literature shows that current methods are able to produce
impressive texture synthesis on various kinds of textures. Our experiments will
illustrate this, and the opposite. Indeed, this literature is still working, in a
sense, on toy examples. Most textures are defined by texture samples of relatively small size and the structures are present in a small range of scales. When
confronting the methods with more challenging data, the quality of the results
degrades rapidly. This can be seen for most natural images of textures, which
are non-stationary, due for example to the presence of illumination changes and
perspective. As a matter of fact large photographs of textures are non-stationary
because even homogeneous material always shows an internal variation of structure. Thus the classic exemplar-based texture synthesis problem can be seen
in this light as an almost impossible Fourier spectrum extrapolation, given a
very small texture example. Hence our exploration not only of the solutions,
but of the problem itself will illustrate the limitations of the current question,
and introduce a more general question: how to emulate the real, non-stationary
textures, for which we dispose of large samples? Then the question is no longer
to “extend” a small patch into a larger texture of the same kind, but rather to
be able to fabricate other examples of a given large and complex texture, given
only one sample of it.
This survey concentrates on the problem of texture synthesis on flat 2D domains. There are several interesting extensions and applications of the basic
problem which are not discussed here. These include surface texture synthesis,
in which a texture is to be placed onto a curved surface, dynamic texture synthesis, when the goal is to generate textures whose appearance evolves over time
such as for videos of time-variant materials, or solid texture synthesis, where the
aim is to generate the color content of 3D blocks of synthesized materials from
which, for example, computer graphics objects can be carved out. Other related
problems include image completion and resolution enhancement by texture syn6
thesis. Also, the computational cost in real-time applications (e.g. games) or
when the data volume is large (e.g. film production) impose further restrictions
leading to particular algorithms. For a discussion of these topics, we refer the
reader to Wei et al. [89] and the references therein.
We now sketch our plan. We shall present the main trends in exemplarbased texture synthesis by describing in detail several methods illustrating the
three main families. In each case, the strength and limitations will be commented as well as some relevant variations. Section 2 introduces the statisticsbased methods which perform statistical optimization and describes several algorithms. Then Section 3 focuses on patch re-arrangement methods, presenting
three works. The third main class of hybrid methods is discussed in Section 4.
The experimental Section 5 first compares the main families of algorithms in a
varied set of textures; then, the limitations of all current methods are revealed
with high-resolution and non-stationary examples. Finally, Section 6 concludes
the paper. All the results displayed were generated for this paper, with the
original code published with the methods [9, 22, 68, 28, 57, 37, 1, 71, 52] or
with the modifications mentioned in this paper.
2
Statistics-based methods
Statistics-based texture synthesis methods follow the general approach proposed
by Julesz, illustrated in Figure 1. The synthesis is performed in two steps: first,
a set of statistics is estimated from the sample texture; second, a random image
is generated, subject to these statistical constraints. Methods in this class differ
in the set of statistics considered and in the optimization method used to impose
them on a random image. We will describe several algorithms of this class with
increasing sophistication. It will appear that the number of statistics enforced
plays a key role in the success.
2.1
Micro-texture synthesis by phase randomization
The Random Phase Noise (RPN) method synthesizes a new texture from a rectangular sample by simply randomizing the phase of the Fourier coefficients of
the input sample. The results are very satisfying for textures that are characterized by their Fourier modulus, a class called micro-texture by some authors.
This method is also able to create a random texture from any input image, not
necessarily a texture sample. It is in spirit quite close to the noise generators
from computer graphics [65, 87]. The rest of this section describes the main
ideas of this method and we refer the reader to [22] for more details and a
catalog of several synthesis examples.
The RPN of an image u defined on a domain Ω is obtained by adding a random phase θ to the Fourier phase of the input sample image. The random phase
is a white noise image uniformly distributed over (−π, π] and is constrained to
be symmetric. In the case of an RGB color image u = (uR , uG , uB ), the RPN
image is obtained by adding the same random phase to the Fourier transform
7
Input
Noise
statistics
optimization
Output
Figure 1: Statistics-based methods. A set of statistics is extracted from an
input sample (analysis step). Then, starting with a noise image, an optimization
procedure is applied to enforce these statistics on the output image (synthesis
step).
of each color channel. Adding the same random phase to the original phases of
each color channel preserves the phase displacements between channels. This is
important as it permits to create new textures without creating false colors [23].
More precisely, a uniform random phase is defined as a random image θ ∈
RM ×N satisfying the following conditions:
• θ is odd: ∀m, n ∈ Ω, θ(−m, −n) = −θ(m, n);
• θ(m, n) is uniform on the interval (−π, π] for (m, n) 6∈ {(0, 0), (M/2, 0), (0, N/2), (M/2, N/2)};
• θ(m, n) is uniform on the set {0, π} for (m, n) ∈ {(0, 0), (M/2, 0), (0, N/2), (M/2, N/2)};
• for every subset S of the Fourier domain which does not contain distinct
symmetric points, the family of random variables {θ(m, n)|(m, n) ∈ S} is
independent.
The RPN of an image u ∈ RM ×N is defined as the random image X where there
exists a uniform random phase θ such that
X̂(ξ, η) = û(ξ, η)eiθ(ξ,η) , (ξ, η) ∈ Ω,
(1)
where û denotes the Fourier transform of u. An equivalent definition is
X̂(ξ, η) = |û(ξ, η)|eiθ(ξ,η) ,
8
(2)
output
output
input
input
Figure 2: Synthesis results of the RPN method [87, 23]. This method works
extremely well for micro-textures including tissues and granular textures with
no geometric structures [22]. For more structured texture images it fails. Two
examples are shown: a successful synthesis on the left and a failure case on the
right.
where θ is a uniform random phase. Given the phase φ of a real-valued image and
a uniform random phase θ, the random image (θ + φ) mod 2π is also a uniform
random phase, which proves this equivalence. The first definition (1) leads to a
natural extension of RPN to color images [23], while the second definition (2)
highlights the fact that the RPN depends only on the Fourier modulus of the
sample image u.
Similarly, an Asymptotic Discrete Spot Noise (ADSN) associated with an
image u is defined as the convolution of a normalized zero-mean copy of u
with a Gaussian white noise. A Gaussian white noise image has a uniform
random phase and its Fourier modulus is a white Rayleigh noise; the phase
and modulus are independent. Thus, the phase of the ADSN is a uniform
random phase whereas its Fourier modulus is the pointwise multiplication of
the Fourier modulus of u by a Rayleigh noise [23]. Both ADSN and RPN
have uniform random phases, but the modulus distributions are different. RPN
keeps the Fourier modulus of the original image, while for ADSN the Fourier
modulus is degraded by a pointwise multiplication by a white Rayleigh noise.
Regardless of their theoretical differences, ADSN and RPN produce results that
are perceptually very similar [23].
The RPN method is the fastest method presented in this review since it
basically needs the computation of two FFTs. Nevertheless, this method is limited to micro-textures and it will fail synthesizing structured textures, namely
macro-textures. In Figure 2 two synthesis examples are shown. The first synthesis (left example in Figure 2) shows outstanding results. This micro-texture
is indeed well represented by its Fourier modulus. However this is not at all
the case for the second texture synthesis (right example in Figure 2). Clearly,
the knowledge of the modulus of the Fourier coefficients of this texture is not
sufficient to recover the strong contrast of the input.
9
2.2
The Heeger and Bergen pyramid based texture synthesis
Heeger and Bergen [33] proposed to characterize a texture by the first order
statistics of both its color and its responses to multiscale and multi-orientation
filters organized in a steerable pyramid [21]. This proposition, motivated by the
study of human texture perception, focuses on the synthesis of microtextures
defined as images that don’t have conspicuous patterns (e.g., granite, bark,
sand).
Let us describe the input texture image u and the synthesized texture v
using the Heeger and Bergen method. First the image u is filtered using a
steerable pyramid decomposition [21, 79] with S scales and Q orientations at
each scale. The steerable pyramid is a linear multiscale and multi-orientation
image decomposition. Given an input image, it is first filtered to provide a high
frequency image and a low frequency image. Band-pass oriented filters are then
sequentially applied to the low frequency image which is also down-sampled.
These band-pass oriented filters are applied S times to the corresponding low
frequency image. This decomposition yields images of different sizes corresponding to the different scales and orientations on which the gray level histograms
are extracted as well as the gray level histogram of u. These histograms define
the set of statistics that characterize u.
The second step consists in generating the output image v, which is initialized with a noise image. Its pixel values are iteratively modified to match the
histograms of u and of its steerable decomposition. These histogram matchings are performed on v alternately in the image domain and in the multiscale
transform domain, until all the output histograms match the ones of u. A third
parameter is introduced here and it is the number of iterations used to achieve
a stabilization of the histogram matching.
To the best of our knowledge, no theorem guarantees that this iteration will
end with an image respecting all statistics; there is of course one solution to it,
namely the example image. But the goal is to create an image different from
the example. Hence the random initialization, which is supposed to lead always
to different samples of the same texture. This remark applies to all texture
synthesis methods we will consider: their success will mainly be judged visually
and experimentally.
To treat RGB color images, instead of applying the method to each color
channel of the input image which are highly correlated, the authors proposed
to change the color space RGB to a more adapted color space. This new color
space is obtained by principal component analysis of the RGB values of the
input image u. In [9] a detailed explanation of the original method of Heeger
and Bergen [33] is provided with a complete analysis of the steerable pyramid
decomposition and the histogram matching step. The authors also provide
in [9] a minor improvement in the edge handling of the convolutions as well
as an experimental section illustrating the influence of the parameters, namely
the number of iterations, the number of scales and the number of orientations.
As we said, there is no theoretical proof of convergence of the method but
10
output
output
input
input
Figure 3: Synthesis results of the Heeger and Bergen method [33]. This method
works for microtextures. For more structured texture images it fails. Two
examples are shown: a successful synthesis on the left and a failure case on the
right.
an experimental study shows that the results tend to stabilize after five to
ten iterations [9]. Increasing the number of orientations changes the results
slightly, but four orientations are enough in general. The number of scales is
very important. Taking the highest number permits to take into account all the
scales of the texture. When the input texture has no evident structure then this
parameter has less influence in the result.
As our experiments here will show, the results yielded by this approach are
convincing for some stochastic textures, but the method fails for most complex texture images. In particular it generally fails (visually) for quasi-periodic
textures, random mosaic textures, textures having more than one dominant orientation, and textures having correlations of high frequency content over large
distances. This demonstrates experimentally that all the spatial information
characterizing a texture is not captured by the first order statistics of a set of
linear filter outputs. In Figure 3 two synthesis examples are shown: a successful
synthesis and a failure case.
2.3
FRAME: a mathematical model for textures
FRAME, which stands for Filters, Random fields And Maximum Entropy, is
a mathematical model of textures developed by Zhu, Wu and Mumford in [97]
and [98]. It is the most mathematical solid work among the stream of work on
texture modeling during that period. It puts the graphics method of Heeger
and Bergen [33] in a mathematical sound setting, i.e., it has a formal statistical
model, and it can match the marginal statistics. It also answers the Julesz quest
by pursuing the minimum statistical constraint that are necessary.
The FRAME model is based on the maximum entropy principle. It starts
with a set of filters that are selected from a general filter bank to capture features
of the texture. These filters are applied to observed texture images, and the
histograms of the filtered images are extracted. Then, the maximum entropy
principle is employed to derive a distribution f , which has in expectation the
same filter responses as the original image, while being of maximum entropy.
11
More precisely, let u be an observed texture image and let F k , k = 1, . . . , K
be a set of filters. Let Huk be the (discrete) histograms of the filter responses
F k ∗ u, and for any image v, let Hvk be the histograms of the filter responses
F k ∗ v. Zhu, Wu and Mumford seek for a distribution f (v) on images v such
that
Ef (Hvk ) = Huk ,
(3)
while being of maximum entropy (i.e. while being “as random as possible”).
By Lagrange Multipliers (maximization under constraints), the solution has the
form
!
L X
K
X
1
k k
exp −
f (v; λ) =
λi Hu (i) ,
Z(λ)
i=1
k=1
where L is the number of bins of the discrete histograms. To find the value of
the parameters λ satisfying Equation (3), Zhu, Wu and Mumford use a gradient
descent to find the right λ and the Gibbs sampler algorithm to sample random
images v from a distribution f (·; λ). The distribution f (·; λ) defines a Markov
Random Field (MRF).
Finally, a stepwise algorithm is proposed to choose the filters from a general filter bank. This “filter pursuit” step is achieved thanks to the minimax
entropy principle: find the set of filters such that the associated distribution f
is of minimum entropy, since it is equivalent to be of minimal Kullback-Leibler
divergence from the “true” underlying distribution. A detailed explanation of
this fact can be found in [98].
The FRAME model was later extended, in particular with non linear filters, using the output of some layers of a neural network. It is then called
DeepFrame [58], and we will talk again about it in Section 2.5.
2.4
The Portilla and Simoncelli algorithm
The key issue in FRAME and in the method of Heeger and Bergen is to choose
the “right” filters and the statistics that will be matched. In [69], Portilla
and Simoncelli proposed an important improvement on Heeger and Bergen’s
method [33]. The texture is again synthesized starting from a noise image and
coercing it to have the same statistics as the input image. As we have seen,
marginal statistics are not enough to capture the relations across scales and
orientations. Portilla and Simoncelli proposed to match a set of joint statistics
measurements of the coefficients of the steerable pyramid decomposition of the
input texture. The statistics used to characterize the input texture are the
autocorrelation and cross-correlation coefficients (inner and intra scales), as well
as the statistical moments of order one, two, three and four of the input sample’s
values. To enforce these statistics on the result, the image under construction is
projected iteratively into the subspace of constraints using a gradient projection
approach until stabilization. The final output image may not have exactly the
same statistics as the input sample. It merely represents a local minimum.
Again there is no proof of a convergence of the method anyway.
12
Portilla and Simoncelli’s technique is based on the theories of human visual
perception, in particular Julesz’ hypothesis stating that two images are perceptually equivalent if and only if they agree on a set of statistic measurements.
The goal is to establish the minimal set of measurements in a way that all types
of textures are correctly synthesized using that set of measurements. In the
same way as Heeger and Bergen’s method, the input texture sample is decomposed with a multiscale oriented linear basis: the steerable pyramid [21, 79].
For each pair of coefficients at nearby positions, orientations and scales, the
average value of their product, of their magnitude product and their relative
phase is measured. In addition to these parameters, some marginal statistics on
the input image pixels distribution are kept: the mean, the variance, the skewness, the kurtosis and the range. The number of parameters will depend on the
number of sub-band images and on the size of the neighbourhood considered to
estimate the statistical constraints of the example texture.
The second part of the algorithm is the synthesis step coercing to a random
noise image, the measurements previously computed. The synthesized image is
initialized with a Gaussian white noise image and then iteratively the algorithm
alternates between: 1) constructing the steerable pyramid and enforcing the
sample statistics of each sub-band image matching those of the corresponding
sub-bands of the target image; 2) reconstructing an image from the pyramid
and then forcing it to have the same marginal statistics as the input texture.
A texture is defined as a two-dimensional stationary random field X(m, n)
on a finite lattice (m, n) ∈ Ω ⊂ Z2 . Julesz’ hypothesis is the basis to connect
this statistical definition to perception: there exists a set of constraint functions
{φk , k = 1, . . . , K} such as if two random fields, X and Y , are identical in
expectation over this set of functions then any two samples drawn from X
and Y will be perceptually equivalent under some fixed comparison conditions.
The importance of human perception as a fundamental criterion of equivalence
between textures can be seen through this hypothesis, as well as the existence
of such a set of statistical measurements capable of capturing this equivalence.
To choose the set of constraint functions Portilla and Simoncelli proceeded as
follows:
1. Set an initial set of constraints and synthesize a large library of texture
examples;
2. Group the synthesis failures classifying them according to visual features
that distinguish them from their original texture examples and keep the
group with the poorest results;
3. Add a new statistical constraint to the set capturing the missing feature
of the failure group;
4. Re-synthesize the failure group and verify the wanted feature is captured;
otherwise go back to the previous point;
5. Verify that the original constraints are still needed; for each constraint,
13
find a texture example that fails when the constraint is removed from the
set;
6. Delete the unnecessary constraint, re-synthesize the library and go back
to the second point.
Following this strategy, the constraint set is adapted to a reference set of
textures and not just to one texture, and it is driven by perceptual criteria.
The set of constraints is composed of:
1. Marginal statistics formed by: skewness and kurtosis of the low-pass images of each level of the pyramid, variance of the high-pass image of the
pyramid, skewness, kurtosis, variance, mean and range of the image. The
marginal statistics set the general degree of pixel intensity and their distribution. This is why they cannot be discarded from the statistics set [69].
2. Autocorrelation of the low-band coefficients. This allows to capture the
periodic structures of a texture as well as long-range correlation. Omitting
this constraint from the original set yields unsatisfying results for textures
having periodic or long-range correlation patterns [69].
3. Autocorrelation and cross-correlation of the magnitude of the sub-bands.
These statistics appear to be relevant because observation reveals that
oriented bands have a particular behaviour concerning certain pattern
and their periodicity whatever the orientation [69]. The cross-correlations
kept are of each sub-band image with others of the same scale (inner
cross-correlation) and of each sub-band with sub-bands at the coarser
scale (intra cross-correlation).
4. Cross-correlation of the real part of the sub-bands with the real and imaginary parts of the coefficients’ phase of the coarser scale. This statistic
is important to capture the strong illumination effects present in some
texture images. In particular, the synthesized image looses its threedimensional effect and the shadows structure if they are not considered [69].
The set of statistics is summarized in Table 1. As mentioned previously, the
number of parameters used depends on the number of scales S and orientations
Q of the steerable decomposition as well as the size of the neighbourhood Na
used to compute the auto-correlations. The total number of parameters is 6+1+
2(S+1)+(S+1)(Na2 +1)/2+SQ(Na2 +1)/2+SQ(Q−1)/2+(S−1)Q2 +2(S−1)Q2 ,
where the terms correspond (from left to right) to: the marginal statistics of
u, the variance of high pass image, the skewness and kurtosis of the low band
images, the auto-correlation of the low band images, the auto-correlation of the
sub band images, the inner cross-correlation of the sub band images, the intra
cross-correlation of the low band images and the cross correlation of the real
part of the sub band images with the real and imaginary part of the phase sub
band images. In general S = 4, Q = 4 and Na = 7 are used, leading to a total
of 710 parameters.
14
range of u
max(u) and min(u)
mean of u
µ1 (u)
variance of u
µ2 (u)
skewness of u
µ3 (u)/(µ2 (u))1.5
kurtosis of u
µ4 (u)/(µ2 (u))2
lowband’s skewness
µ3 (ls )/(µ2 (ls ))1.5 , 1 ≤ s ≤ S + 1
lowband’s kurtosis
µ4 (ls )/(µ2 (ls ))2 , 1 ≤ s ≤ S + 1
highband’s variance
µ2 (h)
<{ls } auto-correlation Γ<{ls } (m, n) , 1 ≤ s ≤ S + 1
|us,q | auto-correlation
inner cross-correlation
intra cross-correlation
cross-correlation with
the real part of the
phase
cross-correlation with
the imaginary part of
the phase
Γ|us,q | (m, n) , 1 ≤ s ≤ S, 0 ≤ q ≤ Q − 1
0
C |us,q | , us,q , 1 ≤ s ≤ S, 0 ≤ q, q 0 ≤ Q − 1
0
C |us,q | , us+1,q , 1 ≤ s ≤ S − 1, 0 ≤ q, q 0 ≤ Q − 1
n
o
0
< us+1,q
s,q
C < {u } , us+1,q0
, 1 ≤ s ≤ S − 1, 0 ≤ q, q 0 ≤
|
|
Q−1
n
o
0
= us+1,q
s,q
C < {u } , us+1,q0
, 1 ≤ s ≤ S − 1, 0 ≤ q, q 0 ≤
|
|
Q−1
(
Central sample mo- µn (u) =
ment
1
MN
1
MN
PM −1 PN −1
j=0 u(i, j)
Pi=0
M −1 PN −1
n
i=0
j=0 (u(i, j) − µ1 (u))
if n = 1
if n > 1
Correlation
τx,y (u) : u(m, n) 7→ u(bm − xcM , bn − ycN )
0 ≤ m ≤ M − 1, 0 ≤ n ≤ N − 1, (x, y) ∈ Ω
PM −1 PN −1
∗
C(u, v) = M1N i=0
j=0 (u(i, j) − m(u)) (v(i, j) − m(v))
Auto-correlation
Γu (x, y) = C (u, τx,y (u))
Translation operator
Table 1: Summary of the set of statistical constraints for the Portilla-Simoncelli
method.
15
After setting the set of statistical constraints, a sample verifying them has
to be generated. Let ck be the corresponding estimated values of the constraint
functions for a particular texture image. Portilla and Simoncelli [69] “samples”
an image from the set of images that yield the same estimated constraints values Aφ,~
~ c = {u : φk (u) = ck , ∀k}. To pick at random from this set the authors
proposed to select at random a sample u0 from R|Ω| and then project it sequentially onto subsets of Aφ,~
~ c . To emulate this the authors proposed a gradient
projection. That is moving in the direction of the gradient of the constraint
φk (v):
→
−
v 0 = v + λk ∇φk (v)
choosing λk such that
φk (v 0 ) = ck .
(4)
→
−
Computing ∇φk (v) is usually simple, and it remains to find the λk that solves (4).
When there are multiple solutions for λk , the one with smaller amplitude is chosen, modifying as little as possible the image. In that way, we stay as close as
possible to the already projected set. When there is no solution, the λk is the
one that comes closest to satisfying (4). Finally this method can be extended to
the adjustment of a subset of constraints. Once the set of statistical measurements is defined and a method to sample from the Julesz’ ensemble of textures,
the synthesis can be performed as explained previously.
In a pre-attentive examination, the results are in general indistinct from the
original texture samples. Nevertheless, on attentive examination the synthesis
of structured textures often present blurry and jammed results. Long range
structures are missed and the method tends to homogenize the output texture.
Figure 4 shows two synthesis results. The first example (left in Figure 4) represents a quasi-periodic image where the method yields excellent results although
it contains some global structures. In the second example (right in Figure 4),
even though we recognize the nature of the input sample, one can observe that
strong structures are missing. It is impossible to recover the lined up tiles.
Increasing the number of orientations Q will improve the results since more
information is captured. However for Q > 4 the improvement is hardly noticeable. The number of levels S of the steerable pyramid is the most influential
parameter. Depending on the nature of the texture, it will need to be increased
to capture the details at all scales. Once again, for microtextures this parameter
is less influential. Finally, the size of the neighborhood Na used to compute the
autocorrelation is important whenever the texture has periodic information.
As we will see in Section 5, even though imperfect, the results are very impressive, as they succeed modeling most textures using a moderately large set
of global statistics. This brings us to the following two questions. Is the set of
statistics considered enough to describe any kind of textures? Is the optimization step enough to enforce these statistics? Fifteen years later, Gatys et al. [27]
proposed a texture synthesis method based on Convolutional Neural Networks
(CNN) which can be seen as an extension of Portilla and Simoncelli’s work,
16
output
output
input
input
Figure 4: Synthesis results of the Portilla and Simoncelli method [69]. It is satisfactory for many small grain textures (left) but may miss the global structure
(right).
where the set of statistics used is much larger and unknown; also, the optimization is performed by the backpropagation method.
2.5
Texture synthesis using CNN
It is hard to define metrics to determine if two textures are similar or not according to human taste. Julesz’ conjecture that humans cannot distinguish two
textures with same second order statistics was invalidated. Yet this does not
rule out a more general hypothesis, according to which there is a set of low-level
filters such that if two textures respect the same statistics for these filters, they
are indistinguishable. Portilla and Simoncelli’s approach [69] and Zhu, Wu, and
Mumford’s FRAME (Filters, Random field, And Maximum Entropy) [97, 98]
can be seen as fixing a set of hand-picked filters and synthesing new textures by
enforcing the response to the filters to have similar statistics. The set of filters is
chosen to match human expectations about textures. However determining the
exact set of filters equivalent to human vision is very hard, and both approaches
use only a subset of them. Portilla and Simoncelli achieve similar statistics by
iterating specific projections, starting from white noise, while FRAME achieves
that with a Gibbs Sampler and some simplifications (quantizing the image intensities, etc). Recently, Convolutional Neural Networks (CNNs) have given a
breath of fresh air to these approaches. CNNs are compositions of layers of convolutions, non-linearities and pooling. In the past few years, CNNs have been
successfully applied in a wide variety of domains, in particular in image related
tasks. Arguably, the win by a large margin of CNNs [45] in the 2012 ILSVRC
challenge [75], an image classification challenge, helped spark interest of the
global community to these methods. We refer the reader to the corresponding
literature for more details on the working of CNNs.
By taking a fully trained CNN on some visual classification task, and restricting to lower layers, one gets a set of low level filters which can directly
be used for synthesizing texture, as shown in several works. The topic is quite
active recently, and the question “how to best synthesize a texture with the
help of neural networks” is far from being solved. In the following, we will focus
17
on two different approaches: Gatys’ texture generator [27] and DeepFrame [58].
Gatys’ approach is to minimize the distance between the Gram matrices defined
by the local filter responses of the network layers, while DeepFrame generates
textures by sampling from an exponential model. The use of CNNs by these new
approaches solves the issues of their ancestors: first the filters do not need to
be handpicked anymore, they are encoded directly by the CNN. A pre-trained
CNN successful on some image-related tasks can be selected for the texture generation. The choice of the CNN and whether it is pre-trained or the weights are
random, affect the result. Second, the architecture of Neural Networks eases
the generation process. The statistics of all the filters can be handled at the
same time, via backpropagation for example. DeepFrame needs no quantization,
unlike its predecessor, and synthesizes textures at a faster speed. Because the
filter responses at a given Neural Network layer also encode the image content,
texture transfer – also named style transfer – can be achieved by applying the
statistics of the filter responses of a source image to a target image while keeping
overall the filter responses similar [29]. While initially both Gatys’ texture generator and DeepFrame used the VGG network [80] trained on ImageNet [15],
more recent work obtained good results with networks with random weights
[85] or by integrating the network training with the generation process [93].
The success of VGG for texture generation seems to stem from its training on
an object classification task. This implies that its trained features are valuable
“textons” able to discriminate shape and object features. One could imagine
using a network trained to distinguish textures directly, instead of VGG. But,
to the best of our knowledge, no network has been trained on an ImageNet
equivalent to textures.
We now take a closer look at Gatys’ texture generator and at DeepFrame.
Gatys’ texture model is a generalization of Julesz’ model. It postulates that textures are described by the correlations between the neural network activations
(features). Thus, by starting from random noise and imposing the correlations
between the features to be the same as for a given input texture, one should get
a new sample of this texture.
More precisely, Gatys’s texture generator seeks to minimize the cost
X
E=
wl ||Gl − T l ||2F
l
where ||.||F is the Frobenius norm, wl are weights and Gl , T l are the Gram
matrices, respectively for the image and the target texture, of the feature maps
of a pretrained neural network at a layer l. In [27], a custom 19-layer VGG
network was used where max pooling was replaced by average pooling and the
network weights were rescaled. Let Nl be the number of feature maps at layer
l (this usually corresponds to the number of “channels”), and Ml the size of
each feature map at layer l (Ml × Nl is the number of outputs of layer l). If
we denote by Fijl , i ∈ {1 · · · , Nl }, j ∈ {1 · · · , Ml }, the j-th output with the i-th
18
feature map at layer l, then
Gl
ij
=
M
1 Xl l l
Fik Fjk .
Ml
k=0
The texture generator minimizes the cost via backpropagation in the network,
and thus falls into a local minimum. Starting from white noise, several thousand
iterations can be needed to reach visual convergence. While in [27] the features
were extracted from VGG [80], a Deep Convolutional Neural Network trained
on image classification tasks, in [85] it is noted that taking a pre-trained network
is not necessary and a network with random weights can give satisfying results.
The minimization of E is done with L-BFGS-B [94] and the bounds are set to
the minima and maxima of the source texture. After convergence, the histogram
of the source is enforced.
To generate the results in this article, we made a few changes compared
to [27]. The 19-layer VGG network used in [27] pads the outputs at every
convolution layer with zeros on each layer (to have the layer outputs be the
same size as the layer inputs). That, plus the fact that pixels on the border are
“seen” by fewer features than the pixels in the center, means that all pixels on
the image are not imposed the same distribution. If we take the same layers
than in [27] (conv1_1, pool1, pool2, pool3, pool4) the top layer’s outputs
(pool4) depend each on a 124 × 124 area of the source. Thus 123 pixels should
be removed on each border in order to have all remaining pixels seen by the same
number of features. Removing 123 pixels on each border is not sufficient however
to get the same constraints on the border and the center since the neighbouring
pixels affect the features, and those neighbouring pixels are not affected by
the same features. Thus to generate the results in this article, we decided to
both remove the padding and generate bigger images – 256 pixels more on each
border – which we then crop. The impact of this change can be seen on Figure 5.
Other than that, we took the same parameters. In [2] the method solves the
same problem by removing the network padding and enforcing periodicity. With
the default network and parameters of Gatys’ texture generator, except for the
boundaries, a pixel is seen by 37504 filters. In Gatys’ method, textures are
only described by the Gram matrices. The number of elements in the Gram
matrices totals 352256, 176640 if we remove the redundant values (the matrices
are symmetric). This number of parameters doesn’t depend on the image size,
and once the Gram matrices of the source computed, the output texture can be
any size.
To fix some of the shortcomings of Gatys’ texture generator [27], several
works complete the objective function. The method in [55] incorporates spectrum constraints to significantly improve the generation of textures with low
frequency patterns. In [6] the proposed method considers spatial co-occurences
of features to help handling long-range consistency constraints. In [91] it is
noticed that the Gram matrices have several particularities that decrease the
quality of the texture obtained in several cases with instabilities, particularly
visible when generating a texture with a size different from the source. In our
19
input
no padding + crop
no padding
padding
Figure 5: This figure shows the impact of the padding in the neural network.
The second image shows the result of a 1024 × 1024 generated texture without
the network padding, cropped to 512 × 512, while the figures on the right show
512 × 512 sized generated results without or with padding. The same random
initialization was used for all three results (and cropped for the last two results).
The differences are particularly visible on the border of the pictures, since it is
where each variant imposes different statistics.
experiments we didn’t notice such an instability problem, although we observed
some instabilities (see for example the fourth column of figure 20 and the first
column of figure 22). It is possible that the instabilities are affected by the
parameter choice. To solve the instability problem, the authors added to the
objective function a term to force the feature maps histograms to be the same as
for the source. The authors of [63] also discussed some insufficiencies of Gram
matrices in the case of style transfer, and in particular proposed to shift the
activations to avoid sparsity. To accelerate the speed of the texture generation,
the method of [83] trains for a given texture a new CNN, which outputs new
samples of the texture. The CNN is trained with the same objective function
as for Gatys’ texture generator. Once the CNN is trained, generation is fast.
DeepFrame’s texture generator samples from an exponential model. The
model is defined by the probability density function
"K
#
XX
1
f (u; w) =
exp
wk Fk (u)(x) g(u),
Z(w)
k=1 x∈Ω
where Fk corresponds to a filter map extracted from a CNN, Ω is the image
domain of u the image, Z(w) is a normalizing constant and g(u) is a reference
distribution, like
1
1
2
g(u) =
exp
−
||u||
.
2σ 2
(2πσ 2 )|Ω|/2
In contrast, the FRAME model defined the probability density function
"K
#
XX
1
f (u; λ) =
exp
λk [Fk ∗ u(x)]
Z(λ)
k=1 x∈Ω
20
where the (Fk )k=1..K were kernels, such as Gabor filters, or Difference of Gaussian filters, and λk was a discretization function with finite number of possible
outputs.
In a first phase, the DeepFrame parameters w = (wk ) are tuned for the
source texture, then in a second phase new samples of the texture are generated
via Langevin dynamics. While in [58] a pre-trained network is used, in the
method of [93] its own network is trained on the source.
While both Gatys’ texture generator and DeepFrame have a fixed texture
model used to generate new samples, for which they learn parameters, a third
successful CNN method to synthesize texture learns directly its model: in [36]
a generative CNN is trained to synthesize new images from one or several samples of a source. The training is based on the adversarial model: a discriminator tries to distinguish the fake generated samples from true ones, while a
generator creates new samples. Spatial invariance assumptions are encoded in
the networks, but else, the texture model is in some sense learned by the two
networks. This method can still be considered as a statistics-based method,
because in some sense the discriminator checks the statistics of the texture are
correct. To generate samples with this method (“SGAN” for Spatial Generative
Adversarial Networks), we took the default network parameters, and applied
the source histogram. We stopped after a few hundred epochs. The outputs
suffer from a sort of noise pattern, which changes after every epoch. When the
noise pattern was too important, we decided to select among the last twenty
epochs the generator’s result with the less noise. SGAN is a recent method, and
there are certainly ways to better select the parameters and reduce this noise,
but this goes beyond our goals here. Recently a new extension called PSGAN
(for Periodic Spatial Generative Adversarial Networks) [7] was introduced to fix
some shortcomings of SGAN, in particular to improve the generation result for
textures with periodic patterns.
CNNs are also successful in the synthesis of images more general than textures [64, 86, 76, 18], in particular with methods relying on Generative Adversarial Networks (GAN) [30, 16, 73, 35, 88], but these methods are out of the
scope of this study, which focuses on synthesizing new texture samples based on
a single reference sample. These methods generally need a database of images.
3
Patch re-arrangement methods
In contrast to the statistics-based methods, the patch re-arrangement methods
do not attempt to characterize textures by a statistical model. Spanning from
the groundbreaking work by Efros and Leung [19], this family of algorithms
consists of clever heuristics to re-arrange parts of the sample texture in a random
way in order to create a new texture. By copying directly from the sample image,
these methods often are able to keep complex structures from the input. By the
same token, the process is frequently limited to copying and the results show
little innovation relative to the sample. We will illustrate the family here by
the original Efros and Leung [19] algorithm, a further extension by Efros and
21
input
Gatys
DeepFrame
SGAN
input
Gatys
DeepFrame
SGAN
Figure 6: Comparison between Gatys’ texture generator [27], DeepFrame [58]
and SGAN [36]. For all three methods, we used the default parameters, except
that in the case of Gatys we used the method we described above where we
remove the network padding and crop the result and in case of DeepFrame and
SGAN, we specified the result’s histogram on the source histogram. Overall,
SGAN looks the best when looking from far, but when zoomed in, Gatys seems
to respect the best the local structures.
Freeman [20] which incorporate more recent techniques, and a more recent CNN
based method [53].
3.1
The Efros and Leung algorithm
In his foundational paper of information theory [78], Claude E. Shannon proposed to approximate the information contents of natural languages by the
entropy of generative stochastic processes. He used a Markov chain to generate
English text sequentially, letter by letter. Given a piece of already generated
text, the next letter is sampled from the probability distribution of English text
conditioned to the previous n letters. The following sequence was generated by
Shannon using a third-order model:
in no ist lat whey cratict froure birs grocid pondenome of demonstures of the reptagin is regoactiona of cre
Although very few words are real English words, this simple model produces
surprisingly good English “textures”. Inspired by Shannon’s method, Efros and
Leung [19] proposed to adapt the same ideas for image texture synthesis.
Efros and Leung in [19] synthesize a new texture image by considering that
a pixel value depends on the values of its neighbouring pixels. The method
is illustrated in Figure 7 and works as follows. For a given input texture, a
22
output
input
(1)
(2)
Figure 7: Overview of the Efros and Leung algorithm [19]. Given a texture
image (left) a new image (right) is being synthesized a pixel at a time. For
a pixel (m, n) (red point in the output) being synthesized the method finds
all neighbourhoods in the left image that match the neighbourhood of (m, n)
(dashed squares) and then chooses randomly one of the neighbourhoods (yellow
square) and assigns its central pixel value to (m, n).
new image is synthesized sequentially, pixel by pixel. For a pixel (m, n) being
synthesized, the algorithm finds all the neighbourhoods in the input image that
are similar to the neighbourhood of (m, n) up to a patch distance tolerance.
Then one of these neighbourhoods is randomly chosen and its central pixel
value is affected to the pixel (m, n). The neighbourhood of (m, n) is a square
patch but only the known pixels (coming from the seed or already synthetized)
of this patch are considered when comparing to the neighbourhoods of the input.
Denoting p1 and p2 two patches of size P × P , the comparison is made using a
Gaussian-weighted distance defined as
2
X
1
p1 (i, j) − p2 (i, j) Gσ (i, j),
i,j Gσ (i, j) i,j
d p1 , p2 = P
(5)
where Gσ is a Gaussian kernel with standard deviation σ.
Levina and Bickel in [51] provided a theoretical justification of Efros and
Leung’s work. The Efros and Leung algorithm is based on resampling from
the random field directly, without constructing an explicit model for the distribution. The authors of [51] formalized this algorithm in the framework of
resampling from random fields and proved that it provides consistent estimates
of the joint distribution of pixels in a window of specified size.
In general the visual results are very impressive, especially for structured
textures. Nevertheless this algorithm suffers from two important drawbacks:
verbatim copies of the input and garbage growing (the algorithm starts reproducing iteratively one part of the example and neglects the rest). Figure 8 shows
two synthesis examples. The first synthesis result illustrates a failure case. In
particular one can observe the effect of garbage growing, which reproduces incoherently the right side of the wood sample texture. The second example shows
23
output
output
input
input
Figure 8: Synthesis results of the Efros and Leung method [19]. Left: the example shows the garbage growing effect. Right: the example shows the strength of
this method to synthesize macrotextures. The patch size used for both synthesis
is P = 40.
the strength of this method when it comes to synthesize textures with conspicuous patterns as in this case the brick patterns. To illustrate the verbatim-copy
regions, position and synthesis maps are used to visualize from which regions
of the input texture each synthetized pixel comes from. A synthesis and the
corresponding map are shown in Figure 9 (obtained with the online demo [1]).
Large continuous zones are identified in the synthesis maps which corresponds
to the verbatim copies produced by the method. This representation also shows
that the synthesized image is indeed a re-arrangement of pieces of the input
sample.
Increasing the patch size P results in increasing the verbatim copied regions.
However if the patch size is too small the local aspect of this method fails in
recovering the global configuration of the input texture in particular for macrotextures. A second parameter of the method is the tolerance parameter ε which
is used to select the most similar patches in the input image. Large tolerance
values increase the garbage growing effect.
The Efros and Leung method also suffers from its high computational complexity. Several optimizations have been proposed to accelerate this algorithm.
Among them Wei and Levoy [90] managed to fix the shape and size of the
learning patch and Ashikhmin [3] proposed to extend existing patches whenever possible instead of searching in the entire sample texture. The following
section describes a particularly important extension of the method.
3.2
The Efros and Freeman algorithm
Efros and Freeman’s method [20] is an extension of Efros and Leung’s. It is
based on the same principle where the pixel values are conditioned to their
neighbourhood values. Efros and Freeman proposed to generate a new image
sequentially, patch by patch (instead of pixel by pixel) in a raster scan order as
illustrated in Figure 10. At each step a patch that is only partially defined on
a region called overlap region is completed. This overlap region is of width wo .
This is the patch under construction. To do so a patch of the input image among
24
Figure 9: From left to right: texture sample, position map, synthesized image
and synthesis map. The synthesis map shows for each synthesized patch its
initial position in the texture sample. It allows then to identify exactly the
verbatim copy regions (they correspond to continuous color areas of the map).
This method reveals the verbatim copies of the input in the generated texture
and the repetitions (garbage).
those who match the patch under construction on its overlap region is randomly
selected (patch selection step). An optimal boundary cut between the chosen
patch (pin ) and the one under construction (pold ) is then computed across the
overlap region (stitching step). This optimal boundary cut is used to construct a
new patch (pnew ) by blending the (pin ) and (pold ) along the cut. There are three
possible overlap regions: vertical overlap for the first row, horizontal overlap for
the first column, and L-shaped overlap everywhere else (Figure 10).
In the patch selection step, to select a patch pin of an input image u one
computes the square distance between the overlap region of the patch pold and
the corresponding regions of all the patches of u. The minimal distance Dmin
is determined and pin is randomly picked among all patches whose distance to
pold is lower than (1 + ε)Dmin where ε is the tolerance parameter. The squared
distance image d contains at each position (m, n) the distance between pold
and the patch from u according to some binary weight t that equals one in the
overlap region and zero otherwise. More precisely, one has
X
d(m, n) =
t(i, j)(pold (i, j) − u(m + i, n + j))2 .
(6)
i,j
The patch pin of u having coordinates (m, n) is similar to the partially defined
patch pold on their overlap region. To get the final patch pnew one must combine
the patches pold and pin . Denoting t the binary weight for the overlap regions
as in (6), then, for any binary image r such that 0 ≤ r(i, j) ≤ t(i, j), (i, j) ∈
{1, . . . , P }2 , P can be defined as the combination
pnew = t pold + (1 − t) pin .
The main contribution of Efros and Freeman [20] is to look for a binary shape M
where the transition between pold and pnew along the boundary of the shape is
25
Vertical overlap
Horizontal overlap
L-shape overlap
Iteration 10
Iteration 115
Iteration 239
Figure 10: Three different iterations of the synthesis process are shown. At each
iteration a patch is being synthesized. This patch is represented by the pink
square in the three iterations shown. From left to right the three overlap cases
are represented: vertical, horizontal and L-shape.
minimal. For simplicity, and to be able to use linear programming, the authors
do not allow for any shape, but only for the ones whose boundaries are simple
forward paths from one end to the other of the overlap region. This results
in two pieces of image being sewn together along some general boundary path,
hence the algorithm’s name “quilting”.
This method yields very impressive visual results, in particular for highly
structured textures. In terms of speed the gain is truly significant with respect
to the methods which synthesize an image pixel by pixel. The patch size being
larger, the risk of garbage growing is reduced compared to the Efros-Leung algorithm. Nevertheless the risk of verbatim copies remains and is even amplified.
Moreover, the respect of the global statistics of the input is not guaranteed and
this is quite visible when the input texture is not stationary (for example if there
is a change of illumination across the image). Figure 11 shows two synthesis examples. The first one (left) shows an excellent synthesis result where the strong
structures of the input are perfectly recovered. The second one (right) puts in
evidence the verbatim copy of parts of the input and the garbage growing effect.
To illustrate this the synthesis map of the second example is shown in Figure 12.
The parameters P and ε play the same role as in Efros and Leung’s method.
A third parameter, the overlap size O is used. Increasing this value tends to
increase the verbatim copies of large regions. However if this value is too small
then garbage growing increases. The value O = 0.25P is generally satisfactory.
3.3
High level patch re-arrangement with Convolutional
Neural Networks
The CNNs texture synthesis methods presented in Section 2.5 typically generated new texture samples by enforcing similar statistics on the feature maps of a
pre-trained CNN. At the crossroad of patch re-arrangement methods and CNNs,
lies CNNMRF [53]. What distinguishes this method to those of Section 2.5 is
26
output
output
input
input
Figure 11: Synthesis results of the Efros and Freeman method [20]. It works
for microtextures but risks losing the example’s global statistics. It works for
macrotextures too, but risks verbatim copies. Two examples are shown: a
success (left) and a failure (right). The parameters used for are P = 80 and
O = P/4.
Figure 12: From left to right: texture sample, position map, synthesized image
and synthesis map. The synthesis map shows for each synthesized patch its
initial position in the texture sample. It puts in evidence the verbatim copy regions (they correspond to continuous color areas of the map) and the repetitions
(corresponding to repeated continuous patches of the same color).
that the texture samples are generated by enforcing similar patches of feature
maps on selected upper layers of a pre-trained CNN. The image is then obtained
with backpropagation and a smoothness constraint.
More precisely, starting from random noise, an image is generated by minimizing the energy:
XX
E=
||ψi (F l ) − ψN N (i) (Fsl )||2 + R
l
i
where l goes among the selected layers (relu3_1 and relu4_1 of the VGG
network [80]), i goes among all the positions in the layer, ψi (F l ) represents the
patch at the i-th position and ψN N (i) (Fsl ) is its best matching patch in the source
according to the normalized cross-correlation. The default patch size is 3 × 3
times the number of feature maps (often referred as the number of “channels” of
the layer). R is a regularizer term to impose smoothness of the resulting image.
In [53], the energy also contains a term to enforce the content of the source if
doing texture transfer. This term isn’t used for texture synthesis.
As noted by the author, a natural seamless patch blending is obtained by
27
input
no padding + crop
no padding
padding
Figure 13: This figure shows the impact of the padding in the neural network
when generating images with CNNMRF [53]. The second image shows the result
of a 1024 × 1024 generated texture without the network padding, cropped to
512 × 512, while the figures on the right show 512 × 512 sized generated results
without or with padding. The same random initialization was used for all three
results (and cropped for the last two results).
performing a patch re-arrangement on the levels of the CNN instead of doing it
directly on the image, like in the other methods of this section.
Similarly to what was done in Section 2.5, we removed the padding of the
VGG network to generate the results of this method. Indeed if the padding
is kept, the spatial invariance assumption is violated. Moreover pixels on the
border of the generated images are seen by fewer features, which reinforces the
violation of the spatial invariance. Thus in addition to removing the network
padding, we generated bigger images and then cropped the result. On figure 13,
the generated texture with the network padding and no border crop kept tends
to reproduce exactly significant parts of the input on the borders. This problem
doesn’t appear on the image with the padding removed and the border cropped.
To generate the figures in Section 5 which features images of size 1024 × 1024,
we couldn’t add 256 pixels more on each border, as was done in Section 2.5, due
to memory constraints. Instead we generated images of size 1280 × 1280, which
were then cropped.
4
Hybrid methods
The two main approaches to texture synthesis are the statistics-based methods
and the patch re-arrangement methods. In the first class, a texture is characterized by a statistical signature; then, a random sampling conditioned to
this signature produces genuinely different texture images. Nevertheless, these
methods often fail for macrotextures. The second class boils down to a clever
“copy-paste” procedure, which stitches together verbatim copies of large regions
of the example. A third kind of hybrid methods combines ideas from both approaches, leading to synthesized textures that are everywhere different from the
original but with better quality than the purely statistics-based methods. We
will describe one such method, its multiscale extension and the explicit combination of complementary algorithms.
28
4.1
Local Gaussian models for texture synthesis
Raad et al.’s method [70] uses locally Gaussian (LG) texture model in the patch
space. Each texture patch is modeled by a multivariate Gaussian distribution
learned from its similar patches. Inspired by [20], the idea of searching for
patches to stitch together in the original sample is maintained. However, instead
of using the exact patch taken in the input texture, the stitched patch is sampled
from its Gaussian model. Locally Gaussian patch models have been proved very
useful in image denoising [10]. This approach permits to maintain the coherence
between patches with respect to the input sample, while creating new patches
that do not exist in the sample texture but are still perceptually equivalent to
it.
The multivariate Gaussian models involved are defined by their mean vector
µ and their covariance matrix Σ. For a given patch p, of size P × P pixels, these
parameters are estimated from the set of the R nearest patches Upu (nearest
neighbours of p taken in u) as defined
P
µ = R1 ρ∈Upu ρ,
(7)
P
t
1
Σ = R−1
(ρ
−
µ)
(ρ
−
µ)
.
u
ρ∈U
p
The sampled vector p0 is defined as
p0 = √
X
1
aρ (ρ − µ) + µ,
R − 1 ρ∈U u
aρ ∼ N (0, 1),
(8)
p
where aρ are scalar random variables associated to each patch and following a
normal distribution. Note that p0 follows the distribution N (µ, Σ). These models have reasonable variances, confirming that effectively the patches simulated
have an acceptable degree of innovation [70].
The new texture image is synthesized by stitching together patches sampled
from multivariate Gaussian distributions (8) in the input sample patch space.
The method is iterative: the patches are synthesized in a raster-scan order (top
to bottom and left to right). The goal of each iteration is to generate a new
patch pm,n
(patch in v placed at (m, n)) that is partially defined on a region
v
called the overlap area (see Figure 10). The known part of the patch defines the
set of patches Upum,n
from which its Gaussian model is inferred. The generated
v
patch pm,n
is
then
sampled
as defined in (8). The last step consists in stitching
v
the patch into the output texture using the quilting method of [20].
This synthesis algorithm generates a texture that is perceptually equivalent
to the sample texture yet not composed of patches existing in the input texture.
Thus, this method reduces some of the drawbacks of the statistics-based and
the patch-based methods. Indeed the method yields satisfying results for microand macro-textures, and reduces the verbatim copies of the input. However,
this method remains local and is (like all patch based approaches) not forced to
respect the global statistics of the texture sample.
29
output
output
input
input
Figure 14: Synthesis results of the locally Gaussian method [70]. It works well
for macrotextures. As one can observe in both examples the result is slightly
blurry, a characteristics of the Gaussian model. The parameters used for are
P = 40, R = 30 and O = P/2.
Figure 14 shows two results of the method. The algorithm remains dependent on the choice of the patch size P and of the number of nearest neighbours
R as illustrated in Figure 15. These values may have to be adjusted for each
texture sample. As for the overlap size a convenient value is O = P/2. If
this value is too small then the region used to infer the Gaussian models is not
enough. The patches used to infer the model can be very different on a high
portion of the patch. The algorithm has a low computational complexity, compared for instance with classic patch-based denoising algorithms [49, 13]. An
alternative to reduce the dependency of the method to the patch size is to work
in a multiscale approach.
4.2
Multiscale texture synthesis methods
Most real textures are organized at multiple scales: the global structure is revealed at coarse scales but important detail are present at finer ones. As we
have seen, the results of patch-based methods depend strongly on the patch size.
Small patch sizes may capture the finer details of the input but the resulting texture will lack global coherence. On the other hand, using large patches will maintain the global structures at the risk of a “copy-paste” effect. Furthermore, with
large patches it becomes impossible to model the patch variability due to the
lack of sufficient samples. This is apparent in the examples of Figure 14, where
modeling patches as multivariate Gaussian vectors leads to a slightly blurry
texture. A natural solution is to use a multiscale approach [46, 81, 50, 34, 70]
using several patch sizes for a single texture synthesis, capturing different levels
of details.
This section illustrates the ideas and difficulties of a multiscale extension
using as example the local Gaussian models for texture synthesis presented in
the previous section [70]. The Multi-Scale Locally Gaussian (MSLG) method
works at S scales and can be summarized in a few sentences. The synthesis
begins at the coarsest scale (s = S − 1) using the local Gaussian method where
the quilting step is replaced by a simple average of the overlapping patches. For
30
input
R = 10, P = 20
R = 10, P = 30
R = 20, P = 10
R = 20, P = 20
R = 20, P = 30
R = 30, P = 10
R = 30, P = 20
R = 30, P = 30
Figure 15: Texture synthesis result for the left top corner texture image. We
show the results obtained for different values of R (the number of similar
patches) and P (the side patch size). From left to right P = 10, 20, 30. From top
to bottom, the number of nearest neighbours is R = 10, 20, 30. All the results
are obtained for an overlap of a half patch size O = P/2.
31
the remaining scales (s = S − 2, . . . , 0), a synthesis is performed by using the
result of the previous scale (s + 1) and the sample image at the corresponding
resolution. At each scale the synthesis is done patch by patch in a raster-scan
order. Each new patch, added to the synthesized image, overlaps part of the
previously synthesized patch and it is the combination of a low resolution patch
and a high resolution one sampled from a multivariate Gaussian distribution.
The Gaussian distribution of the high frequencies of a given patch is estimated
from the high frequencies of its m nearest neighbours in the corresponding scale
input image. The synthesis result of the finer scale is the desired output image.
Let us denote the sample texture by u and us , s = 1, . . . , S − 1 are the
zoomed out versions by a factor 2s , s = 1, . . . , S − 1. The synthesis result
at each scale is denoted by vs , s = 1, . . . , S − 1 and v is the synthesis result
returned by the multiscale algorithm. An additional image ṽs is needed at each
scale, corresponding to a low resolution version of vs obtained by interpolating
vs+1 . To estimate the parameters of the Gaussian distribution of the patch
0
0
pvms ,n being processed, the set U ums 0 ,n0 of R nearest patches in us is considered.
pv s
The R nearest neighbours in us to the current patch are those minimizing the
L2 distance restricted to the overlap area:
0
0
m ,n 2
d(pm,n
) =
us , pvs
1 X
(us (m + i, n + j) − vs (m0 + i, n0 + j))2
|O|
(i,j)∈O
+
1
P2
P
−1
X
(ũs (m + i, n + j) − ṽs (m0 + i, n0 + j))2 ,
(9)
i,j=0
where ũs denotes the low resolution of the image us , ũs = us ∗ Gσ and ṽS−1 =
uS−1 ∗Gσ . In (9), the overlap area is denoted as O and the size of patch overlap is
i,j
fixed to P/2. On the set U ums 0 ,n0 only the high frequency of the patches pi,j
us −pũs
pv s
is considered to infer the multivariate Gaussian distribution N (µH , ΣH ). The
patch pm,n
is synthesized as the combination of a low resolution patch pm,n
vs
ṽs
yield from the previous scale with a high resolution one p ∼ N (µH , ΣH ), thus
m,n
pm,n
vs = pṽs + p. For more details please refer to [70].
Figure 16 shows two synthesis examples. In both cases the result is satisfyingly recovering the details of the different scales for reasonable values of the
patch size P = 20. However one can notice that the results are blurry with
respect to the input and this effect is increased with respect to the single scale
approach.
4.3
Combination of methods
A smart combination of complementary methods may keep the advantages of
each one. We will illustrate the methodology by combining a multiscale approach with three other methods:
MSLG+EF The Multi-Scale Locally Gaussian method combined with the
Efros and Freeman method.
32
output
output
input
input
Figure 16: Synthesis results of the multiscale locally Gaussian method [70].
Both examples show that the details of different scales are correctly synthesized
when using a patch size P = 20. However the results are slightly blurred with
respect to the input. The number of scales is S = 3 for the first example (left)
and S = 2 for the second example (right).
MSLG+PS The Multi-Scale Locally Gaussian method combined with the Portilla and Simoncelli method.
MSLG+Gatys The Multi-Scale Locally Gaussian method combined with the
Gatys et al. method.
The combination of the Multi-Scale Locally Gaussian method with the Efros
and Freeman method (MSLG+EF) consists of two steps. The first step synthesizes the given input u with the Multi-Scale Locally Gaussian method generating
a new texture image that we denote umslg . The second step consists in applying the Efros and Freeman algorithm to the given input sample, initializing the
output image that we denote uef with the image umslg . The method is basically
the same as the one described in section 3.2. The only step of the algorithm
that is modified is the patch selection step. In the method described in [20] at
each iteration the added patch was chosen among those (in the input sample)
whose overlap region was similar to the one of the patch under construction.
When combining the methods, instead of only comparing the overlap areas, the
entire patches are compared. Initializing the output with a first synthesis umslg
enables the method to use the whole patch under construction to find a candidate in the input sample u. The candidate patch taken from u is then quilted
in uef at the corresponding position with the same stitching step as in [20].
This combination allows to recover the lost resolution of the MSLG synthesis
as illustrated in Figure 17. However it is not capable of masking the garbage
growing effects as effectively MSLG+PS combination does.
The combination of the Multi-Scale Locally Gaussian method with the Portilla and Simoncelli method (MSLG+PS) consists of two steps. In the first
step, given the input image u, a new texture umslg is generated using MSLG.
The second step uses PS where the initialization “noise image” is replaced by
umslg generating the output image that we denote ups . As explained in section
2.4, the statistics to impose are learnt on the input u. What follows is a synthesis step where the output image is projected on the subspaces of constraints.
There exist several local solutions to this projection step. When initializing PS
33
MSLG
MSLG+EF
input
Figure 17: Synthesis results of the combination of the Multi-Scale Locally Gaussian method with the Efros and Freeman (MSLG+EF).
MSLG
MSLG+PS
input
Figure 18: Synthesis results of the combination of the Multi-Scale Locally Gaussian method with the Portilla and Simoncelli methods (MSLG+PS).
with the result of MSLG, the initialization image is generally quite close to the
images living in the sub-space of the whole set of constraints. Thus the result obtained is improved compared to PS images starting from a random noise image.
Naturally fixing the initialization of the PS algorithm removes the randomness
of the generated texture. But this is not the case since the initialization is itself
random as it is generated from another random process. This combination is
illustrated in Figure 18.
The combination of the Multi-Scale Locally Gaussian method with Gatys’
texture generator (MSLG+Gatys) is very similar to its combination with the
Portilla and Simoncelli method. The texture generator is initialized with the
result of MSLG umslg , and the statistics of the target image are enforced via
several iterations of backpropagation generating the output image denoted as
ugatys. This combination is illustrated in Figure 19.
34
MSLG
MSLG+Gatys
input
Figure 19: Synthesis results of the combination of the Multi-Scale Locally Gaussian method with Gatys’ texture generator (MSLG+Gatys).
5
Experiments
The first part of this section compares the exemplar-based texture synthesis
methods described before on a set of standard textures. These results illustrate
the advantages and limitations of each one. Then, the second part attempts at
the synthesis of real life and more complex textured images, revealing the shortcomings still present in all the methods when confronted with such a demanding
task.
5.1
Comparative evaluation
We will compare the results of the following texture synthesis methods: Random
Phase Noise (RPN) [87, 23], Heeger and Bergen (HB) [33], Portilla and Simoncelli (PS) [69], Gatys (Gatys) [27], SGAN [36], Efros and Leung (EL) [19], Efros
and Freeman (EF) [20], CNNMRF [53] and MSLG [70]. Figures 20 to 23 show
results for various texture samples, one per column; in each figure, the first row
shows the sample image and the following rows correspond, as indicated, to one
of the algorithms. We focus on these original texture synthesis algorithms, and
do not show the numerous variants. For several of our sample textures, these
variants could get better results, but we think that showing the results of the
original algorithms better underlines their intrinsic strengths and weaknesses.
Similarly we won’t present the results of all the combinations of the different
methods.
The second to sixth rows correspond to statistics-based methods described
in Section 2, namely Random Phase Noise, Heeger-Bergen, Portilla-Simoncellli,
Gatys and SGAN. Early statistics-based methods: Heeger and Bergen (1995),
Portilla and Simoncelli (2000) and Random Phase Noise (1991) yield good results for microtextures, i.e. textures with no conspicuous structures, as can be
seen in the first texture example of Figure 20 and to a lesser extent for the
second and third example. The Heeger and Bergen’s method is inspired on a
model of the early visual cortex; it provides satisfying results in some cases but
35
input
RPN [23]
HB [33]
PS [69]
Gatys [27]
SGAN [36]
Figure 20: Comparison of texture synthesis methods. From top to bottom: input
sample, Random Phase Noise (RPN) [23], Heeger and Bergen (HB) [33], Portilla
and Simoncelli (PS) [69], Gatys (Gatys) [27] and SGAN [36].
36
input
EL [19]
EF [20]
CNNMRF [53]
MSLG [70]
Figure 21: Comparison of texture synthesis methods. From top to bottom: input
sample, Efros and Leung (EL) [19], Efros and Freeman (EF) [20], CNNMRF [53]
and MSLG [70].
37
input
RPN [23]
HB [33]
PS [69]
Gatys [27]
SGAN [36]
Figure 22: Comparison of texture synthesis methods. From top to bottom: input
sample, Random Phase Noise (RPN) [23], Heeger and Bergen (HB) [33], Portilla
and Simoncelli (PS) [69], Gatys (Gatys) [27] and SGAN [36].
38
input
EL [19]
EF [20]
CNNMRF [53]
MSLG [70]
Figure 23: Comparison of texture synthesis methods. From top to bottom: input
sample Efros and Leung (EL) [19], Efros and Freeman (EF) [20], CNNMRF [53]
and MSLG [70].
39
there is no theoretical proof of the convergence of the method. On the other
hand, the RPN method yields a simple and elegant theory with no convergence
issue. The visual results yield by both methods are in general satisfying for microtextures. However, for textures with local structures, the results are blurry
and unsatisfying. Among these three methods, the results obtained by Portilla
and Simoncelli are by far the most remarkable. These results contain recognizable configurations from the sample. This can be observed for the last texture
example in Figure 20 and the first two examples in Figure 22. Notice that the
Heeger and Bergen and RPN methods yield unsatisfying results for these three
examples. Clearly the global statistics considered by these methods are not
enough to characterize these highly structured textures.
The second and third rows of the Figures 21 and 23 correspond to the
patch re-arrangement methods Efros-Leung (1999) and Efros-Freeman (2001)
described in Sections 3. The first three textures in Figure 21 have no conspicuous structures but are not stationary (for example, there are small changes of
illumination). The Efros and Leung method, being too local, fails to recover
the global characteristics of these textures. A similar and attenuated behavior
is observed in Efros-Freeman’s results. The methods are significantly better
than their predecessors in the presence of local structure, but have their specific
problems. Efros and Leung’s results for the third and fourth texture in Figure 21 show two clear examples of garbage growing. The method has repeated
a very small part of the input in an inconsistent way creating “garbage”. In
general this phenomenon is more evident in Efros and Leung’s results, compared
to those of Efros and Freeman. The results of Efros-Leung and Efros-Freeman
for the first texture in Figure 23 show that the global organization is sometimes
missed, mostly due to the fact that these methods work at a single scale. The
second texture example in Figure 23 yields impressive results in the case of EfrosFreeman’s method. Nevertheless, looking carefully one can notice the verbatim
copies of the piece of chalks in the input image. The hybrid method MSLG
(2016) described in Section 4, whose results are on the fifth row of Figures 21
and 23, faces the same issues for the three first examples in Figure 21. This is
less visible though, since the Gaussian models tend to smooth slightly the result.
However, the original granularity of the input sample is lost in MSLG. As mentioned in Section 3, Efros and Leung’s and Efros and Freeman’s results depend
on the patch size, while the multiscale approach (MSLG) is more robust to that
parameter. When the former two methods fail to preserve global organization,
MSLG, working at multiple scales, manages to preserve this organization. In
the second texture example in Figure 23, MSLG avoids the verbatim copy since
the patches are being sampled from their Gaussian model and therefore are different from their original patches. Nevertheless, the Gaussian model strongly
smooths the output. The synthesis of the flower texture (Figure 23 third row)
is very satisfying for the three methods. Finally, the pumpkin texture shows a
clear example of the verbatim copy effect in the Efros-Leung and Efros-Freeman
methods. The fourth row of the Figures 21 and 23 corresponds to the CNN patch
re-arrangement method CNNMRF (2016) described in Section 3. By bringing
the patch re-arrangement to CNN features, the blending between the patches
40
is improved. For example on the fourth texture of Figure 23, the separation
between the patches are visible on the results of Efros-Freeman, which is not
the case for CNNMRF. However CNNMRF suffers particularly from verbatim
copies and fails to recover the global statistics of the image. Efros-Leung and
Efros-Freeman do suffer less from these problems because the patch selection
step for these methods picks randomly among a selection of patches, whereas
CNNMRF takes the most likely patch.
The recent statistics-based CNN methods, discussed in Section 2.5, show significant improvement over their predecessors. Gatys (2015) is the best statisticsbased method at respecting the fine details for all the textures of Figures 20
and 22, which can be well noticed with a zoom-in. However, some low frequencies or structure organizations are missed, as seen on the fourth texture of
Figure 20, and some contrast instabilities can be noticed, for example on the
first texture of Figure 22. As discussed in Section 2.5, some variants were proposed to fix these problems. SGAN (2016), on the other hand, better respects
the low frequencies, and the results often look better than Gatys when zoomedout. However on the fine scale, the results are incomplete and noisy, as seen on
all the textures of the Figures 20 and 22. SGAN fails to generate correctly the
first texture of Figure 20, possibly because this texture has no structure and is
a microtexture. It is likely that better results can be obtained by tuning the
parameters, but as said in Section 2.5, the default parameters were used.
Among all these methods, the CNN based methods are the most expensive in
computational time. Pixel based methods, like Efros-Leung, are more expensive
than patch based methods like Efros-Freeman or MSLG. The speed of statisticsbased methods depends on how global the optimization is, and on the number
of iterations needed. Portilla-Simoncelli’s and Heeger-Bergen’s speeds are comparable to patch based methods, while Random Phase Noise is the cheapest of
the methods reviewed here.
These comparative evaluations show the strengths and weaknesses of the
different original methods described in this survey. As said previously, some
variants of these methods can get better results on some pictures. For example,
a better result for the fourth texture of the Figure 20 can be seen on Figure 19.
For this texture, first generating with MSLG, then refining with Gatys’ texture
generator, enables to combine the best of both algorithms: The fabric elements
are well aligned, and look good at a fine scale. Overall, over the last three
decades, tremendous progress was made to generate convincing new texture
samples from a small and stationary texture sample. However one could argue
that the samples used in this comparison are toy examples. Indeed, except for
the third texture of Figure 20, and fourth texture of Figure 22, the samples
do not suffer much of illumination changes or perspective, and are essentially
stationary. Nevertheless, most textures are not stationary. Think for example
of a wood texture. This leads us to wonder whether the presented algorithms
get acceptable results on these complex scenarios.
41
crop 1
crop 2
Figure 24: Two crops of different parts of a larger wood texture. The cropped
images are of size 500 × 500 pixels. Each one represents a different texture
belonging to a single “big texture”.
5.2
Getting out of toy examples
The previous examples present some quite impressive texture synthesis results
by several algorithms. The texture synthesis problem seems to be almost solved
for “academic” textures. Still, those results were obtained for pictures of relatively small size and taken in almost ideal conditions, in order to get almost
stationary textures. In this section, we discuss the situation for more complex
textures: When the same methods are applied to sample images of real and
non-stationary textures, where long-range structure is present as well as varying detail at every scale. Figures 24 and 25 show some realistic examples of real
world images that nobody would hesitate calling textures. Nevertheless on second thoughts they do have a complex, non-stationary structures, because every
large enough image has it. But these are precisely the examples that need being
emulated! In this endeavor, we can relax the requirement that the synthesis
must make a larger image. Let’s just ask if a method is able to reproduce a
perceptually similar texture at the very same size.
Each of textures in Figures 24 and 25 show different salient sub-textures
within the same image. Since the methods in Section 5.1 usually assume that
the texture is stationary, it is not completely fair to use these methods on
these samples. Several works have investigated ways to handle these complex
cases [74, 4, 44, 56]. In this section we show the results of the state of the art
algorithms presented in this paper, and will show that they are still far from
emulating to real world textures, even without the requirement of building a
larger texture patch from the sample.
Figures 26 and 27 show the results of the presented statistics-based, patch rearrangement and hybrid methods on some of these more complex examples. The
best results are Gatys’ texture generator on the second texture and MSLG+PS
and MSLG+Gatys on the first texture and fourth texture. When applying
RPN or PS to them the results obtained are often too blurry. Gatys’ texture
42
crop 1
crop 2
Figure 25: Two crops of different parts of a larger stone texture. The cropped
images are of size 512 × 512 pixels. Each one represents a different texture
belonging to a single “big texture”.
generator fails to catch the low frequency structures for the last two textures.
EF, CNNMRF and MSLG suffer from garbage growing and verbatim copies on
the first three textures. This is true especially when the input is not stationary.
SGAN fails to generate properly on the first two textures, and while the global
organization of the third and fourth pictures is good, it suffers from the noise at
small scale mentioned previously. As noticed in the previous section, the MSLG
results are slightly blurry.
These results show that while some methods can get good results on some of
these challenging texture samples, no method manages to get satisfying results
for all four textures.
6
Conclusion
With the multiplication of applications in computer graphics to the entertainment industry, the interest in the generation of synthetic objects with realistic
texture has grown rapidly. High budget film sets, computer games, and in
some cases digital art, spend intensive human efforts to imitate the appearance
and feel of real world items. For this reason, exemplar-based texture synthesis
has been the focus of intensive work for three decades. And as the available
computational power increased, so has the sophistication of these methods.
In the end of the last millennium, statistics-based methods, such as RPN,
Heeger-Bergen and Portilla-Simoncelli focused on a reduced set of statistics. The
results were quite satisfactory on micro-textures, but could be blurry and far
from the originals for more complex structures. Patch re-arrangement methods,
such as Efros-Leung and Efros-Freeman, managed to respect significantly better the feel and the low level structures of these textures, but could have issues,
such as discontinuities, verbatim copy, garbage growing or simply not respecting
some essential statistics of the textures, such as the average intensity. Hybrid
43
input
RPN [23]
PS [69]
Gatys [27]
SGAN [36]
EF [20]
CNNMRF [53]
Figure 26: Synthesis results for statistical based and patch re-arrangement
methods on complex texture. They show the current limitations of all best
methods. RPN scrambles the textures. PS loses long range coherence of the
wood veins. EF and CNNMRF’s copy paste is quite visible for all textures and
incurs in garbage growing. PS and Gatys have satisfying results on the left hand
two textures, but miss to emulate long range interactions on the wood textures.
SGAN grows periodic noise patterns. 44
input
MSLG
MSLG+EF
MSLG+PS
MSLG+Gatys
Figure 27: Synthesis results for the hybrid methods. In columns 2) and 3),
MSLG has repetitions and garbage growing; thus all the generated results based
on the MSLG outputs keep this defect. In columns 1) and 4), MSLG respects
well the global statistics of the textures, and the combination with other methods indeed improves the result. MSLG+PS and MSLG+Gatys perform better
on these examples than MSLG+EF.
45
methods, such as MSLG, fix some of the issues of patch re-arrangement methods, but still share some of their issues. Very recently, statistics-based methods
have been revisited with Convolutional Neural Networks (CNNs). CNN based
methods significantly increase the number of texture statistics involved in their
model, for example by a factor of 25 approximately in the case of Gatys’ texture generator compared to Portilla-Simoncelli. The results show a spectacular progress over their predecessors, but no method is perfect yet. Patch rearrangement methods were revisited as well by CNNs. CNNMRF improves the
blending between the image patches, but the results still suffer from the problems mentioned above for patch re-arrangement methods. In this review, we
presented three statistics-based neural methods with different models: Gatys’
texture generator, DeepFrame and SGAN. When zoomed-in, the outputs of
Gatys’ texture generator are the best among the statistics-based methods, but
miss some important low frequency constraints of the texture when zoomed-out.
Some variants aim at fixing this shortcoming. SGAN succeeds better on several
examples to respect the global structure of the texture, but the details of the
texture are poor. While all the other statistics-based methods have an explicit
texture model, the SGAN model is more implicit.
Our experimental results look no doubt sometimes worse than in the original
papers, but precisely we did not select the best examples. Our examination of
the history of the method leads to the following conclusions.
- The exemplar-based texture synthesis problem is implicitly ill-posed, as it requires to extrapolate a Fourier spectrum by enlarging the image given a very
small sample of it. Having very small samples may have been historically interesting in computer graphics, but is no longer a technical issue, given the
available memories and computational power in all computers.
- By working on small texture examples the literature has somehow unrealistically restricted the problem. Indeed it is simply not true that textures are as
stationary as those examples suggest.
- When trying to work on larger examples, we have seen that no texture sample
is really stationary. A realistically large texture sample in fact contains smaller
patches of very different textures.
- This explains first why patch based copy-paste methods are doomed in spite
of some apparent success in some quasi-periodic texture with no conspicuous
detail. On more involved samples, they cannot but reproduce recognizable details.
- This also explains why progress in this topic is linked to the design of methods
enforcing more and more statistical parameters. The number of statistics enforced by statistical models is growing fast: 710 for Portilla-Simoncelli, 176640
for the default model in Gatys’ texture generator. With some results showing
that the filters can be chosen with random weights [85], one can wonder if the
solution is not to just use the highest number of statistics possible to emulate a
texture. One may also wonder where to draw a reasonable limit between synthesizing complex textures and rendering scenes containing textured objects.
- Thus, the Portilla-Simoncelli method, of enforcing a high number of statistics,
wins, but it is somewhat a Pyrrhic victory. Indeed, the more random statis46
tics we pile up, the better the exemplar-based results. But it remains to find
numerical tools applying automatically an Occam’s razor as Portilla and Simoncelli did manually. This is still needed to realize the goal of Julesz’ program,
which was to find the minimal sufficient set of statistics rendering two textures
indistinguishable.
Acknowledgements
We thank Rafael Grompone von Gioi for valuable corrections and suggestions,
Arthur Leclaire and Yang Lu for their feedback and for helping produce some
of the experiments, and the anonymous referees for valuable advice.
References
[1] Cecilia Aguerrebere, Yann Gousseau, and Guillaume Tartavel. Exemplarbased texture synthesis: the Efros-Leung algorithm. Image Processing On
Line, 2013:213–231, 2013.
[2] Miika Aittala, Timo Aila, and Jaakko Lehtinen. Reflectance modeling by
neural texture synthesis. ACM Transactions on Graphics, 35(4), 2016.
[3] Michael Ashikhmin. Synthesizing natural textures. In Proceedings of the
2001 symposium on Interactive 3D graphics, pages 217–226. ACM, 2001.
[4] Abdourrahmane M. Atto, Zhangyun Tan, Olivier Alata, and Maxime Moreaud. Non-stationary texture synthesis from random field modeling. In
Image Processing (ICIP), 2014 IEEE International Conference on, pages
4266–4270. IEEE, 2014.
[5] James R. Bergen and Edward H. Adelson. Visual texture segmentation
based on energy measures (a). Journal of the Optical Society of America
A, 3, 1986.
[6] Guillaume Berger and Roland Memisevic. Incorporating long-range consistency in cnn-based texture generation. In Proceedings of the International
Conference on Learning Representations (ICLR), 2017.
[7] Urs Bergmann, Nikolay Jetchev, and Roland Vollgraf. Learning texture
manifolds with the periodic spatial gan. arXiv preprint arXiv:1705.06566,
2017.
[8] Charles Bordenave, Yann Gousseau, and François Roueff. The dead leaves
model: a general tessellation modeling occlusion. Advances in Applied
Probability, pages 31–46, 2006.
[9] Thibaud Briand, Jonathan Vacher, Bruno Galerne, and Julien Rabin. The
Heeger & Bergen pyramid based texture synthesis algorithm. Image Processing On Line, 4:276–299, 2014.
47
[10] Antoni Buades, Marc Lebrun, and Jean-Michel Morel. Implementation of
the “Non-Local Bayes” (NL-Bayes) Image Denoising Algorithm. Image
Processing On Line, 2013:1–42, 2013.
[11] Terry Caelli and Béla Julesz. Experiments in the visual perception of texture. Biol. Cybernetics, 28:167–175, 1978.
[12] Tianhorng Chang and C. C. Jay Kuo. Texture analysis and classification
with tree-structured wavelet transform. IEEE Transactions on Image Processing, 2(4):429–441, 1993.
[13] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-D transform-domain collaborative filtering. Image Processing, IEEE Transactions on, 16(8):2080–2095, 2007.
[14] Jeremy S. De Bonet. Multiresolution sampling procedure for analysis and
synthesis of texture images. In Proceedings of the 24th annual conference
on Computer graphics and interactive techniques, pages 361–368. ACM
Press/Addison-Wesley Publishing Co., 1997.
[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei.
Imagenet: A large-scale hierarchical image database. In Computer Vision
and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages
248–255. IEEE, 2009.
[16] Emily L Denton, Soumith Chintala, Rob Fergus, et al. Deep generative image models using a laplacian pyramid of adversarial networks. In Advances
in neural information processing systems, pages 1486–1494, 2015.
[17] Agnès Desolneux, Lionel Moisan, and Samuel Ronsin. A compact representation of random phase and Gaussian textures. In 2012 IEEE International
Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE,
2012.
[18] Alexey Dosovitskiy and Thomas Brox. Generating images with perceptual
similarity metrics based on deep networks. In Advances in Neural Information Processing Systems, pages 658–666, 2016.
[19] Alexei Efros, Thomas K. Leung, et al. Texture synthesis by non-parametric
sampling. In Computer Vision, 1999. The Proceedings of the Seventh IEEE
International Conference on, volume 2, pages 1033–1038. IEEE, 1999.
[20] Alexei A. Efros and William T. Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th annual conference on Computer
graphics and interactive techniques, pages 341–346. ACM, 2001.
[21] William T. Freeman and Edward H. Adelson. The design and use of steerable filters. IEEE Transactions on Pattern analysis and machine intelligence, 13(9):891–906, 1991.
48
[22] Bruno Galerne, Yann Gousseau, and Jean-Michel Morel. Micro-texture
synthesis by phase randomization. Image Processing On Line, 2011, 2011.
[23] Bruno Galerne, Yann Gousseau, and Jean-Michel Morel. Random phase
textures: Theory and synthesis. Image Processing, IEEE Transactions on,
20(1):257–267, 2011.
[24] Bruno Galerne and Arthur Leclaire. An Algorithm for Gaussian Texture
Inpainting. submitted to Image Processing Online, (7):1–16, 2017. https:
//doi.org/10.5201/ipol.2017.198.
[25] Bruno Galerne and Arthur Leclaire. Texture Inpainting using Efficient
Gaussian Conditional Simulation. SIAM Journal on Imaging Sciences,
10:1446–1474, 2017. https://doi.org/10.1137/16M1109047.
[26] Bruno Galerne, Arthur Leclaire, and Lionel Moisan. Microtexture inpainting through Gaussian conditional simulation. In 2016 IEEE International
Conference on Acoustics, Speech and Signal Processing (ICASSP), pages
1204–1208. IEEE, 2016.
[27] Leon Gatys, Alexander S. Ecker, and Matthias Bethge. Texture synthesis
using convolutional neural networks. In Advances in Neural Information
Processing Systems, pages 262–270, 2015.
[28] Leon Gatys, Alexander S. Ecker, and Matthias Bethge. Code to synthesise
textures using convolutional neural networks as described in gatys et al.
2015. https://github.com/leongatys/DeepTextures, 2015–2016.
[29] Leon Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer
using convolutional neural networks. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pages 2414–2423, 2016.
[30] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David WardeFarley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative
adversarial nets. In Advances in neural information processing systems,
pages 2672–2680, 2014.
[31] Ulf Grenander. General Pattern Theory. Oxford University Press, 1993.
[32] Baining Guo, Harry Shum, and Ying-Qing Xu. Chaos mosaic: Fast and
memory efficient texture synthesis. Microsoft research paper MSR-TR2000-32, 2000.
[33] David J. Heeger and James R. Bergen. Pyramid-based texture analysis/synthesis. In Proceedings of the 22nd annual conference on Computer
graphics and interactive techniques, pages 229–238. ACM, 1995.
[34] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and
David H. Salesin. Image analogies. In Proceedings of the 28th annual
conference on Computer graphics and interactive techniques, pages 327–
340. ACM, 2001.
49
[35] Daniel Jiwoong Im, Chris Dongjoo Kim, Hui Jiang, and Roland Memisevic.
Generating images with recurrent adversarial networks. arXiv preprint
arXiv:1602.05110, 2016.
[36] Nikolay Jetchev, Urs Bergmann, and Roland Vollgraf. Texture synthesis with spatial generative adversarial networks.
arXiv preprint
arXiv:1611.08207, 2016.
[37] Nikolay Jetchev, Urs Bergmann, and Roland Vollgraf. Spatial generative
adversarial networks. https://github.com/zalandoresearch/spatial_
gan, 2016–2017.
[38] Béla Julesz. Visual pattern discrimination.
Transactions on, 8(2):84–92, 1962.
Information Theory, IRE
[39] Béla Julesz. Textons, the elements of texture perception, and their interactions. Nature, 290(5802):91–97, 1981.
[40] Béla Julesz. A theory of preattentive texture discrimination based on firstorder statistics of textons. Biological cybernetics, 41(2):131–138, 1981.
[41] Béla Julesz. Dialogues on Perception. The MIT Press, 1995.
[42] Béla Julesz, Edgar N. Gilbert, Larry A. Shepp, and Harry L. Frisch. Inability of humans to discriminate between visual textures that agree in
second-order statistics—revisited. Perception, 2(4):391–405, 1973.
[43] Béla Julesz, Edgar N. Gilbert, and Jonathan D. Victor. Visual discrimination of textures with identical third-order statistics. Biological Cybernetics,
31(3):137–140, 1978.
[44] Alexandre Kaspar, Boris Neubert, Dani Lischinski, Mark Pauly, and Johannes Kopf. Self tuning texture optimization. In Computer Graphics
Forum, volume 34, pages 349–359. Wiley Online Library, 2015.
[45] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural
information processing systems, pages 1097–1105, 2012.
[46] Vivek Kwatra, Irfan Essa, Aaron Bobick, and Nipun Kwatra. Texture optimization for example-based synthesis. In ACM Transactions on Graphics
(TOG), volume 24, pages 795–802. ACM, 2005.
[47] Vivek Kwatra, Arno Schödl, Irfan Essa, Greg Turk, and Aaron Bobick.
Graphcut textures: image and video synthesis using graph cuts. In ACM
Transactions on Graphics (ToG), volume 22, pages 277–286. ACM, 2003.
[48] Andrew Laine and Jian Fan. Texture classification by wavelet packet signatures. IEEE Transactions on Pattern Analysis and Machine Intelligence,
15(11):1186–1191, 1993.
50
[49] Marc Lebrun, Antoni Buades, and Jean-Michel Morel. A nonlocal Bayesian
image denoising algorithm. SIAM Journal on Imaging Sciences, 6(3):1665–
1688, 2013.
[50] Sylvain Lefebvre and Hugues Hoppe. Parallel controllable texture synthesis.
ACM Transactions on Graphics (TOG), 24(3):777–786, 2005.
[51] Elizaveta Levina and Peter J. Bickel. Texture synthesis and nonparametric
resampling of random fields. The Annals of Statistics, 34(4):1751–1773,
2006.
[52] Chuan Li and Michael Wand. Code for paper ”combining markov random
fields and convolutional neural networks for image synthesis”. https://
github.com/chuanli11/CNNMRF, 2015–2016.
[53] Chuan Li and Michael Wand. Combining markov random fields and convolutional neural networks for image synthesis. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pages 2479–2486,
2016.
[54] Lin Liang, Ce Liu, Ying-Qing Xu, Baining Guo, and Heung-Yeung Shum.
Real-time texture synthesis by patch-based sampling. ACM Transactions
on Graphics (ToG), 20(3):127–150, 2001.
[55] Gang Liu, Yann Gousseau, and Gui-Song Xia. Texture synthesis through
convolutional neural networks and spectrum constraints. In 23rd International Conference on Pattern Recognition, ICPR 2016, Cancún, Mexico,
December 4-8, 2016, pages 3234–3239. IEEE, 2016.
[56] Yitzchak David Lockerman, Basile Sauvage, Rémi Allègre, Jean-Michel
Dischler, Julie Dorsey, and Holly E Rushmeier. Multi-scale label-map extraction for texture synthesis. ACM Trans. Graph., 35(4):140–1, 2016.
[57] Yang Lu, Song-Chun Zhu, and Ying Nian Wu. Code for Learning
FRAME Models Using CNN Filters. http://www.stat.ucla.edu/~yang.
lu/project/deepFrame/main.html.
[58] Yang Lu, Song-Chun Zhu, and Ying Nian Wu. Learning FRAME Models
Using CNN Filters. In Thirtieth AAAI Conference on Artificial Intelligence,
2016.
[59] Jitendra Malik and Pietro Perona. Preattentive texture discrimination with
early vision mechanisms. Journal of the Optical Society of America A,
7(5):923–932, 1990.
[60] B.S. Manjunathi and W. Y. Ma. Texture features for browsing and retrieval
of image data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 18(8):837–842, 1996.
51
[61] Georges Matheron. Schéma booléen séquentiel de partition aléatoire.
Rapport technique N-83, Centre de Morphologie Mathématique, École des
Mines de Paris, 214, 1968.
[62] David Mumford and Agnès Desolneux. Pattern Theory: The Stochastic
Analysis of Real-World Signals. A K Peters, Ltd., 2010.
[63] Roman Novak and Yaroslav Nikulin. Improving the neural algorithm of
artistic style. arXiv preprint arXiv:1605.04603, 2016.
[64] Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel
recurrent neural networks. arXiv preprint arXiv:1601.06759, 2016.
[65] Ken Perlin. An image synthesizer. ACM Siggraph Computer Graphics,
19(3):287–296, 1985.
[66] Gabriel Peyré. Sparse modeling of textures. Journal of Mathematical Imaging and Vision, 34(1):17–31, 2009.
[67] Gabriel Peyré. Texture synthesis with grouplets. IEEE Trans. Pattern.
Anal. Mach. Intell., 4(32):733–746, 2010.
[68] Javier Portilla and Eero P. Simoncelli. Representation and synthesis of
visual texture. http://www.cns.nyu.edu/~lcv/texture/.
[69] Javier Portilla and Eero P. Simoncelli. A parametric texture model based
on joint statistics of complex wavelet coefficients. International Journal of
Computer Vision, 40(1):49–70, 2000.
[70] Lara Raad, Agnès Desolneux, and Jean-Michel Morel. A conditional multiscale locally Gaussian texture synthesis algorithm. Journal of Mathematical
Imaging and Vision, 56(2):260–279, 2016.
[71] Lara Raad and Bruno Galerne. Efros and freeman image quilting algorithm
for texture synthesis. IPOL Journal, 7:1–22, 2017.
[72] Julien Rabin, Gabriel Peyré, Julie Delon, and Marc Bernot. Wasserstein
barycenter and its application to texture mixing. In Scale Space and Variational Methods in Computer Vision, volume 6667 of Lecture Notes in Computer Science, pages 435–446. Springer Berlin / Heidelberg, 2012.
[73] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks.
arXiv preprint arXiv:1511.06434, 2015.
[74] Amir Rosenberger, Daniel Cohen-Or, and Dani Lischinski. Layered shape
synthesis: automatic generation of control maps for non-stationary textures. In ACM Transactions on Graphics (TOG), volume 28, page 107.
ACM, 2009.
52
[75] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh,
Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual
Recognition Challenge. International Journal of Computer Vision (IJCV),
115(3):211–252, 2015.
[76] Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P Kingma. Pixelcnn++: Improving the pixelcnn with discretized logistic mixture likelihood
and other modifications. In Proceedings of the International Conference on
Learning Representations (ICLR), 2017.
[77] Jean Serra. Image analysis and mathematical morphology, v. 1. Academic
press, 1982.
[78] Claude E. Shannon. A mathematical theory of communication. Bell System
Technical Journal, 27(3):379–423, 1948.
[79] Eero P. Simoncelli and William T. Freeman. The steerable pyramid: a
flexible architecture for multi-scale derivative computation. In ICIP (3),
pages 444–447, 1995.
[80] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks
for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
[81] Guillaume Tartavel, Yann Gousseau, and Gabriel Peyré. Variational texture synthesis with sparsity and spectrum constraints. Journal of Mathematical Imaging and Vision, 52(1):124–144, 2014.
[82] M. R. Turner. Texture discrimination by Gabor functions. Biological Cybernetics, 55:71–82, 1986.
[83] Dmitry Ulyanov, Vadim Lebedev, Andrea Vedaldi, and Victor Lempitsky.
Texture networks: Feed-forward synthesis of textures and stylized images.
In Int. Conf. on Machine Learning (ICML), 2016.
[84] M. Unser. Texture classification and segmentation using wavelet frames.
IEEE Transactions on Image Processing, 4(11):1549–1560, 1995.
[85] Ivan Ustyuzhaninov, Wieland Brendel, Leon Gatys, and Matthias Bethge.
What does it take to generate natural textures? In Proceedings of the
International Conference on Learning Representations (ICLR), 2017.
[86] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, koray
kavukcuoglu, Oriol Vinyals, and Alex Graves. Conditional image generation with pixelcnn decoders. In D. D. Lee, M. Sugiyama, U. V. Luxburg,
I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems 29, pages 4790–4798. Curran Associates, Inc., 2016.
[87] Jarke J. van Wijk. Spot noise texture synthesis for data visualization. In
SIGGRAPH ’91, pages 309–318, New York, NY, USA, 1991. ACM.
53
[88] Xiaolong Wang and Abhinav Gupta. Generative image modeling using style
and structure adversarial networks. In European Conference on Computer
Vision, pages 318–335. Springer, 2016.
[89] Li-Yi Wei, Sylvain Lefebvre, Vivek Kwatra, and Greg Turk. State of the
art in example-based texture synthesis. In Eurographics 2009, State of the
Art Report, EG-STAR, pages 93–117. Eurographics Association, 2009.
[90] Li-Yi Wei and Marc Levoy. Fast texture synthesis using tree-structured vector quantization. In Proceedings of the 27th annual conference on Computer
graphics and interactive techniques, pages 479–488. ACM Press/AddisonWesley Publishing Co., 2000.
[91] Pierre Wilmot, Eric Risser, and Connelly Barnes. Stable and controllable
neural texture synthesis and style transfer using histogram losses. arXiv
preprint arXiv:1701.08893, 2017.
[92] Ying Nian Wu, Song-Chun Zhu, and Xiuwen Liu. Equivalence of Julesz
ensembles and FRAME models. International Journal of Computer Vision,
38(3):247–265, 2000.
[93] Jianwen Xie, Yang Lu, Song-Chun Zhu, and Ying Nian Wu. A theory
of generative convnet. In International Conference on Machine Learning,
pages 2635–2644, 2016.
[94] Ciyou Zhu, Richard H. Byrd, Peihuang Lu, and Jorge Nocedal. Lbfgs-b:
Fortran subroutines for large-scale bound constrained optimization. Report
NAM-11, EECS Department, Northwestern University, 1994.
[95] Song-Chun Zhu, Cheng-en Guo, Yizhou Wang, and Zijian Xu. What are
Textons? International Journal of Computer Vision, 62(1–2):121–143,
2005.
[96] Song-Chun Zhu, Xiu Wen Liu, and Ying Nian Wu. Exploring texture
ensembles by efficient Markov Chain Monte Carlo – toward a “trichromacy”
theory of texture. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 22(6):554–569, 2000.
[97] Song-Chun Zhu, Ying Nian Wu, and David Mumford. Minimax entropy
principle and its application to texture modeling. Neural computation,
9(8):1627–1660, 1997.
[98] Song-Chun Zhu, Ying Nian Wu, and David Mumford. Filters, random
fields and maximum entropy (frame): Towards a unified theory for texture
modeling. International Journal of Computer Vision, 27(2):107–126, 1998.
54
| 1cs.CV
|
TIME-DEPENDENT WAVE EQUATIONS ON GRADED GROUPS
arXiv:1705.03047v2 [math.AP] 31 Jul 2017
MICHAEL RUZHANSKY AND CHIARA TARANTO
Abstract. In this paper we consider the wave equations for hypoelliptic homogeneous left-invariant operators on graded Lie groups with time-dependent Hölder
(or more regular) non-negative propagation speeds. The examples are the timedependent wave equation for the sub-Laplacian on the Heisenberg group or on
general stratified Lie groups, or p-evolution equations for higher order operators
on Rn or on groups, already in all these cases our results being new. We establish
sharp well-posedness results in the spirit of the classical result by Colombini, de
Giorgi and Spagnolo. In particular, we describe an interesting local loss of regularity phenomenon depending on the step of the group (for stratified groups) and on
the order of the considered operator.
1. Introduction
In this paper we are interested in the well-posedness of the following Cauchy problem:
2
∂t u(t, x) + a(t)Ru(t, x) = 0, (t, x) ∈ [0, T ] × G,
(1.1)
u(0, x) = u0 (x), x ∈ G,
∂ u(0, x) = u (x), x ∈ G,
t
1
for the time-dependent propagation speed a = a(t) ≥ 0.
In the case of G = Rn and R = −∆, the equation (1.1) is the usual wave equation
with the time-dependent propagation speed and its well-posedness results for Hölder
regular functions a have been obtained by Colombini, de Giorgi and Spagnolo in their
seminal paper [7]. Moreover, it has been shown by Colombini and Spagnolo in [13]
and by Colombini, Jannelli and Spagnolo in [8] that even in the case of G = R and
d2
∞
if a ∈ C ∞
R = − dx
2 the Cauchy problem (1.1) does not have to be well-posed in C
is not strictly positive or if it is in the Hölder class a ∈ C α for 0 < α < 1.
In this paper we obtain new results for the following situations:
(i) G = Hn is the Heisenberg group and R is the positive Kohn-Laplacian on G.
(ii) G is a stratified Lie group and R is a (positive) sub-Laplacian on G.
(iii) G is a graded Lie group in the sense of Folland and Stein [24] and R is any
positive Rockland operator on G, i.e. any positive left-invariant homogeneous
hypoelliptic differential operator on G.
2010 Mathematics Subject Classification. 35L05, 35L30, 43A70, 42A80.
Key words and phrases. Sub-Laplacian, Rockland operator, Gevrey spaces, wave equation,
Heisenberg group, graded groups.
The first author was supported in parts by the EPSRC grant EP/K039407/1 and by the Leverhulme Grant RPG-2014-02. No new data was collected or generated during the course of research.
1
2
M. RUZHANSKY AND C. TARANTO
In fact, our results are for the latter case (iii), the former two cases (i) and (ii) being
its special cases. In particular, already in the cases of G being the Euclidean space
Rn , the Heisenberg group Hn , or any stratified Lie group, the case (iii) above allows
one to consider R to be an operator of any order, as long as it is a positive left(or right-) invariant homogeneous hypoelliptic differential operator. In the case of
Rn these cases of so-called p-evolution equations have been studied in e.g. [4, 5, 6],
however, for more restrictive conditions on a(t) than those considered in this paper.
For a(t) ≡ 1 and G being the Heisenberg group Hn with R being the positive subLaplacian, the wave equation (1.1) was studied by Müller and Stein [36] and Nachman
[37]. Other noncommutative settings with a(t) ≡ 1 have been analysed as well, see
e.g. Helgason [30]. For G being a compact Lie group and −R any Hörmander’s sum
of squares on G the problem (1.1) was studied in [27], and so the results of the present
paper provide a nilpotent counterpart of the results there.
Apart from an independent interest of the subelliptic setting of stratified or graded
Lie groups, these settings are the model cases for many corresponding problems for
general partial differential operators on manifolds in view of the celebrated lifting
theorem of Rothschild and Stein [41].
From the point of view of the time-dependent coefficient a(t), we aim at carrying
out the comprehensive analysis, thus distinguishing between the following four cases:
Case 1: a ∈ Lip([0, T ]), a(t) ≥ a0 > 0;
Case 2: a ∈ C α ([0, T ]), 0 < α < 1, a(t) ≥ a0 > 0;
Case 3: a ∈ C l ([0, T ]), l ≥ 2, a(t) ≥ 0;
Case 4: a ∈ C α ([0, T ]), with 0 < α < 2, a(t) ≥ 0.
The first case is the simplest situation while in the forth case we have an irregular
coefficient that is allowed to be zero at some points. The second and third situations
are ‘intermediate’ cases, in the sense that we have either the regularity or the strict
positivity. We distinguish between these cases because the results and methods of
proofs are rather different.
We note that if the operator R is not elliptic, the local approach to the Cauchy
problem (1.1) is problematic since the equation is only weakly hyperbolic already in
Case 1 above. Consequently, since the equation (1.1) in local coordinates is the
space-dependent variable multiplicities problem, very little is known about its wellposedness. In this direction, only very special results for some second order operators
are available, see e.g. Nishitani [38] or Melrose [34]. Non-Lipschitz coefficients have
been also much analysed, see e.g. Colombini and Métivier [11] or Colombini and
Lerner [10]. In addition to already mentioned restrictions for the well-posedness, see
also Colombini and Métivier [12] for a recent overview from the point of view of
systems.
In the case of Rn and −R being the Laplacian, the regularity of a less than Hölder
such as discontinuous or measure-valued a have been considered in [28]. However,
such low regularity requires very different methods, and this problem for the general
Cauchy problem (1.1) will be considered elsewhere.
Wave equations with time dependent coefficients for general densely defined operators with discrete spectrum acting in Hilbert spaces have been considered in [42].
However, that setting is different from the present one since the spectrum in our
situation is continuous.
TIME-DEPENDENT WAVE EQUATIONS ON GRADED GROUPS
3
To formulate our results, let us briefly introduce some necessary notation following,
for example, Folland and Stein [24]. Let G be a graded Lie group, i.e. a connected
simply connected Lie group such that its Lie algebra g has a vector space decomposition
g = ⊕∞
j=1 Vj ,
(1.2)
such that all but finitely many of the Vj ’s are {0} and [Vi , Vj ] ⊂ Vi+j . A special case
analysed in detail by Folland [21] is of stratified Lie groups when the first stratum
V1 generates g as an algebra, see also Folland and Stein [23]. A typical example of
such Lie group is the Heisenberg group. In general, graded Lie groups are necessarily
homogeneous and nilpotent. Moreover, any graded Lie group can be viewed as some
Rn with a polynomial group law. We can also refer to [19, Section 3.1] for a detailed
discussion of graded Lie groups and their properties.
Let R be a positive Rockland operator on G, that is, a positive (in the operator
sense) left-invariant differential operator which is homogeneous of degree ν > 0 and
which satisfies the so-called Rockland condition. This means that for each represenb except for the trivial one, the operator π(R) is injective on the space
tation π ∈ G,
of smooth vectors Hπ∞ , i.e.
∀v ∈ Hπ∞
π(R)v = 0 =⇒ v = 0.
(1.3)
Alternative characterisations of such operators have been considered by Rockland
[40] and Beals [1], until the definitive result of Helffer and Nourrigat [29] saying that
Rockland operators are precisely the left-invariant homogeneous hypoelliptic differential operators on G. The existence of Rockland operators on general nilpotent Lie
groups characterises precisely the class of graded Lie groups [35, 45]. An example of
a positive Rockland operator is the positive sub-Laplacian on a stratified Lie group:
if G is a stratified Lie group and {X1 , . . . , Xk } is a basis for the first stratum of its
Lie algebra, then the positive sub-Laplacian
L=−
k
X
Xj2
j=1
is a positive Rockland operator. Moreover, for any m ∈ N, the operator
R = (−1)
m
k
X
Xj2m
j=1
is a positive Rockland operator on the stratified Lie group G. More generally, for any
graded Lie group G ∼ Rn , if X1 , . . . , Xn is the basis of its Lie algebra g with dilation
weights ν1 , . . . , νn , i.e. with
Dr X j = r ν j X j ,
j = 1, . . . , n, r > 0,
where Dr are dilations on g, then the operator
R=
n
X
j=1
ν0
νj
ν
2 ν0
(−1) aj Xj
j
,
aj > 0,
(1.4)
4
M. RUZHANSKY AND C. TARANTO
is a Rockland operator of homogeneous degree 2ν0 , if ν0 is any common multiple
of ν1 , . . . , νn . We refer to [19, Section 4.1.2] for other examples and a detailed discussion of Rockland operators and graded Lie groups. In the case of Rn , all elliptic
homogeneous differential operators with constant coefficients are Rockland operators.
To formulate our results we will need two scales of spaces, namely, Sobolev and
Gevrey spaces, adapted to the setting of graded Lie groups. Thus, let G be a graded
Lie group and let R be a positive Rockland operator of homogeneous degree ν. For
s
any real number s ∈ R, the Sobolev space HR
(G) is the subspace of S ′ (G) obtained
as the completion of the Schwartz space S(G) with respect to the Sobolev norm
s
kf kHRs (G) := k(I + R) ν f kL2 (G) .
(1.5)
For stratified Lie groups such spaces and their properties have been extensively analysed by Folland in [21] and on general graded Lie groups they have been investigated
in [18, 19]. In particular, these spaces do not depend on a particular choice of the
Rockland operartor R used in the definition (1.5), see [19, Theorem 4.4.20]). These
spaces perfectly suit Case 1 described above but already in the Euclidian case, with
the elliptic Laplace operator instead of the hypoelliptic Rockland operator in the
wave equation (1.1), if the coefficient a(t) is not Lipschitz regular or may become
zero, the Gevrey spaces appear naturally (see e.g. Bronshtein [3]) since we can not
expect anymore the well-posedness in C ∞ (G) or D ′ (G). Indeed, Colombini and Spagnolo exhibited a concrete example in [13] of a Cauchy problem for the time-dependent
wave equation on R with smooth a ≥ 0 which is not well-posed in C ∞ (R) or D ′ (R).
Thus, given s ≥ 1, we define the Gevrey type space
1
s
GR
(G) := {f ∈ C ∞ (G) | ∃A > 0 : keAR 2s f kL2 (G) < ∞}.
(1.6)
These spaces provide for a subelliptic version of the usual Gevrey spaces. For example,
for G = Hn being the Heisenberg group with the basis X1 , . . . , X2n of the first stratum,
s
(Hn ) if and only if there exist two constants B, C > 0
it was shown in [20] that f ∈ GR
2n
such that for every α ∈ N0 the following inequality holds
k∂ α f kL2 (Hn ) ≤ CB |α| (α!)s ,
P
(1.7)
where ∂ α = Y1 . . . Y|α| , with Yj ∈ {X1 , . . . , X2n } for every j = 1, . . . , |α| and Yj =Xk 1 =
αk for every k = 1, . . . , 2n.
Gevrey spaces (1.6) and the corresponding spaces of ultradistributions have been
considered on compact Lie groups and on compact manifolds in [16] and in [17],
respectively.
By an argument similar to that in [20] for the sub-Laplacian or in [16, Theorem
2.4] for elliptic operators, it can be shown that if R is a positive Rockland operator of
s
homogeneous degree ν, then f ∈ GR
(G) if and only if there exist constants B, C > 0
such that for every k ∈ N ∪ {0} we have
kRk f kL2 (G) ≤ CB νk ((νk)!)s .
(1.8)
Since Sobolev spaces do not depend on a particular choice of the Rockland operator
used in their definition, the characterisation (1.8) of the Gevrey spaces implies that
s
the same is true for GR
(G).
s
s
Thus, we may drop the subscript R in HR
and GR
but we may also keep using it
to refer to the norms that we may be using.
TIME-DEPENDENT WAVE EQUATIONS ON GRADED GROUPS
5
Let us now formulate the main theorem of our paper, where we consider the following four cases:
Case
Case
Case
Case
1:
2:
3:
4:
a ∈ Lip([0, T ]), a(t) ≥ a0 > 0;
a ∈ C α ([0, T ]), 0 < α < 1, a(t) ≥ a0 > 0;
a ∈ C l ([0, T ]), l ≥ 2, a(t) ≥ 0;
a ∈ C α ([0, T ]), with 0 < α < 2, a(t) ≥ 0.
These are the four cases to which we refer repeatedly throughout this paper.
Theorem 1.1. Let G be a graded Lie group and let R be a positive Rockland operator
of homogeneous degree ν. Let T > 0. Then the following holds, referring respectively
to Cases 1-4 above:
s+ ν
s
(G),
Case 1: Given s ∈ R, if the initial Cauchy data (u0 , u1) are in HR 2 (G) × HR
ν
s+ 2
then there exists a unique solution of (1.1) in the space C([0, T ], HR (G)) ∩
s
C 1 ([0, T ], HR
(G)), satisfying the following inequality for all values of t ∈
[0, T ]:
ku(t, ·)k2 s+ ν2 + k∂t u(t, ·)k2HRs ≤ C(ku0 k2 s+ ν2 + ku1 k2HRs );
HR
HR
(1.9)
s
s
Case 2: If the initial Cauchy data (u0 , u1) are in GR
(G) × GR
(G), then there exists a
2
s
unique solution of (1.1) in C ([0, T ], GR(G)), provided that
α
1≤s<1+
;
1−α
s
s
(G) × GR
(G), then there exists a
Case 3: If the initial Cauchy data (u0 , u1) are in GR
2
s
unique solution of (1.1) in C ([0, T ], GR(G)), provided that
l
1≤s<1+ ;
2
s
s
Case 4: If the initial Cauchy data (u0 , u1 ) are in GR
(G) × GR
(G) then there exists a
2
s
unique solution of (1.1) in C ([0, T ], GR(G)), provided that
1 ≤ s < 1 + α.
As it will follow from the proof, in Cases 2 and 4, one can take the equalities
α
and s = 1 + α, respectively, provided that T > 0 is small enough. We
s = 1 + 1−α
refer to [25, 26] concerning the sharpness of the above Gevrey indices in the case of
G = Rn and R = −∆, and for further relevant references for that case.
Let us formulate a corollary from Theorem 1.1 showing the local loss of regularity
for the Cauchy problem (1.1). We recall that any graded Lie group G can be identified,
for example through the exponential mapping, with the Euclidean space Rn where n
is the topological dimension of G. Then, if ν1 , . . . , νn are the dilation weights on G
as in (1.4), for any s ∈ R we have the local Sobolev embedding theorems:
s/ν
s/ν
s
Hloc 1 (Rn ) ⊂ HR,loc
(G) ⊂ Hloc n (Rn ),
(1.10)
see [19, Theorem 4.4.24]. If G is a stratified Lie group, we have ν1 = 1 and νn is the
step of G, i.e. the number of steps in the stratification of its Lie algebra. In other
6
M. RUZHANSKY AND C. TARANTO
words, if G is a stratified Lie group of step r and H s (G) is the Sobolev space defined
using (any) sub-Laplacian on G, then the embeddings (1.10) are reduced to
s/r
s
s
Hloc
(Rn ) ⊂ Hloc
(G) ⊂ Hloc (Rn ).
(1.11)
These embeddings are sharp, see Folland [21]. Consequently, using the characterisas
tion (1.8) of GR
(G), we also obtain the embeddings
ν1 s
sνn
s
(Rn ),
(Rn ) ⊂ GR,loc
(G) ⊂ Gloc
Gloc
(1.12)
σ
where the space Gloc
(Rn ) is the usual Euclidean Gevrey space, namely, the space of
all smooth functions f ∈ C ∞ (Rn ) such that for every compact set K ⊂ Rn there exist
two constants B, C > 0 such that for every α we have
|∂ α f (x)| ≤ CB |α| (α!)σ
for all x ∈ K.
(1.13)
Consequently, if G is a stratified Lie group of step r we have the embeddings
s
s
sr
Gloc
(Rn ) ⊂ GR,loc
(G) ⊂ Gloc
(Rn ).
(1.14)
Consequently, using these embeddings, we obtain the following local in space wellposedness result using the usual Euclidean Gevrey spaces. Here we may also assume
that the Cauchy data are compactly supported due to the finite propagation speed
of singularities. To emphasise the appearing phenomenon of local loss of Euclidean
regularity we formulate it in the simplified setting of stratified Lie groups, with topological identification G ∼ Rn .
Corollary 1.2. Let G ∼ Rn be a stratified Lie group of step r and let R be a positive
Rockland operator of homogeneous degree ν (for example, R can be a positive subLaplacian in which case we have ν = 2). Assume that the Cauchy data (u0 , u1 ) are
compactly supported. Then the following holds, referring respectively to Cases 1-4
above:
ν
Case 1: Given s ∈ R, if (u0 , u1) are in H s+ 2 (Rn )×H s (Rn ), then there exists a unique
ν
solution of (1.1) in C([0, T ], H (s+ 2 )/r (Rn )) ∩ C 1 ([0, T ], H s/r (Rn )), satisfying
the following inequality for all values of t ∈ [0, T ]:
ku(t, ·)k2H (s+ ν2 )/r + k∂t u(t, ·)k2H s/r ≤ C(ku0k2H s+ ν2 + ku1k2H s );
(1.15)
Case 2: If (u0 , u1 ) are in G s (Rn )×G s (Rn ), then there exists a unique solution of (1.1)
in C 2 ([0, T ], G sr (Rn )), provided that
α
;
1<s<1+
1−α
Case 3: If (u0 , u1 ) are in G s (Rn )×G s (Rn ), then there exists a unique solution of (1.1)
in C 2 ([0, T ], G sr (Rn )), provided that
l
1<s<1+ ;
2
s
n
s
n
Case 4: If (u0 , u1 ) are in G (R ) × G (R ) then there exists a unique solution of (1.1)
in C 2 ([0, T ], G sr (Rn )), provided that
1 < s < 1 + α.
TIME-DEPENDENT WAVE EQUATIONS ON GRADED GROUPS
7
The statements in Cases 2-4 for s = 1 are not so interesting, with the analytic
well-posedness known in these case anyway, see Bony and Shapira [2].
For G = Rn and R being the Laplacian, we have r = 1 and there is no loss of
regularity in any of the Cases 1-4, when the results are known from [7, 9, 25, 26, 33].
However, already on the Heisenberg group with step r = 2, we observe the local
loss of regularity in Euclidean Sobolev and Gevrey spaces in all statements of Cases
1-4 in Corollary 1.2.
We also note that using local Sobolev and Gevrey embeddings (1.10) and (1.12),
it is easy to formulate an extension of Corollary 1.2 to general graded Lie groups.
2. Preliminaries on graded Lie groups and Rockland operators
In this section we recall some preliminaries and fix the notation concerning the
Fourier analysis on graded Lie groups. We refer to [24] and to [19, Chapter 5] for
further details.
Thus, a connected and simply connected Lie group G is called graded when its Lie
algebra is graded in the sense of the decomposition (1.2).
A Lie algebra g is stratified if it is graded and if its first stratum V1 generates g as
an algebra. Thus, in this case every element of the Lie algebra can be written as a
linear combination of elements in V1 and their iterated commutators. A Lie group is
stratified when it is connected, simply connected and its Lie algebra is stratified.
Furthermore, if there are r non zero Vj ’s in the vector space decomposition (1.2),
then the group (respectively the algebra) is said to be stratified of step r.
From the definition of a stratified Lie algebra, it follows that, assuming that V1 has
dimension k, any basis {X1 , . . . , Xk } for V1 forms a Hörmander system, see [31], and
we can consider its associated sub-Laplacian operator that is also a positive Rockland
operator:
L := − X12 + · · · + Xk2 .
(2.1)
Example 2.1 (The Heisenberg group). A classical example of a graded (stratified)
Lie group is the Heisenberg group Hn that might be seen as the manifold R2n+1
endowed with the group law
1
(x, y, t)(x′, y ′ , t′ ) := (x + x′ , y + y ′ , t + t′ + (x · y ′ − x′ · y)),
2
where (x, y, t), (x′ , y ′ , t′ ) ∈ Rn × Rn × R ∼ Hn . The Heisenberg Lie algebra hn
associated with the Heisenberg group is the space of all the left-invariant vector fields
of Hn and it admits the following canonical basis:
yj
Xj = ∂xj − ∂t ,
2
xj
Yj = ∂yj + ∂t ,
2
T = ∂t .
The former vector fields satisfy the canonical commutation relations
[Xj , Yj ] = T,
∀ j = 1, . . . , n,
8
M. RUZHANSKY AND C. TARANTO
and all the other possible commutators are zero. Therefore, the Heisenberg group is
a graded (stratified) Lie group of step 2, whose Lie algebra admits the vector space
decomposition
hn = V 1 ⊕ V 2 ,
where
V1 =
n
X
j=1
RXj ⊕ RYj
and
V2 = RT.
Hypoellipticity and other questions on the Heisenberg group have a long history, see
e.g. Taylor [44], Folland [22], or Thangavelu [46], and many references therein.
From now on, we consider G to be a graded Lie group, even if some of the following
definitions and remarks hold in a more general setting.
Let π be a representation of G on the separable Hilbert space Hπ . A vector v ∈ Hπ
is said to be smooth or of type C ∞ if the function
G ∋ x 7→ π(x)v ∈ Hπ
is of class C ∞ . The space of all smooth vectors of a representation π is denoted by
Hπ∞ . Let g be the Lie algebra of G and let π be a strongly continuous representation
of G on a Hilbert space Hπ . For every X ∈ g and v ∈ Hπ∞ we define
1
dπ(X)v := lim π expG (tX) v − v .
t→0 t
Then dπ is a representation of g on Hπ∞ (see e.g. [19, Proposition 1.7.3]) called the
infinitesimal representation associated to π. By abuse of notation, we will often still
denote it by π, therefore, for any X ∈ g, we write π(X) meaning dπ(X).
Any left-invariant differential operator T on G, according to the Poincaré-BirkhoffWitt theorem, can be written in a unique way as a finite sum
X
T =
cα X α ,
(2.2)
|α|≤M
where all but finitely many of the coefficients cα ∈ C are zero and X α = X1 . . . X|α| ,
with Xj ∈ g. This allows one to look at any left-invariant differential operator T
on G as an element of the universal enveloping algebra U(g) of the Lie algebra of G.
b yields a field of
Therefore, the family of infinitesimal representations π(T ), π ∈ G
operators that turns to be the symbol associated with the operator T .
b and let R be a positive Rockland operator of homogeneous degree ν > 0,
Let π ∈ G
then using formula (2.2), the infinitesimal representation of R associated to π is
X
π(R) =
cα π(X)α ,
[α]=ν
α
α
π(X1α1
· · · Xnαn )
and [α] = ν1 α1 + · · · + νn αn is the homogewhere π(X) = π(X ) =
neous degree of the multiindex α, with Xj being homogeneous of degree νj .
The operator R and its infinitesimal representations π(R) are densely defined on
D(G) ⊂ L2 (G) and Hπ∞ ⊂ Hπ , respectively, see e.g. [19, Proposition 4.1.15]. We
denote by R2 the self-adjoint extension of R on L2 (G) and we keep the same notation π(R) for the self-adjoint extensions on Hπ of the infinitesimal representations.
TIME-DEPENDENT WAVE EQUATIONS ON GRADED GROUPS
9
Recalling the spectral theorem for unbounded operators [39, Theorem VIII.6], we can
consider the spectral measures E and Eπ corresponding to R2 and π(R), so that we
have
Z
Z
R2 =
λdE(λ) and π(R) =
λdEπ (λ).
R
R
Furthermore, for any f ∈ L2 (G) we have
F φ R f (π) = φ π(R) fb(π),
(2.3)
for any measurable bounded function φ on R, see e.g. [19, Corollary 4.1.16]. The
infinitesimal representations π(R) of a positive Rockland operator are also positive,
due to the relations between their spectral measures. In particular, Hulanicki, Jenkins
b \ {1}, is discrete
and Ludwig showed in [32] that the spectrum of π(R), with π ∈ G
and lies in (0, ∞). This implies that we can choose an orthonormal basis for Hπ such
that the infinite matrix associated to the self-adjoint operator π(R) has the form
2
π1 0 . . . . . .
0 π22 0 . . .
..
(2.4)
π(R) = ..
,
.
. 0
..
..
..
.
.
.
b \ {1}.
where πj are strictly positive real numbers and π ∈ G
3. Parameter dependent energy estimates
In this section we prove certain energy estimates for second order ordinary differential equations with explicit dependence on parameters. This will be crucial in the
proof of Theorem 1.1 where the parameters will correspond to the spectral decomposition (2.4) of the infinitesimal representations of the Rockland operators.
Results of the following type have been of use in different estimates related to
weakly hyperbolic partial differential equations, such as [9] and [27]. However, in those
papers the conclusions rely on more general results, see [25]. We partly follow the
argument in [27] based on a standard reduction to a first order system. Consequently,
we carry out different types of arguments depending on assumptions in each of the
cases, altogether allowing us to formulate the precise dependence on parameters for
ordinary differential equations corresponding to the propagation coefficient a(t) as in
Cases 1-4 of Theorem 1.1, to which we refer in the following statement.
Proposition 3.1. Let β > 0 be a positive constant and let a(t) be a function that
behaves according to Cases 1, 2, 3 and 4 in Theorem 1.1. Let T > 0. Consider the
following Cauchy problem:
′′
2
v (t) + β a(t)v(t) = 0 with t ∈ (0, T ],
(3.1)
v(0) = v0 ∈ C,
v ′ (0) = v ∈ C.
1
Then the following holds:
10
M. RUZHANSKY AND C. TARANTO
Case 1: There exists a positive constant C > 0 such that for all t ∈ [0, T ] we have
β 2 |v(t)|2 + |v ′(t)|2 ≤ C(β 2 |v0 |2 + |v1 |2 ).
Case 2: There exist two positive constants C, K > 0 such that for all t ∈ [0, T ] we
have
1
(3.2)
β 2 |v(t)|2 + |v ′ (t)|2 ≤ CeKtβ s (β 2 |v0 |2 + |v1 |2 ),
α
for any 1 ≤ s < 1 + 1−α . Moreover, there exists a constant k > 0 such that
1−α− 1
s
for any β0 ≥ 1 the estimate (3.2) holds for K = kβ0
for all β ≥ β0 .
Case 3: There exist two positive constants C, K > 0 such that for all t ∈ [0, T ] we
have
1
l
β 2 |v(t)|2 + |v ′ (t)|2 ≤ C(1 + β σ )eKβ σ β 2 |v0 |2 + |v1 |2 ,
with σ = 1 + 2l .
Case 4: There exist two positive constants C, K > 0 such that
α
1
β 2 |v(t)|2 + |v ′ (t)|2 ≤ C(1 + β α+1 )eKtβ s (β 2 |v0 |2 + |v1 |2 ),
(3.3)
for any 1 ≤ s < 1 + α. Moreover, there exists a constant k > 0 such that for
1
−1
any β0 ≥ 1 the estimate (3.3) holds for K = kβ01+α s for all β ≥ β0 .
The constants C in the above inequalities may depend on T but not on β.
Proof. First we reduce the problem (3.1) to a first order system. In order to do this
we rewrite it in a standard way as a matrix-valued equation. Thus we define the
column vectors
iβv0
iβv(t)
,
, V0 :=
V (t) :=
v1
∂t v(t)
and the matrix
0 1
A(t) :=
,
a(t) 0
that allow us to reformulate the second order system (3.1) as the first order system
(
Vt (t) = iβA(t)V (t),
(3.4)
V (0) = V0 .
We will now treat each case separately.
3.1. Case 1: a ∈ Lip([0, T]), a(t) ≥ a0 > 0. This is the simplest case that can be
treated by a classical
argument. We observe that the eigenvalues of our matrix A(t)
p
are given by ± a(t). The symmetriser S of A, i.e. the matrix such that
is given by
Thus we define the energy as
SA − A∗ S = 0,
2a(t) 0
S(t) =
.
0
2
E(t) := S(t)V (t), V (t) ,
TIME-DEPENDENT WAVE EQUATIONS ON GRADED GROUPS
11
and we want to estimate its variations in time. A straightforward calculation yields
the following inequality that will help us to get such estimate:
2|V |2 min {a(t), 1} ≤ E(t) ≤ 2|V |2 max {a(t), 1}.
t∈[0,T ]
(3.5)
t∈[0,T ]
In particular, in this case the continuity of a(t) ensures the existence of two strictly
positive constants a0 and a1 such that
a0 = min a(t) and a1 = max a(t).
t∈[0,T ]
t∈[0,T ]
Thus setting c0 := 2 min{a0 , 1} and c1 := 2 max{a1 , 1}, the inequality (3.5) becomes
c0 |V (t)|2 ≤ E(t) ≤ c1 |V (t)|2 .
(3.6)
A straightforward calculation, together with (3.6), gives the following estimate:
Et (t) = St (t)V (t), V (t) + S(t)Vt (t), V (t) + S(t)V (t), Vt (t) =
= St (t)V (t), V (t) + iβ S(t)A(t)V (t), V (t) − iβ S(t)V (t), A(t)V (t) =
= St (t)V (t), V (t) + iβ S(t)A(t) − A∗ (t)S(t) V (t), V (t) =
= St (t)V (t), V (t) ≤ kSt (t)k|V (t)|2 ,
(3.7)
thus setting c′ := c−1
0 supt∈[0,T ] kSt (t)k, we get from (3.7) using (3.6) that
Et (t) ≤ c′ E(t).
(3.8)
Applying the Gronwall lemma to (3.8), we deduce that there exists a constant c > 0
independent of t ∈ [0, T ] such that
E(t) ≤ cE(0).
(3.9)
Therefore, putting together (3.9) and (3.6) we obtain
c0 |V (t)|2 ≤ E(t) ≤ cE(0) ≤ cc1 |V (0)|2.
We can then rephrase this, asserting that there exists a constant C > 0 independent
of t such that |V (t)|2 ≤ C|V (0)|2. Then we write this inequality going back to the
definition of V (t), yielding
β 2 |v(t)|2 + |∂t v(t)|2 ≤ C β 2 |v0 |2 + |v1 |2 ,
as required.
3.2. Case 2: a ∈ C α ([0, T]), with 0 < α < 1, a(t) ≥ a0 > 0. Here we follow the
method developed by Colombini and Kinoshita [9] for n = 1 and subsequently extended [25] for any n ∈ N. We look for solutions of the form
1
V (t) = e−ρ(t)β s (det H(t))−1 H(t)W (t),
(3.10)
where
• s ∈ R depends on α as will be determined in the argument;
• the function ρ = ρ(t) ∈ C 1 ([0, T ]) is real-valued and will be chosen later;
• W (t) is the energy;
12
M. RUZHANSKY AND C. TARANTO
• H(t) is the matrix defined by
H(t) :=
1
1
,
λǫ1 (t) λǫ2 (t)
where for all ǫ > 0, λǫ1 (t) and λǫ2 (t) are regularisations of the eigenvalues of
the matrix A(t) of the form
√
λǫ1 (t) := (− a ∗ ϕǫ )(t),
√
λǫ2 (t) := (+ a ∗ ϕǫ )(t),
with {ϕǫ (t)}ǫ>0 being a family of cut-off
R functions defined starting from a
non-negative function ϕ ∈ Cc∞ (R), with R ϕ = 1, by setting ϕǫ (t) := 1ǫ ϕ ǫt .
By construction, it follows that λǫ1 , λǫ2 ∈ C ∞ ([0, T ]).
Furthermore,p
we can easily check, using the Hölder regularity of a(t) of order α and,
therefore, of a(t) of the same order α, the following inequalities:
p
√
(3.11)
det H(t) = λǫ2 (t) − λǫ1 (t) ≥ 2 a(0) ≥ 2 a0 ,
and for all t ∈ [0, T ] and ǫ > 0 there exist two constants c1 , c2 > 0 such that
p
|λǫ1 (t) + a(t)| ≤ c1 ǫα ,
p
|λǫ2 (t) − a(t)| ≤ c2 ǫα ,
(3.12)
uniformly in t and ǫ. Now we substitute our suggested solution (3.10) in (3.4) yielding
1 H (t)W (t)
1 H(t)W (t)
H(t)W (t)
t
t
+ e−ρ(t)β s
+ e−ρ(t)β s
+
det H(t)
det H(t)
det H(t)
1
1
H(t)W (t)
−ρ(t)β s H(t)W (t)
.
− e−ρ(t)β s (det H)t (t)
2 = iβA(t)e
det H(t)
det H(t)
1
1
− ρ′ (t)β s e−ρ(t)β s
1
Multiplying both sides of this equality by eρ(t)β s det H(t)H −1(t) we get
−1
1
Wt (t) = ρ′ (t)β s W (t) − H −1 (t)Ht (t)W (t) + (det H)t (t) det H(t) W (t)+
+ iβH −1 (t)A(t)H(t)W (t).
(3.13)
This leads to the estimate
d
|W (t)|2 = Wt (t), W (t) + W (t), Wt (t) = 2Re Wt (t), W (t) =
dt
1
= 2 ρ′ (t)β s |W (t)|2 − Re H −1 (t)Ht (t)W (t), W (t) +
−1
+ det H(t) (det H)t (t)|W (t)|2 + βIm H −1 (t)A(t)H(t)W (t), W (t) .
We observe that
2Im H −1AHW, W = H −1 AHW, W − H −1AHW, W =
= H −1AHW, W − W, H −1 AHW = H −1AHW, W − (H −1 AH)∗ W, W =
= H −1 AH − (H −1 AH)∗ W, W ≤ kH −1AH − (H −1 AH)∗ kkW k2.
TIME-DEPENDENT WAVE EQUATIONS ON GRADED GROUPS
13
Thus we obtain
−1
1
d
|W (t)|2 ≤ 2ρ′ (t)β s + 2kH −1(t)Ht (t)k + 2 det H(t) (det H)t (t) +
dt
+βkH −1AH − (H −1 AH)∗ k |W (t)|2 .
(3.14)
To proceed we need to estimate the following quantities:
I) kH −1(t)Ht (t)k;
−1
II) det H(t) (det H)t (t) ;
III) kH −1AH − (H −1AH)∗ k.
In [25] and [9], the authors determine estimates for similar functions in a more general
setting, i.e. starting from an equation of arbitrary order m. In this particular case, we
can proceed by straightforward calculations without relying on the mentioned works.
We deal with these three terms as follows:
ǫ
λ2 −1
0
0
1
−1
I) Since H (t) = λǫ −λǫ
, it follows that
and Ht (t) =
2
1
−λǫ1 1
∂t λǫ1 ∂t λǫ2
∂t λǫ
the entries of the matrix H −1 Ht are given by the functions λǫ −λj ǫ . We have, for
2
1
example for λ2 ,
1Z p
√
1√
′ t
ǫ
=
a(t − ρǫ)ϕ′ (ρ)dρ =
∂t λ2 (t) = a ∗ ∂t ϕǫ (t) = 2 a ∗ ϕ
ǫ
ǫ
ǫ
Z p
Z
p
′
1p
1
a(t − ρǫ) − a(t) ϕ (ρ)dρ +
a(t) ϕ′ (ρ)dρ ≤ kǫα−1 , (3.15)
=
ǫ
ǫ
√
where we are using the Hölder continuity
R ′ of a for the first term and the fact
that the second term is zero, since ϕ = 0. Combining the inequalities (3.11)
and (3.15), we get for a suitable positive constant k1 that
kH −1 (t)Ht (t)k ≤ k1 ǫα−1 .
II) First we can estimate
det H(t)
therefore,
−1
(det H)t (t) =
det H(t)
−1
∂t λǫ2 − ∂t λǫ1
2∂t λǫ2
2kǫα−1
=
≤
√ ,
λǫ2 − λǫ1
λǫ2 − λǫ1
2 a0
(det H)t (t) ≤ k2 ǫα−1 ,
for a constant k2 > 0.
III) Also in this case, we write explicitly the matrix we are interested in, that is
−2a(t)+(λǫ1 )2 +(λǫ2 )2
0
∗
λǫ1 −λǫ2
.
H −1 AH − H −1 AH = 2a(t)− (λǫ )2 +(λǫ )2
1
2
0
λǫ −λǫ
1
2
Observing that, by definition, (λǫ1 )2 = (λǫ2 )2 , and recalling inequality (3.11), to
get the desired norm estimate, it is enough to consider the function |a(t)−(λǫ2 )2 |.
14
M. RUZHANSKY AND C. TARANTO
A straightforward calculation, using inequality (3.12), shows that
p
p
a(t) + λǫ2 | ≤
|a(t) − (λǫ2 )2 | = | a(t) − λǫ2
Z p
p
α
≤ c2 ǫ
a(t) +
a(t − s)ϕǫ (s)ds =
Z p
p
√
α
a(t) + a(t − sǫ) ϕ(s)ds ≤ 2c2 k akL∞ ǫα .
= c2 ǫ +
It follows that
kH −1AH − (H −1 AH)∗ k ≤ k3 ǫα .
Going back to (3.14), combining it with estimates I), II) and III), we get an estimate
for the derivative of the energy, that is
1
d
|W (t)|2 ≤ 2ρ′ (t)β s + 2k1 ǫα−1 + 2k2 ǫα−1 + k3 βǫα |W (t)|2.
(3.16)
dt
At this point we choose ǫ = β1 , observing that we can always consider β large enough,
say β > 1, in order to have a small ǫ ∈ (0, 1]. Indeed, for β ≤ β0 for some fixed
β0 > 0, a modification of the argument below gives estimate (3.2) with constants
depending only on β0 and T . So we may assume that β > β0 for β0 to be specified.
We define also ρ(t) := ρ(0) − Kt for some K > 0 to be specified. Substituting this in
(3.16) we get for a suitable constant k > 0 that
1 1
1
d
1−α
2
′
s
|W (t)|2 = − 2K + 2kβ 1−α− s β s |W (t)|2.
|W (t)| ≤ 2ρ (t)β + 2kβ
dt
If we have
α
1
>1−α
⇐⇒
s<1+
,
s
1−α
and then also
1
1−α− 1s
K := kβ0
(3.17)
≥ kβ 1−α− s ,
then for all t ∈ [0, T ] we have
d
|W (t)|2 ≤ 0.
dt
This monotonicity of the energy W (t) yields the following boundedness for the solution vector V (t):
1
−1
|V (t)| = e−ρ(t)β s det H(t) kH(t)k|W (t)| ≤
1 det H(0) kH(t)k
1
−1
|V (0)|. (3.18)
≤ e−ρ(t)β s det H(t) kH(t)k|W (0)| = eKtβ s
det H(t) kH(0)k
−1
Note that, according to property (3.11), the function det H (t) is bounded. Furthermore, the behaviour of the convolution and the definition of H(t) guarantee the
existence of suitable constants c, c′ > 0 such that kH(t)k ≤ c and kH −1 (0)k ≤ c′ .
Therefore, there exists a constant C > 0 such that
1
|V (t)| ≤ CeKtβ s |V (0)|,
that means, by definition of V (t), that
1
β 2 |v(t)|2 + |vt (t)| ≤ CeKtβ s β 2 |v0 |2 + |v1 |2 ,
TIME-DEPENDENT WAVE EQUATIONS ON GRADED GROUPS
15
proving the statement of Case 2.
3.3. Case 3: a ∈ C l ([0, T]), with l ≥ 2, a(t) ≥ 0. In this case we extend the technique developed for Case 1. First we perturb the symmetriser of the matrix A(t).
This is done considering the so-called quasi-symmetriser of A(t), the idea introduced
for such problems by D’Ancona and Spagnolo in [15].
Consider the quasi-symmetriser of A(t), that is, a family of coercive, Hermitian
matrices of the form
2a(t) 0
(2)
2 1 0
2 1 0
=
Qǫ (t) := S(t) + 2ǫ
,
+ 2ǫ
0 0
0 0
0
2
(2)
(2)
for all ǫ ∈ (0, 1], and such that Qǫ A − A∗ Qǫ goes to zero as ǫ goes to zero. The
associated perturbed energy is given by
Eǫ (t) := Q(2)
V
(t),
V
(t)
.
ǫ
We proceed estimating the energy, calculating its derivatives in time, so that
d
d
(2)
Eǫ (t) =
Q(2)
(t)V
(t),
V
(t)
+ (Q(2)
ǫ Vt (t), V (t)) + (Qǫ V (t), Vt (t)) =
dt
dt ǫ
d
(2)
∗
(2)
Q(2)
(t)V
(t),
V
(t)
+
iβ
Q
A(t)
−
A
(t)Q
V
(t),
V
(t)
.
(3.19)
=
ǫ
ǫ
dt ǫ
To estimate the second term in the right hand side, we set
iβv(t)
v
V (t) =
=: 1 .
∂t v
v2
Algebraic calculations give
Q(2)
ǫ (t)A(t)
∗
−A
(t)Q(2)
ǫ (t)
2
= 2ǫ
0 1
,
−1 0
therefore
Z
Z
(2)
∗
(2)
2
2Im(v2 v1 )dt ≤ 2ǫ 2|ǫv1 ||v2 |dt ≤
i Qǫ (t)A(t) − A (t)Qǫ (t) V (t), V (t) ≤ 2ǫ
Z
Z
≤ 2ǫ (ǫ2 |v1 |2 + |v2 |2 )dt ≤ 2ǫ
ǫ2 + a(t) |v1 |2 + |v2 |2 dt = 2ǫ Q(2)
ǫ (t)V (t), V (t) .
Using this estimate in (3.19), we get
d
d
(2)
∗
(2)
Eǫ (t) =
Q(2)
(t)V
(t),
V
(t)
+
iβ
Q
A(t)
−
A
(t)Q
V
(t),
V
(t)
≤
ǫ
ǫ
dt
dt ǫ
d
(2)
Q (t)V (t), V (t) + 2βǫEǫ (t)
≤
dt ǫ
#
" d (2)
Q
(t)V
(t),
V
(t)
ǫ
dt
(3.20)
=
+ 2βǫ Eǫ (t).
(2)
Qǫ (t)V (t), V (t)
In order to apply the Gronwall lemma, we first estimate the integral
Z T d (2)
Q (t)V (t), V (t)
dt ǫ
dt.
(2)
Qǫ (t)V (t), V (t)
0
(3.21)
16
M. RUZHANSKY AND C. TARANTO
Let us recall that from the definition of the quasi-symmetriser, it follows that
Z
(2)
Qǫ V, V = 2
a(t) + ǫ2 β 2 |v|2 + |∂t v|2 dt.
(3.22)
Thus, setting c1 := max 1, 2(kakL∞ + ǫ2 ) , we obtain a bound from above for (3.22),
that is
Q(2)
V,
V
≤ c1 |V |2 .
ǫ
2 −1
Observing that ǫ2 c−1
1 ≤ 1 and ǫ c1 ≤ c1 for small enough ǫ, we can also deduce an
inequality from below of the form
2 −1
2
(2)
ǫ c1 |V | ≤ Qǫ V, V .
Hence, there exists a constant c1 ≥ 1 such that for t ∈ [0, T ] we have
2
2
(2)
2
c−1
1 ǫ |V (t)| ≤ Qǫ (t)V (t), V (t) ≤ c1 |V (t)| .
(3.23)
The lower bound, together with [26, Lemma 2] (see [33, Lemma 2] for a detailed
proof), allows us to estimate the integral (3.21) as follows
Z T
Z T d (2)
(2)
d
Q (t)V (t), V (t)
Q (t)V (t), V (t)
dt ǫ
dt ǫ
dt ≤
1 dt ≤
1− 1l (2)
(2)
(2)
Qǫ (t)V (t), V (t)
0
0
(Qǫ (t)V (t), V (t) l
Qǫ (t)V (t), V (t)
Z T
(2)
d
1
1
1
Q (t)V (t), V (t)
2
2
(2) l
− 2l
dt ǫ
l −l
l −l
c
kQ
k
.
ǫ
dt
≤
c
≤ c1 ǫ
T
l ([0,T ]) ≤ c3 ǫ
1
1
ǫ
C
2
1− l
(2)
0
|V (t)| l
Qǫ (t)V (t), V (t)
Thus, by the Gronwall lemma and the estimates for the quasi-symmetriser just derived, we obtain
Eǫ (t) ≤ Eǫ (0)ec3 ǫ
−2
l +2βǫT
.
Combining the latter inequality with (3.23) we obtain
2
2
cT (ǫ
c−1
1 ǫ |V (t)| ≤ Eǫ (t) ≤ Eǫ (0)e
−2
l +βǫ)
≤ c1 |V (0)|2 ecT (ǫ
2
l
2
−2
l +βǫ)
.
We choose ǫ such that ǫ− l = βǫ, thus ǫ = β − 2+l and ǫβ = β 2+l . We can assume β is
large enough with a remark for small β similar to Case 2. Setting σ = 1 + 2l , for a
suitable constant K ∈ C it follows that
l
1
|V (t)|2 ≤ Cβ σ eKβ σ |V (0)|2 .
This means that
as required.
1
l
β 2 |v(t)|2 + |v ′ (t)|2 ≤ Cβ σ eKβ σ β 2 |v0 |2 + |v1 |2 ,
TIME-DEPENDENT WAVE EQUATIONS ON GRADED GROUPS
17
3.4. Case 4: a ∈ C α ([0, T]), with 0 < α < 2, a(t) ≥ 0. In this last case we extend
the proof ofp
Case 2. However, under these assumptions the roots of the matrix A(t),
that are, ± a(t) might coincide, and hence they are Hölder of order α2 instead of
α. In order to adapt this proof to the one for Case 2 we
without loss of
√ will assume
2α
α
generality that a ∈ C ([0, T ]) with 0 < α < 1, so that a ∈ C ([0, T ]).
Following the argument developed for Case 2, we look again for solutions of the
form
1
−1
V (t) = e−ρ(t)β s det H(t) H(t)W (t),
with the real-valued function ρ(t), the exponent s and the energy W (t) to be chosen
later, while H(t) is the matrix given by
1
1
H(t) =
,
λǫ1,α (t) λǫ2,α (t)
where the regularised eigenvalues of A(t), λǫ1,α (t) and λǫ2,α (t) differ from the ones
defined in the previous case in the following way
√
λǫ1,α (t) := (− a ∗ ϕǫ )(t) + ǫα ,
√
λǫ2,α (t) := (+ a ∗ ϕǫ )(t) + 2ǫα .
Arguing as in Case 2, we can easily see that the smooth functions λǫ1 (t) and λǫ2 (t)
satisfy uniformly in t and ǫ the following inequalities
• det H(t) = λǫ2,α (t) − λǫ1,α (t) ≥ c1 ǫα ;
p
• |λǫ1,α (t) + a(t)| ≤ c2 ǫα ;
p
• |λǫ2,α (t) − a(t)| ≤ c3 ǫα .
We now look for the energy estimates. In order to do this, recalling the calculations
done before (3.13) and (3.14), we obtain
′
1
d
2
W (t) = 2Re Wt (t), W (t) ≤ 2ρ (t)β s + 2kH −1(t)Ht (t)k+
dt
−1
(3.24)
det Ht (t) + βkH −1AH − (H −1 AH)∗ k |W (t)|2.
+ 2 det H(t)
The same arguments as in Case 2 allow us to get the following bounds
I) kH −1(t)Ht (t)k ≤ k1 ǫ−1 ;
−1
det Ht (t) ≤ k2 ǫ−1 ;
II) det H(t)
III) kH −1AH − (H −1AH)∗ k ≤ k3 ǫα .
Combining (3.24) with I), II) and III) we obtain
1
d
|W (t)|2 ≤ 2ρ′ (t)β s + 2k1 ǫ−1 + 2k2 ǫ−1 + k3 βǫα |W (t)|2 .
dt
1
We choose ǫ−1 = βǫα which yields ǫ = β − α+1 . Thus, setting γ :=
a constant c > 0 the estimate
1
d
|W (t)|2 ≤ 2ρ′ (t)β s + 2cβ γ |W (t)|2.
dt
1
,
α+1
we obtain for
18
M. RUZHANSKY AND C. TARANTO
We take ρ(t) := ρ(0) − Kt with K > 0 to be chosen later. Considering
1
>γ
⇐⇒
s < 1 + α,
s
we get
1
1
d
|W (t)|2 ≤ (−2K + 2cβ γ− s )β s |W (t)|2 ≤ 0,
dt
provided that β is large enough. Similarly to Case 2, we then get
1
−1
−1
|V (t)| ≤ eKtβ s det H(t)
det H(0)kH(t)k kH(0)k |V (0)|.
(3.25)
Since
−1
−1
α
det H(t)
det H(0)kH(t)k kH(0)k
≤ cǫ−α = cβ α+1 ,
the inequality (3.25) becomes
α
which means
1
|V (t)| ≤ cβ α+1 eKtβ s |V (0)|,
1
α
β 2 |v(t)|2 + |v ′ (t)|2 ≤ cβ α+1 eKtβ s β 2 |v0 |2 + |v1 |2 .
Combining this with a remark for small β similar to Case 2 yields the result. Thus
Proposition 3.1 is proved.
4. Proof of Theorem 1.1
In this section we combine the things from Section 2 and Section 3 to prove Theorem
1.1. However, we will need one more ingredient, the Fourier transform on G, that we
now briefly describe.
b By a usual abuse of notation we will identify
Let f ∈ L1 (G) and let π ∈ G.
irreducible unitary representations with their equivalence classes. The group Fourier
transform of f at π is defined by
Z
b
FG f (π) ≡ f (π) ≡ π(f ) :=
f (x)π(x)∗ dx,
G
with integration against the biinvariant Haar measure on G. This gives a linear
mapping fb(π) : Hπ → Hπ that can be represented by an infinite matrix once we
choose a basis for the Hilbert space Hπ . Consequently, we can write
FG Rf (π) = π(R)fb(π).
By Kirillov’s orbit method (see e.g. [14]), one can explicitly construct the Plancherel
b Therefore we can have the Fourier inversion formula. In
measure µ on the dual G.
addition, the operator π(f ) = fb(π) is Hilbert-Schmidt:
kπ(f )k2HS = Tr π(f )π(f )∗ < ∞,
b ∋ π 7→ kπ(f )k2 in integrable with respect to µ. Furthermore,
and the function G
HS
the Plancherel formula holds:
Z
Z
2
|f (x)| dx =
kπ(f )k2HS dµ(π),
(4.1)
G
see e.g. [14] or [19].
b
G
TIME-DEPENDENT WAVE EQUATIONS ON GRADED GROUPS
19
Proof of Theorem 1.1. Our aim is to reduce the Cauchy problem (1.1) to a form
allowing us to apply Proposition 3.1. In order to do this, we take the group Fourier
b that is,
transform of (1.1) with respect to x ∈ G for all π ∈ G,
∂t2 u
b(t, π) + a(t)π(R)b
u(t, π) = 0.
(4.2)
2
∂t2 u
b(t, π)m,k + a(t)πm
u
b(t, π)m,k = 0,
(4.3)
Keeping in mind the form (2.4) of the infinitesimal representation π(R) the equation
(4.2) can be seen componentwise as an infinite system of equations of the form
b and any m, k ∈ N. The key point of the
where we are considering any π ∈ G,
following argument is to decouple the system given by the matrix equation (4.2). In
order to do this, we fix an arbitrary representation π, and a general entry (m, k) and
b(t, π)m,k
we treat each equation given by (4.3) individually. Note that eventually u
is a function only of t. Formally, recalling the notation used in Proposition 3.1, we
write
2
v(t) := u
b(t, π)m,k , β 2 := πm
,
and
v0 := u
b0(π)m,k , v1 := u
b1 (π)m,k .
Therefore, equation (4.3) becomes
v ′′ (t) + a(t)β 2 v(t) = 0.
We proceed discussing implications of Proposition 3.1 separately in each case.
Case 1: a ∈ Lip([0, T]), a(t) ≥ a0 > 0.
Applying Proposition 3.1, we get that there exists a positive constant C > 0 such
that
β 2 |v(t)|2 + |v ′(t)|2 ≤ C(β 2 |v0 |2 + |v1 |2 ),
which is equivalent to
|πm u
b(t, π)m,k |2 + |b
u′ (t, π)m,k |2 ≤ C |πm u
b0 (π)m,k |2 + |b
u1 (π)m,k |2 .
(4.4)
b and m, k ∈ N. We multiply the inequality (4.4) by
This holds uniformly in π ∈ G
4s/ν
πm yielding
2s
2s
1+ 2s
1+ 2s
b(t, π)m,k |2 + |πmν u
b′(t, π)m,k |2 ≤ C |πm ν u
b0 (π)m,k |2 + |πmν u
b1 (π)m,k |2 . (4.5)
|πm ν u
Thus, recalling that for any Hilbert-Schmidt operator A we have
X
kAk2HS =
|(Aϕm , ϕk )|2
m,k
for any orthonormal basis {ϕ1 , ϕ2 , . . . }, we can consider the infinite sum over m, k of
the inequalities provided by (4.5), to get
1
s
s
kπ(R) 2 + ν u
b(t, π)k2HS + kπ(R) ν ∂t u
b(t, π)k2HS ≤
≤ C kπ(R)
1
+s
2 ν
s
ν
(4.6)
u
b0 (π)k2HS + kπ(R) u
b1 (π)k2HS .
b so
We can now integrate both sides of (4.6) against the Plancherel measure µ on G,
that the Plancherel identity yields estimate (1.15).
Case 2: a ∈ C α ([0, T]), with 0 < α < 1, a(t) ≥ a0 > 0.
20
M. RUZHANSKY AND C. TARANTO
The application of Proposition 3.1 implies the existence of two positive constants
b we have
C, K > 0 such that for all m, k ∈ N and for every representation π ∈ G
1
s
b0 (π)m,k |2 + |b
u1(π)m,k |2 ),
|πm u
b(t, π)m,k |2 + |b
u′(t, π)m,k |2 ≤ CeKtπm (|πm u
(4.7)
where
α
.
1−α
s
s
If the Cauchy data (u0 , u1) are in GR
(G) × GR
(G) then there exist two positive
constants A0 and A1 such that
s<1+
1
1
keA0 R 2s u0 kL2 < ∞ and keA1 R 2s u1 kL2 < ∞.
We note that we can restrict to consider πm large enough since the cut-off to bounded
πm produces functions in any Gevrey spaces. Indeed, if a cut-off χ : R → C has a
compact support, then by the same energy estimate, since χ(R) and R commute,
the problem is reduced to the solution χ(R)u(t, x) to the Cauchy problem
2
∂t (χ(R)u(t, x)) + a(t)R(χ(R)u(t, x)) = 0, (t, x) ∈ [0, T ] × G,
(4.8)
χ(R)u(0, x) = χ(R)u0 (x), x ∈ G,
∂ (χ(R)u(0, x)) = χ(R)u (x), x ∈ G,
t
1
with data χ(R)u0 and χ(R)u1 , so it is in any Gevrey class.
Take now A = min{A0 , A1 }, then we can always assume K in Case 2 of Proposition
3.1 is small enough, so that we have some B > 0 such that KT = A − B. Therefore,
we can rewrite inequality (4.7) as
1
1
2s
s
b(t, π)m,k |2 + |b
u′(t, π)m,k |2 ≤ CeAπm (|πm u
b0 (t)m,k |2 + |b
u1 (t)m,k |2 ). (4.9)
eBπm |πm u
b and applying
Summing over m, k, integrating against the Plancherel measure of G
the Plancherel identity, inequality (4.9) becomes
1
1
1
1
1
keBR 2s ukL2 (G) + keBR 2s ∂t ukL2 (G) ≤ keBR 2s R 2 ukL2 (G) + keBR 2s ∂t ukL2 (G) ≤
1
1
1
≤ keBR 2s R 2 u0 kL2 (G) + keBR 2s u1 kL2 (G) .
(4.10)
1
s
s
(G). Therefore, from (4.10)
If a function f belongs to GR
(G), then also R 2 f is in GR
we get the desired well-posedness result.
Case 3 : a ∈ C l ([0, T]), with l ≥ 2, a(t) ≥ 0.
Similarly to the previous cases, the application of Proposition 3.1 yields the existence of two positive constants C, K > 0 such that
l
+2
1
σ
l
1
σ
u0 (π)m,k |2 + Cπmσ eKT πm |b
u1 (π)m,k |2 ≤
|πm u
b(t, π)m,k (t)|2 + |b
u′(t, π)m,k |2 ≤ Cπmσ eKT πm |b
1
s
K ′ πm
≤ Ce
1
2
s
K ′ πm
|b
u0(π)m.k | + Ce
|b
u1 (π)m,k |2 ,
with 1 ≤ s < σ = 1 + 2l , for some K ′ > 0 small enough. Proceeding as in Case 2, we
obtain the desired inequality.
Case 4: a ∈ C α ([0, T]), with 0 < α < 2, a(t) ≥ 0.
TIME-DEPENDENT WAVE EQUATIONS ON GRADED GROUPS
21
In this last case, applying Proposition 3.1 we have that there exist two positive
constants C, K > 0 such that
1
1
s
2
2
|b
u0 (π)m,k |2 + |b
u1 (π)m,k |2 ),
πm
|b
u(t.π)m.k |2 + |b
u′(t, π)m,k |2 ≤ Cπmα+1 eKT πm (πm
with 1 ≤ s < α + 1. Arguing as above, the result follows.
References
[1] R. Beals. Opérateurs invariants hypoelliptiques sur un groupe de Lie nilpotent. Séminaire
Goulaouic-Schwartz 1976/1977: Équations aux dérivées partielles et analyse fonctionnelle, Exp.
No. 19:8pp, 1977.
[2] J.-M. Bony, P. Schapira. Existence et prolongement des solutions holomorphes des équations aux
dérivées partielles. Invent. Math., 17:95–105, 1972.
[3] M. D. Bronšteı̆n. The Cauchy problem for hyperbolic operators with characteristics of variable
multiplicity. Trudy Moskov. Mat. Obshch., 41:83–99, 1980.
[4] M. Cicognani, F. Hirosawa, M. Reissig. The Log-effect for p-evolution type models. J. Math.
Soc. Japan, 60:819–863, 2008.
[5] M. Cicognani, F. Hirosawa, M. Reissig. Loss of regularity for p-evolution type models. J. Math.
Anal. Appl., 347:35–58, 2008.
[6] M. Cicognani, F. Colombini. The Cauchy problem for p-evolution equations, Trans. Amer. Math.
Soc., 362:4853–4869, 2010.
[7] F. Colombini, E. De Giorgi, S. Spagnolo. Sur les équations hyperboliques avec des coefficients
qui ne dépendent que du temps, Ann. Scuola Norm. Sup. Pisa Cl. Sci., 6:511–559, 1979.
[8] F. Colombini, E. Jannelli, S. Spagnolo. Nonuniqueness in hyperbolic Cauchy problems. Ann. of
Math. (2), 126(3):495–524, 1987.
[9] F. Colombini, T. Kinoshita. On the Gevrey well-posedness of the Cauchy problem for weakly
hyperbolic equations of higher order, J. Differential Equations, 186:394–419, 2002.
[10] F. Colombini, N. Lerner. Hyperbolic operators with non-Lipschitz coefficients. Duke Math. J.,
77:657–698, 1995.
[11] F. Colombini, G. Métivier. The Cauchy problem for wave equations with non Lipschitz coefficients; application to continuation of solutions of some nonlinear wave equations. Ann. Sci. Éc.
Norm. Supér., (4) 41:177–220, 2008.
[12] F. Colombini, G. Métivier. Counterexamples to the well posedness of the Cauchy problem for
hyperbolic systems. Anal. PDE, 8:499–511, 2015.
[13] F. Colombini, S. Spagnolo. An example of a weakly hyperbolic Cauchy problem not well posed
in C ∞ , Acta Math., 148:243–253, 1982.
[14] L. J. Corwin, F. P. Greenleaf. Representations of nilpotent Lie groups and their applications,
Cambridge Studies in Advanced Mathematics, Cambridge University Press, Cambridge, 18
(1990). Basic theory and examples.
[15] P. D’Ancona, S. Spagnolo. Quasi-symmetrization of hyperbolic systems and propagation of the
analytic regularity, Bollettino della Unione Matematica Italiana, 1 B:169–186, 1998.
[16] D. Dasgupta, M. Ruzhansky. Gevrey functions and ultradistributions on compact Lie groups
and homogeneous spaces, Bull. Sci. Math., 756–782, 2014.
[17] D. Dasgupta, M. Ruzhansky. Eigenfunction expansions of ultradifferentiable functions and ultradistributions, Trans. Amer. Math. Soc., 368:8481–8498, 2016.
[18] V. Fischer, M. Ruzhansky. Sobolev spaces on graded groups. to appear in Ann. Inst. Fourier,
https://arxiv.org/abs/1311.0192, 2013.
[19] V. Fischer, M. Ruzhansky. Quantization on nilpotent Lie groups, Progress in Mathematics, vol.
314, Birkhäuser, 2016. (open access book)
[20] V. Fischer, M. Ruzhansky, C. Taranto. On the sub-Laplacian Gevrey spaces, preprint.
[21] G. B. Folland. Subelliptic estimates and function spaces on nilpotent Lie groups. Ark. Mat.,
13(2):161–207, 1975.
22
M. RUZHANSKY AND C. TARANTO
[22] G. B. Folland. Harmonic Analysis in Phase Space, Princeton, Princeton University Press, 2nd
ed., 1989.
[23] G. B. Folland, E. M. Stein. Estimates for the ∂b complex and analysis on the Heisenberg group.
Comm. Pure Appl. Math., 27:429–522, 1974.
[24] G. B. Folland, E. M. Stein. Hardy spaces on homogeneous groups, volume 28 of Mathematical
Notes. Princeton University Press, Princeton, N.J.; University of Tokyo Press, Tokyo, 1982.
[25] C. Garetto, M. Ruzhansky. On the well-posedness of weakly hyperbolic equations with timedependent coefficients, J. Differential Equations, 253:1317–1340, 2012.
[26] C. Garetto, M. Ruzhansky. Weakly hyperbolic equations with non-analytic coefficients and
lower order terms, Math. Ann., 357:401–440, 2013.
[27] C. Garetto, M. Ruzhansky. Wave equation for sum of squares on compact Lie groups, J. Differential Equations, 258:4324–4347, 2015.
[28] C. Garetto, M. Ruzhansky. Hyperbolic second order equations with non-regular time dependent
coefficients. Arch. Rational Mech. Anal., 217:113–154, 2015.
[29] B. Helffer, J. Nourrigat. Caracterisation des opérateurs hypoelliptiques homogènes invariants à
gauche sur un groupe de Lie nilpotent gradué. Comm. Partial Differential Equations, 4(8):899–
958, 1979.
[30] S. Helgason. Wave equations on homogeneous spaces. In Lie group representations, III (College
Park, Md., 1982/1983), volume 1077 of Lecture Notes in Math., pages 254–287. Springer, Berlin,
1984.
[31] L. Hörmander. Hypoelliptic second order differential equations, Acta Math., 119:147–171, 1967.
[32] A. Hulanicki, J. W. Jenkins, J. Ludwig. Minimum eigenvalues for positive, Rockland operators,
Proc. Amer. Math. Soc., 94:718–720, 1985.
[33] T. Kinoshita, S. Spagnolo. Hyperbolic equations with non-analytic coefficients, Math. Ann.,
336:551–569, 2006.
[34] R. Melrose. Propagation for the wave group of a positive subelliptic second-order differential
operator. In Hyperbolic equations and related topics (Katata/Kyoto, 1984), pages 181–192. Academic Press, Boston, MA, 1986.
[35] K. G. Miller. Parametrices for hypoelliptic operators on step two nilpotent Lie groups. Comm.
Partial Differential Equations, 5(11):1153–1184, 1980.
[36] D. Müller, E. M. Stein. Lp -estimates for the wave equation on the Heisenberg group. Rev. Mat.
Iberoamericana, 15(2):297–334, 1999.
[37] A. I. Nachman. The wave equation on the Heisenberg group. Comm. Partial Differential Equations, 7(6):675–714, 1982.
[38] T. Nishitani. Sur les équations hyperboliques à coefficients höldériens en t et de classe de Gevrey
en x. Bull. Sci. Math. (2), 107(2):113–138, 1983.
[39] M. Reed, B. Simon. Methods of Modern Mathematical Physics, Vol. 1: Functional Analysis,
revised and enlarged edition, Academic Press, 1980.
[40] C. Rockland. Hypoellipticity on the Heisenberg group-representation-theoretic criteria. Trans.
Amer. Math. Soc., 240:1–52, 1978.
[41] L. P. Rothschild, E. M. Stein. Hypoelliptic differential operators and nilpotent groups. Acta
Math., 137(3-4):247–320, 1976.
[42] M. Ruzhansky, N. Tokmagambetov. Wave equation for operators with discrete spectrum and irregular propagation speed. to appear in Arch. Rational Mech. Anal.
https://arxiv.org/abs/1705.01418
[43] M. Ruzhansky, V. Turunen. Pseudo-Differential Operators and Symmetries: Background Analysis and Advanced Topics, Basel, Birkhäuser, 2009.
[44] M. E. Taylor. Noncommutative harmonic analysis, volume 22 of Mathematical Surveys and
Monographs, American Mathematical Society, 1986.
[45] A. F. M. ter Elst, D. W. Robinson. Spectral estimates for positive Rockland operators. In
Algebraic groups and Lie groups, volume 9 of Austral. Math. Soc. Lect. Ser., pages 195–213.
Cambridge Univ. Press, Cambridge, 1997.
[46] S. Thangavelu. Harmonic analysis on the Heisenberg group, Boston, Birkhäuser, 1998.
TIME-DEPENDENT WAVE EQUATIONS ON GRADED GROUPS
Michael Ruzhansky:
Department of Mathematics
Imperial College London
180 Queen’s Gate, London SW7 2AZ
United Kingdom
E-mail address [email protected]
Chiara Taranto:
Department of Mathematics
Imperial College London
180 Queen’s Gate, London SW7 2AZ
United Kingdom
E-mail address [email protected]
23
| 4math.GR
|
1
Green Data Centers: A Survey, Perspectives, and Future
Directions
Xibo Jin1,2 , Fa Zhang1 , Athanasios V. Vasilakos3 and Zhiyong Liu1
of Computing Technology, Chinese Academy of Sciences
2 University of Chinese Academy of Sciences, Beijing, China
3 University of Western Macedonia, Greece
1,2
Email: {jinxibo, zhangfa, zyliu}@ict.ac.cn 3 [email protected]
arXiv:1608.00687v1 [cs.DC] 2 Aug 2016
1 Institute
Abstract—At present, a major concern regarding data centers
is their extremely high energy consumption and carbon dioxide
emissions. However, because of the over-provisioning of resources,
the utilization of existing data centers is, in fact, remarkably low,
leading to considerable energy waste. Therefore, over the past
few years, many research efforts have been devoted to increasing
efficiency for the construction of green data centers. The goal of
these efforts is to efficiently utilize available resources and to reduce
energy consumption and thermal cooling costs. In this paper, we
provide a survey of the state-of-the-art research on green data center
techniques, including energy efficiency, resource management, thermal control and green metrics. Additionally, we present a detailed
comparison of the reviewed proposals. We further discuss the key
challenges for future research and highlight some future research
issues for addressing the problem of building green data centers.
Keywords—Green data center; Energy efficiency; Resource management; Thermal control; Green metrics, monitoring and experimental
techniques
I.
I NTRODUCTION
As hosts of large-scale service applications, data centers now
play a crucial role in modern Information Technology (IT)
infrastructures. With the development of cloud computing,
data centers are growing exponentially in number and size. A
new study from International Data Corporation (IDC) predicts
that the total number of data centers deployed worldwide will
peak at 8.6 million in 2017 [1]. Consequently, the extremely
high energy consumption and carbon dioxide emissions of
data centers are becoming a major concern worldwide. The
Natural Resources Defense Council (NRDC) reported that 91
billion KWh of electricity was consumed by U.S. data centers in
2013. Continuing this trend, NRDC estimates that the electricity
usage will reach 140 billion KWh by 2020. As a result, data
center electricity consumption will cost American businesses
$13 billion annually in electricity bills and result in the emission
of nearly 100 million metric tons of carbon pollution per year
[2]. Globally, it had revealed that power requirements grew by
63% to 38 Gigawatts in 2012, up from 24 Gigawatts in 2011
[3]. Moreover, data centers will consume 8% of the worldwide
electricity supply by 2020 while this fraction was about 1.3%
in 2010 [4]. Therefore, making data centers “green” can reduce
both costs for energy consumption and carbon dioxide emissions.
Existing data centers often operate at low utilization because
of over-provisioning and fragmentation of resources [5], [6],
leading to considerable energy waste. According to a McKinsey
study in 2008 [7], the typical utilization ratio is approximately
6%. A Gartner report from 2012 [8] found the typical utilization rates to be in the 7% to 12% range, slightly better than
the result of the 2008 McKinsey study. Recently, Google Inc.
reported that they can improve the utilization of their servers
to relatively high rates of 20-40% [9]. Many studies [10], [11],
[12] have also shown that data center networks experience
high underutilization, with typical utilization ranging between
5% and 25%. What makes the matter worse is that the low
utilization of these servers and network resources also cause
the waste of other supporting infrastructure, such as power
distribution and cooling. From these data, we can see that the
present energy efficiency of data centers is extremely low and
that there is great potential for reducing energy consumption
in data centers.
Motivated by the high energy consumption and low utilization of data centers, many research efforts over the past few
years have focused on the design of green data center infrastructures and services. Generally, these approaches can be categorized into two major classes [13]: 1) those that adopt “green”
equipment in the preliminary design and building phase of the
data centers, and 2) those that increase the efficiency of the
daily expenditures incurred during the operation of the data
centers. In this paper, we focus on the latter class of methods,
i.e., those that emphasize managing the procedures for running
and operating data centers in a “green” manner. To this end,
research studies on the development of green data centers
have concentrated on the following aspects: 1) decreasing the
power consumption of data center resources, 2) increasing the
utilization of data centers, 3) controlling the thermal behavior of
data centers, and 4) developing green metrics, monitoring and
experimental techniques. From the perspective of data center
architecture, the fundamental components of data centers are
1) computing servers, 2) connection networks, and 3) cooling
equipment. Correspondingly, green data center technologies
can be applied to a individual component or to a hybrid scheme
of components.
Although obvious progress has been made in the construction of green data centers over the past decade, there is still
a large gap, and therefore a large opportunity for savings,
between the current average and the characteristics of a bestpractice green data center [14]. In this paper, we present a
survey of the current state-of-the-art research on green data
centers, as summarized in Fig. 1. Some recent surveys have
conducted on energy-aware resource allocation for cloud computing [15] or green data center networks [16]. Different from
them, our survey covers many aspects of the development of
green data centers. Our main contributions are two-fold: 1)
we discuss the key insights underlying recent strategies and
compare the relevant proposals, and 2) we note the future
research challenges and directions for various aspects of the
design of green data centers.
The remainder of this paper is organized as follows. First,
we discuss the studies on and challenges of decreasing power
consumption in Section II. We introduce efforts toward and
opportunities for the efficient utilization of data center re-
2
Green Data Center
Energy
Efficiency
Resource
Management
Virtual Machine
Assignment
Dynamic
Speed Scaling
Network Traffic
Engineering
Power-Down
Mechanism
Power Distribution
Hybrid Technology
Green
Metrics
Thermal
Control
Cooling and
Workload
Distribution
Green Metrics
TemperatureReliability Trade-off
Green Monitoring
and Experimental
Techniques
Renewable
Energy Access
Fig. 1: An overview of green data center techniques
sources in Section III. We explore studies and potential cooling
issues related to data center temperature and thermal control in
Section IV, and we describe the development of and guidelines
for green metrics, monitoring and experimental techniques in
Section V. Finally, we conclude the paper in Section VI.
II.
E NERGY E FFICIENCY
The direct approach to the development of green data centers
is to curb their energy use by employing energy efficiency
techniques. The fundamental concept is to exploit power management technology in the Information and Communication
Technology (ICT) devices of data centers. In this regard, the
research community has proposed a number of possible approaches in recent years.
A. Dynamic Speed Scaling (SS)
This scheme, which is also known as Dynamic Voltage
and Frequency Scaling (DVFS), focuses on lowering the frequency/speed of devices to save energy. Dynamic speed scaling allows for power savings as power consumption is approximately proportional to the supply voltage or the device
speed s (e.g., following the cube-root rule, s3 , or a more general
function f (s, α), where α is a constant power parameter) [17],
[18], [19]. The goal is to determine the processing speeds and
job assignment to minimize the total energy cost and guarantee
the prescribed performance constraints. Intensive research, initiated by Yao et al. [17], has been conducted in pursuit of saving
energy by speed scaling. Some efforts address the problem
of scheduling via job deadlines [20] and the optimization of
job flow times [21]. Other approaches consider the problem
of makespan minimization [22], incorporate precedence constraints defined between jobs [23] and account for affinities
between jobs and processors [24]. The research on variablespeed architectures includes single processors [17], [25], parallel
processors [26], [27], and network devices [28], [29], [35] with
preemption or nonpreemption settings [30], [31]. Some studies
also have addressed realistic speed models in which only a finite set of discrete speed levels is available [32], [33] or in which
the speeds are bounded on an interval [34]. In addition, Bampis
et al. [35] studied a heterogeneous multiprocessor preemptive
problem, in which it was assumed that each processor had
a different speed-to-power function. We can present all these
efforts in a consistent fashion by extending the standard threefield notation of [55]; we summarize the algorithmic results
obtained for dynamic speed scaling in Tables I and II, and the
symbols used are defined in Table III.
When a large-scale warehouse is considered, speed scaling is
adopted to seek solutions for optimizing the energy consumption based on the current load [56], [57], [58]. Higher speeds
allow for faster execution but also result in higher power
consumption. Gandhi et al. [56] studied the problem of finding
the optimal power allocating in terms of determining the
optimal frequencies of the servers in a server farm to minimize
mean response time after measuring the power-to-frequency
relationship of each server for a given workload. The authors
noted that use of the optimal power allocation significantly
improved the response time, by a factor of typically 1.4 and,
in some cases, as much as a factor of 5. In [57], the authors
considered how to balance the mean response time and mean
energy consumption in processor sharing scheduling when dynamic speed scaling is applied to reduce energy consumption.
They identified a scheme that provided results that were nearly
identical to the optimal objective of dynamic speed scaling,
i.e., that simultaneously minimized energy consumption and
response time. Speed scaling has also been studied in the
arbitration of power consumption and system throughput. In
[58], the authors provided a probabilistic framework in which
online decisions were made on request for admission control,
routing, and virtual machine (VM) allocation. They modeled a
unified objective function to balance the power consumption of
the servers and the system throughput for application requests.
At the network level, an Adaptive Link Rate (ALR) can be
applied on links to reduce energy consumption [59], [28], [60],
[61]. Gunaratne et al. [28] proposed a Markovian model to
reduce the energy consumption of a typical Ethernet link by
adaptively varying the link data rate in response to utilization.
The authors found that their scheme allows an Ethernet link
to operate at a lower data rate for more than 80% of the time,
yielding approximately $70 million in energy savings per year
TABLE I: Algorithmic results for speed scaling approaches based on the minimization of energy subject to deadline feasibility constraints.
Speed Mode
Environment
Problem
1|r j , d j , pmtn | E
Algorithm Type
Offline
Complexity
Approximation/Competitive Ratio
O ( n3 ) [17], O ( n2 log n ) [20]
-α α α
αα [17], 2( α−
1 ) e [25]
Online
Single processor
Continuous
speed scaling
Multiple
homogeneous
processors
1| agreeable, pmtn | E
1|r j , d j , w j = 1, non-pmtn | E
1| agreeable, non-pmtn | E
1|laminar, non-pmtn | E
1|r j , d j , non-pmtn | E
Offline
P | agreeable, w j = 1, pmtn, non-mig| E
Offline
P | agreeable, pmtn, non-mig| E
P |r j = 0, d j = d, w j = 1, M j , pmtn, non-mig| E
P |r j = 0, d j = d, pmtn, non-mig| E
P |r j , d j , w j = 1, pmtn, non-mig| E
P |r j = 0, d j = 1, w i,j , pmtn, non-mig| E
Single processor
Discrete
speed
scaling
Multiple
homogeneous
processors
Offline
P |r j = 0, d j = d, pmtn, mig| E
P |r j , d j , pmtn, mig| E
P |r j = 0, d j , pmtn, non-mig| E
P |r j , d j , pmtn, non-mig| E
Multiple
heterogeneous
processors
Offline
1
O ( n2 ) [37]
Polynomial [38]
Polynomial [17], [30]
NP-hard [30]
NP-hard
O ( n log n ) [40]
O ( n2 f ( n )) [41]
O ( n f ( n ) log K ) [42]
O ( n log n ) [26]
Online
Offline
Online
Offline
Online
Offline
O ( m f ( n ) log n ) [24]
NP-hard [24], [26]
NP-hard ( m ≥2) [26]
NP-hard
NP-hard
APX-hard [39]
P |r j , d j , non-pmtn, non-mig| E
Offline
R|r ij , dij , w ij , pmtn, mig| E
R|r ij , dij , w ij , pmtn, non-mig| E
Offline
Polynomial ( 1ǫ )
NP-hard
Offline
O ( dn log n ) [32]
1|r j , d j , pmtn | E
P |r j = 0, d j = d, M j , pmtn, non −mig| E
NP-hard
Offline
(12(1 + ǫ)) α−1 [39]
----α α α
2( α −
1 ) e [26]
αα 24α [26]
-PTAS [24], [26]
αα 24α [26], B⌈α⌉ [27]
αα 24α [26]
2(2 − m1 ) α [26], B⌈α⌉ [27]
B⌈α⌉ [27]
α α α
2( α −
1 ) e B⌈α ⌉ [27]
1
m α ( n m ) α−1 [31]
( 25 ) α−1 Beα ((1 + ǫ)(1 + wwmax )) α [39]
min
OPT + ǫ [35]
(1 + ǫ) α Beα [35]
-2 α − 1 ( α −1 ) α − 1 ( δ α −1 ) α
( δ −1)( δ α− δ ) α −1
Online
NP-hard [33]
1
4α /(2e 2 α 2 ) [36]
---QPTAS [38]
25α−4 [30]
(1 + wwmax ) α [31]
min
eα [35]
2α −1 ( 1 + ǫ ) α B
+ 1 (δ = maxi sis+i 1 ) [43]
(2∆ ) α−1 ( ∆ = maxi ( s i +1 /s i )) [33]
3
TABLE II: Algorithmic results for speed scaling approaches based on flow time, makespan, and throughput.
Speed Mode
Environment
Problem
1|r j , w j = 1, non-pmtn, E | F
1|r j , non-pmtn, E | F
1|r j , w j = 1, non-pmtn | E + F
Single processor
Continuous
speed scaling
Single processor
(speed bounded)
1|r j , w j = 1, pmtn | E + F
1|r j , pmtn | E + W F
1|r j , pmtn, E | F
1|r j , pmtn | E + F
1|r j , pmtn/non-pmtn, E |Cmax
1|r j , d j , pmtn, E |U
1|r j , d j , pmtn, E |WU
1|r j , w j = 1, pmtn | E + F
1|r j , pmtn | E + F
Algorithm
Type
Complexity
O ( n2 log L ) [44]
Offline
O ( n3 log K ) [21]
Online
Online
Offline
O ( n log n ) [22]
9 ) [49]
Pseudo-poly, O ( n6 T 9 Vmax
2
4
9
9 ) [49]
Pseudo-poly, O ( n S T Vmax
Online
Multiple
homogeneous
processors
Multiple
heterogeneous
processors
Discrete
speed scaling
Single processor
P = 2|r j , d j , pmtn, mig|U, E
P |r j , pmtn, mig| E + F
P |r j , pmtn, non-mig| E + F
P |r j , w j = 1, pmtn/non-pmtn, non-mig, E | F
P |r j , pmtn, non-mig| E + F
P |r j , w j = 1, pmtn/non-pmtn, non-mig, E |Cmax
P |r j = 0, non-pmtn, non-mig, E |Cmax
P |r j = 0, non-pmtn, non-mig, prec, E |Cmax
P |r j , d j , w j = 1, m, non-pmtn, non-mig, E |WU
P | agreeable, w j , m, non-pmtn, non-mig, E |WU
R|r j , pmtn, non-mig| E + F
R|r j , Pi ( s ) = s αi , pmtn, non-mig| E + W F
R|r j , Pi ( s ) = s αi , pmtn, non-mig| E + F
R|r ij , dij , w ij , Pi ( s ) = s αi , pmtn, non-mig, E |WU
1|r j , pmtn/non-pmtn, E |Cmax
1|r j , d j , pmtn |U, E
-(1 + ǫ) α -energy O ( ǫ1 )-approximate [44]
-√
8.3e ((3 + 5) /2) α [21]
4 [45]
O ( α2 / ln2 α) [45]
not O (1)-competitive [45]
O ( lnαα ) [46]
3 + ǫ [47]
2 + ǫ [48]
---4 [34]
α −1
2( α + 1) / ( α −
1/ ( α −1) ) [46]
( α +1 )
14-competitive ( αα + α2 4α )-competitive [50]
4-competitive ( αα + α2 4α )-competitive [34]
3-approximate O (1)-approximate [43]
3-competitive O (1)-competitive [51]
O (1) [52]
O (1) [52]
arbitrarily good [22]
(3 + ǫ) B⌈α⌉ [27]
arbitrarily good [22]
1|r j , d j , pmtn, |U, E
Offline
Multiple
homogeneous
processors
(speed bounded)
Approximation/Competitive
Ratio
Online
Offline
Online
NP-hard [22]
Offline
NP-hard
pseudo, O ( n12m +7 S2 ) [53]
O ( n2m +2V 2m +1 Sm ) [53]
O ( 1ǫ )-competitive (1 + ǫ )-speedup [54]
O ( α2 ) ( α = maxi αi ) [54]
O ( α) ( α = maxi αi ) [54]
2( Γ + 1)(1 + ǫ ) ( Γ = maxi αi ) [53]
Online
Offline
Offline
Online
2
O (log1+ α m ) [23]
O ( n log n ) [22]
-14-competitive ( ∆ α ( αα + αα 4α ) + 1)-competitive [50]
4
5
TABLE III: Definitions of symbols for dynamic speed scaling
Symbol
Definition
1
P
R
m
n
d
Single process
Homogeneous parallel (multiple) processors
Heterogeneous parallel (multiple) processors
Number of processors
Number of jobs
Number of available discrete speeds
e
α
ǫ
B⌈ α ⌉
eα
B
A mathematical constant, e ≃2.71828
A constant power parameter
An arbitrary parameter
⌈ α⌉-th Bell number
Generalized version of the Bell number Ω(( e lnα α ) α )
The largest ratio of two consecutive (non-zero)
speed levels
∆
f (n)
Time complexity of computing a maximum flow
in a layered graph with n vertices
K
The range of all possible values of speeds divided
by the desired accuracy
L
The range of all possible values of power divided
by the desired accuracy
T
Vmax
V
S
The
The
The
The
span length max j,j ′ ( d j − r j ′ )
maximum processing volume
sum of processing volumes Σ j w j
sum of weights
rj
dj
wj
w max ( w min )
Job j’s release date
Job j’s deadline
Job j’s processing volume
Maximum (Minimum) processing volume of all
jobs
r ij
Job j’s release date on processor i
Job j’s deadline on processor i
dij
w ij
Job j’s processing volume on processor i
Mj
Job j’s eligible processor set
Energy consumption
E
(W ) F
(Weight) Flow time
Makespan
Cmax
(W )U
(Weight) Throughput
pre
Precedence constraints defined between jobs
( non−) pmtn (Not) Allow jobs to be interrupted
( non−) mig
(Not) Allow jobs to be interrupted and resumed
on the same or another processor
agreeable
laminar
For any two jobs j and j′ , if r j ≤r j ′ , then d j ≤ d j ′
If r j ≤r j ′ , then d j ≥d j ′ or d j ≤r j ′
in the U.S. with only a very small increase in packet delay. Abts
et al. [60] proposed a power-optimizing mechanism for links
with the capabilities of a network with the flattened butterfly
topology. The authors scaled down the network links with a
data rate proportional to the traffic intensity while accepting a
trade-off of additional mean latency. They reported an energy
savings of approximately 42%. Wang et al. [61] presented a
rate-adaptation-based solution with the intent of approaching
network-wide energy proportionality via routing optimization.
The simulation results indicated that the scheme could achieve
up to 40% energy savings while introducing a very slight
increase in network delay.
B. Power-Down Mechanism (PDM)
In this scheme, devices are transitioned into a low-power
standby or sleep mode while they are idle to save energy. One
must determine when idle periods exist that are of sufficient
duration to outweigh the costs incurred by the scheme (such
as the cost for transitioning from the sleep state and the delay
cost) and decide when to wake the devices from the powersaving mode to satisfy the workload demand. When a single
device is considered and the transition delay is ignored, the
problem of managing power in a two-state situation (active
state and sleep state) is equivalent to the ski-rental problem.
A competition analysis of this problem was presented by [62],
[63]. Irani et al. [64] extended the study to devices with multiple states. The authors presented a deterministic algorithm
that achieved 2-competitive performance when the transition
energies were assumed to be additive. In [65], the authors
investigated the generalized case in which the state transition
energies could take
√ arbitrary values and achieved a competitiveness of 3 + 2 2. With preemptions allowed and the process
of waking up assumed to require a certain fixed amount of
energy, Baptiste et al. [67] provided an O(n5 )-time algorithm
for offline operation. For the case in which the tasks have
agreeable deadlines, i.e., later-released tasks also have later
deadlines, [68] and [69] proposed algorithms to improve the
time complexity. Given a single processor, tasks with arbitrary
processing time and unit transition costs, [68] proposed an
O(n log n ) greedy algorithm. Later, Angel et al. [69] proposed
an O(n2 ) algorithm for tasks with arbitrary processing time and
arbitrary transition costs in a single-processor environment.
For tasks with unit processing time and unit transition cost in
the multiprocessor case, the authors of [69] also provided an
O(n2 m) algorithm. In [70], Demaine et al. considered the powersaving problem in which each task must be executed within
a specified set of time intervals and proposed a (1 + 32 L c )approximation algorithm, where L c is the transition cost. In
[71], the authors studied the problem of online dynamic power
management in a three-state situation (busy state, standby state
and off state). They first considered the case in which the set of
tasks that could feasibly be scheduled on a single processor was
known in advance (packable) and, for this case, proposed a 4competitive online algorithm that used at most two processors.
The competitive ratio was improved to 3.59 for tasks with
unit processing time. For the general case of tasks with unit
processing time and multiple processors, an O(1)-competitive
algorithm was provided. In Table IV, we summarize all the
algorithmic results mentioned above regarding power-down
mechanisms; the definitions of the symbols are the same as
those given in Table III.
For large hosting centers, several studies recommend energyproportional computing [72], [73], [74], [75], [76], [77], [78], [79].
Here, “proportional” refers to the concept that the size of the
data center should be proportional to its workload demand
[80]. In [72], the authors proposed energy-conscious provisioning to concentrate the request load on a minimal active set
of servers appropriate to the current aggregate load level. The
excess servers would be transitioned into a power-saving state
to reduce the energy cost during periods of light load. These
authors reported that their system could reduce server energy
usage by 29% or more for a typical Web workload. Lin et al.
[73] investigated the energy-saving problem by dynamically
“right-sizing” the target data center by turning off servers
during lighter periods in both offline and online cases. They
6
TABLE IV: Algorithmic results for power-down mechanisms. The definitions of the symbols are the same as those given in Table III, and the
term packable means that the set of tasks that can feasibly be scheduled on a single processor is known in advance.
Environment
Task Model
Algorithm Type
rj , wj
Two states (active state and sleep
state), transition cost Lc is arbitrary
Online
--
rj , wj
Multiple states ({s0 , ..., s k }), transition
cost Lc is additive
Online
--
2 [64]
rj , wj
Multiple states ({s0 , ..., s k }), transition
cost Lc is arbitrary
Online
--
√
3 + 2 2 [65]
rj , dj , wj = 1
Two states (active state and sleep
state), transition cost Lc is arbitrary
Offline
O ( n7 ) [66],
O ( n4 ) [67]
--
r j , d j , w j , pmtn
Two states (active state and sleep
state), transition cost Lc is arbitrary
Offline
O ( n5 ) [67]
--
agreeable, w j ,
pmtn
Two states (active state and sleep
state), transition cost Lc is unit
Offline
O ( n log n) [68]
--
Single processor
Multiple
homogeneous
processors
Multiple
homogeneous
processors,
m=2
Multiple
homogeneous
processors
Complexity
Approximation/Competitive
Ratio
Transition Cost
2 [62], e/ ( e − 1) [63]
3
agreeable, w j = 1
Two states (active state and sleep
state), transition cost Lc is arbitrary
Offline
O ( n ) [68],
O ( n2 ) [69]
--
agreeable, w j ,
pmtn
Two states (active state and sleep
state), transition cost Lc is arbitrary
Offline
O ( n2 ) [69]
--
agreeable, w j = 1
Two states (active state and sleep
state), transition cost Lc is unit
Offline
O ( m2 n3 ) [68],
O ( mn2 ) [69]
--
r j , d j , w j = 1,
r j ∈N, d j ∈N
Two states (active state and sleep
state), transition cost Lc is arbitrary
Offline
O ( m5 n7 ) [70]
--
Execute in a specified
set of time intervals,
wj = 1
Two states (active state and sleep
state), transition cost Lc is arbitrary
Offline
NP-hard [70]
1 + 23 Lc [70]
r j , d j , w j = 1,
pmtn, non − mig,
packable
Three states (busy, standby and off),
transition cost Lc is arbitrary
Online
--
3.59 [71]
r j , d j , w j , pmtn,
non − mig,
packable
Three states (busy, standby and off),
transition cost Lc is arbitrary
Online
--
4 [71]
r j , d j , w j = 1,
pmtn, non − mig
Three states (busy, standby and off),
transition cost Lc is arbitrary
Online
--
modeled the cost of toggling a server back and forth between
the active and power-saving modes as a switching cost. The
authors reported a 50% energy savings for a peak-to-mean ratio
(PMR) of 5. Azar et al. [74] considered turning off machines to
save on computation costs. The authors accounted for the delay
cost as the duration required to turn the machines on or off.
They proposed a bi-objective algorithm to address the problem
of minimizing both the maximum task delay and the total cost.
On the network-device level, power-down mechanisms redirect network traffic to aggregate traffic on only a few network
devices and links while allowing idle devices to sleep for energy conservation [81]. In [75], the authors investigated poweraware network design and routing to dynamically power
on/off line cards and chassis. They modeled the minimization
of network-wide power consumption as a mixed-integer optimization problem and solved it using a branch-and-cut solver
that incorporated several practical considerations (such as the
network structure). They demonstrated that their results were
within 11% of the optimal result. Bolla et al. [76] used energyaware traffic engineering on backbone networks to manage the
standby and wakeup behavior of network devices with respect
to the IP layer. The basic concept driving their work was to
periodically reconfigure the network nodes and links to accommodate incoming traffic volumes and operational constraints
(e.g., quality of service and reconvergence times). The authors
noted that the energy savings depended on the level of traffic
O (1)-competitive [71]
load; specifically, the energy savings was approximately 40%
when the traffic load was less than 50%. Andrews et al. [77]
studied routing and scheduling in a power-down model and
captured the trade-off between energy savings and provisioning performance guarantees such as end-to-end delays. The
authors proposed scheduling algorithms for line and arbitrary
topologies. In the case of periodic scheduling, the schedule determined the active period per element within each frame and
prioritized the packets within each active period. Zhang et al.
[78] proposed GreenTE, an intra-domain energy-aware traffic
engineering mechanism that maximizes the number of links
that can be placed in sleep mode while satisfying performance
constraints (e.g., link utilization and packet delay). The authors
reported an energy savings of 27-42% for a maximum link
utilization below 50%. ElasticTree [79] dynamically adjusts the
set of active network elements to achieve a trade-off between
energy efficiency and performance. The authors used methods
including a formal model, a greedy bin-packer, a topologyaware heuristic and prediction to determine which subset of
links and switches to use. The authors reported a network
energy savings of 50%.
C. Hybrid Technology
Various studies of hybrid technology have also been reported
that have explored both speed scaling and power-down mechanisms to fully exploit the potential for energy savings. In
7
a standard speed scaling setting, there is a tendency to use
low speed levels, subject to prescribed deadline constraints.
By contrast, in combination with the power-down approach,
it can be beneficial to speed up the processing of tasks to
generate idle intervals in which devices can be transitioned
to the power-saving mode. Irani et al. [82] performed the first
theoretical analysis of reducing energy usage by combining
placing the system in a sleep state if it is idle with varying
the speed at which tasks are processed. Assuming that all
tasks could be preempted and resumed at no cost, the authors
proposed a 2-approximation offline algorithm and a constantcompetitive-ratio online algorithm. Several subsequent studies
presented algorithms that improved on these result [83], [84],
[85], [86]. In [83], the authors considered online scheduling of a
processor to manage its energy consumption using both speed
scaling and a sleep state. They proposed an algorithm called
SOA to improve the competitive ratio for energy minimization
from 22α −2 αα + 2α −1 + 2 [82] to max {4, αα + 2}, under the
assumption that the speed could be arbitrarily high. They also
showed that when the maximum speed of the processor was
bounded, their algorithm was 4-competitive for throughput
and (αα + α2 4α + 2)-competitive for energy. In [84], the authors
presented an algorithm called SqOA to further improve the
competitive ratio of SOA [83] in the unbounded-speed setting
to max {4, (2 − 1/α)α 2α −1 + 2}. The authors of [85] designed
superior approximation algorithms for the offline case. They
proposed a 4/3-approximation algorithm for general convex
power and a 137/117-approximation algorithm for power functions of the form P (s) = βsα + γ. Recently, Antoniadis et al. [86]
closed the gap in the approximation factor by presenting a fully
polynomial-time approximation scheme for the deadline-based
preemptive offline scheduling problem. In [87], the authors
extended the study of this strategy to minimize the sum of
energy and flow on a processor with one or multiple levels
of sleep states. They designed O(1)-competitive clairvoyant
and non-clairvoyant algorithms in the unlimited-speed setting
(note: the terms “clairvoyant” and “non-clairvoyant” indicate
that the sizes of the tasks are either known or not known,
respectively, when the tasks arrive). Under a maximum-speed
constraint, they enhanced their algorithm to again achieve
O(1)-competitive performance.
Chen et al. [88] proposed a framework to allow the employment of both approaches to enhance energy savings in
hosting centers. Their solutions were based on queuing models
and control theory and did not require compromise in enduser Service Level Agreements (SLAs). In [89], the authors
considered reducing the power consumed in geographically
distributed data centers. They proposed an approach that incorporated selecting the number of active servers and their service
rates to minimize the time-averaged power cost incurred when
handling delay-tolerant workloads. Liu et al. [90] developed
SleepScale, a power management tool to manage data centers
for power efficiency while fulfilling quality-of-service (QoS)
agreements. They exploited the low-power states and operating
frequency that are built into modern CPUs and proposed a
joint optimization of speed scaling and sleep-state selection to
reduce energy consumption. [91], [92] and [93] presented studies of network elements. Nedevschi et al. [91] explored the use
of hybrid technology to reduce network energy consumption
by putting network components to sleep during idle periods
and adapting the rate of network operation to the requested
workload. They showed that the effect depended on the power
profile of the network equipment and the utilization of the net-
work itself. Using real-world traffic workloads, topologies and
power constants, they reported an energy savings of 50% for
a network utilization of 10-20%, accompanied by a very slight
increase in latency. The developers of EATe [92] formulated a
method of online traffic management to spread the load across
multiple paths while minimizing energy consumption. Their
approach leveraged the rate adaptation and power-down of
links and routers to configure an energy-proportional network
hardware structure. Extensive ns-2 simulations demonstrated
that EATe could place 21% of the links in the sleep state, lower
the link energy levels to achieve 8% energy savings and put
16% of the active routers to sleep. Wang et al. [93] studied
the problem of flow scheduling and routing while meeting
flow deadlines in data center networks. They combined speed
scaling and power-down mechanisms to determine routing
paths and schedules as well as the transmission rate for each
flow.
D. Challenges and Research Issues
The studies discussed above demonstrate that energyefficient mechanisms can assist in making data centers green
and reducing energy usage. Furthermore, they reveal that there
is a trade-off between performance and energy consumption.
The aforementioned efforts also have demonstrated that a considerable amount of the energy consumed as a result of overprovisioning can be saved while still satisfying performance
constraints. However, there are still several other issues to be
addressed concerning the design of energy-efficient algorithms:
• Most studies that have been conducted on energy efficiency thus far have focused on ideal power models.
However, future investigations should consider scenarios
that more closely resemble real systems. For example,
some overhead is incurred when switching speeds because the processor must stop while the voltage is changing, and frequent changes in speed can also harm a processor’s lifetime reliability [94]. An interesting question
is how to incorporate these aspects of real systems into
performance studies.
• In classical parallel scheduling problems, the number
of servers is fixed, and the scheduler primarily decides
which task to process on which processor and determines
the speed of each processor at a given time. However,
data centers operate on a different model in which servers
can be activated and released on demand [95] and in
which it is necessary to exploit an appropriate number
of servers to process the submitted tasks. Thus, the characteristics of data centers pose new challenges and open
problems, of which the key underlying requirements are
scalability, uncertainty and efficiency.
• As mentioned earlier, the latency time and power consumption incurred in the process of rebooting means
that the effects of turning servers or switches off and on
cannot be overlooked. Moreover, these effects are more
serious at the network level because communication data
can be lost after a certain threshold latency time [96].
Fortunately, data center network architectures typically
provide multiple paths for communication among their
servers. The question of how to efficiently manage energy
consumption in data center networks is still a major concern that is worthy of investigation. The well-structured
topology of data center networks can be investigated to
address this issue. A good mechanism design should
8
minimize the effects on the network throughput, latency,
congestion and packet loss while incorporating the structure of the data center network architecture.
• Current results for multiple processors (servers) mostly
focus on environments that consist of homogeneous collections of identical processors (servers). However, many
computer architects believe that architectures based on
heterogeneous processors will dominate future architectural design [97], [98]. The primary advantage of heterogeneous architectures is that the design can include
certain processors that are specialized for particular types
of tasks, with the intent of assigning tasks to those
processors that are best suited for them [99]. It is also
natural to consider the possibility of heterogeneity in data
centers, as the various servers contained in data centers
often differ from each other in their power-performance
relationships. At this time, few results for heterogeneous
environments are extant in energy-efficient algorithm design.
We believe that addressing the above challenges and research
issues will lead to significant advances in the current level of
energy efficiency of data centers.
III.
D ATA C ENTER R ESOURCE M ANAGEMENT
Whereas the previous section discussed schemes to increase
the energy efficiency of computing and communication devices, in this section, we present prominent research works
that exploit virtualization and cloud computing to efficiently
utilize data center resources. The concept of virtualization is
in contrast to the traditional architecture, which leads to low
utilization rates and, in turn, wastes resources because servers
continue to run although they are not processing applications to
their full potential. Through virtualization, the power of servers
can be multiplexed to address many applications, and with
the help of virtualization, more than one application can be
assigned to one server. Thus, this technique enables significant
and cost-effective gains in resource utilization and associated
energy savings. A second innovation that can greatly increase
utilization rates is the development and maturation of cloud
computing. The shift to the cloud encourages economies of
scale because servers can be run on virtual platforms, which
offer higher utilization rates, and these large cloud providers
are highly incentivized to optimize the utilization of data centers and reduce energy consumption to maximize productivity.
In this section, we analyze methods that adopt these two
approaches for virtual machine assignment, network traffic
engineering, power distribution and renewable energy access
with a view to providing better energy management for green
data centers, and we consider the corresponding directions for
future research.
A. Energy-Aware Virtual Machine Assignment
Underutilization in data centers is a major cause of excessive
energy consumption. Moreover, modern servers are sufficiently
powerful to use virtualization to present the illusion of many
smaller virtual machines (VMs), each running independent
applications [100]. Therefore, optimal virtual machine management (including allocation, consolidation, and migration)
is an important technique for data centers that can facilitate
the improvement of resource utilization and the reduction of
energy consumption. The general idea is to efficiently map
VMs to servers based on resource matching and improved
performance metrics. An early study [101] extended virtualization solutions to support rich and effective policies for active
power management, which had not been done before. The
authors integrated “hard” and “soft” power states to achieve
high power savings and showed that substantial benefits could
be derived from the coordination of online methods for server
consolidation using their proposed management techniques. In
[102], Raghavendra et al. proposed a power management solution that coordinated various individual approaches, including
a virtual machine controller. Their design used techniques and
actuators to optimize power at multiple system levels across
both hardware and software. Hermenier et al. [103] proposed
Entropy, a consolidation manager that performed dynamic
consolidation of VMs based on constraint programming and
accounted for migration overhead. The authors showed that
their approach could significantly reduce the number of nodes
and the migration time compared with results obtained using
the FFD heuristic. They reported that the consumption of nodes
per hour could be reduced by 25%.
When exploiting the ability to consolidate several virtual
machines on the same physical server, a key challenge is to explicitly account for the specified QoS in the optimization problem [104], [105]. Van et al. [106] defined the condition of SLA
fulfillment and the resource management cost as a global utility
function. The authors adopted the Constraint Programming
approach to formulate and solve the considered problem. Kusic
et al. [107] considered the problem of consolidating services into
a smaller number of computing resources to achieve higher
server utilization and energy efficiency while maintaining the
desired QoS. The authors implemented a dynamic resource
provisioning framework to treat this problem as a problem of
sequential optimization and solve it using a lookahead control
scheme. They reported 22% power savings compared with a
system without dynamic control. Several other efforts have
also been directed toward the online assignment of virtual
machines to fewer servers. Beloglazov et al. [108] proposed
novel adaptive heuristics for the dynamic consolidation of VMs
based on an analysis of historical data related to VM resource
usage. They applied a modification of the Best Fit Decreasing
algorithm to perform the VM assignment, and they achieved
significant reduction of energy consumption while ensuring a
high level of adherence to the SLA. In [109], the authors considered the problem of consolidating virtual machines when the
network bandwidth demands of the VMs are dynamic. They
converted the problem into a Stochastic Bin Packing problem
and proposed an online packing algorithm.
√ They reported that
the performance ratio was within (1 + ǫ)( 2 + 1) for any ǫ > 0
and that the number of required servers was reduced by 30%
compared with the results of the HARMONIC algorithm [110].
In [111], policies that minimized the total number of occupied
servers in the system were studied. The authors presented
Greedy with sublinear Safety Stocks (GSS) and Modified GSS
(GSS-M) policies. Generally speaking, these GSS and GSS-M
policies scheduled incoming virtual machines in a manner
that greedily minimized the total number of occupied physical
servers. The authors demonstrated the asymptotic optimality
and convergence rate of their policies. However, such beneficial
consolidation also introduces an SLA violation risk. In [112], the
authors presented a risk management framework to address the
problem of consolidating virtual machines for energy efficiency
in data centers. They suggested a two-step approach to virtual
machine scheduling: 1) they solved the problem of resource
allocation for virtual machines using a stochastic program, and
9
2) they proposed an algorithm for dynamic virtual machine
consolidation at run time. Using real-world workloads, they
showed that SLA violations were reduced by a factor of four,
from 25% to 2-5%, compared with other virtual machine consolidation algorithms that do not consider the SLA violation
risk.
Table V summarizes the various approaches to energy-aware
virtual machine assignment and presents a comparison of
the common features identified among the different proposals
mentioned above. The features selected for comparison are as
follows: 1) the resources considered, 2) the techniques used
in the proposals, 3) the traces used for the experiments, 4)
the performance metrics considered, and 5) the results of the
proposals.
B. Data Center Network Traffic Engineering
Traffic engineering is a highly effective approach to making
data centers green based on different traffic patterns and network architectures [113]. Several solutions have been proposed
for achieving energy proportionality by applying traffic aggregation and virtual machine assignment techniques. In [114],
the authors demonstrated that the minimization of network
power consumption for any general data center topology is
NP-complete. They proposed a centralized network power controller program that gathered traffic data and server statistics
from all switches and servers and used this information to
perform traffic aggregation and virtual machine assignment
and migration in the target data center. They demonstrated the
optimal placement for the virtual machines and the bandwidth
savings through trace evaluation. In [115], the authors developed REsPoNse, a framework based on identifying and using
energy-critical paths to redirect traffic to allow large portions of
the network to enter a low-power state. They demonstrated that
energy savings were achieved without frequent recomputation
and deployment of new routing tables, and they showed
that REsPoNse had marginal impact on the application-level
throughput and latency. Their results yielded energy savings
of approximately 30% and 42%. Jiang et al. [116] considered
a joint design incorporating virtual machine assignment and
routing for data centers. They considered the optimization
of both energy cost and network congestion, then proposed
an offline algorithm based on a Markov chain model and
an incremental online solution for a dynamic environment.
The authors reported performance analysis results for various topologies, including clique, fat-tree, HyperX and BCube
topologies, under a spectrum of elephant and mice workloads.
Given that the bandwidth demands of different flows do not
peak at exactly the same time, the authors [117] proposed a
correlation-aware power optimization scheme that dynamically
consolidated traffic flows onto a small set of switches and links
and then shut down unused network devices. They designed a
heuristic algorithm to determine the consolidation and rate configuration, which was fundamentally based on a bin-packing
algorithm. They reported an energy savings of 46% in network
energy for a data center network with only a negligible increase
in delay. To exploit the well-structured topology of data centers,
Zhang et al. [118] adopted a hierarchical perspective to study
the task of optimizing power usage in data center networks
while guaranteeing connectivity and maximum link utilization.
They divided their discussion into two levels: 1) the core level,
for the determination of the core switches that must be powered
to carry outgoing traffic and the aggregation switches to serve
the out-pod traffic in each pod, and 2) the pod level, for
the determination of the aggregation switches that must be
active to carry intra-pod traffic. They designed a hierarchical
energy optimization algorithm and evaluated it on various
traffic patterns, including large traffic flow only, small traffic
flow only, and random traffic patterns. In [119], the authors
adopted the routing perspective to consider reducing the power
consumption of high-density data center networks while satisfying network performance requirements. They proposed a
routing algorithm that used as little network power as possible
to provide routing services, and they reported a 39% energy
savings for a link load of 20% and a reliability requirement of
2 (i.e., the threshold for the ratio of the network throughput
that the data center operator could tolerate after switch and
link elimination to the basic throughput achievable using all
switches was 2). In [120], the authors proposed a new framework to embrace the new opportunities offered by combining
certain unique features of data centers with traffic engineering.
They solved the problem in two steps: first, by allocating
virtual machines to servers to reduce the amount of traffic and
to generate favorable conditions for traffic engineering, and
then, by reducing the number of active switches and balancing
traffic flows to reduce energy consumption. The experimental
results indicated an energy savings of up to 50%. In addition to
considering the network level, the authors of [121] considered
a joint host-network energy optimization scheme involving the
simultaneous optimization of virtual machine assignment and
network flow routing. They modeled the joint optimization
problem as an integer linear program. Because this problem is
NP-complete, they proposed a series of techniques to address
the associated challenges: 1) they converted the problem to
one that considered a single optimization solution, 2) they
proposed a parallelization approach for rapid completion, and
3) they quickly identified efficient routing paths for the flows.
The authors reported that their optimization method reduced
energy consumption by 40% for a system load of 30%. By
utilizing correlation analysis in the joint power optimization
of a data center network and servers, the authors of [122]
proposed PowerNetS to achieve improved energy savings.
Their approach featured a well-designed heuristic algorithm to
incrementally perform virtual machine and traffic consolidation
with lower virtual machine migration overheads, thereby optimizing power consumption while guaranteeing the fulfillment
of network constraint. They reported that their method could
achieve 51.6% energy savings for a data center.
In Table VI, we summarize and compare the proposals mentioned above. The features presented for comparison are 1) the
traffic patterns, 2) the network architectures, 3) the techniques
used in the proposals, 4) the considered QoS specifications, and
5) the results of the proposals.
C. Power Distribution
Whereas virtual machine consolidation allows the total number of active physical servers to be reduced, power distribution
approaches have also been proposed to reduce the high energy
costs of idle systems. Driven by increasing power density, the
increasing number of IT devices, and the capability of ondemand addition and removal of IT devices, there have been
dramatic changes in how power is utilized in data centers
[123]. In general, power distribution is over-provisioned in data
centers to accommodate peaks and to allow for future expansion. However, based on the assumption that simultaneous
10
TABLE V: Comparison of the proposals regarding virtual machine assignment
Proposal
VM Resources
VirtualPower [101]
Cores on chip
No “Power”
Struggles [102]
CPU
Entropy [103]
Processing unit,
memory
Resource
management [106]
Techniques
VM consolidation, DVFS,
resource (time slice)
scheduling, server
power-down
VM consolidation, DVFS,
server power-down
Traces
Performance
Result
RUBiS tiered web service,
transactional load
Processing rate,
throughput
34% reduction in
power
Trace from enterprise
deployment
Throughput, power
budget
64% reduction in
power
VM consolidation
NASGrid benchmark
suite
Migration time
25% reduction in
nodes per hour
CPU, memory
VM provisioning, VM
packing
Multiplayer online game,
web application
Response time,
throughput
Trade-off between
SLA and energy
Power & SLA
management [107]
CPU
VM consolidation, DVFS,
server power-down
Web workload, Trade6
enterprise application
Response time
22% reduction in
power
Consolidation of
VM [108]
CPU
Dynamic consolidation of
VMs
Trace from PlanetLab
Capacity violation,
VM migration
Consolidation VMs
[109]
Bandwidth
VM consolidation
Trace from IBM Global
service and cluster
Capacity violation
Minimize occupied
servers [111]
CPU, memory
VM placement with
packing constraints
--
--
Risk management
for VMs [112]
CPU, memory
Resource allocation, VM
consolidation
Trace from CoMon project
Capacity violation
Significant
reduction in
energy
30% reduction in
required number
of servers
Reduced number
of servers
Factor of 4
reduction in SLA
violations
TABLE VI: Comparison of the proposals regarding network traffic engineering
Proposal
Network Architectures
Techniques
QoS
Result
E-commerce trace
Fat tree
Traffic aggregation, VM
placement
Bandwidth, connectivity
guarantees
55% network
power reduction
REsPoNse [115]
Media stream, web
workload
Fat tree
Energy critical-path
detection
Throughput, latency
40% network
energy savings
Joint controller
[116]
Parallel workload
Fat tree, clique, HyperX,
BCube
Route selection, VM
placement
Congestion
Data center cost
improvement
CARPO [117]
Wikipedia trace
Fat tree
Traffic consolidation,
link-rate adaptation
Packet delay
46% network
energy savings
HERO [118]
Large, small and
random traffic
Fat tree
Switching off network
switches
Connectivity, maximum
link utilization
network power
savings
PRP [119]
one-to-one,
one-to-many,
all-to-all
Fat tree, BCube
Switch and link
elimination
Throughput
39% network
energy savings
GreenDCN [120]
Communicationintensive
Fat tree
VM assignment, switch
and link engineering
Connectivity
50% network
energy savings
Joint optimization
[121]
Iperf UDP
Fat tree
VM placement, flow
routing
--
40% energy savings
PowerNetS [122]
Wikipedia, Yahoo!
and IBM traces
Fat tree
VM and flow
consolidation
Packet delay
51.6% energy
savings
Network power
controller [114]
Traffic Patterns
peak draw across all equipment will occur only rarely, power
over-subscription is intentionally exploited to improve power
utilization in modern data centers [9], [124], [126], [127]. Fan
et al. [124] were the first to study power usage at the scale
of data center workloads, and they reported the use of power
modeling for power provisioning. Their work yielded several
key findings: 1) The gap between the maximum power actually
used by large groups of servers and their aggregate theoretical
peak usage can be as large as 40% in data centers, which
suggests a significant opportunity to host additional servers
under the same power budget; 2) power capping is more
useful as a safety mechanism to prevent overload situations;
3) large groups of servers are operating near peak power
levels in certain time intervals, which suggests that power
gaps and power management techniques can be more easily
exploited at the data center level than at the rack level; and 4)
when large groups of servers are considered, frequency scaling
has the potential to be moderately effective at reducing peak
power consumption. Their evaluation results indicated energy
savings of 35-40%. Meisner et al. [125] proposed PowerNap,
an energy-conservation approach that supported an entire
system in transitioning rapidly between a high-performance
active state and a near-zero-power idle state in response to
instantaneous variations in load. Based on the PowerNap concept, the authors developed requirements and mechanisms to
eliminate idle power waste in enterprise blade servers. The
11
authors also introduced another power provisioning approach
named RAILS to be applied when PowerNap was operating
in the low-efficiency regions of current blade center power
supplies with the intent of providing high conversion efficiency
throughout the entire range. They reported that PowerNap and
RAILS could reduce the average server power consumption
by 74%. To overcome local utilization spikes that could force
the throttling of server performance to enforce safe power
budgets, Pelley et al. [126] developed mechanisms to better
utilize the installed power infrastructure. They proposed Power
Routing to centrally schedule servers dynamically across redundant power feeds, and they proved that the problem of
assigning servers to power distribution units (PDUs) was NPcomplete. The authors first optimally solved the relaxation
problem through linear programming, assuming the servers
to be assigned fractionally across feeds, and then constructed
an approximate solution to the original problem. Using traces
from production systems, they demonstrated that capital costs
could be reduced by 32% without performance degradation.
For the case in which servers are shared by virtual machines
belonging to different applications, a power budgeting solution
named VPS was presented by [127] to guarantee the quality
of service. VPS dynamically shifted power among various
distributed components as workloads and power availability
varied to efficiently utilize the power budget. The authors
combined hardware-based and software-based power control
knobs to optimize the performance. They used traces to evaluate the following performance metrics: data-center-level and
application-level power budgeting errors, application performance differentiation, and performance achieved within the
power budget. In [128], the authors explored the feasibility
of low-latency power states for enterprise server systems and
proposed an end-to-end virtualization power management solution. They implemented the capability of a low-power and
low-latency state for virtualization servers and demonstrated
this technology could overcome the traditional barriers (e.g.,
incurred latencies) to achieve substantial energy savings via
dynamic consolidation. Based on real-system evaluations, they
reported that data center power efficiency could be improved
by 30%.
In addition to power capping, several studies have introduced Uninterruptible Power Supplies (UPSs) to store energy
and reduce power budget violations [129], [130], [131]. In
[129], the authors pioneered the leveraging of a UPS in a
data center as an energy buffer (eBuff) to address the peak
power draw problem. They used the energy stored in UPS
batteries to provide energy during peak demand, resulting in a
reduction in capital expenses and operating expenses without
performance degradation. They reported that eBuff yielded
a 15-45% peak reduction, which corresponded to a 6-18%
savings in operation expenses. Because eBuff was centralized
and limited by the capacity of the UPS battery, Kontorinis et
al. [130] presented an architecture for distributed per-server
UPSs that stored energy during low-activity periods and used
this energy during power spikes. They proposed details regarding the sizing as well as technological alternatives and
approaches that managed battery charging and discharging
behavior while addressing reliability and availability concerns.
They demonstrated that their policies prolonged the duration of
UPS batteries’ usage and reduced their total cost of ownership
per server by 6.3%. In [131], the authors studied the possibility
of leveraging stored energy to address power emergencies.
Their work demonstrated that mechanisms for power capping
had performance-degrading ramifications, and they proposed
an offline theoretical framework and several online heuristics
to temporarily augment utility supplies during emergencies.
They reported that battery-based solutions 1) could cope with
emergencies of short duration, 2) supplemented existing reaction strategies to enhance efficacy for longer emergencies,
and 3) provided feasible options when other strategies (power
states, migration, etc.) did not suffice. In the case of aggressive
over-subscription, whereas the aforementioned work suggested
power capping and stored energy to reduce the risk of power
peaks, the authors of [132] suggested that the power hierarchy
should be treated as a data center resource of comparable
importance to that of other computing resources. To this end,
the authors proposed vPower, a software system to virtualize
power distribution. vPower combined computing parameters
(such as DVFS, consolidation, etc.) and energy storage devices
to establish a virtual power hierarchy for each application.
The primary advantages of vPower were that it 1) allowed
applications to specify their power needs, 2) performed admission control and assignment, 3) dynamically monitored power
usage, and 4) enforced the allocation of power for fairness and
system efficiency. The results of a power hierarchy prototype
indicated the achieved improvement in system utilization was
over 50% compared with the conservative scheduling approach
based on peak load requirements.
D. Renewable Energy Access
Expanding the use of renewable energy is another route
toward green data centers. Some data center owners have
already investigated and accepted alternative energies, such
as solar, geothermal and wind energy [133], [134]. However,
utilizing these green energies is challenging because of their
restricted, intermittent and unstable nature. Therefore, the main
question that must be addressed is how to exploit renewable
energy and overcome the associated hindrances by exploiting the unique features of data centers. Sharma et al. [135]
developed Blink to address intermittent power constraints.
Based on workload characteristics and energy constraints, Blink
determined the power state of the servers and minimized
performance degradation. Because the power varied and the
energy stored in the batteries was for short-term use, Blink’s
policy had potential benefits for distributed applications. The
authors constructed a cluster of 10 low-power motherboards
powered by an array of micro wind turbines and solar panels to
evaluate the performance under intermittent power constraints.
They demonstrated that an asymmetric load-proportional policy could increase fairness without significantly sacrificing the
cache’s hit rate to achieve a blinking version of memcached.
Considering the location restrictions of renewable energy, Akoush et al. [136] proposed a computation architecture to avoid
the necessity of expensive infrastructure investment. Their design consisted of three relevant components: 1) data centers collocated with renewable energy sources, 2) data centers linked
together over a dedicated communication network, and 3) a
software framework that supported the seamless execution and
migration of virtual machines depending on power availability.
Although they demonstrated the viability of the design in a
case study, the authors also noted the technical challenges and
inherent limitations of their approach, such as the efficiency of
virtual machine migration between data centers and suitable
workload concerns. In [137], the authors leveraged tunable
energy, such as that provided by fuel cells and gas turbines,
12
to propose a power management approach that enabled data
center operation with high performance and low overhead.
Unlike previous work, in which the workload had been forced
to track the variable power budget, their approach explored
the possibility of achieving a renewable energy supply to
follow the data center power demand. To overcome issues of
performance degradation, they proposed two adaptive load
tuning schemes, named DGR Boost and UPS Boost, based on
mature computer tuning approaches (such as DVFS and UPS
batteries). Evaluations performed on real-world data center
traces and industry data on distributed generation systems
indicated that their technique achieved a 37% improvement in
performance over existing supply-tracking-based designs and
saved 100 metric tons of carbon emissions annually, given that
the power consumption of the data center was 10MW.
If powered only by variable sources of green energy, a
data center will experience unpredictable performance for applications. The most reasonable approach is to incorporate
renewable energy into the electrical grid. In [138] and [139],
the authors considered a data center powered by a solar array
and the electrical grid. Two scheduler frameworks, named
GreenSlot and GreenHadoop, were proposed for parallel batch
jobs and data-processing jobs, respectively. Both schedulers
used historical data and weather forecasts to predict the
amount of solar energy that would be available in the future,
with the intent of maximizing the use of green energy. Because
of the inherent natural variability in the availability of solar
energy, grid energy was used to avoid deadline violations.
The results obtained using a real-world trace demonstrated
that GreenSlot could increase green energy consumption by up
to 117% compared with a conventional backfilling scheduler,
whereas GreenHadoop could increase green energy consumption by up to 31% compared with the Hadoop algorithm.
However, the inefficient and redundant load-matching activities of these algorithms can incur performance losses. iSwith
[140] explored the design trade-offs between energy utilization
and load tuning (e.g., DVFS and CPU power states) for a
data center partially powered by intermittent renewable energy.
iSwith combined optimization of the supply-side fluctuation
with minimization of the load-fluctuation-induced overhead
to mitigate performance loss. This system could reduce job
waiting time by 80% and also mitigate peak network traffic by
95% and average network traffic by 75% while still maintaining
96% renewable energy utilization.
To avoid transmitting renewable energy over long distances,
which leads to significant attenuation losses [136], several
proposed solutions have suggest dispatching traffic to multiple data centers in different geographical locations to use
the renewable energy available nearby [141], [142], [4], [143].
In [141], the authors noted that geographical traffic routing
could significantly reduce the use of brown energy if the price
of energy was dynamically determined in proportion to the
instantaneous fraction of the total energy that was brown.
Moreover, the degree of use of renewable energy was dependent on the form of the pricing model and the acceptability of
this dynamic energy pricing. In [142], the authors exploited
multiple uncorrelated wind energy sources to significantly
reduce the effects of intermittency and to achieve almost entirely green Internet-scale data centers (IDCs). They proposed
WPA, a policy that routed jobs based on the current states
of the workloads and wind power availabilities at different
data centers. The authors showed that more than 95% of the
energy consumption in IDCs could be satisfied by wind power
without delaying the processing of jobs. Similarly, Gao et al. [4]
dynamically controlled the fraction of user traffic directed to
each geographically distributed data center with the objective
of navigating the three-way trade-off among carbon footprint,
access latency and electricity cost. The authors showed that
carbon emissions could be reduced by 10% without increasing
either the mean latency or the electricity cost. However, at
present, renewable energy is more expensive to use than is
brown energy produced with fossil-based fuel. Zhang et al.
[143] considered the problem of maximizing the percentage of
renewable energy used to reduce the negative environmental
implications given certain operation budgets. They proposed a
novel middleware system that dynamically distributed requests
among different data centers. The authors formulated their
objective as a constrained optimization problem and solved it
via linear-fractional programming. The authors experimentally
demonstrated that their system could significantly increase
the use of renewable energy without violating the desired
operational cost budget, despite the intermittent nature of the
supplies of renewable energy at different locations and the time
variation in electricity prices and workload traces.
E. Future Research Directions
To reduce the energy consumption of data centers through
efficient utilization of resources, many proposals have been put
forward for the green implementation data centers. Because of
their exponential growth and underutilization, data centers still
pose several challenges, and several inadequately addressed research directions remain with regard to resource management:
• A major reason why the resource utilization of data
centers remains low is that operators are concerned about
potential QoS violations. Recently, several studies have
suggested the existence of interference between collocated applications [144], [145] when servers are virtualized and shared to improve utilization. The challenge of
virtual machine management and data center right-sizing
is to improve resource utilization while preserving QoS
guarantees, which benefits both data center operators and
users.
• In a multi-tenant data center environment, different
tenants may desire different levels of application performance. This characteristic requires heterogeneous resource management mechanisms, which will introduce
additional complexity and overhead. Moreover, tenants
possess their own resources, which may not be controlled
by the data center operators in a centralized manner. Thus
far, few studies have focused on energy-aware resource
management in multi-tenant data centers, and the task
of providing efficient polices for such scenarios requires
future exploration.
• Although traffic consolidation technologies offer considerable energy savings, they can also have severe effects
on network performance. The virtualization of network
functions is a promising approach to providing efficient
support for the deployment and management of network services. However, current implementations such as
CloudNaaS [146] do not account for energy consumption.
Finding a good balance between energy consumption and
network performance remains a challenging problem for
current data center networks.
• Most existing studies concerning power distribution are
focused on intra-data-center networking. However, geo-
13
•
graphically distributed data centers can also offer opportunities for power distribution. For example, a workload
can be preferentially scheduled to data centers with high
power capping or a greater amount of stored energy
by virtue of being located nearer to power stations. The
question of how to reduce energy consumption by combining power distribution, application performance and
workload scheduling remains an open research problem.
A demand-response mechanism is an energy-efficient
method of overcoming the instability that is characteristic
of renewable energy [141]. In SaaS data centers, such a
mechanism also causes service providers to lose users in
response to supply conditions. A promising solution is to
design beneficial pricing schemes to incentivize service
providers to reduce brown energy consumption.
IV.
T EMPERATURE AND T HERMAL C ONTROL
A large portion of the energy used in computing devices
is converted into heat. In chip design, the heat problem is
extremely important, as it can threaten to cause the chips to
fracture at extreme temperatures [147]. An ambient temperature range of 68◦ F to 75◦ F is optimal for device reliability [148].
Data centers house tens of thousands of servers and communication networks, which causes their cooling costs for heat
removal to increase exponentially with increasing energy consumption. It is reported that more than a third and sometimes
as much as one half of the total cost of a data center consists
of cooling costs [149]. As mentioned earlier, virtualizing and
consolidating servers can increase processor utilization rates.
To reduce management expenses and floor-space requirements,
operators prefer to install many multi-core-processor server
blades in a single rack-mounted enclosure. These trends toward
concentration and higher-density computing in combination
with the abovementioned issues mean that the problem of
temperature and thermal control must play a critical role in
the attempt to establish green data centers.
A. Cooling and Workload Distribution
Optimization of the delivery of cooling is one strategy for
temperature control [150]. Patel et al. [150] proposed a system
to fully control a data center environment by linking all its
attributes together, including distributed sensing, variable air
conditioning and data aggregation. In addition, they exploited
variable opening plenum tiles to assist in achieving more flexible control of the cooling distribution and in directing cooling
resources where and when needed. The authors reported an energy savings of 50%. In [151], the authors proposed TE-Shave,
a generalized framework that exploited thermal energy storage
(TES) tanks to reduce data center power. TE-Shave switched to
the TES for stored cold water or ice to supplement the chillers
and for heat exchange during peak power periods. Moreover,
TE-Shave charged the TES through the preparation of cold
water or ice by increasing the cooling power when the data
center experienced power valleys. The authors demonstrated
that their proposed framework could lead to a 28% savings in
operational expenses compared with existing work that focused
only on reducing the server-side power demand.
Another approach to addressing temperature modulation is
based on workload assignment and migration between servers
to achieve thermal balance [152], [153], [154]. In [153], the
authors presented temperature-aware workload assignment
algorithms to minimize the energy expended by the cooling
infrastructure. Their algorithms produced the recommended
theoretical heat distribution to maximize cooling efficiency and
minimize the amount of heat recirculating within a data center.
The proposed algorithms achieved cooling savings of more
than 25%. Heath et al. [154] developed Freon, a user-space
system to achieve thermal management without introducing
unnecessary performance degradation. The main policy driving
the operation of Freon was periodic temperature monitoring
and the exertion of feedback control over the utilization of
the servers by distributing requests based on load weights.
Generally speaking, Freon dynamically shifted load away from
hot servers and increased the load on other, “cool” servers.
An extended version named Freon-EC was proposed to combine energy conservation and thermal management. Freon-EC
turned off servers based on their temperatures and physical
locations in the computer room. In fact, Freon-EC turned hot
servers off and replaced them with servers from different
regions that were unaffected by the thermal threat. Through
experiments, Freon was proven to manage temperature with as
little potential throughput degradation as possible. In [155], the
authors considered the problem of temperature-aware workload distribution in geo-distributed data centers. They explored
two key aspects: 1) utilizing the geographical diversity of
temperature to reduce cooling energy consumption and 2)
exploiting the elastic nature of batch workloads to dynamically
adjust capacity allocation. The problem was formulated as a
joint optimization of request routing for interactive workloads
and capacity allocation for batch workloads, and it was solved
using a distributed m-block Alternating Direction Method of
Multipliers (ADMM) algorithm. Extensive simulations using
real-world traces revealed that their approach saved 15-20%
in cooling energy and 5-20% in overall energy cost. In addition
to these job-allocation-centric techniques, T ∗ [156] provided a
data-centric approach to reducing cooling energy costs in big
data analytics cloud computing. Files were proactively allocated in a thermal- and energy-aware manner based on knowledge of server-profile, data-semantics, and cluster information.
T ∗ proposed predictive and non-predictive temperature-aware
file allocation algorithms to reduce cooling energy costs and
ensure high data-local performance. Evaluations using onemonth-long real-world traces demonstrated a reduction of up
to 42% in cooling energy costs and a 9× better average response
time compared with techniques that were data-allocationagnostic in nature.
B. Temperature-Reliability Trade-off
Several other efforts have focused on supplying data centers with relatively warm air to reduce cooling costs, i.e.,
increasing/controlling the temperature setpoint. Studies show
that increasing the temperature setpoint by merely one degree
can yield an energy savings of 2-5% [157]. However, raising
the temperature of the server’s components can reduce the
system reliability. Based on the critical observation that hot
spots can be selectively cooled, [158] proposed a cooling-power
management mechanism to allow global data center coolers to
operate at a higher temperature while achieving the same level
of chip reliability. The authors integrated thermal models of
various components, such as the silicon in the chip, embedded
thermo-electric coolers, and air conditioners, and quantified the
expected savings in cooling power for the entire data center.
Upon evaluating over 43 applications from the SPECcup 2000
benchmark, they reported a cooling power savings of 12% on
14
average. However, there is a limited understanding of how
higher temperatures will affect data center systems. In [159],
the authors presented a multi-faceted study of temperature
management in data centers. By analyzing a large collection of
field data from different production environments, the authors
studied the influence of higher temperatures on 1) hardware
reliability, such as storage subsystem reliability, memory subsystem reliability, and server reliability as a whole, as well as 2)
server performance and power consumption. They concluded
that the operators could run their data centers hotter than they
currently were to save energy while still limiting the resultant
negative effects on system reliability and performance.
C. Open Questions and Research Directions
The continued growth of data centers is resulting in sharply
increasing levels of energy consumption required for both
powering these data centers and their thermal management.
Several methodologies have recently been introduced that offer
promising approaches to reducing the escalating energy costs
associated with meeting their cooling requirements. However,
there are still several challenges and research issues that must
be addressed regarding the temperature and thermal control of
data centers:
• Most existing studies have focused on heat dissipation and cooling load management within data centers.
Chillers are used to cool the hot water that returns
from the Computer Room Air Conditioners (CRACs) via
mechanical refrigeration cycles, and the compressors for
these chillers consume large amounts of energy [160].
Recently, various so-called free cooling technologies, such
as air-side economizing and water-side economizing,
have been adopted to reduce the dependence on mechanical chillers. Air-side economizing refers to directing
cold outside air into the data center to cool the servers
and ejecting the hot exhaust air back outside instead
of cooling and recirculating it. Water-side economizing
refers to cooling down the warmer chilled water from the
computer room using water streams. However, these free
cooling technologies heavily depend on air conditions
and geographical locations. An interesting question is
whether and how these technologies can be coupled to
address power and cooling demands. A related study
has already been performed [155], but more thorough
theoretical and practical studies are needed to explore this
possibility.
• As mentioned earlier, service providers may own geographically distributed data centers to better meet their
QoS needs, such as delay guarantees. Many of these
providers, such as Google [161], have been relying upon
fresh air for cooling to reduce the energy costs of operating their data centers. When renewable energy is
considered, the problem becomes more interesting and
sophisticated. For example, for systems that rely on airside economization, it is typically preferable to operate
in a cold-temperature environment. However, the generation of renewable energy using solar panels requires
hot, sunny days, which will adversely affect the cooling
energy consumption. Addressing this issue will require
holistic methods that simultaneously consider cooling
energy costs, the availability of renewable energy, and
performance requirements.
• In the current thermal control approaches implemented in
most data centers, the CRAC fans are operating at their
peak speed. Lowering CRAC fan speeds allows energy
consumption to be reduced. However, the challenge is
to develop CRAC fan speed controllers that can supply
a data center with only the necessary cooling demands.
The existing work of [162] represents an initial effort to
address the problem of fan speed management.
V.
G REEN M ETRICS , M ONITORING AND E XPERIMENTAL
T ECHNIQUES
In this section, we present the metrics relevant to building
a “green” data center and practical efforts to monitor and
test data centers for the optimization of energy consumption
and carbon emissions. The primary goals of developing green
metrics are to 1) quantitatively describe the degree of greenness
of a data center, 2) track and report the energy usage and
carbon emissions of a data center to increase its green efficiency,
and 3) provide a clear visualization of the efficiency of a data
center to help its operators to reduce their electricity bills and
carbon taxes. In the meantime, monitoring efforts provide data
center operators with a real-time, detailed understanding of
operation parameters, including green metrics. As a result,
operators can adopt appropriate green measures based on the
output parameters, for example, the consolidation of virtual
machines when the monitoring data indicate that the data center is experiencing low utilization. Moreover, experimentation
and simulation techniques are also critical for designing test
tools and demonstrating energy-efficient policies.
A. Green Data Center Metrics
Various metrics focused on data center energy efficiency have
arisen over the past years. We summarize the most popular
green metrics below:
1) Power Usage Effectiveness (PUE):
PUE =
Total Facility Power
IT Equipment Power
The IT equipment power is the power consumed by the servers,
network, storage, and supplemental equipment (i.e., workstations/laptops used for monitoring); the total facility power
includes the aforementioned power plus auxiliary consumption
in the data center [163].
2) Data Center infrastructure Efficiency (DCiE):
DCiE =
IT Equipment Power
1
=
PUE
Total Facility Power
This metric is simply the inverse of the PUE [163], and both
were developed by The Green Grid to measure energy efficiency in data centers. Barroso et al. [9] suggested that the
average value of the PUE is 1.83, i.e., the DCiE is 0.54.
3) Data Center energy Productivity (DCeP):
DCeP =
Use f ul Work Produced
Total Energy Consumed to Per f orm that Work
This is a metric to characterize the computing efficiency of a
data center [164]. Useful work refers to the jobs performed by
the IT equipment during the time period. The denominator is
the total energy consumed during the assessment window.
15
Set benchmarking goals
Share results
Prioritize metrics
Create follow-up
action plan
Identify required data and
develop data collection plan
Benchmark metrics and
identify potential actions
Obtain and install
monitoring equipment
Analyze data and
compute metrics
Collect data
Fig. 2: The benchmarking process for data center metrics [165]
4) Heating, Ventilation, and Air Conditioning (HVAC):
HVAC =
IT Electrical Energy
HVAC + ( Fuel + Steam + Chilled)×293
The HVAC system energy is the sum of the electrical energy
used for cooling, fan movement, and any other HVAC energy
usage such as steam or chilled water, and it can be reduced
via free cooling with air economizers and/or high-efficiency
chillers [165]. When the value of this metric is high, it suggests
a high potential to reduce HVAC energy usage.
5) Carbon Usage Effectiveness (CUE):
CUE =
Total CO2 Emissions
IT Equipment Energy
The total carbon emissions include direct greenhouse gas emissions and indirect greenhouse gas emissions from the consumption of purchased electricity, heat, and steam. The denominator
is the energy consumption of the IT equipment [166].
In addition to these metrics, many other metrics have been
proposed by The Green Grid and Lawrence Berkeley National
Laboratory that have different practical uses, such as the CEF
(Carbon Emission Factor), AEU (Air Economizer Utilization),
WEU (Water Economizer Utilization) and CSE (Cooling System
Efficiency) [165], [166]. Table VII presents the taxonomy of the
green metrics listed above. Note that the present values shown
in the table were obtained from [165], [166] if no other reference
is indicated.
Now, we focus on works that have studied how to measure
these metrics. A technical report [165] provided an approach
for obtaining the particular values of the metrics. The specific
implementation process for benchmarking a data center is
illustrated in Fig. 2. The authors of this report also provided
guidance for calculating the required data. For example, electrical energy use can be derived from meter data or utility
bills. In [169], the authors’ intent was to present the “what”
and “how” for the measurement of metrics of energy efficiency.
They attempted to introduce a cheap and effective continuous
measurement method based on the power demand profiles
recorded at various points in a facility. Their power measuring
point system is illustrated in Fig. 3. When the required data
have been obtained from the various registries, the metrics can
be calculated based on the power consumed integrated over 15minute or 1-hour periods. By using standard measuring equipment, the reliability and accuracy of the results can be guaranteed. Another work [170] adopted machine learning techniques
to study the PUE. The authors developed a neural network
framework that learned from actual operations data to predict
the PUE. They reported that the PUE was within 0.004+/-0.005,
with an error of approximately 0.4% for a PUE of 1.1. DCeP, the
metric of energy productivity, was studied in [171]. The authors
presented an approach to calculating the DCeP by designing
an experiment using a highly instrumented, high-performance
computing data center. Their evaluation results demonstrated
that the DCeP could be used to successfully distinguish among
different operational states in the data center and validated
its utility as a metric for identifying the configurations of
hardware and software that could improve energy productivity.
Unfortunately, the improvement of one specific green metric
may lead to the worsening of another metric. Therefore, holistic
frameworks for addressing all metrics have also been proposed
in recent works [172], [173]. These approaches offer a combined
visualization of all metrics to guide data center operators in
worst-metric improvement.
Main Power Input
Cooling
Cooling
System
UPS
UPS
System
Lighting
Lighting
Others
Others
IT Equipment Load
Servers
Network
Storage
Supplemental
Equipment
Fig. 3: Power measuring point registries
B. Green Monitoring and Experimental Technique
Monitoring is an effective method of detecting and proactively mitigating the failure of data centers. Moreover, we
will explore efforts that yield the tracks and traces of energy
usage, thermal emission, and power distribution for individual
devices. Based on the collected information, intelligent mechanisms can then react to increase the energy efficiency and
productivity of a data center. In [174], the authors presented a
comprehensive online monitoring service in their GreenCloud
architecture. The monitoring service monitored and collected
data concerning resource utilization, application workloads directed through virtual machine hypervisors, and power usage
using an intelligent power distribution unit. Analyzing and
mining this collected information provided a clear understanding of the data center energy usage and temperature behavior.
Furthermore, the monitoring service helped to consolidate the
16
TABLE VII: Green metrics taxonomy
Metric Definition
Total Facility Power
I T Equipment Power
PUE =
1
PUE
DCiE =
=
I T Equipment Power
Total Facility Power
DCeP =
Use f ul Work Produced
Total Energy Consumed to Per f orm that Work
HVAC =
CEF =
I T Electrical Energy
HVAC +( Fuel + Steam+ Chilled)×293
Total CO2 Emissions
Total Facility Energy
CUE = CEF× PUE =
CSE =
Total CO2 Emissions
I T Equipment Energy
Average Cooling System Power Usage
Average Cooling Load
Objective
Proposer
Ideal
Present Value
Characterize the total energy efficiency
of a data center
Green Grid, 2008
→1
1.83
Characterize the total energy efficiency
of a data center
Green Grid, 2008
→1
0.54
Characterize the IT computing
efficiency
Green Grid, 2008
Larger
866658( NormalizedTasks
)
KWhr
Characterize the energy efficiency of
the HVAC system
Lawrence Berkeley
National Laboratory, 2009
Higher
--
Assess the carbon emissions per unit of
energy used
Green Grid, 2010
Lower
0.147 (Washington)
[167]
Characterize the overall efficiency of
the cooling system
Green Grid, 2010
Lower
0.27 (Washington)
Represent the carbon emission
efficiency of IT energy use
Lawrence Berkeley
National Laboratory, 2009
Lower
0.6-0.8( KW
Ton )
AEU =
Air Economizer Hours
24 ×365
Measure the percentage of hours in a
year that an air-side economizer system
is used to provide “free cooling”
Lawrence Berkeley
National Laboratory, 2009
→100%
76% (New York)
W EU =
Water Economizer Hours
24 ×365
Measure the percentage of hours in a
year that a water-side economizer
system is used to provide “free
cooling”
Lawrence Berkeley
National Laboratory, 2009
→100%
33% =
virtual machine workload and achieve significant energy savings while guaranteeing the real-time performance required for
sensitive applications. Using a different approach, the authors
of [175] developed PowerTracer, a request-tracing strategy for
diagnosing energy inefficiencies in multi-tier service systems.
They first profiled the patterns of requests through kernel
instrumentation and then measured the server-side latency and
service time of each tier. Finally, PowerTracer collected the
resource consumption information for individual requests and
analyzed the features of each tier and each pattern. By thus
gaining insights into the root causes of energy inefficiency, the
authors could devise efficient power-saving methods for multitier applications.
Even in the initial phases of the design and development of a
data center, an effective experimental and simulation platform
enables the operator to identify and understand the design
aspects that are critical to energy and resource efficiency. EEFSim, a virtualized data center simulator for cloud computing
research, was designed and developed in [176]. The simulator
was demonstrated to be suitable for testing, as it reproduced
the behavior of a real cloud framework and the execution
information that it offered. Another advantage of this simulator
was that it could be easily used to evaluate the power performance of a system by testing different solutions and approaches
within merely a few minutes. As a result, the simulator offered
the possibility to validate different scheduling and migration
policies for their power consumption associated with virtual
machines in a cloud data center. Eliazovich et al. [177] presented
GreenCloud, a packet-level simulator for energy-aware studies
of cloud computing data centers. This simulator could be used
to capture the details of the energy consumption of various
data center components (e.g., servers, switches, and links) as
well as packet-level communication patterns. Its effectiveness
in utilizing power management schemes was demonstrated
for two-tier, three-tier, and three-tier high-speed data center
architectures. A more powerful tool for analyzing green data
center design and resource management, named GDCSim,
was proposed in [178]. This simulation tool was used for the
2900
24 ×365
[168]
iterative design of green data centers. Specifically, it captured
the interdependencies between online resource management
schemes and the physical behavior of data centers. The objective of the simulator was to generate data center configurations
for particular purposes, such as CPU sleep-state transitions and
DVFS for power management, as well as thermal behavior
characterization. In [179], the authors proposed a simulator,
named SimWare, to analyze the power consumption of servers,
cooling units, and fans as well as the effects of heat recirculation and air supply timing. SimWare considered the time
required for air to travel from the CRACs to the servers and
could evaluate user-defined job scheduling and virtual machine
management algorithms. Data center designers could use this
simulator to evaluate energy-conserving policies and to assess
mechanical design options such as server placement, airflow
management and CRAC control strategies.
C. Challenges and Design Guidelines
Currently, the PUE is a widely accepted and measured metric
in data centers, whereas the use of other green metrics is
much less widespread. There is no doubt that green efficiency
metrics will continue to be developed and drive improvements
in sustainability and environmental friendliness. However, an
effective metric to characterize the energy and resource usage
efficiency of a data center must satisfy the following design
guidelines:
• able to track the important and actionable parameters to
assist operators in making data centers more green;
• consistent, easy and cost effective to implement and
report for management purposes;
• feasible, simple and inexpensive to monitor.
Thus, developing metrics that are well suited for reducing
energy consumption is a challenging task with a significant
contribution to the construction of “green” data centers.
Moreover, the monitoring of green metrics in data centers
assists in diagnosing energy inefficiencies and in the implementation of proactive mechanisms to improve energy efficiency.
17
However, this issue has received little attention thus far, and
existing monitors have been proposed for either particular
architectures [174] or special applications [175]. The challenges
are the large number of hosted servers or virtual machines and
the availability of communication resources in data centers.
Consequently, centralized monitoring techniques suffer from
low scalability and elasticity. Monitoring techniques adopted
from cooperative monitoring [180] can be applied to overcome
these drawbacks by enabling distributed and robust monitoring
solutions for large-scale, complex data centers. The key research
question of interest is how to minimize the costs incurred while
guaranteeing monitoring accuracy.
Finally, although platforms such as GreenCloud [177] and
GDCSim [178] have provided data center energy-aware simulation environments, the primary disadvantage of these platforms is that most of them are coarse grained and focused
only on certain components or functions. For example, when
a simulation of virtual machine migration considers only CPU
resources, the results for virtual machine allocation may suffer from a failure to consider the bandwidth requirements.
Therefore, the task of designing a system-wide experimental
and simulation platform that integrates all components of a
data center, such as the CPU, memory, cache, I/O, disk and
communication network, requires further exploration.
[11]
D. Abts, M. Marty, P. Wells, P. Klausler and H. Liu. Energy proportional datacenter networks. In Proceedings of ISCA’10, pages
338-347, 2010.
[12] T. Benson, A. Akella and D. Maltz. Network traffic characteristics
of data centers in the wild. In Proceedings of IMC’10, pages 267280, 2010.
[13] S. Rivoire, M. A. Shah, P. Ranganathan, C. Kozyrakis and J. Meza.
Models and metrics to enable energy-efficiency optimizations. In
Computer, 40(12), pages 39-48, 2007.
[14] Natural Resources Defense Council. Data center efficiency assessment. Issue paper, 2014.
[15] A. Beloglazov, J. Abawajy and R. Buyya. Energy-aware resource
allocation heuristics for efficient management of data centers for
cloud computing. In Future Generation Computer Systems, 28(5),
pages 755-768, 2012.
[16] K. Bilal, S.U.R. Malik, O. Khalid, A. Hameed, E. Alvarez, V.
Wijaysekara, R. Irfan, S. Shrestha, D. Dwivedy, M. Ali, U.S.
Khan, A. Abbas, N. Jalil and S.U. Khan. A taxonomy and survey
on green data center networks. In Future Generation Computer
Systems, 36, pages 189-208, 2014.
[17] F. Yao, A. Demers, and S. Shenker. A scheduling model for
reduced CPU energy. In Proceedings of FOCS’95, pages 374–382,
1995.
[18] T. Mudge. Power: A first-class architecture design constraint.
Computer, 34(4), pages 52–58, 2001.
[19] S. Albers. Energy-efficient algorithms. In Communications of the
ACM, 53(5), pages 86-96, 2010.
[20] M. Li, A. C. Yao and F. F. Yao. Discrete and continuous minenergy schedules for variable voltage processors. In Proceedings
of the National Academy of Sciences of USA, 103(11), pages 39833987, 2006.
VI. C ONCLUSION
[21] S. Albers and H. Fujiwara. Energy-efficient algorithms for flow
time minimization. In ACM Transactions on Algorithms, 3(4):49,
The development of green data centers has garnered consid2007.
erable attention because of their financial and environmental
[22] D. P. Bunde. Power-aware scheduling for makespan and flow. In
impact. Extensive research is being conducted in an attempt
Scheduling, 12(5), pages 489-500, 2009.
[23] K. Pruhs, R. V. Stee and P. Uthaisombut. Speed scaling of tasks
to reduce energy consumption and carbon emissions. In this
with precedence constraints. In Theory of Computing Systems,
article, we present a comprehensive review of the current state43(1), pages 67-80, 2008.
of-the-art research in the area of green data center management.
[24] X. Jin, F. Zhang, Y. Song, L. Fan and Z. Liu. Energy-efficient
This survey offers significant insights into various aspects of
scheduling with time and processors eligibility restrictions. In
the problem, including energy efficiency, resource management,
Proceedings of EuroPar’13, pages 66-77, 2013.
[25] N. Bansal, T. Kimbrel and K. Pruhs. Speed scaling to manage
thermal control and green metrics. We summarize and compare
energy and temperature. In Journal of the ACM, 54(1):3, 2007.
the existing schemes and highlight the challenges and potential
[26] S. Albers, F. Müller, and S. Schmelzer. Speed scaling on parallel
directions for future research with regard to each of these
processors. In Proceedings of SPAA’07, pages 289–298, 2007.
aspects.
[27] G. Greiner, T. Nonner, and A. Souza. The bell is ringing in speedscaled multiprocessor scheduling. In Proceedings of SPAA’09,
pages 11–18, 2009.
R EFERENCES
[28] C. Gunaratne, K. Christensen, B. Nordman and S. Suen. Reducing
the energy consumption of ethernet with adaptive link rate
[1] R. L. Villars. Worldwide datacenter census and construction 2014(ALR). In IEEE Transactions on Computers, 57(4), pages 448-461,
2018 forecast: aging enterprise datacenters and the accelerating
2008.
service provider buildout. 2014.
[2] http://www.nrdc.org/energy/data-center-efficiency-assessment.asp. [29] M. Andrews, A. F. Anta, L. Zhang and W. Zhao. Routing for
power minimization in the speed scaling model. In IEEE/ACM
[3] A. Venkatraman. Global census shows datacentre power demand
Transactions on Networking, 20(1), pages 285-294, 2012.
grew 63% in 2012.
[30] A. Antoniadis and C. C. Huang. Non-preemptive speed scaling.
http://www.computerweekly.com/news/2240164589/Datacentre-power-demand-grew-63-in-2012-Global-datacentre-census.
In Scheduling, 16(4), pages 385-394, 2013.
2012.
[31] E. Bampis, A. Kononov, D. Letsios, G. Lucarelli and I. Nemparis.
[4] P.X. Gao, A.R. Curtis, B. Wong and S. Keshav. It’s not easy being
From preemptive to non-preemptive speed-scaling scheduling. In
green. In Proceedings of SIGCOMM’12, pages 211-222, 2012.
Journal of Discrete Applied Mathematics, 181, pages 11-20, 2015.
[5] A. Greenberg, J. Hamilton, D. A. Maltz and P. Patel. The cost
[32] M. Li and F. F. Yao. An efficient algorithm for computing optimal
of a cloud: research problems in data center networks. In ACM
discrete voltage schedules. In SIAM Journal on Computing, 35(3),
SIGCOMM Computer Communication Review, 39(1), pages 68-73,
pages 658-671, 2005.
2009.
[33] X. Jin, F. Zhang and Z. Liu. Discrete Min-Energy Scheduling on
[6] P. Patel-Predd. Energy-efficient ethernet. In IEEE Spectrum, pages
Restricted Parallel Processors. In Proceedings of IPDPS’13, pages
13, 2008.
2226-2229, 2013.
[7] J. M. Kaplan, W. Forrest and N. Kindler. Revolutionizing data
[34] N. Bansal, H. Chan, T. Lam and L. Lee. Scheduling for speed
center energy efficiency. Technical report, McKinsey&Company,
bounded processors. In Proceedings of ICALP’08, pages 409-420,
2008.
2008.
[8] G. James. Power, pollution, and the internet. New York Times,
[35] E. Bampis, A. Kononov, D. Letsios, G. Lucarelli and M. Sviri2012.
denko. Energy efficient scheduling and routing via randomized
[9] L. A. Barroso, J. Clidaras and U. Holzle. The data center as
rounding. In Proceedings of FSTTCS’13, pages 449-460, 2013.
a computer: an introduction to the design of warehouse-scale
[36] N. Bansal, H.-L. Chan, K. Pruhs and D. Katz. Improved Bounds
machines. Morgan&Claypool publish, 2013.
for speed scaling in devices obeying the cube-root rule. In
[10] A. Carrega, S. Singh, R. Bruschi and R. Bolla. Traffic merging for
Proceedings of ICALP’09, pages 144-155, 2009.
energy-efficient datacenter networks. In Proceedings of SPECTS’12,
[37] W. Wu, M. Li and E. Chen. Min-energy scheduling for aligned
pages 1-5, 2012.
18
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
jobs in accelerate model. In Theoretical Computer Science, 412(1214), pages 1122-1139, 2011.
C.-C. Huang and S. Ott. New results for non-preemptive speed
scaling. In Proceedings of MFCS’14, pages 360-371, 2014.
V. Cohen-Addad, Z. Li, C. Mathieu and I. Milis. Energy-efficient
algorithms for non-preemptive speed-scaling. In WAOA’14, 2014.
J.-J. Chen, H.-R. Hsu, K.-H. Chuang, C.-L. Yang, A.-C. Pang and
T.-W. Kuo. Multiprocessor energy-efficient scheduling with task
migration considerations. In Proceedings of ECRTS’04, pages 101108, 2004.
S. Albers, A. Antoniadis and G. Greiner. On multi-processor
speed scaling with migration: extended abstract. In Proceedings
of SPAA’11, pages 279-288, 2011.
E. Angel, E. Bampis, F. Kacem and D. Letsios. Speed scaling on
parallel processors with migration. In Proceedings of EuroPar’12,
pages 128-140, 2012.
M. Li. Approximation algorithms for variable voltage processors:
min energy, max throughput and online heuristics. In Theoretical
Computer Science, 412(32), pages 4074-4080, 2011.
K. Pruhs, P. Uthaisombut and G. Woeginger. Getting the best
response for your erg. In Proceedings of SWAT’04, pages 14-25,
2004.
N. Bansal, K. Pruhs and C. Stein. Speed scaling for weighted flow
time. In Proceedings of SODA’07, pages 805-813, 2007.
T.-W. Lam, L.-K. Lee, I.K. To and P.W. Wong. Speed scaling
functions for flow time scheduling based on active job count.
In Proceedings of ESA’08, pages 647-659, 2008.
N. Bansal, H.-L. Chan and K. Pruhs. Speed scaling with an
arbitrary power function. In Proceedings of SODA’09, pages 693701, 2009.
L.L.H. Andrew, A. Wierman and A. Tang. Optimal speed scaling
under arbitrary power functions. In ACM SIGMETRICS Performance Evaluation Review, 37(2), pages 39-41, 2009.
E. Angel, E. Bampis and V. Chau. Throughput maximization in
the speed-scaling setting. In Proceedings of STACS’14, pages 53-62,
2014.
H.-L. Chan, W.-T. Chan, T.-W. Lam, L.-K. Lee, K.-S. Mak and
P.W.H. Wong. Energy efficient online deadline scheduling. In
Proceedings of SODA’07, pages 795-804, 2007.
T.-W. Lam, L.-K. Lee, I.K.K. To and P.W.H. Wong. Energy efficient
deadline scheduling in two processor systems. In Proceedings of
ISAAC’07, pages 476-487, 2007.
T.-W. Lam, L.-K. Lee, I.K.K. To and P.W.H. Wong. Improved
multi-processor scheduling for flow time and energy. In Scheduling, 15(1), pages 105-116, 2012.
E. Angle, E. Bampis, V. Chau and N.K. Thang. Throughput
maximization in multiprocessor speed-scaling. In Proceedings of
ISAAC’14, pages 247-258, 2014.
A. Gupta, R. Krishnaswamy and K. Pruhs. Scalably scheduling power-heterogeneous processors. In Proceedings of ICALP’10,
pages 312-323, 2010.
R.L. Graham, E.L. Lawler, J.K. Lenstra and A.H.G. Rinnooy Kan.
Optimization and approximation in deterministic sequencing and
scheduling: a survey. In Annals of Discrete Mathematics, vol.5,
pages 287-326, 1979.
A. Gandhi, M. Harchol-Balter, R. Das and C. Lefurgy. Optimal
power allocation in server farms. In Proceedings of SIGMETRICS’09, pages 157-168, 2009.
A. Wierman, L. L. Andrew and A. Tang. Power-aware speed scaling in processor sharing systems. In Proceedings of INFOCOM’09,
pages 2007-2015, 2009.
F. Liu, Z. Zhou, H. Jin, B. Li, B. Li and H. Jiang. On Arbitrating the
Power-Performance Tradeoff in SaaS Clouds. In IEEE Transactions
on Parallel and Distributed Systems, 25(10), pages 2648-2658, 2014.
K. Bilal, S. U. Khan, S. A. Madani, K. Hayat, M. I. Khan, N. MinAllah, J. Kologziej, L. Wang, S. Zeadally and D. Chen. A survey
on green communications using adaptive link rate. In Journal of
Cluster Computing, 16(3), pages 575-589, 2013.
D. Abts, M. R. Marty, P. M. Wells, P. Klausler and H. Liu. Energy proportional datacenter networks. In Proceedings of ISCA’10,
pages 338-347, 2010.
L. Wang, F. Zhang, C. Hou, J. A. Aroca and Z. Liu. Incorporating
rate adaptation into green networking for future data centers. In
Proceedings of NCA’13, pages 106-109, 2013.
S. Irani and A.R. Karlin. On online computation. In Approximation
Algorithms for NP-Hard Problems, chapter 13, pages 521-564, 1997.
A.R. Karlin, M.S. Manasse, L.A. McGeoch and S. Owicki. Com-
[64]
[65]
[66]
[67]
[68]
[69]
[70]
[71]
[72]
[73]
[74]
[75]
[76]
[77]
[78]
[79]
[80]
[81]
[82]
[83]
[84]
[85]
[86]
[87]
[88]
[89]
petitive randomized algorithms for nonuniform problems. In
Algorithmica, 11(6), pages 542-571, 1994.
S. Irani, S. Shukla and R. Gupta. Online strategies for dynamic
power management in systems with multiple power-saving
states. In ACM Transaction on Embedded Computing Systems, 2(3),
pages 325-346, 2003.
J. Augustine, S. Irani and C. Swamy. Optimal power-down
strategies. In SIAM Journal on Computing, 37(5), pages 1499-1516,
2008.
P. Baptiste. Scheduling unit tasks to minimize the number of idle
periods: a polynomial time algorithm for offline dynamic power
management. In SODA’06, pages 364-367, 2006.
P. Baptiste, M. Chrobak and C. Dürr. Polynomial time algorithms
for minimum energy scheduling. In Proceedings of ESA’07, pages
136-150, 2007.
Gururaj et al.. See survey of Marek Chrobak.
http://www.cs.pitt.edu/ kirk/cs3150spring2010
/10071.ChrobakMarek.Slides.pdf, 2010.
E. Angel, E. Bampis and Vincent Chau. Low complexity scheduling algorithms minimizing the energy for tasks with agreeable
deadlines. In Discrete Applied Mathematics, 175, pages 1-10, 2014.
E.D. Demaine, M. Ghodsi, M.T. Hajiaghayi, A.S. Sayedi-Roshkhar
and M. Zadimoghaddam. Scheduling to minimize gaps and
power consumption. In Scheduling, 16(2), pages 151-160, 2013.
J.-J. Chen, M.-J. Kao, D.T. Lee, I. Rutter and D. Wagner. Online
dynamic power management with hard real-time guarantees. In
STACS’14, pages 226-238, 2014.
J. S. Chase, D. C. Anderson, P. N. Thakar, A. M. Vahdat and R. P.
Doyle. Managing energy and server resources in hosting centers.
In Proceedings of SOSP’01, pages 103-116, 2001.
M. Lin, A. Wierman, L. Andrew and E. Thereska. Dynamic
right-sizing for power-proportional data centers. In Proceedings
of INFOCOM’11, pages 1098-1106, 2011.
Y. Azar, N. Ben-Aroya, N. R. Devanur and N. Jain. Cloud
scheduling with setup cost. In Proceedings of SPAA’13, pages 298304, 2013.
J. Chabarek, J. Sommers, P. Barford, C. Estan, D. Tsiang and
S. Wright. Power awareness in network design and routing. In
Proceedings of INFOCOM’08, pages 1130-1138, 2008.
R. Bolla, R. Bruschi, A. Cianfrani and M. Listanti. Enabling
backbone networks to sleep. In IEEE Network, 25(2), pages 2631, 2011.
M. Andrews, A. F. Anta, L. Zhang and W. Zhao. Routing and
scheduling for energy and delay minimization in the powerdown
model. In Proceedings of INFOCOM’10, pages 21-25, 2010.
M. Zhang, C. Yi, B. Liu and B. Zhang. GreenTE: power-aware
traffic engineering. In Proceeding of ICNP’10, pages 21-30, 2010.
B. Heller, S. Seetharaman, P. Mahadevan, Y. Yiakoumis, P.
Sharma, S. Banerjee and N. McKeown. ElasticTree: saving energy
in data center networks. In Proceedings of NSDI’10, pages 249-264,
2010.
J. Shuja, S. A. Madani, K. Bilal, K. Hayat, S. U. Khan and S.
Sarwar. Energy-efficient data centers. In Computing, 94(12), pages
973-994, 2012.
M. Gupta and S. Singh. Greening of the internet. In Proceedings
of SIGCOMM’03, pages 19-26, 2003.
S. Irani, S. Shukla and R. Gupta. Algorithms for power savings.
In ACM Transactions on Algorithms, 2(4):41, 2007.
X. Han, T.-W. Lam, L.-K. Lee, I.K.K. To and P.W.H. Wong.
Deadline scheduling and power management for speed bounded
processors. In Theoretical Computer Science, 411(40-42), pages 35873600, 2010.
G. Kumar and S. Shannigrahi. New online algorithm for dynamic
speed scaling with sleep state. In arXiv:1303.0377, 2013.
S. Albers and A. Antoniadis. Race to idle: new algorithms for
speed scaling with a sleep state. In ACM Transactions on Algorithms 10(2):9, 2014.
A. Antoniadis, C.-C. Huang and S. Ott. A fully polynomialtime approximation scheme for speed scaling with sleep state.
In arXiv:1407.0892, 2014.
T.W. Lam, L.-K. Lee, H.-F. Ting, I.K.-K. To and P.W.H Wong.
Sleep with guilt and work faster to minimize flow plus energy.
In Proceedings of ICALP’09, pages 665-676, 2009.
Y. Chen, A. Das, W. Qin, A. Sivasubramaniam, Q. Wang and N.
Gautam. Managing server energy and operation costs in hosting
centers. In Proceedings of SIGMETRICS’05, pages 303-314, 2005.
Y. Yao, L. Huang, A.B. Sharma, L. Golubchik and M.J. Neely. Data
19
[90]
[91]
[92]
[93]
[94]
[95]
[96]
[97]
[98]
[99]
[100]
[101]
[102]
[103]
[104]
[105]
[106]
[107]
[108]
[109]
[110]
[111]
[112]
centers power reduction: a two time scale approach for delay
tolerant workloads. In Proceedings of INFOCOM’12, pages 14311439, 2012.
Y. Liu, S.C. Draper and N.S. Kim. SleepScale: runtime joint speed
scaling and sleep states management for power efficient data
centers. In Proceedings of ISCA’14, pages 313-324, 2014.
S. Nedevschi, L. Popa, G. Iannaccone, S. Ratnasamy and D.
Wetherall. Reducing network energy consumption via sleeping
and rate-adaptation. In Proceedings of NSDI’08, pages 323-336,
2008.
N. Vasić and D. Kostić. Energy-aware traffic engineering. In
Proceedings of e-Energy’10, pages 169-178, 2010.
L. Wang, F. Zhang, K. Zheng, A. V. Vasilakos, S. Ren and Z. Liu.
Energy-efficient flow scheduling and routing with hard deadlines
in data centers networks. In Proceedings of ICDCS’14, pages 248257, 2014.
A.K. Coskun, R. Strong, D.M. Tullsen and T. S. Rosing. Evaluating the impact of job scheduling and power management
on processor lifetime for chip multiprocessors. In Proceedings of
SIGMETRICS’09, pages 169-180, 2009.
R. Buyya, C.S. Yeo and S. Venugopal. Market-oriented cloud
computing: vision, hype, and reality for delivering IT services as
computing utilities. In Proceedings of HPCC’08, pages 5-13, 2008.
J. F. Kurose and K. W. Ross. Computer networking: a top-down
approach. Addison-Wesley publish, 2013.
R. Kumar, D. M. Tullsen and N. P. Jouppi. Core architecture optimization for heterogeneous chip multiprocessors. In Proceedings
of PACT’06, pages 23-32, 2006.
F. A. Bower, D. J. Sorin and L. P. Cox. The impact of dynamically
heterogeneous multicore processors on thread scheduling. In
IEEE Micro, 28(3), pages 17-25, 2008.
A. Gupta, S. Im, R. Krishnaswamy, B. Moseley and K. Pruhs.
Scheduling heterogeneous processors isn’t as easy as you think.
In Proceedings of SODA’12, pages 1242-1253, 2012.
P. Barham, B. Dragovic, K. Fraser, S. Hand, T. Harris, A. Ho,
R. Neugebauer, I. Pratt and A. Warfield. Xen and the art of
virtualization. In Proceedings of SOSP’03, pages 164-177, 2003.
R. Nathuji, K. Schwan. VirtualPower: coordinated power management in virtualized enterprise systems. In Proceedings of
SOSP’07, pages 265-278, 2007.
R. Raghavendra, P. Ranganathan, V. Talwar, Z. Wang and X.
Zhu. No “power” struggles: coordinated multi-level power management for the data center. In Proceedings of ASPLOS’08, pages
48-59, 2008.
F. Hermenier, X. Lorca, J. M. Menaud, G. Muller and J. Lawall.
Entropy: a consolidation manager for clusters. In Proceedings of
VEE’09, pages 41-50, 2009.
N. Bobroff, A. Kochut and K. Beaty. Dynamic placement of
virtual machines for managing SLA violations. In Proceedings of
IM’07, pages 119-128, 2007.
T. Wood, P. Shenoy, A. Venkataramani and M. Yousif. Blackbox and gray-box strategies for virtual machine migration. In
Proceedings of NSDI’07, pages 229-242, 2007.
H.N. Van, F.D. Tran and J.M. Menaud. SLA-aware virtual resource management for cloud infrastructures. In Proceedings of
CIT’09, pages 357-362, 2009.
D. Kusic, J. O. kephart, J. E. Hanson, N. Kandasamy and
G. Jiang. Power and performance management of virtualized
computing environments via lookahead control. Journal of Cluster
Computing, 12(1), pages 1-15, 2009.
A. Beloglazov and R. Buyya. Optimal online deterministic algorithms and adaptive heuristics for energy and performance
efficient dynamic consolidation of virtual machines in cloud data
centers. In Journal of Concurrency and Computation: Practice and
Experience, 24(3), pages 1397-1420, 2011.
M. Wang, X. Meng and L. Zhang. Consolidating virtual machines with dynamic bandwidth demand in data centers. In
Proceedings of INFOCOM’11, pages 71-75, 2011.
C. C. Lee and D. T. Lee. A simple on-line bin-packing algorithm.
In Journal of the ACM, 32(3), pages 562-572, 1985.
A.L. Stolyar and Y. Zhong. A large-scale service system with
packing constraints: minimizing the number of occupied servers.
In ACM SIGMETRICS Performance Evaluation Review, 41(1), pages
41-52, 2013.
X. Jin, F, Zhang, S. Hu and Z. Liu. Risk management for
virtual machines consolidation in data centers. In Proceedings of
Globecom’13, pages 2872-2878, 2013.
[113]
[114]
[115]
[116]
[117]
[118]
[119]
[120]
[121]
[122]
[123]
[124]
[125]
[126]
[127]
[128]
[129]
[130]
[131]
[132]
[133]
[134]
[135]
[136]
[137]
P. Mahadevan, P. Alto, S. Banerjee, P. Sharma, A. Shah and P.
Ranganathan. On energy efficiency for enterprise and data center
networks. In IEEE Communications Magazine, 49(8), pages 94-100,
2011.
R. McGeer, P. Mahadevan and S. Banerjee. On the complexity
of power minimization schemes in data center networks. In
Proceedings of Globecom’10, pages 1-5, 2010.
N. Vasić, P. Bhurat, D. Novaković, M. Canini, S. Shekhar and D.
Kostić. Identifying and using energy-critical paths. In Proceedings
of CoNEXT’11, No. 18, 2011.
J. Jiang, T. Lan, S. Ha, M. Chen and M. Chiang. Joint VM
placement and routing for data center traffic engineering. In
Proceedings of INFOCOM’12, pages 2876-2880, 2012.
X. Wang, Y. Yao, X. Wang, K. Lu and Q. Cao. CARPO:
correlation-aware power optimization in data center networks.
In Proceedings of INFOCOM’12, pages 1125-1133, 2012.
Y. Zhang and N. Ansari. HERO: hierarchical energy optimization for data center networks. In Proceedings of ICC’12, pages 29242928, 2012.
M. Xu, Y. Shang, D. Li and X. Wang. Greening data center
networks with throughput-guaranteed power-aware routing. In
Computer Networks, 57(15), pages 2880-2899, 2013.
L. Wang, F. Zhang, J.A. Aroca, A.V. Vasilakos, K. Zheng, C. Hou,
D. Li and Z. Liu. GreenDCN: a general framework for achieving
energy efficiency in data center networks. In IEEE Journal on
Selected Areas in Communications, 32(1), pages 4-15, 2014.
H. Jin, T. Cheocherngngarn, D. Levy, A. Smith, D. Pan, J. Liu and
N. Pissinou. Joint host-network optimization for energy-efficient
data center networking. In Proceedings of IPDPS’13, pages 623-634,
2013.
K. Zheng, X. Wang, L. Li and X. Wang. Joint power optimization
of data center network and servers with correlation analysis. In
Proceedings of INFOCOM’14, 2014.
N. Rasmussen. A scalable, reconfigurable, and efficient data center power distribution architecture. American Power Conversion,
White paper #129, 2009.
X. Fan, W.-D. Weber and L.A. Barroso. Power provisioning for
a warehouse-sized computer. In Proceedings of ISCA’07, pages 1323, 2007.
D. Meisner, B.T. Gold and T.F. Wenisch. PowerNap: eliminating
server idle power. In Proceedings of ASPLOS’09, pages 205-216,
2009.
S. Pelley, D. Meisner, P. Zandevakili, T.F. Wenisch and J. Underwood. Power routing: dynamic power provisioning in the data
center. In Proceedings of ASPLOS’10, pages 231-242, 2010.
H. Lim, A. Kansal and J. Liu. Power budgeting for virtualized
data centers. In Proceedings of USENIX ATC’11, 2011.
C. Isci, S. McIntosh, J. Kephart, R. Das, J. Hanson, S. Piper, R.
Wolford, T. Brey, R. Kantner, A. Ng, J. Norris, A. Traore and M.
Frissora. Agile, efficient virtualization power management with
low-latency server power states. In Proceedings of ISCA’13, pages
96-107, 2013.
S. Govindan, A. Sivasubramaniam and B. Urgaonkar. Benefits
and limitations of trapping into stored energy for datacenters. In
Proceedings of ISCA’11, pages 341-352, 2011.
V. Kontorinis, L.E. Zhang, B. Aksanli, J. Sampson, H. Homayoun, E. Pettis, D.M. Tullsen and T. S. Rosing. Managing distributed ups energy for effective power capping in data centers.
In Proceedings of ISCA’12, pages 488-499, 2012.
S. Govindan, D. Wang, A. Sivasubramaniam and B. Urgaonkar.
Leveraging stored energy for handling power emergencies in aggressively provisioned datacenters. In Proceedings of ASPLOS’12,
pages 75-86, 2012.
D. Wang, C. Ren and A. Sivasubramaniam. Virtualizing power
distribution in datacenters. In Proceedings of ISCA’13, pages 595606, 2013.
B. Stone. Google’s next frontier: renewable energy. In New York
Times, 2007.
P. Gupta. Google to use wind energy to power data centers. In
New York Times, 2010.
N. Sharma, S. Barker, D. Irwin and P. Shenoy. Blink: managing
server clusters on intermittent power. In Proceedings of ASPLOS’11, pages 185-198, 2011.
S. Akoush, R. Sohan, A. Rice, A.W. Moore and A. Hopper. Free
lunch: exploiting renewable energy for computing. In Proceedings
of HotOS’11, 2011.
C. Li, R. Zhou and T. Li. Enabling distributed generation pow-
20
[138]
[139]
[140]
[141]
[142]
[143]
[144]
[145]
[146]
[147]
[148]
[149]
[150]
[151]
[152]
[153]
[154]
[155]
[156]
[157]
[158]
[159]
[160]
[161]
ered sustainable high-performance data center. In Proceedings of
ming. http://tinyurl.com/89ros64, 2012.
HPCA’13, pages 35-46, 2013.
[162] V. Sundaralingam, P. Kumar and Y. Joshi. Server heat load based
CRAC fan controller paired with rear door heat exchanger. In
Í. Goiri, K. Le, M.E. Haque, R. Beauchea, T.D. Nguyen, J.
Proceedings of the ASME 2011 Pacific Rim Technical Conference and
Guitart, J. Torres and R. Bianchini. GreenSlot: scheduling energy
Exhibition on Packaging and Integration of Electronic and Photonic
consumption in green datacenters. In Proceedings of SC’11, pages
Systems, pages 489-496, 2011.
1-11, 2011.
[163] C. Belady, A. Rawson, D. Pfleuger and S. Cader. The green grid
Í. Goiri, K. Le, T.D. Nguyen, J. Guitart, J. Torres and R. Biandata center power efficiency metrics: PUE and DCiE. Consortium
chini. GreenHadoop: leveraging green energy in data-processing
Green Grid, 2008.
frameworks. In Proceedings of EuroSys’12, pages 57-70, 2012.
[164] D. Anderson et al.. A framework for data center energy producC. Li, A. Qouneh and T. Li. iSwitch: coordinating and optimiztivity. The Green Grid, 2008.
ing renewable energy powered server clusters. In Proceedings of
[165] P. Mathew, S. Ganguly, S. Greenberg and D. Sartor. SelfISCA’12, pages 512-523, 2012.
benchmarking guide for data centers: metrics, benchmarks, acZ. Liu, M. Lin, A. Wierman, S.H. Low and L.L.H. Andrew.
tions. The Lawrence Berkeley National Laboratory, 2009.
Greening geographical load balancing. In Proceedings of SIGMET[166] D. Azevedo et al.. Carbon usage effectiveness (CUE): A green
RICS’11, pages 233-244, 2011.
grid data center sustainability metric. The Green Grid, 2010.
Y. Gao, Z. Zeng, X. Liu and P.R. Kumar. The answer is blowing
[167] U.S. Department of Energy, Energy Information Administration.
in the wind: analysis of powering Internet data centers with wind
Appendix F. Electricity emission factors. EIA-1605, 2007.
energy. In Proceedings of INFOCOM’13, pages 520-524, 2013.
[168] 42U. WEU: Water-side economizers utilization.
Y. Zhang, Y. Wang and X. Wang. GreenWare: greening cloudhttp://www.42u.com/cooling/water-side-economizers.htm,
scale data centers to maximize the use of renewable energy. In
2011.
Proceedings of Middleware’11, pages 143-164, 2011.
[169] E. Jaureguialzo. PUE: the green grid metric for evaluating the
J. Mars, L. Tang, R. Hundt, K. Skadron and M. L. Soffa. Bubbleenergy efficiency in DC (Data Center). Measurement method
Up:increasing utilization in modern warehouse scale computers
using the power demand. In Proceedings of INTELEC’11, pages
via sensible co-locations. In Proceedings of MICRO’11, pages 2481-8, 2011.
259, 2011.
[170] J. Gao. Machine learning applications for data center optimizaH. Yang, A. Breslow, J. Mars and L. Tang. Bubble-flux: precise
tion. Google Report, 2014.
online QoS management for increased utilization in warehouse
[171] L.H. Sego, A. Márquez, A. Rawson, T. Cader, K. Fox, W.I.
scale computers. In Proceedings of ISCA’13, pages 607-618, 2013.
Gustafson JR. and C.J. Mundy. Implementing the data center enT. Benson, A. Akella, A. Shaikh and S. Sahu. CloudNaaS: a cloud
ergy productivity metric. In ACM Journal on Emerging Technologies
networking platform for enterprise applications. In Proceedings of
in Computing System, 8(4):30, 2012.
SOCC’11, No.8, 2011.
[172] Green IT Promotion Council. DPPE: holistic framework for data
K. Skadron, M.R. Stan, K. Sankaranarayanan, W. Huang, S.
center energy efficiency. 2012.
Velusamy and D. Tarjan. Temperature-aware microarchitecture:
[173] Harmonizing global metrics for data center energy efficiency. In
modeling and implementation. In ACM Transactions on ArchitecMetrics and measurements, 2014.
ture and Code Optimization, 1(1), pages 94-125, 2004.
[174] L. Liu, H. Wang, X. Liu, X. Jin, W. He, Q. Wang and Y.
R. Grundy. Recommended data center temperature & humidity.
Chen. GreenCloud:
a new architecture
for green data
center. In
Computer
Room Temperature
Humidity.htm,
http://www.avtech.com/About/Articles/AVT/NA/All/-/DD-NN-AN-TN/Recommended
Proceedings of ICAC-INDST’09, pages 29-38, 2009.
2005.
[175] G. Lu, J. Zhang, H. Wang, L. Yuan and C. Weng. PowerTracer:
C. Belady, A. Rawson, J. Pfleuger and T. Cader. The green grid
tracing requests in multi-tier services to diagnose energy ineffidata center power efficiency metrics: PUE & DCiE. Green Grid,
ciency. In Proceedings of ICAC’12, pages 97-102, 2012.
Technical Report, 2008.
C.D. Patel, C.E. Bash, R. Sharma and M. Beitelmal. Smart cooling
[176] F. Julià, J. Roldan, R. Nou, J.O. Fitó, A. Vaqué, Í. Goiri and
of data centers. In Proceedings of IPACK’03, pages 129-137, 2003.
J.L. Berral. EEFSim: energy efficiency simulator. Research Report:
W. Zheng, K. Ma and X. Wang. Exploiting thermal energy
UPC-DAC-RR-CAP-2010-15, 2010.
storage to reduce data center capital and operating expenses. In
[177] D. Kliazovich, P. Bouvry, Y. Audzevich and S.U. Khan. GreenProceedings of HPCA’14, pages 132-141, 2014.
Cloud: a packet-level simulator of energy-aware cloud computR.K. Sharma, C.L. Bash, C.D. Patel, R.J. Friedrich and J.S. Chase.
ing data centers. In Proceedings of Globecom’10, pages 1-5, 2010.
Balance of power: dynamic thermal management for Internet
[178] S.K.S. Gupta, R.R. Gilbert, A. Banerjee, Z. Abbasi, T. Mukherjee
Data centers. In IEEE Internet Computing, 9(1), pages 42-49, 2005.
and G. Varsamopoulos. GDCSim: a tool for analyzing green
J. Moore, J. Chase, P. Ranganathan and R. Sharma. Making
data center design and resource management techniques. In
scheduling “cool”: temperature-aware workload placement in
Proceedings of IGCC’11, pages 1-8, 2011.
data centers. In Proceedings of ATEC’05, pages 61-74, 2005.
[179] S. Yeo and H.-H.S. Li. SimWare: a holistic warehouse-scale
T. Heath, A.P. Centeno, P. George, L. Ramos, Y. Jaluria and
computer simulator. In Computer, 45(9), pages 48-55, 2012.
R. Bianchini. Mercury and freon: temperature emulation and
[180] K. Xu and F. Wang. Cooperative monitoring for internet data
management for server systems. In Proceedings of ASPLOS’06,
centers. In Proceedings of IPCCC’08, pages 111-118, 2008.
pages 106-116, 2006.
H. Xu, C. Feng and B. Li. Temperature aware workload management in geo-distributed datacenters. In Proceedings of ICAC’13,
pages 303-314, 2013.
R.T. Kaushik and K. Nahrstedt. T ∗ : a data-centric cooling energy costs reduction approach for big data analytics cloud. In
Proceedings of SC’12, pages 1-11, 2012.
J. Brandon. Going green in the data center: practical steps for
your sme to become more environmentally friendly. In Processor,
29(39), pages 1-30, 2007.
S. Biswas, M. Tiwari, T. Sherwood, L. Theogarajan and F.T.
Chong. Fighting fire with fire: modeling the datacenter-scale effects of targeted superlattice thermal management. In Proceedings
of ISCA’11, pages 331-340, 2011.
N. El-Sayed, I.A. Stefanovici, G. Amvrosiadis, A.A. Hwang and
B. Schroeder. Temperature management in data centers: why
some (might) like it hot. In Proceedings of SIGMETRICS’12, pages
163-174, 2012.
R. Zhou, Z. Wang, A. McReynolds, C.E. Bash, T.W. Christian
and R. Shih. Optimization and control of cooling microgrids for
data centers. In Proceedings of ITherm’12, pages 338-343, 2012.
R. Miller. Too hot for humans, but Google servers keep hum-
| 8cs.DS
|
CELLULAR RESOLUTIONS FROM MAPPING CONES
arXiv:1311.4599v1 [math.AC] 19 Nov 2013
ANTON DOCHTERMANN AND FATEMEH MOHAMMADI
Abstract. One can iteratively obtain a free resolution of any monomial ideal I by considering the mapping cone of the map of complexes associated to adding one generator at
a time. Herzog and Takayama have shown that this procedure yields a minimal resolution
if I has linear quotients, in which case the mapping cone in each step cones a Koszul complex onto the previously constructed resolution. Here we consider cellular realizations of
these resolutions. Extending a construction of Mermin we describe a regular CW-complex
that supports the resolutions of Herzog and Takayama in the case that I has a ‘regular decomposition function’. By varying the choice of chain map we recover other known cellular
resolutions, including the ‘box of complexes’ resolutions of Corso, Nagel, and Reiner and the
related ‘homomorphism complex’ resolutions of Dochtermann and Engström. Other choices
yield combinatorially distinct complexes with interesting structure, and suggests a notion
of a ‘space of cellular resolutions’.
1. Introduction
Let R = k[x1 , x2 , . . . , xn ] be the graded polynomial ring over a field k, and let I ⊂ R be
an ideal generated by monomials I = hm1 , m2 , . . . , mk i. A popular game in combinatorial
commutative algebra is to understand minimal graded resolutions of the monomial ideal I
under various restrictions. In [HT02] Herzog and Takayama describe a general ‘mapping
cone’ procedure for constructing a free resolution of a monomial ideal I. The basic idea is to
utilize the short exact sequences that arise from adding one generator of I at a time, and to
iteratively build the resolution as a mapping cone of an appropriate map between previously
constructed complexes.
If the ideal I has linear quotients with respect to some ordering of its generators, the
complexes and maps in question are particularly well-behaved, and this case the mapping
cone construction leads to a minimal resolution of I. The class of ideals with linear quotients
includes stable ideals, squarefree stable ideals, as well as matroidal ideals (relevant definitions below). Stable monomial ideals themselves have a well-known minimal resolution first
described in [EK90], and the mapping cone construction generalizes this so-called EliahouKervaire (EK) resolution. In this context there is a natural choice of homogeneous basis for
each free module in the resolution of I. In the case that the ideal I has linear quotients, and
furthermore has a regular decomposition function, the differentials in the free resolution can
also be described explicitly. We review these concepts in Section 2.1.
One way to describe the resolution of a (monomial) ideal is through the construction of a
CW-complex whose vertices index the generators of I, and whose higher dimensional faces
index the syzygies. These so-called cellular resolutions were first introduced by Bayer and
Sturmfels in [BS98]. A natural question to ask is if the mapping cone construction can be
Date: March 22, 2018.
Key words and phrases. cellular resolution, mapping cone, linear quotients.
1
realized cellularly. In [Mer10] Mermin shows that the EK resolution is indeed cellular, and
in the case that I is generated in a fixed degree d, is in fact supported on a subcomplex of
a suitably subdivided dilated simplex.
In this paper we combine methods from [HT02] and [Mer10] to show that one can obtain a
larger class of cellular resolutions via the mapping cone construction. Although the mapping
cones of [HT02] are purely algebraic objects, they have a geometric interpretation which,
in the context of ideals with linear quotients, we seek to combine with the decomposition
function approach from [HT02]. The basic idea is that if I is an ideal with linear quotients
then the mapping cone construction can be realized as an iterative procedure where in each
step a geometric simplex is glued in a certain way onto the existing cellular resolution.
In Section 3 we extend Mermin’s result to the case of ideals with linear quotients. Our
main result from that section is the following1.
Theorem 1.1. (Theorem 3.10) Suppose I has linear quotients with respect to some ordering
(m1 , . . . , mk ) of the generators, and furthermore suppose that I has a regular decomposition
function. Then the minimal resolution of I obtained as an iterated mapping cone is cellular
and supported on a regular CW -complex.
In Section 4 we investigate other combinatorial types of cellular resolutions that can be
recovered from the mapping cone construction. In particular, we are interested in a family
of cellular resolutions in the literature ([CN09], [NR09], [Sin08], [DE12]) that all have similar
combinatorial structure. These complexes go by various names and are applied to different
classes of ideals, including the complexes of boxes resolutions of strongly stable and squarefree
strongly stable ideals from [NR09], as well as the homomorphism complex resolutions of
cointerval hypergraph edge ideals of [DE12]. For a fixed ideal I the combinatorial types
of these complexes more or less coincide, but they look much different than the polyhedral
complex obtained as the realization of the EK resolution described above (see examples
below).
We seek to relate these constructions. In Section 4 we show that the homomorphism
complex resolution can be obtained as an iterated mapping cone with a different choice of
‘decomposition function’; in essence a different way to glue in the simplex at each step in
the construction. In this way we provide a uniform description of various cellular cellular
resolutions from the literature, answering a question of Mermin from [Mer10]. Our main
result from that section is the following (see Section 4 for definitions).
Theorem 1.2. (Theorem 4.13) Suppose I = IH is a cointerval ideal associated to a cointerval
hypergraph H, and let XH be the homomorphism complex supporting its minimal resolution.
Then under the lexicographic ordering of the generators of I, the iterated mapping cone
resolution of I is supported on XH .
The choices involved in realizing the mapping cone as a cellular resolution lead to a natural
question of what all possible realizations look like. In Section 4.2 we fix a choice of basis
for each free module in the mapping cone construction and consider the family of geometric
1While
preparing this paper we learned of the recent preprint [Goo13] of Goodarzi, where similar results
were independently obtained.
2
realizations obtained by choosing different regular decomposition functions. The gluing of
the simplex at each step amounts to the choice of a map of chain complexes that lifts the
given map of R-modules. Finally, in Section 4.3 we investigate the extent to which these
choices can be realized as a single space of resolutions.
Acknowledgments. The authors would like to thank Alexander Engström, Jürgen Herzog,
and Volkmar Welker for helpful conversations.
2. Preliminaries
For some fixed field k, we let R = k[x1 , . . . , xn ] denote the polynomial ring on n variables
with its usual Zn -grading, and let I ⊂ R be a monomial ideal. We will be interested in
describing (minimal) graded free resolutions of the ideal I.
2.1. Mapping cones and linear quotients. We begin by recalling the mapping cone
resolutions from [HT02]. For this, let I ⊂ R be a monomial ideal with an ordered set of
generators G(I) = (m1 , . . . , mk ) and define Ij = (m1 , . . . , mj ). Then for each j there are
exact sequences
0 → R/(Ij−1 : mj ) → R/Ij−1 → R/Ij → 0.
Suppose we have we have resolutions G and F of, respectively, the R-modules in the first
two positions of the sequence. We then obtain a resolution of R/Ij as a mapping cone of
a homomorphism of complexes ψ : G → F , where ψ is a lift of the map of R-modules
R/(Ij−1 : mj ) → R/Ij−1 .
Recall that if ψ : G → F is a map of chain complexes then C(ψ), the mapping cone of ψ,
is the chain complex defined by
C(ψ) = G[1] ⊕ F,
with differential
dC(ψ) (g n+1 , f n ) = − dG (g n+1 ), ψ(g n+1 ) + dF (f n ) .
In general the resolution of R/Ij obtained from a mapping cone is not minimal, and so
we search for a class of ideals where one can inductively describe both the resolution of
R/(Ij−1 : mj ) as well as the comparison map ψ. In [HT02] the authors restrict to a class of
ideals for which each of the colon ideals Ij−1 : mj are generated by subsets of the variables, in
which case the resolution of R/(Ij−1 : mj ) is resolved by a Koszul complex. This motivates
the following.
Definition 2.1. A monomial ideal I ⊂ R is said to have linear quotients if there exists an
ordering of the generators I = (m1 , . . . , mk ) such that for each j the colon ideal Ij : mj is
generated by a subset of the variables, so that
Ij : mj = hxj1 , . . . , xjr i.
One can check (see [HT02], attributed to Skölberg) that an ideal I has linear quotients if
and only if the first syzygy module of I has a quadratic Gröbner basis. In the case that I
is squarefree, and hence I is the Stanley-Reisner ideal I∆ of a simplicial complex ∆, this is
equivalent to the Alexander dual ∆∗ being (nonpure) shellable.
3
The property of having linear quotients is itself a generalization of the notion of a stable
ideal, as introduced in [EK90]. If w is a monomial in the polynomial ring R = k[x1 , . . . , xn ]
we let max(w) denote the largest index of the variables dividing w. A monomial ideal I ⊂ R
is called stable if for every monomial w ∈ I and index i < m = max(w), we have that the
monomial xi w/xm belongs to I. Stable ideals are in turn generalizations of strongly stable
(or sometimes shifted, or 0-Borel fixed ) ideals. An ideal I is strongly stable if whenever w ∈ I
with xj dividing I, we have xi w/xj ∈ I for all i < j. Note that j is not required to be the
largest index of variables dividing w, as in the definition of stable ideals.
To summarize, we have the following hierarchy of ideal containments:
{strongly stable} ⊂ {stable} ⊂ {having linear quotients}.
In [HT02] the authors seek minimal resolutions of ideals obtained from the mapping
cone construction; this amounts to finding an explicit description of the comparison maps
ψ : G → F introduced above. The approach taken in [HT02] is to first restrict to ideals
with linear quotients since in this case one can provide a description of the bases of the free
modules in a free resolution. We first establish some terminology.
Definition 2.2. Suppose I has linear quotients with respect to the sequence (m1 , . . . , mk )
of its generators, and for each j with 1 ≤ j ≤ k let Ij = hm1 , . . . , mj i. Let M (I) denote the
set of all monomials in I. For each generator mj , with j = 1, . . . , k, we define
set(mj ) = {k ∈ [n] : xk ∈ hm1 , . . . , mj−1 i : mj }.
In the case that I has linear quotients, it is shown in [HT02] that the mapping cone
construction produces a minimal free resolution of I, and furthermore that a basis for each
free module Fi in the minimal free resolution can be explicitly described as follows.
Lemma 2.3. [HT02, Lemma 1.5] Let I be a monomial ideal with linear quotients. Then the
iterated mapping cone F , derived from the sequence m1 , . . . , mk , is a minimal graded free
resolution of R/I, and for all i > 0, the symbols
(m; α),
where m ∈ G(I)
α ⊂ set(m),
|α| = i − 1, ,
form a homogeneous basis of the ith module in the minimal resolution of R/I.
In [HT02] the authors also provide an explicit description of the differentials in these resolutions for a certain subclass of ideals that satisfy an extra condition. In this context, let
M (I) denote the set of all monomials in I, and define the decomposition function of I to be
the assignment b : M (I) → G(I) given by: define b(m) = mj if j is the smallest number such
that m ∈ hm1 , . . . , mj i. The decomposition function is similar to the description of the differentials in the EK resolution of stable ideals introduced in [EK90]. Indeed, when restricted
to the class of stable ideals, the mapping cone construction recovers the EK resolution. To
describe the differentials from [HT02] we will need to assume one further condition.
Definition 2.4. The decomposition function of I is said to be regular if for each m ∈ G(I)
and every t ∈ set(m) we have
set(b(xt m)) ⊆ set(m).
In this case for each m ∈ G(I) and t, s ∈ set(m) we have
(∗)
b(xs b(xt m)) = b(xt b(xs m)).
4
Note that in degree two for t, s ∈ set(m), in the case that b(xt b(xs m)) 6= b(xt m), we have
b(xt b(xs m)) = b(xs b(xt m)) = xt xs .
Example 2.5. Consider the ideal I = hx1 x3 x4 , x1 x3 x5 , x1 x2 x4 , x1 x4 x5 , x2 x3 x4 , x2 x3 x5 i. Then
I has linear quotients with respect to this order of generators, where for instance I5 :
(x2 x3 x5 ) = hx1 , x4 i. In addition, one can check that the decomposition function b for this
ideal is regular, where for example b(x2 x3 (x1 x4 x5 )) = b(x3 x2 (x1 x4 x5 )). One can also check
that I is not stable and also not cointerval (see Section 4.1 for a definition of the latter).
For the class of ideals with linear quotients and regular decomposition functions we have
the following result.
Theorem 2.6. [HT02, Theorem 1.12.] Let I be a monomial ideal with linear quotients,
and F the graded minimal free resolution of R/I. Suppose the decomposition function b :
M (I) → G(I) is regular. Then the chain map d of F is given by
X
X
xj i m
(b(xji m); α \ ji )
d(m; α) = −
(−1)i−1 xji (m; α \ ji ) +
(−1)i−1
b(xji m)
ji ∈α
ji ∈α
X
X
xj i m
=
(−1)i xji (m; α \ ji ) −
(−1)i
(b(xji m); α \ ji )
b(xji m)
j ∈α
j ∈α
i
i
if σ 6= ∅, where α = {j1 , . . . , jp } ⊂ set(m) with j1 < · · · < jp , and d(m, ∅) = m otherwise.
There are several classes of ideals that are known to have linear quotients that admit
regular decomposition functions. These include stable ideals and matroidal ideals (as shown
in ([HT02]), as well as completely lexsegment ideals [EOS10] and the Alexander dual of the
generalized Hibi ideals from [EHM11].
2.2. Cellular resolutions. One natural way to describe a resolution of an ideal I is through
the construction of a polyhedral (or more general CW-) complex XI whose faces (vertices,
edges, and higher dimensional cells) are labeled by monomials in such a way that the chain
complex determining the cellular homology of X realizes a graded free resolution of I. The
study of cellular resolutions was first explicitly initiated in [BS98] (to where we refer for
further details). Cellular resolutions have the advantage that algebraic resolutions can in
some sense be given a global description, and they also lead to combinatorially interesting
geometric complexes.
The well-known Taylor resolution [Tay66] guarantees that all monomial ideals have a
(usually far from minimal) cellular resolution supported on a simplex. In this case that I
is generated by variables, the Taylor resolution is in fact minimal, and recovers the Koszul
resolution of I. This fact will be used heavily in our construction of the resolution of ideals
with linear quotients, where by definition certain colon ideals are generated by variables.
Not all minimal free resolutions of monomial ideals are supported by a CW-complex, as is
shown in [Vel08], but a natural question to ask is which algebraic complexes can be realized
cellularly (and to provide a geometric/combinatorial description). In [BW02] Batzies and
Welker develop an algebraic version of Discrete Morse theory and construct CW complexes
that support minimal cellular resolutions of shellable ideals (which can be seen to coincide
with the class of ideals with linear quotients). However, there construction is not explicit
and they make no claims regarding the regularity of their complexes.
5
As mentioned in the introduction, Mermin [Mer10] has shown that the Eliahou-Kervaire
resolution of a stable ideal I is cellular and supported on a regular CW-complex. In the
case that I is generated in a fixed degree one can realize the supporting complex as a
subcomplex of a certain subdivision of a dilated simplex (see the next section for details).
In [Sin08] Sinefakopoulos describes cellular resolutions of the class of strongly stable ideals
generated in a fixed degree but obtains combinatorially distinct complexes. In the case that
I = hx1 , . . . xn id , the dth power of the maximal graded ideal in R, the complexes are different
subdivisions of a dilated simplex. In [DE12] the authors describe cellular resolutions of what
they call cointerval ideals via spaces of graph homomorphisms, extending the ‘complexes of
boxes’ constructions from [NR09] and [CN09]. In [NPS02] the authors construct minimal
cellular resolutions of monomial ideals arising from matroids and oriented matroids. In
[DJS12] methods from tropical convexity are employed to build cellular resolutions of a
certain class of monomial ideals. For a particular ideal, these constructions can lead to
(geometric) complexes with different combinatorial structure.
As we will see in Section 4, many of these constructions can be realized as different
instances of the cellular mapping cone construction.
3. The generalized Eliahou Kervaire cellular resolution
In this section we show that the generalized Eliahou-Kervaire (EK) resolution described in
Theorem 2.6 is supported on a cellular complex. We will extend the construction of Mermin
[Mer10] (where the case of stable ideals was considered) to realize the supporting space as
a CW -complex obtained by gluing together certain simplices corresponding to data coming
from the generators of I.
Throughout this section we will use the notation and terminology from Section 2.1. In
particular we assume that our ideal I is a monomial ideal with linear quotients with respect
to the sequence of generators G(I) = (m1 , . . . , mk ), where we use m1 > m2 > · · · > mk to denote the ordering of the generators defining the linear quotient (this will be convenient when
we no longer have indices on the generators). We furthermore assume that its decomposition
function b : M (I) → G(I) is regular (see 2.4).
3.1. Cellular construction. In a description of any cellular resolution we need to label the
0-cells of a CW-complex with monomials corresponding to generators of I. We first note that
each monomial m = xa11 xa22 · · · xann ∈ G(I) can be regarded as a point (a1 , a2 , . . . an ) ∈ Rn
by considering its exponent vector. We will often abuse notation and use m to denote this
point in Rn .
Now, for each m ∈ G(I) and α = {j1 , . . . , jp } ⊂ set(m) with j1 < · · · < jp , we let
σ = (σ1 , . . . , σp ) be a permutation of {j1 , . . . , jp }. We define ch(m, α, σ) to be the subset of
Rn obtained as the convex hull of the subset of generators of I that we reach by applying
the decomposition function in the order prescribed by the permutation σ:
ch(m, α, σ) = Conv m, b(xσ1 m), b(xσ2 xσ1 m), . . . , b(xσp · · · xσ1 m)
Here we use the shorthand notation b(xσ2 xσ1 m) = b(xσ2 b(xσ1 m)). We say that ch(m, α, σ)
is nondegenerate if there are no repetitions of monomials involved in the description of
ch(m, α, σ); otherwise we say that ch(m, α, σ) is degenerate (see Example 3.2). We will see
in Corollary 3.4 that in this case ch(m, α, σ) is in fact a p-dimensional simplex.
6
x1x4x5
3
2
x1x4x5
x1x2x4
3
x1x3x4
x1x3x4
Figure 1. The degenerate simplex ch(x1 x4 x5 , {2, 3}, (32)) on the left, and
the nondegenerate ch(x1 x4 x5 , {2, 3}, (23)) on the right.
Lemma 3.1. Let m ∈ G(I), α ⊂ set(m), and σ a permutation of α. If ch(m, α, σ) is
degenerate then there exists a permutation σ 0 of α such that ch(m, α, σ 0 ) is nondegenerate
and ch(m, α, σ) is a face of ch(m, α, σ 0 ).
Proof. Suppose that ch(m, α, σ) is degenerate. Let s be the first index with the property
that applying b does not lead to a new generator, so that
b(xσs+1 b(xσs · · · xσ1 m)) = b(xσs · · · xσ1 m).
Let σ 0 = (σ1 , . . . , σs−1 , σs+1 , σs , . . . , σp ) be the permutation obtained from σ from switching
positions s and s + 1.
If b(xσs+1 b(xσs−1 · · · xσ1 m)) 6= b(xσs−1 · · · xσ1 m) then we get that ch(m, α, σ) is a face of
ch(m, α, σ 0 ). Otherwise we push σs+1 further back by switching σs−1 and σs+1 to get the
permutation σ 00 = (σ1 , . . . , σs+1 , σs−1 , σs , . . . , σp ). We then apply the above argument for
s − 1 (instead of s) and in order to find a permutation σ 00 such that ch(m, α, σ 00 ) contains
ch(m, α, σ 0 ) as a face. In the worst case we consider the permutation (σs+1 , σ1 , . . . , σp ) and
we have b(xσs+1 m) 6= m as desired.
By continuing the same argument, we keep increasing dimensions and we get a nondegenerate ch(m, α, σ 0 ) containing ch(m, α, σ) as a face.
Example 3.2. We return to the ideal I from Example 2.5, where
I = hx1 x3 x4 , x1 x3 x5 , x1 x2 x4 , x1 x4 x5 , x2 x3 x4 , x2 x3 x5 i
has linear quotients with respect to the given ordering of the generators. One can check
that set(x1 x4 x5 ) = {2, 3}. The simplices corresponding to the permutations σ = (23) and
σ 0 = (32) are described below.
(i) The simplex ch(x1 x4 x5 , {2, 3}, (32)) is degenerate since it is the convex hull of the
vectors corresponding to x1 x4 x5 and x1 x3 x4 = b(x3 (x1 x4 x5 )). Here the fact that
2 6∈ set(x1 x3 x4 ) implies degeneracy (see Figure 3.1).
(ii) Switching the positions of 2 and 3 gives us the permutation σ 0 = (32). The resulting simplex ch(x1 x4 x5 , {2, 3}, (32)) is then nondegenerate since it is the convex
hull of the vectors corresponding to x1 x4 x5 , x1 x2 x4 = b(x2 (x1 x4 x5 )), and x1 x3 x4 =
b(x3 (x1 x2 x4 )). Note that ch(x1 x4 x5 , {2, 3}, σ) is a face of ch(x1 x4 x5 , {2, 3}, σ 0 ).
7
Lemma 3.3. Suppose ch(m, α, σ) is nondegenerate. Then for each ` and j with j < ` we
have
b(xσj xσj−1 · · · xσ1 m) 6= xσj b(xσj−1 · · · xσ1 m)/xσ` .
Proof. By contradiction assume that for some j < ` we have
b(xσj m0 ) = xσj m0 /xσ` ,
where m0 = b(xσj−1 · · · xσ1 m). Then by our assumption of regularity (∗) we can push σ` back
until it meets σj as follows:
b(xσ` · · · xσ1 m) = b(xσ` b(xσ`−1 · · · xσj+1 b(xσj m0 )))
= b(xσ`−1 b(xσ` xσ`−2 · · · xσj+1 b(xσj m0 )))
= ···
= b(xσ`−1 b(xσ`−2 · · · xσj+1 xσ` b(xσj m0 )))
= b(xσ`−1 · · · xσj+1 xσ` (xσj m0 /xσ` ))
= b(xσ`−1 · · · xσ1 m),
which is a contradiction by our assumption that ch(m, α, σ) is nondegenerate.
As a consequence of Lemma 3.3 we have
Corollary 3.4. Suppose that | set(m)| = p and ch(m, set(m), σ) is nondegenerate. Then
ch(m, set(m), σ) is a simplex of dimension p.
Proof. Let v0 = (v0,1 , v0,2 , . . . , v0,n ) be the exponent vector of m and vi = (vi,1 , vi,2 , . . . , vi,n )
be the exponent vector of b(xσi · · · xσ1 m) for each i > 0. We will show that v0 , v1 , . . . , vp are
affinely independent. Assume that λ0 v0 + λ1 v1 + · · · + λp vp = 0 and λ0 + λ1 + · · · + λp = 0.
Then by Lemma 3.3 we know that for each `,
(
v`−1,σ` ,
if i < `;
vi,σ` =
v`−1,σ` + 1, if i ≥ `.
In particular vp,σp = v`,σp + 1 for all ` < p. Therefore, we have
(vp,σp − 1)(λ0 + λ1 + · · · + λp−1 ) + vp,σp λp = 0
which implies that λp = 0. A similar argument shows that λi = 0 for all i.
As we have seen, for each subset α ⊂ set(m) of size q we have some permutation σ of α
such that the simplex ch(m, α, σ) is of dimension q − 1.
Definition 3.5. Suppose m is a generator of I and α ⊂ set(m). Let F be a facet of the
(possibly degenerate) simplex ch(m, α, σ). We say that F is an exterior facet if it is not
a facet of ch(m, α, σ 0 ) for some other permutation σ 0 of α. Otherwise we say that F is an
interior facet.
Corollary 3.6. Let α ∈ set(m), let σ be some permutation of α, and suppose F is a facet of
the simplex ch(m, α, σ) = Conv{m, m1 , . . . , mp }. Then F is exterior if and only if m` 6∈ F
for some 1 < ` < p and σ` ∈ set(b(xσ`+1 m`−1 )). If F is interior, then there exist exactly two
nondegenerate simplices containing F : the simplex ch(m, α, σ) and the simplex ch(m, α, σ 0 ),
where σ 0 = (σ1 , . . . , σ`−1 , σ`+1 , σ` , . . . , σp ).
8
Proof. Let ch(m, α, σ) = hm, m1 , . . . , mp i be a nondegenerate simplex which contains F as a
facet with m` 6∈ F for some 1 < ` < p. Now we consider the following two cases:
Case (i). Suppose σ` ∈ set(b(xσ`+1 m`−1 )). Then b(xσ` xσ`+1 m`−1 ) 6= b(xσ`+1 m`−1 ) and so
F is also a facet of ch(m, set(m), σ 0 ), where σ 0 = (σ1 , . . . , σ`−1 , σ`+1 , σ` , . . . , σp ). Therefore F
is an interior facet.
By contradicton assume that ch(m, set(m), σ 00 ) is another nondegenerate simplex containing F as a facet such that σi00 = σ` . If i < `, then the corresponding simplex does not contain
m`−1 . If i > `, then m`+1 does not appear among the vertices, a contradiction.
Case (ii). Suppose σ` 6∈ set(b(xσ`+1 m`−1 )). In this case we have b(xσ` xσ`+1 m`−1 ) =
b(xσ`+1 m`−1 ). Assume that ch(m, set(m), σ 0 ) = hw0 , w1 , . . . , wp i is another simplex containing
F as a facet and let wi 6∈ {m, m1 , . . . , mp }. We wish to show that i = 0. By contradiction
we suppose i ≥ 1 and note that b(xσ` wi−1 ) = wi . Our assumption that b(xσ` xσ`+1 m`−1 ) =
b(xσ`+1 m`−1 ) implies that σ` ∈ set(mi ) for all i ≤ ` and σ` 6∈ set(mi ) for all i > `. On the other
hand, σ` ∈ set(wi−1 ) which implies that i < ` and σ`+1 ∈ set(wi ). Thus wi+1 = b(xσ`+1 wi ) is
not equal to m`+1 which implies that wi , wi+1 6∈ F , a contradiction.
Therefore we have ` = 0 so that F = hm1 , . . . , mp i is the facet of ch(m, α, σ) obtained by
removing the vertex m. We conclude that F is not a facet of any other simplex of the form
ch(m, α, σ 0 ) = hm, m01 , . . . , m0p i, and hence F is exterior.
We next construct the cells that will serve as basis elements of the free modules in our
resolution. We obtain these by gluing together the simplices ch(m, α, σ) corresponding to
the different choices of the permutation σ. For this we define the cell U (m, α) as the union
over all permutations σ of α.
[
U (m, α) =
ch(m, α, σ).
σ a permutation of α
Note that by Lemma 3.1, U (m, α) can be written (as a subset of Rn ) as the union of
nondegenerate simplices ch(m, α, σ).
3.2. Orientation of ch(m, set(m), σ):
Definition 3.7. For each p ≥ 1 we fix the permutation (p, . . . , 1). Then for each permutation
σ of {1, . . . , p} we define
(σ) = sgn(σ, (p, . . . , 1))
where sgn(−, −) denotes the standard sign function for permutations.
Lemma 3.8. There exists an orientation on the simplices ch(m, set(m), σ) such if F in an
interior facet belonging to both ch(m, set(m), σ) and ch(m, set(m), σ 0 ), then F has the same
induced orientation if and only if the the simplices themselves have opposite orientations.
Proof. By Lemma 3.6 each interior facet F belongs to exactly two simplices ch(m, set(m), σ)
and ch(m, set(m), σ 0 ). Therefore the orientation of each interior facet F can be determined
uniquely by the simplices containing F .
For each two nondegenerate simploices G = ch(m, set(m), σ) and G0 = ch(m, set(m), σ 0 )
there exists a chain G0 = G, G1 , . . . , Gr = G0 such that Gi+1 is obtained from Gi by swapping
two suitable indices.
9
These two facts show that the orientations on the simplices can be determined uniquely.
Notation. Assume that F is a facet of ch(m, set(m), σ) = hm = m0 , m1 , . . . , mp i. When F
is an exterior facet, or when the simplex containing F is clear, we denote o(F ) to denote the
the orientation of F that is determined by sgn(σ) and the missing vertex in F .
The following is our main technical lemma regarding the differential maps.
Lemma 3.9. The topological differentials of each cell U (m, α) with subset α = {j1 , . . . , jp }
of set(m) are
X
X
d(U (m, α)) =
(−1)i U (m, α\ji ) −
(−1)i U (b(xji m), α\ji ) .
i
i
Proof. For the simplex hm0 , m1 , . . . , mp i, the boundary map is given by
d(hm0 , m1 , . . . , mp i) =
p
X
(−1)p−i hm0 , . . . , m
ci , . . . , mp i
i=0
For α ⊂ set(m), the cell U (m, α) for α ⊆ set(m) is the union of all nondegenerate simplices
ch(m, α, σ), and is oriented as
X
U (m, α) =
o(F )F,
F =ch(m,α,σ)
nondegenerate
where o(F ) denotes the unique orientation given by Lemma 3.8. So the topological differentials of the cell α ⊆ set(m) are given by
X
d(U (m, α)) =
o(F )d(F ).
F =ch(m,α,σ)
nondegenerate
Now corresponding to the fixed simplex ch(m, α, σ) = hm0 , m1 , . . . , mp i we have the following terms in the differential of U (m, α):
Case 1. i = 0. The term corresponding to i = 0 is (−1)p sgn(σ)hm1 , . . . , mp i.
Let σ 0 = (σ2 , . . . , σp ) and π be the permutation such that πσ 0 = (p, . . . , σb1 , . . . , 1). Then
(σ1 , p)(σ1 , p − 1) · · · (σ1 , σ1 + 1)πσ 0 = (p, . . . , 1)
shows that sgn(σ) = sgn(σ 0 )(−1)p−σ1 . Thus the corresponding term is
(−1)σ1 sgn(σ 0 ) ch(b(xjσ1 m), α\jσ1 , σ 0 ).
This face can not be obtained from the differential of any other simplex in U (m, α) by
Lemma 3.6.
Case 2. i = p. In this case the term corresponding to i = p is sgn(σ)hm0 , . . . , mp−1 i.
Then again this face can not be obtained from the differential of any other simplex in U (m, α).
As in Case 1 we have sgn(σ) = sgn(σ 0 )(−1)σp +1 , where σ 0 = (σ1 , . . . , σp−1 ). Thus this term
can be written as
(−1)σp +1 sgn(σ 0 ) ch(m, α\jσp , σ 0 ).
10
Case 3. 0 < i < p. The corresponding term is
(−1)p−i sgn(σ)hm0 , . . . , m
ci , . . . , mp i.
Now we have two subcases:
Case 3.1. First suppose σi ∈ set(b(xjσi+1 mi−1 )). Thus b(xσi b(xjσi+1 mi−1 )) 6= b(xjσi+1 mi−1 ).
We set σ 0 = (σ1 , . . . , σi−1 , σi+1 , σi , . . . , σp ). Our condition guarantees that ch(m, α, σ 0 ) is
nondegenerate. Note that ch(m, α, σ 0 ) = hm0 , . . . , mi−1 , m0i , mi+1 , . . . , mp i for some m0i . Now
by removing the ith vertex of ch(m, α, σ 0 ) we get
c0 , . . . , mp i.
(−1)p−i sgn(σ 0 )hm0 , . . . , m
i
Since sgn(σ 0 ) = − sgn(σ), when we take the sum over all possible permutations of α given
nondegenerate cells, these two terms will be canceled.
Case 3.2. Next suppose σi 6∈ set(b(xjσi+1 mi−1 )). Then b(xσi b(xjσi+1 mi−1 )) = b(xjσi+1 mi−1 )
and the corresponding term is
(−1)p−i sgn(σ) ch(m, α\jσi , σ 0 ),
where σ 0 = (σ1 , . . . , σbi , . . . , σp ). Since (−1)p−i−σi +1 sgn(σ 0 ) = sgn(σ), we can write this term
as
(−1)σi +1 ch(m, α\jσi , σ 0 ).
P
Now by considering
d(o(F )F ) over all non-degenerate cells F we find that the remaining terms are the sum of
o(σ 0 ) ch(b(xji m), α\ji , σ 0 ),
(1)
P
i
(2)
P
i+1
i (−1)
(3.2)
i (−1)
o(σ 0 ) ch(m, α\ji , σ 0 ), where ji ∈ set(b(xji+1 b(xji−1 · · · xj1 m))),
i+1
o(σ 0 ) ch(m, α\ji , σ 0 ),
i (−1)
P
where ji 6∈ set(b(xji+1 b(xji−1 · · · xj1 m)))
over all σ 0 P
where the corresponding facet is nondegenerate. Then the first sum can be
i
written
P as i+1i (−1) U (b(xji m), α\ji ) and the sums coming from (2) and (3.2) can be written
as i (−1) U (m, α\ji ). Therefore
X
X
d(U (m, α)) =
(−1)i U (b(xji m), α\ji ) +
(−1)i+1 U (m, α\ji ),
i
i
as desired. This completes the proof.
With these preliminaries in place we can establish the main result of this section.
Theorem 3.10. Suppose I has linear quotients with respect to some ordering (m1 , . . . , mk )
of the generators, and furthermore suppose that I has a regular decomposition function. Then
the minimal resolution of I obtained as an iterated mapping cone is cellular and supported
on a regular CW -complex.
11
x3x5
x2x5
2
1
x1x5
x1x5
2
x1x2
2
x1x3
x2x3
x1x5
x2x5
3
3
x2x3
1
x1x3
x1x2
x1x3
x1x2
x4x5
2
x1x5
1
3
1
x3x5
x2x5
x1x2
x1x3
x2x3
x1x5
x1x3
x1x2
x2x3
x1x2
x1x3
Figure 2. The resolution of I = hx1 x2 , x1 x3 , x1 x5 , x2 x3 , x2 x5 , x3 x5 , x4 x5 i
built from an iterated mapping cone. At each step we have labeled the newly
added edges with the elements of set(mj ).
Proof. Adding the monomial coefficients to the differential map from Lemma 3.9 we obtain
X
X
xj i m
d(U (m, α)) =
(−1)i (
(−1)i xji U (m, α\ji ).
)U (b(xji m), α\ji ) −
b(x
m)
j
i
i
i
This is precisely the minimal free resolution of I described in Theorem 2.6. Therefore the
complex constructed as the union of the cells U (m, α) supports the minimal free resolution
of I, as desired. Moreover this resolution in the closed form looks like the Eliahou-Kervaire
resolution.
Note that for any two simplices G = ch(m, set(m), σ) and G0 = ch(m, set(m), σ 0 ) there
exists the chain G0 = G, G1 , . . . , Gr = G0 where Gi+1 can be obtained from Gi by swapping
two suitable indices. This implies that U (m, set(m)) is a shellable for all m. Hence the
constructed resolution is regular, by Lemma 3.6 and [DK73, Proposition 1.2].
Remark 3.11. We note that in the proofs of Theorem 3.10 and the related lemmas, the only
property of the decomposition function that we use is its regularity, and not its definition
in terms of assigning a particular generator to a monomial. Hence we obtain similar results
for any decomposition type function that satisfies the regularity property. We return to this
point in Section 4.2 where we vary the decomposition function to obtain combinatorially
distinct cellular resolutions.
4. Other cellular realizations of the mapping cone
In each step of the mapping cone construction we need to choose the homomorphism of
complexes ψ : G → F that lifts the map of R-modules R/(Ij−1 : fj ) → R/Ij−1 . In the context
of the generalized EK resolution described above, this choice was encoded in the definition of
12
the decomposition function ψ. We will see that varying the decomposition function leads to
combinatorially distinct geometric complexes supporting cellular resolutions, recovering the
results from the papers mentioned above. We view the cellular mapping cone construction
as a means of unifying the various constructions of cellular resolutions from the literature.
4.1. The complex of boxes (homomorphism) resolution. In this section we show how
a different choice of decomposition function in the mapping cone construction recovers the
cellular resolutions of [Sin08], of [CN09] and [NR09] (where they are called ‘complex of
boxes’ resolutions), and of [DE12] (where they are constructed as ‘homomorphism complex’
resolutions).
In this context we restrict our attention to cointerval ideals, a class of hypergraph edge
ideals introduced in [DE12] and [MKM10] that generalize squarefree strongly stable ideals.
Recall that a (regular) d–graph H on vertex set [n] is a collection of subsets of [n] =
{1, 2, . . . , }, each of cardinality d. A d-graph H naturally gives rise to a (square-free) monomial ideal IH by taking generators to be the edges of H. To describe the class of cointerval
graphs we need the following notion.
Definition 4.1. Let H be a d–graph and let v ∈ V (H) ⊆ Z be some vertex. Then the
v–layer of H is a (d − 1)–graph on V \ v with edge set
{v1 v2 · · · vd−1 | vv1 v2 · · · vd−1 ∈ E(H) and v < v1 , v2 , . . . , vd−1 }.
Definition 4.2. The class of cointerval d–graphs is defined recursively as follows.
Any 1–graph is cointerval. For d > 1, a finite regular d–graph H with vertex set V (H) ⊆ Z
is cointerval if
(1) for every i ∈ V (H) the i–layer of H is cointerval;
(2) for every pair i < j of vertices, the j–layer of H is a subgraph of the i–layer of H.
When d = 2 the class of cointerval graphs can be seen to coincide with the well-studied
complements of interval graphs of structural graph theory (hence the name). One can see
that cointerval d–graphs generalize the class of (pure) shifted simplicial complexes. Given a
cointerval d-graph H, we will often refer to the associated edge ideal IH as a cointerval ideal.
In [MKM10] the authors work with a class of ideals they call generalized Ferrers ideals
which can be seen to coincide with the class of cointerval ideals. We recall the equivalent
definition here.
Lemma 4.3. A squarefree monomial ideal I (generated by monomials of degree d) is cointerval if and only if for any monomial m = xi1 xi2 · · · xid ∈ I we also have
xj1 xj2 · · · xjt xit+1 · · · xid ∈ I,
where (i1 , i2 , . . . , id ) and (j1 , j2 , . . . , jd ) are such that j1 ≤ i1 , j2 ≤ i2 , . . . , jt ≤ it for some
t ≤ d.
From [MKM10] we also take the following result.
Theorem 4.4. [MKM10, Theorem 2.5] Let IH be the edge ideal of a generalized d-Ferrers
hypergraph H, so that IH is a cointerval ideal. Then IH is weakly polymatroidal, and in
particular has linear quotients with respect to lexicographic order on its generators.
13
Next we recall the construction of the polyhedral complex XH that supports a minimal
free resolution of the cointerval ideal associated to the hypergraph H. For subsets σi , σj ⊂
[n] = {1, 2, . . . , n} we say that σi < σj if x < y for all x ∈ σi and y ∈ σj . We use the notation
∆S to denote the simplex with vertex set S.
Definition 4.5. Let H be a d–graph on vertex set V (H) = {v1 , v2 . . . , vn }. The polyhedral
complex XH is defined to be the subcomplex of the product
d
Y
∆V (H)
i=1
satisfying
(1) The vertices of XH are vi1 × vi2 × · · · × vid , where vi1 vi2 · · · vid is an edge of H;
(2) For σi ⊆ V (H), the cells σ1 × σ2 × · · · × σd satisfy σ1 < σ2 < · · · < σd .
Note that for any d-graph H, the faces of the complex XH are naturally labeled by
monomials. In particular, the vertices are labeled by monomials corresponding to the edges
of H (i.e. the generators of IH ), and the higher dimensional faces F = σ1 × σ2 × · · · × σd are
labeled by
d Y
Y
xj ,
i=1 vj ∈σi
which can be seen as equal to the least common multiple of the monomial labels on the
vertices of F .
Remark 4.6. Viewing H as a directed d-graph (with orientation on the edges given by the
integer labels on the vertices), one can regard XH as a ‘space of directed edges’ of H. Indeed,
if we let E denote the d-graph with vertex set [d] consisting of a single edge {1, 2, . . . , d},
then XH = Hom(E, H), a space of directed graph homomorphisms from E to H analogous to the undirected Hom complexes of [BK06]. This perspective was also employed in
[BBK12] where the authors study ideals arising from more general (nondegenerate) simplicial
homomorphisms.
The main result from [DE12] is that these complexes support minimal cellular resolutions
of cointerval ideals.
Theorem 4.7. [DE12, Theorem 4.1] Let H be a cointerval d–graph on vertex set [n]. Then
the polyhedral complex XH supports a minimal cellular resolution of the edge ideal IH .
One nice thing about the spaces XH is tdhat the differential maps are so easy to describe.
Indeed, if σ1 × σ2 × · · · × σd is a cell of XH , we can write d σ1 × σ2 × · · · × σd as:
(1)
d
X
`=1
(−1)`−1
X
(−1)j+|σ1 |+···+|σ`−1 | σ1 × σ2 × · · · × {σ`1 , σ`2 , . . . , σc
`j , . . . , σ`k } × · · · × σd .
j
Here we remove the element σij only if it leaves a nonempty subset, that is if |σi | ≥ 2.
These are the differential maps we wish to recover in Theorem 4.13, where we show that the
homomorphism complex XH can in fact also be realized as an iterated geometric mapping
cone construction. We first need another description of the basis elements.
14
Lemma 4.8. Let I = IH be a cointerval ideal associated to the cointerval d–graph H, and
let F be its minimal free resolution. Then the basis elements for each free R-module Fi
determined by the cellular resolution XH correspond to the symbols
(m; α),
where m ∈ G(I),
α ⊂ set(m),
|α| = i − 1.
Proof. Suppose IH is a cointerval ideal with monomial generators m1 , m2 , . . . , mk listed in
lexicographical order. From Lemma 4.3 we see that for any generator m = xi1 xi2 · · · xid we
have
set(m) = {j ∈ [n] : j < i1 or for some t ≥ 1 we have it < j < it+1 , xi1 · · · xit xj xit+2 · · · xid ∈ I}.
Since XH supports a minimal cellular resolution of IH , we have that a basis for Fi is given
by the number of i − 1 dimensional faces of XH . From Definition 4.5 we have an explicit
description of these faces, which we now want to show are naturally labeled by the symbols
(m; α).
Suppose m = xi1 xi2 · · · xid ∈ G(I) and α ⊂ set(m). We associate the symbol (m; α) to the
face σ1 × σ2 × · · · × σd , where for 1 ≤ ` ≤ d we define σ` = {i` } ∪ {j ∈ α : i`−1 < j < i` }
(by convention we set i0 = 0). To see that this assignment defines a bijection we describe
the inverse. For this, suppose σ1 × σ2 × · · · × σd is a face of XH . Define m = xi1 xi2 · · · xid ,
where for each ` with 1 ≤ ` ≤ d we set i` = max(σ` ). Then define
α=
[
(σ` \i` ).
`
One can check that these assignments are inverses of one another.
The differentials of the resolution supported on the homomorphism complex are described
by the incidence face structure of the polyhedral complex XH . To recover these differentials
as an iterated mapping cone, we need a new notion of decomposition function.
Let I be a cointerval monomial ideal with the sequence of generators m1 , m2 , . . . , mk listed
in lexicographical order. As above let G(I) be the set of all monomials in I. Let xG(I) be
the set of all products xi mj , where mj ∈ G(I) is a generator of I, and xi is a variable in the
polynomial ring such that i ∈ set(mj ).
Definition 4.9. With the notation established a above, the map c : xG(I) → G(I) is
defined as follows. Suppose mj = xj1 xj2 · · · xjd , with j1 < j2 < · · · < jd and let xi such that
i ∈ set(m). Pick the minimum jk such that jk−1 < i ≤ jk , and define c(xi m) = xi m/xjk .
We note that this assignment is different than the decomposition function from [HT02].
In particular, our assignment does not in general satisfy
c(xs c(xt m)) = c(xt c(xs m))
(the edge ideal of the complete 3-graph on n = 5 vertices provides a counterexample,
where set(x3 x4 x5 ) = {1, 2}).
We will use this new decomposition function c to encode the map of chain complexes ψ
involved in the mapping cone construction. We first set up some further notation. For any
generator m = xi1 xi2 · · · xid of I we decompose set(m) as
set(m) = A1 ∪ A2 ∪ · · · ∪ Ad ,
15
where
A` = {j ∈ [n] : i`−1 < j < i` such that we have xi1 · · · xi`−1 xj xi`+1 · · · xid ∈ I}.
Notation. Let α ⊂ set(m). For each 1 ≤ ` ≤ d, let s` denote the maximum element of
α ∩ A` , and s0` denote the maximum element of α ∩ (A` \{s` }). Let T (α) = {s1 , . . . , sd }.
Example 4.10. Consider again our running example
I = hx1 x2 , x1 x3 , x1 x5 , x2 x3 , x2 x5 , x3 x5 , x4 x5 i.
We see that for m = x2 x5 and α = set(x2 x5 ) = {1, 3}, we have T (α) = {1, 3}. On the other
hand, for m = x3 x5 and α = set(x3 x5 ) = {1, 2}, we get T (α) = {2}.
Remark 4.11. Let t, s ∈ A` for some ` and t > s. Then c(xs (xt m)) = c(xs m), since i`−1 <
s < t < j` . Here we again employ the shorthand notation c(xs (xt m)) = c(xs c(xt m)).
Remark 4.12. Note that (1) can be written as
d
X
`=1
`−1
(−1)
s
X̀
(−1)j+|σ1 |+···+|σ`−1 | σ1 × σ2 × · · · × {σ`1 , σ`2 , . . . , σc
`j , . . . , σ`k } × · · · × σd
j=1
+
d
X
(−1)`−1+|σ1 |+···+|σ` | σ1 × σ2 × · · · × σ` \ max(σ` ) × · · · × σd .
`=1
|A` |>0
Recall from Lemma 4.8 that each σ1 × σ2 × · · · × σd corresponds to a symbol (m; α). The
first summand above is taken over all elements of α and the second summand is over the
indices of the elements of supp(m). The second summand can be also considered over the
elements of T (α) since in case that |T (α) ∩ A` | > 0 we have that c(xs` m) = xs` m/xmax(σ` ) .
Theorem 4.13. Let I = IH be the monomial edge ideal associated to a cointerval d-graph H,
and F the graded minimal free resolution of R/I obtained from the homomorphism complex
XH . Then F is realized as an iterated mapping cone, with the basis for each module Fi as
above. The chain map d of F is given by
X
X
xj i m
(c(xji m); α \ ji )
d(m; α) =
(−1)i xji (m; α \ ji ) +
(−1)i−1
c(xji m)
j ∈α
ji ∈T (α)
i
if σ 6= ∅, where α = {j1 , . . . , jp } ⊂ set(m) with j1 < · · · < jp , and d(m, ∅) = m otherwise.
Proof. We follow the strategy of the proof of Theorem 1.12 in [HT02]. Let I = IH be a cointerval interval ideal associated to the cointerval d-graph H, and let XH be its homomorphism
complex. We show by induction on j that the complex F (j) has the desired boundary map.
By definition F (j+1) is the mapping cone of ψ (j) : K (j) → F (j) so that F (j) is a subcomplex
of F (j+1) and hence it is enough to check the formula on the basis elements (m, α).
The definition of the mapping cone differential gives d(mj+1 ; α) = −dK (mj+1 ; α)+ψ (j) (mj+1 ; α),
where dK is the differential of the relevant Koszul complex K (j) .
Hence it is enough to show that we can define ψ (j) according to
X
xji mj+1
ψ (j) (mj+1 ; α) =
(−1)i−1
(c(xji mj+1 ); α \ ji ),
c(xji mj+1 )
ji ∈T (α)
16
x3x5
2
x2x5
1
x1x5
x1x5
x2x3
x1x3
x1x2
x1x3
x2x3
1
2
x1x3
x1x2
x4x5
x1x2
x1x5
2
3
x2x5
3
1
x3x5
3
x1x2
2
x1x5
1
x2x5
x1x3
x2x3
x1x5
x1x3
x2x3
x1x2
x1x3
x1x2
Figure 3. The resolution of I = hx1 x2 , x1 x3 , x1 x5 , x2 x3 , x2 x5 , x3 x5 , x4 x5 i
built from an iterated mapping cone as a realization of the homomorphism
complex.
if α 6= ∅, and otherwise ψ (j) (mj+1 ; ∅) = mj+1 .
For this we must verify that ψ (j) ◦ dK = d ◦ ψ (j) . To simplify notation we let m = mj+1
and ψ = ψj+1 .
For t ∈ set(m) a singleton element we have
ψ ◦ dK
(m; {t}) = ψ xt (m; ∅) = xt m,
and on the other hand
xt m
xt m
(c(xt m); ∅) =
c(xt m) = xt m.
c(xt m)
c(xt m)
For larger subsets of set(m), next consider σ ⊂ set(m) with |σ| ≥ 2. In this case we have
d ◦ ψ (m; {t}) = d
X
ψ ◦ dk (m; α) =
(−1)(α;t) xt ψ (m; α\t)
t∈α
(2)
=
X
=
X X
(−1)(α;t) xt (
t∈α
X
(−1)(α\t;s)
s∈T (α\t)
(−1)(α;t)+(α\t;s)
t∈α s∈T (α\t)
17
xs m
(c(xs m); α\{s, t})
c(xs m)
xt xs m
(c(xs m); α\{s, t}).
c(xs m)
Here we use the notation (α; t) = |{s ∈ α : s < t}|.
The other composition gives us
X
xt m
d ◦ ψ) (m; α) =
(−1)(α;t)
d (c(xt m); α\t) ,
c(xt m)
(3)
t∈T (α)
where
X
d (c(xt m); α\t) =
(−1)(α\t;s) xs (c(xt m); α\{s, t})
s∈α\t
X
+
xs c(xt m)
c(xs c(xt m)); α\{s, t} .
c(xs c(xt m))
(−1)(α\t;s)
s∈T (α\t)
Therefore Equation 3 becomes
X
(4)
(−1)(α;t)
t∈T (α)
X
xt m
xs c(xt m)
(−1)(α\t;s)
c(xs c(xt m)); α\{t, s}
c(xt m)
c(xs c(xt m))
s∈T (α\{t})
X
xt m X
+
(−1)(α;t)
(−1)(α\t;s) xs c(xt m); α\{t, s} .
c(xt m)
t∈T (α)
s∈α\{t}
Now note that the first summand expressing d ◦ ψ) (m; α) in Equation (4) can be written
X
X
xs xt m
c(xs c(xt m)); α\{t, s} ,
(−1)(α;t)+(α\t;s)
c(xs c(xt m))
t∈T (α) s∈T (α\{t})
which can be expanded as
d XX
X
`=1 t=s`
(5)
+
(−1)(α;t)+(α\t;s)
s=sk
k6=`
d XX
X
(−1)(α;t)+(α\t;s)
`=1 t=s` s=s0`
xs xt m
c(xs c(xt m)); α\{t, s}
c(xs c(xt m))
xs xt m
c(xs c(xt m)); α\{t, s} .
c(xs c(xt m))
Exchanging the role of ` and k in the first summand of 5 shows that it is ‘zero’.
The second summand of 4 can be written
X X
xt xs m
(−1)(α;t)+(α\t;s)
c(xt m); α\{t, s} ,
c(xt m)
t∈T (α) s∈α\{t}
and after applying the observation from Remark 4.11 we obtain
(6)
d XX
X
xs x t m
d ◦ ψ) (m; α) =
(−1)(α;t)+(α\t;s)
c(xs m); α\{t, s}
c(xs m)
`=1 t=s` s=s0`
X X
xt xs m
+
(−1)(α;t)+(α\t;s)
c(xt m); α\{t, s} .
c(xt m)
t∈T (α) s∈α\{t}
18
Next we compare the indices
appearing in the non-zero summands
corresponding to Equa
tion 2 for ψ ◦ dk ) (m; α) and Equation
6
for
d
◦
ψ)
(m;
α)
,
The
indices
appearing in the
non-zero summands of ψ ◦ dk ) (m; α) consist of
[
[
(t, sk ) : t ∈ A` , ` 6= k
(t, s` ) : t ∈ A` \s`
(s` , s0` )}.
Recall our notation, here sk denotes the maximum element of α ∩ Ak , and s0` denotes the
maximum element of α ∩ A` \s` .
On the other hand the indices appearing in the non-zero summands of d ◦ ψ) (m; α)
consist of
[
[
(sk , j) : j ∈ A` , ` 6= k
(s` , j) : j ∈ A` \s` .
(s` , s0` )
Exchanging the roles of s and t completes the proof.
Remark 4.14. The cells that serve as basis elements of the free modules in the resolution
constructed in Theorem 4.13 can also be obtained by gluing together the simplices ch(m, α, σ)
corresponding to the different choices of permutations σ of α from the set A, where
A = {σ = (σ1 , . . . , σp ) : if σi , σj ∈ A` for some ` and σi > σj then i < j}.
For this we define the cell U (m, α) as the union over all permutations σ ∈ A of α.
U (m, α) =
[
ch(m, α, σ).
σ∈A
4.2. Other regular decomposition functions. Describing all possible cellular realizations of the mapping cone resolution of an ideal I with linear quotients seems to be difficult
(even for a fixed ordering of the generators). Perhaps a more manageable task would be to
restrict to those cellular resolutions one obtains from regular decomposition functions.
For an ideal I with linear quotients and a regular decomposition function, Herzog and
Takayama in [HT02, Theorem 1.12] provide an explicit description of the differential map
of the resolution of an ideal I obtained from an iterated mapping cone. One can check
that the proof of the theorem (and the related lemmas) relies only on the regularity of
the decomposition function b, and not the definition of b in terms of assigning a particular
generator of I to every monomial. Similarly, our cellular realization in Theorem 3.10 only
relies on the regularity property. Hence we can vary the decomposition function in each step
and apply those results to obtain combinatorially distinct geometric complexes that support
resolutions, as the next example illustrates.
Example 4.15. Consider the ideal given by I = hx1 x2 , x1 x3 , x1 x5 , x2 x3 , x2 x5 , x3 x5 , x4 x5 i.
One can check that I has linear quotients with respect to this ordering. We construct a
resolution of I via the iterated mapping cone procedure, and by choosing different regular decomposition functions at each step we arrive at a family of combinatorially distinct
complexes supporting the resolution.
19
45
45
35
35
25
23
15
13
12
35
25
23
15
23
15
13
13
12
45
35
15
12
25
45
35
23
13
12
45
25
45
35
25
15
23
13
12
25
15
23
13
12
Figure 4. The family of cellular resolutions obtained by considering all
possible regular decompositions of the ideal I with linear quotient order
(x1 x2 , x1 x3 , x1 x5 , x2 x3 , x2 x5 , x3 x5 , x4 x5 ). Here we use the notation ij to denote the variable xi xj . The first complex is the generalized EK resolution
from Section 3 and the last is the homomorphism complex of Section 4.
4.3. Further questions - a space of resolutions? In each step of the mapping cone
construction we have a choice of homomorphism of complexes ψ : G → F that lifts the map
of R-modules R/(Ij−1 : fj ) → R/Ij−1 . As we have seen, this choice of homomorphism can
be encoded in the cellular structure of the Koszul simplex that is glued onto the previously
constructed resolution. After fixing a basis the collection of all such choices of bluings forms
a finite set, but can we understand them as comprising some geometric object and hence
obtain a ‘space of mapping cone resolutions’ ?
We note that if I = hx1 , x2 , . . . , xn id is a power of the graded maximal ideal, a certain space
of cellular resolutions of I is in fact described in [DJS12]. In this context a cellular resolution
of I is obtained by a generic arrangement of tropical hyperplanes, which in turn corresponds
to a regular triangulation of the product of simplices ∆n−1 × ∆d−1 . The collection of all
regular triangulations of ∆n−1 × ∆d−1 (or any polytope) has a natural polyhedral structure
known as a secondary polytope.
A natural question arrises: if we fix the linear quotient order on the generators of an ideal
I, what are the possible combinatorial types of complexes that we see as we attach a simplex
in each step of the cellular mapping cone construction? How many different choices do we
have to glue in the simplex? At one extreme sits the maximal ideal I = hx1 , x2 , . . . xn i, where
20
we have no choice but to build another simplex of one more dimension in each step. The
space of resolutions in this case is a single point. As seen in [DJS12], already for the square
of the maximal ideal in 4 variables we see distinct complexes arising.
References
[BBK12] Benjamin Braun, Jonathan Browder, and Steven Klee, Cellular resolutions of ideals defined by
nondegenerate simplicial homomorphisms, Israel Journal of Mathematics (2012), 1–24. ↑4.6
[BK06] Eric Babson and Dmitry N. Kozlov, Complexes of graph homomorphisms, Israel Journal of Mathematics 152 (2006), no. 1, 285–312. ↑4.6
[BS98] Dave Bayer and Bernd Sturmfels, Cellular resolutions of monomial modules, J. Reine Angew.
Math. 502 (1998), 123–140. MR1647559 (99g:13018) ↑1, 2.2
[BW02] Ekkehard Batzies and Volkmar Welker, Discrete Morse theory for cellular resolutions, J. Reine
Angew. Math. 543 (2002), 147–168. MR1887881 (2003b:13017) ↑2.2
[CN09] Alberto Corso and Uwe Nagel, Monomial and toric ideals associated to Ferrers graphs, Transactions of the American Mathematical Society 361 (2009), no. 3, 1371–1395. ↑1, 2.2, 4.1
[DE12] Anton Dochtermann and Alexander Engström, Cellular resolutions of cointerval ideals, Mathematische Zeitschrift 270 (2012), no. 1-2, 145–163. ↑1, 2.2, 4.1, 4.1, 4.7
[DJS12] Anton Dochtermann, Michael Joswig, and Raman Sanyal, Tropical types and associated cellular
resolutions, Journal of Algebra 356 (2012), no. 1, 304–324. ↑2.2, 4.3
[DK73] Gopal Danaraj and Victor Klee, Shellings of spheres and polytopes, DTIC Document, 1973. ↑3.2
[EHM11] Viviana Ene, Jürgen Herzog, and Fatemeh Mohammadi, Monomial ideals and toric rings of Hibi
type arising from a finite poset, Eur. J. Comb. 32 (2011), no. 3, 404–421. ↑2.1
[EK90] Shalom Eliahou and Michel Kervaire, Minimal resolutions of some monomial ideals, J. Algebra
129 (1990), no. 1, 1–25. MR1037391 (91b:13019) ↑1, 2.1, 2.1
[EOS10] Viviana Ene, Anda Olteanu, and Loredana Sorrenti, Properties of lexsegment ideals, Osaka J.
Math. 47 (2010), no. 1, 67–87. ↑2.1
[Goo13] Afshin Goodarzi, Cellular structure for the Herzog–Takayama Resolution, 2013. Preprint available
at arXiv:1305.4302. ↑1
[HT02] Jürgen Herzog and Yukihide Takayama, Resolutions by mapping cones, Homology, Homotopy and
Applications 4 (2002), no. 2, 277–294. ↑1, 2.1, 2.1, 2.1, 2.3, 2.1, 2.6, 2.1, 4.1, 4.1, 4.2
[Mer10] Jeffrey Mermin, The Eliahou-Kervaire resolution is cellular, Journal of Commutative Algebra 2
(2010), no. 1, 55–78. ↑1, 1, 2.2, 3
[MKM10] Fatemeh Mohammadi, Dariush Kiani, and Somayeh Moradi, On the Betti numbers of the edge
ideals of some hypergraphs, Contributions to Algebra and Geometry 51 (2010), no. 2, 467–475.
↑4.1, 4.1, 4.1, 4.4
[NPS02] Isabella Novik, Alexander Postnikov, and Bernd Sturmfels, Syzygies of oriented matroids, Duke
Math. J. 111 (2002), no. 2, 287–317. MR1882136 (2003b:13023) ↑2.2
[NR09] Uwe Nagel and Victor Reiner, Betti numbers of monomial ideals and shifted skew shapes, Electron.
J. Combin 16 (2009), no. 2, 59. ↑1, 2.2, 4.1
[Sin08] Achilleas Sinefakopoulos, On Borel fixed ideals generated in one degree, Journal of Algebra 319
(2008), no. 7, 2739–2760. ↑1, 2.2, 4.1
[Tay66] Diana K. Taylor, Ideals generated by monomials in an R-sequence, Ph.D. thesis, University of
Chicago, Department of Mathematics, 1966. ↑2.2
[Vel08] Mauricio Velasco, Minimal free resolutions that are not supported by a CW-complex, Journal of
Algebra 319 (2008), no. 1, 102–114. ↑2.2
Department of Mathematics, University of Miami, USA
E-mail address: [email protected]
Institut für Mathematik, Universität Osnabrück, Germany
E-mail address: [email protected]
21
| 0math.AC
|
MapReduce Algorithms for Inferring Gene Regulatory Networks from
Time‐Series Microarray Data Using an Information‐Theoretic Approach
Yasser Abduallah, Turki Turki, Kevin Byron, Zongxuan Du, Miguel Cervantes-Cervantes
Jason T. L. Wang
Computer Science Department, New Jersey Institute of Technology, Newark, NJ 07102, USA
1
Abstract
Gene regulation is a series of processes that control gene expression and its extent. The
connections among genes and their regulatory molecules, usually transcription factors, and a
descriptive model of such connections, are known as gene regulatory networks (GRNs).
Elucidating GRNs is crucial to understand the inner workings of the cell and the complexity of
gene interactions. To date, numerous algorithms have been developed to infer gene regulatory
networks. However, as the number of identified genes increases and the complexity of their
interactions is uncovered, networks and their regulatory mechanisms become cumbersome to
test. Furthermore, prodding through experimental results requires an enormous amount of
computation, resulting in slow data processing. Therefore, new approaches are needed to
expeditiously analyze copious amounts of experimental data resulting from cellular GRNs. To
meet this need, cloud computing is promising as reported in the literature. Here we propose new
MapReduce algorithms for inferring gene regulatory networks on a Hadoop cluster in a cloud
environment. These algorithms employ an information-theoretic approach to infer GRNs using
time-series microarray data. Experimental results show that our MapReduce program is much
faster than an existing tool while achieving slightly better prediction accuracy than the existing
tool.
2
1. Introduction
Current biotechnology has allowed researchers in various fields to obtain immense amounts of
experimental information, ranging from macromolecular sequences, gene expression data to
proteomics and metabolomics. In addition to large-scale genomic information obtained through
such methods as third generation DNA sequencing, newer technology, such as RNA-seq and
ChIP-seq, has allowed researchers to fine tune the analysis of gene expression patterns [1-3].
More information on interactions between transcription factors and DNA, both qualitative and
quantitative, is increasingly emerging from microarray data [4-6]. Although microarrays alone
do not provide direct evidence of functional connections among genes, the attachment of
transcription factors (TFs) and their binding sites (TFBSs), located at specific gene promoters,
influences transcription and modulates RNA production from a particular gene, thus establishing
a first level of functional interaction. Since the TFs are gene-encoded polypeptides and the target
TFBSs belong to different genes, analyses of TFs-TFBSs interactions could reveal gene
networks and may even contribute to elucidate unknown GRNs [7]. Besides contributing to infer
and understand these interactions, determining GRNs also provides models of such connections
[8]. GRNs could be the basis to infer more complex networks, encompassing gene, protein, and
metabolic spaces, as well as the entangled and often overlooked signaling pathways that
interconnect them [9-13].
The core GRN apparatus consists of the sum of cis-regulatory modular DNA sequence elements
that interact with TFs. These sequences read and process information incoming from the cell,
transducing it into the formation of gene products while modulating their abundance [14]. To
make them easier to understand, GRN models must be genome-oriented and viewable at
different levels, from global patterns of gene expression, down to cis-regulatory DNA and
nucleotide sequences [15].
Interactions among genes are mediated by gene products such as DNA-binding proteins
(including TFs) and miRNAs. The analyses of gene interactions can be difficult if time-series
data are part of the experimental design [16]. Analysis of genes, gene products and metabolism
(the Three Spaces of gene networks) would require additional computing resources. Among the
previously ignored components of gene networking are miRNAs [17, 18]. In addition to their
3
importance as regulatory elements in gene expression, the capacity of miRNAs to be transported
from cell to cell implicates them in a panoply of pathophysiological processes that include
antiviral defense, tumorigenesis, lipometabolism and glucose metabolism [16]. This role in
disease complicates our understanding of translational regulation via endogenous miRNAs. In
addition, miRNAs seem to be present in different types of foods [19] with potential implications
on human health and disease. Understanding the biogenesis, transport and mechanisms of action
of miRNAs on their target RNA would result in possible therapies, requiring large amounts of
computational power, which can be attained by cloud computing and process parallelizing.
Detailed experimental analysis of several functional regulatory elements has revealed that they
consist of dense clusters of unique, short DNA sequences specifically recognized by a range of
TFs. Biochemically, protein-binding to these sequences controls the regulatory output of the
clusters and, from an informational perspective, clustered specific target sites determine the type
of regulatory outcome and the cellular functions that will be performed. GRNs are encoded in the
DNA and can be thought of as a sequence-dependent regulatory genome, given that TFs
recognize specific short DNA motifs. The small length of these motifs means that they will occur
frequently but randomly within the enormity of the total genome of a particular organism [2022]. Therefore, to parse functional regulatory elements using bioinformatics requires the analysis
of copious amounts of genomic data.
Analyses of time-series data from microarrays can show the chronological expression of specific
genes or groups of genes. These temporal patterns can be used to infer or propose causal
relationships in gene regulation [23]. Thus, genes in logical networks can modulate the extent of
each other’s gene expression over the life span of a cell or a whole organism. Time-series
microarray data sheds light on a complex but measurable regulatory system, allowing for a more
precise inference of gene interaction.
Numerous algorithms have been developed for inferring GRNs [24-27]. In this paper, we present
a new approach, tailored to cloud computing, to infer GRNs using time-series microarray data.
Using time as a variable in the analysis of GRNs permits the study of changes in cellular
phenotype, as opposed to a snapshot in a limited time frame that may reveal static interactions
but not progression of gene expression phenomena. The time-series datasets used in this work
4
come from DREAM4 challenges [28-31] and an Affymetrix Yeast Genome 2.0 Array
downloaded from NCBI’s Gene Expression Omnibus. The array contains 5,744 probe sets and
includes 10,928 Saccharomyces cerevisiae genes with 49 time points and transcriptional
oscillations of about 4 hours. These oscillations reveal cell redox states, which in turn result from
changes in metabolic fluxes and cell cycle phases [32].
As knowledge in several biological fields leads to an ever-expanding accumulation in gene
expression data, the main consideration in data processing is that analysis of information
becomes increasingly time-consuming, thus creating a demand to speed up the analytical
process. In order to obtain results more expeditiously, we develop information-theoretic
algorithms using MapReduce that run on a distributed, multi-node Apache Hadoop cluster in a
cloud environment. Cloud resources are increasingly more flexible and affordable compared with
local traditional computing resources. Cloud computing advantages in the field of bioinformatics
research are well known [33-39].
2. Materials and Methods
2.1. Framework
Previous information-theoretic algorithms for network inference were implemented in the R
programming language using steady state data [40] and time-series data [23]. The tool using
steady state data is named ARACNE [40] and the tool using time-series data is named
TimeDelay-ARACNE [23]. ARACNE infers an undirected network, which basically shows
whether two genes are mutually dependent rather than the regulatory relationship between the
genes. In contrast, TimeDelay-ARACNE infers a directed network, in which an edge from gene
A to gene B indicates that A regulates the expression of B.
In contrast to the R-based ARACNE and TimeDelay-ARACNE, our proposed informationtheoretic framework is tailored to the MapReduce programming paradigm. Like TimeDelayARACNE [23], the input of our framework is a set of time-series gene expression data and the
output is an inferred gene regulatory network. The input dataset contains m genes, and each gene
has n expression values recorded at n different time points respectively. Our framework consists
of three steps. Step 1 aims to detect, for each gene g, the first time point t ( t 1 ) at which a
5
substantial change in the gene expression of g with respect to the gene expression of g at time
point 1 takes place. This t is referred to as the time point of Substantial Change of Expression
(ScE) of gene g, denoted ScE ( g ). Step 2 calculates, for two genes g x and g y , the influence of
g x on g y , denoted influence( g x , g y ), based on the ScE values of the genes. Step 3 determines the
edges between genes using their influence values. Below we present details of the proposed
framework.
Step 1. Calculation of ScE
Let g(t) be the expression value of gene g at time point t. We say g is activated (or induced) at
time point t ( t 1 ) if
g (t )
where 1 is a threshold. We say g is inhibited (or repressed) at
g (1)
time point t ( t 1 ) if
g (t ) 1
. For each gene g, we maintain two sets of time points: g (t ) and
g (1)
g (t ); g (t ) contains all time points at which g is induced and g (t ) contains all time points at
which g is repressed. Initially, g (t ) and g (t ) . The two sets of time points are then
updated as follows. For each time point t ( t 1 ),
If
1
if
g (t )
, then g (t ) g (t ) {t},
g (1)
if
g (t ) 1
, then g (t ) g (t ) {t}.
g (1)
g (t )
, then g is neither induced nor repressed at time point t. In this case, we simply
g (1)
ignore this time point t without adding t to g (t ) or g (t). The value of used in this study is set
to 1.2. With this threshold value and datasets used in the study (DREAM4 [28-31]), there is a
significant difference between the mean of the gene expression values of the time points at which
g is induced and the mean of the gene expression values of the time points at which g is
repressed according to Student’s t-test (p < 0.05).
Let ScE ( g ) represent the first time point t ( t 1 ) at which g is either induced or repressed, i.e.,
ScE( g ) min{g (t ) g (t )} .
6
For any two genes ga and gb , there are three cases to be considered.
Case 1: ScE ( ga ) ScE ( gb ). We send the ordered pair ( ga , gb ) and the expression values of the
two genes to step 2.
Case 2: ScE ( gb ) ScE ( ga ). We send the ordered pair ( gb , ga ) and the expression values of the
two genes to step 2.
Case 3: ScE ( ga ) ScE ( gb ). We send ga , gb with a tag indicating both of the ordered pairs ( ga ,
gb ) and ( gb , ga ) should be considered, together with their gene expression values.
Step 2. Calculation of influence values
For each pair of genes ( g x , g y ) received from step 1, we calculate the time-delayed mutual
information [41] between the genes as follows:
I
k
(k )
(gx , g y )
1 i n k
p ( g ix ,
g iy k ) log
i
i k
p(g x , g y
i
)
i k
p(g x ) p(g y
)
(1)
where n is the total number of time points, p( g ix ) is the marginal distribution of g ix , and
p( g ix , g iyk ) is the joint distribution of g ix and g iy k . (In our implementation, a hash table is used
to calculate the joint distribution to save time and space.) The parameter k, 1 k h, represents
the length of delayed time and h is the maximum length of delayed time. (In the study presented
here, h is set to 3.) The notation g ix denotes the gene expression of g x at time point i and g iy k
is the gene expression of g y at time point i + k. Figure 1 illustrates how to calculate time-delayed
mutual information. There are 21 time points in Figure 1. The length of delayed time is 2 (i.e., k
= 2). Each rectangle represents the gene expression value obtained at some time point. Mutual
7
Figure 1: Illustration of how to calculate time-delayed mutual information.
information of rectangles with the same color is computed. Then, the influence of g x on g y is
calculated as follows:
k
(k )
influence ( g x , g y ) max { I ( g x , g y )} .
1 k h
(2)
Referring to the three cases in step 1, for case 1, we calculate influence( ga , gb ), and send ( ga , gb )
and influence( ga , gb ) to step 3. For case 2, we calculate influence( gb , ga ), and send ( gb , ga ) and
influence( gb , ga ) to step 3. For case 3, if influence( ga , gb ) influence( gb , ga ), then we send ( ga ,
gb ) and influence( ga , gb ) to step 3; otherwise we send ( gb , ga ) and influence( gb , ga ) to step 3.
Step 3. Determination of edges between genes
Let be a threshold. For each pair ( g x , g y ) received from step 2, if influence( g x , g y ) then we
create an edge from g x to g y indicating g x substantially influences g y or g x regulates the
expression of gy , i.e., there is a predicted present edge from g x to g y . If influence( g x , g y )
then we do not create an edge from g x to gy , i.e., there is a predicted absent edge from g x to
g y . With the predicted present and absent edges, we are able to infer or reconstruct a gene
regulatory network. The value of used in this study is set to 0.96. With this threshold value
and datasets used in the study (DREAM4 [28-31]), there is a significant difference between the
8
mean of the influence values of the predicted present edges and the mean of the influence values
of the predicted absent edges according to Student’s t-test (p < 0.05).
2.2. MapReduce Algorithms
Figure 2 presents a high-level conceptual description of the Hadoop MapReduce implementation
of our proposed framework. The system includes a driver, and one or more mappers and
reducers. The driver takes the input from the user, including the locations of input and output
files, as well as algorithm parameters and thresholds. The driver prepares a job with the required
configuration, sends the job to Hadoop to start it, and calculates the time the job takes to
complete. The mappers are user-defined programs (UDPs), which prepare data and perform
calculations, if needed, and then send the processed data (key-value pairs) to the reducers. The
reducers are also UDPs, which perform the final processing and write the results into the output
file. Hadoop optimizes the number of mappers for a job. The user can control the number of
reducers needed for completing the job. The Hadoop distributed file system (HDFS) is a global
repository for storage of the input flat file (in plain text format) with gene expression data and
the output file with an inferred gene regulatory network.
Each gene has an identifier. (We use g x to represent both a gene and its identifier when the
context is clear.) Each line in the input file contains a pair of genes and their expression values.
Genes are sorted based on their identifiers. Each pair of genes g x , g y occurs in the input file only
once; specifically, the gene pair in which the identifier of g x is less than the identifier of g y
occurs in the input file. Suppose there are m genes. There are (m × m−1)/2 lines in the input file.
We develop four MapReduce algorithms, named M0, M1, M2, and M3 respectively. These
algorithms differ in which steps, as described in Section 2.1, are performed by mappers.
Algorithm M0. In this algorithm, mappers perform zero steps. Reducers have to do steps 1, 2
and 3. In the key-value pairs transmitted between mappers and reducers, the key is a pair of
genes, and the value contains the expression profiles of the genes.
9
Figure 2: Conceptual description of the Hadoop MapReduce implementation of our proposed
algorithms.
Algorithm M1. In this algorithm, mappers perform step 1. Reducers have to do steps 2 and 3. In
the key-value pairs transmitted between mappers and reducers, the key is a pair of genes and the
value contains the expression profiles of the genes. Part of the value is a tag indicating which
case in step 1 applies to the pair of genes.
Algorithm M2. In this algorithm, mappers have to do steps 1 and 2. Reducers perform step 3. In
the key-value pairs transmitted between mappers and reducers, the key is an ordered pair of
genes ( g x , g y ) and the value is influence( gx , g y ).
Algorithm M3. In this algorithm, mappers have to do steps 1, 2 and 3. Reducers perform zero
steps. In the key-value pairs transmitted between mappers and reducers, the key is the edge
gx
g y and the value is the influence of g x on g y that exceeds the threshold .
10
The time needed by mappers is bounded by O(m2 / M ) and the time needed by reducers is
bounded by O(m2 / R), where m is the number of genes, M is the number of mappers and R is the
number of reducers. Thus, the time complexity of our MapReduce algorithms is O(m2 / M ) +
O(m2 / R) . Note that this is a very pessimistic upper bound since reducers often work in parallel
with mappers, and hence the actual time needed by the algorithms is much less. Note also that, in
practice, M > R, and hence the time complexity of our algorithms is bounded by O(m2 / R) .
3. Results and Discussion
3.1. Experimental Results
The four algorithms described in Section 2.2 were implemented in MapReduce and Java on a
Hadoop infrastructure (cloud), which is a virtual environment based on VMware Big Data
Extensions (BDE). The infrastructure hardware cluster associated with BDE is comprised of two
IBM iDataPlex dx360 M4 servers. Each dx360 M4 server is comprised of two Intel Xeon
2.7GHz E5-2680 (8 Core) CPUs for a total of 16 cores per server. With the enabling of
hyperthreading, the number of logical processors is doubled to provide 32 logical processors per
server. Each server has 128 GB RAM.
The dataset used in the experiments was GSE30052 [32], downloaded from the Gene Expression
Omnibus (GEO) at http://www.ncbi.nlm.nih.gov/geo/. This dataset was created using an
Affymetrix Yeast Genome 2.0 Array containing 5,744 probe sets for S. cerevisiae gene
expression analysis. The dataset contains 10,928 genes with 49 time points. The dataset is split
into key-value pairs as described in Section 2.2 and the input file has (10,928 × 10,927)/2 lines,
taking up 26.8 GB of disk space. Hadoop assigns 254 mappers to this dataset. The default value
for the number of reducers is set to 20.
We divided GSE30052 into smaller datasets that were subsets of GSE30052 with varying
numbers of genes. Figure 3 compares the running times of the four MapReduce algorithms
described in Section 2.2 for varying dataset sizes. It can be seen from the figure that all the four
algorithms scale well when the dataset size becomes large (i.e., the number of genes increases).
Algorithm M2 performs the best. This happens because with M2, the mappers, working in
11
Figure 3: Performance comparison of the four MapReduce algorithms.
parallel, share some workload with the reducers, which perform a relatively smaller amount of
computation while writing results into the output file. It is worth noting that M2 is better than
M3, in which the mappers have to do all the computation. Algorithm M0 performs the worst.
With M0, the reducers have to do all the calculations and become too busy to quickly complete
the job.
We then fixed the algorithm and used M2 in all subsequent experiments. Figure 4 shows running
times of the M2 algorithm with 1, 20, and 100 reducers respectively. It can be seen that the
optimal number of reducers is 20. With this configuration, the reducers work at their maximum
limit. When there are too many (e.g., 100) reducers, the overhead is too large, and as a
consequence the system is slowed down. On the other hand, when only one reducer is employed,
the reducer is overloaded and the overall performance of the system degrades.
Finally, we conducted experiments to compare the MapReduce implementation of the M2
algorithm running on the Hadoop cluster (denoted MRC), the MapReduce implementation of the
M2 algorithm running on a standalone single-node server (denoted MRS), the Java
implementation of the M2 algorithm running on a single-node server, and the R implementation
12
Figure 4: Effect of the number of reducers on the performance of the M2 algorithm.
of the related time-delayed mutual information algorithm, TimeDelay-ARACNE [23]. In Figure
5, it can be seen that MRC is highly scalable and that it outperforms the other three programs.
Notably, due to Hadoop’s overhead, MRS is even slower than the Java program. The R program
is not scalable; its running time dramatically increases as the dataset becomes large.
3.2. Discussion
Our information-theoretic algorithms for inferring gene regulatory networks are implemented in
MapReduce and run on a Hadoop cluster. A tool that is closely related to our work is the
TimeDelay-ARACNE program in R [23], which also infers gene regulatory networks from timeseries gene expression profiles using an information-theoretic approach. As shown in Figure 5,
the TimeDelay-ARACNE program in R does not scale well whereas our MapReduce program is
highly scalable when running on the Hadoop cluster. Furthermore, our MapReduce program
differs from the TimeDelay-ARACNE R program in that our algorithm is deterministic
whereas the R program is implemented based on a non-deterministic algorithm, specifically
Markov random fields. For the same dataset and parameter values, the R program produces
different results in different executions. In contrast, our MapReduce program always produces
13
Figure 5: Performance comparison of the MapReduce algorithm (M2) and related programs.
the same result, in different executions, for the same dataset.
To evaluate the accuracy of these programs, we adopted the five time-series gene expression
datasets available in the DREAM4 100-gene in silico network inference challenge [28-31]. Each
dataset contains 10 times series, where each time series has 21 time points, for 100 genes. Thus,
in each time series, each gene has 21 gene expression values. Totally, there are 50 time series in
the five datasets. Each time-series dataset is associated with a gold standard file, where the gold
standard represents the ground truth of the network structure for the time-series data. Each edge
in the gold standard represents a true regulatory relationship between two genes. Our
experimental results showed that the average accuracy of TimeDelay-ARACNE on the five
datasets is approximately 92.4%. The average accuracy of our deterministic algorithm (M2) is
about 93.1%, which is slightly better than TimeDelay-ARACNE. It was observed that, in a
dataset, when M2 is better than TimeDelay-ARACNE, it has a higher accuracy than TimeDelayARACNE for every execution of TimeDelay-ARACNE on the dataset. Furthermore, our
MapReduce program (M2) takes, on average, 20 seconds to infer a gene regulatory network on a
DREAM4 dataset while the average time used by the TimeDelay-ARACNE R program is 3,500
seconds.
14
We also tested on different values for the parameters , , and the maximum length of delayed
time, h, used in the proposed algorithms. Experimental results showed that the default values for
these parameters ( 1.2, 0.96, h 3 ) achieve the highest accuracy. When compared with other
parameter values (e.g., the maximum length of delayed time h 6, 0.35 or 2 ), the accuracy
achieved by the default parameter values is significantly higher than the accuracy achieved by
the other parameter values according to Wilcoxon signed rank tests [42] (p < 0.05).
4. Conclusions
We have presented four MapReduce algorithms for reconstructing gene regulatory networks
from time-series microarray data using an information-theoretic approach. Our experimental
results showed that the algorithm (M2) that uses mappers to perform a large portion of work and
reducers to perform a relatively small amount of computation achieves the best performance.
This M2 algorithm is faster than an algorithm in which the mappers have to do all the
computation. Moreover, the M2 algorithm is much faster than another algorithm in which the
reducers have to do all the computation and become too busy to quickly complete the job.
When tested on DREAM4 datasets with 100 genes in each dataset, our MapReduce program
(M2) is slightly better than a closely related R program (TimeDelay-ARACNE [23]) in terms of
accuracy; furthermore, our MapReduce program is much faster than the existing R program.
When tested on a big dataset (GSE30052 [32]) with 10,928 genes, our MapReduce program was
found to be highly scalable whereas the R program was not (cf. Figure 5). It should be pointed
out, however, that the comparison with the TimeDelay-ARACNE R program is not completely
fair. Our MapReduce program is based on a parallel algorithm whereas the TimeDelayARACNE R program is based on a sequential algorithm. Further study would be needed to
investigate parallel versions of the R program or a new TimeDelay-ARACNE R package that
supports parallelization.
The work presented here shows that distributing highly parallel tasks in a cloud environment
achieves higher performance than running the tasks in a standalone or non-cloud environments.
In general, cloud computing can provide the power to integrate the ever-increasing information
about the Three Spaces of gene networks [8] as well as the multi-pronged signal transduction
15
pathways traversing these spaces. Comprehending systems biology and functional genomics
could eventually contribute to a better grasp of organismal physiology. Thus, the cloud would
provide computing power that is needed as the analysis of multilevel processes becomes more
complicated. Cloud computing will enable genome-scale network inference as demonstrated in
this study.
Epigenetics [43] is an emerging aspect of gene regulation whose study would require enormous
computing capacity. This type of posttranslational regulation cross-talk involves chemical
modifications of DNA and histones in a process known as chromatin remodeling. The role of
genetic and epigenetic networks in a variety of health conditions is now coming into view. For
example, there are at least 450 different genes associated with intellectual disability and related
cognitive disorders. Some of these genes are involved in synaptic plasticity and cell signaling
whereas others are epigenetic genes involved in chromatin modifications [44]. Analysis of the
interactions across these genes and networks, as well as finding new mutations, will require the
development of highly expeditious bioinformatics tools to mine the anticipated high amounts of
data.
Genome-scale metabolic models are becoming essential in biomedical applications, and
researchers are moving towards building such models [45]. MapReduce algorithms could
become a powerful tool in the analyses of all aspects of gene networking in the Three Spaces
paradigm. In general, cloud computing could facilitate the handling of the vast amounts of
information (big data) that such analyses require.
Acknowledgments
The authors thank N. Patel and L. Zhong for useful conversations concerning gene network
inference.
16
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
Schadt, E.E., S. Turner, and A. Kasarskis, A window into third-generation sequencing.
Hum Mol Genet, 2010. 19(R2): p. R227-40.
Angelini, C. and V. Costa, Understanding gene regulatory mechanisms by integrating
ChIP-seq and RNA-seq data: statistical solutions to biological problems. Front Cell Dev
Biol, 2014. 2: p. 51.
Qin, J., Y. Hu, F. Xu, H.K. Yalamanchili, and J. Wang, Inferring gene regulatory
networks by integrating ChIP-seq/chip and transcriptome data via LASSO-type
regularization methods. Methods, 2014. 67(3): p. 294-303.
Brown, J.B. and S.E. Celniker, Lessons from modENCODE. Annu Rev Genomics Hum
Genet, 2015. 16: p. 31-53.
Filkov, V., S. Skiena, and J. Zhi, Analysis techniques for microarray time-series data. J
Comput Biol, 2002. 9(2): p. 317-30.
Zhu, F., L. Shi, J.D. Engel, and Y. Guan, Regulatory network inferred using expression
data of small sample size: application and validation in erythroid system. Bioinformatics,
2015. 31(15): p. 2537-44.
Werner, T., S.M. Dombrowski, C. Zgheib, F.A. Zouein, H.L. Keen, M. Kurdi, and G.W.
Booz, Elucidating functional context within microarray data by integrated transcription
factor-focused gene-interaction and regulatory network analysis. Eur Cytokine Netw,
2013. 24(2): p. 75-90.
Brazhnik, P., A. de la Fuente, and P. Mendes, Gene networks: how to put the function in
genomics. Trends Biotechnol, 2002. 20(11): p. 467-72.
Ideker, T. and N.J. Krogan, Differential network biology. Mol Syst Biol, 2012. 8: p. 565.
Margolin, A.A., K. Wang, W.K. Lim, M. Kustagi, I. Nemenman, and A. Califano,
Reverse engineering cellular networks. Nat Protoc, 2006. 1(2): p. 662-71.
Kibinge, N., N. Ono, M. Horie, T. Sato, T. Sugiura, M. Altaf-Ul-Amin, A. Saito, and S.
Kanaya, Integrated pathway-based transcription regulation network mining and
visualization based on gene expression profiles. J Biomed Inform, 2016. 61: p. 194-202.
Altaf-Ul-Amin, M., F.M. Afendi, S.K. Kiboi, and S. Kanaya, Systems biology in the
context of big data and networks. Biomed Res Int, 2014. 2014: p. 428570.
Altaf-Ul-Amin, M., T. Katsuragi, T. Sato, and S. Kanaya, A glimpse to background and
characteristics of major molecular biological networks. Biomed Res Int, 2015. 2015: p.
540297.
Karlebach, G. and R. Shamir, Modelling and analysis of gene regulatory networks. Nat
Rev Mol Cell Biol, 2008. 9(10): p. 770-80.
Longabaugh, W.J., E.H. Davidson, and H. Bolouri, Visualization, documentation,
analysis, and communication of large-scale gene regulatory networks. Biochim Biophys
Acta, 2009. 1789(4): p. 363-74.
Xu, L., B.F. Yang, and J. Ai, MicroRNA transport: a new way in cell communication. J
Cell Physiol, 2013. 228(8): p. 1713-9.
Chen, K. and N. Rajewsky, The evolution of gene regulation by transcription factors and
microRNAs. Nat Rev Genet, 2007. 8(2): p. 93-103.
17
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
Zhong, L., J.T.L. Wang, D. Wen, V. Aris, P. Soteropoulos, and B.A. Shapiro, Effective
classification of microRNA precursors using feature mining and AdaBoost algorithms.
OMICS, 2013. 17(9): p. 486-93.
Wagner, A.E., S. Piegholdt, M. Ferraro, K. Pallauf, and G. Rimbach, Food derived
microRNAs. Food Funct, 2015. 6(3): p. 714-8.
Longabaugh, W.J., E.H. Davidson, and H. Bolouri, Computational representation of
developmental genetic regulatory networks. Dev Biol, 2005. 283(1): p. 1-16.
Levine, M. and E.H. Davidson, Gene regulatory networks for development. Proc Natl
Acad Sci U S A, 2005. 102(14): p. 4936-42.
Davidson, E.H. and D.H. Erwin, Gene regulatory networks and the evolution of animal
body plans. Science, 2006. 311(5762): p. 796-800.
Zoppoli, P., S. Morganella, and M. Ceccarelli, TimeDelay-ARACNE: reverse engineering
of gene networks from time-course data by an information theoretic approach. BMC
Bioinformatics, 2010. 11: p. 154.
Maetschke, S.R., P.B. Madhamshettiwar, M.J. Davis, and M.A. Ragan, Supervised, semisupervised and unsupervised inference of gene regulatory networks. Brief Bioinform,
2014. 15(2): p. 195-211.
Marbach, D., R.J. Prill, T. Schaffter, C. Mattiussi, D. Floreano, and G. Stolovitzky,
Revealing strengths and weaknesses of methods for gene network inference. Proc Natl
Acad Sci U S A, 2010. 107(14): p. 6286-91.
Patel, N. and J.T.L. Wang, Semi-supervised prediction of gene regulatory networks using
machine learning algorithms. J Biosci, 2015. 40(4): p. 731-40.
Turki, T. and J.T.L. Wang. A new approach to link prediction in gene regulatory
networks. in Proceedings of the 16th International Conference on Intelligent Data
Engineering and Automated Learning. 2015.
Greenfield, A., A. Madar, H. Ostrer, and R. Bonneau, DREAM4: combining genetic and
dynamic information to identify biological networks and dynamical models. PLoS One,
2010. 5(10): p. e13397.
Marbach, D., T. Schaffter, C. Mattiussi, and D. Floreano, Generating realistic in silico
gene networks for performance assessment of reverse engineering methods. J Comput
Biol, 2009. 16(2): p. 229-39.
Prill, R.J., D. Marbach, J. Saez-Rodriguez, P.K. Sorger, L.G. Alexopoulos, X. Xue, N.D.
Clarke, G. Altan-Bonnet, and G. Stolovitzky, Towards a rigorous assessment of systems
biology models: the DREAM3 challenges. PLoS One, 2010. 5(2): p. e9202.
Schaffter, T., D. Marbach, and D. Floreano, GeneNetWeaver: in silico benchmark
generation and performance profiling of network inference methods. Bioinformatics,
2011. 27(16): p. 2263-70.
Chin, S.L., I.M. Marcus, R.R. Klevecz, and C.M. Li, Dynamics of oscillatory phenotypes
in Saccharomyces cerevisiae reveal a network of genome-wide transcriptional
oscillators. FEBS J, 2012. 279(6): p. 1119-30.
Langmead, B., K.D. Hansen, and J.T. Leek, Cloud-scale RNA-sequencing differential
expression analysis with Myrna. Genome Biol, 2010. 11(8): p. R83.
Nguyen, T., W. Shi, and D. Ruden, CloudAligner: A fast and full-featured MapReduce
based tool for sequence mapping. BMC Res Notes, 2011. 4: p. 171.
18
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
Schatz, M.C., CloudBurst: highly sensitive read mapping with MapReduce.
Bioinformatics, 2009. 25(11): p. 1363-9.
Zhao, S., K. Prenger, L. Smith, T. Messina, H. Fan, E. Jaeger, and S. Stephens, Rainbow:
a tool for large-scale whole-genome sequencing data analysis using cloud computing.
BMC Genomics, 2013. 14: p. 425.
Mohammed, E.A., B.H. Far, and C. Naugler, Applications of the MapReduce
programming framework to clinical big data analysis: current landscape and future
trends. BioData Min, 2014. 7: p. 22.
Zou, Q., X.B. Li, W.R. Jiang, Z.Y. Lin, G.L. Li, and K. Chen, Survey of MapReduce
frame operation in bioinformatics. Brief Bioinform, 2014. 15(4): p. 637-47.
O'Driscoll, A., V. Belogrudov, J. Carroll, K. Kropp, P. Walsh, P. Ghazal, and R.D.
Sleator, HBLAST: Parallelised sequence similarity--A Hadoop MapReducable basic
local alignment search tool. J Biomed Inform, 2015. 54: p. 58-64.
Margolin, A.A., I. Nemenman, K. Basso, C. Wiggins, G. Stolovitzky, R. Dalla Favera,
and A. Califano, ARACNE: an algorithm for the reconstruction of gene regulatory
networks in a mammalian cellular context. BMC Bioinformatics, 2006. 7 Suppl 1: p. S7.
Fraser, A.M. and H.L. Swinney, Independent coordinates for strange attractors from
mutual information. Phys Rev A 1986. 33: p. 1134.
Song, Y., L. Hua, B.A. Shapiro, and J.T.L. Wang, Effective alignment of RNA pseudoknot
structures using partition function posterior log-odds scores. BMC Bioinformatics, 2015.
16: p. 39.
Li, E., Chromatin modification and epigenetic reprogramming in mammalian
development. Nat Rev Genet, 2002. 3(9): p. 662-73.
van Bokhoven, H., Genetic and epigenetic networks in intellectual disabilities. Annu Rev
Genet, 2011. 45: p. 81-104.
Zhang, C., B. Ji, A. Mardinoglu, J. Nielsen, and Q. Hua, Logical transformation of
genome-scale metabolic models for gene level applications and analysis. Bioinformatics,
2015. 31(14): p. 2324-31.
19
| 5cs.CE
|
arXiv:1706.08329v3 [cs.LO] 18 Dec 2017
Faculty of Computer Science
Institute of Artificial Intelligence
Knowledge Representation and Reasoning
The Boolean Solution Problem
from the Perspective of Predicate Logic
– Extended Version –
Christoph Wernhard
KRR Report 17-01
Mail to
Technische Universität Dresden
01062 Dresden
Office
Room 2006
Nöthnitzer Straße 46
01187 Dresden
Internet
http://www.wv.inf.tu-dresden.de
The Boolean Solution Problem
from the Perspective of Predicate Logic
– Extended Version –
Christoph Wernhard
Technische Universität Dresden
Abstract. Finding solution values for unknowns in Boolean equations
was a principal reasoning mode in the Algebra of Logic of the 19th century. Schröder investigated it as Auflösungsproblem (solution problem). It
is closely related to the modern notion of Boolean unification. Today it is
commonly presented in an algebraic setting, but seems potentially useful
also in knowledge representation based on predicate logic. We show that
it can be modeled on the basis of first-order logic extended by secondorder quantification. A wealth of classical results transfers, foundations
for algorithms unfold, and connections with second-order quantifier elimination and Craig interpolation show up. Although for first-order inputs
the set of solutions is recursively enumerable, the development of constructive methods remains a challenge. We identify some cases that allow
constructions, most of them based on Craig interpolation, and show a
method to take vocabulary restrictions on solution components into account.
Revision: December 18, 2017
2
TABLE OF CONTENTS
Table of Contents
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
2
Notation and Preliminaries . . . . . . . . . . . . . . . . . . . . . . . .
4
2.1
2.2
Notational Conventions . . . . . . . . . . . . . . . . . . . . . . .
Substitution with Terms and Formulas . . . . . . . . . . . . . . .
4
5
The Solution Problem from Different Angles . . . . . . . . . . . . . . .
6
3.1
3.2
3.3
3.4
.
.
.
.
6
8
8
9
The Method of Successive Eliminations – Abstracted . . . . . . . . . .
10
3
4
4.1
4.2
4.3
4.4
Basic Formal Modeling . . . . . . . . . . . . . .
View as Unification . . . . . . . . . . . . . . . .
View as Construction of Elimination Witnesses
View as Related to Definientia and Interpolants
.
.
.
.
.
.
.
.
.
.
.
.
Reducing n-ary to 1-ary Solution Problems . . . . .
Solving on the Basis of Second-Order Formulas . . .
Solving with the Method of Successive Eliminations
Solving by Inside-Out Witness Construction . . . . .
.
.
.
.
5.1
5.2
5.3
Conditions for the Existence of Solutions . . . . . . . . . . . . . .
Characterization of SOL-Witnessed in Terms of ELIM-Witness .
The Elimination Result as Precondition of Solution Existence . .
15
17
17
Reproductive Solutions as Most General Solutions . . . . . . . . . . .
18
6.1
6.2
6.3
6.4
.
.
.
.
19
23
26
28
Approaching Constructive Solution Techniques . . . . . . . . . . . . .
29
7.1
7.2
7.3
7.4
Background: Craig Interpolation, Definability and Independence
Cases Related to Definability and Interpolation . . . . . . . . . .
The EHW-Combination of ELIM-Witnesses for Disjuncts . . . .
Relational Monadic Formulas and Relaxed Substitutibility . . . .
30
31
32
34
Solutions in Restricted Vocabularies . . . . . . . . . . . . . . . . . . .
35
8.1
8.2
8.3
An Example: Synthesizing Definitional Equivalence . . . . . . . .
Modeling with Two Consecutive Solution Problems . . . . . . . .
Expressing a Vocabulary Restriction on all Unknowns . . . . . .
35
35
36
A Herbrand View on the Solution Problem . . . . . . . . . . . . . . .
36
10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
7
8
9
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
15
.
.
.
.
.
.
.
.
.
.
.
.
Existence of Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . .
Parametric, General and Reproductive Solutions
The Rigorous Solution . . . . . . . . . . . . . . .
Schröder’s Reproductive Interpolant . . . . . . .
From Unary to n-ary Reproductive Solutions . .
.
.
.
.
.
.
.
.
10
12
13
14
6
.
.
.
.
.
.
.
.
.
.
.
.
5
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Introduction
1
3
Introduction
Finding solution values for unknowns in Boolean equations was a principal reasoning mode in the Algebra of Logic of the 19th century. Schröder [34] investigated it as Auflösungsproblem (solution problem). It is closely related to the
modern notion of Boolean unification. For a given formula that contains unknowns formulas are sought such that after substituting the unknowns with
them the given formula becomes valid or, dually, unsatisfiable. Of interest are
also most general solutions, condensed representations of all solution substitutions. A central technique there is the method of successive eliminations, which
traces back to Boole. Schröder investigated reproductive solutions as most general solutions, anticipating the concept of most general unifier. A comprehensive
modern formalization based on this material, along with historic remarks, is
presented by Rudeanu [30] in the framework of Boolean algebra. In automated
reasoning variants of these techniques have been considered mainly in the late
80s and early 90s with the motivation to enrich Prolog and constraint processing
by Boolean unification with respect to propositional formulas handled as terms
[26,9,27,28,21,22]. An early implementation based on [30] has been also described
in [37]. An implementation with BDDs of the algorithm from [9] is reported in
[10]. The ΠP
2 -completeness of Boolean unification with constants was proven
only later in [21,22] and seemingly independently in [2]. Schröder’s results were
developed further by Löwenheim [24,25]. A generalization of Boole’s method beyond propositional logic to relational monadic formulas has been presented by
Behmann in the early 1950s [6,7]. Recently the complexity of Boolean unification in a predicate logic setting has been investigated for some formula classes,
in particular for quantifier-free first-order formulas [16]. A brief discussion of
Boolean reasoning in comparison with predicate logic can be found in [8].
Here we remodel the solution problem formally along with basic classical results and some new generalizations in the framework of first-order logic extended
by second-order quantification. The main thesis of this work is that it is possible
and useful to apply second-order quantification consequently throughout the formalization. What otherwise would require meta-level notation is then expressed
just with formulas. As will be shown, classical results can be reproduced in this
framework in a way such that applicability beyond propositional logic, possible algorithmic variations, as well as connections with second-order quantifier
elimination and Craig interpolation become visible. Of course, methods to solve
Boolean equations on first-order formulas do not necessarily terminate. However, the set of solutions is recursively enumerable. By the modeling in predicate
logic we try to pin down the essential points of divergence from propositional
logic. Special cases that allow solution construction are identified, most of them
related to definiens computation by Craig interpolation. In addition, a way to
express a generalization of the solution problem where vocabulary restrictions
are taken into account in terms of two related solution problems is shown.
The envisaged application scenario is to let solving “solution problems”, or
Boolean equation solving, on the basis of predicate logic join reasoning modes
like second-order quantifier elimination (or “semantic forgetting”), Craig interpo-
4
Section 2
lation and abduction to support the mechanized reasoning about relationships
between theories and the extraction or synthesis of subtheories with given properties. On the practical side, the aim is to relate it to reasoning techniques such
as Craig interpolation on the basis of first-order provers, SAT and QBF solving, and second-order quantifier elimination based on resolution [19] and the
Ackermann approach [15]. Numerous applications of Boolean equation solving
in various fields are summarized in [31, Chap. 14]. Applications in automated
theorem proving and proof compression are mentioned in [16, Sect. 7]. The prevention of certain redundancies has been described as application of (concept)
unification in description logics [4]. Here the synthesis of definitional equivalences
is sketched as an application.
The rest of the paper is structured as follows: Notation, in particular for
substitution in formulas, is introduced in Sect. 2. In Sect. 3 a formalization of
the solution problem is presented and related to different points of view. Section 4
is concerned with abstract properties of and algorithmic approaches to solution
problems with several unknowns. Conditions under which solutions exist are
discussed in Sect. 5. Adaptions of classical material on reproductive solutions
are given in Sect. 6. In Sect. 7 various techniques for solution construction in
particular cases are discussed. The solution problem with vocabulary restrictions
is discussed in Sect. 8. The solution problem is displayed in Sect. 9 as embedded
in a setting with Skolemization and Herbrand expansion. Section 10 closes the
paper with concluding remarks.
The material in Sect. 2–5 has also been published as [40].
2
2.1
Notation and Preliminaries
Notational Conventions
We consider formulas in first-order logic with equality extended by second-order
quantification upon predicates. They are constructed from atoms (including
equality atoms), constant operators >, ⊥, the unary operator ¬, binary operators ∧, ∨ and quantifiers ∀, ∃ with their usual meaning. Further binary operators →, ←, ↔, as well as n-ary versions of ∧ and ∨ can be understood as
meta-level notation. The operators ∧ and ∨ bind stronger than →, ← and ↔.
The scope of ¬, the quantifiers, and the n-ary connectives is the immediate subformula to the right. A subformula occurrence has in a given formula positive
(negative) polarity if it is in the scope of an even (odd) number of negations.
A vocabulary is a set of symbols, that is, predicate symbols (briefly predicates),
function symbols (briefly functions) and individual symbols. (Individual symbols
are not partitioned into variables and constants. Thus, an individual symbol is –
like a predicate – considered as variable if and only if it is bound by a quantifier.)
The arity of a predicate or function s is denoted by arity(s). The set of symbols
that occur free in a formula F is denoted by free(F ). The property that no
member of free(F ) is bound by a quantifier occurrence in F is expressed as
CLEAN(F ). Symbols not present in the formulas and other items under discussion
Notation and Preliminaries
5
are called fresh. The clean variant of a given formula F is the formula G obtained
from F by successively replacing all bound symbols with fresh symbols such that
CLEAN(G). The replacement is done in some way not specified here further such
that each formula has a unique clean variant.
We write F |= G for F entails G; |= F for F is valid ; and F ≡ G for F is
equivalent to G, that is, F |= G and G |= F .
We write sequences of symbols, of terms and of formulas by juxtaposition.
Their length is assumed to be finite. The empty sequence is written . A sequence
with length 1 is not distinguished from its sole member. In contexts where a set
is expected, a sequence stands for the set of its members. Atoms are written
in the form p(t), where t is a sequence of terms whose length is the arity of
the predicate p. Atoms of the form p(), that is, with a nullary predicate p, are
written also as p. For a sequence of fresh symbols we assume that its members
are distinct. A sequence p1 . . . pn of predicates is said to match another sequence
q1 . . . qm if and only if n = m and for all i ∈ {1, . . . , n} it holds that arity(pi ) =
arity(qi ). If s = s1 . . . sn is a sequence of symbols, then ∀s stands for ∀s1 . . . ∀sn
and ∃s for ∃s1 . . . ∃sn .
If F = F1 . . . Fn is a sequence of formulas,
then CLEAN(F ) states CLEAN(Fi )
Sn
for all i ∈ {1, . . . , n}, and free(F ) = i=1 free(Fi ). If G = G1 . . . Gn is a second
sequence of formulas, then F ≡ G stands for F1 ≡ G1 and . . . and Fn ≡ Gn .
As explained below, in certain contexts the individual symbols in the set
X = {xi | i ≥ 1} play a special role. For example in the following shorthands for a
predicate p, a formula F and x = x1 . . . xarity(p) : p ⇔ F stands for ∀x (p(x) ↔ F );
p 6⇔ F for ¬(p ⇔ F ); p ⇒ F for ∀x (p(x) → F ); and p ⇐ F for ∀x (p(x) ← F ).
2.2
Substitution with Terms and Formulas
To express systematic substitution of individual symbols and predicates concisely
we use the following notation:
– F (c) and F (t) – Notational Context for Substitution of Individual Symbols.
Let c = c1 . . . cn be a sequence of distinct individual symbols. We write F as
F (c) to declare that for a sequence t = t1 . . . tn of terms the expression F (t)
denotes F with, for i ∈ {1, . . . , n}, all free occurrences of ci replaced by ti .
– F [p], F [G] and F [q] – Notational Context for Substitution of Predicates. Let
p = p1 . . . pn be a sequence of distinct predicates and let F be a formula. We
write F as F [p] to declare the following:
• For a sequence G = G1 (x1 . . . xarity(p1 ) ) . . . Gn (x1 . . . xarity(pn ) ) of formulas
the expression F [G] denotes F with, for i ∈ {1, . . . , n}, each atom occurrence pi (t1 . . . tarity(pi ) ) where pi is free in F replaced by Gi (t1 . . . tarity(pi ) ).
• For a sequence q = q1 . . . qn of predicates that matches p the expression F [q] denotes F with, for i ∈ {1, . . . , n}, each free occurrence of pi
replaced by qi .
• The above notation F [S], where S is a sequence of formulas or of predicates, is generalized to allow also pi at the ith position of S, for example
F [G1 . . . Gi−1 pi . . . pn ]. The formula F [S] then denotes F with only those
6
Section 3
predicates pi with i ∈ {1, . . . , n} that are not present at the ith position
in S replaced by the ith component of S as described above (in the
example only p1 , . . . , pi−1 would be replaced).
– F [p] – Notational Context for Substitution in a Sequence of Formulas. If
F = F1 . . . Fn is a sequence of formulas, then F [p] declares that F [S], where
S is a sequence with the same length as p, is to be understood as the sequence
F1 [S] . . . Fn [S] with the meaning of the members as described above.
In the above notation for substitution of predicates by formulas the members
x1 , . . . , xarity(p) of X play a special role: F [G] can be alternatively considered as
obtained by replacing predicates pi with λ-expressions λx1 . . . λxarity(pi ) .Gi followed by β-conversion. The shorthand p ⇔ F can be correspondingly considered
as p ↔ λx1 . . . λxarity(p) .G. The following property substitutible specifies preconditions for meaningful simultaneous substitution of formulas for predicates:
Definition 1 (SUBST(G, p, F ) – Substitutible Sequence of Formulas).
A sequence G = G1 . . . Gm of formulas is called substitutible for a sequence
p = p1 . . . pn of distinct predicates in a formula F , written SUBST(G, p, F ), if
and only if m = n and for all i ∈ {1, . . . , n} it holds that (1.) No free occurrence
of pi in F is in the scope of a quantifier occurrence that binds a member of
free(Gi ); (2.) free(Gi ) ∩ p = ∅; and (3.) free(Gi ) ∩ {xj | j > arity(pi )} = ∅.
The following propositions demonstrate the introduced notation for formula substitution. It is well known that terms can be “pulled out of” and “pushed in to”
atoms,
justified by the equivalences p(t1 . .W. tn ) ≡ ∃x1 . . . ∃xn (p(x1 . . . xn ) ∧
Vn
n
x
i=1 i = ti ) ≡ ∀x1 . . . ∀xn (p(x1 . . . xn ) ∨ i=1 xi 6= ti ), which hold if no member of {x1 , . . . , xn } does occur in the terms t1 , . . . , tn . Analogously, substitutible
subformulas can be “pulled out of” and “pushed in to” formulas:
Proposition 2 (Pulling-Out and Pushing-In of Subformulas). Let G =
G1 . . . Gn be a sequence of formulas, let p = p1 . . . pn be a sequence of distinct
predicates and let F = F [p] be a formula such that SUBST(G, p, F ). Then
Vn
Wn
(i) F [G] ≡ ∃p (F ∧ i=1 (pi ⇔ Gi )) ≡ ∀p (F ∨ i=1 (pi 6⇔ Gi )).
(ii) ∀p F |= F [G] |= ∃p F.
Ackermann’s Lemma [1] can be applied in certain cases to eliminate second-order
quantifiers, that is, to compute for a given second-order formula an equivalent
first-order formula. It plays an important role in many modern methods for
elimination and semantic forgetting – see, e.g., [15,12,18,33,23,41]:
Proposition 3 (Ackermann’s Lemma, Positive Version). Let F, G be formulas and let p be a predicate such that SUBST(G, p, F ), p ∈
/ free(G) and all free
occurrences of p in F have negative polarity. Then ∃p ((p ⇐ G) ∧ F [p]) ≡ F [G].
3
3.1
The Solution Problem from Different Angles
Basic Formal Modeling
Our formal modeling of the Boolean solution problem is based on two concepts,
solution problem and particular solution:
The Solution Problem from Different Angles
7
Definition 4 (F [p] – Solution Problem (SP), Unary Solution Problem
(1-SP)). A solution problem (SP) F [p] is a pair of a formula F and a sequence p
of distinct predicates. The members of p are called the unknowns of the SP. The
length of p is called the arity of the SP. A SP with arity 1 is also called unary
solution problem (1-SP).
The notation F [p] for solution problems establishes as a “side effect” a context
for specifying substitutions of p in F by formulas as specified in Sect. 2.2.
Definition 5 (Particular Solution). A particular solution (briefly solution)
of a SP F [p] is defined as a sequence G of formulas such that SUBST(G, p, F )
and |= F [G].
The property SUBST(G, p, F ) in this definition implies that no member of p
occurs free in a solution. Of course, particular solution can also be defined on the
basis of unsatisfiability instead of validity, justified by the equivalence of |= F [G]
and ¬F [G] |= ⊥. The variant based on validity has been chosen here because
then the associated second-order quantifications are existential, matching the
usual presentation of elimination techniques.
Being a solution is aside of the substitutibility condition a semantic property,
that is, applying to formulas modulo equivalence: If G is a solution of F [p], then
all sequences H of formulas such that SUBST(H, p, F ) and G ≡ H (that is, if
G = G1 . . . Gn and H = H1 . . . Hn , then Hi ≡ Gi holds for all i ∈ {1, . . . , n})
are also solutions of F [p].
Solution problem and solution as defined here provide abstractions of computational problems in a technical sense that would be suitable, e.g., for complexity
analysis. Problems in the latter sense can be obtained by fixing involved formula
and predicate classes. The abstract notions are adequate to develop much of the
material on the “Boolean solution problem” shown here. On occasion, however,
we consider restrictions, in particular to propositional and to first-order formulas,
as well as to nullary predicates. As shown in Sect. 6, further variants of solution,
general representations of several particular solutions, can be introduced on the
basis of the notions defined here.
Example 6 (A Solution Problem and its Particular Solutions). As an
example of a solution problem consider F [p1 p2 ] where
F = ∀x (a(x ) → b(x )) →
(∀x (p1 (x ) → p2 (x )) ∧ ∀x (a(x ) → p2 (x )) ∧ ∀x (p2 (x ) → b(x ))).
The intuition is that the antecedent ∀x (a(x ) → b(x )) specifies the “background
theory”, and w.r.t. that theory the unknown p1 is “stronger” than the other
unknown p2 , which, in addition, is “between” a and b. Examples of solutions are:
a(x1 )a(x1 ); a(x1 )b(x1 ); ⊥a(x1 ); b(x1 )b(x1 ); and (a(x1 ) ∧ b(x1 ))(a(x1 ) ∨ b(x1 )).
No solutions are for example b(x1 )a(x1 ); a(x1 )⊥; and all members of {>, ⊥} ×
{>, ⊥}.
Assuming a countable vocabulary, the set of valid first-order formulas is recursively enumerable. It follows that for an n-ary SP F [p] where F is first-order the
8
Section 3
set of those of its particular solutions that are sequences of first-order formulas
is also recursively enumerable: An n-ary sequence G of well-formed first-order
formulas that satisfies the syntactic restriction SUBST(G, p, F ) is a solution of
F [p] if and only if F [G] is valid.
In the following subsections further views on the solution problem will be
discussed: as unification or equation solving, as a special case of second-order
quantifier elimination, and as related to determining definientia and interpolants.
3.2
View as Unification
Because |= F [G] if and only if F [G] ≡ >, a particular solution of F [p] can be
seen as a unifier of the two formulas F [p] and > modulo logical equivalence as
equational theory. From the perspective of unification the two formulas appear
as terms, the members of p play the role of variables and the other predicates
play the role of constants.
Vice versa, a unifier of two formulas can be seen as a particular solution,
justified by the equivalence of L[G] ≡ R[G] and |= (L ↔ R)[G], which holds
for sequences G and p of formulas and predicates, respectively, and formulas
L = L[p], R = R[p], (L ↔ R) = (L ↔ R)[p] such that SUBST(G, p, L) and
SUBST(G, p, R). This view of formula unification can be generalized to sets
with a finite cardinality k of equivalences, since for all i ∈ {1, . . . , k} it holds
Vk
that Li ≡ Ri can be expressed as |= i=1 (Li ↔ Ri ).
An exact correspondence between solving a solution problem F [p1 . . . pn ]
where F is a propositional formula with ∨, ∧, ¬, ⊥, > as logic operators and
E-unification with constants in the theory of Boolean algebra (with the mentioned logic operators as signature) applied to F =E > can be established:
Unknowns p1 , . . . , pn correspond to variables and propositional atoms in F correspond to constants. A particular solution G1 . . . Gn corresponds to a unifier
{p1 ← G1 , . . . , pn ← Gn } that is a ground substitution. The restriction to ground
substitutions is due to the requirement that unknowns do not occur in solutions.
General solutions Sect. 6 are expressed with further special parameter atoms,
different from the unknowns. These correspond to fresh variables in unifiers.
A generalization of Boolean unification to predicate logic with various specific
problems characterized by the involved formula classes has been investigated
in [16]. The material presented here is largely orthogonal to that work, but a
technique from [16] has been adapted to more general cases in Sect. 7.3.
3.3
View as Construction of Elimination Witnesses
Another view on the solution problem is related to eliminating second-order
quantifiers by replacing the quantified predicates with “witness formulas”.
Definition 7 (ELIM-Witness). Let p = p1 . . . pn be a sequence of distinct
predicates. An ELIM-witness of p in a formula ∃p F [p] is defined as a sequence
G of formulas such that SUBST(G, p, F ) and ∃p F [p] ≡ F [G].
The Solution Problem from Different Angles
9
The condition ∃p F [p] ≡ F [G] in this definition is equivalent to |= ¬F [p] ∨ F [G].
If F [p] and the considered G are first-order, then finding an ELIM-witness is
second-order quantifier elimination on a first-order argument formula, restricted
by the condition that the result is of the form F [G]. Differently from the general case of second-order quantifier elimination on first-order arguments, the set
of formulas for which elimination succeeds and, for a given formula, the set of
its elimination results, are then recursively enumerable. Some well-known elimination methods yield ELIM-witnesses, for example rewriting a formula that
matches the left side of Ackermann’s Lemma (Prop. 3) with its right side,
which becomes evident when considering that the right side F [G] is equivalent to ∀x1 . . . ∀xarity(p) (G ← G) ∧ F [G]. Finding particular solutions and finding
ELIM-witnesses can be expressed in terms of each other:
Proposition 8 (Solutions and ELIM-Witnesses). Let F [p] be a SP and let
G be a sequence of formulas. Then:
(i) G is an ELIM-witness of p in ∃p F if and only if G is a solution of the
SP (¬F [q] ∨ F )[p], where q is a sequence of fresh predicates matching p.
(ii) G is a solution of F [p] if and only if G is an ELIM-witness of p in ∃p F
and it holds that |= ∃p F .
Proof (Sketch). Assume SUBST(G, p, F ). (8.i) Follows since ∃p F [p] ≡ F [G]
iff ∃p F [p] |= F [G] iff F [p] |= F [G] iff |= ¬F [q] ∨ F [G]. (8.ii) Left-To-Right:
Follows since |= F [G] implies |= ∃p F [p] and |= F [G], which implies ∃p F [p] ≡
> ≡ F [G]. Right-to-left: Follows since ∃p F [p] ≡ F [G] and |= ∃p F [p] together
imply |= F [G].
t
u
3.4
View as Related to Definientia and Interpolants
The following proposition shows a further view on the solution problem that
relates it to definitions of the unknown predicates:
Proposition 9 (Solution as Entailed by a Definition). A sequence G =
G1 . . . Gn of formulas is a particular
solution of a SP F [p = p1 . . . pn ] if and
Vn
only if SUBST(G, p, F ) and i=1 (pi ⇔ Gi ) |= F .
Proof. Follows from the definition of particular solution and Prop. 2.i.
t
u
In the special case where F [p] is a 1-SP with a nullary unknown p, the characterization of a solution G according to Prop. 9 can be expressed with an entailment
where a definition of the unknown p appears on the right instead of the left side:
If p is nullary, then ¬(p ⇔ G) ≡ p ⇔ ¬G. Thus, the statement p ⇔ G |= F is
for nullary p equivalent to
¬F |= p ⇔ ¬G.
(i)
The second condition of the characterization of solution according to Prop. 9,
that is, SUBST(G, p, F ), holds if it is assumed that p is not in free(G), that
free(G) ⊆ free(F ) and that no member of free(F ) is bound by a quantifier occurrence in F . A solution is then characterized as negated definiens of p in the
10
Section 4
negation of F . Another way to express (i) along with the condition that G is
semantically independent from p is as follows:
∃p (¬F ∧ ¬p) |= G |= ¬∃p (¬F ∧ p).
(ii)
The second-order quantifiers upon the nullary p can be eliminated, yielding the
following equivalent statement:
¬F [⊥] |= G |= F [>].
(iii)
Solutions G then appear as the formulas in a range, between ¬F [⊥] and F [>].
This view is reflected in [30, Thm. 2.2], which goes back to work by Schröder.
If F is first-order, then Craig interpolation can be applied to compute formulas G that also meet the requirements free(G) ⊆ free(F ) and p ∈
/ free(F ) to
ensure SUBST(G, p, F ). Further connections to Craig interpolation are discussed
in Sect. 7.
4
The Method of Successive Eliminations – Abstracted
4.1
Reducing n-ary to 1-ary Solution Problems
The method of successive eliminations to solve an n-ary solution problem by
reducing it to unary solution problems is attributed to Boole and has been formally described in a modern algebraic setting in [30, Chapter 2, § 4]. It has
been rediscovered in the context of Boolean unification in the late 1980s, notably with [9]. Rudeanu notes in [30, p. 72] that variants described by several
authors in the 19th century are discussed by Schröder [34, vol. 1, §§ 26,27]. To
research and compare all variants up to now seems to be a major undertaking on
its own. Our aim is here to provide a foundation to derive and analyze related
methods. The following proposition formally states the core property underlying the method in a way that, compared to the Boolean algebra version in [30,
Chapter 2, § 4], is more abstract in several aspects: Second-order quantification
upon predicates that represent unknowns plays the role of meta-level shorthands
that encode expansions; no commitment to a particular formula class is made,
thus the proposition applies to second-order formulas with first-order and propositional formulas as special cases; it is not specified how solutions of the arising
unary solution problems are constructed; and it is not specified how intermediate
second-order formulas (that occur also for inputs without second-order quantifiers) are handled. The algorithm descriptions in the following subsections show
different possibilities to instantiate these abstracted aspects.
Proposition 10 (Characterization of Solution Underlying the Method
of Successive Eliminations). Let F [p = p1 . . . pn ] be a SP and let G =
G1 . . . Gn be a sequence of formulas. Then the following statements are equivalent:
(a) G is a solution of F [p].
The Method of Successive Eliminations – Abstracted
11
(b) For i ∈ {1, . . . , n}: Gi is a solution of the 1-SP
(∃pi+1 . . . ∃pn F [G1 . . . Gi−1 pi . . . pn ])[pi ]
such that free(Gi ) ∩ p = ∅.
Proof. Left-to-right: From (a) it follows that |= F [G]. Hence, for all i ∈ {1, . . . , n}
by Prop. 2.ii it follows that
|= ∃pi+1 . . . ∃pn F [G1 . . . Gi pi+1 . . . pn ].
From (a) it also follows that SUBST(G, p, F ). This implies that for all i ∈
{1, . . . , n} it holds that
SUBST(Gi , pi , ∃pi+1 . . . ∃pn F [G1 . . . Gi−1 pi . . . pn ]) and free(Gi ) ∩ p = ∅.
We thus have derived for all i ∈ {1, . . . , n} the two properties that characterize
Gi as a solution of the 1-SP as stated in (b).
Right-to-left: From (b) it follows that Gn is a solution of the 1-SP
(F [G1 . . . Gn−1 pn ])[pn ].
Hence, by the characteristics of solution it follows that |= F [G1 . . . Gn ]. The property SUBST(G, p, F ) can be derived from free(G)∩p = ∅ and the fact that for all
i ∈ {1, . . . , n} it holds that SUBST(Gi , pi , (∃pi+1 . . . ∃pn F [G1 . . . Gi−1 pi . . . pn ])).
The properties |= F [G1 . . . Gn ] and SUBST(G, p, F ) characterize G as a solution
of the SP F [p].
t
u
This proposition states an equivalence between the solutions of an n-ary SP
and the solutions of n 1-SPs. These 1-SPs are on formulas with an existential
second-order prefix. The following gives an example of this decomposition:
Example 11 (Reducing an n-ary Solution Problem to Unary Solution
Problems). Consider the SP F [p1 p2 ] of Examp. 6. The 1-SP with unknown p1
according to Prop. 10 is
(∃p2 F [p1 p2 ])[p1 ],
whose formula is, by second-order quantifier elimination, equivalent to ∀x (a(x ) →
b(x )) → ∀x (p1 (x ) → b(x )). Take a(x1 ) as solution G1 of that 1-SP. The 1-SP
with unknown p2 according to Prop. 10 is
(F [G1 p2 ])[p2 ].
Its formula is then, by replacing p1 in F as specified in Examp. 6 with a and
removing the duplicate conjunct obtained then, equivalent to
∀x (a(x ) → b(x )) → (∀x (a(x ) → p2 (x )) ∧ ∀x (p2 (x ) → b(x ))).
A solution of that second 1-SP is, for example, b(x1 ), yielding the pair a(x1 )b(x1 )
as solution of the originally considered SP F [p1 p2 ].
12
4.2
Section 4
Solving on the Basis of Second-Order Formulas
The following algorithm to compute particular solutions is an immediate transfer
of Prop. 10. Actually, it is more an “algorithm template”, since it is parameterized
with a method to compute 1-SPs and covers a nondeterministic as well as a
deterministic variant:
Algorithm 12 (SOLVE-ON-SECOND-ORDER). Let F be a class of formulas
and let 1 -SOLVE be a nondeterministic or a deterministic algorithm that outputs for 1-SPs of the form (∃p1 . . . ∃pn F [p])[p] with F ∈ F solutions G such that
free(G) ∩ {p1 , . . . , pn } = ∅ and F [G] ∈ F.
Input: A SP F [p1 . . . pn ], where F ∈ F, that has a solution.
Method: For i := 1 to n do: Assign to Gi an output of 1 -SOLVE applied to
the 1-SP (∃pi+1 . . . ∃pn F [G1 . . . Gi−1 pi . . . pn ])[pi ].
Output: The sequence G1 . . . Gn of formulas, which is a particular solution of
F [p1 . . . pn ].
The solution components Gi are successively assigned to some solution of the
1-SP given in Prop. 10, on the basis of the previously assigned components
G1 . . . Gi−1 . Even if the formula F of the input problem does not involve secondorder quantification, these 1-SPs are on second-order formulas with an existential
prefix ∃pi+1 . . . ∃pn upon the yet “unprocessed” unknowns.
The algorithm comes in a nondeterministic and a deterministic variant, just
depending on whether 1 -SOLVE is instantiated by a nondeterministic or a deterministic algorithm. Thus, in the nondeterministic variant the nondeterminism
of 1 -SOLVE is the only source of nondeterminism. With Prop. 10 it can be verified that if a nondeterministic 1 -SOLVE is “complete” in the sense that for each
solution there is an execution path that leads to the output of that solution,
then also SOLVE-ON-SECOND-ORDER based on it enjoys that property, with
respect to the n-ary solutions G1 . . . Gn .
For the deterministic variant, from Prop. 10 it follows that if 1 -SOLVE is
“complete” in the sense that it outputs some solution whenever a solution exists,
then, given that F [p1 . . . pn ] has a solution, which is ensured by the specification
of the input, also SOLVE-ON-SECOND-ORDER outputs some solution G1 . . . Gn .
This method applies 1 -SOLVE to existential second-order formulas, which
prompts some issues for future research: As indicated in Sect. 3.4 (and elaborated
in Sect. 7) Craig interpolation can in certain cases be applied to compute solutions of 1-SPs. Can QBF solvers, perhaps those that encode QBF into predicate
logic [35], be utilized to compute Craig interpolants? Can it be useful to allow
second-order quantifiers in solution formulas because they make these smaller
and can be passed between different calls to 1 -SOLVE ?
As shown in Sect. 6, if 1 -SOLVE is a method that outputs so-called reproductive solutions, that is, most general solutions that represent all particular solutions, then also SOLVE-ON-SECOND-ORDER outputs reproductive solutions.
Thus, there are two ways to obtain representations of all particular solutions
whose comparison might be potentially interesting: A deterministic method that
The Method of Successive Eliminations – Abstracted
13
outputs a single reproductive solution and the nondeterministic method with an
execution path to each particular solution.
4.3
Solving with the Method of Successive Eliminations
The method of successive eliminations in a narrower sense is applied in a Boolean
algebra setting that corresponds to propositional logic and outputs reproductive solutions. The consideration of reproductive solutions belongs to the classical material on Boolean reasoning [34,25,30] and is modeled in the present
framework in Sect. 6. Compared to SOLVE-ON-SECOND-ORDER, the method
handles the second-order quantification by eliminating quantifiers one-by-one,
inside-out, with a specific method and applies a specific method to solve 1-SPs,
which actually yields reproductive solutions. These incorporated methods apply
to propositional input formulas (and to first-order input formulas if the unknowns are nullary). Second-order quantifiers are eliminated by rewriting with
the equivalence ∃p F [p] ≡ F [>] ∨ F [⊥]. As solution of an 1-SP F [p] the formula
(¬F [⊥] ∧ t) ∨ (F [>] ∧ ¬t) is taken, where t is a fresh nullary predicate that is
considered specially. The intuition is that particular solutions are obtained by
replacing t with arbitrary formulas in which p does not occur (see Sect. 6 for a
more in-depth discussion).
The following algorithm is an iterative presentation of the method of successive eliminations, also called Boole’s method, in the variant due to [9]. The
presentation in [28, Sect. 3.1], where apparently minor corrections compared
to [9] have been made, has been taken here as technical basis. We stay in the
validity-based setting, whereas [30,9,28] use the unsatisfiability-based setting.
Also differently from [9,28] we do not make use of the xor operator.
Algorithm 13 (SOLVE-SUCC-ELIM).
Input: A SP F [p1 . . . pn ], where F is propositional, that has a solution and a
sequence t1 . . . tn of fresh nullary predicates.
Method:
1. Initialize Fn [p1 . . . pn ] with F .
2. For i := n to 1 do: Assign to Fi−1 [p1 . . . pi−1 ] the formula Fi [p1 . . . pi−1 >] ∨
Fi [p1 . . . pi−1 ⊥].
3. For i := 1 to n do: Assign to Gi the formula (¬Fi [G1 . . . Gi−1 ⊥] ∧ ti ) ∨
(Fi [G1 . . . Gi−1 >] ∧ ¬ti ).
Output: The sequence G1 . . . Gn of formulas, which is a reproductive solution
of F [p1 . . . pn ] with respect to the special predicates t1 . . . tn .
The formula assigned to Fi−1 in step (2.) is the result of eliminating ∃pi in
∃pi Fi [p1 . . . pi ] and the formula assigned to Gi in step (3.) is the reproductive
solution of the 1-SP (Fi [G1 . . . Gi−1 pi ])[pi ], obtained with the respective incorporated methods indicated above. The recursion in the presentations of [9,28]
is translated here into two iterations that proceed in opposite directions: First,
14
Section 4
existential quantifiers of ∃p1 . . . ∃pn F are eliminated inside-out and the intermediate results, which do not involve second-order quantifiers, are stored. Solutions
of 1-SPs are computed in the second phase on the basis of the stored formulas.
In this presentation it is easy to identify two “hooks” where it is possible to
plug-in alternate methods that produce other outputs or apply to further formula classes: In step (2.) the elimination method and in step (3.) the method to
determine solutions of 1-SPs. If the plugged-in method to compute 1-SPs outputs particular solutions, then SOLVE-SUCC-ELIM computes particular instead
of reproductive solutions.
4.4
Solving by Inside-Out Witness Construction
Like SOLVE-SUCC-ELIM, the following algorithm eliminates second-order quantifiers one-by-one, inside-out, avoiding intermediate formulas with existential
second-order prefixes of length greater than 1, which arise with SOLVE-ONSECOND-ORDER. In contrast to SOLVE-SUCC-ELIM, it performs elimination by
the computation of ELIM-witnesses.
Algorithm 14 (SOLVE-BY-WITNESSES). Let F be a class of formulas and
ELIM -WITNESS be an algorithm that computes for formulas F ∈ F and predicates p an ELIM-witness G of p in ∃p F [p] such that F [G] ∈ F.
Input: A SP F [p1 . . . pn ], where F ∈ F, that has a solution.
Method: For i := n to 1 do:
1. Assign to Gi [p1 . . . pi−1 ] the output of ELIM -WITNESS applied to
∃pi F [p1 . . . pi Gi+1 . . . Gn ].
2. For j := n to i+1 do: Re-assign to Gj [p1 . . . pi−1 ] the formula Gj [p1 . . . pi−1 Gi ].
Output: : The sequence G1 . . . Gn of formulas, which provides a particular
solution of F [p1 . . . pn ].
Step (2.) in the algorithm expresses that a new value is assigned to Gj and that
Gj can be designated by Gj [p1 . . . pi−1 ], justified because the new value does not
contain free occurrences of pi , . . . , pn . In step (1.) the respective current values of
Gi+1 . . . Gn are used to instantiate F . It is not hard to see from the specification
of the algorithm that for input F [p] and output G it holds that ∃p F ≡ F [G]
and that SUBST(G, p, F ). By Prop. 8.ii, G is then a solution if |= ∃p F . This
holds indeed if F [p] has a solution, as shown below with Prop. 15.
If ELIM -WITNESS is “complete” in the sense that it computes an elimination witness for all input formulas in F, then SOLVE-BY-WITNESSES outputs a
solution. Whether all solutions of the input SP can be obtained as outputs for different execution paths of a nondeterministic version of SOLVE-BY-WITNESSES
obtained through a nondeterministic ELIM -WITNESS , in analogy to the nondeterministic variant of SOLVE-ON-SECOND-ORDER, appears to be an open
problem.
15
Existence of Solutions
5
5.1
Existence of Solutions
Conditions for the Existence of Solutions
We now turn to the question under which conditions there exists a solution of
a given SP, or, in the terminology of [30], the SP is consistent. A necessary
condition is easy to see:
Proposition 15 (Necessary Condition for the Existence of a Solution).
If a SP F [p] has a solution, then it holds that |= ∃p F .
Proof. Follows from the definition of particular solution and Prop. 2.ii.
t
u
Under certain presumptions that hold for propositional logic this condition is
also sufficient. To express these abstractly we use the following concept:
Definition 16 (SOL-Witnessed Formula Class). A formula class F is called
SOL-witnessed for a predicate class P if and only if for all p ∈ P and F [p] ∈ F
the following statements are equivalent:
(a) |= ∃p F .
(b) There exists a solution G of the 1-SP F [p] such that F [G] ∈ F.
Since the right-to-left direction of that equivalence holds in general, the leftto-right direction alone would provide an alternate characterization. The class
of propositional formulas is SOL-witnessed (for the class of nullary predicates).
This follows since in propositional logic it holds that
∃p F [p] ≡ F [F [>]],
(iv)
which can be derived in the following steps: F [F [>]] ≡ ∃p (F [p]∧(p ↔ F [>])) ≡
(F [>] ∧ (> ↔ F [>])) ∨ (F [⊥] ∧ (⊥ ↔ F [>])) ≡ F [>] ∨ F [⊥] ≡ ∃p F [p].
The following definition adds closedness under existential second-order quantification and dropping of void second-order quantification to the notion of SOLwitnessed, to allow the application on 1-SPs matching with item (b) in Prop. 10:
Definition 17 (MSE-SOL-Witnessed Formula Class). A formula class F
is called MSE-SOL-witnessed for a predicate class P if and only if it is SOLwitnessed for P, for all predicates in p ∈ P and F ∈ F it holds that ∃p F ∈ F,
and, if ∃p F ∈ F and p ∈
/ free(F ), then F ∈ F.
The class of existential QBFs (formulas of the form ∃p F where F is propositional) is MSE-SOL-witnessed (like the more general class of QBFs – secondorder formulas with only nullary predicates). Another example is the class of
first-order formulas extended by second-order quantification upon nullary predicates, which is MSE-SOL-witnessed for the class of nullary predicates. The
following proposition can be seen as expressing an invariant of the method of
successive eliminations that holds for formulas in an MSE-SOL-witnessed class:
16
Section 5
Proposition 18 (Solution Existence Lemma). Let F be a formula class
that is MSE-SOL-witnessed for predicate class P. Let F [p = p1 . . . pn ] ∈ F with
p ∈ P n . If |= ∃p F [p], then for all i ∈ {0, . . . , n} there exists a sequence G1 . . . Gi
of formulas such that free(G1 . . . Gi ) ∩ p = ∅, SUBST(G1 . . . Gi , p1 . . . pi , F ),
|= ∃pi+1 . . . ∃pn F [G1 . . . Gi pi+1 . . . pn ] and ∃pi+1 . . . ∃pn F [G1 . . . Gi pi+1 . . . pn ] ∈ F.
Proof. By induction on the length i of the sequence G1 . . . Gi . The conclusion
of the proposition holds for the base case i = 0: The statement SUBST(, , F )
holds trivially, |= ∃p F is given as precondition, and ∃p F ∈ F follows from
F ∈ F. For the induction step, assume that the conclusion of the proposition
holds for some i ∈ {0, . . . n − 1}. That is, there exists a sequence G1 . . . Gi
of formulas such that free(G1 . . . Gi ) ∩ p = ∅, SUBST(G1 . . . Gi , p1 . . . pi , F ),
|= ∃pi+1 . . . ∃pn F [G1 . . . Gi pi+1 . . . pn ] and ∃pi+1 . . . ∃pn F [G1 . . . Gi pi+1 . . . pn ] ∈
F. Since F is MSE-SOL-witnessed for P and p1 , . . . , pi ∈ P it follows that there
exists a solution Gi+1 of the 1-SP
(∃p1 . . . ∃pi ∃pi+2 . . . ∃pn F [G1 . . . Gi pi+1 . . . pn ])[pi+1 ]
such that ∃p1 . . . ∃pi ∃pi+2 . . . ∃pn F [G1 . . . Gi+1 pi+2 . . . pn ] ∈ F. From the characteristics of solution it follows that
SUBST(Gi+1 , pi+1 , ∃p1 . . . ∃pi ∃pi+2 . . . ∃pn F [G1 . . . Gi pi+1 . . . pn ]),
which implies (since all members of p with exception of pi+1 are in the quantifier
prefix of the problem formula) that free(Gi+1 ) ∩ p = ∅, hence
free(G1 . . . Gi+1 ) ∩ p = ∅.
Given the induction hypothesis SUBST(G1 . . . Gi , p1 . . . pi , F ), it also implies
SUBST(G1 . . . Gi+1 , p1 . . . pi+1 , F ).
From the characteristics of solution it follows in addition that
|= ∃p1 . . . ∃pi ∃pi+2 . . . ∃pn F [G1 . . . Gi+1 pi+2 . . . pn ],
which, since free(G1 . . . Gi+1 ) ∩ p = ∅, is equivalent to
|= ∃pi+2 . . . ∃pn F [G1 . . . Gi+1 pi+2 . . . pn ].
Finally, we conclude from ∃p1 . . . ∃pi ∃pi+2 . . . ∃pn F [G1 . . . Gi+1 pi+2 . . . pn ] ∈ F,
established above, and the definition of MSE-SOL-witnessed that
∃pi+2 . . . ∃pn F [G1 . . . Gi+1 pi+2 . . . pn ] ∈ F,
which completes the proof of the induction step.
t
u
A sufficient and necessary condition for the existence of a solution of formulas
in MSE-SOL-witnessed classes now follows from Prop. 18 and Prop. 15:
Existence of Solutions
17
Proposition 19 (Existence of a Solution). Let F be a formula class that
is MSE-SOL-witnessed on predicate class P. Then for all F [p] ∈ F where the
members of p are in P the following statements are equivalent:
(a) |= ∃p F .
(b) There exists a solution G of the SP F [p] such that F [G] ∈ F.
Proof. Follows from Prop. 18 and Prop. 15.
t
u
From that proposition it is easy to see that for SPs with propositional formulas
the complexity of determining the existence of a solution is the same as the
complexity of deciding validity of existential QBFs, as proven in [21,22,2], that
is, ΠP
2 -completeness: By Prop. 19, a SP F [p] where F is propositional has a
solution if and only if the existential QBF ∃p F [p] is valid and, vice versa, an
arbitrary existential QBF ∃p F [p] (where F is quantifier-free) is valid if and only
if the SP F [p] has a solution.
5.2
Characterization of SOL-Witnessed in Terms of ELIM-Witness
The following proposition shows that under a minor syntactic precondition on
formula classes, SOL-witnessed can also be characterized in terms of ELIMwitness instead of solution as in Def. 16:
Proposition 20 (SOL-Witnessed in Terms of ELIM-Witness). Let F be
a class of formulas that satisfies the following properties: For all F [p] ∈ F and
predicates q with the same arity of p it holds that F [p] ∨ ¬F [q] ∈ F, and for all
F ∨ G ∈ F it holds that F ∈ F. The class F is SOL-witnessed for a predicate
class P if and only if for all p ∈ P and F [p] ∈ F there exists an ELIM-witness
G of p in F [p] such that F [G] ∈ F.
Proof. Left-to-right: Assume that F is meets the specified closedness conditions
and is SOL-witnessed for P, p ∈ P and F [p] ∈ F. Let q be a fresh predicate with
the arity of p. The obviously true statement |= ∃p F [p] ∨ ¬∃p F [p] is equivalent to
|= ∃p F [p]∨¬F [q] and thus to |= ∃p (F [p]∨¬F [q]). By the closedness properties of
F it holds that F [p] ∨ ¬F [q] ∈ F. Since F is SOL-witnessed for P it thus follows
from Def. 16 that there exists a solution G of the SP (F [p] ∨ ¬F [q])[p] such that
(F [G] ∨ ¬F [q]) ∈ F, and, by the closedness properties, also F [G] ∈ F. From
the definition of solution it follows that |= F [G] ∨ ¬F [q], which is equivalent
to ∃p F [p] ≡ F [G], and also that SUBST(G, p, F [G] ∨ ¬F [q]), which implies
SUBST(G, p, F [G]). Thus G is an SO-witness of p in F [p] such that F [G] ∈ F.
Right-to-left: Easy to see from Prop. 8.ii.
t
u
5.3
The Elimination Result as Precondition of Solution Existence
Proposition 19 makes an interesting relationship between the existence of a solution and second-order quantifier elimination apparent that has been pointed
out by Schröder [34, vol. 1, § 21] and Behmann [6], and is briefly reflected in
18
Section 6
[30, p. 62]: The formula ∃p F is valid if and only if the result of eliminating the
existential second-order prefix (called Resultante by Schröder [34, vol. 1, § 21]) is
valid. If it is not valid, then, by Prop. 19, the SP F [p] has no solution, however,
in that case the elimination result represents the unique (modulo equivalence)
weakest precondition under which the SP would have a solution. The following
proposition shows a way to make this precise:
Proposition 21 (The Elimination Result is the Unique Weakest Precondition of Solution Existence). Let F be a formula class and let P be a
predicate class such that F is MSE-SOL-witnessed on P. Let F [p] be a solution
problem where F ∈ F and all members of p are in P. Let A be a formula such
that (A → F ) ∈ F, A ≡ ∃p F , and no member of p does occur in A. Then
(i) The SP (A → F )[p] has a solution.
(ii) If B is a formula such that (B → F ) ∈ F, no member of p occurs in B,
and the SP (B → F )[p] has a solution, then B |= A.
Proof. (21.i) From the specification of A it follows that |= A → ∃pF and thus
|= ∃p (A → F ). Hence, by Prop. 19, the SP (A → F )[p] has a solution. (21.ii)
Let B be a formula such that the left side of holds. With Prop. 19 it follows that
|= B → ∃pF . Hence B |= ∃pF . Hence B |= A.
t
u
The following example illustrates Prop. 21:
Example 22 (Elimination Result as Precondition for Solvability). Consider the SP F [p1 p2 ] where
F = ∀x (p1 (x ) → p2 (x )) ∧ ∀x (a(x ) → p2 (x )) ∧ ∀x (p2 (x ) → b(x )).
Its formula is the consequent of the SP considered in Examp. 6. Since ∃p1 ∃p2 F ≡
∀x (a(x ) → b(x )) 6≡ >, from Prop. 19 it follows that F [p1 p2 ] has no solution.
If, however, the elimination result ∀x (a(x ) → b(x )) is added as an antecedent
to F , then the resulting SP, which is the SP of Examp. 6, has a solution.
6
Reproductive Solutions as Most General Solutions
Traditionally, concise representations of all particular solutions have been central to investigations of the solution problem. This section presents adaptions of
classic material to this end, due in particular to Schröder and Löwenheim, and
presented in a modern algebraic formalization by Rudeanu [30]. The idea is that
a general solution G[t] has parameter predicates t such that each instantiation
G[T ] with a sequence T of formulas is a particular solution and that for all particular solutions H there exists a sequence T of formulas such that H ≡ G[T ].
In this way, a general solution represents all solutions. A remaining difficulty is
to determine for a given particular solution H the associated T . This is remedied
with so-called reproductive solutions, for which H itself can be taken as T , that
is, it holds that G[H] ≡ H.
Reproductive Solutions as Most General Solutions
19
We give formal adaptions in the framework of predicate logic that center
around the notion of reproductive solution. This includes precise specifications
of reproductive solution and two further auxiliary types of solution. A technique
to construct a reproductive solution from a given particular solution, known
as Schröder’s rigorous solution or Löwenheim’s theorem and a construction of
reproductive solutions due to Schröder, which succeeds on propositional formulas
in general, is adapted. Finally, a way to express reproductive solutions of n-ary
SPs in terms of reproductive solutions of 1-SPs in the manner of the method of
successive eliminations is shown.
6.1
Parametric, General and Reproductive Solutions
The following definitions give adaptions of the notions of parametric, general and
reproductive solution for predicate logic, based on the modern algebraic notions
in [30,14] as starting point.
Definition 23 (Parametric and Reproductive Solution Problem (PSP,
RSP, 1-RSP)). A parametric solution problem (PSP) F [p]:t is a pair of a
solution problem F [p] and a sequence t of distinct predicates such that (free(F )∪
p) ∩ t = ∅. The members of t are called the solution parameters of the PSP.
If the sequences of predicates p and t are matching, then the PSP is called a
reproductive solution problem (RSP). A RSP with arity 1 is also called unary
reproductive solution problem (1-RSP).
Definition 24 (Parametric, General and Reproductive Solution). Define the following notions:
(i) A parametric solution of a PSP F [p]:t is a sequence G[t] of formulas such
that CLEAN(G), SUBST(G, p, F ) and for all sequences of formulas H such that
SUBST(H, t, G) and SUBST(H, p, F ) it holds that if there exists a sequence T
of formulas such that SUBST(T , t, G), SUBST(G[T ], p, F ) and
H ≡ G[T ],
then
|= F [H].
(ii) A general solution of a PSP F [p]:t is a sequence G[t] of formulas such that
the characterization of parametric solution (Def. 24.i) applies, with the if-then
implication supplemented by its converse.
(iii) A reproductive solution of a RSP F [p]:t is a sequence G[t] of formulas
such that
1. G is a parametric solution of F [p]:t and
2. For all sequences H of formulas such that SUBST(H, t, G) and
SUBST(H, p, F ) it holds that if
|= F [H],
then
H ≡ G[H].
20
Section 6
Parametric solution can be characterized more concisely than in Def. 24.i, but
not showing the syntactic correspondence to the characterization of general solution in Def. 24.ii:
Proposition 25 (Compacted Characterization of Parametric Solution).
A parametric solution of a PSP F [p]:t is a sequence G[t] of formulas such
that CLEAN(G), SUBST(G, p, F ) and for all sequences T of formulas such that
SUBST(T , t, G), SUBST(G[T ], p, F ) it holds that
|= F [G[T ]].
Proof. The left side of the proposition can be expressed as:
(1) CLEAN(G),
(2) SUBST(G, p, F ),
and for all sequences H, T of formulas it holds that
if
(3) SUBST(H, t, G), (III)
(4) SUBST(H, p, F ),
(5) SUBST(T , t, G),
(6) SUBST(G[T ], p, F ) and
(7) H ≡ G[T ],
then
(8) |= F [H].
The right side of the proposition can be expressed as:
(9) CLEAN(G),
(10) SUBST(G, p, F ),
and for all sequences T of formulas it holds that
if
(11) SUBST(T , t, G) and
(12) SUBST(G[T ], p, F ),
then
(13) |= F [G[T ]].
Left-to-right: If H = G[T ], then H ≡ G[T ]. Thus, this direction of the proposition follows if statements (9)–(12) imply (1)–(6), with H instantiated to G[T ].
Statements (1), (2), (5) and (6) are (9), (10), (11) and (12), respectively. The
instantiation of (3), that is, SUBST(G[T ], t, G), follows from (10) and (11). The
instantiation of (4) is SUBST(G[T ], p, F ), which is, like (6), identical to (12).
Right-to-left: Statements (1)–(7) imply (9)–(12). This holds since (1), (2), (5)
and (6) are (9), (10), (11) and (12), respectively. Hence, assuming the right side
of the proposition, statements (1)–(7) then imply (13), that is, |= F [G[T ]]. Statement (13), (7) and (6) imply (8), that is |= F [H], which concludes the proof. t
u
The essential relationships between particular, parametric, general and reproductive solutions, as well as an alternate characterization of reproductive solution
implied by these, are gathered in the following proposition:
Reproductive Solutions as Most General Solutions
21
Proposition 26 (Relationships Between the Solution Types). Let G =
G[t] be a sequence of formulas. Then:
(i) G is a parametric solution of the PSP F [p]:t if and only if CLEAN(G)
and G is a particular solution of the SP F [p].
(ii) If G is a parametric solution of the PSP F [p]:t and T is sequence of formulas such that SUBST(T , t, G), SUBST(G[T ], p, F ), then G[T ] is a particular
solution of the SP F [p].
(iii) A general solution of a PSP is also a parametric solution of that PSP.
(iv) If G is a general solution of the PSP F [p]:t and H is a particular solution
of the SP F [p] such that SUBST(H, t, G), then there exists a sequence T of
formulas such that SUBST(T , t, G), SUBST(G[T ], p, F ) and
H ≡ G[T ].
(v) A reproductive solution of a RSP is also a general solution of that RSP.
(vi) If G is a parametric solution of the RSP F [p]:t, then for all sequences H
of formulas such that SUBST(H, t, G) and SUBST(H, p, F ) it holds that if
H ≡ G[H],
then
|= F [H].
(vii) G is a reproductive solution of the RSP F [p]:t if and only if
1. G is a parametric solution of F [p]:t and
2. For all sequences H of formulas such that SUBST(H, t, G) and
SUBST(H, p, F ) it holds that
|= F [H]
if and only if
H ≡ G[H].
Before we come to the proof of Prop. 26, let us observe that the conclusion
of Prop. 26.vi is item (2.) of the definiens of reproductive solution (Def. 24.iii)
after replacing the if-then implication there by its converse, and that Prop. 26.vii
characterizes reproductive solution like its definition (Def. 24.iii), except that the
definiens is strengthened by turning the if-then implication in item (2.) into an
equivalence.
Proof (Proposition 26).
(26.i) Left-to-right: Let q be a sequence of fresh predicates that matches t and
assume that G[t] is a parametric solution of F [p]:t. Hence SUBST(G, p, F ) and
|= F [G[q]], which implies |= F [G]. Thus G is a particular solution of F [p]. Note
that this direction of the proposition requires the availability of fresh predicates
22
Section 6
in the vocabulary. Right-to-left: Can be derived in the following steps explained
below:
(1) G[t] is a particular solution of F [p].
(2) CLEAN(G)
(3) SUBST(G, p, F )
(4) |= F [G].
(5) SUBST(T , t, G).
(6) SUBST(G[T ], p, F ).
(7) SUBST(T , t, F [G]).
(8) |= ∀t F [G].
(9) |= F [G[T ]].
(10) G is a parametric solution of F [p]:t.
Step (1) and (2), where t is some sequence of distinct predicates such that
(free(F ) ∪ p) ∩ t = ∅, form the left side of the proposition. Steps (3) and (4)
follow from (1) and the characteristics of particular solution. Let T be a sequence
of formulas such that (5) and (6) hold, conditions on the left side of Prop. 25.
Step (7) follows from (5) and (6). Step (8) follows from (4). Step (9) follows from
(7) and (8) by Prop. 2.ii. Finally, step (10), the right side of the proposition,
follows from Prop. 25 with (9), (2) and (3).
(26.ii) The left side of the proposition includes SUBST(G[T ], p, F ) and, by
Prop. 25, implies |= F [G[T ]], from which the right side follows.
(26.iii) Immediate from the definition of general solution (Def. 24.ii).
(26.iv) The left side of the proposition implies SUBST(H, t, G),
SUBST(H, p, F ) and |= F [H]. The right side then follows from the definition
of general solution (Def. 24.ii).
(26.v) By definition, a reproductive solution is also a parametric solution.
Let G be a reproductive solution of F [p]:t. Let COND stand for the following
conjunction of three statements:
SUBST(H, t, G), SUBST(H, p, F ) and |= F [H].
From the definition of reproductive solution it immediately follows that for all
sequences H of formulas such that COND it holds that H ≡ G[H]. From this
it follows that for all sequences H of formulas such that COND it holds that
SUBST(H, t, G), SUBST(G[H], p, F ) and H ≡ G[H], which can be derived as
follows: The first of the statements on the right, SUBST(H, t, G), is included
directly in the left side, that is, COND. The second one, SUBST(G[H], p, F ), follows from SUBST(H, t, G) and SUBST(H, p, F ) that are in COND together with
SUBST(G, p, F ), which holds since G is a parametric solution. The above implication also holds if H on its right side is replaced by a supposedly existing T .
It then forms the remaining requirement to show that G is a general solution:
For all sequences H of formulas such that COND there exists a sequence T of
formulas such that SUBST(T , t, G), SUBST(G[T ], p, F ) and H ≡ G[T ].
Reproductive Solutions as Most General Solutions
23
(26.vi) Can be shown in the following steps, explained below:
(1)
(2)
(3)
(4)
(5)
(6)
(7)
SUBST(G, p, F ).
SUBST(H, t, G).
SUBST(H, p, F ).
H ≡ G[H].
SUBST(G[H], p, F ).
F [G[H]] |= ⊥.
F [H] |= ⊥.
Assume that G is a parametric solution of the RSP F [p]:t, which implies (1).
Let H be a sequence of formulas such that (2) and (3), the preconditions of the
converse of (as well as the unmodified) item (2.) in the definition of reproductive
solution (Def. 24.iii), hold. Further assume (4), the right side of item (2.). We
prove the proposition by deriving the left side of item (2.). Step (5) follows
from (1), (2) and (3). Step (6) follows from (2) and (5) by Prop. 25 since G is a
parametric solution. Finally, step (7), the left side of item (2.), follows from (6)
and (4) with (3) and (5).
(26.vii) Follows from Prop. 26.vi.
t
u
Rudeanu [30] notes that the concept of reproductive solution seems to have
been introduced by Schröder [34], while the term reproductive is due to Löwenheim [25]. Schröder calls the additional requirement that a reproductive solution
must satisfy in comparison with general solution Adventivforderung (adventitious requirement) and discusses it at length in [34, vol. 3, § 12], describing it
with reproduzirt. [34, vol. 3, p. 171].
6.2
The Rigorous Solution
From any given particular solution G, a reproductive solution can be constructed, called here, following Schröder’s terminology [34, vol. 3, § 12], the
rigorous solution associated with G. In the framework of Boolean algebra, the
analogous construction is [30, Theorem 2.11].
Proposition 27 (The Rigorous Solution). Let F [p]:t = t1 . . . tn be a RSP.
For i ∈ {1, . . . , n} let xi stand for x1 . . . xarity(ti ) . Assume free(F ) ∩ X = ∅,
SUBST(t1 (x1 ) . . . tn (xn ), p, F ) and SUBST(F [t], p, F ). If G = G1 . . . Gn is a
particular solution of that RSP , then the sequence R = R1 . . . Rn of formulas
defined as follows is a reproductive solution of that RSP:
Ri is the clean variant of (Gi (x) ∧ ¬F [t]) ∨ (ti (xi ) ∧ F [t]).
In the specification of Ri the formula Gi is written as Gi (x) to indicate that members of X may occur there literally without being replaced. In the unsatisfiabilitybased setting, the Ri would be characterized as the clean variant of
(Gi (x) ∧ F [t]) ∨ (ti (xi ) ∧ ¬F [t]).
(v)
24
Section 6
The proof of this proposition is based on the following lemma, a predicate logic
analog to [30, Lemma 2.3] for the special case n = 1, which is sufficient to prove
Prop. 27: The effect of the lemma for arbitrary n is achieved by an application
of Prop. 27 within an induction.
Proposition 28 (Subformula Distribution Lemma). Let p be a predicate
(with arbitrary arity ≥ 0), let F [p] be a formula and let V, W, A be formulas
such that SUBST(V, p, F ), SUBST(W, p, F ), SUBST(A, p, F ) and, in addition,
free(A) ∩ X = ∅. It then holds that
F [(A ∧ V ) ∨ (¬A ∧ W )] ≡ (A ∧ F [V ]) ∨ (¬A ∧ F [W ]).
Proof. Assume the preconditions of the proposition. It follows that SUBST((A ∧
V ) ∨ (¬A ∧ W ), p, F ). Making use of Prop. 2.i, the conclusion of the proposition
can be then be shown in the following steps:
Left side
≡ ∃p (F [p] ∧ (p ⇔ ((A ∧ V ) ∨ (¬A ∧ W ))))
≡ (A ∧ ∃p (F [p] ∧ (p ⇔ ((A ∧ V ) ∨ (¬A ∧ W ))))) ∨
(¬A ∧ ∃p (F [p] ∧ (p ⇔ ((A ∧ V ) ∨ (¬A ∧ W )))))
≡ (A ∧ ∃p (F [p] ∧ (p ⇔ ((> ∧ V ) ∨ (⊥ ∧ W ))))) ∨
(¬A ∧ ∃p (F [p] ∧ (p ⇔ ((⊥ ∧ V ) ∨ (> ∧ W )))))
≡ (A ∧ ∃p (F [p] ∧ (p ⇔ V ))) ∨ (¬A ∧ ∃p (F [p] ∧ (p ⇔ W )))
≡ Right side.
t
u
The preconditions in Prop. 28 permit that x1 , . . . , xarity(p) may occur free in V
and W , whereas in A no member of X is allowed to occur free. We are now ready
to prove Prop. 27:
Proof (Proposition 27). By item (1.) of the definition of reproductive solution
(Def. 24.iii), R[t] = R1 [t] . . . Rn [t] is required to be a parametric solution for
which by Prop. 25 three properties have to be shown: The first one, CLEAN(R),
is immediate since each member of R is the clean variant of some formula. The
second one, SUBST(R, p, F ), is easy to derive from the preconditions and the
definition of R. The third one is an implication that can be shown in the following
steps, explained below:
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
SUBST(T , t, R).
SUBST(R[T ], p, F ).
SUBST(G, p, F ).
|= F [G].
¬F [T ] ∧ ¬F [R[T ]] |= ¬F [G].
F [T ] ∧ ¬F [R[T ]] |= ¬F [T ].
F [G] |= F [R[T ]].
|= F [R[T ]].
Let T be a sequence of formulas such that statements (1) and (2), which are on
the left side of the implication to show, do hold. We derive the right side of the
Reproductive Solutions as Most General Solutions
25
implication, that is |= F [R[T ]]. Steps (3) and (4) hold since G is a particular solution. Steps (5) and (6) can be shown by induction based on the equivalences (9)
and (10), respectively, below, which hold for all i ∈ {0, . . . , n − 1} and follow
from Prop. 28:
(9)
¬F [G1 . . . Gi Ri+1 [T ] . . . Rn [T ]]
≡ ¬F [G1 . . . Gi ((Gi+1 ∧ ¬F [T ]) ∨ (Ti+1 ∧ F [T ])) Ri+2 [T ] . . . Rn [T ]]
≡ (¬F [T ] ∧ ¬F [G1 . . . Gi+1 Ri+2 [T ] . . . Rn [T ]]) ∨
(F [T ] ∧ ¬F [G1 . . . Gi Ti+1 Ri+2 [T ] . . . Rn [T ]]).
(10)
¬F [T1 . . . Ti Ri+1 [T ] . . . Rn [T ]]
≡ ¬F [T1 . . . Ti ((Gi+1 ∧ ¬F [T ]) ∨ (Ti+1 ∧ F [T ])) Ri+2 [T ] . . . Rn [T ]]
≡ (¬F [T ] ∧ ¬F [T1 . . . Ti Gi+1 Ri+2 [T ] . . . Rn [T ]]) ∨
(F [T ] ∧ F [T1 . . . Ti+1 Ri+2 [T ] . . . Rn [T ]]).
The required preconditions of Prop. 28 are justified there as follows, where F 0
stands for F after the substitutions indicated in (9) or (10), that is, the formula
matched with the left side of Prop. 28:
– SUBST(Gi+1 , pi+1 , ¬F 0 ): Follows from (3).
– SUBST(Ti+1 , pi+1 , ¬F 0 ): Follows from (1) and (2).
– SUBST(F [T ], pi+1 , ¬F 0 ): Follows from (1), (2) and the precondition
SUBST(F [t], p, F ).
– free(F [T ]) ∩ X = ∅:
Follows from (1), (2) and the precondition
free(F ) ∩ X = ∅.
Step (7) follows from (6) and (5) and, finally, step (8) follows from (7) and (4).
Item (2.) of the definition of reproductive solution follows since for all sequences of formulas H such that SUBST(H, t, R) and SUBST(H, p, F ) (note
that SUBST(H, t, G) is implied by SUBST(H, t, R)) it holds that if |= F [H],
then H ≡ R[H], or, equivalently, but more explicated, it holds for all i ∈
{1, . . . , n} that
Ri [H]
≡ (Gi [H] ∧ ¬F [H]) ∨ (Hi ∧ F [H])
≡ (Gi [H] ∧ ⊥) ∨ (Hi ∧ >)
≡ Hi .
t
u
The algebraic version [30, Theorem 2.11] is attributed there and in most of
the later literature to Löwenheim [24,25], thus known as Löwenheim’s theorem
for Boolean equations. However, at least the construction for unary problems
appears to be in essence Schröder’s rigorose Lösung [34, vol. 3, § 12]. (Löwenheim
remarks in [24] that the rigorose Lösung can be derived as a special case of his
theorem.) Behmann comments that Schröder’s discussion of rigorose Lösung
starts only in a late chapter of Algebra der Logik mainly for the reason that only
then suitable notation was available [6, Footnotes on p. 22f]. Schröder [34, vol. 3,
p. 168] explains his term rigoros as adaption of à la rigueur, that is, if need be,
because he does not consider the rigorous solution as a satisfying representation
26
Section 6
of all particular solutions. He notes that to detect all particular solutions on the
basis of the rigorose Lösung, one would have to test all possible formulas T as
parameter value. As remarked in [27, p. 382], Löwenheim’s theorem has been
rediscovered many times, for example in [26].
6.3
Schröder’s Reproductive Interpolant
For 1-RSPs of the form
((A ⇒ p) ∧ (p ⇒ B))[p]:t,
(vi)
A ∨ (B ∧ t(x)),
(vii)
the formula
where xs = x1 . . . xarity(p) , is a reproductive solution. This construction has been
shown by Schröder and is also discussed in [6]. For the notion of solution based on
unsatisfiability instead of validity, the analogous construction applies to 1-RSPs
of the form
((A ∧ p) ∨ (B ∧ ¬p))[p]:t
(viii)
and yields
B ∨ (¬A ∧ t).
(ix)
We call the solution interpolant because with the validity-based notion of solution assumed here the unknown p, and thus also the solution, is “between” A and
B, that is, implied by A and implying B. The following proposition makes the
construction precise and shows its justification. The proposition is an adaption
of [30, Lemma 2.2], where [34, vol. 1, § 21] is given as source.
Proposition 29 (Schröder’s Reproductive Interpolant). Let
(F = ∀y (A(y) → p(y)) ∧ ∀y (p(y) → B(y)))[p]:t,
where y is a sequence with the arity of p as length of distinct individual symbols not in X , be a 1-RSP that has a solution. Let x = x1 . . . xarity(p) . Assume
SUBST(A(x), p, F ), SUBST(B(x), p, F ) and SUBST(t(x), p, F ). Then the clean
variant of following formula is a reproductive solution of that 1-RSP:
A(x) ∨ (B(x) ∧ t(x)).
That p does not occur free in A or in B is ensured by the preconditions
SUBST(A, p, F ) and SUBST(B, p, F ). The symbols y for the quantified variables
indicate that these are independent from the special meaning of the symbols
in X .
Proof (Proposition 29). Assume the preconditions of the proposition and let
G[t] stand for the clean variant of A(x) ∨ (B(x) ∧ t(x)). By item (1.) of the
definition of reproductive solution (Def. 24.iii), G is required to be a parametric
solution for which by Prop. 25 three properties have to be shown: The first one,
Reproductive Solutions as Most General Solutions
27
CLEAN(G), is immediate since G is a clean variant of some formula. The second
one, SUBST(G, p, F ), easily follows from the preconditions. The third one is an
implication that can be shown in the following steps, explained below:
SUBST(A(x) ∨ (B(x) ∧ T (x)), p, F ).
|= ∃p (∀y (A(y) → p(y)) ∧ ∀y (p(y) → B(y))).
|= ∀y (A(y) → B(y)).
|= ∀y (A(y) → (A(y) ∨ (B(y) ∧ T (y)))) ∧
∀y ((A(y) ∨ (B(y) ∧ T (y))) → B(y)).
(4) |= F [A(x) ∨ (B(x) ∧ T (x))].
(1)
(2)
(3)
(4)
Let T (x) be a formula such that statement (1), which is on the left side of the
implication to show, does hold. We derive the right side of the implication, that is,
|= F [B(x) ∨ (A(x) ∧ T (x))]: Step (2) follows with Prop. 15 from the precondition
that the considered 1-RSP has a solution. Step (3) follows from (2) by secondorder quantifier elimination, for example with Ackermann’s lemma (Prop. 3).
The formulas to the right of |= in both statements are equivalent. Step (4) follows
from (3) by logic. Justified by (1), we can express (4) as (5), the right side of the
implication to show. Item (2.) of the definition of reproductive solution follows
since for all formulas H(x) such that SUBST(H(x), t, G), SUBST(H(x), p, F ),
it holds that |= F [H(x)] implies H(x) ≡ G[H(x)], which can be derived in the
following steps:
iff
iff
implies
implies
iff
|= F [H(x)]
|= ∀y (A(y) → H(y)) ∧ ∀y (H(y) → B(y))
|= ∀y (H(y) ↔ (A(y) ∨ H(y))) and
|= ∀y (H(y) ↔ (B(y) ∧ H(y)))
H(x) ≡ A(x) ∨ H(x) and H(x) ≡ B(x) ∧ H(x)
H(x) ≡ A(x) ∨ (B(x) ∧ H(x))
H(x) ≡ G[H(x)].
t
u
As shown by Schröder, the (clean variants of the) following two formulas are
further reproductive solutions in the setting of Prop. 29:
B(x) ∧ (A(x) ∨ t(x))
(x)
(A(x) ∧ t(x)) ∨ (B(x) ∧ ¬t(x)).
(xi)
and
These two formulas and the solution according to Prop. 29 are all equivalent
under the assumption that a solution exists, that is, |= ∃p F , which, by secondorder quantifier elimination, is equivalent to
|= ∀y (A(y) → B(y)).
(xii)
Any 1-RSP F [p]:t where F is a propositional formula or, more generally,
where the unknown p is nullary, can be brought into the form matching Prop. 29
28
Section 6
by systematically renaming bound symbols and rewriting F [p] with the equivalence
F [p] ≡ (¬F [⊥] → p) ∧ (p → F [>]).
(xiii)
For the notion of solution based on unsatisfiability, the required form can be
obtained with the Shannon expansion
F [p] ≡ (F [>] ∧ p) ∨ (F [⊥] ∧ ¬p).
6.4
(xiv)
From Unary to n-ary Reproductive Solutions
If the solution of a RSP is composed as suggested by Prop. 10 from reproductive
solutions of unary solution problems, then it is itself a reproductive solution:
Proposition 30 (Composing a Reproductive Solution from Unary Reproductive Solutions). Let F [p = p1 . . . pn ]:t = t1 . . . tn be a RSP. If G[t] =
G1 [t] . . . Gn [t] is a sequence of formulas such that for all i ∈ {1, . . . , n} it holds
that Gi is a reproductive solution of the 1-RSP
(∃pi+1 . . . ∃pn F [G1 . . . Gi−1 pi . . . pn ])[pi ]:ti
and free(Gi ) ∩ (p ∪ ti+1 . . . tn ) = ∅, then G is a reproductive solution of the
considered RSP F [p]:t.
Proof. Assume the preconditions and the left side of the proposition. We show
the two items of the definition of reproductive solution (Def. 24.iii) for G. Item (1.),
that is, G is a parametric solution of F [p]:t, can be derived as follows: Each Gi ,
for i ∈ {1, . . . , n}, is a reproductive solution of the associated 1-RSP. Hence, by
Prop. 26 it is a general, hence parametric, hence particular solution. By Prop. 10
it follows that G is a particular solution of F [p]. By Prop. 26.i it is then also a
parametric solution of F [p]:t. Item (2.) of the definition of reproductive solution
can be shown as follows: First we note the following statement that was given
as precondition:
(1)
For i ∈ {1, . . . , n} it holds that free(Gi ) ∩ ti+1 . . . tn = ∅.
For i ∈ {1, . . . , n} let
Fi [pi t] def
= ∃pi+1 . . . ∃pn F [G1 [t] . . . Gi−1 [t]pi . . . pn ],
that is, Fi is the formula of the 1-SP of which Gi is a reproductive solution. By
the definition of reproductive solution and the left side of the proposition it holds
for all formulas Hi that if
(2) SUBST(Hi , ti , Gi )
SUBST(Hi , pi , Fi [pi t]), and
|= Fi [Hi t],
then
(3) Hi ≡ Gi [t1 . . . ti−1 Hi ti+1 . . . tn ].
Approaching Constructive Solution Techniques
29
From this and (1) it follows that all for all sequences of formulas H1 . . . Hi it
holds that if
(4) SUBST(Hi , ti , Gi )
SUBST(Hi , pi , Fi [pi H1 . . . Hi−1 ti . . . tn ]), and
|= Fi [Hi H1 . . . Hi−1 ti . . . tn ],
then
(5) Hi ≡ Gi [H1 . . . Hi ti+1 . . . tn ].
Now let H ≡ H1 . . . Hn be a sequence of formulas such that
(6) SUBST(H, t, G)
SUBST(H, p, F ), and
|= F [H].
We prove the item (2) of Def. 24.iii by showing H ≡ G[H], which is equivalent to the statement that for all i ∈ {1, . . . , n} it holds that Hi ≡ Gi [H],
and, because of (1), to the statement that for all i ∈ {1, . . . , n} it holds that
Hi ≡ Gi [H1 . . . Hi ti+1 . . . tn ], which matches (5). We thus can prove H ≡
G[H] by showing that (4), which implies (5), holds for all i ∈ {1, . . . , n}.
The substitutivity conditions in (4) follow from the substitutivity conditions
in (6). The remaining condition |= Fi [Hi H1 . . . Hi−1 ti . . . tn ] can be proven by
induction. As induction hypothesis assume that for all j ∈ {1, . . . i − 1} it
holds that Hj ≡ Gj [H]. From |= F [H] in (6) it follows by Prop. 2.ii that
|= ∃pi+1 . . . ∃pn F [H1 . . . Hi pi+1 . . . pn ]. With the induction hypothesis it follows
that
|= ∃pi+1 . . . ∃pn F [G1 [H] . . . Gi−1 [H]Hi pi+1 . . . pn ],
which, given the substitutivity conditions of (6) and SUBST(G, p, F ), which
holds since G is a parametric solution, can be expressed as
|= Fi [Hi H1 . . . Hi−1 ti . . . tn ],
such that all conditions of (4) are satisfied and Hi ≡ Gi [H] can be concluded.
t
u
This suggests to compute reproductive solutions of propositional formulas for
a n-ary SP by constructing Schröder interpolants for 1-SPs. Since second-order
quantifier elimination on propositional formulas succeeds in general, the construction of the Schröder interpolant can there be performed on the basis of
formulas that are just propositional, without second-order quantifiers.
7
Approaching Constructive Solution Techniques
On the basis of first-order logic it seems that so far there is no general constructive method for the computation of solutions. We discuss various special
cases where a construction is possible. Some of these relate to applications of
Craig interpolation. Recent work by Eberhard, Hetzl and Weller [16] shows a constructive method for quantifier-free first-order formulas. A generalization of their
technique to relational monadic formulas is shown, which, however, produces solutions that would be acceptable only under a relaxed notion of substitutibility.
30
7.1
Section 7
Background: Craig Interpolation, Definability and Independence
By Craig’s interpolation theorem [13], if F and G are first-order formulas such
that F |= G, then there exists an a Craig interpolant of F and G, that is, a
first-order formula H such that
free(H) ⊆ free(F ) ∩ free(G)
(xv)
F |= H |= G.
(xvi)
and
Craig interpolants can be constructed from proofs of |= F → G, as, for example, shown for tableaux in [36,17]. Lyndon’s interpolation theorem strengthens
Craig’s theorem by considering in addition that predicates in the interpolant H
occur only in polarities in which they occur in both side formulas, F and G. In
fact, practical methods for the construction of interpolants from proofs typically
compute such Craig-Lyndon interpolants.
One of the many applications of Craig interpolation is the construction of
a definiens for a given predicate: Let F [pq1 . . . qk ] be a first-order formula such
that free(F ) ∩ X = ∅ and pq1 , . . . qk is a sequence of distinct predicates and let
x stand for x1 . . . xarity(p) . Then p is definable in terms of (free(F ) \ {q1 , . . . , qk })
within F , that is, there exists a first-order formula G such that
free(G) ⊆ (free(F ) \ {p, q1 , . . . qk }) ∪ x
(xvii)
F |= p ⇔ G,
(xviii)
∃p∃q1 . . . ∃qk (F ∧ p(x)) |= ¬∃p∃q1 . . . ∃qk (F ∧ ¬p(x)).
(xix)
and
if and only if
That entailment holds if and only if the following first-order formula is valid:
F ∧ p(x) → ¬(F [p0 q10 . . . qk0 ] ∧ ¬p0 (x)),
(xx)
where p0 q 0 . . . qk0 is a sequence of fresh predicates that matches pq . . . qk . The
definientia G of p with the stated characteristics are exactly the Craig interpolants of the two sides of that implication. Substitutibility SUBST(G, p, F ) can
be ensured by presupposing CLEAN(F ) and that no members of X are bound
by a quantifier occurrence in F .
Another application of Craig interpolation concerns the independence of formulas from given predicates: Second-order quantification allows to express that
a formula F [p] is semantically independent from the set of the predicates in p
as
∃p F ≡ F,
(xxi)
which is equivalent to ∃p F |= F , and thus, if q is a sequence of fresh predicates
that matches p, also equivalent to
|= F [q] → F.
(xxii)
31
Approaching Constructive Solution Techniques
As observed in [29], any interpolant of F [q] and F is then equivalent to F but its
free symbols do not contain members of p, that is, it is syntactically independent
of p. Thus, for a given first-order formula semantic independence from a set of
predicates can be expressed as first-order validity and, if it holds, an equivalent
formula that is also syntactically independent can be constructed by Craig interpolation. With Craig-Lyndon interpolation this technique can be generalized
to take also polarity into account, based on encoding of polarity sensitive independence as shown here for negative polarity: That F [p] is independent from
predicate p in negative polarity but may well depend on p in positive polarity
can be expressed as
∃q (F [q] ∧ ∀x (q(x) → p(x))),
(xxiii)
where x = x1 . . . xarity(p) and q is a fresh predicate with the same arity as p.
7.2
Cases Related to Definability and Interpolation
The following list shows cases where for an n-ary SP F [p = p1 . . . pn ] with firstorder F and which has a solution a particular solution can be constructed. Each
of the properties that characterize these cases is “semantic” in the sense that if
it holds for F , then it also holds for any first-order formula equivalent to F . In
addition, each property is at least “semi-decidable”, that is, the set of first-order
formulas with the property is recursively enumerable. Actually in the considered
cases, for each property a first-order formula can be constructed from F that is
valid if and only if F has the property. For two of the listed cases, (3.) and (5.),
the characterizing property implies the existence of a solution.
1. Each unknown occurs free in F only with a single polarity. A sequence of
> and ⊥, depending on whether the respective unknown occurs positively
or negatively, is then a solution. That F is semantically independent of unknowns in certain polarities, that is, is equivalent to a formula in which the
unknowns do not occur in these polarities, can be expressed as first-order
validity and a corresponding formula that is syntactically independent can
be constructed by Craig-Lyndon interpolation.
2. Each unknown is definable in the formula. A sequence of definientia, which
can be constructed with Craig interpolation, is then a solution. Rationale:
Let G1 . . . Gn be definientia of p1 . . . pn , respectively, in F . Under the assumption that there exists a solution H of F [p] it holds that
> |= F [H] |= ∃p F [p] ≡ ∃p (F [p] ∧
n
^
(pi ⇔ Gi )) ≡ F [G].
i=1
3. Each unknown is definable in the negated formula. The sequence of negated
definientia, which can be constructed with Craig interpolation, is then a
solution. Rationale: It holds in general that p ⇔ G |= p 6⇔ ¬G. Hence,
if G1 . . . Gn are definientia of p1 . . . pn , respectively, in ¬F , then ¬F |=
32
Section 7
Vn
i=1 (pi
⇔ Gi ) |=
Wn
i=1 (pi
⇔ Gi ) ≡
n
^
Wn
i=1 (pi
6⇔ ¬Gi ). Thus
(pi ⇔ ¬Gi ) |= F,
i=1
matching the characterization of solution in Prop. 9.
4. Each unknown is nullary. This specializes case (3.): If a solution exists, then a
nullary unknown is definable in the negated formula: For nullary predicates p
it holds in general that
p ⇔ ¬G ≡ p 6⇔ G.
Thus p ⇔ ¬G |= F (which matches Prop. 9) holds if and only if ¬F |= p ⇔
G.
5. Each unknown has a ground instance that is definable in the negated formula.
The sequence of negated definientia is a solution. If p1 (t1 ) . . . pn (tn ) are the
definable ground instances, then optionally in each solution component Gi ,
under the assumption CLEAN(Gi ), each member tij of ti = ti1 . . . tiarity(pi )
can be replaced by xj . The construction of the definientia can be performed
with Craig interpolation, as described above for predicate definientia, except
that an instance p(t) takes the place of p(x). The difficulty is to find suitable
instantiations t1 . . . tn . A way to avoid guessing might be to let the formula
whose proof serves as basis for interpolant extraction follow the schema
∃y (F ∧ p(y) → ¬(F [p0 ] ∧ ¬p0 (y))),
where y = y1 . . . yarity(p) and take the instantiation of y found by the prover. If
the proof involves different instantiations of y it has to be rejected. Rationale:
Similar to the case (4.) since for ground atoms p(t) it holds in general that
p(t) ↔ G ≡ ¬(p(t) ↔ ¬G).
These cases suggest to compute particular solutions based on Prop. 10 by computing solutions for 1-SPs for each unknown, which is inspected for matching the
listed cases or other types of solvable cases, for example the forms required by
Schröder’s interpolant or by Ackermann’s lemma. If that fails for an unknown,
an attempt with the unknowns re-ordered is made. For propositional problems,
an interpolating QBF solver would be a candidate to compute solutions. Encodings of QBF into predicate logic, e.g., [35], could possibly be applied for general
first-order formulas with nullary unknowns.
7.3
The EHW-Combination of ELIM-Witnesses for Disjuncts
Eberhard, Hetzl and Weller show in [16] that determining the existence of a
Boolean unifier (or, in, our terms, particular solution) for quantifier-free predicate logic is ΠP
2 -complete, as for propositional logic [2]. Their proof rests on the
existence of an EXPTIME function wit from quantifier-free formulas to quantifierfree formulas such that ∃pF [p] ≡ F [wit(F [p])]. The specification of wit(F [p])
33
Approaching Constructive Solution Techniques
is presented there as a variant of the DLS algorithm [15,11] for second-order
quantifier elimination: The input is converted to disjunctive normal form and
a specialization of Ackermann’s lemma is applied separately to each disjunct.
The results for each disjunct are then combined in a specific way to yield the
overall witness formula. The following proposition states a generalized variant
of this technique that is applicable also to other classes of inputs, beyond the
quantifier-free case.
Proposition
Wn31 (EHW-Combination of ELIM-Witnesses for Disjuncts).
Let F [p] = i=1 Fi be a formula and let G1 , . . . , Gn be formulas such that for
i ∈ {1, . . . n} it holds that SUBST(Gi , pi , Fi ) and ∃p Fi [p] ≡ Fi [Gi ]. Assume
that there are no free occurrences of X in F and, w.l.o.g, that no members of
free(F ) ∪ X are bound by a quantifier occurrence in F . Let
G(x) def
=
n
^
((
i=1
i−1
^
¬Fj [Gj ]) ∧ Fi [Gi ] → Gi (x)).
j=1
Then it holds that SUBST(G, p, F ) and ∃p F [p] ≡ F [G].
Formulas G and Gi are written as Gi (x) and G(x) where they occur as formula
constituents instead of substituents to emphasize that x may occur free in them.
Proof (Proposition 31). This proof is an adaption of the proof of Theorem 2 in
[16]. We write here I is a model of F symbolically as I |= F . That SUBST(G, p, F )
follows from the preconditions of the proposition and the construction of G. The
right-to-left direction of the stated equivalence, that is,
n
_
i=1
Fi [G] |= ∃p
n
_
Fi [p],
i=1
then follows from Prop. 2.ii. The left-to-right direction of the equivalence can be
show in the following steps, explained below.
Wn
(1) I |= ∃p
Wn i=1 Fi [p].
(2) I |= Wi=1 ∃p Fi [p].
n
(3) I |= i=1 Fi [Gi ].
Vk−1
(4) I |= ( j=1 ¬Fj [Gj ]) ∧ Fk [Gk ].
(5) I |= ∀x (G(x) ↔ Gk (x)).
(6) I |= F
Wkn[G].
(7) I |= i=1 Fi [G].
Let I be an interpretation such that (1) holds. Step (2) is equivalent to (1).
Assume the precondition
of the proposition that for all i ∈ {1, . . . n} it holds
Wn
that ∃p Fi [p] ≡ i=1 Fi [Gi ]. Step (3) follows from this and (1). By (3) there
is a smallest member k of {1, . . . , n} such that I |= Fk [Gk ]. This implies (4).
The left-to-right direction of the equivalence in (5) follows since if I |= G(x)
then by (4) and the definition of G(x) it is immediate that I |= Gk (x). The
34
Section 7
right-to-left direction of the equivalence in (5) can be shown as follows: Assume
I |= Gk (x). Then I is a model of the kth conjunct of G(x) since Gk (x) is in the
conclusion of that conjunct, and I is a model of each jth conjunct of G(x) with
j 6= k, because the antecedent of such a conjunct contradicts with (4). Step (6)
follows from (4) and (5). Step (7) follows from (6).
t
u
The following proposition is another variant of the EHW-combination of
witnesses for disjuncts. It can be proven similarly to Prop. 31.
Proposition 32 (Alternate Variant of EHW-Combination). Let F [p] be
a 1-SP and let G1 , . . . , Gn be formulas such
Wn that for i ∈ {1, . . . , n} it holds that
SUBST(Gi , p, F ) and such that ∃p F ≡ i=1 F [Gi ]. Assume that there are no
free occurrences of {xi | i ≥ 1} in F and, w.l.o.g, that no members of free(F ) ∪ x
are bound in by a quantifier occurrence in F . Let
G def
=
n
^
i=1
((
i−1
^
¬F [Gj ]) ∧ F [Gi ] → Gi ).
j=1
Then SUBST(G, p, F ) and ∃p F ≡ F [G].
Proposition 32 is also applicable to 1-SPs of the form handled by Prop. 31,
but for this case leads to a more clumsy result G: Assume the additional
preWn
condition
that
for
all
j
∈
{1,
.
.
.
,
n}
it
holds
that
SUBST(G
,
p,
F
).
Let
j
i=1 i
Wn
F [p] def
F
[p].
Then
∃pF
[p]
≡
F
[G
]
∨
.
.
.
∨
F
[G
]
≡
F
[G
]
∨
.
.
.
∨
F
[G
= i=1 i
1
1
n
n
1
n ].
7.4
Relational Monadic Formulas and Relaxed Substitutibility
The class of relational monadic formulas with equality, called here MON= , is the
class of first-order formulas with equality, with unary predicates and with individual constants but no other functions (without equality it is the Löwenheim
class). It is decidable and permits second-order quantifier elimination, that is,
each formula in MON= extended by predicate quantification is equivalent to a
formula in MON= . As shown in [39] it has interesting relationships with ALC.
Behmann [5] gave a decision method for MON= that performs second-order
quantifier elimination by equivalence preserving formula rewriting [39,38]. Almost three decades later he published an adaption of these techniques to the
solution problem for Klassenlogik [6,7], which in essence seems to be MON= .
It still remains open to assess this and apparently related works by Löwenheim
[25].
Under a relaxed notion of substitutibility the construction of ELIM-witnesses
for MON= is possible by joining Behmann’s rewriting technique [5] with the
EHW-combination (Prop. 31). Let F [p] be a MON= formula and let p be a
unary predicate. Assume that free(F ) ∩ X = ∅. The reconstruction of Behmann’s
normalization shown in the proofs of LemmaW14 and Lemma 16 of [38] can be
n
slightly modified to construct a formula F 0 = i=1 Fi00 that is equivalent to ∃p F
and such that each Fi00 is of the form
Fi00 = Ci ∧∃ui (Di (ui )∧∃p (∀y (Ai (ui y) → p(y))∧∀y (p(y) → Bi (ui y)))), (xxiv)
35
Solutions in Restricted Vocabularies
where ui is a sequence of individual symbols such that ui ∩ free(Ci ) = ∅, predicate p has only the two indicated occurrences and free(Fi0 ) ⊆ free(∃p F ). Let
Fi000 [p] = Ci ∧ Di (ui ) ∧ ∀y (Ai (ui y) → p(y)) ∧ ∀y (p(y) → Bi (ui y)).
(xxv)
Then Fi00 ≡ ∃ui ∃p Fi000 ≡ ∃ui Fi00 [A(ui x1 )], where the last equivalence follows
from Ackermann’s lemma (Prop. 3). It holds that SUBST(Fi00 , p, A(ui x1 )) but,
since the quantified symbols u may occur in A(ui x1 ), the substitutibility condition SUBST(∃u Fi00 , p, A(ui x1 )) does not hold in general. The variables ui can
be gathered to a single global prefix u (assuming w.l.o.g. that noneWof them ocn
curs free in any of the Ci ) such that F ≡ ∃u F 0000 [p] where F 0000 = i=1 Fi000 . By
Prop. 31 we can construct a formula G(u) such that SUBST(G(u), p, F 0000 ) and
∃p F 0000 [p] ≡ F 0000 [G(u)]. This implies ∃p F [p] ≡ F [G(u)]. However, substitutibility of G(u) holds only with respect to F 0000 , while SUBST(G(u), p, F ) does not
hold in general. Thus, under a relaxed notion of substitutibility that permits the
existentially quantified u in the witness the EHW-combination can be applied
to construct witnesses for MON= formulas.
8
Solutions in Restricted Vocabularies
8.1
An Example: Synthesizing Definitional Equivalence
In some applications it is useful to restrict the allowed vocabulary of the solution
components. Consider for example the task of finding a mapping that establishes
a definitional equivalence [20] between two formulas A and B where the predicates occurring free in A are in a set VA = {a1 , . . . , an } and those occurring in
B are in another set VB = {b1 , . . . , bm }, disjoint with VA . The objective is then
to find a solution GH of the SP F [pq] where G = G1 . . . Gm , H = H1 . . . Hn ,
p = p1 . . . pm , q = q1 . . . qn ,
F = (A ∧
m
^
i=1
∀y i (bi (y i ) ↔ pi (y i ))) ↔ (B ∧
n
^
∀z i (ai (z i ) ↔ qi (z i ))), (xxvi)
i=1
y i = y1 . . . yarity(bi ) , z i = z1 . . . zarity(ai ) , and the restriction is satisfied that all
predicates in free(G) are in VA and all predicates in free(H) are in VB .
8.2
Modeling with Two Consecutive Solution Problems
This can be achieved by solving consecutively two SPs followed by interpolant
computation: First, compute a reproductive solution R[t] of F [pq]:t. Since it is
a most general solution, if there is a particular solution, say GH, that meets the
vocabulary restrictions, there must be a sequence T of “instantiation formulas”
such that R[T ] ≡ GH. Each member of R[T ] is then “semantically” in the
required predicate vocabulary, that is, equivalent to a formula in which all free
predicates are members of the given set of predicates. Craig-Lyndon interpolation
can be applied on each of these formulas, if they are first-order, to construct
36
Section 9
equivalent formulas that are also syntactically in the required vocabulary as
explained in Sect. 7.1.
The remaining issue is to find suitable instantiation formulas T . These can
again be determined as solutions of a SP: Consider, for example, the case where
for a formula R[tb] a sequence T of formulas should be found such that Ri [T b]
is semantically independent from the members of b, that is, it should hold that
∃b R[T b] ≡ R[T b]. This is equivalent to R[T c] |= R[T b], where c is a sequence
of fresh predicates matching b, and hence also equivalent to |= R[T c] → R[T b].
Thus, suitable T can be obtained as solutions of the SP
(R[tc] → R[tb])[t].
(xxvii)
For an n-ary solution problem where R1 [tb1 ] . . . Rn [tbn ] is given and the requirement is that for all i ∈ {1, . . . , n} it holds that ∃bi Ri [T bi ] ≡ Ri [T bi ], a single
SP that combines the requirements can be used:
(
n
^
(Ri [tci ] → Ri [tbi ]))[t],
(xxviii)
i=1
where, for i ∈ {1, . . . , n}, ci is a sequence of fresh predicates that matches bi .
In fact, if m, n ≥ 1 the first SP of this method can be trivially solved: As
reproductive solution take the rigorous solution based on a particular solution
where G1 = ¬b1 (x1 . . . xarity(p1 ) ), H1 = ¬q1 (x1 . . . xarity(q1 ) ), and G2 . . . Gm and
H2 . . . Hn have arbitrary values, for example >. The actual effort to construct
the vocabulary restricted solution is then required for the second SP.
8.3
Expressing a Vocabulary Restriction on all Unknowns
A different technique applies to solution problems where there is only a single set
of predicates, say the set of members of the sequence b of predicates, that are not
permitted to occur free in the solution components: The vocabulary restriction
can then be directly encoded into the solution problem by means of secondorder quantification, justified by the equivalence of the following statements,
which follows from the requirement of substitutibility for solutions:
9
G is a solution of the SP F [p] and free(G) ∩ b = ∅.
(xxix)
G is a solution of the SP (∀b F )[p].
(xxx)
A Herbrand View on the Solution Problem
The characterization of solution in Prop. 9 is by an entailment of F . In presence
of the result of [16] for quantifier free predicate logic this brings up the question
whether Skolemization and Herbrand’s theorem justify some “instance-based”
technique for computing solutions that succeed on large enough quantifier expansions. The following is a formal account of that scenario which, so far, shows
no positive result but might be useful as a basis for further investigations.
Conclusion
37
Consider a 1-SP F [p] that has a solution G. By Prop. 9 it then holds that
p ⇔ G |= F . Let D stand for p ⇔ G. By Herbrand’s theorem we know that
there are formulas D0 , F 0 , Dh and F h and sequences of fresh functions d and f
such that
p ⇔ G = D ≡ ∀f ∃d D0 |= ∀f ∃d Dh |= ∀f ∃d F h |= ∀f ∃d F 0 ≡ F,
(xxxi)
and Dh |= F h . The functions d and f are the functions introduced by Skolemizing D (w.r.t. ∃) and F (w.r.t. ∀), respectively (this is the only place in the paper
where we consider second-order quantification upon functions). Formulas D0
and F 0 are the universal and existential, resp., first-order formulas, obtained
from Skolemizing D0 and F 0 , respectively. Formulas Dh and F h are quantifier free, obtained from D0 and F 0 , resp., by instantiating their matrices with
terms constructed from d, f and the free individual symbols in D or F (it is
assumed w.l.o.g. that at least one individual symbol is among the symbols available for term construction). We thus know p ⇔ G |= ∀f ∃d F h , or, equivalently,
p ⇔ G |= ∃ d F h . Thus G must be the solution to the 1-SP (∀f ∃d F h )[p], where
F h is quantifier-free and ∀f ∃d is a second-order prefix with quantifiers upon
functions.
As a sufficient condition for solutions it can be derived from this setting that
a solution H of the quantifier-free formula F h in which no member of f occurs
free is, under the assumption CLEAN(F ), also a solution of F , which follows
since p ⇔ H |= F h |= ∃dF h |= ∀f ∃dF h |= F .
10
Conclusion
The solution problem and second-order quantifier elimination were interrelated
tools in the early mathematical logic. Today elimination has entered automatization with applications in the computation of circumscription, in modal logics,
and for semantic forgetting and modularizing knowledge bases, in particular for
description logics. Since the solution problem on the basis of first-order logic is,
like first-order validity, recursively enumerable there seems some hope to adapt
techniques from first-order theorem proving.
The paper makes the relevant scenario accessible from the perspective of
predicate logic and theorem proving. It shows that a wealth of classical material
on Boolean equation solving can be transferred to predicate logic and only few
essential diverging points crystallize, like the constructability of witness formulas
for quantified predicates, and “Schröder’s reproductive interpolant” that does not
apply in general to first-order logic. An abstracted version of the core property
underlying the classical method of successive eliminations provides a foundation
for systematizing and generalizing algorithms that reduce n-ary solution problems to unary solution problems. Special cases based on Craig interpolation have
been identified as first steps towards methods for solution construction.
Beyond the presented core framework there seem to be many results from different communities that are potentially relevant for further investigation. This includes the vast amount of techniques for equation solving on the basis of Boolean
38
Section 10
algebra and its variants, developed over the last 150 years. For description logics
there are several results on concept unification, e.g., [4,3]. Variants of Craig interpolation such as disjunctive interpolation [32] share with the solution problem
at least the objective to find substitution formulas such that the overall formula
becomes valid (or, dually, unsatisfiable).
Among the issues that immediately suggest themselves for further research
are the parallel between nondeterministic methods with execution paths for each
particular solution and methods that compute a most general solution, the exploration of formula simplifications and techniques such as definitional normal forms
to make constructions like rigorous solution and reproductive interpolant feasible, and the investigation of the relaxed notion of substitutibility under which
solutions for relational monadic formulas can be constructed. The possible characterization of solution by an entailment also brings up the question whether
Skolemization and Herbrand’s theorem justify some “instance-based” technique
for computing solutions that succeeds on large enough quantifier expansions.
Acknowledgments
The author thanks anonymous reviewers for their helpful comments. This work
was supported by DFG grant WE 5641/1-1.
References
39
References
1. Ackermann, W.: Untersuchungen über das Eliminationsproblem der mathematischen Logik. Math. Ann. 110, 390–413 (1935)
2. Baader, F.: On the complexity of Boolean unification. Inf. Process. Lett. 67(4),
215–220 (Aug 1998)
3. Baader, F., Morawska, B.: Unification in the description logic EL. Logical Methods
in Computer Science 6(3), 1–31 (2010)
4. Baader, F., Narendran, P.: Unification of concept terms in description logics. J.
Symb. Comput. 31, 277–305 (2001)
5. Behmann, H.: Beiträge zur Algebra der Logik, insbesondere zum Entscheidungsproblem. Math. Ann. 86(3–4), 163–229 (1922)
6. Behmann, H.: Das Auflösungsproblem in der Klassenlogik. Archiv für Philosophie
4(1), 97–109 (1950), (First of two parts, also published in Archiv für mathematische
Logik und Grundlagenforschung, 1.1 (1950), pp. 17-29)
7. Behmann, H.: Das Auflösungsproblem in der Klassenlogik. Archiv für Philosophie
4(2), 193–211 (1951), (Second of two parts, also published in Archiv für mathematische Logik und Grundlagenforschung, 1.2 (1951), pp. 33-51)
8. Brown, F.M.: Boolean Reasoning. Dover Publications, second edn. (2003)
9. Büttner, W., Simonis, H.: Embedding Boolean expressions into logic programming.
J. Symb. Comput. 4(2), 191–205 (1987)
10. Carlsson, M.: Boolean constraints in SICStus Prolog. Tech. Rep. SICS T91:09,
Swedish Institute of Computer Science, Kista (1991)
11. Conradie, W.: On the strength and scope of DLS. J. Applied Non-Classical Logic
16(3–4), 279–296 (2006)
12. Conradie, W., Goranko, V., Vakarelov, D.: Algorithmic correspondence and completeness in modal logic. I. The core algorithm SQEMA. LMCS 2(1:5), 1–26 (2006)
13. Craig, W.: Linear reasoning. A new form of the Herbrand-Gentzen theorem. J.
Symb. Log. 22(3), 250–268 (1957)
14. Deschamps, J.P.: Parametric solutions of Boolean equations. Discrete Mathematics
3, 333–342 (1972)
15. Doherty, P., Łukaszewicz, W., Szałas, A.: Computing circumscription revisited: A
reduction algorithm. J. Autom. Reasoning 18(3), 297–338 (1997)
16. Eberhard, S., Hetzl, S., Weller, D.: Boolean unification with predicates. J. Logic
and Computation 27(1), 109–128 (2017)
17. Fitting, M.: First-Order Logic and Automated Theorem Proving. Springer, 2nd
edn. (1995)
18. Gabbay, D.M., Schmidt, R.A., Szałas, A.: Second-Order Quantifier Elimination: Foundations, Computational Aspects and Applications. College Publications
(2008)
19. Gabbay, D., Ohlbach, H.J.: Quantifier elimination in second-order predicate logic.
In: KR’92. pp. 425–435. Morgan Kaufmann (1992)
20. Hodges, W.: A Shorter Model Theory. Cambridge University Press (1997)
21. Kanellakis, P.C., Kuper, G.M., Revesz, P.Z.: Constraint query languages. In:
PODS’90. pp. 299–313. ACM Press (1990)
22. Kanellakis, P.C., Kuper, G.M., Revesz, P.Z.: Constraint query languages. J. Comput. Syst. Sci. 51(1), 26–52 (1995)
23. Koopmann, P., Schmidt, R.A.: Uniform interpolation of ALC-ontologies using fixpoints. In: FroCoS 2013. LNCS (LNAI), vol. 8152, pp. 87–102. Springer (2013)
40
24. Löwenheim, L.: Über das Auflösungsproblem im logischen Klassenkalkül. In:
Sitzungsberichte der Berliner Mathematischen Gesellschaft. vol. 7, pp. 89–94. Teubner (1908)
25. Löwenheim, L.: Über die Auflösung von Gleichungen im logischen Gebietekalkül.
Math. Ann. 68, 169–207 (1910)
26. Martin, U., Nipkow, T.: Unification in Boolean rings. In: CADE-8. LNCS (LNAI),
vol. 230, pp. 506–513. Springer (1986)
27. Martin, U., Nipkow, T.: Unification in Boolean rings. J. Autom. Reasoning 4(4),
381–396 (1988)
28. Martin, U., Nipkow, T.: Boolean unification – The story so far. J. Symb. Comput.
7, 275–293 (1989)
29. Otto, M.: An interpolation theorem. Bulletin of Symbolic Logic 6, 447–462 (2000)
30. Rudeanu, S.: Boolean Functions and Equations. Elsevier (1974)
31. Rudeanu, S.: Lattice Functions and Equations. Springer (2001)
32. Rümmer, P., Hojjat, H., Kuncak, V.: Disjunctive interpolants for Horn-clause verification. In: CAV 2013. LNCS, vol. 8044, pp. 347–363. Springer (2013)
33. Schmidt, R.A.: The Ackermann approach for modal logic, correspondence theory
and second-order reduction. J. Applied Logic 10(1), 52–74 (2012)
34. Schröder, E.: Vorlesungen über die Algebra der Logik. Teubner (vol. 1, 1890; vol. 2,
pt. 1, 1891; vol. 2, pt. 2, 1905; vol. 3, 1895)
35. Seidl, M., Lonsing, F., Biere, A.: bf2epr: A tool for generating EPR formulas from
QBF. In: PAAR-2012. EPiC, vol. 21, pp. 139–148 (2012)
36. Smullyan, R.M.: First-Order Logic. Springer, New York (1968), also republished
with corrections by Dover publications, New York, 1995
37. Sofronie, V.: Formula-handling computer solution of Boolean equations. I. Ring
equations. Bull. of the EATCS 37, 181–186 (1989)
38. Wernhard, C.: Heinrich Behmann’s contributions to second-order quantifier elimination. Tech. Rep. KRR 15–05, TU Dresden (2015)
39. Wernhard, C.: Second-order quantifier elimination on relational monadic formulas
– A basic method and some less expected applications. In: TABLEAUX 2015.
LNCS (LNAI), vol. 9323, pp. 253–269. Springer (2015)
40. Wernhard, C.: The Boolean solution problem from the perspective of predicate
logic. In: FroCoS 2017. LNCS (LNAI), vol. 10483, pp. 333–350 (2017)
41. Zhao, Y., Schmidt, R.A.: Concept forgetting in ALCOI-ontologies using an Ackermann approach. In: ISWC 2015. LNCS, vol. 9366, pp. 587–602. Springer (2015)
| 2cs.AI
|
arXiv:1709.03028v2 [cs.CV] 23 Oct 2017
Convolutional Neural Networks: Ensemble Modeling,
Fine-Tuning and Unsupervised Semantic Localization
for Intraoperative CLE Images
Mohammadhassan Izadyyazdanabadia,b,∗, Evgenii Belykha,b,c , Michael
Mooneyb , Nikolay Martirosyanb , Jennifer Eschbacherb , Peter Nakajib , Mark
C. Preulb , Yezhou Yanga
a Arizona State University, Tempe AZ 85281, USA
of Neurosurgery, Barrow Neurological Institute, St Joseph’s Hospital and
Medical Center, Phoenix, AZ 85013
c Irkutsk State Medical University, Krassnogo vosstaniya 1, Irkutsk, Russia 664003
b Department
Abstract
Confocal laser endomicroscopy (CLE) is an advanced optical fluorescence
technology undergoing assessment for applications in brain tumor surgery. Many
of the CLE images can be distorted and interpreted as nondiagnostic. However,
just one neat CLE image might suffice for intraoperative diagnosis of the tumor.
While manual examination of thousands of nondiagnostic images during surgery
would be impractical, this creates an opportunity for a model to select diagnostic
images for the pathologists or surgeons review. In this study, we sought to
develop a deep learning model to automatically detect the diagnostic images.
We explored the effect of training regimes and ensemble modeling and localized
histological features from diagnostic CLE images. The developed model could
achieve higher agreement with the ground truth than the other human observers.
With the speed and precision of the proposed method, it has potential to be
integrated into the operative workflow in the brain tumor surgery.
Keywords: Neural network, Unsupervised localization, Ensemble Modeling,
Brain, Confocal Laser Endomicroscopy, Surgical vision.
∗ Corresponding
author:
Email address: [email protected] (Mohammadhassan Izadyyazdanabadi)
Preprint submitted to Journal of Visual Communication and Image Representation October 25, 2017
1. Introduction
Handheld, portable Confocal Laser Endomicroscopy (CLE) is being explored
in neurosurgery because of its ability to image histopathological features of tissue
in real time [1, 2, 3, 4]. CLE provides cellular resolution imaging during brain
tumor surgery and thus may provide the surgeon with precise histopathological
information during tumor resection in order to interrogate regions that may
harbor malignant or spreading tumor, especially at the tumor border.
(a)
(b)
(c)
(d)
Figure 1: Diagnostic and nondiagnostic CLE images (field of view = 475 × 475 µm). (a,b)
Diagnostic images from glioma cases. (b) Unsupervisd localization of histopathological features of gliomas such as pleomorphism and hypercellularity detected by our model. For more
results see Fig. 5. (c,d) Nondiagnostic images from meningioma cases occluded with motion
(c) and blood artifact (d).
Current CLE systems are able to image more than one image per second, and
2
thus over the course of examination of the surgical tumor resection or inspection
area, hundreds to thousands of images may be collected. The number of images
may become rapidly overwhelming for the neurosurgeon and neuropathologist
when trying to quickly select a diagnostic or meaningful image or group of
images as the surgical inspection progresses. CLE is designed to be used on
the fly in real time while the surgeon operates the brain. Thus overcoming
the barriers involved in image selection is a key component for making CLE a
practical and advantageous technology for the neurosurgical operating room.
A wide range of fluorophores are able to be used for CLE in gastroenterology,
but fluorophore options are limited for in vivo human brain use due to potential
toxicity [1, 3, 5]. In addition, motion and blood artifacts that are present in
many of the images acquired with CLE using fluorescein sodium (FNa) are a
barrier for revealing underlying meaningful histology. The display of suboptimal
images or nondiagnostic frames interferes with the selection of and focus upon
diagnostic images by the neurosurgeon and pathologist throughout the operation
in order to make a correct intraoperative diagnosis. Previous assessment [6]
of CLE in human brain tumor surgery found that about half of the acquired
images were interpreted as nondiagnostic due to abundance of motion and blood
artifacts or lack of discernible or characteristic histopathological features.
Filtering out the nondiagnostic images before making an intraoperative diagnosis is challenging due to the high number of images acquired, the novel
and frequently unfamiliar appearance of tissue features compared to conventional histology, great variability among images from the same tumor type, and
potential similarity between images from other tumor types for the untrained
interpreter (Fig. 1).
Applications of machine learning in medical imaging have greatly increased
in the last ten years, resulting in numerous computer-aided detection(CADe)
and diagnosis(CADx) systems in ultrasound, magnetic resonance imaging (MRI),
and computed tomography (CT) [7]. Applications of machine learning for CLE
imaging in neurosurgery are yet to be performed. In this study, we developed an
ensemble of deep convolutional neural networks that can automatically evaluate
3
the diagnostic value of CLE frames within milliseconds.
Due to the limited number of images in our dataset, we sought to transfer
learning benefits by using pretrained models, fine-tuning them in shallow and
deep manner and compare results with the models trained from scratch. Our
results demonstrated that a shallow fine-tuned model, although performs better
than trained from scratch, is not enough for the optimal performance and that
a deep fine-tuned model detects the diagnostic CLE frames better. We also
investigated the effect of ensemble modeling by creating an ensemble of models
which were crafted at the development stage and produced the minimum loss on
validation dataset. Finally, we compared the performance among the ensemble
of models and each single model.
2. Related Works
2.1. Convolutional Neural Networks
Convolutional neural networks (CNNs), a subcategory of deep learning methods, have proven useful in visual imagery analysis from numerous fields, including medical images. This is mainly due to the deep multilayer architecture of
CNNs which enables extracting abstract discriminant features, both local and
global, present in the images.
In the recent years, deep learning has been vastly applied in medical image
or exam classification. According to a survey done by Litjens et al. [8], exam
and object classification together make up the number one task of interest in
medical image analysis followed by object detection and organ segmentation
(exam classification alone is the third task of interest). Most of these studies in
medical imaging field use one of the three following imaging modalities: MRI,
microscopy or CT.
Histopathological microscopic images and brain MRI scans were the first
two areas where deep learning has been explored in medical imaging [8]. In
histopathology, deep learning has been used for mitosis detection [9], classification of leukocytes [10] and nuclei detection and classification [11]. In brain MRI,
4
several studies have concluded that CNN benefits the diagnosis of Alzheimers
disease [12, 13] as well as brain extraction [14] and lesion detection, classification, and tumor grading [15, 16].
No-reference image quality assessment has been formulated as a classification
problem as employed in retinal [17] and echocardiographic[18] images. CNNs
may also be exercised in the detection of key frames from a temporal sequence
of frames in a video. Two studies demonstrated the use of classification scheme
on ultrasound (US) stream video to label the frames [19, 20].
2.2. Transfer Learning vs. Deep Training
One of the major limitations in medical imaging is the small size of datasets.
The number of images employed for deep learning applications in medical imaging is usually much smaller than those in computer vision. Therefore, two forms
of transfer learning have gained great interest: 1. Application of a pretrained
network on large-size natural images (i.e. ImageNet) as a feature extractor.
2. Initializing model parameters (weights and biases) using the data from a
pretrained model [21] instead of random initialization. A previous study by
Tajbaksh et al. [22] showed that a sufficiently fine-tuned AlexNet model could
produce equal or better results than a deeply trained one for colonoscopy image
quality assessment and few other medical applications. Here, we’ll study the
fine tuning effect by extending it to Inception network architectures in single
and ensemble mode.
2.3. Ensemble Modeling
Ensemble modeling is a well established method for increasing the model
performance and reducing its variance in machine learning [23, 24, 25]. Kumar et al. [26] created an ensemble of 5 different models to classify the image
modality from ImageCLEF 2016 medical image dataset. Specifically, 2 classifiers were created by fine-tuning AlexNet and GoogLeNet with softmax and 3
other classifiers by training an SVM on top of the features extracted by AlexNet,
GoogLeNet, and their combination. Their results showed the ensemble could
5
improve the top-1 accuracy of the classifier compared to single models, however
it is not clear that how much of the improvement was because of the AlexNet
and GoogLeNet combination or the 5 classifiers ensemble.
Christodoulidis et al.
[27] created an ensemble of multi-source transfer
learning using an automatic model selection approach described in [ensembleselection]. After creating a pool of pretrained CNNs on several public texture
datasets and fine-tuning them on the lung CT dataset, the top models which
iterative grouping would produce the highest F-scores on the validation dataset
were aggregated, creating an ensemble model. 5 ensemble models were developed and their output was then averaged to make the final output. Despite its
computational complexity, it enhanced the lung disease pattern classification
accuracy only by 2%.
To generate diversity in our models while using the whole training dataset,
we trained different neural networks on different data using cross-validation.
Although previous studies have tried to create variant deep learning models using different network architectures, none of them have employed training data
diversification through cross-validation as described in [28]. Our proposed ensemble employed model diversification both in the network architectures and in
the training and validation datasets following [28].
2.4. Confocal Laser Endomicroscopy in Neurosurgery
Handheld, portable CLE has demonstrated its value for brain tumor surgery
due to its ability to provide rapid intraoperative information regarding histopathological features of the tumor tissue [6]. Convenience, portability, and speed of
CLE are significant advantages in surgery. A decision support system aiding and
accelerating analysis of CLE images by the neuropathologist or neurosurgeon
would improve the workflow in the neurosurgical operating room [29].
Potentially used at any time during the surgery, CLE interrogation of the
tissue generates images with a speed of 0.8 - 1.2 frames per second. The frames
are considered nondiagnostic when the histological features are obscured by
the red blood cells or motion artifacts, are out of focus, or lack any useful
6
histopathological information. Acquired images are then exported from the
instrument as JPEG or TIFF files for review. Currently, the pathologist reviews
all images, including nondiagnostic ones, trying to explore the diagnostic frames
for the diagnosis. Manual selection and review of thousands of images acquired
at some point in surgery by the CLE operator is tedious and impractical for
widespread use. Previously, we have presented [29] the first deeply trained
CNN model for automatic detection of diagnostic CLE frames. In this work, we
extend our previous work with the following contributions and advancements:
1. Dataset. Our dataset contains CLE images which is a novel technology
in contrast to commonly used MRI or CT scans. The dataset used includes
20,734
CLE
images
from
intracranial
neoplasms.
2. Deep training, shallow fine-tuning or deep fine-tuning? The CNN
architectures were trained in three regimes: I. deeply trained (train the
network from scratch with model weights randomly initialized) II. shallow
fine-tuned (fine-tune only the fully connected layer(s) of the model which
are responsible for the classification) III. deeply fine-tuned (fine-tune the
whole network using our dataset). In this study we report model accuracy
on the test dataset for the best 5 models from each network architecture
and training regime. Our work is different from [22] since it considers the
fine tuning effect on two different network architectures and its effect on
the ensemble models.
3. Ensemble modeling. Prior to the test phase, we created an ensemble
of the best 5 models from each network and training regime. We explored
the effect of ensemble modeling in all circumstances by comparing the
ensemble performance with the average of single models. Our work is
different from [26] since our ensemble generates diversity in single models
by using different training and validation data achieved from nested-cross
validation. Further, we studied the effect of ensemble modeling on different training scenarios rather than one.
7
4. Unsupervised localizing of histological features. We visualized the
shallow neurons’ activation to depict the broad histological patterns; visualization of deep neurons activation could localize specific histopathological lesions for diagnostic images. The neural response of convolutional
layers to the diagnostic images are visualized and analyzed by a neurosurgeon. We also extracted the CNN’s deepest neural activation in response
to patches of the diagnostic images using a sliding-window. Qualitative
assessment of the localized regions was performed by a neurosurgeon with
further analysis of the histopathological features.
5. Interobserver study. We compared the interobserver agreement between physician-physician and ensemble of models-physician to compare
our ensemble model performance with human performance. We also reported the kappa statistic for this observer study.
3. Methods
3.1. Image Acquisition
In the following sections we briefly explain the confocal imaging instrument
instrument specifications and the intraoperative data collection process.
3.1.1. Instrument specifications
The CLE image acquisition system (Optiscan 5.1, Optiscan Pty, Ltd.) included a rigid pen-sized optical laser scanner with a 6.3 mm outer diameter and
a working length of 150 mm. A 488 nm diode laser provided excitation light
and the fluorescent emission signal was detected with a 505-585 nm band-pass
filter. A single optical fiber acted as both the excitation pinhole and the detection pinhole for confocal isolation of the image plane. The detector signal was
digitized synchronously with the scanning to construct images parallel to the
tissue surface (en face optical sections).
8
Laser power was typically set to 550-900 µW and maximum power was limited to 1000 µW when applied to the brain tissue. A field of view of 475 × 475
µm was scanned at 1024 × 1024 pixels (1.2/second frame rate), with a lateral
resolution of 0.7 µm and an axial resolution (i.e., effective optical slice thickness)
of approximately 4.5 µm.
3.1.2. Intraoperative CLE imaging
Seventy-four adult patients (31 male and 43 female) were enrolled in the
study (mean age 47.5 years). Intraoperative CLE images were acquired both in
vivo and ex vivo by 4 neurosurgeons. For in vivo imaging, multiple locations of
the tissue around the lesion were imaged and excised from the patient. For ex
vivo imaging, tissue samples suspicious for tumor were excised, placed on gauze
and imaged on a separate work station in the operating room. Multiple images
were obtained from each biopsy location.
Co-registration of the CLE probe with the image guided surgical system
allowed precise intraoperative mapping of CLE images with regard to the site
of the biopsy. The only fluorophore administered was FNa (5mL, 10%) that
was injected intravenously during the surgery.
Precise location of the areas imaged with the CLE was marked with tissue
ink. Imaged tissue was sent to the pathology laboratory for formalin fixation,
paraffin embedding and histological sections preparation. Final histopathological assessment was performed by standard light microscopic evaluation of 10µm-thick hematoxylin and eosin (H & E)-stained sections.
3.2. Image annotation
The image annotation process was done in two distinct stages: initial review
and .
3.2.1. Initial review
Initially all images were reviewed. A neuropathologist and 2 neurosurgeons
who were not involved in the surgeries reviewed the CLE images. For each
patient, the histopathological features of corresponding CLE images and H
9
& E-stained frozen and permanent sections were reviewed and the diagnostic
value of each image was examined. When CLE image revealed clear identifiable
histopathological feature, it was labeled as diagnostic; otherwise it was labeled
as nondiagnostic.
3.2.2. Validation review
The database of images was divided into development and test datasets
(explained in dataset preparation 4.1). Test dataset composed of 4171 CLE images randomly chosen from various patients. The validation review (val review)
dataset consists of 540 images randomly chosen from the test dataset. Following
this separation, two neurosurgeons reviewed val-review dataset without having
access to the corresponding H & E-stained slides and labeled them as diagnostic
or nondiagnostic .
3.3. Convolutional Neural Networks
Convolutional Neural Networks (CNNs) are multilayer learning frameworks
and may consist of an input layer, a few convolutional layers, pooling layers, fully
connected layers and the output. The goal of a CNN is to learn the hierarchy
of underlying feature representations. We explain the fundamental elements of
a CNN below.
3.3.1. Convolutional layer
Convolutional layers, first introduced in [30] are the substitute of previous
hand-crafted feature extractors. At each convolutional layer three dimensional
matrices (kernels) are slid over the input and set the dot product of kernel
weights with the receptive field of the input as the corresponding local output. This helps to retain the relative position of features to each other. The
multi-kernel characteristic of convolutional layers enables them to prospectively
extract several distinct feature maps from the same input image.
10
3.3.2. Activation layer
The convolutional layer output then goes through an activation function to
adjust the negative values. We employed the rectified linear unit (ReLU) which
is usually the preferred choice because of its simplicity, higher speed, reduced
likelihood of vanishing gradients (especially in deep networks) and tendency
to add sparsity over other nonlinear functions such as sigmoid function. The
in
output of j th ReLU layer (aout
j ), given its input (aj ), was calculated in-place
(to consume less memory) by following:
aout
= max (ain
j
j , 0)
(1)
3.3.3. Normalization layer
Following the ReLU layer, a local response normalization (LRN) map is
applied after the initial convolutional layers. This layer inhibits the local ReLU
neurons’ activations since there’s no bound to limit them (Eq. 1). By using the
Caffe implemented [31] LRN, the local regions are expanded across neighbor
feature maps at each spatial location. The output of j th LRN layer (aout
j ),
given its input (ain
j ), is calculated as:
aout
=
j
ain
j
(1 +
α
L
PL
n=1
2
β
ain
j (n) )
(2)
th
in
where ain
element of the ain
j (n) is the n
j and L is the length of aj vector
(number of neighbor maps employed in the normalization). α, β and L are
the layer’s hyperparameters and are set to their default values obtained from
[32](α = 1, β = 0.75 and L = 5).
3.3.4. Pooling layer
After rectification and normalization of convolutional layer output, it’s further down-sampled by pooling operations. Pooling operations accumulate values
in a smaller region by subsampling operations such as max, min, and average
sampling. Here, max pooling was applied.
11
3.3.5. Fully connected Layer
Following several convolutional and pooling layers, the network lateral layers
are fully connected. Each neuron of the layer’s output is greedily connected to
all the layer’s input neurons. It can be thought of as a convolutional layer
with kernel size of the layer input. The layer output is also passed through a
ReLU layer. The fully connected layers are generally thought of as the classifier
of a CNN model because they intake the most abstract features extracted in
convolutional layers and make the output, which is the model prediction.
3.3.6. Dropout Layer
Fully connected layers are usually followed by a dropout layer, except the last
fully connected layer that produces the class-specific probabilities. In dropout
layers, a subset of input neurons as well as all their connections are temporarily
removed from the network. Srivastava et al.[33] have demonstrated this method
efficiency at improving the CNN performance in numerous computer vision tasks
through reducing the overfitting.
3.3.7. Learning
The learning of a CNN is through Stochastic Gradient Descent (SGD) which
stands on two major menhirs: Forward and Back Propagation. In forward
propagation, the model makes predictions using the images in the training batch
and the current model parameters. Once the prediction for all training images
is made, the loss is calculated using the truth label provided by the experts in
the initial review (explained in 3.2.1). In this work we adopt the softmax loss
function given by:
L(t, y) = −
n
N
C
1 XX n
eyk
tk log( PC
)
n
ym
N n=1
m=1 e
(3)
k=1
where tnk is the nth training image’s k th ground truth output, and ykn is the
value of the k th output layer unit in response to the nth input training image.
N is the number of training images in the minibatch, and since we consider 2
diagnostic value categories, C = 2.
12
Through the back propagation, the loss gradient with respect to all model
weights aids upgrading the weights as follows:
W (j, i + 1) = W (j, i) + µ∆W (j, i) − α(j, i)
∂L
∂W (j)
(4)
where W (j, i), W (j, i + 1) and ∆W (j, i) are the weights of j th convolutional
layer at iteration i and i + 1 and the weight update of iteration i, µ is the
momentum and α(j, i) is the learning rate and is dynamically lowered as the
training progresses.
3.4. Evaluation Metrics
In model performance estimation (explained in 4.3) we calculated the loss,
accuracy, sensitivity, specificity and area under receiver operating characteristics
(ROC) curve (AUC). Here, we assumed the state of being a diagnostic image as
positive and being nondiagnostic as negative. This way, sensitivity determines
the model ability to detect diagnostic images and specificity determines its ability to detect nondiagnostic images. Accuracy determines general capability of
a model to detect diagnostic and nondiagnostic images correctly [34].
4. Experimental Setup
4.1. Dataset Preparation
Our dataset included 20,734 CLE images from 74 brain tumor cases. For
each CLE image, the diagnostic quality was determined by the experts in the
initial review. The dataset was divided into two main subsets on patient level:
development (dev) and test. Our pilot study revealed that when the division
is on image level (mixing the images from all the patients and dividing them
randomly) the model would produce poor results on images from new patients.
The total number of patients and images used at each stage are presented in
Table 1. Each subset contains images from various tumor types (mainly from
gliomas and meningiomas). The dev set will be available online. The test set
was isolated all through the model development and was accessed only in the
test phase.
13
4.2. Model Development
After the initial data split, we employed a patient-based k-fold cross validation for model development. Fifty nine cases that were allocated for model
development were divided into 5 groups. Since CNNs require a large set of
hyperparameters to be defined optimally (i.e. initial value of the learning rate
and its lowering policy, momentum, batch size, etc.), we used different values
with grid searching throughout the model development process. For every set
of feasible parameters, we trained the model on 4 folds and validated on the
fifth left-out group of patients. The set of hyperparameters which produced the
minimum average loss was employed for each set of experiments.
The small dataset size was a main limitation of our study for using CNNs, especially with the patient-level data preparation. Therefore, we counterbalanced
this limitation by fine-tuning the pretrained publicly available CNN architectures trained on large computer vision datasets (i.e. ImageNet).
Though the question about how deep should we fine-tune the pretrained models for optimal results still remains unanswered, one study tried to answer this
question using endoscopy and ultrasound images[22]. Due to the substantial intrinsic dissimilarities between the images in the 2 studies, we performed a similar
Table 1: Dataset preparation: Patient-based allocation of diagnostic and nondiagnostic images
from various neoplasms to model development and testing. Number of patients for each tumor
type is also provided.
Development
Test
59
15
Gliomas
16
5
Meningiomas
24
6
Other neoplasms
19
4
16,366
4,171
Diagnostic
8,023
2,071
Nondiagnostic
8,343
2,100
No. of Patients (total)
No. of Images (total)
14
Figure 2: Ensemble effect on network 1 (top) and network 2 (bottom) while using diverse
training regimes. For both networks, the improvement was more noticeable with DT and
DFT regimes. The arithmetic and geometric ensemble performed similarly. Neither of the
two ensembles could improve network 2 trained with SFT.
investigation. Our confocal images have a much higher spatial resolution and
are fluorescent images from the brain.
In total, we developed 42 models (30 single models and 12 ensemble models)
using two network architectures and three training regimes (deep training, shallow fine-tuning and deep fine-tuning). The experiments are designed in order
to practically find the optimal model development pathway that produces the
highest performance in the considered clinical application.
4.2.1. Network architectures
Two CNN architectures were applied in this study. Network 1 had 5 convolutional layers. The first two convolutional layers had 96 and 256 filters of size
11 × 11 and 5 × 5 with maximum pooling. The third, fourth and fifth convolu-
15
tional layers were connected back to back without any pooling in between. The
third convolutional layer had 384 filters of size 3 × 3 × 256, the fourth layer had
384 filters of size 3 × 3 × 192 and the fifth layer had 256 filters of size 3 × 3 × 192
with maximum pooling. For more details please refer to [32].
Network 2 had 22 layers with parameters and 9 inception modules. Each
inception module was a combination of filters of size 1×1, 3×3, 5×5 and a 3×3
max pooling, put together in parallel and the output filter banks concatenated
into an input single vector for the next stage. For more details please refer to
[35].
The pretrained model for network 1, exploited in fine-tuning experiments,
was the iteration 360,000 snapshot of training the model on ImageNet classification with 1000 classes. The pretrained model for network 2 was iteration
2,400,000 of training on ImageNet classification dataset. Both models are publicly available in Caffe libraries [31].
4.2.2. Training regimes
We exercised various training regimes to see how deep fine-tuning should be
done in CLE image classification for optimal results. Depending on which layers
of the network are being learned through training, we had three regimes.
In regime 1, deep training (DT), the whole model weights were initialized
randomly (training from scratch) and got modified all through the training with
nonzero learning rates.
In regime 2, shallow fine-tuning (SFT), the whole model weights, except
the last fully connected layer, were initialized with the corresponding values
from the pretrained model and their values were fixed for the period of training.
The last fully connected layer was initialized randomly and got tuned during
training.
In regime 3, deep fine-tuning (DFT), all model weights, except for the
last fully connected layer, were initialized with the corresponding values from
the pretrained model and last fully connected layer was initialized randomly.
Throughout the training, all the CNN layers, including the last fully connected
16
layer, were tuned with nonzero learning rates. Our hyperparameter optimization
showed that the SFT and DFT experiments required 10 times smaller initial
learning rates (0.001) compared to the DT regime (0.01). To avoid overfitting,
the training process was stopped after 3 epochs of consistent loss increment
on the validation dataset. We also used dropout layer (ratio = 0.5) and L2
regularization (λ = 0.005).
4.2.3. Ensemble Modeling
Let’s assume ykn (j) is the the value of the k th output layer unit of the j th
CNN model in response to the nth input test image. The linear and log-linear
ensemble classifier output for the same input would be:
Ensnlinear = arg max
k
l
X
ykn (j)
(5)
j=1
Ensnlog−linear = arg max
k
l
Y
ykn (j)
(6)
j=1
where l is the number of CNN models combined to generate the ensemble
models.
Model selection was done in two forms: single models and ensemble of models. We selected the top model (with minimum loss on the validation dataset)
from each fold of the 5-fold cross validation (Model 1-5 in Table 2). In each
network architecture and training regime, we combined the top-5 developed
single models to produce two ensembles of models using the arithmetic (5)
and geometric mean (6) of their outputs. We created 12 ensemble models
(2network architectures × 3training regimes × 2ensemble types ) in total and compared
their performance with single models.
4.3. Interobserver Study
Each solo and ensemble model developed was tested on the test dataset.
The ensemble of network 2 models trained with DFT was also tested on the
17
val review images (3.2.2) to compare human-human and model-human interobserver agreements. The resulting agreement rate (val-rater 1 and 2) was further
compared with the initial image review results. The agreement of the model
prediction with the initial review was also calculated. The general agreements
are compared and discussed in section 5.4. Kappa analysis was also done for
further validation.
Gold standard ground truth for the val review images was defined by majority voting (see Fig. 3). The agreements of the third rater with the gold
standard and the proposed ensemble model with the gold standard is calculated
and compared as well in Table 3.
4.4. Unsupervised Histological Feature Localization
For localization of the histological features, we examined the neural activation at two sites. First, the activation of neurons in the first convolutional layer
of the network 1 were visualized and the 96 feature planes were saved for review
by a neurosurgeon. Neurons that presented high activation to the location of
cellular structures in the input image were selected and seemed to be consistent
with diverse diagnostic images. Secondly, we applied a sliding window of size
227 × 227 pixels (size of network 1 input after input cropping) with stride of 79
pixels over the diagnostic CLE images (1024 × 1024 pixels). The result was a
10 × 10 matrix that provided the diagnostic value of different locations of the
input image (diagnostic map). The locations of input images corresponding to
the highest activations of the diagnostic map were detected and marked with a
bounding box. The detected features using each of these two ways were further
reviewed by a neurosurgeon.
5. Results and Discussion
We developed 42 models and tested them on 4,171 test images; accuracy
rates (agreement with the initial review) are presented in Table 2. We found that
network 2 resulted in more precise predictions about the diagnostic quality of
18
Figure 3: Interobserver study using gold standard ground truth. Gold standard was defined
using the initial review and one of the val-raters (here val-rater 1). Then, the agreement of the
ensemble model and the other val-rater (here val-rater 2) with the gold standard is calculated
to compare the human-human with model-human agreement. *If the initial review and the
val-rater 1 agreed on an image, it is added to the gold standard, otherwise it is disregarded.
images than network 1 when DT and DFT training regimes were used, while SFT
training regime resulted in slightly better accuracy of network 1, compared to
network 2. Therefore, network 2 architecture is a better feature extraction tool
for CLE images, since it concatenates multi-scale features inside its inception
modules.
5.1. Ensemble or solo model?
We did an ROC analysis for each of the two networks and three training
regimes to see how the ensemble of models performed compared to the single
19
Table 2: The accuracy of different models on the test dataset. The top-5 models crafted
from each training regime, as well as their arithmetic and geometric ensembles, were exerted.
For each network, the ensemble of DFT models makes the most accurate predictions. The
difference between arithmetic and geometric ensemble AUC was negligible.
Network
Training
Network 1
Network 2
DT
SFT
DFT
DT
SFT
DFT
Model 1
0.685
0.760
0.760
0.731
0.746
0.746
Model 2
0.658
0.749
0.755
0.750
0.746
0.805
Model 3
0.677
0.751
0.765
0.715
0.747
0.797
Model 4
0.681
0.754
0.771
0.739
0.743
0.811
Model 5
0.699
0.753
0.775
0.721
0.747
0.777
Mean
0.680
0.753
0.765
0.731
0.746
0.787
0.704
0.755
0.788
0.754
0.750
0.816
0.703
0.758
0.786
0.755
0.751
0.818
Regime
Arithmatic
Ensemble
Geometric
Ensemble
models. Fig. 2 presents the ROC curves and corresponding AUC values for each
ensemble model and the mean of single models. The AUC value increased by
2% for both networks with DT and DFT when the ensemble is applied instead
of the single model. This effect gets smaller with network 1 SFT and becomes
negligible with network 2 SFT. The two arithmetic and geometric ensemble
models produced roughly similar results (paired t-test: P value < 0.05 ).
SFT models display less sensitivity to the ensemble effect compared to DT
and DFT. This is not surprising since they represent identical models except in
the softmax classifier layer which has been adjusted through training.
5.2. Which training regime: DT, SFT and DFT?
Fig. 4 displays the results of ROC analysis when comparing the three training regimes in each network architecture and single/ensemble states. In all
20
Figure 4: Training regime effect on network 1 (left) and network 2 (right) while using single
(top) or ensemble of models (bottom). In all circumstances the AUC for DFT regime was
greater than the SFT and SFT is greater than DT, although the effect size varied.
paired comparisons, DFT outperformed SFT and SFT outperformed the DT
regime (paired t-test: P value < 0.05 ).
We traced the AUC elevation from DT to DFT regime to see how much of it
corresponded to the transformation of DT to SFT and SFT to DFT. For network
1, 70-80 % of the improvement occurred in the DT to SFT transformation,
depending on whether it’s a single or ensemble model. For network 2 ensemble
model (right bottom of Fig. 4), however, the AUC improvement caused by
transforming the training regime from DT to SFT (2%) is only 25% of the
total improvement from DT to DFT. For network 2 single model the AUC
improvement was evenly divided between the two transformations.
Our results from this experiment indicated that for our dataset, network 1
mainly required fine-tuning the classification layer and fine-tuning other layers
21
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
(j)
(k)
(l)
Figure 5: Unsupervised localization of the histopathological features from shallow and deep
neurons inside the network. First column (a, e, i) shows the input CLE images from human
glioblastoma obtained intraoperatively. Second column (b, f, j) displays activation of neurons
from the first layer (conv1, neuron 24) (shallow features); it highlights some of the cellular areas
present in the image. Third column (c, g, k) illustrates diagnostic regions of interest identified
with sliding window approach. The boxed regions represent high activation of the deepest
network neuron. Fourth column (d, h, l) contains images extracted from conv1 activation
(neuron 22), representative of the high fluorescence signal, a diagnostic sign of blood-brain
barrier disruption and leakage of fluorescent agent from the vessels into the extracellular space.
22
(feature extractors) had a smaller contribution. However, for network 2, finetuning the feature extractors was more important than modifying the classifier
layer. Though, further experiments on more datasets are necessary to generalize
this observation.
5.3. Histological features localization
8 out of total 384 reviewed colored neuron activation maps from the first
layer were selected for 4 diagnostic CLE images representative for glioma. Selected activation maps highlighted diagnostic tissue architecture patterns in
warm colors. Particularly, several maps emphasized regions of optimal image
contrast, where hypercellular and abnormal nuclear features could be identified,
and would serve as diagnostic features for image classification (Fig. 5, columns 2
and 4). Additionally, sliding window method was able to identify diagnostic aggregates of abnormally large malignant glioma cells and atypically hypercellular
areas (Fig. 5, third column).
Activation of the neurons in the first convolutional layer (conv1) were found
to highlight areas with increased fluorescein signal, a sign specific to brain tumor
regions. Increased fluorescent signal on CLE images represent areas with blood
brain barrier disruption which correspond to the tumor areas visible on a contrast enhanced MR imaging. Interestingly, sliding window method and selected
colored activation maps were not distracted or deceived by the red blood cells
contamination, as they mostly highlighted tumor and brain cells rather then
hypercellular areas due to bleeding. The proposed feature localization approach
may be useful in the future to aid in the identification of not only the diagnostic
frames, but also directing the surgeon’s attention to the image parts containing
major histopathological features.
5.4. Inter-rater agreement
Table 3 demonstrates the agreement between each of the val-raters and the
initial review on the whole val review dataset and the gold standard subset
(explained in Fig. 3). The model agreement with the initial review is larger
23
Table 3: Interobserver study results.
The model-human agreement was higher than the
human-human agreement both on whole val review dataset and the gold standard subset.
Dataset
Whole Val
Gold-Standard
Review
Rater
Val-Rater 1
General
Cohen’s
General
Agreement
Kappa
Agreement
0.32,
66 %
67%
Fair
Val-Rater 2
0.47,
73 %
75 %
Moderate
Model
0.47,
76 %
85 %
Moderate
than each val-rater’s agreement with the initial review. This suggests that the
model has successfully learned the histological features of the CLE images that
are more probable to be noticed by the neurosurgeons when the corresponding
H & E-stained histological slides were also provided for reference.
To consider images from the val review set that the majority of raters agreed
on, that is one of the val-raters agreed on with the initial review, we used the
gold standard subset. The gap between the model-human and human-human
agreements became even more evident (19% for val-rater 1 and 9% for val-rater
2) with the gold standard subset (Table 3, column 4).
6. Conclusion and future work
This paper presents a deep CNN based approach that can automatically
detect the diagnostic CLE images from brain tumor surgery. We used a manually annotated in-house dataset to train and test this approach. Our results
showed that both deep fine-tuning and creating an ensemble of models could
enhance the performance; but only their combination could reach the maximum
accuracy. The ensemble effect was stronger in DT and DFT than SFT devel-
24
oped models. The proposed method was also able to localize some histological
features of diagnostic images. Ultimately, Table 3 indicates that the proposed
ensemble of deeply fine-tuned models could detect the diagnostic images with
a higher agreement than the trained human observers. Other confocal imaging
techniques may be aided by such deep learning models. Confocal reflectance
microscopy (CRM) has been studied [36] for rapid, fluorophore-free evaluation
of pitutary adenoma biopsy specimens ex vivo. CRM allows preserving the
biopsy tissue for future permanent analysis, immunohistochemical studies, and
molecular studies. Continued use of unsupervised image segmentation methods
to detect meaningful histological features from confocal brain tumor images will
likely allow for more rapid and detailed diagnosis.
Acknowledgement
This work was supported by the Newsome Family Endowed Chair of Neurosurgery Research at the Barrow Neurological Institute held by Dr. Preul and
by funds from the Barrow Neurological Foundation.
References
[1] E. Belykh, N. L. Martirosyan, K. Yagmurlu, E. J. Miller, J. M. Eschbacher,
M. Izadyyazdanabadi, L. A. Bardonova, V. A. Byvaltsev, P. Nakaji, M. C.
Preul, Intraoperative fluorescence imaging for personalized brain tumor
resection: Current state and future directions, Frontiers in Surgery 3.
[2] P. Charalampaki, M. Javed, S. Daali, H.-J. Heiroth, A. Igressa, F. Weber, Confocal laser endomicroscopy for real-time histomorphological diagnosis: Our clinical experience with 150 brain and spinal tumor cases,
Neurosurgery 62 (2015) 171–176.
[3] S. Foersch, A. Heimann, A. Ayyad, G. A. Spoden, L. Florin, K. Mpoukouvalas, R. Kiesslich, O. Kempski, M. Goetz, P. Charalampaki, Confocal
25
laser endomicroscopy for diagnosis and histomorphologic imaging of brain
tumors in vivo, PLoS One 7 (7) (2012) e41760.
[4] N. Sanai, J. Eschbacher, G. Hattendorf, S. W. Coons, M. C. Preul, K. A.
Smith, P. Nakaji, R. F. Spetzler, Intraoperative confocal microscopy for
brain tumors: a feasibility analysis in humans, Neurosurgery 68 (2011)
ons282–ons290.
[5] A. Zehri, W. Ramey, J. Georges, M. Mooney, N. Martirosyan, M. Preul,
P. Nakaji, Neurosurgical confocal endomicroscopy: A review of contrast
agents, confocal systems, and future imaging modalities, Surgical neurology
international 5 (2014) 60.
[6] N. L. Martirosyan, J. M. Eschbacher, M. Y. S. Kalani, J. D. Turner, E. Belykh, R. F. Spetzler, P. Nakaji, M. C. Preul, Prospective evaluation of
the utility of intraoperative confocal laser endomicroscopy in patients with
brain neoplasms using fluorescein sodium: experience with 74 cases, Neurosurgical focus 40 (3) (2016) E11.
[7] H. Greenspan, B. van Ginneken, R. M. Summers, Guest editorial deep
learning in medical imaging: Overview and future promise of an exciting
new technique, IEEE Transactions on Medical Imaging 35 (5) (2016) 1153–
1159.
[8] G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi, M. Ghafoorian, J. A. van der Laak, B. van Ginneken, C. I. Sánchez, A survey on deep
learning in medical image analysis, arXiv preprint arXiv:1702.05747.
[9] D. Ciresan, A. Giusti, L. M. Gambardella, J. Schmidhuber, Deep neural
networks segment neuronal membranes in electron microscopy images, in:
Advances in neural information processing systems, 2012, pp. 2843–2851.
[10] J. Zhao, M. Zhang, Z. Zhou, J. Chu, F. Cao, Automatic detection and
classification of leukocytes using convolutional neural networks, Medical &
biological engineering & computing (2016) 1–15.
26
[11] K. Sirinukunwattana, S. E. A. Raza, Y.-W. Tsang, D. R. Snead, I. A. Cree,
N. M. Rajpoot, Locality sensitive deep learning for detection and classification of nuclei in routine colon cancer histology images, IEEE transactions
on medical imaging 35 (5) (2016) 1196–1206.
[12] H.-I. Suk, D. Shen, Deep ensemble sparse regression network for alzheimers
disease diagnosis, in: International Workshop on Machine Learning in Medical Imaging, Springer, 2016, pp. 113–121.
[13] J. Shi, X. Zheng, Y. Li, Q. Zhang, S. Ying, Multimodal neuroimaging
feature learning with multimodal stacked deep polynomial networks for
diagnosis of alzheimer’s disease, IEEE journal of biomedical and health
informatics.
[14] S. S. M. Salehi, D. Erdogmus, A. Gholipour, Auto-context convolutional
neural network (auto-net) for brain extraction in magnetic resonance imaging, IEEE Transactions on Medical Imaging.
[15] M. Ghafoorian, N. Karssemeijer, T. Heskes, M. Bergkamp, J. Wissink,
J. Obels, K. Keizer, F.-E. de Leeuw, B. van Ginneken, E. Marchiori, et al.,
Deep multi-scale location-aware 3d convolutional neural networks for automated detection of lacunes of presumed vascular origin, NeuroImage:
Clinical 14 (2017) 391–399.
[16] L. Zhao, K. Jia, Multiscale cnns for brain tumor segmentation and diagnosis, Computational and mathematical methods in medicine 2016.
[17] D. Mahapatra, P. K. Roy, S. Sedai, R. Garnavi, Retinal image quality
classification using saliency maps and cnns, in: International Workshop on
Machine Learning in Medical Imaging, Springer, 2016, pp. 172–179.
[18] A. Abdi, C. Luong, T. Tsang, J. Jue, D. Hawley, S. Fleming, K. Gin,
J. Swift, R. Rohling, P. Abolmaesumi, Automatic quality assessment of
echocardiograms using convolutional neural networks: Feasibility on the
apical four-chamber view, IEEE Transactions on Medical Imaging.
27
[19] Y. Gao, M. A. Maraci, J. A. Noble, Describing ultrasound video content
using deep convolutional neural networks, in: Biomedical Imaging (ISBI),
2016 IEEE 13th International Symposium on, IEEE, 2016, pp. 787–790.
[20] A. Kumar, P. Sridar, A. Quinton, R. K. Kumar, D. Feng, R. Nanan, J. Kim,
Plane identification in fetal ultrasound images using saliency maps and
convolutional neural networks, in: Biomedical Imaging (ISBI), 2016 IEEE
13th International Symposium on, IEEE, 2016, pp. 791–794.
[21] J. Yosinski, J. Clune, Y. Bengio, H. Lipson, How transferable are features
in deep neural networks?, in: Advances in neural information processing
systems, 2014, pp. 3320–3328.
[22] N. Tajbakhsh, J. Y. Shin, S. R. Gurudu, R. T. Hurst, C. B. Kendall, M. B.
Gotway, J. Liang, Convolutional neural networks for medical image analysis: full training or fine tuning?, IEEE transactions on medical imaging
35 (5) (2016) 1299–1312.
[23] T. G. Dietterich, et al., Ensemble methods in machine learning, Multiple
classifier systems 1857 (2000) 1–15.
[24] Z.-H. Zhou, J. Wu, W. Tang, Ensembling neural networks: many could be
better than all, Artificial intelligence 137 (1-2) (2002) 239–263.
[25] D. Ciregan, U. Meier, J. Schmidhuber, Multi-column deep neural networks
for image classification, in: Computer Vision and Pattern Recognition
(CVPR), 2012 IEEE Conference on, IEEE, 2012, pp. 3642–3649.
[26] A. Kumar, J. Kim, D. Lyndon, M. Fulham, D. Feng, An ensemble of finetuned convolutional neural networks for medical image classification, IEEE
journal of biomedical and health informatics 21 (1) (2017) 31–40.
[27] S.
Christodoulidis,
S. Mougiakakou,
M.
Anthimopoulos,
L.
Ebner,
A.
Christe,
Multisource transfer learning with convolutional
neural networks for lung pattern analysis, IEEE journal of biomedical and
health informatics 21 (1) (2017) 76–84.
28
[28] A. Krogh, J. Vedelsby, Neural network ensembles, cross validation, and
active learning, in: Advances in neural information processing systems,
1995, pp. 231–238.
[29] M. Izadyyazdanabadi, E. Belykh, N. Martirosyan, J. Eschbacher, P. Nakaji,
Y. Yang, M. C. Preul, Improving utility of brain tumor confocal laser endomicroscopy: objective value assessment and diagnostic frame detection
with convolutional neural networks, in: SPIE Medical Imaging, International Society for Optics and Photonics, 2017, pp. 101342J–101342J.
[30] Y. LeCun, Y. Bengio, The handbook of brain theory and neural networks,
MIT Press, Cambridge, MA, USA, 1998, Ch. Convolutional networks for
images, speech, and time series, pp. 255–258.
URL http://dl.acm.org/citation.cfm?id=303568.303704
[31] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick,
S. Guadarrama, T. Darrell, Caffe: Convolutional architecture for fast feature embedding, arXiv preprint arXiv:1408.5093.
[32] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with
deep convolutional neural networks, in: Advances in neural information
processing systems, 2012, pp. 1097–1105.
[33] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov,
Dropout: A simple way to prevent neural networks from overfitting, The
Journal of Machine Learning Research 15 (1) (2014) 1929–1958.
[34] C. E. Metz, Basic principles of roc analysis, in: Seminars in nuclear
medicine, Vol. 8, Elsevier, 1978, pp. 283–298.
[35] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, A. Rabinovich, Going deeper with convolutions, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 1–9.
29
[36] M. A. Mooney, J. Georges, M. I. Yazdanabadi, K. Y. Goehring, W. L.
White, A. S. Little, M. C. Preul, S. W. Coons, P. Nakaji, J. M. Eschbacher, Immediate ex-vivo diagnosis of pituitary adenomas using confocal
reflectance microscopy: a proof-of-principle study, Journal of Neurosurgery.
30
| 1cs.CV
|
T HE LOSS SURFACE AND EXPRESSIVITY OF
DEEP CONVOLUTIONAL NEURAL NETWORKS
Quynh Nguyen and Matthias Hein
Department of Mathematics and Computer Science
Saarland Informatics Campus, Saarland University, Germany
arXiv:1710.10928v1 [cs.LG] 30 Oct 2017
A BSTRACT
We analyze the expressiveness and loss surface of practical deep convolutional
neural networks (CNNs) with shared weights and max pooling layers. We show
that such CNNs produce linearly independent features at a “wide” layer which
has more neurons than the number of training samples. This condition holds e.g.
for the VGG network. Furthermore, we provide for such wide CNNs necessary
and sufficient conditions for global minima with zero training error. For the case
where the wide layer is followed by a fully connected layer we show that almost
every critical point of the empirical loss is a global minimum with zero training
error. Our analysis suggests that both depth and width are very important in deep
learning. While depth brings more representational power and allows the network
to learn high level features, width smoothes the optimization landscape of the
loss function in the sense that a sufficiently wide network has a well-behaved loss
surface with potentially no bad local minima.
1
I NTRODUCTION
It is well known that the optimization problem for training neural networks can have exponentially
many local minima (Auer et al., 1996; Safran & Shamir, 2016) and NP-hardness has been shown in
many cases (Blum & Rivest., 1989; Sima, 2002; Livni et al., 2014; Shamir, 2017; Shalev-Shwartz
et al., 2017). However, it has been empirically observed (Dauphin et al., 2014; Goodfellow et al.,
2015) that the training of state-of-the-art deep CNNs (LeCun et al., 1990; Krizhevsky et al., 2012),
which are often overparameterized, is not hampered by suboptimal local minima.
In order to explain the apparent gap between hardness results and practical performance, many interesting theoretical results have been recently developed (Andoni et al., 2014; Sedghi & Anandkumar,
2015; Janzamin et al., 2016; Haeffele & Vidal, 2015; Gautier et al., 2016; Brutzkus & Globerson,
2017; Soltanolkotabi, 2017; Soudry & Hoffer, 2017; Goel & Klivans, 2017; Du et al., 2017; Zhong
et al., 2017; Tian, 2017; Li & Yuan, 2017) in order to identify conditions under which one can guarantee that local search algorithms like gradient descent converge to the globally optimal solution.
However, it turns out that these approaches are either not practical as they require e.g. knowledge
about the data generating measure, or a modification of network structure and objective, or they
are for quite restricted network structures, mostly one hidden layer networks, and thus are not able
to explain the success of deep networks in general. For deep linear networks one has achieved a
quite complete picture of the loss surface as it has been shown that every local minimum is a global
minimum (Baldi & Hornik, 1988; Kawaguchi, 2016; Freeman & Bruna, 2017; Hardt & Ma, 2017;
Yun et al., 2017). By randomizing the nonlinear part of a feedforward network with ReLU activation function and making some additional simplifying assumptions, Choromanska et al. (2015a) can
relate the loss surface of neural networks to a certain spin glass model. In this model the objective of
local minima is close to the global optimum and the number of bad local minima decreases quickly
with the distance to the global optimum. This is a very interesting result but is based on a number of unrealistic assumptions (Choromanska et al., 2015b). More recently, Nguyen & Hein (2017)
have analyzed deep fully connected networks with general activation functions and could show that
almost every critical point is a global minimum if one layer has more neurons than the number
of training points. While this result holds for networks in practice, it requires a quite extensively
overparameterized network.
1
In this paper we overcome the restriction of previous work in several ways. This paper is one of
the first ones, which studies deep CNNs. CNNs are of high practical interest as they learn very
useful representations (Zeiler & Fergus, 2014; Mahendran & Vedaldi, 2015; Yosinski et al., 2015)
with a small number of parameters. We are only aware of Cohen & Shashua (2016) who study
the expressiveness of CNNs with max-pooling layer and ReLU activation but with rather unrealistic
filters (just 1 × 1) and no shared weights. In our setting we allow as well max pooling and general
activation functions. Moreover, we can have an arbitrary number of filters and we study general
convolutions as the filters need not be applied to regular structures like 3 × 3 but can be patch-based
where the only condition is that all the patches have the size of the filter. Convolutional layers,
fully connected layers and max-pooling layers can be combined in almost arbitrary order. We study
in this paper the expressiveness and loss surface of such a CNN where one layer is wide, in the
sense that it has more neurons than the number of training points. While this assumption sounds at
first quite strong, we want to emphasize that the VGG network (Simonyan & Zisserman, 2015) and
other CNNs, see Table 1, fulfills this condition. We show that such wide CNNs produce linearly
independent feature representations at the wide layer and thus are able to fit the training data exactly
(universal finite sample expressivity). This is even true if the bottom layers (from input to the wide
layer) are chosen randomly with probability one. We think that this explains partially the results
of Zhang et al. (2017) where they show experimentally for several CNNs that they are able to fit
random labels. Moreover, we provide necessary and sufficient conditions for global minima with
zero squared loss and show for a particular class of CNNs that almost all critical points are globally
optimal, which to some extent explains why such wide CNNs can be optimized so efficiently.
2
D EEP C ONVOLUTIONAL N EURAL N ETWORKS
We first introduce our notation and definition of CNNs. Let N be the number of training samples
and denote by X = [x1 , . . . , xN ]T ∈ RN ×d , Y = [y1 , . . . , yN ]T ∈ RN ×m the input resp. output
matrix for the training data (xi , yi )N
i=1 , where d is the input dimension and m the number of classes.
Let L be the number of layers of the network, where each layer is either a convolutional, maxpooling or fully connected layer. The layers are indexed from k = 0, 1, . . . , L which corresponds to
input layer, 1st hidden layer, . . ., and output layer. Let nk be the width of layer k and fk : Rd → Rnk
the function that computes for every input its feature vector at layer k.
The convolutional layer consists of a set of patches of equal length where every patch is a subset of
neurons from the same layer. Throughout this paper, we assume that the patches of every layer cover
the whole layer, i.e. every neuron belongs to at least one of the patches, and that there are no patches
that contain exactly the same subset of neurons. This means that if one patch covers the whole layer
then it must be the only patch of the layer. Let Pk and lk be the number of
patches resp. the size
of each patch at layer k for every 0 ≤ k <nL. For every input o
x ∈ Rd , let x1 , . . . , xP0 ∈ Rl0
denote the set of patches at input layer and fk1 (x), . . . , fkPk (x) ∈ Rlk the set of patches at layer
k. We denote by Tk the number of convolutional filters and by Wk = [wk1 , . . . , wkTk ] ∈ Rlk−1 ×Tk
the corresponding parameter matrix of the convolutional layer k for every 1 ≤ k < L. Each column
of Wk corresponds to one filter whose size is equal to the patch size of previous layer, that is, lk−1 .
Furthermore, bk ∈ Rnk denotes the bias vector and σk : R → R the activation function for each
layer. Note that one can use the same activation function for all layers but we use the general form
to highlight the role of different layers. In this paper, all functions are applied componentwise, and
we denote by [a] the set of integers {1, 2, . . . , a} and by [a, b] the set of integers from a to b.
Definition 2.1 (Convolutional layer) A layer k is called a convolutional layer if its output fk (x) ∈
Rnk is defined for every x ∈ Rd as
p
fk (x)h = σk wkt , fk−1
(x) + (bk )h
(1)
for every p ∈ [Pk−1 ], t ∈ [Tk ], h := (p − 1)Tk + t.
The value of each neuron at layer k is computed by first taking the inner product between a filter
of layer k and a patch at layer k − 1, adding the bias and then applying the activation function.
The number of neurons at layer k is thus nk = Tk Pk−1 , which we denote as the width of layer k.
2
Our definition of a convolutional layer is quite general as every patch can be an arbitrary subset of
neurons of the same layer and thus covers most of existing variants in practice.
Definition 2.1 includes the fully connected layer as a special case by using Pk−1 = 1, lk−1 =
1
nk−1 , fk−1
(x) = fk−1 (x) ∈ Rnk−1 , Tk = nk , Wk ∈ Rnk−1 ×nk , bk ∈ Rnk . Thus we have only one
patch which is the whole feature vector at this layer.
Definition 2.2 (Fully connected layer) A layer k is called a fully connected layer if its output
fk (x) ∈ Rnk is defined for every x ∈ Rd as
fk (x) = σk WkT fk−1 (x) + bk .
(2)
For some results we also allow max-pooling layers.
Definition 2.3 (Max-pooling layer) A layer k is called a max-pooling layer if its output fk (x) ∈
Rnk is defined for every x ∈ Rd and p ∈ [Pk−1 ] as
p
p
fk (x)p = max (fk−1
(x))1 , . . . , (fk−1
(x))l
.
(3)
k−1
A max-pooling layer just computes the maximum element of every patch from the previous layer.
Since there are Pk−1 patches at layer k − 1, the number of output neurons at layer k is nk = Pk−1 .
Reformulation of Convolutional Layers: For each convolutional or fully connected layer, we
denote by Mk : Rlk−1 ×Tk → Rnk−1 ×nk the linear map that returns for every parameter matrix
Wk ∈ Rlk−1 ×Tk the corresponding full weight matrix Uk = Mk (Wk ) ∈ Rnk−1 ×nk . For convolutional layers, Uk can be seen as the counterpart of the weight matrix Wk in fully connected layers.
We define Uk = Mk (Wk ) = Wk if layer k is fully connected. Note that the mapping Mk depends on the patch structure of each convolutional layer k. For example, suppose that layer k has
"
#
a d
1
2
two filters of length 3, that is, Wk = [wk , wk ] = b e , and nk−1 = 5 and patches given by
c f
a b c 0 0
d e f 0 0
0 a b c 0
T
T
a 1D-convolution with stride 1 and no padding then: Uk = Mk (Wk ) =
,
0 d e f 0
0 0 a b c
0 0 d e f
The above ordering of the rows of UkT of a convolutional layer is determined by Equation (1), in
particular, the row index h of UkT is calculated as h = (p − 1)Tk + t, which means for every given
patch p one has to loop over all the filters t and compute the corresponding value of the output unit
by taking the inner product of the h-th row of UkT with the whole feature vector of the previous layer.
We assume throughout this paper that that there is no non-linearity at the output layer. By ignoring
max-pooling layers for the moment, the feature maps fk : Rd → Rnk can be written as
f0 (x) = x,
fk (x) = σk gk (x) , where gk (x) = UkT fk−1 (x) + bk , ∀ 1 ≤ k ≤ L − 1
fL (x) = gL (x) = ULT fL−1 (x) + bL ,
where gk : Rd → Rnk is the pre-activation output at layer k. By stacking the feature vectors of layer
k of all training samples, before and after applying the activation function, into a matrix, we define:
Fk = [fk (x1 ), . . . , fk (xN )]T ∈ RN ×nk ,
and
Gk = [gk (x1 ), . . . , gk (xN )]T ∈ RN ×nk .
In this paper, we refer to Fk as the output matrix at layer k. It follows from above that
F0 = X,
Fk = σk (Gk ), where Gk = Fk−1 Uk + 1N bTk ,
FL = GL = FL−1 UL +
1N bTL .
∀1 ≤ k ≤ L − 1
(4)
(5)
In this paper, we assume the following general condition on the structure of convolutional layers.
3
Assumption 2.4 (Convolutional Structure) For every convolutional layer k, there exists at least
one parameter matrix Wk ∈ Rlk−1 ×Tk for which the corresponding weight matrix Uk =
Mk (Wk ) ∈ Rnk−1 ×nk has full rank.
It is straightforward to see that Assumption 2.4 is satisfied if every neuron of a convolutional layer
belongs to at least one patch and there are no identical patches.
Lemma 2.5 If Assumption 2.4 holds, then for every convolutional layer k, the set of Wk ∈ Rlk−1 ×Tk
for which Uk = Mk (Wk ) ∈ Rnk−1 ×nk does not have full rank has Lebesgue measure zero.
Proof: Since Uk = Mk (Wk ) ∈ Rnk−1 ×nk and Mk is a linear map, every entry of Uk is a linear
function of the entries of Wk . Let m = min(nk−1 , nk ), then the set of low rank matrices Uk
k−1 ,nk )
is characterized by a system of equations where the max(nm
determinants of all m × m
sub-matrices of Uk are zero. As the determinant is a polynomial in the entries of the matrix and thus
a real analytic function, and the composition of analytic functions is again analytic, we get that each
determinant is a real analytic function of Wk . By Assumption 2.4, there exists at least one Wk such
that one of these determinants is non-zero. Thus by Lemma A.2, the set of Wk where this determinant is zero has Lebesgue measure zero. As all the sub-matrices need to have low rank in order
that Uk has low rank, we get that the set of Wk where Uk has low rank has Lebesgue measure zero.
3
W IDE CNN S CAN LEARN LINEARLY INDEPENDENT FEATURES
In this section, we show that a class of standard CNN architectures with convolutional layers, fully
connected layers and max-pooling layers plus standard activation functions like ReLU, sigmoid,
softplus, etc are able to learn linearly independent features at any hidden layer if that layer has more
neurons than the number of training samples. Our assumption on training data is the following.
Assumption 3.1 (Training data) The patches of different training samples are non-identical, that
is, xpi 6= xqj for every p, q ∈ [P0 ], i, j ∈ [N ], i 6= j.
Assumption 3.1 is quite weak, especially if the size of the input patches is larger. If the assumption
does not hold, one can add a small perturbation to the training samples: {x1 + 1 , . . . , xN + N } .
N
The set of {i }i=1 where Assumption 3.1 is not fulfilled for the new dataset has Lebesgue measure
N
zero. Moreover, {i }i=1 can be chosen arbitrarily small so that the influence of the perturbation is
negligible. Our main assumptions on the activation function of the hidden layers is the following.
Assumption 3.2 (Activation function) The activation function σ is continuous, non-constant, and
satisfies one of the following conditions:
• ∃µ+ , µ− ∈ R s.t. lim σk (t) = µ− and lim σk (t) = µ+ and µ+ µ− = 0
t→−∞
t→∞
• ∃ρ1 , ρ2 , ρ3 , ρ4 ∈ R+ s.t. |σ(t)| ≤ ρ1 eρ2 t for t < 0 and |σ(t)| ≤ ρ3 t + ρ4 for t ≥ 0
Assumption 3.2 covers several standard activation functions.
Lemma 3.3 The following activation functions satisfy Assumption 3.2:
ReLU: σ(t) = max(0, t), Sigmoid: σ(t) =
1
1
, Softplus: σα (t) = ln(1 + eαt ) for α > 0.
1 + e−t
α
The softplus function is a smooth approximation of ReLU. It holds:
1
ln(1 + eαt ) = max(0, t).
α
The first main result of this paper is the following.
lim σα (t) = lim
α→∞
α→∞
(6)
Theorem 3.4 (Linearly Independent Features) Let Assumption 3.1 hold for the training sample.
Consider a deep CNN architecture for which there exists some layer 1 ≤ k ≤ L − 1 such that
4
1. Layer 1 and layer k are convolutional or fully connected while all the other layers can be
convolutional, fully connected or max-pooling
2. The width of layer k is larger than the number of training samples, nk = Tk Pk−1 ≥ N
3. (σ1 , . . . , σk ) satisfy Assumption 3.2
Then there exist a set of parameters of the first k layers (Wl , bl )kl=1 such that the set of feature
vectors {fk (x1 ), . . . , fk (xN )} are linearly independent. Moreover, (Wl , bl )kl=1 can be chosen in
such a way that all the weight matrices Ul = Ml (Wl ) have full rank for every 1 ≤ l ≤ k.
Theorem 3.4 implies that a large class of CNNs employed in practice with convolutional, fully
connected and max-pooling layers and standard activation functions like ReLU, sigmoid or softplus
can produce linearly independent features at any hidden layer if its width is larger than the number
of training samples. Figure 1 shows an example of a CNN architecture that satisfies the conditions
of Theorem 3.4 at the first convolutional layer (i.e. k = 1). Note that if a set of vectors are linearly
independent then they are also linearly separable. In this sense, Theorem 3.4 suggests that deep and
wide CNNs can produce linearly separable features at every wide hidden layer.
Linear separability in neural networks has been recently studied by An et al. (2015), where the
authors show that a two-hidden-layer fully connected network with ReLU activation function can
transform any training set to be linearly separable while approximately preserving the distances of
the training data at the output layer. Compared to An et al. (2015) our Theorem 3.4 is derived for
CNNs with a wider range of activation functions. Moreover, our result shows even linear independence of features which is stronger than linear separability. Recently, Nguyen & Hein (2017) have
shown a similar result for deep fully connected networks and analytic activation functions.
We note that, in contrast to fully connected networks, for CNNs the condition nk ≥ N of Theorem
3.4 does not necessarily imply that the network has a huge number of parameters as the layers k and
k + 1 can be chosen to be convolutional. In particular, the condition nk = Tk Pk−1 ≥ N can be
fulfilled by increasing the number of filters Tk or by using a large number of patches Pk−1 (however
Pk−1 is upper bounded by nk ), which is however only possible if lk−1 is small as otherwise our
condition on the patches cannot be fulfilled. Interestingly, such a architecture has been used in the
VGG-Net (Simonyan & Zisserman, 2015), where they use small 3 × 3 filters and minimal stride 1 in
the first layer and thus they fulfill the condition nk ≥ N for k = 1, see Table 1, for ImageNet. Also
note that other state-of-the-art-networks fulfill the condition in Table 1. Overall, Theorem 3.4 can be
seen as a theoretical support for the usage of small filters and strides in practical CNN architectures
as it increases the chance of achieving linear separability at early hidden layers in the network and
also reduces the total number of training parameters. The reason why linear separability helps will
be discussed in Section 4 when we analyze the loss surface of the CNNs. Note also that the condition
nk ≥ N is a sufficient condition but not necessary to prove our results. In particular, we conjecture
that linear separability might hold with far less number of neurons in practical applications.
One might ask now how difficult it is to find such parameters which generate linearly independent
features at a hidden layer? Our next result shows that once analytic activation functions 1 , e.g.
sigmoid or softplus, are used at the first k hidden layers of the network, the linear independence of
features at layer k holds with probability 1 even if one draws the parameters of the first k layers
(Wl , bl )k randomly for any distribution which is absolutely continuous w.r.t. the Lebesgue measure.
Theorem 3.5 Let Assumption 3.1 hold for the training samples. Consider a deep CNN for which
there exists some layer 1 ≤ k ≤ L − 1 such that
1. Every layer from 1 to k is convolutional or fully connected
2. The width of layer k is larger than number of training samples, that is, nk = Tk Pk−1 ≥ N
3. (σ1 , . . . , σk ) are real analytic functions and satisfy Assumption 3.2.
Then the set of parameters of the first k layers (Wl , bl )kl=1 for which the set of feature vectors
{fk (x1 ), . . . , fk (xN )} of layer k are not linearly independent has Lebesgue measure zero.
1
A function σ : R → R is real analytic if its Taylor series about x0 converges to σ(x0 ) on some neighborhood of x0 for every x0 ∈ R (Krantz & Parks, 2002).
5
Note that Theorem 3.5 is a much stronger statement than Theorem 3.4, as it shows that for almost
all weight configurations one gets linearly independent features at the wide layer. While Theorem
3.5 does not hold for the ReLU activation function as it is not an analytic function, we want to note
again that one can approximate the ReLU function arbitrarily well using the softplus function (see
Equation 6), which is an analytic function for any α > 0 and thus Theorem 3.5 applies. It is an open
question if the result holds also for the ReLU activation function itself.
It is very interesting to note that Theorem 3.5 explains previous empirical observations. In particular,
Czarnecki et al. (2017) have shown empirically that linear separability is often obtained already in
the first few hidden layers of the trained networks. This is done by attaching a linear classifier probe
(Alain & Bengio, 2016) to every hidden layer in the network after training the whole network with
backpropagation. The fact that Theorem 3.5 holds even if the parameters of the bottom layers up
to the wide layer k are chosen randomly is also in line with recent empirical observations for CNN
architectures that one has little loss in performance if the weights of the initial layers are chosen
randomly without training (Jarrett et al., 2009; Saxe et al., 2011; Yosinski et al., 2014).
An application of Theorem 3.4 yields the following universal finite sample expressivity for CNNs.
In particular, a deep CNN architecture with scalar output can perfectly express the values of any
scalar-valued function over a finite number of inputs as long as the width of the last hidden layer is
larger than the number of training samples.
Corollary 3.6 (Finite Sample Expressivity) Let Assumption 3.1 hold for the training samples.
Consider a standard CNN with scalar output which satisfies the conditions of Theorem 3.4 at the
last hidden layer k = L − 1. Let fL : Rd → R be the output of the network given as
fL (x) =
m
X
λj f(L−1)j (x) ∀x ∈ Rd
j=1
where λ ∈ RnL−1 is the weight vector of the last layer. Then for every target output y ∈ RN , there
always exist λ, (Wl , bl )L−1
so that it holds fL (xi ) = yi for every i ∈ [N ].
l=1
Proof: Since the network satisfies the conditions of Theorem 3.4 for k = L − 1, there exist a set of
T
N
parameters (Wl , bl )L−1
l=1 such that rank(FL−1 ) = N. Let FL = [fL (x1 ), . . . , fL (xN )] ∈ R then
−1
T
T
it follows that FL = FL−1 λ. Pick λ = FL−1 (FL−1 FL−1 ) y then it holds FL = FL−1 λ = y.
The expressivity of neural networks have been well-studied in the literature, in particular in the
universal approximation theorems for one hidden layer networks (Cybenko, 1989; Hornik et al.,
1989). Recently, many results have be shown why deep networks are superior to shallow networks
in terms of expressiveness (Delalleau & Bengio, 2011; Telgarsky, 2016; 2015; Eldan & Shamir,
2016; Safran & Shamir, 2017; Yarotsky, 2016; Poggio et al., 2016; Liang & Srikant, 2017; Mhaskar
& Poggio, 2016; Montufar et al., 2014; Pascanu et al., 2014; Raghu et al., 2017). While most of
these results are derived for fully connected networks, it seems that Cohen & Shashua (2016) are the
first ones who study expressivity of CNNs. In particular, they show that CNNs with max-pooling
and ReLU units are universal in the sense that they can approximate any given function if the size
of the networks is unlimited. However, the number of convolutional filters in this result has to grow
exponentially with the number of patches and they do not allow shared weights in their result, which
is a standard feature of CNNs. Our Corollary 3.6 shows universal finite sample expressivity, instead
of universal function approximation, even for L = 2 and k = 1, that is a single convolutional layer
network can perfectly fit the training data as long as the number of hidden units is not smaller than
the number of samples. To the best of our knowledge, this is the first result on universal finite sample
expressivity for a large class of practical CNNs.
For fully connected networks, universal finite sample expressivity has been studied by Zhang et al.
(2017); Nguyen & Hein (2017); Hardt & Ma (2017). They show that a network with a single hidden
layer with N hidden units can express any training set of size N . While the number of training
parameters of a single hidden layer CNN with N hidden units and scalar output is just T1 l0 + 2N ,
where T1 is the number of convolutional filters and l0 is the length of each filter, it is N d + 2N for
fully connected networks. If we set the width of the hidden layer of the CNN as n1 = T1 P0 = N
in order to fulfill the condition of Corollary 3.6, then the number of training parameters of the CNN
becomes 2N + N l0 /P0 , which is less than 3N if l0 ≤ P0 compared to (d + 2)N for the fully
connected case. In practice one almost always has l0 ≤ P0 as l0 is typically a small integer and P0
6
Figure 1: An example of CNN for a given training set of size N ≤ 100 × 26 × 26 = 67600.
The width of each layer is d = n0 = 784, n1 = 67600, n2 = 16900, n3 = 2880, n4 = 720, n5 =
100, n6 = m = 10. One can see that n1 ≥ N and the network has pyramidal structure from layer 2
till the output layer, that is, n2 ≥ . . . ≥ n6 .
Table 1: The width of the first convolutional layer (n1 ) and the maximum width of all the hidden layers (max1≤k≤L−1 nk ) of state-of-the-art CNN architectures in comparison with the size of
ImageNet (N ≈ 1200K). All numbers are lower bounds on the true width.
CNN Architecture
VGG(A-E) (Simonyan & Zisserman, 2015)
InceptionV3 (Szegedy et al., 2015b)
InceptionV4 (Szegedy et al., 2016)
SqueezeNet (Iandola et al., 2016)
Enet (Paszke et al., 2016)
GoogLeNet (Szegedy et al., 2015a)
ResNet (He et al., 2016)
Xception (Chollet, 2016)
n1
3000K
700K
700K
1180K
1000K
800K
800K
700K
maxk nk
3000K
1300K
1300K
1180K
1000K
800K
800K
700K
is on the order of the dimension of the input. Thus, the number of parameters of the CNN to achieve
universal finite sample expressivity is significantly smaller than that of fully connected networks.
Obviously, in practice it is more important that the network generalizes rather than just fitting the
training data. By using shared weights and sparsity structure, CNNs seem to implicitly regularize
the model class in order to achieve good generalization performance. Thus even though they can
fit also random labels or noise (Zhang et al., 2017) due to the universal finite sample expressivity
shown in Corollary 3.6, they seem still to be able to generalize well (Zhang et al., 2017).
4
T HE L OSS S URFACE OF C ONVOLUTIONAL N EURAL N ETWORKS
In this section, we restrict our analysis to the use of least squares loss. However, as we show later
that the network can produce exactly the target output (i.e. FL = Y ) for some choice of parameters,
all our results can also be extended to any other loss function where the global minimum is attained
at FL = Y , for instance the squared Hinge-loss analyzed in Nguyen & Hein (2017). Let P denote
the space of all parameters of the network. The final training objective Φ : P → R is given as
1
2
kFL − Y kF
(7)
Φ (Wl , bl )L
l=1 =
2
where FL is defined as in (5), which is also the same as
FL = σL−1 (. . . σ1 (XU1 + b1 ) . . .)UL + bL ,
where Ul = Ml (Wl ) for every 1 ≤ l ≤ L.
Our assumptions on the architecture of CNNs is given below.
Assumption 4.1 (CNN Architecture) Every layer in the network is a convolutional layer or fully
connected layer and the output layer is fully connected. Moreover, there exists some hidden layer
1 ≤ k ≤ L − 1 such that the following holds:
7
• The width of layer k is larger than number of training samples, that is, nk = Tk Pk−1 ≥ N
• All the activation functions of the hidden layers (σ1 , . . . , σL−1 ) satisfy Assumption 3.2
• (σk+1 , . . . , σL−1 ) are strictly increasing or strictly decreasing, and differentiable
• The network is pyramidal from layer k + 1 till the output layer, that is, nk+1 ≥ . . . ≥ nL
A typical example that satisfies Assumption 4.1 is the following (see Figure 1 for an illustration):
•
•
•
•
•
The first layer is a convolutional layer with n1 = T1 P0 ≥ N
Every layer from layer 2 till the output layer is convolutional or fully connected
(σ1 , . . . , σk ) can be ReLU, sigmoid or softplus
(σk+1 , . . . , σL−1 ) can be sigmoid or softplus
n2 ≥ n3 ≥ . . . ≥ nL
One can easily check that the above example satisfies Assumption 4.1 for k = 1.
In the following, let us define for every 1 ≤ k ≤ L − 1 the subset Sk ⊆ P such that
Sk := (Wl , bl )L
rank(Fk ) = N and Ul has full rank for every l ∈ [k + 2, L] .
l=1
The set Sk is the set of parameters where the feature vectors at layer k are linearly independent and
all the weight matrices from layer k + 2 till the output layer have full rank. In the following, we
examine conditions for global optimality in Sk . It is important to note that Sk covers almost the
whole parameter space under an additional mild condition on the activation function.
Lemma 4.2 Let Assumption 3.1 hold for the training sample and let the CNN architecture satisfy
Assumption 4.1 for some layer 1 ≤ k ≤ L − 1. We assume further that the activation functions of
the first k layers (σ1 , . . . , σk ) are real analytic. Then the set P \ Sk has Lebesgue measure zero.
Proof: One can see that
P \ Sk ⊆ (Wl , bl )L
l=1
rank(Fk ) < N ∪ (Wl , bl )L
Ul has low rank for some layer l .
l=1
By Theorem 3.5, it holds that the set (Wl , bl )L
l=1 rank(Fk ) < N has Lebesgue measure zero.
Moreover, it follows from Lemma 2.5 that the set (Wl , bl )L
Ul has low rank for some layer l
l=1
also has measure zero. Thus, P \ Sk has Lebesgue measure zero.
In the next key lemma, we bound the objective function in terms of its gradient magnitude w.r.t. the
weight matrix of layer k for which nk ≥ N. For every matrix A ∈qRm×n , let σmin (A) and σmax (A)
P
2
denote the smallest and largest singular value of A. Let kAkF =
i,j Aij , kAkmin := min |Aij |
i,j
and kAkmax := max |Aij |. From Equations (4), (5) and (7), it follows that Φ can be seen as a function
i,j
of (Ul , bl )L
l=1 , and thus we can use ∇Uk Φ. If layer k is fully connected then Uk = Mk (Wk ) = Wk
and thus ∇Uk Φ = ∇Wk Φ. Otherwise, if layer k is convolutional then we note that ∇Uk Φ is “not”
the true gradient of the training objective because Uk is not the true optimization parameter but
Wk . In this case, the true gradient of Φ w.r.t. to the true parameter matrix Wk which consists of
convolutional filters can be computed via chain rule as
X ∂Φ ∂(Uk )ij
∂Φ
=
∂(Wk )rs
∂(Uk )ij ∂(Wk )rs
i,j
Please note that even though we write the partial derivatives with respect to the matrix elements,
∇Wk Φ resp. ∇Uk Φ are the matrices of the same dimension as Wk resp. Uk in the following.
Lemma 4.3 Consider a standard deep CNN which satisfies Assumption 4.1 for some hidden layer
1 ≤ k ≤ L − 1. Then it holds
∇Uk+1 Φ
F
≥ σmin (Fk )
L−1
Y
σmin (Ul+1 ) kσl0 (Gl )kmin kFL − Y kF
l=k+1
8
and
∇Uk+1 Φ
F
L−1
Y
≤ σmax (Fk )
σmax (Ul+1 ) kσl0 (Gl )kmax kFL − Y kF .
l=k+1
Our next main result is motivated by the fact that empirically when training over-parameterized
neural networks with shared weights and sparsity structure like CNNs, there seem to be no problems
with sub-optimal local minima. In many cases, even when training labels are completely random,
local search algorithms like stochastic gradient descent can converge to a solution with almost zero
training error (Zhang et al., 2017). To understand better this phenomenon, we first characterize in
the following Theorem 4.4 the set of points in parameter space with zero loss, and then analyze in
Theorem 4.5 the loss surface for a special case of the network. We emphasize that our results hold
for standard deep CNNs with convolutional layers with shared weights and fully connected layers.
Theorem 4.4 (Necessary and Sufficient Condition for Zero Training Error) Let
Assumption
3.1 hold for the training sample and suppose that the CNN architecture satisfies Assumption 4.1
for some hidden layer 1 ≤ k ≤ L −1. Let Φ :
P → R be defined as in (7). Given any point
L
L
(Wl , bl )l=1 ∈ Sk . Then it holds that Φ (Wl , bl )l=1 = 0 if and only if ∇Uk+1 Φ
= 0.
L
(Wl ,bl )l=1
Proof: If Φ (Wl , bl )L
l=1 = 0 then it follows from the upper bound of Lemma 4.3 that ∇Uk+1 Φ =
0. Now, we suppose that ∇Uk+1 Φ = 0. Since (Wl , bl )L
l=1 ∈ Sk it holds rank(Fk ) = N and Ul
has full rank for every l ∈ [k + 2, L]. Thus it holds σmin (Fk ) > 0 and σmin (Ul ) > 0 for every
l ∈ [k + 2, L]. Moreover, (σk+1 , . . . , σL−1 ) have non-zero derivative by Assumption 4.1 and thus
kσl0 (Gl )kmin > 0 for every l ∈ [k +
1, L − 1]. This
combined with the lower bound in Lemma 4.3
leads to kFL − Y kF = 0 and thus Φ Wl , bl )L
l=1 = 0.
Lemma 4.2 shows that the set of points which are not covered by Theorem 4.4 has just measure zero
under a mild condition. The necessary and sufficient condition of Theorem 4.4 is rather intuitive as
it requires the gradient of the training objective to vanish w.r.t. the full weight matrix of layer k + 1
regardless of the architecture of this layer. It turns out that if layer k + 1 is fully connected, then this
condition is always satisfied at a critical point, in which case we obtain that every critical point in Sk
is a global minimum with exact zero training error. This is shown in the next Theorem 4.5, where
we consider a classification task with m classes. Z ∈ Rm×m is the full rank class encoding matrix
e.g. the identity matrix and (X, Y ) the training sample such that Yi: = Zj: whenever xi belongs to
class j for every i ∈ [N ], j ∈ [m].
Theorem 4.5 (Loss Surface of CNNs) Let (X, Y, Z) be a training set for which Assumption 3.1
holds, the CNN architecture satisfies Assumption 4.1 for some hidden layer 1 ≤ k ≤ L − 1, and
layer k + 1 is fully connected. Let Φ : P → R be defined as in (7). Then
L
• Every critical point (Wl , bl )L
∈
S
is
a
global
minimum
with
Φ
(W
,
b
)
k
l
l
l=1
l=1 = 0
L
• There exist infinitely many global minima (Wl , bl )L
l=1 ∈ Sk with Φ (Wl , bl )l=1 = 0
Theorem 4.5 indicates that the loss surface for this type of CNNs has a rather simple structure in
the sense that almost every critical point is a global minimum with zero training error. It remains
an interesting open problem if this result can be transferred to the case where layer k + 1 is also
convolutional. In any case whether layer k + 1 is fully connected or not, one might still assume that
a solution with zero training error still exists. However, note that Theorem 4.4 shows that at those
points where the loss is zero, the gradient of Φ w.r.t. Uk+1 must be zero as well. An interesting
special case of Theorem 4.5 is when the network is fully connected, in which case all the results of
Theorem 4.5 hold without any modifications.
Corollary 4.6 (Loss Surface of Fully Connected Nets) Let (X, Y, Z) be a training set with nonidentical training samples, i.e. xi 6= xj for every i 6= j and the fully connected network satisfies
Assumption 4.1 for some layer 1 ≤ k ≤ L − 1. Let Φ : P → R be defined as in (7). Then the
following holds
9
L
• Every critical point (Wl , bl )L
∈
S
is
a
global
minimum
with
Φ
(W
,
b
)
k
l
l
l=1
l=1 = 0
L
• There exist infinitely many global minima (Wl , bl )L
l=1 ∈ Sk with Φ (Wl , bl )l=1 = 0
Corollary 4.6 can be seen as a formal proof for the implicit assumption used in the recent work
(Nguyen & Hein, 2017) that there exists a global minimum with zero training error for the class of
fully connected, deep and wide networks.
5
C ONCLUSION
We have analyzed the expressiveness and loss surface of CNNs in realistic and practically relevant
settings. As state-of-the-art networks fulfill exactly or approximately the condition to have a sufficiently wide convolutional layer, we think that our results help to understand why current CNNs can
be trained so effectively. It would be interesting to discuss the loss surface for cross-entropy loss,
which currently does not fit into our analysis as the global minimum does not exist when the data is
linearly separable.
R EFERENCES
G. Alain and Y. Bengio. Understanding intermediate layers using linear classifier probes. In ICLR
Workshop, 2016.
S. An, F. Boussaid, and M. Bennamoun. How can deep rectifier networks achieve linear separability
and preserve distances? In ICML, 2015.
A. Andoni, R. Panigrahy, G. Valiant, and L. Zhang. Learning polynomials with neural networks. In
ICML, 2014.
P. Auer, M. Herbster, and M. K. Warmuth. Exponentially many local minima for single neurons. In
NIPS, 1996.
P. Baldi and K. Hornik. Neural networks and principle component analysis: Learning from examples
without local minima. Neural Networks, 2:53–58, 1988.
A. Blum and R. L Rivest. Training a 3-node neural network is np-complete. In NIPS, 1989.
A. Brutzkus and A. Globerson. Globally optimal gradient descent for a convnet with gaussian inputs,
2017. arXiv:1702.07966.
F. Chollet.
Xception:
arXiv:1610.02357.
Deep learning with depthwise separable convolutions, 2016.
A. Choromanska, M. Hena, M. Mathieu, G. B. Arous, and Y. LeCun. The loss surfaces of multilayer
networks. In AISTATS, 2015a.
A. Choromanska, Y. LeCun, and G. B. Arous. Open problem: The landscape of the loss surfaces of
multilayer networks. COLT, 2015b.
N. Cohen and A. Shashua. Convolutional rectifier networks as generalized tensor decompositions.
In ICML, 2016.
G. Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of Control,
Signals, and Systems, 2:303–314, 1989.
W. M. Czarnecki, G. Swirszcz, M. Jaderberg, S. Osindero, O. Vinyals, and K. Kavukcuoglu. Understanding synthetic gradients and decoupled neural interfaces. In ICML, 2017.
Y. Dauphin, R. Pascanu, C. Gulcehre, K. Cho, S. Ganguli, and Y. Bengio. Identifying and attacking
the saddle point problem in high-dimensional non-convex optimization. In NIPS, 2014.
O. Delalleau and Y. Bengio. Shallow vs. deep sum-product networks. In NIPS, 2011.
10
S. S. Du, J. D. Lee, and Y. Tian.
arXiv:1709.06129.
When is a convolutional filter easy to learn?, 2017.
R. Eldan and O. Shamir. The power of depth for feedforward neural networks. In COLT, 2016.
C. D. Freeman and J. Bruna. Topology and geometry of half-rectified network optimization. In
ICLR, 2017.
A. Gautier, Q. Nguyen, and M. Hein. Globally optimal training of generalized polynomial neural
networks with nonlinear spectral methods. In NIPS, 2016.
S. Goel and A. Klivans.
arXiv:1709.06010.
Learning depth-three neural networks in polynomial time, 2017.
I. J. Goodfellow, O. Vinyals, and A. M. Saxe. Qualitatively characterizing neural network optimization problems. In ICLR, 2015.
B. D. Haeffele and R. Vidal. Global optimality in tensor factorization, deep learning, and beyond,
2015. arXiv:1506.07540v1.
M. Hardt and T. Ma. Identity matters in deep learning. In ICLR, 2017.
K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, 2016.
K. Hornik, M. Stinchcombe, and H. White. Multilayer feedforward networks are universal approximators. Neural Networks, 2:359–366, 1989.
F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally, and K. Keutzer.
Squeezenet: Alexnet-level accuracy with 50x fewer parameters and < 0.5mb model size, 2016.
arXiv:1602.07360.
M. Janzamin, H. Sedghi, and A. Anandkumar. Beating the perils of non-convexity: Guaranteed
training of neural networks using tensor methods. arXiv:1506.08473, 2016.
K. Jarrett, K. Kavukcuoglu, and Y. LeCun. What is the best multi-stage architecture for object
recognition? In CVPR, 2009.
K. Kawaguchi. Deep learning without poor local minima. In NIPS, 2016.
S. G. Krantz and H. R. Parks. A Primer of Real Analytic Functions. Birkhäuser, Boston, second
edition, 2002.
A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional
neural networks. In NIPS, 2012.
Y. LeCun, B. Boser, J.S. Denker, D. Henderson, R.E. Howard, W. Hubbard, and L.D. Jackel. Handwritten digit recognition with a back-propagation network. In NIPS, 1990.
Y. Li and Y. Yuan. Convergence analysis of two-layer neural networks with relu activation, 2017.
arXiv:1705.09886.
S. Liang and R. Srikant. Why deep neural networks for function approximation? In ICLR, 2017.
R. Livni, S. Shalev-Shwartz, and O. Shamir. On the computational efficiency of training neural
networks. In NIPS, 2014.
A. Mahendran and A. Vedaldi. Understanding deep image representations by inverting them. In
CVPR, 2015.
H. Mhaskar and T. Poggio. Deep vs. shallow networks : An approximation theory perspective, 2016.
arXiv:1608.03287.
B. Mityagin. The zero set of a real analytic function, 2015. arXiv:1512.07276.
G. Montufar, R. Pascanu, K. Cho, and Y. Bengio. On the number of linear regions of deep neural
networks. In NIPS, 2014.
11
Q. Nguyen and M. Hein. The loss surface of deep and wide neural networks. In ICML, 2017.
V. D. Nguyen. Complex powers of analytic functions and meromorphic renormalization in qft, 2015.
arXiv:1503.00995.
R. Pascanu, G. Montufar, and Y. Bengio. On the number of response regions of deep feedforward
networks with piecewise linear activations. In ICLR, 2014.
A. Paszke, A. Chaurasia, S. Kim, and E. Culurciello. Enet: A deep neural network architecture for
real-time semantic segmentation, 2016. arXiv:1606.02147.
T. Poggio, H. Mhaskar, L. Rosasco, B. Miranda, and Q. Liao. Why and when can deep – but not
shallow – networks avoid the curse of dimensionality: a review, 2016. arXiv:1611.00740.
M. Raghu, B. Poole, J. Kleinberg, S. Ganguli, and J. Sohl-Dickstein. On the expressive power of
deep neural networks. In ICML, 2017.
I. Safran and O. Shamir. On the quality of the initial basin in overspecified networks. In ICML,
2016.
I. Safran and O. Shamir. Depth-width tradeoffs in approximating natural functions with neural
networks. In ICML, 2017.
A. Saxe, P. W. Koh, Z. Chen, M. Bhand, B. Suresh, and A. Y. Ng. On random weights and unsupervised feature learning. In ICML, 2011.
H. Sedghi and A. Anandkumar. Provable methods for training neural networks with sparse connectivity. In ICLR Workshop, 2015.
S. Shalev-Shwartz, O. Shamir, and S. Shammah. Failures of gradient-based deep learning. In ICML,
2017.
O. Shamir. Distribution-specific hardness of learning neural networks, 2017. arXiv:1609.01037.
J. Sima. Training a single sigmoidal neuron is hard. Neural Computation, 14:2709–2728, 2002.
K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015.
M. Soltanolkotabi. Learning relus via gradient descent, 2017. arXiv:1705.04591.
D. Soudry and E. Hoffer. Exponentially vanishing sub-optimal local minima in multilayer neural
networks, 2017. arXiv:1702.05777.
C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In CVPR, 2015a.
C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. Rethinking the inception architecture
for computer vision, 2015b. arXiv:1512.00567.
C. Szegedy, S. Ioffe, V. Vanhoucke, and A. Alemi. Inception-v4, inception-resnet and the impact of
residual connections on learning, 2016. arXiv:1602.07261.
M. Telgarsky. Representation benefits of deep feedforward networks, 2015. arXiv:1509.08101v2.
M. Telgarsky. Benefits of depth in neural networks. In COLT, 2016.
Y. Tian. An analytical formula of population gradient for two-layered relu network and its applications in convergence and critical point analysis. In ICML, 2017.
D. Yarotsky. Error bounds for approximations with deep relu networks, 2016. arXiv:1610.01145.
J. Yosinski, J. Clune, Y. Bengio, and H. Lipson. How transferable are features in deep neural
networks? In NIPS, 2014.
12
J. Yosinski, J. Clune, A. Nguyen, T. Fuchs, and H. Lipson. Understanding neural networks through
deep visualization. In ICML, 2015.
C. Yun, S. Sra, and A. Jadbabaie. Global optimality conditions for deep neural networks, 2017.
arXiv:1707.02444.
M. D. Zeiler and R. Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014.
C. Zhang, S. Bengio, M. Hardt, B. Recht, and Oriol Vinyals. Understanding deep learning requires
re-thinking generalization. In ICLR, 2017.
K. Zhong, Z. Song, P. Jain, P. Bartlett, and I. Dhillon. Recovery guarantees for one-hidden-layer
neural networks. In ICML, 2017.
A
A.1
M ISSING P ROOFS OF S ECTION 3
P ROOF OF L EMMA 3.3
• ReLU: It holds for every t < 0 that σ(t) = max(0, t) = 0 < et , and for t ≥ 0 that
σ(t) = t < t + 1. Thus ReLU satisfies the second condition of Assumption 3.2.
• Sigmoid: It holds that
lim
t→−∞
1
= 0,
1 + e−t
lim
t→∞
1
= 1.
1 + e−t
Thus σ satisfies the first condition of Assumption 3.2.
• Softplus: Since 1 + eαt ≤ 2eαt for every t ≥ 0, it holds for every t ≥ 0 that
0 ≤ σα (t) =
1
log(2)
1
log(1 + eαt ) ≤ log(2eαt ) =
+ t.
α
α
α
Moreover, since log(1 + t) ≤ t for t > 0, it holds log(1 + eαt ) ≤ eαt for every t ∈ R. Thus
αt
it holds that 0 ≤ σα (t) ≤ eα for every t < 0. Therefore σα satisfies the second condition
of Assumption 3.2 for ρ1 = 1/α, ρ2 = α, ρ3 = 1, ρ4 = log(2)/α.
A.2
P ROOF OF T HEOREM 3.4
To prove Theorem 3.4, we first show that Assumption 3.1 can be transported from the input to the
output layer.
Lemma A.1 Let Assumption 3.1 hold for the training sample. Consider a standard deep CNN
architecture which satisfies the following
1. The first layer is either convolutional or fully connected while all the other layers can be
convolutional, fully connected or max-pooling
2. (σ1 , . . . , σL ) are continuous and non-constant activation functions
Then for every layer 1 ≤ k ≤ L, there exist a set of parameters of the first k layers (Wl , bl )kl=1 such
that it holds fkp (xi ) 6= fkq (xj ) for every p, q ∈ [Pk ], i, j ∈ [N ], i 6= j. Moreover, (Wl , bl )kl=1 can
be chosen in such a way that, except for max-pooling layers, all the weight matrices Ul = Ml (Wl )
have full rank for every 1 ≤ l ≤ k.
Proof: Since Assumption 3.1 holds for the training inputs, one can first pick (W1 , b1 ) such that
Assumption 3.1 holds at the first layer. Then given that Assumption 3.1 is satisfied at the first layer,
the selection of the parameters of the higher layers can be done similarly.
Since our definition of a convolutional layer includes fully connected layer as a special case, it
is sufficient to prove the result for the general convolutional structure. Since the first layer is a
convolutional layer by our assumption, we denote by Q = [a1 , . . . , aT1 ] ∈ Rl0 ×T1 a matrix that
13
contains the set of convolutional filters of the first layer. Note here that there are T1 filters, namely
a1 , . . . , aT1 , where each filter at ∈ Rl0 for every t ∈ [T1 ]. Let us define the set
D 0
E
o
[ n
at , xpi − at , xqj = 0 .
Q ∈ Rl0 ×T1
S := Q ∈ Rl0 ×T1 M1 (Q) has low rank ∪
i6=j
p,q∈[P0 ]
t,t0 ∈[T1 ]
Basically, S is the set of “true” parameter matrices of the first layer where the corresponding weight
matrix M1 (Q) has low rank or there exists two patches of two different training samples that have
the same inner product with some corresponding two filters. By Assumption 3.1 it holds that xpi 6=
xqj for every p, q ∈ [P0 ], i 6= j, and thus the right hand side in the above formula of S is just the
union of a finite number of hyperplanes which has Lebesgue measure zero. For the left hand side,
it follows from Lemma 2.5 that the set of Q for which M1 (Q) does not have full rank has measure
zero. Thus the left hand side of S is a set of measure zero. Since S is the union of two measure zero
sets, it has also measure zero, and thus the complementary set Rl0 ×T1 \ S must be non-empty and
we choose W1 ∈ Rl0 ×T1 \ S.
Since σ1 is a continuous and non-constant function, there exists an interval (µ1 , µ2 ) such that σ1 is
bijective on (µ1 , µ2 ). We select and fix some matrix Q = [a1 , . . . , aT1 ] ∈ Rl0 ×T1 \ S and select
some β ∈ (µ1 , µ2 ). Let α > 0 be a free variable and W1 = [w11 , . . . , w1T1 ] where w1t denotes the
t-th filter of the first layer. Let us pick
w1t = αQ:t = αat ,
(b1 )h = β,
∀t ∈ [T1 ], h ∈ [n1 ].
It follows that W1 = αQ and thus M1 (W1 ) = M1 (αQ) = αM1 (Q) as M1 is a linear map by our
definition. Since Q ∈
/ S by construction, it holds that M1 (W1 ) has full rank for every α 6= 0. By
Definition 2.1, it holds for every i ∈ [N ], p ∈ [P0 ], t ∈ [T1 ], h = (p − 1)T1 + t that
f1 (xi )h = σ1 ( w1t , xpi + (b1 )h ) = σ1 (α at , xpi + β).
Since β ∈ (µ1 , µ2 ), one can choose a sufficiently small positive α such that it holds α ha, xpi i + β ∈
(µ1 , µ2 ) for every i ∈ [N ], p ∈ [P0 ], t ∈ [T1 ]. Under this construction, we will show that every entry
of f1 (xi ) must be different from every entry of f1 (xj ) for i 6= j. Indeed, let us compare f1 (xi )h
and f1 (xj )v for some h = (p − 1)T1 + t, v = (q − 1)T1 + t0 and i 6= j. It holds for sufficient small
α > 0 that
D 0
E
f1 (xi )h − f1 (xj )v = σ1 α at , xpi + β − σ1 α at , xqj + β 6= 0
(8)
D 0
E
where the last inequality follows from three facts. First, it holds hat , xpi i =
6
at , xqj since Q ∈
/ S.
Second, for the chosen α the values of the arguments of the activation function σ1 lie within (µ1 , µ2 ).
Third, since σ1 is bijective on (µ1 , µ2 ), it maps different inputs to different outputs.
Now, since the entries of f1 (xi ) and that of f1 (xj ) are already pairwise different from each other,
their corresponding patches must be also different from each other no matter how the patches are
organized in the architecture, that is,
f1p (xi ) 6= f1q (xj ) ∀p, q ∈ [P1 ], i, j ∈ [N ], i 6= j.
Now, if the network has only one layer, i.e. L = 1, then we are done. Otherwise, we will prove via
induction that this property can be translated to any higher layer. In particular, suppose that one has
already constructed (Wl , bl )kl=1 so that it holds
fk (xi )h − fk (xj )v 6= 0 ∀h, v ∈ [nk ], i, j ∈ [N ], i 6= j.
(9)
This is true for k = 1 due to (8). We will show below that (9) can also hold at layer k + 1.
1. Case 1: Layer k + 1 is convolutional or fully connected.
Since (9) holds for k by our induction assumption, it must hold that
fkp (xi ) 6= fkq (xj ) ∀p, q ∈ [Pk ], i, j ∈ [N ], i 6= j.
which means Assumption 3.1 also holds for the set of features at layer k. Thus one can
follows the similar construction as done for layer 1 above by considering the output of
layer k as input to layer k + 1. Then one obtains that there exist (Wk+1 , bk+1 ) where
Uk+1 = Mk+1 (Wk+1 ) has full rank so that the similar inequality (8) now holds for layer
k + 1, which thus implies (9) holds for k + 1.
14
2. Case 2: layer k + 1 is max-pooling
By Definition 2.3, it holds nk+1 = Pk and one has for every p ∈ [Pk ]
fk+1 (x)p = max (fkp (x))1 , . . . , (fkp (x))lk .
Since (9) holds at layer k by our induction assumption, every entry of every patch of fk (xi )
must be different from every entry of every patch of fk (xj ) for every i 6= j, that is,
(fkp (xi ))r 6= (fkq (xj ))s for every r, s ∈ [lk ], p, q ∈ [Pk ], i 6= j. Therefore, their maximum elements cannot be the same, that is,
fk+1 (xi )p 6= fk+1 (xj )q
∀p, q ∈ [nk+1 ], i, j ∈ [N ], i 6= j.
which proves (9) for layer k + 1.
So far, we have proved that (9) holds for every 1 ≤ k ≤ L. Thus it follows that for every
layer k, there exists a set of parameters of the first k layers for which the patches at layer k of
different training samples are pairwise different from each other, that is, fkp (xi ) 6= fkq (xj ) for every
p, q ∈ [Pk ], i 6= j. Moreover, except for max-pooling layers, all the weight matrices up to layer k
have been chosen to have full rank.
Proof of Theorem 3.4 Let A = Fk = [fk (x1 )T , . . . , fk (xN )T ] ∈ RN ×nk . Since our definition
of a convolutional layer includes fully connected layer as a special case, it is sufficient to prove the
result for convolutional structure. By Theorem’s assumption, layer k is convolutional and thus it
holds by Definition 2.1 that
p
(xi ) + (bk )j
Aij = fk (xi )j = σ wkt , fk−1
for every i ∈ [N ], t ∈ [Tk ], p ∈ [Pk−1 ] and j = (p, t) := (p − 1)Tk + t ∈ [nk ].
In the following, we show that there exists a set of parameters of the network such that rank(A) = N
and all the weight matrices Ul = Ml (Wl ) have full rank.
First, one observes that the subnetwork consisting of all the layers from the input layer till layer
k − 1 satisfies the conditions of Lemma A.1. Thus by applying Lemma A.1 to this subnetwork,
one obtains that there exist (Wl , bl )k−1
which all the matrices (Ul )k−1
l=1 for
l=1 , except for max-pooling
p
q
layers, have full rank and it holds that fk−1
(xi ) 6= fk−1
(xj ) for every p, q ∈ [Pk−1 ], i 6= j. The
main idea now is to fix the parameters of these layers and pick (Wk , bk ) such that Uk = Mk (Wk )
has full rank and it holds rank(A) = N . Let us define the set
[ [
p
p
S=
a ∈ Rlk−1
a, fk−1
(xi ) − fk−1
(xj ) = 0 .
i6=j p∈[Pk−1 ]
p
p
From the above construction, it holds that fk−1
(xi ) 6= fk−1
(xj ) for every p ∈ [Pk−1 ], i 6= j, and
thus S is the union of a finite number of hyperplanes which thus has measure zero. Let us denote by
Q = [a1 , . . . , aTk ] ∈ Rlk−1 ×Tk a parameter matrix that contains all the convolutional filters of layer
k in its columns. Pick at ∈ Rlk−1 \ S for every t ∈ [Tk ], so that it holds that Uk = Mk (Q) has
full rank. Note here that such matrix Q always exists. Indeed, Q is chosen from a positive measure
set as its columns (i.e. at ) are picked from a positive measure set. Moreover, the set of matrices Q
for which Mk (Q) has low rank has just measure zero due to Lemma 2.5. Thus there always exists
at least one matrix Q so that all of its columns do not belong to S and that Mk (Q) has full rank.
In the rest of the proof, the value of matrix Q is fixed. Let α ∈ R be a free parameter. Since σk
is a continuous and non-constant function, there exist a β ∈ R such that σk (β) 6= 0. Let the value
of β be fixed as well. We construct the convolutional filters Wk = [wk1 , . . . , wkTk ] and the biases
bk ∈ Rnk of layer k as follows. For every p ∈ [Pk−1 ], t ∈ [Tk ], j = (p, t), we define
wkt = −αat ,
p
(bk )j = α at , fk−1
(xj ) + β.
It follows that Wk = −αQ and thus Uk = Mk (Wk ) = −αMk (Q) as Mk is a linear map.
Moreover, since Mk (Q) has full rank by construction, it holds that Uk has full rank for every
15
α 6= 0. As α varies, we get a family of matrices A(α) ∈ RN ×nk where it holds for every i ∈
[N ], j = (p, t) ∈ [nk ] that
p
p
p
A(α)ij = σk wkt , fk−1
(xi ) + (bk )j = σk α at , fk−1
(xj ) − fk−1
(xi ) + β .
(10)
Note that each row of A(α) corresponds to one training sample and that permutations of the rows of
A(α) do not change the rank of A(α). We construct a permutation γ of {1, 2, . . . , N } as follows.
For every j = 1, 2, . .l. , Nm, let (p, t) be the
by j (the inverse transformation for
l tuple
m determined
j
j
given j ∈ [nk ] is p = Tk and t = j −
Tk − 1 Tk ) and define
γ(j) =
arg min
i∈{1,2,...,N } \ {γ(1),...,γ(j−1)}
p
at , fk−1
(xi ) .
Note that γ(j) is unique for every 1 ≤ j ≤ N since at ∈
/ S. It is clear that γ constructed as above
is a permutation of {1, 2, . . . , N } since at every step we take a different element from the index set
[N ]. From the definition of γ(j), it holds that for every j = (p, t) ∈ [N ], i ∈ [N ], i > j that
p
p
at , fk−1
(xγ(j) ) < at , fk−1
(xγ(i) ) .
We can relabel the training data according to the permutation so that w.l.o.g we can assume that γ
is the identity permutation, that is, γ(j) = j for every j ∈ [N ], in which case it holds for every
j = (p, t) ∈ [N ], i ∈ [N ], i > j that
p
p
at , fk−1
(xj ) < at , fk−1
(xi ) .
(11)
Under the above construction of (Wl , bl )kl=1 , we are ready to show that there exist α for which
rank(A(α)) = N. Since σk satisfies Assumption 3.2, we consider the following cases.
1. Case 1: There are finite constants µ+ , µ− ∈ R s.t. lim σk (t) = µ− and lim σk (t) = µ+
t→−∞
t→∞
and µ+ µ− = 0.
Let us consider the first case where µ− = 0. From (10) and (11) one obtains
j=i
σk (β)
lim A(α)ij = µ− = 0 i > j
α→+∞
ηij
i<j
(12)
where ηij is given for every i < j where j = (p, t) as
p
p
µ− , at , fk−1
(xj ) − fk−1
(xi ) < 0
ηij =
p
p
µ+ , at , fk−1 (xj ) − fk−1
(xi ) > 0
Note that ηij cannot be zero for i 6= j because at ∈
/ S. In the following, let us denote
A(α)1:N,1:N as a sub-matrix of A(α) that consists of the first N rows and columns. By the
Leibniz-formula one has
det(A(α)1:N,1:N ) = σk (β)N +
X
sign(π)
π∈SN \{γ}
N
Y
A(α)π(j)j
(13)
j=1
where SN is the set of all N ! permutations of the set {1, . . . , N } and γ is the identity
permutation. Now, one observes that for every permutation π 6= γ, there always exists at
least one component j where π(j) > j in which case it follows from (12) that
lim
α→∞
N
Y
A(α)π(j)j = 0.
j=1
Since there are only finitely many such terms in (13), one obtains
lim det(A(α)1:N,1:N ) = σk (β)N 6= 0
α→∞
where the last inequality follows from our choice of β. Since det(A(α)1:N,1:N ) is
a continuous function of α, there exists α0 ∈ R such that for every α ≥ α0 it
16
holds det(A(α)1:N,1:N ) 6= 0 and thus rank(A(α)1:N,1:N ) = N which further implies
rank(A(α)) = N. Thus the corresponding set of feature vectors {fk (x1 ), . . . , fk (xN )} are
linearly independent.
For the case where µ+ = 0, one can argue similarly. The only difference is that one
considers now the limit for α → −∞. In particular, (10) and (11) lead to
i=j
σk (β)
lim A(α)ij = µ+ = 0 i > j
α→−∞
ηij = 0 i < j.
For every permutation π 6= γ there always exists at least one component j where π(j) > j,
in which case it holds that
lim
α→−∞
N
Y
A(α)π(j)j = 0.
j=1
and thus it follows from the Leibniz formula that
lim det(A(α)1:N,1:N ) = σk (β)N 6= 0.
α→−∞
Since det(A(α)1:N,1:N ) is a continuous function of α, there exists α0 ∈ R such that for
every α ≤ α0 it holds det(A(α)1:N,1:N ) 6= 0 and thus rank(A(α)1:N,1:N ) = N which
further implies rank(A(α)) = N. Thus the set of feature vectors at layer k are linearly
independent.
2. Case 2: There are positive constants ρ1 , ρ2 , ρ3 , ρ4 s.t. |σk (t)| ≤ ρ1 eρ2 t for t < 0 and
|σk (t)| ≤ ρ3 t + ρ4 for t ≥ 0.
Our proof strategy is essentially similar to the previous case. Indeed, for every permutation
π 6= γ there always exist at least one component j = (p, t) ∈ [N ] where π(j) > j in which
p
p
case δj := at , fk−1
(xj ) − fk−1
(xπ(j) ) < 0 due to (11). For sufficiently large α > 0, it
holds that αδj + β < 0 and thus one obtains from (10) that
|A(α)π(j)j | = |σk (αδj + β)| ≤ ρ1 eρ2 β e−αρ2 |δj | .
If π(j) = j then |A(α)π(j)j | = |σk (β)| which is a constant. For π(j) < j, one notices
p
p
that δj := at , fk−1
(xj ) − fk−1
(xπ(j) ) can only be either positive or negative as at ∈
/ S.
In this case, if δj < 0 then it can be bounded by the similar exponential term as above for
sufficiently large α. Otherwise it holds αδj + β > 0 for sufficiently large α > 0 and we get
|A(α)π(j)j | = |σ(αδj + β)| ≤ ρ3 δj α + ρ3 β + ρ4 .
Overall, for sufficiently large α > 0, there must exist positive constants P, Q, R, S, T such
that it holds for every π ∈ SN \ {γ} that
N
Y
A(α)π(j)j ≤ R(P α + Q)S e−αT .
j=1
The upper bound goes to zero as α goes to ∞. This combined with the Leibniz formula
from (13), we get limα→∞ det(A(α)1:N,1:N ) = σk (β)N 6= 0. Since det(A(α)1:N,1:N )
is a continuous function of α, there exists α0 ∈ R such that for every α ≥ α0 it holds
det(A(α)1:N,1:N ) 6= 0 and thus rank(A(α)1:N,1:N ) = N which implies rank(A(α)) = N.
Thus the set of feature vectors at layer k are linearly independent.
Overall, we have shown that there always exist (Wl , bl )kl=1 such that the set of feature vectors
{fk (x1 ), . . . , fk (xN )} at layer k are linearly independent. Moreover, (Wl , bl )kl=1 have been chosen
so that all the weight matrices Ul = Ml (Wl ), except for max-pooling layers, have full rank for
every 1 ≤ l ≤ k.
A.3
P ROOF OF T HEOREM 3.5
To prove Theorem 3.5, the following key property of real analytic functions is required.
17
Lemma A.2 Nguyen (2015); Mityagin (2015) If f : Rn → R is a real analytic function which is
not identically zero then the set {x ∈ Rn | f (x) = 0} has Lebesgue measure zero.
Any linear function is real analytic and the set of real analytic functions is closed under addition,
multiplication and composition, see e.g. Prop. 2.2.2 and Prop. 2.2.8 in Krantz & Parks (2002).
As we assume that all the activation functions of the first k layers are real analytic, we get that
the function fk is a real analytic function of (Wl , bl )kl=1 as it is the composition of real analytic
functions. Now, we recall from our definition that Fk = [fk (x1 ), . . . , fk (xN )]T ∈ RN ×nk is the
output matrix at layer k for all training samples. One observes that the set of low rank matrices
Fk can be characterized by a system of equations such that all the nNk determinants of all N × N
sub-matrices of Fk are zero. As the determinant is a polynomial in the entries of the matrix and
thus an analytic function of the entries and composition of analytic functions are again analytic, we
conclude that each determinant is an analytic function of the network parameters of the first k layers.
By Theorem 3.4, there exists at least one set of parameters of the first k layers such that one of these
determinant functions is not identically zero and thus by Lemma A.2, the set of network parameters
where this determinant is zero has Lebesgue measure zero. But as all sub-matrices need to have low
rank in order that rank(Fk ) < N , it follows that the set of parameters where rank(Fk ) < N has
Lebesgue measure zero.
B
B.1
M ISSING P ROOFS OF S ECTION 4
P ROOF OF L EMMA 4.3
To prove Lemma 4.3, we first derive standard backpropagation in Lemma B.1. In the following
we use the Hadamard product ◦, which for A, B ∈ Rm×n is defined as A ◦ B ∈ Rm×n with
∂Φ
(A ◦ B)ij = Aij Bij . Let δkj (xi ) = ∂gkj
(xi ) be the derivative of Φ w.r.t. the value of unit j at layer
k evaluated at a single sample xi . We arrange these vectors for all training samples into a single
matrix ∆k = [δk: (x1 ), . . . , δk: (xN )]T ∈ RN ×nk . The following lemma is a slight modification of
Lemma 2.1 in (Nguyen & Hein, 2017) for which we provide the proof for completeness.
Lemma B.1 Given some hidden layer 1 ≤ k ≤ L − 1. Let (σk+1 , . . . , σL−1 ) be differentiable.
Then the following hold:
FL − Y,
l=L
1. ∆l =
T
(∆l+1 Ul+1
) ◦ σl0 (Gl ), k + 1 ≤ l ≤ L − 1
T
2. ∇Ul Φ = Fl−1
∆l , ∀ k + 1 ≤ l ≤ L
Proof:
1. By definition, it holds for every i ∈ [N ], j ∈ [nL ] that
∂Φ
(∆L )ij = δLj (xi ) =
= fLj (xi ) − yij
∂gLj (xi )
and hence, ∆L = FL − Y.
For every k + 1 ≤ l ≤ L − 1, the output of the network for a single training sample can
be written as the composition of differentiable functions (i.e. the outputs of all layers from
l + 1 till the output layer), and thus the chain rule yields for every i ∈ [N ], j ∈ [nl ] that
(∆l )ij = δlj (xi ) =
nl+1
X
∂g(l+1)s (xi ) ∂flj (xi )
∂Φ
∂Φ
=
∂glj (xi )
∂g
(x
)
∂flj (xi ) ∂glj (xi )
i
(l+1)s
s=1
nl+1
=
X
δ(l+1)s (xi )(Ul+1 )js σ 0 (glj (xi ))
s=1
nl+1
=
X
s=1
and hence ∆l =
T
(∆l+1 Ul+1
)
0
◦ σ (Gl ).
18
T
(∆(l+1) )is (Ul+1 )sj σ 0 ((Gl )ij )
2. For every k + 1 ≤ l ≤ L − 1, r ∈ [nl−1 ], s ∈ [nl ], one has
N
N
N
X ∂Φ ∂gls (xi ) X
X
∂Φ
T
=
=
δls (xi )f(l−1)r (xi ) =
(Fl−1
)ri (∆l )is
∂(Ul )rs
∂g
(x
)
∂(U
)
ls
i
l
rs
i=1
i=1
i=1
T
= Fl−1
∆l rs
T
and hence ∇Ul Φ = Fl−1
∆l .
The following straightforward inequalities are also helpful to prove Lemma 4.3. Let λmin (·) and
λmax (·) denotes the smallest and largest eigenvalue of a matrix.
Lemma B.2 Let A ∈ Rm×n with m ≥ n. Then it holds σmax (A) kxk2 ≥ kAxk2 ≥ σmin (A) kxk2
for every x ∈ Rn .
q
p
T T
kAxk
Proof: Since m ≥ n, it holds that σmin (A) = λmin (AT A) = min x xAT xAx = min kxk 2
2
p
kAxk
and thus σmin (A) ≤ kxk 2 for every x ∈ Rn . Similarly, it holds σmax (A) = λmax (AT A) =
2
q
T T
kAxk
kAxk
max x xAT xAx = max kxk 2 and thus σmax (A) ≥ kxk 2 for every x ∈ Rn .
2
2
Lemma B.3 Let A ∈ Rm×n , B ∈ Rn×p with m ≥ n. Then it holds σmax (A) kBkF ≥ kABkF ≥
σmin (A) kBkF .
Proof: Since m ≥ n, it holds that λmin (AT A) = σmin (A)2 and λmax (AT A) = σmax (A)2 . Thus
2
2
we have kABkF = tr(B T AT AB) ≥ λmin (AT A) tr(B T B) = σmin (A)2 kBkF . Similarly, it holds
2
2
kABkF = tr(B T AT AB) ≤ λmax (AT A) tr(B T B) = σmax (A)2 kBkF .
Proof of Lemma 4.3 We first prove the lower bound. Let Im denotes an m-by-m identity matrix
and ⊗ the Kronecker product. From Lemma B.1 it holds ∇Uk+1 Φ = FLT ∆k+1 and thus
vec(∇Uk+1 Φ) = (Ink+1 ⊗ FkT ) vec(∆k+1 ).
It follows that
∇Uk+1 Φ F = (Ink+1 ⊗ FkT ) vec(∆k+1 )
2
≥ σmin (Fk ) kvec(∆k+1 )k2 = σmin (Fk ) k∆k+1 kF
(14)
where the inequality follows from Lemma B.2 for the matrix (Ink+1 ⊗ FkT ) ∈ Rnk nk+1 ×N nk+1 with
nk ≥ N by Assumption 4.1. Using Lemma B.1 again, one has
T
0
k∆k+1 kF = (∆k+2 Uk+2
) ◦ σk+1
(Gk+1 )
0
≥ σk+1
(Gk+1 )
F
min
T
∆k+2 Uk+2
0
= σk+1
(Gk+1 )
min
Uk+2 ∆Tk+2
0
≥ σk+1
(Gk+1 )
σ (Uk+2 ) k∆k+2 kF
min min
F
F
where the last inequality follows from Lemma B.3 for the matrices Uk+2 ∈ Rnk+1 ×nk+2 and ∆Tk+2
with nk+1 ≥ nk+2 by Assumption 4.1. By repeating this for k∆k+2 kF , . . . , k∆L−1 kF , one gets
k∆k+1 kF ≥
L−1
Y
L−1
Y
kσl0 (Gl )kmin σmin (Ul+1 ) k∆L kF =
kσl0 (Gl )kmin σmin (Ul+1 ) kFL − Y kF
l=k+1
l=k+1
(15)
From (14), (15), one obtains
∇Uk+1 Φ
F
L−1
Y
≥ σmin (Fk )
kσl0 (Gl )kmin σmin (Ul+1 ) kFL − Y kF
l=k+1
19
which proves the lower bound.
The proof for upper bound is similar. Indeed one has
∇Uk+1 Φ
F
= (Ink+1 ⊗ FkT ) vec(∆k+1 )
2
≤ σmax (Fk ) kvec(∆k+1 )k2 = σmax (Fk ) k∆k+1 kF
(16)
where the inequality follows from Lemma B.2 Now, one has
T
0
k∆k+1 kF = (∆k+2 Uk+2
) ◦ σk+1
(Gk+1 )
≤
=
≤
0
σk+1
(Gk+1 ) max
0
σk+1 (Gk+1 ) max
0
σk+1
(Gk+1 ) max
F
T
∆k+2 Uk+2 F
Uk+2 ∆Tk+2 F
σmax (Uk+2 ) k∆k+2 kF
where the last inequality follows from Lemma B.3. By repeating this chain of inequalities for
k∆k+2 kF , . . . , k∆L−1 kF , one obtains:
k∆k+1 kF ≤
L−1
Y
L−1
Y
kσl0 (Gl )kmax σmax (Ul+1 ) k∆L kF =
kσl0 (Gl )kmax σmax (Ul+1 ) kFL − Y kF .
l=k+1
l=k+1
(17)
From (16), (17), one obtains that
∇Uk+1 Φ
F
L−1
Y
kσl0 (Gl )kmax σmax (Ul+1 ) kFL − Y kF
≤ σmax (Fk )
l=k+1
which proves the upper bound.
B.2
P ROOF OF T HEOREM 4.5
1. Since layer k + 1 is fully connected, it holds at every critical point in Sk that ∇Wk+1 Φ =
0 = ∇Uk+1 Φ. This combined with Theorem 4.4 yields the result.
L
2. One
basically needs to show that there exist (Wl , bl )l=1 such that it holds:
L
Φ (Wl , bl )l=1 = 0, rank(Fk ) = N and Ul = Ml (Wl ) has full rank ∀l ∈ [k + 2, L]
Note that the last two conditions are fulfilled by the fact that (Wl , bl )L
l=1 ∈ Sk .
By Assumption 4.1, the subnetwork consisting of the first k layers satisfies the condition of
Theorem 3.4. Thus by applying Theorem 3.4 to this subnetwork, one obtains that there exist
(Wl , bl )kl=1 such that rank(Fk ) = N . In the following, we fix these layers and show how
to pick (Wl , bl )L
l=k+1 such that FL = Y. The main idea now is to make the output of all the
training samples of the same class become identical at layer k + 1 and thus they will have
the same network output. Since there are only m classes, there would be only m distinct
outputs for all the training samples at layer L − 1. Thus if one can make these m distinct
outputs become linearly independent at layer L−1 then there always exists a weight matrix
WL that realizes the target output Y as the last layer is just a linear map by assumption.
Moreover, we will show that, except for max-pooling layers, all the parameters of other
layers in the network can be chosen in such a way that all the weight matrices (Ul )L
l=k+2
achieve full rank. Our proof details are as follows.
Case 1: k = L − 1
T
T
It holds rank(FL−1 ) = N. Pick bL = 0 and WL = FL−1
(FL−1 FL−1
)−1 Y .
Since the output layer is fully connected, it follows from Definition 2.2 that
FL = FL−1 WL + 1N bTL = Y. Since rank(Fk ) = N and the full rank condition on
L
(Wl )L
/ Sk which finishes the
l=k+2 is not active when k = L − 1, it holds that (Wl , bl )l=1 ∈
proof.
Case 2: k = L − 2
It holds rank(FL−2 ) = N. Let A ∈ Rm×nL−1 be a full row rank matrix such that Aij ∈
range(σL−1 ). Note that nL−1 ≥ nL = m due to Assumption 4.1. Let D ∈ RN ×nL−1
20
be a matrix satisfying Di: = Aj: whenever xi belongs to class j for every i ∈ [N ], j ∈
[m]. By construction, FL−2 has full row rank, thus we can pick bL−1 = 0, WL−1 =
−1
T
T
FL−2
(FL−2 FL−2
)−1 σL−1
(D). Since layer L − 1 is fully connected by assumption, it
follows from Definition 2.2 that FL−1 = σL−1 (FL−2 WL−1 + 1N bTL−1 ) = D and thus
(FL−1 )i: = Di: = Aj: whenever xi belongs to class j.
So far, our construction of the first L − 1 layers has led to the fact that all the training
samples belonging to the same class will have identical output at layer L − 1. Since A has
full row rank by construction, we can pick for the last layer bL = 0, WL = AT (AAT )−1 Z
where Z ∈ Rm×m is our class embedding matrix with rank(Z) = m. One can easily check
that AWL = Z and that FL = FL−1 WL + 1N bTL = FL−1 WL where the later follows from
Definition 2.2 as the output layer is fully connected. Now one can verify that FL = Y.
Indeed, whenever xi belongs to class j one has
(FL )i: = (FL−1 )Ti: WL = (A)Tj: WL = Zj: = Yi: .
−1
Moreover, since rank(Fk ) = N and rank(WL ) = rank(AT (AAT )
Z) = m, it holds that
L
(Wl , bl )L
/ Sk . Therefore, there exists (Wl , bl )L
l=1 ∈
l=1 ∈ Sk with Φ (Wl , bl )l=1 = 0.
Case 3: k ≤ L − 3
It holds rank(Fk ) = N. Let E ∈ Rm×nk+1 be any matrix such that Eij ∈ range(σk+1 )
and Eip 6= Ejq for every 1 ≤ i 6= j ≤ N, 1 ≤ p, q ≤ nk+1 . Let D ∈ RN ×nk+1 satisfies
−1
Di: = Ej: for every xi from class j. Pick bk+1 = 0, Wk+1 = FkT (Fk FkT )−1 σk+1
(D).
Note that the matrix is invertible as Fk has been chosen to have full row rank. Since layer
k + 1 is fully connected by our assumption, it follows from Definition 2.2 that Fk+1 =
σk+1 (Fk Wk+1 + 1N bTk+1 ) = σk+1 (Fk Wk+1 ) = D and thus it holds
(Fk+1 )i: = Di: = Ej:
(18)
for every xi from class j.
So far, our construction has led to the fact that all training samples belonging to the same
class have identical output at layer k + 1. The idea now is to see E as a new training
data matrix of a subnetwork consisting of all layers from k + 1 till the output layer L. In
particular, layer k + 1 can be seen as the input layer of this subnetwork and similarly, layer
L can be seen as the output layer. Moreover, every row of E ∈ Rm×nk+1 can be seen as
a new training sample to this subnetwork. One can see that this subnet together with the
training data matrix E satisfy the conditions of Theorem 3.4 at the last hidden layer L − 1.
In particular, it holds that
• The rows of E ∈ Rm×nk+1 are componentwise different from each other, and thus the
input patches must be also different from each other, and thus E satisfies Assumption
3.1
• Every layer from k +1 till L−1 is convolutional or fully connected due to Assumption
4.1
• The width of layer L − 1 is larger than the number of samples due to Assumption 4.1,
that is, nL−1 ≥ nL = m
• (σk+2 , . . . , σL−1 ) satisfy Assumption 3.2 due to Assumption 4.1
By applying Theorem 3.4 to this subnetwork and training data E, we obtain that there must
L−1
exist (Wl , bl )L−1
l=k+2 for which all the weight matrices (Ul )l=k+2 have full rank such that
the set of corresponding m outputs at layer L − 1 are linearly independent. In particular,
let A ∈ Rm×nL−1 be the corresponding outputs of E through this subnetwork then it holds
that rank(A) = m. Intuitively, if one feeds Ej: as an input at layer k + 1 then one would
get Aj: as an output at layer L − 1. This combined with (18) leads to the fact that if one
now feeds (Fk+1 )i: = Ej: as an input at layer k + 1 then one would get at layer L − 1 the
output (FL−1 )i: = Aj: whenever xi belongs to class j.
Last, we pick bL = 0, WL = AT (AAT )−1 Z. It follows that AWL = Z. Since the output
layer L is fully connected, it holds from Definition 2.2 that FL = FL−1 WL + 1N bTL =
FL−1 WL .
21
One can verify now that FL = Y. Indeed, for every sample xi from class j it holds that
(FL )i: = (FL−1 )Ti: WL = ATj: WL = Zj: = Yi: .
Overall, we have shown that Φ = 0. In addition, it holds rank(Fk ) = N from the conL−1
struction of the first k layers. All the matrices (Ul )l=k+2
have full rank by the construction of the subnetwork from k + 1 till L. Moreover, UL = WL also has full rank since
rank(WL ) = rank(AT (AAT )−1 Z) = m. Therefore it holds (Wl , bl )L
l=1 ∈ Sk .
B.3
P ROOF OF C OROLLARY 4.6
It is clear that the network structure satisfies the conditions of Theorem 4.4 as every layer is fully
connected. Moreover, the input patches also satisfy Assumption 3.1 because every input patch is
simply one training sample in this case but since there are no identical training samples, one derives
that the input patches must be different from each other.
Since all the conditions of Theorem 4.4 are met, the application of Theorem 4.4 yields the result.
22
| 7cs.IT
|
Deep Tracking on the Move: Learning to Track the World from a
Moving Vehicle using Recurrent Neural Networks
Ground Truth
Comparison
Occupancy
Prediction
Neural
Network
Abstract— This paper presents an end-to-end approach for
tracking static and dynamic objects for an autonomous vehicle driving through crowded urban environments. Unlike
traditional approaches to tracking, this method is learned
end-to-end, and is able to directly predict a full unoccluded
occupancy grid map from raw laser input data. Inspired by
the recently presented DeepTracking approach ([1], [2]), we
employ a recurrent neural network (RNN) to capture the
temporal evolution of the state of the environment, and propose
to use Spatial Transformer modules to exploit estimates of the
egomotion of the vehicle. Our results demonstrate the ability
to track a range of objects, including cars, buses, pedestrians,
and cyclists through occlusion, from both moving and stationary
platforms, using a single learned model. Experimental results
demonstrate that the model can also predict the future states
of objects from current inputs, with greater accuracy than
previous work.
h0
I. I NTRODUCTION
The safe and effective operation of an autonomous vehicle
depends on its ability to interpret its surroundings and track
and predict the state of the environment over time. Many
tracking systems employ multiple hand-engineered stages
(e.g. object detection, semantic classification, data association, state estimation and motion modelling, occupancy grid
generation) in order to represent the state and evolution of the
world ([3], [4], [5]). However, as the tasks assigned to robots
become more complex, this approach becomes increasingly
infeasible.
Recent advances in machine learning, particularly those
of deep neural networks, have demonstrated the ability to
capture complex structure in the real world, and have led
to significant improvements in numerous computer vision
and natural language processing applications ([6], [7], [8]).
Such approaches would however typically require large, taskspecific corpora of annotated ground-truth labels to master
the desired task. This becomes difficult when learning a
model of the environment without access to corresponding
ground truth, as is often the case for object tracking in
crowded urban environments.
In recent work, [2] took an alternative approach and
presented an end-to-end fully and efficiently trainable framework for learning a model of the world dynamics, building on
the original DeepTracking work by [1]. We considered the
specific problem of learning to track and classify moving
objects in a complex and only partially-observable realworld scenario, as viewed from a static sensor. Here, we
Authors are from the Mobile Robotics Group at the
University of Oxford, United Kingdom: julie, dushyant,
[email protected]
…
h5
Raw Sensor
Input
arXiv:1609.09365v3 [cs.CV] 19 Apr 2017
Julie Dequaire, Dushyant Rao, Peter Ondrúška, Dominic Zeng Wang and Ingmar Posner
t
visible space
occluded space
…
true positive
t+5
dynamic object
false positive
static object
false negative
field of view mask
Fig. 1. A typical training sequence. The unoccluded occupancy map is
predicted directly from the input grid data, allowing objects to be tracked in
occlusion and in future frames. The observed false positives are therefore
beneficial. Comparison to the visible ground truth shows that the model is
able to capture the dynamics of the moving vehicle (pink rectangle) and
accurately predicts its track.
advance this work and address the problem of tracking from a
moving platform. We extend the neural network architecture
proposed in [2] to account for the egomotion of the sensor
frame as it moves in the world frame, and demonstrate
improved tracking accuracy as compared to previous work.
We demonstrate the system on laser point cloud data
collected in a busy urban environment, with an array of static
and dynamic objects, including cars, buses, pedestrians and
cyclists. The model not only bypasses the sequence of handengineered steps typical of traditional tracking approaches,
but is empirically shown to successfully predict the future
evolution of objects in the environment, even when they are
completely occluded.
The rest of the paper is structured as follows. Section II
highlights related work and Section III summarises the problem definition and DeepTracking framework first presented
in ([1], [2]). Section IV describes the models used to perform
tracking in real-world scenarios considering both static and
dynamic sensors. Section V presents an empirical evaluation
of our methods, and Section VI concludes the paper and
discusses the future implications of our findings.
II. R ELATED W ORK
A number of previous works have explored the problem of
model-free tracking of objects in the environment of an autonomous vehicle ([3], [4], [5]). Typically, these approaches
follow the traditional paradigm of a multi-component
pipeline, with separate components to parametrise and detect
objects, associate new measurements to existing tracks, and
estimate the state of each individually tracked object. The
use of multiple stages in the framework is cumbersome and
introduces extra unnecessary failure modes for the tracking
algorithm.
Recent work proposes to replace these multiple stages with
an end-to-end learning framework known as DeepTracking,
by leveraging neural networks to directly learn the mapping
from raw laser input to an unoccluded occupancy grid ([1],
[2]), even with relatively small amounts of data. The approach utilises an RNN architecture using gated recurrent
units [9] to capture the state and evolution of the world
in a sequence of laser scan frames. Another work [10]
considers recurrent flow networks and takes a different angle
to predicting occupancy in dynamic environments. They
explicitly encode a range of velocities in the hidden layers of
a recurrent network, and use Bayesian optimization to learn
the network parameters which update velocity estimation
and occupancy prediction. However, the model does not
explicitly track objects through occlusion.
DeepTracking shares similarities with deep learning approaches to predictive video modelling ([11], [12]), in that
it is trained to predict the future state of the world based on
current input data. This is particularly important, because in
order to successfully capture the future location of dynamic
objects in the scene, the model must implicitly store the
position and velocity of each object in its internal memory
state.
While this eliminates the need to design individual components by hand, the model assumes a static vehicle [2].
Extending the problem to a moving platform is a challenging
task, as it introduces an array of complex relative motions
between the vehicle and objects in its environment. As
DeepTracking ignores the motion of the vehicle, the model is
forced to learn all possible motion interactions between the
vehicle and its environment as if the vehicle were stationary.
For a moving platform, we leverage estimates of egomotion
as a proxy for vehicle motion. We scale up the RNN-based
models proposed by ([1], [2]) for real-world application on
dynamic vehicles, and exploit Spatial Transformer modules
[13], which allow the internal memory state representations
to be spatially transformed according to the estimated egomotion.
The main contributions of this work are as follows:
1) A more in-depth analysis of the performance of DeepTracking in the case of a static vehicle, building on the
experiments presented in [2].
2) The use of Spatial Transformer modules to exploit
estimates of visual egomotion in the DeepTracking
framework.
3) Demonstration of end-to-end tracking of a variety of
object classes through occlusion, on a moving vehicle
in crowded urban scenes.
III. T RACKING P ROBLEM F ORMULATION
The problem we address in this paper is to uncover the
true state of the world, in terms of a 2D occupancy grid
yt , given a sequence of partially observed states of the
environment x1:t computed from raw sensor measurements.
In particular, we solve for P (yt |x1:t ), the probability of
the true unoccluded state of the world at time t given a
sequence of partial observations at all previous time steps.
This formulation can also be used to predict future states by
solving for P (yt+n |x1:t ), given empty input for xt+1:t+n .
P (yt |x1:t ) = P (yt |ht )
(1)
Evolution of this latent state ht , which includes propagating model dynamics and integrating new sensor measurements, is modelled by the update equation:
ht = f (ht−1 , xt )
(2)
The key element here is that both the latent state update
f (ht−1 , xt ), and the decoding step to produce the output
P (yt |ht ) are modelled as parts of a single neural network and
are trained jointly. Equations 1 and 2 can then be performed
repeatedly as the building block of a recurrent neural network
that continuously updates the belief state ht , and uses it as
network memory to predict yt . This makes it suitable for
online stream filtering of sensor input.
When the output ground-truth yt is not readily available,
as is the case in real-world scenarios, the network can be
trained in an self-supervised fashion. This is made possible
by considering that predicting the movement of objects in
occlusion at time t is similar to predicting a future state yt+n
provided no input is given to the network, i.e xt+n = .
Lack of input observation equates to complete occlusion of
the scene. As only observable ground truth is available, we
reduce the problem of predicting yt+n to that of predicting
the directly observable input xt+n . Training the network to
predict P (xt+n |x1:t ) is then equivalent to computing and
backpropagating the errors only on the observable parts of
the scene. We refer the reader to [1] and [2] for further details
on the RNN and the training procedure.
Each input observation xt ∈ {0, 1}2×M ×M is represented
as a pair of discretised 2D binary grids of size M × M ,
parallel to the ground and locally built around the sensor. The
first matrix encodes whether a cell is directly observable by
the sensor at time t, while the second matrix encodes whether
a cell is observed to be free (value of 0) or occupied (value
Odometry
1
16
Spatial Transformer
Module
dilation: 4
dilation: 2
GRU
16
GRU
16
GRU
dilation: 1
h
B. Deep Tracking from a Moving Vehicle
yt
h
t
2
t-1
xt
Fig. 2. The proposed network architecture features dilated convolutions,
gated recurrent units, a spatial transformer module, and outputs cell occupancy in the sensor’s surroundings. The spatial transformer module is only
utilised in a moving vehicle scenario.
of 1). We refer to these two matrices as xt,vis and xt,occ ,
the visibility and occupancy grids respectively. The output
we wish to obtain is an occlusion-free state of the world
yt ∈ {0, 1}M ×M , and is represented by an occupancy
matrix similar to that of the input occupancy grid of xt .
In the next section, we build upon [2] to deploy the
DeepTracking paradigm on a real-world moving platform.
IV. T ECHNICAL OVERVIEW
A. Deep Tracking from a Static Sensor
First, we revisit [2] to detail the baseline DeepTracking
architecture for real world application, which we extend to
a moving platform.
At each time step t, the partially observed grid xt used as
input to the network is computed from raw 2D laser scans by
ray tracing. Cells in which a measurement ends are marked
as occupied, all cells from the sensor origin up to the end
of the ray are marked as free, and cells beyond the ray are
marked to be unobserved.
The input xt is processed by a multi-layer network illustrated in Figure 2, with the Spatial Transformer module
only utilised in the moving vehicle scenario. The architecture
originally proposed in [1] is scaled up with dilated convolutions [14] and a variant of gated recurrent units ([9], [15]) to
allow for the simultaneous tracking of different-sized objects
such as cars and pedestrians. Each layer l at time t is updated
considering its own activations at time t−1 and those of layer
l −1 below at time t, thus implementing the recurrence. This
allows the network to extract and remember the information
from the past and to use it for prediction of occluded objects
or future states of the world. An additional static memory
can also be utilised, in which the network is able to learn
individual pixel-wise biases to add to the output of every
convolutional operation. The output of the final layer is then
converted into the output cell occupancy yt via a simple
convolutional decoder. As explicited in the results section,
we also experiment with architectures that decode the entire
hidden state to the output.
Tracking a dynamic scene from a moving platform poses
the challenge of decoupling the motion of the vehicle from
the motion of objects in the global environment. In the static
scenario, information related to an object located at index
{i, j} in the input xt would typically be stored at the corresponding neighbouring spatial location {i+∆i, j+∆j} in the
layers of the latent representation, with the neighbourhood
∆i,j based on the receptive field of each neuron in the
hidden state. The latent state update would then pass the
information along spatially in accordance with the observed
relative motion of the input object (Figure 4). In the static
scenario, the dynamics of the scene as viewed from the world
frame are coherent with that viewed from the local sensor
frame.
When tracking from a moving vehicle however, the spatial
update of information within the latent representation would
additionally need to account for the sensor’s egomotion as
it will affect the position of the object relative to the sensor
frame. Although this is a major drawback of the baseline
DeepTracking architecture, it can be compensated by transforming the memory state in accordance to the egomotion. In
other words, a static obstacle situated at position {it−1 , jt−1 }
in the sensor frame at t−1, will be moved to position {it , jt }
in the sensor frame at time t such that:
[xt , yt , 1]T = Tt,t−1 × [xt−1 , yt−1 , 1]T
(3)
where Tt,t−1 is the SE(2) forward transformation of the
sensor source frame at t−1 into the sensor destination frame
at t. This formulation naturally extends to 3D motion with
SE(3).
We aid the network in decoupling egomotion and object
motion in this way by introducing a Spatial Transformer [13]
module (STM) in the hidden state (Figure 2). In the original
work by [13], the STM is introduced as a learnable module
for actively spatially transforming feature maps to aid tasks
such as the classification of distorted datasets. However,
in the context of tracking from a moving platform where
egomotion is readily available (e.g. visual odometry), the
STM can be used just to forward transform the hidden feature
maps centred in the sensor source frame at time t − 1, into
the sensor destination frame at time t, using transformation
Tt,t−1 (Equation 3).
Thus, the STM is introduced in the hidden state and
performs a transformation on all feature maps of ht−1 , set
in frame t − 1, into frame t before update with the new
incoming input xt to form the new memory at t.
C. Training
In both static and dynamic cases, the network is presented
the input sequence x1:t and trained to predict the next n input
frames xt+1:t+n . The binary cross-entropy loss is calculated
and backpropagated only on the visible part of the output,
which is achieved by simply masking the prediction yt
with xt,vis and multiplying the resulting grid element-wise
with the occupancy part grid xt,occ . By using a loss that
V. E XPERIMENTAL R ESULTS
In this section, we perform experimental validation of
both the baseline DeepTracking and the STM-based variant
proposed in this paper. With a stationary vehicle, no spatial
transform is necessary and the two are identical.
A. Static Vehicle
For the static case, we consider an architecture with three
hidden layers of 16 GRU feature maps each. Computation of
the hidden state consists of 3×3 convolutional filters, applied
on the three layers (from bottom up) with strides of 1, 2, and
4 pixels, and receptive fields of 3 × 3, 7 × 7, and 15 × 15,
respectively. We consider an input field of view of 20 × 20
m2 discretised into cells of 20 × 20 cm2 , which translates
into a H × W = 101 × 101 input grid. With a hidden state
consisting of 48 feature maps, the additional static memory
contributes to H ×W ×D = 489, 648 of the total 1, 506, 865
parameters of the network.
The evaluation dataset consists of a 75 minute log collected from a stationary Hokuyo UMT-30LX 2D laser scanner, positioned in the middle of a busy urban intersection.
The area features dense traffic composed of buses, cars, cyclists and pedestrians, which results in extensive amounts of
occlusion. Consequently, at no point in time is the complete
scene fully observable. The dataset is subsampled at 8Hz and
split into a 65 minute set for training, and a 10 minute set
for testing occupancy prediction.
The full model was trained on an Nvidia Tesla K40 GPU
until convergence, using the unsupervised training procedure
described in Section III. The training set is split into minibatch sequences of 40 frames (5 seconds). For every minibatch, the network is shown 10 frames and trained to predict
F1 on Test data epoch 58- St Giles
0.9
0.9
0.80.8
GRU3DilConvBias_16
0.70.7
GRU3DilConvBias_48
GRU3DilConv_16
0.60.6
F1 measure
Prediction Performance [F1 score]
encourages the model to predict the state of the environment
in the future, the model is forced to capture the motion of
each object in the hidden representation.
In the static sensor scenario, the error is only backpropagated on the ground truth available, i.e the visible part of the
space. In the case of a moving sensor, however, an additional
constraint needs to be imposed to account for the fact that as
the robot moves in future frames, it will discover new space
that falls outside the field of view of the current frame. Given
this fact, the model should not be falsely penalised for failing
to accurately guess objects located within this new space
when the input is blanked out. This is similar in nature to
the static case, where the input grid also represents a frontier
between what the robot can perceive and understand of the
scene and the unknown world outside its field of view.
To address this, we apply an additional mask at training
time on the cost computation and error backpropagation to
represent the predictable space in future frames. Accounting
for this field of view drift is crucial in terms of tracking
performance, as it corrects for an objective function that
is otherwise skewed towards the incredibly difficult task
of predicting objects outside the field of view. This mask
has been overlaid in transparency over the ground truth
comparison outputs of Figure 1, and indicates the predictable
free space shrinking on future timesteps.
GRU3DilConv_48
GRU3_16_A
0.5
0.5
RNN48
RNN16 [2]
0.4
0.4
Model-free tracker [4]
0.3
0.3
0.2
0.2
11
12
0.25
13
14
15
0.5
16
0.75
17
18
1
19
20
Timesteps
Prediction time horizon [sec]
Fig. 3. F1 scores of network architectures when attempting to predict
the future occupancy of the scene in a 1.25 second time horizon. The F1
measure is computed with a threshold of 0.5 when considering a cell to be
predicted as occupied or free.
the next 10 frames, leading to two such sequences per 40frame mini-batch. This length of sequence covers the typical
lengths of occlusions observed but can be tuned accordingly.
1) Quantitative results: We first look to quantify the
gain in performance achieved by scaling up the original
DeepTracking network with the proposed architecture. We
compare a number of different architectures ranging from
the original [1] to the proposed model in [2], and compare
performance on the task of predicting the observable nearfuture scene occupancy given the input sequence x1:t . The
predicted output occupancy P (yt+n |x1:t ) is compared to
xt+n,occ , which corresponds to the observed occupancy of
the world at time t. A threshold of 0.5 is used to determine
whether a cell is predicted as occupied or free, and an F1
measure computed for each frame.
Figure 3 compares the F1 measures computed on each
of n = 10 blacked out future frames, given the 10 frames
in the past. All model predictions decrease over time as
would be expected, as the uncertainty of the state of the
world increases with the prediction horizon. There is a
notable step change in performance with neural architectures
compared to a state-of-the art model-free tracking pipeline
approach by [4]. If we increase the capacity of the original
RNN architecture of [1] (RNN16) from 16 to 48 feature
maps (RNN48), we obtain marginal performance increase.
In comparison, replacing the standard RNN48 hidden unit
state with three layers of 16 GRU units (GRU3 16 A)
provides significant improvement in prediction ability. The
affix 16 signifies that we decode only the last hidden unit
(composed of 16 features maps) to the output. Incorporating
dilated convolutions (GRU3DilConv 16) in place of traditional dense convolutions achieves comparable performance
to GRU3DilConv 16 for a similar output receptive field of
15×15. The model with dilated convolutions additionally
requires less computation and nearly one third fewer model
parameters than its dense counterpart GRU3DilConv 16.
Further, we experiment with GRU3DilConv 48 which decodes the full hidden state (composed of 48 features maps)
to the output. Performance of the model is maintained as
t
Input
i
GRU 1
ii
Hidden State
GRU 2
GRU 3
Output
Sequence
true positive
visible space
false positive
false negative
occluded space
car
static background
pedestrian
Fig. 4. Example of outputs produced by the system along with a selection of activations in the hidden layers. As highlighted in colour-coded circles, the
network is able to propagate the assumed motion of the objects even when in complete occlusion. The sample hidden layer activations shown highlight the
fact that lower layers in the hidden units (corresponding to a low dilation of the convolutions) capture the motion of small and slow moving objects such
as pedestrians (e.g pink circle) and static background (e.g yellow squares), whereas a higher level layer learns to detect moving vehicles (orange oval).
illustrated in Figure 3. This is a departure from traditional
architectures where information from the different scales of
the hidden units is directly passed to the output. This may
assist positively for tasks such as semantic labelling, where
scale information is essential. Lastly, performance of the
full model with added static biases (GRU3DilConvBias 16
and GRU3DilConvBias 48) remains commensurate to that
of GRU3DilConv 16 and GRU3DilConv 48 (Figure 3), and
the learned static bias values may convey useful information
such as that of the static background layout. We consider the
B. Moving vehicle
In this section we compare our baseline model
GRU3DilConv 48 (BaselineDT) with an equivalent architecture that incorporates the Spatial Transformer module into
the hidden belief state (STM). As with the static case, we
evaluate the ability to predict future frames given blacked
out input, and illustrate the achieved occlusion-free tracking
performance on a series of examples selected from the test
set.
The evaluation dataset was collected over a 35 minute
period, from a moving vehicle equipped with two Velodyne
HDL64E lasers, resulting in a 360 degree field of view. The
3D point clouds were reduced to a 2D scan by considering
the range of points within 0.6-1.5 meters height from the
ground.
The network was trained on mini-batches of 40 sequences
with a frame rate of 10Hz, alternating between 5 inputs
shown, and 5 inputs hidden. This higher frequency is better
adapted to the moving vehicle case given the input field of
view of 18 × 18 m2 and a vehicle mean velocity of 20 miles
per hour. Longer occluded sequences would lead to increased
loss in useful memory due to the aforementioned drift of
field of view. Finally, we performed an 80/20 split of the
data into training and test set with no overlap in location,
and trained our model on an Nvidia Tesla K40 GPU until
convergence. Our architecture is implemented on Torch and
uses the Spatial Transformer GPU-implementation of [16].
a) Quantitative Results: Figure 5 represents the F1
measure as computed with STM and with the baseline DT.
0.85
Prediction Performance [F1 score]
GRU3DilConv 48 model in the rest of this paper.
2) Qualitative results: To better understand what the network has learned, we also qualitatively analyse a typical test
sequence of length 3 seconds from GRU3DilConv 48, along
with the network output and selected hidden state feature
maps in Figure 4. The network is able to track pedestrians
through full occlusion and the unobserved hallucinated tracks
are represented in blue in the output sequence.
It can be seen that the GRU1 feature maps appear to
have learned to capture the static background, as activations
remain stationary during the sequence (i), and track pedestrians, as highlighted through the pink circles (ii). The GRU2
feature map also captures the motion of pedestrians moving
upwards to the left, while, interestingly, the GRU3 unit seems
to activate only on the car that appears to the top right at
frame 2 (indicated by the orange box). This provides empirical support for the use of Dilated Convolutions which allow
the model to capture patterns of increasing receptive fields
in the hidden state units, while requiring fewer parameters
than more traditional dense convolutional kernels.
In general, we observe that information regarding objects
in the scene is captured in the hidden state, and moves
spatially through the feature maps according to the motion
of the object in the input. This can be problematic when
extending tracking to a moving platform, as the object motion
and vehicle motion are coupled. We address this concern in
the following section.
Baseline
0.8
STM
0.75
0.7
0.65
0.6
0.15
0.3
0.45
0.6
Prediction time horizon [sec]
Fig. 5. Positive contribution of the Spatial Transformer to the network’s
ability to correctly predict the future occupancy of the scene in a 0.6 sec
time horizon. The baseline DT does surprisingly well which we attribute
to the benign test set where the vehicle mostly evolves at constant velocity
down straight roads.
The baseline DT is identical to STM with the exception
that no egomotion is taken into account. In other words, no
additional mask is applied to the cost computation and backpropagation, and hidden states are not forward transformed
into the next sensor frame.
With no egomotion information, one might expect the
F1 measure for the baseline to be very poor. Surprisingly
however, it does very well as illustrated in Figure 5. We
suggest two explanations for this. Firstly, we posit that this is
due to the dataset being relatively benign in terms of vehicle
motion patterns. As most of the driving occurs down straight
roads and at relatively constant velocity (∼ 20 mph) the
baseline may have learned a constant velocity model that
could be used to correct the hidden state update. Secondly,
the F1 measure here may be less informative, as the dataset
is dominated by static objects such as walls and buildings.
As such, a large fraction of the F1 score can be attained
by learning the vehicle’s motion and merely capturing static
scenes. Nonetheless, the STM offers a clear improvement
over the baseline DT in all future frames.
1) Qualitative Results: To qualitatively evaluate the
model, we show a selection of sequences where the model
does well, and where it does more poorly. As the dataset does
not contain as many occlusions and for as long as the static
set due to the setting of a vehicle driving through an urban
environment, we look at the ability of the network to predict
what happens in occlusion by maintaining the blacking out
of the input every 5 frames at test time.
Figure 6 shows a compelling example of STM accurately
tracking both dynamic and static objects through occlusion.
In particular, the model accurately predicts the position of
two static objects of different sizes when future frames are
blacked out, and is also able to maintain the tracks when
both objects are fully occluded.
Figure 7 illustrates how STM accurately predicts the tracks
of both a moving vehicle (red circle) and two occluded
pedestrians (orange rectangle) whereas the baseline model
fails. For the latter, the predicted track of the vehicle gradually shifts from the ground truth until complete failure in
frame 5 (false negative area), and it fails to separate and
Sequence
0.5 sec
t
Output
Input
0.25 sec
1 sec
(continued)
t
static
visible space
Input
dynamic
true positive
Output
occluded space
false positive
false negative
Fig. 6. Left to Right, Top to Bottom: Example sequence (1 second) of dynamic and static object tracking through occlusion by the STM. The output of
the network is evaluated against the visible ground truth. We highlight several objects of interest that are correctly tracked through complete occlusion.
Sequence
0.6 sec
t
STM
Output
Baseline
Output
Input
0.3 sec
Fig. 7. Left to Right: Output sequence (0.6 seconds) of dynamic and static object tracking through occlusion as predicted by the Spatial Transformer
Model (Top) and the Baseline Model (Bottom). We show the colour-coded comparison to the visible ground truth, with false negatives in red, false positives
in blue, and true positives in green. Where the Spatial Transformer network consistently accurately tracks two pedestrians (orange rectangle) occluded by
a moving vehicle (red), the Baseline model fails to capture the dynamics of both the vehicle and the pedestrians. This is particularly visible on the last
frame, with most of the area around the actual location of these objects showing false negative occupancy prediction for the Baseline model.
track the pedestrians through occlusion by the vehicle.
VI. C ONCLUSION
In this paper, we have proposed an approach to perform
object tracking for a mobile robot travelling in crowded
urban environments, building on the previously proposed
DeepTracking framework ([1], [2]). Crucially, unlike classical techniques which employ a multi-stage pipeline, this
approach is learned end-to-end with limited architectural
choices. By employing a Spatial Transformer module, the
model is able to exploit noisy estimates of visual egomotion
as a proxy for true vehicle motion. Experimental results
demonstrate that our method performs favourably to DeepTracking in terms of accurately predicting future states, and
show that the model can capture the location and motion of
cars, pedestrians, cyclists, and buses, even when in complete
occlusion.
Future work will look to estimate ego-motion, explore
modalities such as radar, and extend the approach to 3D.
ACKNOWLEDGEMENT
The authors would like to gratefully acknowledge support
of this work by the UK Engineering and Physical Sciences
Research Council (EPSRC) Doctoral Training Programme
(DTP) and Programme Grant DFR-01420, as well as the
Advanced Research Computing services at the University of
Oxford for providing access to their computing cluster.
R EFERENCES
[1] P. Ondrúška and I. Posner, “Deep tracking: Seeing beyond seeing
using recurrent neural networks,” in The Thirtieth AAAI Conference on
Artificial Intelligence (AAAI), Phoenix, Arizona USA, February 2016.
[2] P. Ondrúška, J. Dequaire, D. Z. Wang, and I. Posner, “End-to-end
tracking and semantic segmentation using recurrent neural networks,”
arXiv preprint arXiv:1604.05091, 2016.
[3] T.-D. Vu, O. Aycard, and N. Appenrodt, “Online Localization and
Mapping with Moving Object Tracking in Dynamic Outdoor Environments,” in Intelligent Vehicles Symposium, 2007 IEEE, June 2007, pp.
190–195.
[4] D. Z. Wang, I. Posner, and P. Newman, “Model-free detection and
tracking of dynamic objects with 2d lidar,” The International Journal
of Robotics Research, vol. 34, no. 7, pp. 1039–1063, 2015.
[5] A. Petrovskaya and S. Thrun, “Model based vehicle detection and
tracking for autonomous urban driving,” Autonomous Robots, vol. 26,
no. 2, pp. 123–139, 2009.
[6] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in neural
information processing systems, 2012, pp. 1097–1105.
[7] G. E. Dahl, D. Yu, L. Deng, and A. Acero, “Context-dependent pretrained deep neural networks for large-vocabulary speech recognition,”
Audio, Speech, and Language Processing, IEEE Transactions on,
vol. 20, no. 1, pp. 30–42, 2012.
[8] T. Wang, D. J. Wu, A. Coates, and A. Y. Ng, “End-to-end text recognition with convolutional neural networks,” in Pattern Recognition
(ICPR), 2012 21st International Conference on. IEEE, 2012, pp.
3304–3308.
[9] K. Cho, B. Van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares,
H. Schwenk, and Y. Bengio, “Learning phrase representations using
rnn encoder-decoder for statistical machine translation,” arXiv preprint
arXiv:1406.1078, 2014.
[10] S. Choi, K. Lee, and S. Oh, “Robust modeling and prediction in
dynamic environments using recurrent flow networks.”
[11] V. Patraucean, A. Handa, and R. Cipolla, “Spatio-temporal
video autoencoder with differentiable memory,” arXiv preprint
arXiv:1511.06309, 2015.
[12] W. Lotter, G. Kreiman, and D. Cox, “Deep predictive coding networks for video prediction and unsupervised learning,” arXiv preprint
arXiv:1605.08104, 2016.
[13] M. Jaderberg, K. Simonyan, A. Zisserman, et al., “Spatial transformer
networks,” in Advances in Neural Information Processing Systems,
2015, pp. 2017–2025.
[14] F. Yu and V. Koltun, “Multi-scale context aggregation by dilated
convolutions,” arXiv preprint arXiv:1511.07122, 2015.
[15] S. Xingjian, Z. Chen, H. Wang, D.-Y. Yeung, W.-k. Wong, and W.-c.
WOO, “Convolutional lstm network: A machine learning approach
for precipitation nowcasting,” in Advances in Neural Information
Processing Systems, 2015, pp. 802–810.
[16] M. Oquab, “Module for spatial transformer networks,” https://github.
com/qassemoquab/stnbhwd/.
| 2cs.AI
|
Maximal subgroups of SL(n, Z)
arXiv:1511.05767v2 [math.GR] 17 Apr 2016
Tsachik Gelander and Chen Meiri
April 19, 2016
Dedicated to the 75’th birthday of Herbert Abels
1
Introduction
More than three decades ago, in a remarkable paper [MS81], Margulis and Soifer proved
existence of maximal subgroups of infinite index in SL(n, Z), answering a question of Platonov.
Moreover, they proved that there are uncountably many such subgroups. Since then, it is
expected that there should be examples of various different nature. However, as the proof is
non-constructive and relies on the axiom of choice, it is highly non-trivial to lay one’s hands
on specific properties of the resulting groups.
Our purpose here is to show that indeed, maximal subgroups ∆ ≤ SL(n, Z) of different
nature do exist. However, our methods say nothing about the intrinsic algebraic structure of
∆. We do not gain any understanding about the abstract groups ∆, instead we are focusing
on the way it sits inside SL(n, Z). The two point of views that we consider are:
• The associated permutation representation Γ y Γ/∆.
• The action of ∆ on the associated projective space P = Pn−1 (R).
The main results of this paper are:
Theorem 1.1. Let n ≥ 3. There are 2ℵ0 infinite index maximal subgroups in SL(n, Z).
Theorem 1.2. Let n ≥ 3. There exists a maximal subgroup ∆ of SL(n, Z) which does not have
a dense orbit in P. In particular, the limit set of ∆ (in the sense of [CG00]) is nowhere-dense.
Theorem 1.3. Let n ≥ 3. There exists an infinite index maximal subgroup M of PSL(n, Z)
and an element g ∈ PSL(n, Z) such that M ∩ gM g −1 = {id}.
Theorem 1.4. Let n ≥ 3. There exists a primitive permutation action of SL(n, Z) which is
not 2-transitive.
1
Remark 1.5. The theorems remain true also for SL(n, Q) instead of SL(n, Z).
Remark 1.6. Recall that the Margulis–Soifer theorem is much more general, i.e. holds for any
finitely generated non-virtually-solvable linear group Γ. Our results however relies on special
properties of SL(n, Z), n ≥ 3. In particular one important ingredient for us is the beautiful
result of Venkataramana about commuting unipotents, Theorem 2.8. Another ingredient is
the result of Conze and Guivarc’h, Theorem 2.12. Some of our results can be extended to the
class of arithmetic groups of higher Q-rank.
Acknowledgment: The first author was partially supported by ISF-Moked grant 2095/15.
The second author was partially supported by ISF grant 662/15.
2
Preliminaries
In this paper we always assume that n ≥ 3.
2.1
Projective space
Let n ≥ 3. The letter P denotes the (n − 1)-dimensional real projective space and fix some
compatible metric distP on P. For every 0 ≤ k ≤ n − 1, the set Lk of k-dimensional subspaces
of P can be endowed with the metric defined by
distLk (L1 , L2 ) := max{distP (x, Li ) | x ∈ L3−i for 1 ≤ i ≤ 2}
for every L1 , L2 ∈ Lk . Note that Lk is naturally homeomorphic to the Grassmannian Gr(k +
1, Rn ). For ε > 0 and a subset A ⊆ P we denote (A)ε := {x ∈ P | distP (x, A) < ε} and
[A]ε := {x ∈ P | distP (x, A) ≤ ε}. If A = {p} then we usually write (p)ε and [p]ε instead of
(A)ε and [A]ε .
2.2
Unipotent elements
Definition 2.1 (Rank 1 unipotent elements). We say that a unipotent element u has rank 1 if
rank(u−In ) = 1. The point pu ∈ P which is induced by the euclidean line {ux − x | x ∈ Rn } is
called the point of attraction of u. The (n − 2)-dimensional subspace Lu ⊆ P which is induced
by the euclidean (n − 1)-dimensional space {x ∈ Rn | ux = x} is called the fixed hyperplane of
of u. The set of rank-1 unipotent elements in SL(n, Z) is denoted by U.
The following two lemmas follow directly from the definition of U and are stated for future
reference.
2
Lemma 2.2 (Structure of unipotent elements). The set U can be divided into equivalence
classes in the following way: u, v ∈ U are equivalent if there exist non-zero integers r and s
such that us = v r . The map u 7→ (pu , Lu ) is a bijection between equivalence classes in U and
the set of pairs (p, L) where p ∈ P is a rational point and L ⊆ Ln−2 is an (n − 2)-dimensional
rational subspace which contains p.
Lemma 2.3 (Dynamics of unipotent elements). Let u ∈ U. For every ε > 0 and every
δ > 0 there exists a constant c such that if m ≥ c and v = um , then v k (x) ∈ (pu )ε for every
x ∈ P \ (Lu )δ and every k 6= 0. Note that the previous lemma implies that pu = pv and
Lu = Lv .
2.3
Schottky systems
Definition 2.4. Assume that S is a non-empty subset of U and A ⊆ R are closed subsets of
P. We say that S is a Schottky set with respect to the attracting set A and the repelling set
R and call the triple (S, A, R) a Schottky system if for every u ∈ S there exist two positive
numbers δu ≥ εu such that the following properties hold:
1. uk (x) ∈ (pu )εu for every x ∈ P \ (Lu )δu and every k 6= 0;
2. If u 6= v ∈ S then (pu )εu ∩ (Lv )δv = ∅;
3. ∪u∈S (pu )εu ⊆ A;
4. ∪u∈S (Lu )δu ⊆ R.
Definition 2.5. The Schottky system (S, A, R) is said to be profinitely-dense if S generates a
profinitely-dense subgroup of SL(n, Z). We say that the Schottky system (S+ , A+ , R+ ) contains
the Schottky system (S, A, R) if S+ ⊇ S, A+ ⊇ A and R+ ⊇ R.
Lemma 2.6. Let (S, A, R) be a Schottky system. Assume that [p]ε ∩ A = ∅ and [L]δ ∩ R = ∅
where δ ≥ ε > 0 and p is a rational point which is continued in a rational subspace L ∈ Ln−2 .
Denote A+ = A ∪ [p]ε and R+ = R ∪ [L]δ . Then there exist v ∈ U with p = pv , L = Lv such
that (S+ , A+ , R+ ) is Schottky system which contains (S, A, R) where S+ := S ∪ {v}.
Proof: Lemma 2.2 implies that there exists u ∈ U such that pu = p and Lu = L. Lemma 2.3
implies that there exists m ≥ 1 such that v := um satisfies the required properties.
The following lemma is a version of the well known ping-pong lemma:
Lemma 2.7 (Ping-pong). Let (S, A, R) be a Schottky system. Then the natural homomorphism ∗u∈S hui → hSi is an isomorphism.
3
An important ingredient for our methods is the following beautiful result:
Theorem 2.8 (Venkataramana, [Ve87]). Let Γ be a Zariski-dense subgroup of SL(n, Z). Assume that u ∈ U ∩ Γ, v ∈ Γ is unipotent and hu, vi ' Z2 . Then Γ has finite index in SL(n, Z).
In particular, if Γ is profinitely-dense then Γ = SL(n, Z).
Note that if u, v ∈ SL(n, Z) ∩ U and pu = pv then (u − 1)(v − 1) = (v − 1)(u − 1) = 0 and
in particular uv = vu. Thus we get the following lemma:
Lemma 2.9. Let g ∈ SL(n, Z) and u1 , u2 ∈ U. Assume that pu2 = gpu1 and Lu2 6= gLu1 .
Then hu1 , g −1 u2 gi ' Z2 .
Lemma 2.10. Assume that g is an element of SL(n, Z), (S, A, R) is a profinitely-dense
Schottky system, δ ≥ ε > 0, p1 and p2 are rational points and L1 and L2 are rational (n − 2)dimensional subspaces such that the following conditions hold:
1. ([p1 ]ε ∪ [p2 ]ε ) ∩ R = ∅ and ([L1 ]δ ∪ [L2 ]δ ) ∩ A = ∅;
2. [p1 ]ε ∩ [L2 ]δ = ∅ and [p2 ]ε ∩ [L1 ]δ = ∅;
3. p1 = gp2 and L1 6= gL2 .
Denote A+ = A ∪ [p1 ]ε ∪ [p2 ]ε and R+ = R ∪ [L1 ]δ ∪ [L2 ]δ . Then there exists a set S+ ⊇ S such
that (S+ , A+ , R+ ) is a Schottky system which contains (S, A, R) and hS+ , gi = SL(n, Z).
Proof: For every 1 ≤ i ≤ 2 choose ui ∈ U such that pui = pi and Lui = Li . Lemma 2.9 implies
that hu1 , g −1 u2 gi ' Z2 . Lemma 2.3 implies that there exists m ≥ 1 such that (S+ , A+ , R+ ) is
m
Schottky system where v1 := um
1 , v2 := u2 and S+ := S ∪ {v1 , v2 }. Theorem 2.8 implies that
hS+ , gi = SL(n, Z).
Definition 2.11. Let 1 ≤ k ≤ n. A k-tuple (p1 , . . . , pk ) of projective points is called generic
if p1 , . . . , pk span a (k − 1)-dimensional subspace of P. Note that the set of generic k-tuples of
P is an open subset of the product of k copies of the projective space, indeed it is even Zariski
open.
Theorem 2.12 (Conze-Guivarc’h, [CG00]). Assume that n ≥ 3 and that Γ ≤ SL(n, R) is a
lattice. Then Γ acts minimally of the set of generic (n − 1)-tuples.
Corollary 2.13. Assume that n ≥ 3 and Γ ≤ SL(n, R) is a lattice. For every 1 ≤ i ≤ 2
let pi ∈ Li ∈ Ln−2 . Then for every positive numbers ε and δ there exists g ∈ Γ such that
gp1 ∈ (p2 )ε and gL1 ∈ (L2 )δ .
The proof of the following Proposition is based on the proof of the main result of [AGS14].
4
Proposition 2.14. Assume that n ≥ 3 and p ∈ L ∈ Ln−2 . Then for every δ ≥ ε > 0 there
exists a finite subset S ⊆ U such that (S, A, R) is a profinitely-dense Schottky system where
A := [p]ε and R := [L]δ .
Proof: We recall some facts about Zariski-dense and profinitely-dense subgroups. For a
positive integer d ≥ 2 let πd : SL(n, Z) → SL(n, Z/dZ) be the modulo-d homomorphism and
denote Kd := ker πd .
(a) If H ≤ SL(n, Z) and πp (H) = SL(n, Z/pZ) for some odd prime p then H is Zariski-dense,
[We96] and [Lu99].
(b) The strong approximation theorem of Weisfeiler [We84] and Nori [No87] implies that if
a subgroup H of SL(n, Z) is Zariski-dense then there exists some positive integer q such
that πd (H) = SL(n, Z/dZ) whenever gcd(q, d) = 1.
(c) If H ≤ SL(n, Z), π4 (H) = SL(n, Z/4Z) and πp (H) = SL(n, Z/pZ) for all odd primes p
then H is profintiely-dense in SL(n, Z).
Fix δ ≥ ε > 0 and set A := [p]ε and R := [L]δ . For every 1 ≤ i ≤ 2n2 − n, fix a point pi
beloning to an (n − 2)-dimensional subspace Li and positive numbers δi ≥ εi > 0 such that
the following two conditions hold:
1. ∪1≤i≤2n2 −n (pi )εi ⊆ A and ∪1≤i≤2n2 −n (Li )δi ⊆ R;
2. For every 1 ≤ i 6= j ≤ 2n2 − n, (pi )εi ∩ (Lj )δj = ∅.
For every 1 ≤ i 6= j ≤ n, let ei,j ∈ SL(n, Z) be the matrix with 1 on the diagonal and on the
(i, j)-entry and zero elsewhere and let e1 , . . . , en2 −n be an enumeration of the ei,j ’s. Denote
the exponent of SL(n, Z/3Z) by t. If g1 , . . . , gn2 −n ∈ K3 and k1 , . . . , kn2 −n are positive integers
then π3 (H1 ) = SL(n, Z/3Z) where ui := gi eitki +1 gi−1 and H1 := hui | 1 ≤ i ≤ n2 − ni. Note
that for every u ∈ U and g ∈ SL(n, Z), pgug−1 = gpu and Lgug−1 = gLu . Thus, Lemma 2.3 and
Corollary 2.13 imply that it is possible to choose gi ’s and ki ’s such that:
3. uki (x) ∈ (pi )εi for every 1 ≤ i ≤ n2 − n, every x 6∈ (Li )δi and every k 6= 0.
In particular, {u1 , . . . , un2 −n } is a Schottky set with respect to A and R which generates a
Zariski-dense subgroup H1 .
The strong approximation theorem implies that there exists some positive integer q such
that πd (H1 ) = SL(n, Z/dZ) whenever gcd(q, d) = 1. Denote the exponent of SL(n, Z/q 2 Z) by
r. As before, there exist gn2 −n+1 , . . . , g2n2 −2n ∈ Kq2 and positive integers kn2 −n+1 , . . . , k2n2 −2n
i +1 −1
such that the elements of the form ui := gi erk
gi satisfy:
i
4. πq2 (H2 ) = SL(n, Z/q 2 Z) where H2 := hui | n2 − n + 1 ≤ i ≤ 2n2 − 2ni;
5
5. uki (x) ∈ (pi )εi for every n2 − n + 1 ≤ i ≤ 2n2 − 2n, every x 6∈ (Li )δi and every k 6= 0.
Denote S := {u1 , . . . , u2n2 −2n }. Item (c) implies that πd (hSi) = SL(n, Z/dZ) for every
d ≥ 1. Thus, (S, A, R) is the required profinitely-dense Schottky system.
The following lemma will be needed in Section 5.
Lemma 2.15. Assume that k is an element of SL(n, Z), δ ≥ ε > 0, p1 , p2 and p3 are rational
points and L1 , L2 and L3 are rational (n − 2)-dimensional subspaces such that the following
conditions hold
• For every 1 ≤ i 6= j ≤ 3, pi ∈ Li and [pi ]ε ∩ [Lj ]δ = ∅;
• p2 = kp1 , p3 := k 2 p1 , L2 = kL1 , L3 6= k 2 L1 ;
Denote A := ∪1≤i≤3 [pi ]ε and R := ∪1≤i≤3 [Li ]δ . Then there exists a profinitely-dense Schottky
system (S, A, R) such that hSi ∩ khSik −1 6= {id} and hS, ki = SL(n, Z).
Proof: Proposition 2.14 implies that there exists a finite S0 = {u1 , . . . , ur } ⊆ U such that
(S0 , A0 , R0 ) is profinitely-dense Schottky system where A0 := [p1 ]ε and R0 := [L1 ]δ . A quick
look at the proof of this proposition implies that we can assume that pu1 = p1 and Lu1 = L1 .
Denote w2 := ku1 k −1 and choose w3 ∈ U such that pw3 = p3 , Lw3 = L3 . Lemma 2.9 implies
that hu1 , k −2 w3m k 2 i ∼
= Z2 for every m 6= 0. Lemma 2.3 implies that for large enough value of
m, (S, A, R) is a profinitely-dense Schottky system where S := S0 ∪ {w2m , w3m }. Note that
w2 ∈ hSi ∩ khSik −1 and that Theorem 2.8 implies that hS, ki = SL(n, Z).
3
Proof of Theorem 1.1
Zorn’s lemma implies that every proper subgroup H of SL(n, Z) is contained in a maximal
subgroup M (Since SL(n, Z) is finitely generated an increasing union of proper subgroups is a
proper subgroup). If H is profinitely-dense then so is M ; hence M should have infinite index.
Thus, Theorem 1.1 follows for the following proposition:
Theorem 3.1. Let n ≥ 3. There exist 2ℵ0 infinite-index profinitely-dense subgroups of
SL(n, Z) such that the union of any two of them generates SL(n, Z).
Proof: For every non-negative integer i fix a rational point pi belonging to a rational (n − 2)dimensional subspace Li and two numbers δi ≥ εi > 0 such that [pi ]εi ∩ [Lj ]δj = ∅ for every
i 6= j. Let A and R be the closures of ∪i≥0 (pi )i and ∪i≥0 (Li )δi respectively. Proposition 2.14
implies that there exists a finite subset S0 ⊆ U such that (S0 , A0 , R0 ) is a profinitely-dense
Schottky system where A0 = [p0 ]ε0 and R0 = [L0 ]δ0 . Lemmas 2.2 and 2.3 imply that for every
i ≥ 1 there are ui,1 , ui,2 ∈ U such that:
6
1. pi = pui,1 = pui,2 and Lui,1 6= Lui,2 ⊆ (Li )δi (hence, hui,1 , ui,2 i ∼
= Z2 );
2. uki,j (x) ∈ (pi )ε for every 1 ≤ j ≤ 2, every x 6∈ (Li )δi and every k 6= 0.
For every function f from the positive integers to {0, 1} the set Sf := S0 ∪ {ui,f (i)|i≥1 } is a
Schottky set with respect to the attracting set A and the repelling set R. If f and g are
distinct function then Sf ∪ Sg contains {ui,1 , ui,2 } for some i ≥ 1 so Theorem 2.8 implies that
hSf ∪ Sg i = SL(n, Z).
Theorem 1.1 implies that there are non-conjugate infinite-index maximal subgroups in
SL(n, Z). Indeed, since SL(n, Z) is countable so is the conjugacy class of every infinite index
subgroup. We can say a little more:
Proposition 3.2. There are infinite-index maximal subgroups whose actions on P have different topological properties. For example, when n ≥ 4 some but not all infinite-index maximal
subgroups M of SL(n, Z) have the following property:
There exists a finitely-generated subgroup H ≤ M and a 2-dimensional subspace l ∈ L2
which is contained in the closure of every H-orbit.
The existence of maximal subgroups with this property follows form the existence of maximal subgroup M which contains an element g ∈ SL(n, Z) with eigenvalues α1 , . . . , αn such
that α1m and αnm are not real for all m ≥ 1 and |α1 | < |αi | < |αn | for all 2 < i < n − 1. The
construction of maximal subgroups without this property can be done by using similar ideas
to the ones used in the proofs of Theorems 1.2, 1.3 and 1.4 below.
4
Proof of Theorem 1.2
In order to prove the Theorem 1.2 it is enough to find an infinite-index maximal subgroup ∆
and two open subsets U and V such that gU ∩ V = ∅ for all g ∈ ∆. Let V ∨ be the dual of
V := Rn . The natural map form P = Gr(1, V ) to Gr(n − 1, V ∨ ) is an SL(n, Z)-equivariant
homeomorphism. Thus, Theorem 1.2 follows from the following proposition:
Proposition 4.1. Let L1 and L2 be distinct (n − 2)-dimensional subspaces of P. There exist
ρ > 0 and a free profinitely-dense subgroup H such that if L ⊆ (L2 )ρ and gL ⊆ (L1 )ρ for some
g ∈ SL(n, Z) and some (n − 2)-dimensional subspace L then hg, Hi = SL(n, Z).
The rest of this section is devoted to the proof of Proposition 4.1. The basic idea is to
find ρ > 0 for which Lemma 2.10 allows us to inductively construct an ascending sequence
(Sk , AK , Rk )k≥0 of profintely-dense Schottky system such that hSk , gk i = SL(n, Z) for every
k ≥ 1 where (gk )k≥1 is an enumeration of the g’s such that g(L2 )ρ ∩ (L1 )ρ 6= ∅.
7
Before starting with the formal proof let us briefly explain the main idea of the proof. Let
(S0 , A0 , R0 ) be a profintly-dense Schottky system such that L1 ∩ A0 = ∅ and L2 \ R0 6= ∅.
Assume for simplicity that g1 L2 = L1 . Our goal is to show that we can find a finite Schottky
system (S1 , A1 , R1 ) which contains (S0 , A0 , R0 ) such that hS1 , g1 i = SL(n, Z). Choose a point
w ∈ L2 \ R0 . We would like to apply Lemma 2.10 with respect to w and g1 w and suitable
hyperplanes w ∈ Lw and gw1 ∈ Lg1 w . The problem is that it is possible that g1 w ∈ R0 , so the
assumptions of Lemma 2.10 do not hold no matter what Lw and Lg1 w are. Therefore we first
construct a Schottky system (S∗ , A∗ , R∗ ) such that w 6∈ R∗ , S∗ := S ∪ {u} and g1 w ∈ Lu .
Note that there is a lot of freedom in choosing the fixed hyperplane Lu and the attracting
point pu of u. The existense of (S∗ , A∗ , R∗ ) is guaranteed by some technical assumptions on
(S0 , A0 , R0 ); One of these assumptions is the requirement that A0 ∩ L1 = ∅ which guarantees
that g1 w 6∈ A0 . Denote h := g1−1 ug1 so h(w) = w. Note that if (S1 , A1 , R1 ) is a finite Schottky
system such that hS1 , hi = SL(n, Z) then also hS1 , g1 i = SL(n, Z). The advantage now is that
we can choose a rational point w1 which is very close to w such that hw1 6= w1 and hw1 is also
very close to w. In particular w1 , hw1 6∈ R∗ . Once more we use some technical assumptions
on (S∗ , A∗ , R∗ ) to show that there exist rational hyperplanes w1 ∈ Lw1 and hw1 ∈ Lgw1 and
positive numbers ε and δ for which the assumptions of Lemma 2.10 hold. Thus, there exists
a finite Schottky system (S1 , A1 , R1 ) such that hS1 , hi = SL(n, Z).
We now start the formal proof. Let L0 ,L1 and L2 be fixed distinct rational (n − 2)dimensional subspaces of P and let p0 ∈ L0 \ L1 ∪ L2 be a fixed point. We will use p0 and L0
in the last part of the proof to construct the profinitely-dense Schottky system (S0 , A0 , R0 ).
Claim 4.2. There exists ρ > 0 with the following properties:
1. For every x ∈ [L1 ]ρ there exists an (n − 2)-dimensional subspace Lx containing x such
that Lx ∩ [p0 ]ρ = ∅ and Lx \ [L0 ∪ L1 ∪ L2 ]ρ 6= ∅.
2. For every (n − 2)-dimensional subspace L ⊆ [L2 ]ρ there exist x ∈ L \ [L0 ∪ L1 ]ρ and an
(n − 2)-dimensional subspace Lx containing x such that Lx is not contained in [L2 ]ρ and
Lx ∩ [p0 ]ρ = ∅.
Proof of Claim 4.2: Fix ρ0 > 0 such that:
• L0 and L1 are not contained in [L2 ]ρ0 ;
• p0 6∈ [L1 ]ρ0 .
For every x ∈ [L1 ]ρ0 there exist an (n − 2)-dimensional subspace Lx containing x and ρx > 0
such that Lx ∩ [p0 ]ρx = ∅ and Lx \ [L0 ∪ L1 ∪ L2 ]ρx 6= ∅. Since [p0 ]ρx and [L0 ]ρx are closed sets, if
y is close enough to x and the hyperplane Ly containing y is chosen to be close enough to Lx
8
then Ly ∩ [p0 ]ρx = ∅ and Ly \ [L0 ∪ L1 ∪ L2 ]ρx 6= ∅. Thus, the compactness of [L1 ]ρ0 implies that
there exists a uniform ρ1 > 0 such that for every x ∈ [L1 ]ρ0 there exists an (n − 2)-dimensional
subspace Lx containing x such that Lx ∩ [p0 ]ρ1 = ∅ and Lx \ [L0 ∪ L1 ∪ L2 ]ρ1 6= ∅.
The set of (n − 2)-dimensional subspaces of P which are contained in [L2 ]ρ0 is compact.
Thus, a similar argument to the one above implies that there exist ρ2 > 0 such that for every
(n−2)-dimensional subspace L ⊆ [L2 ]ρ0 there exist x ∈ L\[L0 ∪L1 ]ρ2 and an (n−2)-dimensional
subspace Lx containing x such that Lx is not contained in [L2 ]ρ2 and Lx ∩ [p0 ]ρ2 = ∅. Then
ρ := min(ρ0 , ρ1 , ρ2 ) satisfies the requirements.
We fix ρ > 0 which satisfies the requirements of Claim 4.2.
Claim 4.3. Assume that A ⊆ R are closed sets such that:
1. [L0 ∪ L1 ]ρ ⊆ R and [p0 ]ρ ⊆ A;
2. For every x ∈ [L1 ]ρ there exists an (n − 2)-dimensional subspace Lx containing x such
that Lx ∩ A = ∅ and Lx \ (R ∪ [L2 ]ρ ) 6= ∅;
3. For every (n − 2)-dimensional subspace L ⊆ [L2 ]ρ there exist x ∈ L \ R and an (n − 2)dimensional subspace Lx containing x such that Lx is not contained in [L2 ]ρ and Lx ∩A =
∅.
Then for every (n − 2)-dimensional subspace L which is disjoint from A, not contained in
[L2 ]ρ and contains a point p ∈ L \ R there exists δ > 0 such that:
4. Items 2 and 3 hold for A∗ := A ∪ [p]δ and R∗ := R ∪ [L]δ instead of A and R.
5. [p]δ ∩ R = ∅ and [L]δ ∩ A = ∅.
Proof of Claim 4.3: For every x ∈ [L1 ]ρ there exists an (n − 2)-dimensional subspace
Lx containing x such that item 2 holds. By a small deformation of Lx we can assume that
p 6∈ Lx and Lx 6= L (so Lx in not contained in L ∪ R). Thus, there exists δx > 0 such that
Lx ∩ (A ∪ [p]δx ) = ∅ and Lx \ (R ∪ [L2 ]ρ ∪ [L]δx ) 6= ∅. If y is close enough to x and the
hyperplane Ly containing y is chosen to be close enough to Lx then Ly ∩ (A ∪ [p]δx ) = ∅ and
Ly \ (R ∪ [L2 ]ρ ∪ [L]δx ) 6= ∅. Thus, the compactness of [L1 ]ρ implies that there exists a uniform
δ1 > 0 such that for every x ∈ [L1 ]ρ there exists an (n−2)-dimensional subspace Lx containing
x such that Lx ∩ (A ∪ [p]δ1 ) = ∅ and Lx \ (R ∪ [L2 ]ρ ∪ [L]δ1 ) 6= ∅
Let L3 be an (n − 2)-dimensional subspaces of P which is contained in [L2 ]ρ so L3 6= L.
Item 3 states that there exists x ∈ L3 \ R and an (n − 2)-dimensional subspace Lx containing
x such that Lx is not contained in [L2 ]ρ and Lx ∩ A =
6 ∅. The set R is closed so by a small
deformation we can assume that x 6∈ L and p 6∈ Lx . Hence, there exists δL3 > 0 such that
x 6∈ [L]δL3 and Lx ∩ [p]δL3 = ∅. If L4 is very close to L3 then we can pick y ∈ L4 \ R which
9
is very close to x and an (n − 2)-dimensional subspace Ly containing y which is close to Lx
such that Ly is not contained in [L2 ]ρ , Ly ∩ A = ∅, y 6∈ [L]δL3 and Ly ∩ [p]δL3 = ∅. The set
of (n − 2)-dimensional subspaces of P which are contained in [L2 ]ρ is compact and thus there
exists a uniform δ2 > 0 such that for every (n − 2)-dimensional subspace L ⊆ [L2 ]ρ there exist
x ∈ L \ R ∪ [L0 ]δ2 and an (n − 2)-dimensional subspace Lx containing x such that Lx is not
contained in [L2 ]ρ and Lx ∩ (A ∪ [p]δ2 ) = ∅. Finally, if 0 < δ3 < min{dist(A, L), dist(R, p)}
then δ := min(δ1 , δ2 , δ3 ) satisfies the requirements.
Claim 4.4. Assume that ρ > 0, A and R satisfies the assumptions of Claim 4.3 and (S, A, R)
is a profinitely-dense Schottky system. Let g ∈ SL(n, Z) be a non-identity element for which
there exists an (n − 2) dimensional subspace L such that L ⊆ [L2 ]ρ and gL ⊆ [L1 ]ρ . Then
there exists (S+ , A+ , R+ ) which still satisfies the assumptions of Claim 4.3 (with respect to
the same ρ) and hS+ , gi = SL(n, Z).
Proof of Claim 4.4: Item 3 of Claim 4.3 implies that there exists a point w ∈ L\A∪R and
an (n − 2)-dimensional subspace Lw containing w which is not contained in [L2 ]ρ such that
Lw ∩ A = ∅. Denote w1 := gw and note that w 6= w1 since w 6∈ [L1 ]ρ while w1 ∈ [L1 ]ρ . Item 2
of Claim 4.3 implies that there exists an (n − 2)-dimensional subspace Lw1 containing w1 and
a point p1 ∈ Lw1 such that Lw1 ∩ A = ∅ and p1 ∈ Lw1 \ (R ∪ [L2 ]ρ ). By a small deformations
we can assume that w, w1 , p1 , Lw and Lw1 are rational and that w 6∈ Lw1 and p1 6∈ Lw . Set
r := min{dist(w, Lw1 ), dist(p1 , Lw ))}. Claim 4.3 implies that there exists 0 < δ1 < r such that
items 4 and 5 of Claim 4.3 hold for A∗ := A ∪ [p1 ]δ1 and R∗ := R ∪ [Lw1 ]δ1 instead of A and R.
Lemma 2.6 implies that there exists u ∈ U with pu = p1 and Lu = Lw1 such that (S∗ , A∗ , R∗ )
is Schottky system where S∗ := S ∪ {u}. Note that w 6∈ A∗ ∪ R∗ and that Lw ∩ A∗ = ∅.
Denote h := g −1 ug. Since h(w) = w and h is not the identity, there exists a point p2
such that p2 and p3 := hp3 are distinct and close as we wish to w. Moreover, we can assume
that p2 , p3 6∈ R∗ ∪ A∗ are rational points and that there exist rational (n − 2)-dimensional
subspaces Lp2 and Lp3 containing p2 and p3 such that Lp2 ∩ A = ∅ and Lp3 ∩ A = ∅ (just
choose L2 and L3 to be very close to Lw ). Finally by slightly deforming Lp2 and Lp3 we can
guarantee that Lp3 6= hLp2 , p2 6∈ Lp3 and p3 6∈ Lp2 . Claim 4.3 applied to A∗ and R∗ implies
that there exists 0 < δ2 < min(dist(p3 , L2 ), dist(p2 , L3 )) for which items 4 and 5 hold. An
additional use of Claim 4.3 with respect to A∗ ∪ [p2 ]δ2 and R∗ ∪ [L2 ]δ2 implies that there exists
a positive number δ3 such that items 4 and 5 hold with respect to A+ := A∗ ∪ [p2 ]δ2 ∪ [p3 ]δ3
and R+ := R∗ ∪ [L2 ]δ2 ∪ [L3 ]δ3 . Lemma 2.10 implies that there exists S∗ ⊆ S+ such that
(S+ , A+ , R+ ) is a Schottky system and hS+ , gi = SL(n, Z).
We are ready to complete the proof of Proposition 4.1. Let (gi )i≥1 be an enumeration
of the elements g of SL(n, Z) for which there exists an (n − 2) dimensional subspace L such
10
that L ⊆ [L2 ]ρ and gL ⊆ [L1 ]ρ . Proposition 2.14 and Claim 4.2 imply that there exists a
profinitely-dense Schottky system (S0 , A0 , R0 ) such that A0 := [p0 ]ρ and R0 := [L0 ∪ L1 ]ρ .
Claim 4.2 allows us to recursively construct an ascending sequence (Si , Ai , Ri )i≥0 of profinitelydense Schottky systems such that hSi , gi i = SL(n, Z) for every i ≥ 1. The set ∪k≥0 Sk freely
generates a free profinitely-dense subgroup H with the required property.
5
Proof of Theorems 1.3 and 1.4
It is enough to show that there exists elements g, k ∈ SL(n, Z) and an infinite-index profinitelydense subgroup H such that kHk −1 ∩ H 6= {id}, hH, ki = SL(n, Z) and gM g −1 ∩ M = {id}
for every H ≤ M SL(n, Z). Indeed, if M is any maximal subgroup which contains H then
SL(n, Z) acts primitively on SL(n, Z)/M but the action is not 2-transitive since the stabiliser
of the pair (M, kM ) is not trivial while the stabiliser of the pair (M, gM ) is trivial.
Definition 5.1. We say that the quadruple (S, A, N , R) is a (profinitely-dense) Schottky
system if the triple (S, A, R) is a (profinitely-dense) Schottky system and N is an open set
which contains A.
Before starting with the formal proof let us explain the main idea of the proof. Choose
g ∈ SL(n, Z) which fixes some point p ∈ P. Let (S0 , A0 , N0 , R0 ) be a profintiely dense
Schottky system such that gN0 ∩ N0 = ∅. Given a non-identity h ∈ SL(n, Z) we would like
to construct a Schottky system (S1 , A1 , N1 , R1 ) which contains (S0 , A0 , N0 , R0 ) such that
either hS0 , hi = SL(n, Z) or hS0 , g −1 hgi = SL(n, Z). In particular, every maximal subgroup
of SL(n, Z) which contains S1 does not contain {h, g −1 hg}. Since gN0 ∩ N0 = ∅, either
h(p) 6∈ N0 or g −1 hg(p) 6∈ N0 . Assume for example that the first case hold. Then we can use
the same technique as in the proof of Theorem 1.2 to construct (S1 , A1 , N1 , R1 ) such that
hS0 , hi = SL(n, Z). In particular, we need to add an element u ∈ U to S0 such that h(p) ∈ Lu
and Lu ∩ A0 = ∅. This is the place where we need the open set N0 ; A0 might contain some
closed ball B for which there exists a sequence of points (zn )n≥1 such that z := limn→∞ zn ∈ B
and zn 6∈ A0 for every n ≥ 1. If (Ln )n≥1 is any sequence of hyperplanes such that zn ∈ Ln and
Ln ∩ B = ∅ for every n ≥ 1 then limn→∞ Ln is the tangent hyperplane L to B at the point z.
The problem is that L might intersect the interior of A0 . Therefore, if h(p) is very close to B
it might happen that L ∩ A0 6= ∅ for every hyperplane L which contains h(p). In particular,
it is not possible to find an element u with the required properties. In order to avoid this
problem we use the open set N0 to guarantee that h(p) is not too close to A0 . Note that we
don’t need the open sets in the proof of Theorem 1.2 because we have a better control on the
projective actions of the elements we deal with (they send some line in a given open set to
some other given open set).
11
Claim 5.2. Let p ∈ P be fixed by a non-identity element g ∈ SL(n, Z). Let k ∈ SL(n, Z) be
an element such that id, k, k 2 , g, gk, gk 2 are pairwise distinct elements. Then there exists a
profinitely-dense Schottky system (S, A, N , R) which satisfies:
1. p 6∈ N ∪ R ∪ gN ∪ gR and N ∩ gN = ∅;
2. If x 6∈ N then there exists x ∈ Lx ∈ Ln−2 and px ∈ Lx such that px 6∈ N ∪ R, gpx 6∈ N
and Lx ∩ A = ∅;
3. khSik −1 ∩ hSi =
6 {id} and hS, ki = SL(n, Z).
Proof: Choose rational points p1 , p2 and p3 and rational (n − 2)-dimensional subspaces L1 ,
L2 and L3 such that:
• For every 1 ≤ i 6= j ≤ 3, pi ∈ Li and pi 6∈ Lj ;
• p2 = kp1 , p3 := k 2 p1 , L2 = kL1 , L3 6= kL1 and p1 , p2 , p3 , p4 , p5 , p6 are pairwise distinct
where p4 := gp1 , p5 := gp2 and p6 := gp3 ;
• p 6∈ ∪1≤i≤3 Li (so p 6∈ ∪1≤i≤3 Li ∪1≤i≤3 gLi since gp = p).
A compactness argument implies that there exist δ > 0 such that:
a) For every 1 ≤ i 6= j ≤ 3, [pi ]δ ∩ [Lj ]δ = ∅;
b) N ∩ gN = ∅ where N := ∪1≤i≤3 (pi )δ ;
c) p 6∈ N ∪ R ∪ gN ∪ gR where R := ∪1≤i≤3 [Li ]δ ;
d) For every x ∈ P there exists x ∈ Lx ∈ Ln−2 and px ∈ Lx such that px 6∈ N ∪ R and
gpx 6∈ N .
An additional compactness argument implies that there exists δ > ε > 0 such that:
e) For every x 6∈ N there exists x ∈ Lx ∈ Ln−2 and px ∈ Lx such that px 6∈ N ∪R, gpx 6∈ N
and Lx ∩ A = ∅ where A := ∪1≤i≤3 [pi ]ε .
Items (a), (b), (c), (e) and Corollary 2.15 imply that the required (S, A, N , R) exists.
Claim 5.3. Let p ∈ P be fixed by a non-identity element g ∈ SL(n, Z). Assume that
(S, A, N , R) is a profinitely-dense Schottky system which satisfies:
1. p 6∈ N ∪ R ∪ gN ∪ gR and N ∩ gN = ∅;
12
2. If x 6∈ N then there exists x ∈ Lx ∈ Ln−2 and px ∈ Lx such that px 6∈ N ∪ R, gpx 6∈ N
and Lx ∩ A = ∅;
Then for every non-identity h ∈ SL(n, Z), (S, A, N , R) is contained in a Schottky system
(S+ , A+ , N+ , R+ ) which still satisfies assumptions 1 and 2 and such that hS+ , hi = SL(n, Z)
or hS+ , g −1 hgi = SL(n, Z).
Proof: By replacing h with g −1 hg if necessary we can assume that y := hp 6∈ N . Under
this assumption we will show that there exists a Schottky system (S+ , A+ , N+ , R+ ) which
satisfies assumption 1 and 2 and hS+ , hi = SL(n, Z). Assumption 2 implies that there exist
y ∈ Ly ∈ Ln−2 and py ∈ Ly such that py 6∈ N ∪ R, gpy 6∈ N and Ly ∩ A = ∅. By making
small deformations we may assume that: y and Ly are rational, gpy 6= py and p 6∈ Ly . A
compactness argument implies that there exists δ > 0 such that:
a) [py ]δ ∩ R = ∅ and [Ly ]δ ∩ A = ∅;
b) N0 ∩ gN0 = ∅ where N0 := N ∪ (py )δ ;
c) p 6∈ N0 ∪ R0 ∪ gN0 ∪ gR0 where R0 := R0 ∪ [Ly ]δ ;
d) For every x 6∈ N there exists x ∈ Lx ∈ Ln−2 and px ∈ Lx such that px 6∈ N0 ∪ R0 ,
gpx 6∈ N0 and Lx ∩ A = ∅.
An additional compactness argument implies that there exists δ > ε > 0 such that:
e) If x 6∈ N0 then there exists x ∈ Lx ∈ Ln−2 and px ∈ Lx such that px 6∈ N0 ∪R0 , gpx 6∈ N0
and Lx ∩ A0 = ∅ where A0 := A ∪ [py ]ε .
Corollary 2.6 and items (a), (b), (c) and (e) implies that there exists u ∈ U such that
pu = py , Lu = Ly and (S0 , A0 , N0 , R0 ) is a Schottky system which still satisfies Assumptions
1 and 2 where S0 = S ∪ {u}. Note that h−1 uhp = p and that if (S+ , A+ , N+ , R+ ) contains
(S0 , A0 , N0 , R0 ) and hS+ , h−1 uhi = SL(n, Z) then also hS+ , hi = SL(n, Z).
Since h−1 uh has infinite order there exists some m ≥ 1 such that id, f, g, gf are pairwise
distinct elements where f := h−1 um h. Thus, every open neighbourhood of p contains a
rational point p1 such that p1 , p2 := f p1 , p3 := gp1 and p4 := gf p1 are 4 distinct points
which are contained in this neighbourhood. If this neighbourhood does not intersect N0 then
assumption 2 implies that for every 1 ≤ i ≤ 2 there exists an (n − 2)-subspace pi ∈ Li ∈ Ln−2
such that Li ∩ A0 = ∅. By using small deformations we can assume that the Li ’s and pi ’s are
rational, L2 6= f L1 , p 6∈ Li and pj 6∈ Li for every 1 ≤ i 6= j ≤ 4.
A compactness arguments implies that there exists δ > 0 such that:
a) For every 1 ≤ i 6= j ≤ 2, [pi ]δ ∩ (R0 ∪ [Lj ]δ ) = ∅ and [Li ]δ ∩ A0 = ∅;
13
b) N+ ∩ gN+ = ∅ where N+ := N0 ∪ (p1 )δ ∪ (p2 )δ ;
c) p 6∈ N+ ∪ R+ ∪ gN+ ∪ gR+ where R+ := R0 ∪ [L1 ]δ ∪ [L2 ]δ ;
d) If x 6∈ N0 then there exists x ∈ Lx ∈ Ln−2 and px ∈ Lx such that px 6∈ N+ ∪ R+ ,
gpx 6∈ N+ and Lx ∩ A0 = ∅.
An additional compactness argument implies that there exists δ > ε > 0 such that:
e) If x 6∈ N+ then there exists x ∈ Lx ∈ Ln−2 and px ∈ Lx such that px 6∈ N+ ∪ R0 ,
gpx 6∈ N+ and Lx ∩ A+ = ∅ where A+ := A ∪ [p1 ]ε ∪ [p2 ]ε .
Corollary 2.10 and items (a), (b), (c) and (e) imply that there exists S+ ⊇ S such that
(S+ , A+ , N+ , R+ ) is a Schottky system which satsfies items 1 and 2 and hS+ , f i = SL(n, Z).
We are now ready to complete the proof of Theorems 1.3 and 1.4. Let (hi )i≥1 be an
enumeration of the non-identity elements of SL(n, Z). Let p ∈ P be fixed by a non-identity
element g ∈ SL(n, Z). Let k ∈ SL(n, Z) be an element such that id, k, k 2 , g, gk, gk 2 are pairwise
distinct elements. Claims 5.2 and 5.3 allow us the recursively construct ascending sequence
(Si , Ai , Ni , Ri )i≥0 of profintely-dense Schottky systems such that khS0 ik −1 ∩ hS0 i =
6 {id} and
hS0 , ki = SL(n, Z) and for every i ≥ 1 either hhi , Si i = SL(n, Z) or hg −1 hi g, Si i = SL(n, Z).
The subgroup H := hSi | i ≥ 0i has the required properties stated in the first paragraph of
this section.
6
The SL(n, Q) case
Margulis’ and Soifer’s paper is concerned only with finitely generated group. In [GG08] the
first author and Glasner studied infinite index maximal subgroups in general linear groups
and proved a criterion for their existence. The goal of this section is to show that Theorems
1.3, 1.4 and 1.2 hold also for SL(n, Q). Since the proofs are very similar to the ones given for
SL(n, Z), we only sketch the required modifications needed for proving the SL(n, Q) version
of Theorem 1.3.
The first modification concerns the elements in the generating set S. It is not enough
anymore to deal only with unipotent elements. We say that an element g ∈ SL(n, Q) is
strongly-semisimple if it has a unique eigenvalue λ+
g of maximal absolute value and a unique
+
eigenvalue λ−
g of minimal absolute value. Let pg be the projective point corresponding to
+
+
the λ+
g -eigenspace and Lg ∈ Ln−1 the projective hyperplane corresponding to Im(Id − λg ).
+
+
−
Denote p−
g := pg −1 and Lg := Lg −1 . To simplify the notation, if u ∈ U (in particular u is not
−
+
−
strongly-semisimple) we denote p+
u = pu := pu and Lu = Lu := Lu .
14
Definition 6.1. Assume that S is a non-empty subset of SL(n, Q) consisting of elements
which are either strongly-semisimple or unipotent of rank 1. Let A ⊆ R be closed subsets of
P. We call the triple (S, A, R) a Schottky system if for every g ∈ S there exist two positive
numbers δg ≥ εg such that the following properties hold:
−
+
−
1. g k (x) ∈ (p+
g , pg )εg for every x ∈ P \ (Lg ∪ Lg )δg and every k 6= 0;
+
−
−
2. If g 6= h ∈ S then (p+
g , pg )εg ∩ (Lh ∪ Lh )δh = ∅;
−
3. ∪g∈S (p+
g , pg )εg ⊆ A;
−
4. ∪g∈S (L+
h , Lh )δv ⊆ R.
The basic idea in the proof of Theorem 1.2 is to pick two small enough disjoint open subset
U, V ⊆ P and to construct enumeration (gk )k≥1 of all the elements g ∈ SL(n, Z) such that
gU ∩ V = ∅. (In the proof we passed to the dual space but it is merely a convenience.) The
second step is to construct a unipotent Schottky system (S0 , A0 , R0 ) such that the following
items hold for k = 0:
(1) hSk i is profintiely dense in SL(n, Z);
(2) hSk i ∩ {gi | i ≥ 1} = ∅;
(3) There exists Sk+1 ⊇ Sk , Ak+1 ⊇ Ak and Rk+1 ⊇ Rk such that hSk+1 , gk+1 i = SL(n, Z)
and the first two items still hold for (Sk+1 , Ak+1 , Rk+1 ).
The main part of the proof of Theorem 1.2 is to inductively construct an ascending sequence
(Sk , Ak , Rk )k≥0 such that for every k ≥ 0 items (1)–(3) hold. Any infinite index maximal
subgroup ∆ of SL(n, Z) which contains Λ := h∪k≥1 Sk i has empty intersection with {gk | k ≥
1}. The existence of ∆ follows from Zorn’s lemma.
The fact that the gk ’s belong to SL(n, Z) is not used in the construction, and it is possible
to take enumeration (gk )k≥1 of all the elements g ∈ SL(n, Q) such that gU ∩ V = ∅ and
to inductively construct an ascending sequence (Sk , Ak , Rk )k≥0 such that for every k ≥ 0
items (1)–(3) hold where the equality hSk+1 , gk+1 i = SL(n, Z) in item (3) is replaced with
hSk+1 , gk+1 i ⊇ SL(n, Z). However, it is not possible to directly apply Zorn’s lemma since
SL(n, Q) is not finitely generated and the union of ascending sequence of proper subgroups of
SL(n, Q) might be equal to SL(n, Q). Thus, Λ might not be contained in a maximal subgroup.
An additional problem is that even if Λ is contained in a maximal subgroup then this maximal
subgroup might contain SL(n, Z) and have a non-empty intersection with {gk | k ≥ 1}.
In order to overcome these problems we take enumeration (Ck )k≥1 of the left cosets of
SL(Z[1/p]) in SL(n, Q) where p is some odd prime. One start with a unipotent Schottky
system (S0 , A0 , R0 ) in SL(Z[1/p]) such that the following items hold for k = 0:
15
(1∗ ) hSk ∩ SL(n, Z)i is profintiely dense in SL(n, Z) and Sk \ SL(n, Z) 6= ∅;
(2∗ ) hSk i ∩ {gi | i ≥ 1} = ∅;
(3∗ ) There exists Sk+1 ⊇ Sk , Ak+1 ⊇ Ak and Rk+1 ⊇ Rk such that:
(a) The first two items still hold for (Sk+1 , Ak+1 , Rk+1 );
(b) hSk+1 , gk+1 i ⊇ SL(n, Z);
(c) If k 6= 0, hSk+1 i ∩ Ck 6= ∅.
It is possible to inductively construct an ascending sequence (Sk , Ak , Rk ) such that for
every k items (1∗ )–(3∗ ) hold. The new requirement, Part (c) of item (3∗ ), brings almost no
difficulty since SL(n, Z[1/p]) is dense in SL(n, R) so in every coset it is easy to find an element
which still plays ping-pong with the elements of Sk . Note that this is the place where we
need to deal with strongly semi-simple elements since an element which is very close to a
strongly semi-simple element is strongly semi-simple while an element which is very close to
a unipotent element might not be unipotent. Denote Λ := hSk i. An ascending sequence of
proper subgroups which contain Λ cannot be equal to SL(n, Q). Indeed, if this happens then
one of the subgroups in the union contains the finitely generated group SL(n, Z[1/p]) and
has non-trivial intersections with all the cosets of SL(n, Z[1/p]) so it is equal to SL(n, Q), a
contradiction. Zorn’s lemma implies that Λ is contained in some maximal subgroup ∆. We
need to show that ∆ ∩ {gk | k ≥ 1} = ∅. Assume otherwise, then item (3∗ ) implies that ∆
contains SL(n, Z). Since SL(n, Z) is a maximal subgroup of SL(n, Z[1/p]), item (1∗ ) implies
that ∆ contains SL(n, Z[1/p]). But ∆ non-trivially intersects all the cosets of SL(n, Z[1/p]),
so ∆ = SL(n, Q), a contradiction.
7
Questions
Question 7.1. Does SL(n, Z) have a finitely generated infinite-index maximal subgroup? This
question appears in Margulis and Soifer’s paper and it is still open. Note that general finitely
generated linear groups which are not virtually-solvable might but don’t have to contain such
subgroups. For example, non-commutative free groups do not have finitely generated infinite
index maximal subgroups. On the other hand, as pointed out by Yair Glasner, the finitely
generated linear group SL(n, Z[1/p]), where p is a prime integer, contains the finitely generated
subgroup SL(n, Z) as a maximal subgroup.
Question 7.2. Margulis and Soifer proved that every finitely generated linear group which is
not virtually-solvable contains uncountably many maximal subgroups. The proof of Theorem
1.1 uses Venkatramana’s theorem which in turn is based on the congruence subgroup property.
16
Thus, our techniques do not give an answer to the following question: Does every finitely
generated linear group which is not virtually-solvable contains 2ℵ0 maximal subgroups?
Question 7.3. Does SL(n, Z) contain two infinite-index non-isomorphic maximal subgroups?
When n ≥ 4 Margulis and Soifer proved that SL(n, Z) contains an infinite-index maximal
subgroup which has a subgroup isomorphic to Z2 . It is also not hard to construct infinite-index
maximal subgroups of SL(n, Z) which have non-trivial torsion elements for n ≥ 3. Thus, a
positive answer to any one of the following three questions would provide an example of two
infinite-index non-isomorphic maximal subgroups:
Question 7.4. Does SL(n, Z) contain a torsion-free infinite-index maximal subgroup?
Question 7.5. Does SL(n, Z) contain an infinite-index maximal subgroup which does not have
a subgroup isomorphic to Z2 ?
Question 7.6. Does SL(n, Z) contains a maximal subgroup which is a free group?
Note that it is possible that all the infinite-index maximal subgroups of a given finitely
generated linear group are isomorphic. For example, this is the case for a non-commutative
free group.
Question 7.7. An element g ∈ SL(3, Z) is called complex if for every m ≥ 1 the matrix
g m has a non-real eigenvalue. Is it possible that an infinite-index Zariski-dense subgroup
of SL(3, Z) contain a complex element? The thin subgroups that arise from standard pingpong arguments on projective space do not contain complex elements. More surprisingly, the
thin surface subgroups constructed in the work of Long–Reid–Thistlethwaite [LRT11] do not
contain complex elements. A positive answer to the above question would provide the first
example (to the best of our knowledge) of a thin group whose limit set is P (In fact this group
would act minimally on pairs of distinct elements of projective space). A negative answers
would provide a very strong restriction on the structure of thin groups.
Question 7.8. Very little is known about permutation actions of SL(n, Z). In light of theorem
1.4 it is natural to ask if SL(n, Z) have 2-transitive actions? More generally, does SL(n, Z)
have highly transitive actions? In an upcoming joint paper with Glasner we prove that all
convergence groups (and in particular, linear groups of real-rank 1) have highly transitive
permutation actions. Another question of this nature is: Does SL(n, Z) have non-equivalent
primitive permutation actions?
Question 7.9. Call a subset S ⊆ SL(n, Z) fat if [SL(n, Z) : hSi] = ∞ but every Zariski-dense
subgroup of SL(n, Z) which contains S is of finite index. Venkataraman’s theorem (Theorem
2.8 above) implies that if u, v are unipotent elements, u has rank 1 and hu, vi is isomorphic to
Z2 then {v, u} is a fat subset. Question 7.7 asks if {g} is a fat subset whenever g is a complex
element. What are the fat subsets of SL(n, Z)?
17
References
[AGS14] M. Aka, T. Gelander and G.A. Soifer.Homogeneous number of free generators. J.
Group Theory 17 (2014), no. 4, 525539.
[CG00] J.-P. Conze and Y. Guivarc’h, Limit sets of groups of linear transformations. Ergodic
theory and harmonic analysis (Mumbai, 1999). Sankhya Ser. A 62 (2000), no. 3, 367–385.
[GG08] T. Gelander and Y. Glasner, Countable primitive groups. Geom. Funct. Anal. 17
(2008), no. 5, 1479–1523.
[LRT11] D.D. Long, A. Reid and M. Thistlethwaite, Zariski dense surface subgroups in
SL(3,Z). Geom. Topol. 15 (2011), no. 1, 1–9.
[Lu99] A. Lubotzky, One for almost all: generation of SL(n,p) by subsets of SL(n,Z). Algebra,
K-theory, groups, and education (New York, 1997), 125–128, Contemp. Math., 243, Amer.
Math. Soc., Providence, RI, 1999.
[MS81] G. Margulis and G. A. Soifer, Maximal subgroups of infinite index in finitely generated
linear groups. J. Algebra 69 (1981), no. 1, 1-23.
[No87] M.V. Nori, On subgroups of GLn (Fp ). Invent. Math. 88 (1987), no. 2, 257–275.
[We96] T. Weigel, On the profinite completion of arithmetic groups of split type. Lois
d’algébres et variétés algébriques (Colmar, 1991), 79–101, Travaux en Cours, 50, Hermann, Paris, (1996).
[We84] B. Weisfeiler, Strong approximation for Zariski-dense subgroups of semisimple algebraic groups. Ann. of Math. (2) 120 (1984), no. 2, 271–315.
[Ve87] T.N. Venkataramana, Zariski dense subgroups of arithmetic groups. J. Algebra 108
(1987), no. 2, 325–339.
18
| 4math.GR
|
arXiv:1711.04973v1 [math.OC] 14 Nov 2017
A Robust Variable Step Size Fractional Least Mean
Square (RVSS-FLMS) Algorithm
Shujaat Khana,b , Muhammad Usmanb , Imran Naseemc,d,∗, Roberto Tognerid,
Mohammed Bennamoune
a Department
of Bio and Brain Engineering, Korea Advanced Institute of Science and
Technology (KAIST), Daejeon, Republic of Korea.
b Faculty of Engineering Science and Technology (FEST), Iqra University, Defence View,
Karachi-75500, Pakistan.
c College of Engineering, Karachi Institute of Economics and Technology, Korangi Creek,
Karachi 75190, Pakistan.
d School of Electrical, Electronic and Computer Engineering, The University of Western
Australia, 35 Stirling Highway, Crawley, Western Australia 6009, Australia.
e School of Computer Science and Software Engineering, The University of Western
Australia, 35 Stirling Highway, Crawley, Western Australia 6009, Australia.
Abstract
In this paper, we propose an adaptive framework for the variable step size of
the fractional least mean square (FLMS) algorithm. The proposed algorithm
named the robust variable step size-FLMS (RVSS-FLMS), dynamically updates
the step size of the FLMS to achieve high convergence rate with low steady
state error. For the evaluation purpose, the problem of system identification is
considered. The experiments clearly show that the proposed approach achieves
better convergence rate compared to the FLMS and adaptive step-size modified
FLMS (AMFLMS).
Keywords: Least mean square (LMS), fractional calculus, plant identification,
channel equalization, fractional LMS (FLMS), modified fractional LMS
(MFLMS), adaptive step-size modified fractional LMS (AMFLMS), robust
variable step size FLMS (RVSS-FLMS), robust variable step size (RVSS), high
convergence, low steady state error, adaptive filter.
∗ Corresponding
author
Email addresses: [email protected] (Shujaat Khan), [email protected]
(Muhammad Usman), [email protected] (Imran Naseem),
[email protected] (Roberto Togneri), [email protected]
(Mohammed Bennamoun)
Preprint submitted to 13th IEEE Colloquium on Signal Processing & its ApplicationsNovember 15, 2017
1. Introduction
The standard integral and derivative are essential tools of calculus used by
the professionals dealing with the natural or artificial systems. The fractional
order calculus (FOC) however extends the traditional definitions which are limited to integer values. The FOC is as old as classical calculus, however, the
applications of the FOC in the areas of engineering and research have become
more popular in the last decade. New areas are emerging as both theoretical
and practical aspects of the FOC operators are now well established. Fractional
calculus has turned out to be an important tool in various fields [1], [2].
Liouville, Reimann and Leibniz, amongst others, are considered to be the
pioneers of fractional calculus [3]. In the late 19th century Heaviside made the
use of fractional derivatives in transmission lines and gave the solution for the
diffusion equation. In the recent years fractional calculus has been applied to
the problem of path tracking in autonomous vehicles [4]. In [5] the concept
of pulses propagating through porous media is proposed with an experimental
proof for a theoretical model based on fractional derivatives. Fractional calculus is also applied in the theory of viscoelasticity [6]. In the field of image
processing fractional differentiation has been used for a number of tasks such as
edge detection [7]. In the domain of signal processing FOC has found a number
of useful applications. Adaptive filtering for instance, has been the main focus
of research in this context. In [8] an FOC based least mean square algorithm
named the fractional least mean square (FLMS) is proposed. The FLMS is compared with the least mean square (LMS) algorithm for the problem of system
identification.
In the recent past, various modifications have been proposed for the conventional FLMS [9, 10, 11, 12, 13, 14, 15, 16]. In this paper a modified FLMS based
algorithm is proposed. The proposed method is named the robust variable step
size fractional least mean square (RVSS-FLMS) and is inspired by the variable
learning rate LMS (RVSS-LMS) [17]. The RVSS-FLMS algorithm dynamically
2
adapts the learning rate of FLMS using the instantaneous error energy to achieve
high convergence rate and low steady state error. The proposed algorithm is
evaluated on the problem of system identification and is compared with the conventional FLMS and adaptive step-size modified fractional least mean square
(AMFLMS) algorithm. The rest of the paper is organized as follows : In section
2 a brief literature review of modified least mean square algorithms is provided.
In section 3 the proposed RVSS-FLMS algorithm is discussed followed by the
experiments in section 4. The paper is concluded in section 5.
2. Least Mean Square
The statistics in the LMS are estimated continuously, therefore it is an adaptive filter belonging to the group of stochastic gradient methods. Extensive
research is done towards optimization of the LMS algorithm by numerous researchers [9, 18, 19, 20, 21]. LMS is applied in diversified applications such as
plant identification [8], noise cancellation [22], echo cancellation [23], ECG signal analysis [24], time series prediction [25] etc. Achieving fast convergence and
low steady state error is a challenge, various researchers have proposed different
solutions for this. One of the disadvantages of the LMS is that it is sensitive
to the scaling of its input [26]. A variant of the LMS algorithm the normalised
least mean squares filter (NLMS) solves this problem through normalization.
In [27], an optimized NLMS filter is proposed to achieve a tradeoff between the
convergence rate and tracking. In [28] q-LMS is proposed which utilizes the
q-gradient from the Jackson’s derivative so that secant of the cost function is
computed instead of the tangent. The algorithm therefore takes larger steps
towards the optimum solution and therefore achieves the higher convergence
rates. For the optimal processing of complex signals, linear modelling in complex numbers is utilized by [29]. For the complex domain adaptive filtering an
augmented complex least mean square (ACLMS) is proposed.
In [21] fractional least mean square (FLMS) is proposed for parameter estimation of input non-linear control autoregressive (INCAR) or Hammerstein
3
non-linear controlled auto regression models based on fractional signal processing approach. Unknown parameters are estimated by applying the FLMS algorithm having different step sizes. The algorithm provides better convergence
results when compared to the volterra least mean square (VLMS) and kernel
least mean square (KLMS). A modified structure of the FLMS, for the prediction of chaotic and non-stationary time series, named the modified fractional
least mean square (MFLMS) is proposed in [10]. The algorithm incorporates
the adjustable gain parameter thereby avoiding some complex calculations and
with reduced computational expense. The MFLMS is tested on stationary and
non-stationary time series with different noise levels and has shown improved
results compared to the LMS and FLMS. In [11] the adaptive weight gain parameters are incorporated by implementing a gradient-based approach on the
variable learning scheme. Implementation of this approach changes the nature
of the order of fractional derivative in MFLMS from fixed to adaptive. The
adaptive nature of the adjustable parameters may help the filters in handling
the problems with non-linear nature. They proposed adaptive step-size modified
fractional least mean square (AMFLMS) and demonstrated the improvements
by comparing their algorithm with the LMS, FLMS and MFLMS.
In this research we propose to utilize the concept of robust variable step size
(RVSS) [17] for the variable learning rate of the FLMS algorithm. The proposed
scheme is robust and computationally less expensive. For the performance evaluation we consider the problem of the system identification and compare our
results with the FLMS and AMFLMS algorithm.
3. Proposed RVSS FLMS
In [8] the weight update equation for the fractional LMS (FLMS) is given
as:
wk (n + 1) = wk (n) − ν
∂J(n)
− νf
∂wk
∂
∂wk
f
J(n)
(1)
where wk (n) is the weight of the k th tap at the nth iteration, f is the fractional power of derivative, and ν and νf are the step sizes.
4
J(n) =
1
1
e(n)2 = (d(n) − y(n))2
2
2
(2)
J(n) is the cost function defined in Eq. (2), where e(n) is the instantaneous
error between the desired output d(n) and the estimated output y(n) at the nth
iteration.
Now, the
∂J(n)
is defined as :
∂wk
∂J(n) ∂e(n) ∂y(n)
∂J(n)
=
∂wk
∂e
∂y ∂wk
(3)
f
∂
J(n), the methods of chain rule
∂wk
can be defined in various ways. The modified Rieman-Lioville based definition of
For the fractional derivative term
the fractional chain rule i.e Dxν f (g(x)) = (Dg1 f (g))g=g(x) Dxν g(x), as suggested
in [30], is one of the well known techniques. This definition has shown good
results in various applications of fractional calculus. For example, in [30] for the
fractional Taylor series of non differentiable functions, [31] for the derivation of
further fractional order derivatives, [32] for system modeling and others [33, 34,
35, 36].
Using the modified Rieman-Lioville chain rule for fractional derivatives, the
f
∂
J(n) term is defined as :
∂wk
∂
∂wk
f
J(n) =
∂J(n) ∂e(n)
∂e
∂y
∂
∂wk
f
y(n)
(4)
solving Eq. (3) results in :
∂J(n)
= −e(n)x(n)
∂wk
(5)
According to the Rieman-Lioville fractional derivative method :
1
D f (t) =
Γ (n − v)
v
d
dt
n Zt
0
5
(t − τ )(n−v−1) f (τ )dτ
(6)
Dv (t − a)α =
Γ (1 + α)
(t − a)(α−v)
Γ (1 + α − v)
(7)
where α−v +1 > 0, D is the differential operator and v is a real number defining
the fractional power of the derivative.
Using Eq. (7), Eq. (4) can be reduced to:
∂
∂wk
f
(1−f )
J(n) = −e(n)x(n)
wk
(n)
Γ (2 − f )
(8)
Using Eq. (5) and (8) the weight update Eq. (1) becomes :
wk (n + 1) = wk (n) + νe(n)x(n) + νf e(n)x(n)
wk1−f (n)
Γ (2 − f )
(9)
Note that the result in Eq. (9) is similar to the FLMS algorithm results [10, 11,
21], obtained from the alternative method of fractional derivatives to avoid the
chain rule.
For simplicity, in Eq. (9), we consider νf = νΓ (2 − f ) and this results in :
wk (n + 1) = wk (n) + νe(n)x(n) + νe(n)x(n)wk1−f (n)
(10)
wk (n + 1) = wk (n) + νe(n)x(n) 1 + wk1−f (n)
(11)
For the time varying step size ν can be replaced by ν(n)
wk (n + 1) = wk (n) + ν(n)e(n)x(n) 1 + wk1−f (n)
(12)
For the adaptation of learning rate, we proposed to use the concept of error energy correlation of RVSS-LMS algorithm [17]. The update rule for the
time varying learning rate ν(n) in the proposed robust variable step size-FLMS
(RVSS-FLMS) is defined as:
ν(n + 1) = βν(n) + γp2 (n)
6
(13)
where (0 < β < 1), (γ > 0), p(n) is the average error energy correlation and
ν(n + 1) is set to νmin or νmax
p(n) = αp(n − 1) + (1 − α)e(n)e(n − 1)
(14)
The positive constant α (0 < α < 1) is a weighting parameter that governs
the averaging time constant. It is referred to as the forgetting factor. The limits
on ν(n + 1) are given by :
ν(n + 1) =
where νmax > νmin > 0.
νmax
if ν(n + 1) > νmax
νmin if ν(n + 1) < νmin
ν(n + 1)
otherwise
(15)
4. Simulation Setup and Results
For the evaluation of the proposed method the problem of plant identification
is considered. The modeling of a linear system is an interesting problem. Sometimes it is desired to model the system as a linear filter to avoid unnecessary
complexity. The adaptive learning methods like LMS, achieves good performance in this regard [8, 18, 21, 27]. To evaluate the efficacy of the proposed
robust variable step size-FLMS (RVSS-FLMS), we consider a linear system,
shown in Fig. 1:
y(n) = a1 x(n) + a2 x(n − 1) + a3 x(n − 2) + d(n)
(16)
Eq. (16) shows the mathematical model of the system, where x(n) is the
input and y(n) is the output of the system, d(n) is the disturbance model
assumed to be N (0, σd2 ), ai ’s represent the polynomial coefficients depicting the
zeros of the system. For the said experiment, x(n) is a binary shift keying
(BPSK) modulated random sequence comprising 600 samples. In Fig. 1, the
impulse response of the system is defined by h(n) while ŷ(n) is the estimated
output, ĥ(n) is the estimated impulse response and e(n) is the estimation error.
7
Figure 1: System identification using adaptive learning algorithm.
The simulation parameters selected for the experiments are: a1 = 0.9, a2 =
0.3, a3 = −0.1. The experiments are performed on four noise levels with the
following SNR values 10 dB, 20 dB, 30 dB, and 40 dB.
For the model design, the number of taps were chosen to be 3. The weights
were initialized with a small value of 1×10−20 . For the fractional LMS (FLMS),
adaptive step-size modified fractional least mean square (AMFLMS) and RVSSFLMS the initial values of ν(n) and νf (n) for n = 0 set to 1 × 10−4 and the
fractional power f (n) for n = 0 was taken to be 0.5. For the AMFLMS, Ψk , Ψk,f
and Φk,f were initialized with a small value of 1 × 10−15 . The mixing parameter
β of AMFLMS was set to 0.5 and the adaptation rates α, αf and αf r were all
set to 1 × 10−13 . For the proposed RVSS-FLMS the values of α, β and γ were all
set to 0.5. The value of νmax and νmin are chosen to be 3 × 10−4 and 1 × 10−4
respectively.
The experiments were performed on four noise levels and the mean square
error (MSE) curves are depicted in Fig. 2. For the SNR values of 10 dB, 20 dB,
30 dB and 40 dB the RVSS-FLMS produced the best performance, achieving
an MSE of −10.22 dB, −20.26 dB, −29.70 dB and −37.68 dB in around 20,
38, 60 and 65 iterations respectively. The conventional FLMS converged to the
MSE value of −10.21 dB, −20.25 dB, −29.71 dB and −37.58 dB in around 45,
60, 80 and 90 iterations respectively. Whereas the AMFLMS converged to the
MSE value of −10.22 dB, −20.27 dB, −29.71 dB and −37.60 dB in around 80,
8
100, 120 and 140 iterations respectively. The convergence rate of the proposed
RVSS-FLMS algorithm is therefore found to be superior the FLMS and the
AMFLMS algorithms.
For the fair evaluation of the proposed algorithm, another well known performance measurement parameter called the normalized weight difference (NWD),
was used. The NWD is a comparative measure between the actual impulse
response and the identified weights. The experiments were performed on four
noise levels for which the NWD curves are depicted in Fig. 3. For the SNR
value of 10 dB, 20 dB, 30 dB and 40 dB the RVSS-FLMS produced high convergence and low steady state error, achieving an NWD value of −16.21 dB,
−20.00 dB, −25.06 dB and −28.52 dB in approximately 35, 60, 95 and 100
iterations respectively. The FLMS converges to the NWD value of −15.81 dB,
−20.09 dB, −25.06 dB and −28.91 dB in approximately 60, 85, 120 and 105
iterations respectively. Whereas the AMFLMS converges to the NWD value
of −15.67 dB, −20.12 dB, −25.00 dB and −29.02 dB in approximately 120,
175, 200 and 200 iterations respectively. The proposed algorithm has therefore
shown to comprehensively outperform the FLMS and AMFLMS algorithms.
For the purpose of time-complexity comparison of the proposed algorithm
with the FLMS and AFLMS, we investigated the training time of FLMS, AMFLMS and RVSS-FLMS for 200 iterations. The proposed method utilized 2.05
seconds whereas the FLMS and AMFLMS took 1.87 seconds and 11.06 seconds
respectively. All the experiments were conducted using Matlab on an Intel(R)
Core(TM) i5-4690 CPU @ 3.5GHz machine with 16GB memory. Although the
proposed algorithm took a few milliseconds more than the FLMS but the experiments clearly show that the proposed robust approach dynamically adapts
the learning rate to achieve the minimum steady state error in lesser number of
iterations. The summary results for all simulations is showed in Table 1.
9
0
FLMS (10dB)
AMFLMS (10dB)
RVSS−FLMS (10dB)
−5
−10
FLMS (20dB)
AMFLMS (20dB)
RVSS−FLMS (20dB)
−5
MSE (dB)
MSE (dB)
0
−10
−15
−20
−15
0
50
100
150
Iterations
−25
0
200
100
150
Iterations
200
0
FLMS (30dB)
AMFLMS (30dB)
RVSS−FLMS (30dB)
−10
−20
FLMS (40dB)
AMFLMS (40dB)
RVSS−FLMS (40dB)
−10
MSE (dB)
MSE (dB)
0
50
−20
−30
−30
0
50
100
150
Iterations
−40
0
200
50
100
150
Iterations
200
0
FLMS (10dB)
AMFLMS (10dB)
RVSS−FLMS (10dB)
−5
−10
−15
−20
0
50
100
150
Iterations
200
0
FLMS (30dB)
AMFLMS (30dB)
RVSS−FLMS (30dB)
−10
−20
−30
0
50
100
150
Iterations
200
Normalized weight difference (dB) Normalized weight difference (dB)
Normalized weight difference (dB) Normalized weight difference (dB)
Figure 2: System identification: The MSE curves for various methods.
0
FLMS (20dB)
AMFLMS (20dB)
RVSS−FLMS (20dB)
−5
−10
−15
−20
−25
0
50
100
150
Iterations
200
0
FLMS (40dB)
AMFLMS (40dB)
RVSS−FLMS (40dB)
−10
−20
−30
0
50
100
150
Iterations
200
Figure 3: System identification: The normalized weight difference curves for various methods.
10
Table 1: Results for plant identification problem
(Iterations, MSE) for SNR values
Algorithm
10 dB
20 dB
(Iterations, NWD) for SNR values
30 dB
40 dB
10 dB
Time (seconds)
20 dB
30 dB
40 dB
FLMS
(45, -10.21 dB)
(60, -20.25 dB)
(80, -29.71 dB)
(90, -37.58 dB)
(60, -15.81 dB)
(85, -20.09 dB)
(120, -25.06 dB)
(105, -28.91 dB)
1.87
AMFLMS
(80, -10.22 dB)
(100, -20.27 dB)
(120, -29.71 dB)
(140, -37.6 dB)
(120, -15.67 dB)
(175, -20.12 dB)
(200, -25.00 dB)
(200, -29.02 dB)
11.06
RVSS-FLMS
(20, -10.22 dB)
(38, -20.26 dB)
(60, -29.70 dB)
(65, -37.68 dB)
(35, -16.21 dB)
(60, -20.00 dB)
(95, -25.06 dB)
(100, -28.52 dB)
2.05
5. Conclusion
In this research an adaptive least mean square algorithm based on fractional
derivative is proposed. In particular, the step size of the fractional LMS (FLMS)
is made adaptive using the concept of robust variable step size. For this purpose
the concept of robust variable step size is utilized. The performance evaluation
of the proposed approach is done by implementing it for the problem of system
identification with different noise levels. The results of the proposed algorithm
are compared to the FLMS and adaptive step-size modified fractional least
mean square (AMFLMS) approaches. The proposed algorithm attains better
convergence rate and steady-state error and is therefore found to be superior to
the FLMS and AMFLMS algorithms.
References
References
[1] A. Loverro, Fractional calculus: history, definitions and applications for
the engineer, Rapport technique, Univeristy of Notre Dame: Department
of Aerospace and Mechanical Engineering.
[2] R. E. Gutiérrez, J. M. Rosário, J. Tenreiro Machado, Fractional order calculus: basic concepts and engineering applications, Mathematical Problems
in Engineering 2010.
[3] J. Liouville, Mémoire sur le changement de la variable indépendante, dans
le calcul des différentielles a indices quelconques, 1835.
[4] J. I. Suárez, B. M. Vinagre, A. Calderón, C. Monje, Y. Chen, Using fractional calculus for lateral and longitudinal control of autonomous vehicles,
11
in: International Conference on Computer Aided Systems Theory, Springer,
2003, pp. 337–348.
[5] Z. Fellah, C. Depollier, M. Fellah, Application of fractional calculus to the
sound waves propagation in rigid porous materials: validation via ultrasonic
measurements, Acta Acustica united with Acustica 88 (1) (2002) 34–39.
[6] E. Soczkiewicz, Application of fractional calculus in the theory of viscoelasticity, Molecular and Quantum Acoustics 23 (2002) 397–404.
[7] B. Mathieu, P. Melchior, A. Oustaloup, C. Ceyral, Fractional differentiation
for edge detection, Signal Processing 83 (11) (2003) 2421–2432.
[8] R. M. A. Zahoor, I. M. Qureshi, A modified least mean square algorithm
using fractional derivative and its application to system identification, European Journal of Scientific Research 35 (1) (2009) 14–21.
[9] Y. Tan, Z. He, B. Tian, A novel generalization of modified lms algorithm to
fractional order, Signal Processing Letters, IEEE 22 (9) (2015) 1244–1248.
[10] B. Shoaib, I. M. Qureshi, I. , S. , A modified fractional least mean square
algorithm for chaotic and nonstationary time series prediction, Chinese
Physics B 23 (3) (2014) 030502.
[11] B. Shoaib, I. M. Qureshi, S. , I. , Adaptive step-size modified fractional least
mean square algorithm for chaotic time series prediction, Chinese Physics
B 23 (5) (2014) 050503.
[12] S. Zhou, H. Zhao, W. Wang, A fraction normalized subband adaptive filter
algorithm, in: Control Conference (CCC), 2016 35th Chinese, TCCT, 2016,
pp. 3095–3098.
[13] S. Cheng, Y. Wei, Y. Chen, L. Xiaojian, Y. Wang, A novel fractional order
normalized lms algorithm with direction optimization, IFAC-PapersOnLine
49 (9) (2016) 180–185.
12
[14] N. I. Chaudhary, M. A. Z. Raja, M. S. Aslam, N. Ahmed, Novel generalization of volterra lms algorithm to fractional order with application to
system identification, Neural Computing and Applications (2016) 1–18.
[15] M. Geravanchizadeh, S. Ghalami Osgouei, Speech enhancement by modified convex combination of fractional adaptive filtering, Iranian Journal of
Electrical and Electronic Engineering 10 (4) (2014) 256–266.
[16] J. Ahmad, M. Usman, S. Khan, I. Naseem, H. J. Syed, Rvp-flms : A
robust variable power fractional lms algorithm, in: 2016 IEEE International
Conference on Control System, Computing and Engineering (ICCSCE),
IEEE, 2016.
[17] T. Aboulnasr, K. Mayyas, A robust variable step-size lms-type algorithm:
analysis and simulations, IEEE Transactions on signal processing 45 (3)
(1997) 631–639.
[18] M. A. Z. Raja, N. I. Chaudhary, Two-stage fractional least mean square
identification algorithm for parameter estimation of carma systems, Signal
Processing 107 (2015) 327–339.
[19] A. Khalili, A. Rastegarnia, S. Sanei, Quantized augmented complex leastmean square algorithm: Derivation and performance analysis, Signal Processing 121 (2016) 54–59.
[20] A. Khalili, A. Rastegarnia, W. M. Bazzi, Z. Yang, Derivation and analysis
of incremental augmented complex least mean square algorithm, Signal
Processing, IET 9 (4) (2015) 312–319.
[21] N. I. Chaudhary, M. A. Z. Raja, J. A. Khan, M. S. Aslam, Identification
of input nonlinear control autoregressive systems using fractional signal
processing approach, The Scientific World Journal 2013.
[22] J. M. Górriz, J. Ramı́rez, S. Cruces-Alvarez, C. G. Puntonet, E. W. Lang,
D. Erdogmus, A novel lms algorithm applied to adaptive noise cancellation,
IEEE Signal Processing Letters 16 (1) (2009) 34–37.
13
[23] J. Benesty, S. L. Gay, An improved pnlms algorithm, in: Acoustics, Speech,
and Signal Processing (ICASSP), 2002 IEEE International Conference on,
Vol. 2, IEEE, 2002, pp. II–1881.
[24] N. V. Thakor, Y.-S. Zhu, Applications of adaptive filtering to ecg analysis: noise cancellation and arrhythmia detection, IEEE transactions on
biomedical engineering 38 (8) (1991) 785–794.
[25] W. Liu, P. P. Pokharel, J. C. Principe, The kernel least-mean-square algorithm, IEEE Transactions on Signal Processing 56 (2) (2008) 543–554.
[26] S. C. Douglas, A family of normalized lms algorithms, IEEE signal processing letters 1 (3) (1994) 49–51.
[27] S. Ciochin, C. Paleologu, J. Benesty, An optimized {NLMS} algorithm for
system identification, Signal Processing 118 (2016) 115 – 121.
[28] U. M. Al-Saggaf, M. Moinuddin, A. Zerguine, An efficient least mean
squares algorithm based on q-gradient, in: Signals, Systems and Computers, 2014 48th Asilomar Conference on, IEEE, 2014, pp. 891–894.
[29] S. Javidi, M. Pedzisz, S. L. Goh, D. P. Mandic, The augmented complex
least mean square algorithm with application to adaptive prediction problems 1, in: Cognitive Information Processing, 2008 1st IAPR Workshop on,
EURASIP, 2008.
[30] G. Jumarie, Modified riemann-liouville derivative and fractional taylor series of nondifferentiable functions further results, Computers & Mathematics with Applications 51 (9) (2006) 1367–1376.
[31] G. Jumarie, Table of some basic fractional calculus formulae derived from
a modified riemann-liouville derivative for non-differentiable functions, Applied Mathematics Letters 22 (3) (2009) 378–385.
[32] G. Jumarie, On the derivative chain-rules in fractional calculus via fractional difference and their application to systems modelling, Open Physics
11 (6) (2013) 617–633.
14
[33] G. Jumarie, The minkowski’s space-time is consistent with differential geometry of fractional order, Physics Letters A 363 (1) (2007) 5–11.
[34] G. Jumarie, Lagrangian mechanics of fractional order, hamilton-jacobi fractional pde and taylors series of nondifferentiable functions, Chaos, Solitons
& Fractals 32 (3) (2007) 969–987.
[35] G. Jumarie, Probability calculus of fractional order and fractional taylors
series application to fokker-planck equation and information of non-random
functions, Chaos, Solitons & Fractals 40 (3) (2009) 1428–1448.
[36] G. Jumarie, An approach via fractional analysis to non-linearity induced
by coarse-graining in space, Nonlinear Analysis: Real World Applications
11 (1) (2010) 535–546.
15
| 10math.ST
|
arXiv:1711.10177v1 [cs.AI] 28 Nov 2017
Gradual Tuning: a better way of Fine Tuning the
parameters of a Deep Neural Network
Guglielmo Montone
Laboratoire Psychologie de la Perception
Université Paris Descartes, Paris
[email protected]
J.Kevin O’Regan
Laboratoire Psychologie de la Perception
Université Paris Descartes, Paris
[email protected]
Alexander V. Terekhov
Laboratoire Psychologie de la Perception
Université Paris Descartes, Paris
and
Redwood Center for Theoretical Neuroscience
University of California, Berkeley
[email protected]
Abstract
In this paper we present an alternative strategy for fine-tuning the parameters of a
network. We named the technique Gradual Tuning. Once trained on a first task, the
network is fine-tuned on a second task by modifying a progressively larger set of the
network’s parameters. We test Gradual Tuning on different transfer learning tasks,
using networks of different sizes trained with different regularization techniques.
The result shows that compared to the usual fine tuning, our approach significantly
reduces catastrophic forgetting of the initial task, while still retaining comparable
if not better performance on the new task.
1
Introduction
Due to the huge amount of data needed to train a network it is always more common in the Deep
Learning community to use Transfer Learning techniques. Such techniques mainly consist in training
a network on a new task after the same network has been trained in solving one or more tasks similar
to the new one. Transfer Learning has been proven to be effective in many domains like image
recognition, natural language processing and reinforcement learning [3, 1, 5]. Two strategies are
particularly common in Transfer Learning [6]. In one strategy, referred to as Fine Tuning, training
on a new task involves allowing modification of all the parameters of the network. The second
strategy instead consists in adding some new parameters to the network, often by adding new layers
on top of the old network. The network is then trained on the new task by modifying only the new
parameters [2]. While Fine Tuning is often the more effective of the two techniques, it has the major
drawback of suffering from Catastrophic Forgetting (CF): training on the new task considerably
reduces performance on the first task.
In this paper we propose a very simple variation of fine tuning that we called Gradual Tuning: during
the training of the network on the second task we involve the parameters of the network only gradually,
first allowing just the output layer of the network to be updated and then progressively involving
the rest of the parameters. The justification for this procedure derives from the following simple
consideration: the gradient of the error function with respect to one of the features of the network
(here the word feature refers to the vector of incoming weights of a neuron) is a function of the
feature itself but also of the way the feature is connected to the other features of the network. If the
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
network has good features at level L for solving a given task but these features are not appropriately
connected together by the upper layer, then both the features of layer L and the upper layers will have
a gradient different from zero. So good features of layer L may be changed just because the upper
layer does not combine them in the proper way.
We tested Fine Tuning and Gradual Tuning on different tasks. A first set of experiments was run on a
transfer learning task build on the MNIST dataset [4]. The rest of the experiments were involved more
challenging synthetic tasks. These tasks, all consisting in binary classification of images involving
the concepts of line and angle, were developed to test Fine Tuning and Gradual Tuning in different
conditions. In a first experiment we compared the two strategies when the transfer is between two
very similar tasks or when it is between two more different ones. In order to test how much the quality
of the transfer and the amount of CF was affected by the quality of the initial features we developed a
second group of experiments. In these experiments the network was at first trained on a group of 5
tasks, then the training on a new task was performed by using both Fine Tuning and Gradual Tuning.
The paper is organized as follows: In the next section we will more clearly describe the Gradual
Tuning procedure. In section 3 and 4 we will present all the experiments we developed and the dataset
used. Finally in 5 we will present the results of our experiments.
2
Gradual training procedure
In this paragraph we will give a precise description of the Gradual Tuning procedure that we used in
all the experiments presented in this paper. Let’s consider a network trained on a first task (Task-A)
that we wish to train on a second task, Task-B. To do so we connect the last hidden layer of the
network with some new output neurons, randomly initializing the weights of this new layer. Then we
start the training on the second task. The training with the Fine Tuning procedure simply consists in
evaluating the gradient for all the parameters of the network and updating them with the usual update
rule. In the Gradual Tuning procedure we start by allowing just the parameters of the output layer to
be updated, then, when the performance of the network stops improving, we allow the hidden layer
closest to the output to be trained as well. We continue in this way allowing the next hidden layer to
be updated each time the performance of the network stops improving. The process stops when all
the layers of the network have been modified by the training.
3
Experiments
The experiments described in this paper are all meant to examine the amount of CF in transfer learning
tasks when a network is trained with the standard Fine Tuning procedure as compared to training with
the Gradual Tuning. All the experiments were developed on FFW networks (the characteristics of the
network can be found at the end of this paragraph). The procedure of the first group of experiments is
the following. We first train the network on a first task (Task-A). Then the last hidden layer of the
trained network, other than the layer that remains connected to the output layer used for Task-A, is
additionally connected to a new output layer that will be used for training the network on a second
task (Task-B). The weights of this new output layer are randomly initialized. The network is then
trained on Task-B, while the performance of the network is evaluated both on Task-A and Task-B.
The training of the network on Task-B is done either with the usual Fine Tuning or with the Gradual
Tuning procedure. We compare the two training procedures in the following situations:
• when Task-A and Task-B are both either simple or more complex;
• when Task-A and Task-B are similar or different.
In the first experiment we did on MNIST, we split MNIST into two datasets, one containing the digits
from zero to four and one containing the digits from five to nine. Task-A and Task-B consisted both
in 5-class classification problems.
In order to have more challenging tasks for the network we build a series of synthetic binary
classification tasks. All the tasks involve the concepts of line, angle and triangle. In each task
the network receives an image as input and has to answer a question like: "is there an angle or a
triangle?". A precise description of each dataset and the relative binary classification task can be
found in paragraph 4. We develop such tasks for different reasons. The first is simply that such
tasks are more complex than MNIST, meaning that the network needs much more data to have good
performance (less than 10% misclassified). The second reason is that we wanted to have both tasks
2
a
b
c
d
e
f
g
h
Figure 1: In each pictures some samples from the used datasets. (a) ac (b) acl (c) at (d) atl (e) sbl (f)
sbt (g) sb2l (f) cnc
that were very similar to each other and tasks that were more different. In this way we would be able
to check to what extent similarity of the tasks influences CF: This was our second experiment where
a network trained on Task-A was then trained on Task-B in two different conditions, with Task-B
more or less similar to Task-A.
A third reason for having a set of similar tasks was that we wished to investigate how much CF
depends on the quality of the features the network learned during Task-A. To test this we developed
the following experiment. During the first part of the experiment, the network was trained on several
tasks (5-tasks) at the same time. This was done by connecting the last hidden layer of the network to
five different output layers. The training was executed at the same time, meaning that in each epoch
the batches of data relative to the different datasets were regularly alternated. Paragraph 5 provides
the details of the training procedure. Once the training was over, a new output layer was connected to
the last hidden layer of the network and the network was trained on a second task (Task-B). During
the training the performance of the network was evaluated for the five classification tasks of Task-A
and the classification of Task-B.
Two kinds of FFW networks were used during the experiments. The first kind has two fully connected
hidden layers of 500 units each. The second kind of network has three fully connected layers of
1000 units each. The activation function of the nodes of the hidden layers is a ReLU function
(f (x) = max(0, x)). The activation function of the output units is a softmax function. In the
following we will refer to the two architectures as the (500-500) and the (1000-1000-1000) network.
The error used to train the network was the categorical cross-entropy.
We will frequently refer to two different regularization techniques, namely L1 regularization and
Dropout. In the case of L1 regularization the sum of the absolute value of the weights of the network
were multiplied by a coefficient λ1 = 10−4 and added to the training error of the network. In the
case of Dropout, the percentage of neurons that were dropped out during the training was 20% for
the first hidden layer of the network and 50% for the other layers.
4
The tasks
The experiments consisted in training a network on classification tasks. In this section we illustrate
all the tasks we used. We split the MNIST dataset into two smaller datasets that we named MNIST-04
and MNIST-59. The MNIST-04 dataset contains the MNIST digits between 0 and 4 while the
Mnist-59 contains the digits between 5 and 9. Both MNIST-04 and MNIST-59 are composed of
approximately 30.000 gray-scale images of 28 × 28 pixels. Each of the datasets was split into three
smaller sets containing 20.000, 5.000 and 5.000 data points, respectively used as training, validation
and test sets.
The tasks we built by creating synthetic datasets are all binary classification tasks. In each task
the stimuli were gray scale images, 32 × 32 pixels in size. Each image contained two to four line
segments, each at least 13 pixels long (30% of the image diagonal). The segments were white on a
3
dark random background or black on a light random background. Examples of the 8 tasks can be
seen in figure 1):
ac: requires classifying the images into those containing an angle (between 20◦ and 160◦ ) and a pair
of crossing line segments (the crossing point must lie between 20% and 80% along each segment’s
length).
acl: the same as ac, but in each picture there is an additional distracting line crossing neither of the
other line segments.
at: distinguishes between images containing an angle (between 20◦ and 160◦ ) and a triangle (with
each angle between 20◦ and 160◦ ).
atl: the same as at, but with an additional distracting line, crossing neither of the line segments
forming the angle and the triangle.
sbl: requires classifying the images into those having blunt (between 100◦ and 160◦ ) and those
having sharp (between 20◦ and 80◦ ) angles in them. Each image also contains an additional line
segment, crossing neither of the line segments forming the angle.
sbt: requires classifying the images into those having blunt (between 100◦ and 160◦ ) and those
having sharp (between 20◦ and 80◦ ) triangles in them, in the picture there is also a distracting line
that does not intersect any of the two triangles.
sb2l: the same as sbl, but has two additional distracting lines, crossing neither of the line segments
forming the angle.
cnc: distinguishes between a pair of crossing and a pair of non-crossing lines (the crossing point must
lie between 20% and 80% of each segment length).
Each stimulus was generated by randomly selecting the parameters describing each figure (length
and orientation of straight segments, amplitude and orientation of the angles) and verifying that all
conditions were satisfied. Four different types of random backgrounds were generated with four
patterns changing at different rates. For our experiments we generated 350.000 stimuli for each
condition. Each dataset was split into three sets containing respectively 330.000, 10.000 and 10.000
data points. These sets were used as training, validation and test sets.
5
Results
We were interested in comparing the amount of CF for networks trained with the Fine Tuning
procedure versus with the Gradual Tuning procedure. In the first section (5.1) we present all the
experiments where the Task-A is a classification performed on a single dataset. We tested the two
procedures in the cases where Task-A and Task-B are rather simple, when Task-A and Task-B are
more complex but similar, and when Task-A and Task-B are more complex and rather different. In
section 5.2 we study the amount of CF for a network trained on a task for which we suppose the
network has good features. To guarantee this, Task-A was actually composed of 5 binary classification
tasks on which the network was trained at the same time. The network was then trained on Task-B.
In the first experiment Task-B is very similar to one of the tasks the network was trained on during
Task-A. In the second experiment instead the classification task of Task-B is quite different from the
classification tasks of Task-A. The results of all the experiments presented in paper are reported as
the percentage of misclassified data on the test set. Frequently in the next paragraph we will refer to
the error on the validation set, this refers to the percentage of misclassified data on the validation set.
5.1
Single task experiment
The first experiment was on MNIST. During Task-A a network was trained on MNIST-04. The
training was stopped if the error on the validation set did not go below its minimum for 20 consecutive
epochs. The percentage of misclassification on the test set was reported. The training was done using
a fixed learning rate λ = 0.01. Once the training was over, the output of the last hidden layer of the
network was connected to five extra output neurons for the classification on MNIST-59.
The output layer parameters were randomly initialized. Training on Task-B was executed with both
the Fine Tuning procedure and the Gradual Tuning procedure. In both cases the training was stopped
4
Table 1: MNIST-04 → MNIST-59 (500-500). In the column named Task-B the mean percentage
of misclassified on the test set calculated over 5 repetitions is reported together with its standard
deviation. In the column named Task-A the first error reported for each entry is the percentage of
misclassified cases on Task-A after the training on that task. The second and third value are the mean
and the standard deviation of the percentage of misclassified data on Task-A after the training on
Task-B. Column Epochs gives the mean and standard deviation of the number of epochs the network
was trained on Task-B.
Reg
No
L1
Drop
Task-B
1.56 ± 0.07
1.51 ± 0.03
1.60 ± 0.07
Fine Tuning
Task-A
1.2 → 4.5 ± 1.2
1.2 → 3.5 ± 0.6
1.3 → 8.2 ± 0.9
Epochs
91 ± 14
79 ± 21
61 ± 11
Task-B
1.54 ± 0.05
1.44 ± 0.03
1.41 ± 0.06
Gradual Tuning
Task-A
1.2 → 1.83 ± 0.05
1.2 → 2.02 ± 0.02
1.3 → 4.4 ± 1.1
Epochs
77 ± 13
95 ± 7
113±38
when the error on the validation didn’t go below its minimum in 20 epochs. Moreover in the Gradual
Tuning procedure, the training on a lower layer was allowed when the difference in the error on
the validation set between two consecutive epochs was less than 0.1%. The network was trained
on Task-B five times for each of the two training procedures, each time re-initializing the output
layer. The entire experiment, training on Task-A and then on Task-B was executed using different
regularization techniques. In particular we tried to train the network without any regularization, with
L1, and with Dropout regularization. In tables 1 and 2 the results relative to this experiment are
reported for the two network architectures. In particular in the section of the table named Fine Tuning
we report the results relative to the training with the Fine Tuning procedure, while in the part of the
table named Gradual Tuning we report the results obtained with the Gradual Training technique. The
column labeled Task-B contains the error on the test set for Task-B. The value reported in the table is
the mean error obtained over the five repetitions and its standard deviation. Each entry of the column
Task-A contains two error values. They refer to the error of the network on Task-A at the end of the
training on Task-A and at the end of the training on Task-B. This second value is actually the mean
error obtained over the five repetitions and it is reported with its standard deviation. Finally in the
column labeled Epochs, we report the mean number of epoch the network was trained on Task-B
with its standard deviation.
Looking at the two tables it is evident that the network trained with the Fine Tuning procedure tends
to forget much more than the one trained with the Gradual Tuning procedure. If we compare the
error on Task-A after training on Task-B, we notice that the percentage of misclassifications in the
case of Gradual Tuning is generally around 2 times less than the percentage obtained with the Fine
Tuning. This happens despite the fact that in many cases the network trained with the Gradual
Tuning procedure is trained for more epochs than the one trained with the Fine Tuning procedure.
At the same time the network tends to perform similarly on the new task (Task-B) independently
of the training procedure used. Finally it is also clear that the network trained with the Dropout
regularization tends to forget much more than the network trained with L1 regularization or the one
trained without regularization. In order to verify the results obtained on MNIST on more complex
Table 2: MNIST-04 → MNIST-59 (1000-1000-1000). See Table 1 for a detailed presentation of the
values reported.
Reg
No
L1
Drop
Task-B
2.01 ± 0.03
1.92 ± 0.05
1.66 ± 0.05
Fine Tuning
Task-A
1.5 → 4.3 ± 0.9
1.6 → 3.4 ± 1.3
1.3 → 12.1±0.6
Epochs
117±70
188±54
102±17
Task-B
1.99 ± 0.05
1.88 ± 0.02
1.52 ± 0.05
Gradual Tuning
Task-A
1.5 → 2.6 ± 0.1
1.6 → 2.0 ± 0.1
1.3 → 3.9 ± 0.3
Epochs
147±39
245±59
138±35
datasets we did two other experiments using the synthetic datasets. In the first experiment Task-A and
Task-B are very similar, while in the second the two tasks are more different. In the first experiment
we used as Task-A the classification of the dataset acl and as Task-B the classification on ac. These
are very similar tasks, with the first being more complex than the second because of the presence
of a distracting line in the pictures. The tests were executed on the two network architectures and
5
the results are reported in tables 3, 4. In this case the difference between the Fine Tuning and the
Table 3: acl → ac (500-500). See Table 1 for a detailed presentation of the values reported.
Reg
No
L1
Drop
Task-B
2.50 ± 0.01
2.10 ± 0.10
3.87 ± 0.18
Fine Tuning
Task-A
9.0 → 18.8±0.3
7.1 → 22.6±1.7
7.7 → 16.7±1.3
Epochs
47 ± 5
38 ± 7
115±40
Task-B
2.50 ± 0.05
2.11 ± 0.16
3.76 ± 0.12
Gradual Tuning
Task-A
9.0 → 16.1 ± 0.3
7.1 → 17.1 ± 2.4
7.7 → 17.5 ± 0.5
Epochs
92 ± 37
45 ± 23
151±17
Gradual Tuning is less evident. The network performs almost the same on Task-B, but this time the
amount of forgetting on Task-A is also nearly the same for the two procedures. The only exception is
the big network (1000-1000-1000) when trained with L1 regularization. In that case we have again
that the amount of forgetting is almost twice for Fine Tuning as compared to Gradual Tuning. The
Table 4: acl → ac (1000-1000-1000). See Table 1 for a detailed presentation of the values reported.
Reg
No
L1
Drop
Task-B
2.25 ± 0.04
1.60 ± 0.05
1.58 ± 0.10
Fine Tuning
Task-A
6.7 → 13.3±0.2
4.7 → 20.5±2.9
5.9 → 20.5±1.1
Epochs
73 ± 52
23 ± 13
113±21
Task-B
2.33 ± 0.09
1.61 ± 0.06
1.55 ± 0.18
Gradual Tuning
Task-A
6.7 → 9.5 ± 0.3
4.7 → 11.9 ± 3.2
5.9 → 20.0 ± 1.5
Epochs
47 ± 16
37 ± 15
145±46
results presented in tables 5 and 6 are relative to a network that was first trained on the dataset acl
(Task-A) and then on atl (Task-B). Task-A is the same as for the previous experiment. Task-B is
more complex when compared to Task-B of the previous experiment. Moreover in this experiment
Task-A and Task-B are much less similar than in the previous experiment. This time the amount of
forgetting seems to be different for the big network when trained without regularization or with L1
regularization (see Table 6). In these two cases the network trained with Gradual Tuning forgets
approximately half as much as the network trained with Fine Tuning. No particular difference seems
to be present in the small network.
Table 5: acl → atl (500-500). See Table 1 for a detailed presentation of the values reported.
Reg
No
L1
Drop
Task-B
4.31 ± 0.12
2.32 ± 0.33
8.21 ± 1.10
Fine Tuning
Task-A
9.1 → 27.1±0.3
7.1 → 40.2±3.7
7.7 → 38.0±2.6
Epochs
78 ± 16
77 ± 48
158±46
Task-B
4.65 ± 0.12
2.09 ± 0.28
4.29 ± 0.14
Gradual Tuning
Task-A
9.1 → 24.1 ± 0.1
7.1 → 36.6 ± 1.9
7.7 → 46.2 ± 0.2
Epochs
120±28
108±20
293 ± 4
Considering all the experiments we described so far we found that the network trained with dropout
tends to suffer more from CF than the one trained with L1 regularization, both for Fine Tuning and
Gradual Tuning procedures. Moreover we often find that performance on the second task (Task-B)
is often similar for the two regularization techniques. This may be due to the fact that the networks
were trained with a fixed learning rate, while often dropout regularization is used either with a fixed
learning rate, but bigger than the one we used, or with update of the weights with momentum. Such
solutions were not adopted here because we wanted to use the same training procedure across all the
experiments.
5.2
Multiple tasks
In this section we describe the experiments for which Task-A consisted in training the network on
several tasks at the same time. It is well known that a network trained at the same time on a set
of similar tasks develops better features for each of the tasks than a network trained on just one of
the tasks. This is in part due to the fact the network trained on several tasks is exposed to more
data and in part to the ability of the network to transfer learning from the different tasks. So we
6
Table 6: acl → atl (1000-1000-1000). See Table 1 for a detailed presentation of the values reported.
Reg
No
L1
Drop
Task-B
3.96 ± 0.09
1.68 ± 0.17
1.95 ± 0.34
Fine Tuning
Task-A
6.7 → 25.1±0.8
4.7 → 40.0±1.0
5.9 → 48.6±0.2
Epochs
120±56
56 ± 9
167±57
Task-B
5.92 ± 0.07
1.50 ± 0.02
1.28 ± 0.17
Gradual Tuning
Task-A
6.7 → 12.0 ± 0.1
4.7 → 23.4 ± 1.9
5.9 → 39.8 ± 0.6
Epochs
191±12
56 ± 6
205±57
Table 7: acl + sb2l + sbl + sbt + cnc → ac (1000-1000-1000). The results are reported following the
same schema as in table 3, but this time Task-A consisted in training the network on 5 datasets. Each
row in column Task-A refers to one of the datasets. The three numbers reported are respectively the
error on the task at the end of Task-A, the mean error and its standard deviation at the end of Task-B
over 5 repetitions.
Fine Tuning
Task-A
7.0 → 11.6 ± 0.4 (acl)
5.5 → 7.4 ± 0.2 (sb2l)
No
1.66 ± 0.06 2.4 → 3.9 ± 0.4 (sbl)
1.5 → 2.3 ± 0.1 (sbt)
0.9 → 1.2 ± 0.1 (cnc)
4.4 → 21.0 ± 1.0 (acl)
3.3 → 6.3 ± 0.5 (sb2l)
L1
0.92 ± 0.02 1.4 → 5.6 ± 0.5 (sbl)
0.7 → 2.6 ± 0.4 (sbt)
0.4 → 0.7 ± 0.1 (cnc)
4.1 → 11.5 ± 1.6 (acl)
2.4 → 6.0 ± 0.7 (sb2l)
Drop 0.73 ± 0.08 0.8 → 4.1 ± 0.7 (sbl)
0.2 → 1.0 ± 0.2 (sbt)
0.4 → 0.5 ± 0.0 (cnc)
Reg
Task-B
Epochs
42 ± 29
26 ± 5
58 ± 22
Gradual Tuning
Task-A
7.0 → 10.2 ± 0.3
5.5 → 6.6 ± 0.3
1.82 ± 0.08 2.4 → 2.8 ± 0.1
1.5 → 1.8 ± 0.1
0.9 → 1.1 ± 0.1
4.4 → 14.8 ± 0.8
3.3 → 5.4 ± 0.3
0.89 ± 0.04 1.4 → 2.1 ± 0.0
0.7 → 1.6 ± 0.1
0.4 → 0.8 ± 0.0
4.1 → 11.6 ± 1.7
2.4 → 6.1 ± 0.6
0.70 ± 0.10 0.8 → 4.1 ± 0.7
0.2 → 1.1 ± 0.2
0.4 → 0.6 ± 0.1
Task-B
Epochs
98 ± 54
36 ± 5
88 ± 23
were curious to see how much a network trained on several tasks would suffer from CF, and how
the performance would differ for the Gradual Tuning and the Fine Tuning procedures. In the two
experiments, during Task-A the network was trained on 5 datasets at the same time. The tasks we
chose were the following: acl, sb2l, sbl, sbt, cnc. The network we used for this experiment was a
network of three hidden layers, each layer having 1000 nodes. The last layer was connected to five
output layers each one having two nodes. During the training, batches of 20 data points from the
different datasets were alternated in a fixed order. At the end of each epoch the error on the validation
set was calculated for all of the five datasets. If all of the errors on the validation were below their
minimum the network configuration was stored. If any of the 5 errors did not go below its minimum
in 20 epochs the training was stopped. Once the training was over a sixth output layer was connected
to the last hidden layer and the network was trained on Task-B. The training on Task-B was repeated
5 times in the two conditions (Fine Tuning and Gradual Tuning) each time by randomly initializing
the new output layer. In the experiments Task-B always consisted of a single binary classification
task on one of the datasets described in 4.
In the first experiment we evaluated the amount of CF in the case where Task-B is a simplified
version of one of the tasks proposed in Task-A; we used as Task-B the ac task. The results are
reported in table 7. The results are presented in the table following the same schema as in the previous
table, the only difference consists in the fact that the column named Task-A now has the results
relative to the five datasets before and after the training on Task-B. We can now compare the amount
of forgetting on acl and the performance on ac obtained in this experiment with the performance
obtained in the experiment described in the previous section where the network was trained first
on acl and then on ac(see table 4). It is evident that the performance on ac is much better in this
case than before. Regarding the amount of CF that the two networks suffer, the results seem very
similar and it is not easy to say whether CF for the network that has been trained on several tasks
7
Table 8: acl + sb2l + sbl + sbt + cnc → atl (1000-1000-1000). See Table 7 for a detailed presentation
of the values reported.
Fine Tuning
Task-A
7.0 → 17.6 ± 0.7 (acl)
5.5 → 16.7 ± 0.9 (sb2l)
No
2.68 ± 0.11 2.4 → 21.8 ± 2.1 (sbl)
1.5 → 4.2 ± 0.2 (sbt)
0.9 → 7.1 ± 0.7 (cnc)
4.4 → 17.8 ± 4.0 (acl)
3.3 → 12.6 ± 1.2 (sb2l)
L1
1.29 ± 0.06 1.4 → 7.5 ± 1.3 (sbl)
0.7 → 6.5 ± 2.4 (sbt)
0.4 → 19.3 ± 4.8 (cnc)
4.1 → 12.4 ± 1.9 (acl)
2.4 → 9.5 ± 1.1 (sb2l)
Drop 0.59 ± 0.09 0.8 → 8.0 ± 0.3 (sbl)
0.2 → 3.0 ± 0.5 (sbt)
0.4 → 29.6 ± 6.2(cnc)
Reg
Task-B
Epochs
116 ± 66
17 ± 2
95 ± 40
Gradual Tuning
Task-A
7.0 → 11.8 ± 0.1
5.5 → 11.1 ± 0.1
3.46 ± 0.12 2.4 → 9.1 ± 0.1
1.5 → 2.6 ± 0.1
0.9 → 1.5 ± 0.1
4.4 → 11.3 ± 0.6
3.3 → 9.0 ± 0.2
1.14 ± 0.01 1.4 → 5.3 ± 0.4
0.7 → 1.5 ± 0.1
0.4 → 5.1 ± 0.6
4.1 → 10.1 ± 0.7
2.4 → 8.5 ± 0.5
0.52 ± 0.06 0.8 → 5.0 ± 0.4
0.2 → 2.0 ± 0.1
0.4 → 17.6 ± 2.5
Task-B
Epochs
153 ± 59
33 ± 5
138 ± 22
actually is less strong. The effect of the Gradual Tuning in this case is also quite similar to the effect
we obtained on the network trained on a single task. As before we observe that the network trained
with the Gradual Tuning procedure has very similar performances on the new task when compared
with the network trained with the Fine Tuning procedure. Also the network trained with the Gradual
Tuning procedure tends to forget less than the one trained normally, although the improvement in this
case is less evident than it was in other cases.
In this second experiment Task-B is very different from any of the classifications performed during
Task-A because we chose Task-B to be the atl task. The results are reported in table 8. This
experiment is similar to the last experiment of the previous section. In that case we trained a network
first on acl and then on atl(see table 6). The results on Task-B are much better for the network that
was trained on several tasks during Task-A, when compared to the network that was trained just on
one task 6). Concerning the amount of CF on task acl two considerations are interesting. First, this
network forgets much less than the network trained on a single task during Task-A. This suggests
that having better features helped reducing CF. Regarding the second consideration let’s compare the
results in the tables 7 and 8. These were obtained with the same Task-A. The amount of forgetting on
acl is much less in the experiment reported in table 7, where the new task is very similar to the old
one. Lastly the Gradual Tuning technique seems to be quite effective in this case, the amount of CF
being reduced in all the cases, most of the time being from 1.5 to 5 times less.
6
Discussion
The Gradual Tuning technique proved to be a good strategy for reducing the amount of CF in
a transfer learning task. The CF was smaller in almost all the experiments we described in the
paper. Regarding the performance on a new task, we can say that the networks trained with the two
procedures showed similar performance in all the experiments except in the MNIST experiment,
where the networks trained with Gradual Tuning showed better results. Although we didn’t formally
verify it, it is most probably the case that compared to Fine Tuning, Gradual Tuning tends to preserve
more of the features of the original task. For this reason we would have expected that, at least when
the two tasks are very similar, the network trained with Gradual Tuning, in addition to forgetting less
on the old task, would also outperform the network trained with Fine Tuning on the new task. Further
experiments are necessary to understand why this was not the case.
Acknowledgments
This work was funded by ERC Advanced Grant 323674 “FEEL” and ERC proof of concept grant
692765 "FeelSpeech" to Kevin O’Regan.
8
References
[1] Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep
neural networks with multitask learning. In Proceedings of the 25th international conference on
Machine learning, pages 160–167. ACM, 2008.
[2] Guglielmo Montone, J Kevin O’Regan, and Alexander V Terekhov. The usefulness of past
knowledge when learning a new task in deep neural networks. Cognitive Computation Workshop,
NIPS, 2015.
[3] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick,
Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv
preprint arXiv:1606.04671, 2016.
[4] Rupesh K Srivastava, Jonathan Masci, Sohrob Kazerounian, Faustino Gomez, and Jürgen Schmidhuber. Compete to compute. In Advances in neural information processing systems, pages
2310–2318, 2013.
[5] Alexander V Terekhov, Guglielmo Montone, and J Kevin O’Regan. Knowledge transfer in deep
block-modular neural networks. In Conference on Biomimetic and Biohybrid Systems, pages
268–279. Springer, 2015.
[6] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in
deep neural networks? CoRR, abs/1411.1792, 2014.
9
| 9cs.NE
|
arXiv:1603.04000v1 [cs.CV] 13 Mar 2016
1
Learning Typographic Style
Shumeet Baluja
[email protected]
Google, Inc.
Abstract. Typography is a ubiquitous art form that affects our understanding, perception, and trust in what we read. Thousands of different
font-faces have been created with enormous variations in the characters.
In this paper, we learn the style of a font by analyzing a small subset
of only four letters. From these four letters, we learn two tasks. The
first is a discrimination task: given the four letters and a new candidate
letter, does the new letter belong to the same font? Second, given the
four basis letters, can we generate all of the other letters with the same
characteristics as those in the basis set? We use deep neural networks to
address both tasks, quantitatively and qualitatively measure the results
in a variety of novel manners, and present a thorough investigation of
the weaknesses and strengths of the approach.
Keywords: Style Analysis, Typography, Image Generation, Learning
Introduction
The history of fonts and typography is vast, originating back at least to fifteenth century Germany with the creation of the movable type press by Johannes
Gutenberg, and the first font “Blackletter.” This was based on the handwriting
style of the time, and was used to print the first books [1,2]. Centuries later,
numerous studies have consistently shown the large impact that fonts have on
not only the readability of text, but also the comprehensibility and trustability
of what is written [3,4,5].
Despite the prevalence of just the few standard fonts used throughout academic literature, there are innumerable creative, stylized and unique fonts available. Many have been created by individual designers as hobbies, or for particular
applications such as logos, movies or print advertisements. A small sample of a
few of the over 10,000 fonts [6] used in this study are shown in Figure 1.
The seminal work of Tenenbaum and Freeman [7] towards separating style
from content was applied to letter generation. Our motivation and goals are
similar to theirs: we hope that a learner can exploit the structure in samples of
related content to extract representations necessary for modeling style. The endgoal is to perform tasks, such as synthesis and analysis, on new styles that have
never been encountered. In contrast to [7], we do not attempt to explicitly model
style and content separately; rather, through training a learning model (a deep
neural network) to reproduce style, content and style are implicitly distinguished.
2
Shumeet Baluja
Fig. 1. Fonts with varying shapes, sizes and details. As a preview, half of these were
entirely synthesized using our approach. (synthesized rows: A,B,D,E,G,J,K).
We also extend their work in four directions. First, we demonstrate that a very
small subset of characters is required to effectively learn both discriminative and
generative models for representing typographic style; we use only 4 instead of
the 54 alpha-numeric characters used previously. Second, with these 4 letters, we
learn individual-letter and combined-letter models capable of generating all the
remaining letters. Third, we broaden the results to thousands of unseen fonts
encompassing a far more expansive set of styles than seen previously. Finally,
we present novel methods for quantitatively analyzing the generated results that
are applicable to any image-creation task.
Fonts are particularly well suited for the examination of style: they provide
diversity along many axes: shape, size, stroke weight, slant, texture, and serif
details — all within a constrained environment in which the stylistic elements
can be readily distinguished from content. Additionally, unlike many image generation tasks, this task has the enormous benefit of readily available groundtruth data, thereby allowing quantitative measurement of performance. Recently,
growing attention has been devoted to style, not only in terms of fonts [8,9,10],
but in perceptual shape similarity for architecture and rigid objects [11], computer graphics [12], cursive text [13], photographs [14,15,16], artwork [17], and
music [18].
The primary goal of this paper can be succinctly stated as follows: given
a small set of letters (the basis set) of a particular font, can we generate the
remaining letters in the same style? Before we can address this question, we
need to ensure that there is enough information in a small basis set to ascertain
style. This is the subject of the next section.
Learning Typographic Style
2
3
A Discriminative Task: Same Font or Not?
By addressing the question of whether a set of basis-letters contain enough information to extract “style”, we immediately work towards identifying sources
of potential difficulties for the overarching goal of generating characters – i.e.
will the generative process have enough information to extract from the basis
letters to successfully complete the task? Further, the networks created here will
be a vital component in validating the final generated results (Section 3.3).
The discriminative task is as follows: Given a set of four basis letters1 , e.g.
B,A,S,Q, in font-A and another letter Φ, can we train a classifier to correctly
identify whether Φ is generated in font-A’s style? See Figure 2. Next, we describe
how the data is pre-processed and the neural networks used to address the task.
Fig. 2. Positive (Left) and Negative (Right) samples for the classification task. In the
left group, the fifth letter is a member of the same font. On the right, the fifth letter
is not. Note that some cases on the right are easy. However, the top two are more
challenging, requiring the slant and subtle edge weights to be analyzed, respectively.
2.1
Data Specifics
Due to the large variation in the font-faces examined, normalization of the font
images was vital, though minimal. Each letter was input to the network as a
36×36 gray-scale pixel image. In total, 5 letters were used as inputs: the first
4 were the basis letters (all from the same font), and the fifth letter was to be
categorized. The output of the network was a single binary value, indicating
whether the fifth letter belonged to the same font as the other 4.
All the fonts used in this study were True-Type-Font format (TTF), which
allows for scaling. Due to the stylistic variations present in the data, the ratio
of widths to heights of different font faces vary dramatically. The size of the
characters within the 36×36 image is set as follows. For each font, all 26 capital
letters are generated with the same point-setting until the bounding box of
any one of the 26 characters reaches the 36×36 limit in either dimension. All 26
characters are then generated with that point setting; this ensures that if the font
style specifies having a large ’R’ while having a little ’L’, that stylistic decision
is preserved in the training example and that the largest letter fits within the
1
B,A,S,Q comprised the basis set throughout this study. They were chosen because
they contained a diverse set of edges, angles and curves that can be rearranged/analyzed to reveal hints for composing many of the remaining letters.
4
Shumeet Baluja
36×36 bounding box. Note, that this sometimes creates non-standard positioning
of letters when seen together. For example, the character ’Q’ sometimes has a
descender [19] that extends below the font’s baseline; this will appear raised up
in the training examples (see Figure 2).
A training set of approximately 200,000 examples was created from 10,000
randomly chosen unique fonts. The training set was composed of 100,000 positive
examples (5th letter was the same font) and 100,000 negative examples (5th
letter was a different font). When generating the examples, each character was
randomly perturbed from the center location by up to ±2 pixels to introduce
variation (this also allowed us to create enough unique positive examples). A
disjoint testing set with 1,000 font families not present in the training set was
created in exactly the same manner. For both training and testing, negative
examples were chosen by randomly pairing different font-families2 .
2.2
Network Architectures
Building on the successes of deep neural networks for image recognition tasks (see
[20] for a recent summary with ImageNet), we explore a variety of neural network
architectures for this task. Numerous experiments were conducted to discover
effective network architectures to use. Of the over 60 architectures tried, the top
performing seven were selected. Though a full description of the experiments is
beyond the scope of this paper, a few of the general principles found are provided
here to help guide future studies.
1. Treating inputs as a single large image or as 5 individual images: We compared using the full 36×(36×5) pixel image as a single input vs. creating
individual ’towers’ for each of the 5 characters (see [21,22,23,24] for a review of column/tower architectures). As a single image, it may be easier to
capture the relative sizes of characters. As multiple images, individualized
character transformations can be created with potentially cleaner derivatives. Through numerous experiments, the results consistently favored using
individual towers, both in terms of training time and final accuracy. See
Figure 3.
2. Network Depth: In contrast to the trend of increasing network depth, deeper
nets did not lead to improved performance. Unlike in general object detection tasks in which deep neural networks are able to exploit the property
that images can be thoughts of as compositional hierarchies [25], such deep
hierarchies may not be present for this task. For this study, two sets of convolution layers were used with fully connected layers prior to the final output
(more details in Section 2.3). Additional depth did not increase performance.
2
To create negative examples, two fonts from the same font-family, i.e. Times-RomanBold and Times-Roman were never used in the same negative example. Because the
hierarchy of fonts is not apriori known, font families were estimated by a simple
lexicographic analysis of the font names. Importantly, it should be noted that this did
not preclude fonts that appear visually nearly indistinguishable (but with dissimilar
names) from being used as negative examples.
Learning Typographic Style
5
Fig. 3. Left & Middle: Two tower-based network architectures. Left: Layers within
the towers are fully connected. Middle: Layers within the towers are convolutions.
Right: Fully connected net that treated the 5 character inputs as a single image (not
shown: experiments with single-image-input networks that employ convolutions were
also conducted). The figure also shows three different fonts used in training all the
networks.
3. RELU or Logistic? A variety of activation functions were attempted, including exclusively logistic activations throughout the network (recall that
the networks were not as deep as is often used in image-recognition tasks,
a condition often cited for using Rectified Linear (RELU) activations [26]).
Except for the final logistic-output unit, the rest of the activations throughout the network worked best as RELU. Though the performance difference
was not consistent across trials, RELU units trained faster.
4. Convolutions vs. Fully Connected Layers: Convolution layers are frequently
employed for two purposes: achieving translation invariance and free parameter reduction, especially when used in conjunction with pooling layers. For
this task, translation invariance is not as crucial as in general object detection tasks, as the object of interest (the character) is centered. Further,
this task worked well across a number of network sizes and free-parameter
ranges. Networks that employed convolutions as well as those that used fully
connected layers exclusively performed equally well.
2.3
Individual and Ensembles Network Results
As mentioned in the previous section, over 60 architectures were experimented
with for this task. Gradient descent with momentum (SGD+Momentum) was
used to train the networks and no domain-specific prenormalization of the weights
was necessary. From the 60 architectures, seven were chosen (see Table 1). All
are based on the tower architecture (Figure 3). The performance on the discrimination task was measured on an independent test set — the 4-basis letters
and Φ were drawn from fonts that were not used for training. The results were
consistent across a wide variety of free parameters; in the seven networks, the
number of parameters varied by a factor of 18×, with similar performance.
It is illuminating to visualize the types of errors that the networks made. Figure 4 provides examples that were both correctly (Left) and incorrectly (Right)
6
Shumeet Baluja
Table 1. Seven top performing architectures found from the 60 networks tried.
network
1
2
3
4
5
6
7
Voting Ensemble
Description of
hidden layers
in each tower
Fully Connected
Total free
hidden units in
performance
parameters
aggregation layers
1 Fully Connected (50 units)
2 (250,200)
3 Fully Connected (50,50,50 units)
2 (250,200)
3 Fully Connected (100,100,100 units)
2 (250, 200)
3 Fully Connected (100,100,100 units)
2 (50, 50)
2 Conv Paths, 2 Deep
2 (50, 50)
(3x3 → 3x3, 4x4 → 3x3)
2 Conv Paths, 2 Deep
4 (50, 50, 50, 50)
(3x3 → 3x3, 4x4 → 3x3)
2 Conv Paths, 2 Deep
4 (10, 10, 10, 10)
(3x3 → 3x3, 4x4 → 3x3)
216,952
90.3%
89.2%
88.2%
89.3%
90.7%
222,052
91.1%
437,602
463,102
925,352
777,202
52,612
90.0%
92.1%
classified by all 7 networks. In the leftmost column, examples in which the 5th
character was correctly recognized as belonging to the same font are shown. In
the second column, correctly recognized negative examples (the 5th letter was a
different font) are given. Note that the 6th letter, the “ground-truth”, was not
used as input to the network; it is provided here to show the actual letter of the
same font. In the second column, note the similarity of the correctly discerned
letters from the ground-truth. In particular, in this column, row 1 had very few
distinguishing marks between the proposed and real ’H’; in row 4, the ’U’ was
recognized as not being a member of the same font, as was the ’V’ in row 7 –
based solely on the weight of the strokes.
Fig. 4. Same Font or Not? 10 binary classification results. In columns 2 & 4, which show
examples in which the 5th character is taken from a different font than the basis set, the
corresponding actual letter from the basis set is also given. This sixth character, “the
ground-truth,” is not used in training/testing; it is only shown here for comparison.
The third column shows the false-negatives. Several of the mistakes are readily explained: rows 5,6,7 are non-alphabetic fonts (no pre-filtering was done in
the font selection process). Rows 2 and 9 are explicitly designed to have a diverse
set of characters, e.g. for creating ’Ransom Note’ like artistic effects. Also shown
Learning Typographic Style
7
are color-inverted fonts (row 4); the likely cause of the mistakes was the sparsity
of such fonts in the training set.
Finally, in the last column, false-positives are shown – different fonts that are
mistakenly recognized as the same. Three types of errors were seen. When the
fonts have extremely thin strokes (row 2), the failure rate increases. This is due
to two factors: under-representation in the training set and less information in
the pixel maps from which to make correct discriminations. Second, as before,
non-alphabetic fonts appear. The third source of mistakes is that many fonts
have characters that appear similar or match the basis font’s style (rows 3,7,8).
It is worth emphasizing here that the font design process is not an algorithmic
one; the artist/designer can create fonts with as much, or as little, variation as
desired. There may be multiple acceptable variations for each character.
Despite the similar performance of the 7 networks, enough variation in the
outputs exists to use them as an ensemble [27]. In 76.7% of examples, all networks
were correct. In 8.2%, only 1 network was mistaken, in 4.1%, 2/7 were mistaken
and in 3.0%, 3/7 were mistaken. Therefore, by using a simple majority voting
scheme, the seven networks can be employed as an ensemble that yields 92.1%
accuracy. This ensemble will be used throughout the remainder of the paper.
3
A Generative Task: Creating Style-Specific Characters
In the previous section, we determined that there was sufficient information in
the 4-Basis letters to correctly determine whether a fifth letter was a member of
the same font. In this section, we attempt to construct networks that generate
characters. The experiments are broadly divided into two approaches: single
and multiple-letter generation. In the first approach, a network is trained to
generate only a single letter. In the multiple-letter generation networks, all letters
are generated simultaneously by the same network. Though the single letter
networks are conceptually simpler, training a network to generate multiple letters
simultaneously may allow the hidden units to share useful representations and
features – i.e. serif style, line width, angles, etc. This is a form of transfer learning
with a strong basis in multi-task learning [28,29,30,31].
3.1
Single Letter Generation
As in the previous section, experiments with numerous network architectures
were conducted. The architectures varied in the number of layers (2-10), units
(100s-1000s) and connectivity patterns. The final architecture used is shown in
see Figure 6(Left). In the discrimination task described in the previous section,
there was a single output node. This allowed the use of large penultimate hidden
layers since the number of connections to the final output remained small. In
contrast, in image generation tasks, the number of output nodes is the number
of pixels desired; in this case it is the size of a full input character (36×36). This
drastically increases the number of connections in the network. To alleviate this
growth in the number of connections, reverse-retinal connections were used in
8
Shumeet Baluja
which each hidden unit in the penultimate layer connects to a small patch in
the output layer. Various patch sizes were tried; patches of both sizes 3 and 4
were finally employed. Unlike other convolution and de-convolution networks, the
connection weights are not shared. Sharing was not necessary since translation
invariance is not needed in the generated image; each letter should be generated
in the center of the output. Similar architectures have been used for superresolution and image deconvolution, often with weight sharing [32,33].
The letter ’R’ was chosen as the first test case, since, for many fonts, the
constituent building blocks for ’R’ are present in the basis letters (e.g. copy the
’P’ shape from the ’B’-basis letter and combine it with the bottom right leg of
the ’A’). As in the previous experiments, the network was trained with the same
set of 9,000 unique fonts with L2 loss in pixel-space, using SGD+Momentum.
Numerous experiments, both with and without batch normalization [34], were
conducted – no consistent difference in final error was observed. Results for fonts
from the test set are shown in Figure 5.
Inputs
Gen.
SSE
shape
serif
accept?
comments by rater
1
1574
Yes
Yes
Yes
2
1527
Yes
Yes
No
3
1358
Yes
Yes
Yes
4
1996
No
Yes
Yes
5
1728
Yes
Yes
Yes
6
1955
No
Yes
Yes
7
1399
Yes
Yes
Yes
8
2258
Yes
No
No
Serifs missing
9
3290
Yes
Yes
No
Similar, but not quite
10
1355
Yes
Yes
Yes
11
1359
Yes
Yes
Yes
12
2593
Yes
Yes
Yes
13
2428
Yes
Yes
Yes
14
2144
No
No
No
15
1934
Yes
Yes
Yes
16
1512
Yes
Yes
Yes
Actual
Missing divider
Top of R should not be closed
Not Same
Shape very close, maybe tilted?
Fig. 5. 16 Results shown for ’R’ generation. Results are judged by an external human
evaluator. Actual font characters shown on the right. 12/16 were deemed as acceptable
replacements for the original.
Learning Typographic Style
9
Unfortunately, simply measuring the pixel-wise difference to the actual target
letter (SSE) does not correlate well to subjective evaluations of performance.
Yet, as in all real-world image generation tasks, aesthetic judgment is vital. We
asked a human rater3 to ascertain the quality of the result along 3 axis: (1) is
the shape generally correct, (2) are the serif appropriately captured, (3) is this
an acceptable replacement for the actual letter? The last metric is the hardest
to measure, but perhaps the most relevant. The outcome was overall positive,
with the caveats noted in Figure 5.
3.2
Simultaneous Letter Generation - Multi-Task Learning
In this section, we explore the task of generating all the upper-case letters simultaneously. Unlike the single-letter generation process described in the previous
section, the hidden units of the network can share extracted features. This is
particularly useful in the character generation process, as many letters have
common components (e.g. ’P’ & ’R’ , ’O’ & ’Q’, ’T’ & ’I’). See Figure 6(Right).
Fig. 6. Single Vs. Multi-Letter Generation. The left networks show individual letter
creation (only 2 of 22 shown). The right network generates all letters simultaneously,
thereby allowing the hidden units to share extracted information. The tower/column
architectures are employed to transform the basis letters (as was done for the discrimination task). The generative network also auto-encodes the basis letters (Right).
Once again, numerous architectures were explored (over 30 in total). Comparing the best single-letter-generation networks with the best multi-letter-generation
networks, the SSE error was repeatedly reduced (by 5%-6%) by using the multiletter generation networks. Qualitatively, however, the characters generated by
both networks appeared similar. Nonetheless, because of the small SSE error
improvement coupled with the ease of deployment, the multi-letter-generation
network are used going forward. A large set of results are shown in Figure 7.
Subjectively evaluating these results would be error prone and yield inconsistent
results given the difficulty in evaluating the individual consistency of a large set
3
The subjective evaluation was conducted by an independent User Experience Researcher (UER) volunteer not affiliated with this project. The UER was given a
paper copy of the input letters, the generated letters, and the actual letter. The
UER was asked to evaluate the ’R’ along the 3 dimensions listed above. Additionally, for control, the UER was also given examples (not shown here) which included
real ’R’s in order to minimize bias. The UER was not paid for this experiment.
10
Shumeet Baluja
Fig. 7. Raw Output of the multi-letter-generation networks, with no postprocessing. 18 examples shown. Top good results. Bottom, results where some of
the letters are smudged. Non-alphabet/picture fonts not shown.
Learning Typographic Style
11
of characters with each other and with respect to the basis set. Instead, in the
next section, we describe novel methods for validating the results.
3.3
Validating Results
In the previous section, the results of the font generation network were measured
by the network’s SSE error and subsequently by an external human evaluator.
Here, we present alternate methods based on using the discriminative networks
developed in Section 2. The discriminative networks were used to determine
whether an input character Φ was the same font as the basis characters. Recall
that by employing a voting-ensemble of 7 networks, 92.1% accuracy was attained.
Here, to test the generative network G on font f , the results of G(f ) are
passed into the discriminative network ensemble (D) along with the original
BASQ-basis letters, D(Bf , Af , Sf , Qf , G(f )). Consequently, the output of D is
used to determine whether the generated letter is the same font as the basis
letters. Novel variants of this method, using pairs of generator/discriminator
networks have been independently proposed in [35], which is based on the architectures of generative-adversarial models [36,37].
In the most straightforward formulation, we test whether the letters generated by G appear the same as the original font. See Table 2 – column Original
Basis, Synthetic Test. Formally, the test is:
For each font , f , in the test set :
generate all characters :
G(Bf , Af , Sf , Qf ) → [Af 0 , Bf 0 , Cf 0 , Df 0 , ..., Zf 0 ]
test each GENERATED character :
D(Bf , Af , Sf , Qf , Af 0 ) → [Same/Dif f erent]
...
D(Bf , Af , Sf , Qf , Zf 0 ) → [Same/Dif f erent]
Recall that the discriminative network ensemble, D, was trained with tuples
of letters {BasisSet, Φ} drawn from the training set of fonts and tested on an
entirely separate set of fonts. Thus, we are not limited in the set of fonts that we
can use as the basis-set. In the next examination, we reverse the question asked
above. We generate the letters as before, but ask the question, if the generated
letters are used as the basis set, will the original letters be recognized as the
same font? See Table 2 – column Synthetic Basis, Original Test. The generation
portion remains the same, but the test is revised to:
test each ORIGINAL character :
D(Bf 0 , Af 0 , Sf 0 , Qf 0 , Af ) → [Same/Dif f erent]
...
D(Bf 0 , Af 0 , Sf 0 , Qf 0 , Zf ) → [Same/Dif f erent]
12
Shumeet Baluja
In the previous two experiments, we compared the generated characters to
the original font. This gives a measure of how closely the generated characters
resemble the original font. In the final experiment, we ask how consistent the
generated letters are with each other. If we are given the basis letters from the
generated font, will the other generated letters from the same font be classified
as the same? See Table 2 – column Synthetic Basis, Synthetic Test.
test GENERATED characters against GENERATED basis :
D(Bf 0 , Af 0 , Sf 0 , Qf 0 , Af 0 ) → [Same/Dif f erent]
...
D(Bf 0 , Af 0 , Sf 0 , Qf 0 , Zf 0 ) → [Same/Dif f erent]
Looking at Table 2, from the column Original Basis, Synthetic Test, it can be
seen that the generated fonts appear similar to the original fonts on which they
are based. An interesting question arises: why, then, does the second test Synthetic Basis, Original Test have a lower recognition rate? The answer lies in the
fact that the font designer may introduce non-systematic variations that add visual interest, but may not adhere to strict stylistic cues present in other letters.
In the most obvious cases, the ’ransom note fonts’, each character is intentionally not representative of other characters. Other themed fonts introduce objects
such as trees, animals, and vehicles into the design that reflect each artist’s individual creativity. For example, some fonts replace the empty spaces in the letter
’B’ with mini-airplanes, but may not do so in the letter ’A’, etc. The generated
fonts, when they capture such shapes, will synthesize more consistent characters
that reuse many of the same elements. It is likely that the original font’s glyphs
may appear outside the more cohesive set generated by the networks.4
The third column, Synthetic Basis, Synthetic Test, shows that the characters
generated are extremely consistent with each other. Normally, this would be considered a good attribute; however, when viewed in terms of the baseline (Column
4: Original Basis, Original Test), it raises an important question. Why are the
generated fonts more homogeneous than the original fonts? Is it for the reason
mentioned above, or is it for the, potentially more troubling, reason that all the
generated characters (across all fonts) are too much alike? Have all the characters
“regressed to the mean?” As a final test, we examine this possibility explicitly
using only network generated fonts. First, we set the basis letters (BASQ) from
a randomly selected generated font. Second, we randomly select another generated font and character as the candidate test character. 50,000 randomly paired
samples are created. Unlike the previous tests, in which accuracy was measured
by how many matches were found, the accuracy is measured by how many nonmatches are detected. This explicitly tests whether the generated fonts look too
similar to each other. Running this test with the discriminative network ensemble, D, yields a different-font detection rate of 90.7%. For a baseline, we repeat
4
For completeness, we also analyzed the ’R’s generated by the one-letter-at-a-time
networks. They had similar performance (when measured with D) to the ’R’ row
shown in Table 2, with (6%) higher SSE.
Learning Typographic Style
13
Table 2. Performance by Letter: (SSE & Discriminative Network Ensemble
letter
SSE
C
D
E
F
G
H
I
J
K
L
M
N
O
P
R
T
U
V
W
X
Y
Z
1851.0
1987.7
2006.9
1943.0
2096.7
2066.6
1495.3
1968.8
2146.3
1799.7
2697.1
2162.8
1828.5
1936.3
2135.6
1714.8
1910.5
1950.6
2611.0
1985.5
1998.3
1901.4
Discriminative Network Ensemble Based Evaluation
Baseline:
Original Basis Synthetic Basis Synthetic Basis
Original Basis
Synthetic Test Original Test Synthetic Test
Original Test
Average
99%
99%
99%
98%
99%
99%
94%
92%
99%
99%
97%
98%
98%
99%
99%
97%
99%
97%
97%
99%
98%
97%
90%
90%
89%
84%
89%
89%
83%
84%
88%
88%
83%
87%
89%
89%
90%
85%
89%
85%
82%
86%
85%
84%
100%
100%
100%
100%
100%
100%
99%
97%
100%
100%
99%
99%
99%
100%
100%
99%
99%
99%
99%
100%
99%
99%
96%
97%
95%
93%
95%
96%
90%
91%
95%
94%
92%
94%
95%
95%
97%
94%
96%
93%
92%
94%
92%
91%
97%
86%
98%
94%
this experiment with 50,000 pairs generated from the original (non-synthesized)
fonts. D yields a correct different-font detection rate of 90.1%. The very close results indicate that the generated fonts have not regressed to the mean, in terms
of recognizable serifs and style, and remain as distinguishable as the original
fonts. The difference in performance between column 3 and 4 in Table 2 is likely
due to the variability introduced by smaller artistic variances inserted by the
designers, as described in the paragraph above.
4
Future Work
Beyond the straightforward explorations of varying the number and selection of
basis letters and also generating lower-case letters, there are many conceptually
interesting avenues for future work. First, an alternative approach to using the
generator networks is to use discriminative-networks and propagate derivatives
back to the inputs to modify the input pixels to maximize the similarity of
the hidden states to specified values. This has recently been proposed to create
natural images as well as dream like images “Deep Dreams” [38,17].
Second, we made a concerted effort to use the simplest networks possible.
An extensive empirical search through the space of networks and learning algorithms was conducted to find the most straightforward approach to addressing
this task. Recent preprints [9,35] describe concurrent explorations of similar and
related problems with much more complex architectures that yield comparably
promising results. Extending this work to other architectures is easily done; for
14
Shumeet Baluja
example, the evaluation mechanisms used in this study are akin to Generative
Adversarial Nets (GAN) [36,39,37] in which a synthetic-vs.-real distinguisher
network and a font generator are trained to outperform each other. This study
provides a strong, and easily implemented, baseline to which new architectures
and learning approaches can be compared.
Third, the trained networks can be used in novel manners beyond synthesizing characters from existing fonts. Will it be possible to take the attributes of
multiple fonts and combine them? For example, if the the basis set is composed
of 2 characters from font-A and 2 characters from font-B, will the resulting characters be a combination of the two? Preliminary evidence (see Figure 8) suggests
that it may be possible; though perhaps more than 2 examples from each font
will be necessary for artistically more innovative combinations to be produced.
Fig. 8. Novel combinations of two fonts. In each of the two results, the first row shows
the 4 input letters and characters generated from Gf ontA , second row for Gf ontB . In
the Row 3, two characters from each font are used as input and the resulting characters
shown. Top Example: Note that the weight (thickness of lines) from the first font are
combined with the size of the second font. Bottom Example: Note that the “hollow”
look of the first font is combined with the weight and shapes of the second font.
Fourth, we have taken an image based approach to font generation. An alternate, more speculative direction, is to use a non-image based approach with
recurrent networks, such as Long-Short-Term-Memory (LSTMs). Can LSTMs be
used to generate characters directly in TrueType language? For this task, network’s inputs would include the TTF of the basis letters. Similar programmatic
learning in which LSTMs compose simple programs and sequences of music have
been recently attempted [40,41,42]. If it is possible to generate results directly
in TTF encodings, this will produce another, orthogonal, result to complement
this study. The feasibility of this approach is open for future research.
5
Conclusions
In this paper, we have presented a learning method to analyze typographic style
based on a small set of letters and to employ the learned models to both distinguish fonts and produce characters in the same style. The results are quite
promising, but fully capturing artistic visual intent is just beginning. Many of
the overall shapes, weights, angles and serifs were successfully modeled. In the
Learning Typographic Style
15
future, as learning progresses, some of the more individualized nuances and repeated intricate design patterns, unique to individual fonts, will be captured.
6
Acknowledgments
Fig. 9. Acknowledgments displaying 24 synthesized fonts (from test set only).
References
1. Tschichold, J.: Treasury of Alphabets and Lettering: A Source Book of the Best
Letter Forms of Past and Present for Sign Painters, Graphic Artists, Commercial
Artists, Typographers, Printers, Sculptors, Architects, and Schools of Art and
Design. A Norton professional book. Norton (1995)
2. Hutchings, E.:
Typeface timeline shows us the history of fonts.
http://www.psfk.com/2012/04/history-of-fonts.html (2014)
3. Beymer, D., Russell, D., Orton, P.: An eye tracking study of how font size and type
influence online reading. In: Proceedings of the 22nd British HCI Group Annual
Conference on People and Computers: Culture, Creativity, Interaction-Volume 2,
British Computer Society (2008) 15–18
4. Ha, S.: The influence of design factors on trust in a bank’s website. Digital
Repository@ Iowa State University (2009)
5. Li, Y.M., Yeh, Y.S.: Increasing trust in mobile commerce through design aesthetics.
Computers in Human Behavior 26(4) (2010) 673–684
6. 10000Fonts.com: Download. http://www.10000fonts.com/catalog/ (2016)
7. Tenenbaum, J.B., Freeman, W.T.: Separating style and content with bilinear models. Neural computation 12(6) (2000) 1247–1283
8. Bowey, M.: A fontastic voyage: Generative fonts with adversarial networks.
http://multithreaded.stitchfix.com/blog/2016/02/02/a-fontastic-voyage (2016)
9. Upchurch, P., Snavely, N., Bala, K.: From A to Z: Supervised Transfer of Style and
Content Using Deep Neural Network Generators. ArXiv e-prints (March 2016)
16
Shumeet Baluja
10. Bernhardsson, E.: Analyzing 50k fonts using deep neural networks (2016)
11. Lun, Z., Kalogerakis, E., Sheffer, A.: Elements of style: learning perceptual shape
style similarity. ACM Transactions on Graphics (TOG) 34(4) (2015) 84
12. Willats, J., Durand, F.: Defining pictorial style: Lessons from linguistics and computer graphics. Axiomathes 15(3) (2005) 319–351
13. Graves, A.: Generating sequences with recurrent neural networks. CoRR
abs/1308.0850 (2013)
14. Karayev, S., Hertzmann, A., Winnemoeller, H., Agarwala, A., Darrell, T.: Recognizing image style. CoRR abs/1311.3715 (2013)
15. Gygli, M., Grabner, H., Riemenschneider, H., Nater, F., Gool, L.: The interestingness of images. In: Proceedings of the IEEE International Conference on Computer
Vision. (2013) 1633–1640
16. Khosla, A., Das Sarma, A., Hamid, R.: What makes an image popular? In:
Proceedings of the 23rd international conference on World wide web, ACM (2014)
867–876
17. Gatys, L.A., Ecker, A.S., Bethge, M.: A neural algorithm of artistic style. CoRR
abs/1508.06576 (2015)
18. Aucouturier, J.J., Pachet, F.: Representing musical genre: A state of the art.
Journal of New Music Research 32(1) (2003) 83–93
19. Bowey,
M.:
A
20
minute
intro
to
typography
basics.
http://design.tutsplus.com/articles/a-20-minute-intro-to-typography-basics–
psd-3326 (2009)
20. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z.,
Karpathy, A., Khosla, A., Bernstein, M.S., Berg, A.C., Li, F.: Imagenet large scale
visual recognition challenge. CoRR abs/1409.0575 (2014)
21. Karpathy, A., Toderici, G., Shetty, S., Leung, T., Sukthankar, R., Fei-Fei, L.: Largescale video classification with convolutional neural networks. In: Proceedings of the
IEEE conference on Computer Vision and Pattern Recognition. (2014) 1725–1732
22. Ciresan, D., Meier, U., Schmidhuber, J.: Multi-column deep neural networks for
image classification. In: Computer Vision and Pattern Recognition (CVPR), 2012
IEEE Conference on, IEEE (2012) 3642–3649
23. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D.,
Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition. (2015) 1–9
24. Wang, M.: Multi-path convolutional neural networks for complex image classification. CoRR abs/1506.04701 (2015)
25. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. Nature 521(7553) (2015) 436–
444
26. Zeiler, M.D., Ranzato, M., Monga, R., Mao, M., Yang, K., Le, Q.V., Nguyen, P.,
Senior, A., Vanhoucke, V., Dean, J., et al.: On rectified linear units for speech
processing. In: Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE
International Conference on, IEEE (2013) 3517–3521
27. Schapire, R.E., Freund, Y., Bartlett, P., Lee, W.S.: Boosting the margin: A new
explanation for the effectiveness of voting methods. Annals of statistics (1998)
1651–1686
28. Bengio, Y.: Deep learning of representations for unsupervised and transfer learning.
Unsupervised and Transfer Learning Challenges in Machine Learning 7 (2012) 19
29. Huang, J.T., Li, J., Yu, D., Deng, L., Gong, Y.: Cross-language knowledge transfer
using multilingual deep neural network with shared hidden layers. In: Acoustics,
Speech and Signal Processing (ICASSP), 2013 IEEE International Conference on,
IEEE (2013) 7304–7308
Learning Typographic Style
17
30. Oquab, M., Bottou, L., Laptev, I., Sivic, J.: Learning and transferring mid-level
image representations using convolutional neural networks. In: Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition. (2014) 1717–1724
31. Caruana, R.: Multitask learning. Machine learning 28(1) (1997) 41–75
32. Xu, L., Ren, J.S., Liu, C., Jia, J.: Deep convolutional neural network for image
deconvolution. In: Advances in Neural Information Processing Systems. (2014)
1790–1798
33. Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network
for image super-resolution. In: Computer Vision–ECCV 2014. Springer (2014)
184–199
34. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by
reducing internal covariate shift. arXiv preprint arXiv:1502.03167 (2015)
35. Jiwoong Im, D., Dongjoo Kim, C., Jiang, H., Memisevic, R.: Generating images
with recurrent adversarial networks. ArXiv e-prints (February 2016)
36. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S.,
Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in Neural
Information Processing Systems. (2014) 2672–2680
37. Dosovitskiy, A., Brox, T.: Generating Images with Perceptual Similarity Metrics
based on Deep Networks. ArXiv e-prints (February 2016)
38. Mordvintsev, A., Olah, C., Tyka, M.: Inceptionism: Going deeper into neural networks. http://googleresearch.blogspot.com/2015/06/inceptionism-going-deeperinto-neural.html (2015)
39. Denton, E.L., Chintala, S., szlam, a., Fergus, R.: Deep generative image models
using a laplacian pyramid of adversarial networks. In Cortes, C., Lawrence, N., Lee,
D., Sugiyama, M., Garnett, R., eds.: Advances in Neural Information Processing
Systems 28. Curran Associates, Inc. (2015) 1486–1494
40. Zaremba, W., Sutskever, I.: Learning to execute. arXiv preprint arXiv:1410.4615
(2014)
41. Eck, D., Schmidhuber, J.: A first look at music composition using lstm recurrent
neural networks. Istituto Dalle Molle Di Studi Sull Intelligenza Artificiale 103
(2002)
42. Graves, A., Wayne, G., Danihelka, I.: Neural turing machines. arXiv preprint
arXiv:1410.5401 (2014)
| 9cs.NE
|
REGULAR SEQUENCES AND LOCAL COHOMOLOGY MODULES
WITH RESPECT TO A PAIR OF IDEALS
arXiv:1305.0429v1 [math.AC] 2 May 2013
SH. PAYROVI AND M. LOTFI PARSA
A BSTRACT. Let R be a Noetherian ring, I and J two ideals of R and t an integer. Let S be the class of Artinian R-modules, or the class of all R-modules N
i (M) ∈
with dimR N ≤ k, where k is an integer. It is proved that inf{i : HI,J
/ S} =
e J)}, where M is a finitely generated R-module, or
inf{S − deptha (M) : a ∈ W(I,
e J). Let SuppR H i (M) be
is a ZD-module such that M/aM ∈
/ S for all a ∈ W(I,
I,J
a finite subset of Max(R) for all i < t. It is shown that there are maximal ideals
i (M) ∼ H i (M) ⊕ H i (M) ⊕ · · · ⊕ H i (M) for
m1 , m2 , . . . , mk of R such that HI,J
= m1
mk
m2
all i < t.
1. I NTRODUCTION
Throughout this paper, R is a commutative Noetherian ring with non-zero identity, I and J are two ideals of R, M is an R-module and s and t are two integers. For
notations and terminologies not given in this paper, the reader is referred to [4], [5]
and [16] if necessary.
The theory of local cohomology, which was introduced by Grothendieck [10], is
a useful tool for attacking problems in commutative algebra and algebraic geometry. Bijan-Zadeh [3] introduced the local cohomology modules with respect to a
system of ideals, which is a generalization of ordinary local cohomology modules.
As a special case of these extend modules, Takahashi, Yoshino and Yoshizawa [16]
defined the local cohomology modules with respect to a pair of ideals. To be more
precise, let W(I, J) = {p ∈ Spec(R) : I t ⊆ J + p for some positive integer t}. The
set of elements x of M such that SuppR Rx ⊆ W(I, J), is said to be (I, J)-torsion
submodule of M and is denoted by ΓI,J (M). ΓI,J (−) is a covariant, R-linear functor
from the category of R-modules to itself. For an integer i, the local cohomology
i (−) with respect to (I, J), is defined to be the i-th right derived functor
functor HI,J
i (M) is called the i-th local cohomology module of M with reof ΓI,J (−). Also HI,J
i (−) coincides with the ordinary local cohomology
spect to (I, J). If J = 0, then HI,J
2010 Mathematics Subject Classification. 13C15, 13D07, 13D45.
Key words and Phrases. Local cohomology, Regular sequence, Serre subcategory, ZD-module.
1
2
SH. PAYROVI AND M. LOTFI PARSA
e J) = {a E R : I t ⊆ J + a for some positive integer t}. It is
functor HIi (−). Let W(I,
easy to see that
e J), ax = 0} =
ΓI,J (M) = {x ∈ M : ∃ a ∈ W(I,
[
(0 :M a).
e
a∈W(I,J)
In section 2, we study extension functors of local cohomology modules with
respect to a pair of ideals. Let S be a Melkersson subcategory with respect to I,
and M a finitely generated R-module. The current authors, in [13, Theorem 2.11],
i (M) ∈ S for all i < t, then H i (M) ∈ S for all i < t. In 2.5, we
showed that if HI,J
I
j
i (M)) ∈ S for all i < t and all
improve this result and we show that if ExtR (N, HI,J
j < t − i, then Hai (M) ∈ S for all i < t, where M is an arbitrary R-module and N is a
e J).
finitely generated R-module with SuppR N = V(a) for some a ∈ W(I,
Let S be a Serre subcategory of the category of R-modules. Aghapournahr and
Melkersson [1] introduced the notion of S-sequences on M as a generalization of
regular sequences. Suppose that S is a Melkersson subcategory with respect to I, M
is a ZD-module and M/IM ∈
/ S. In [14, Theorem 2.9] it is proved that all maximal
S-sequences on M in I, have the same length. If this common length is denoted
by S − depthI (M), then S − depthI (M) = inf{i : HIi (M) ∈
/ S}; see [14, Corollary
2.12]. In 2.25, we generalize this result as follows. Let S be the class of Artinian
R-modules, or the class of all R-modules N with dimR N ≤ k, where k is an integer.
e J)} = inf{i : H i (M) ∈
Then inf{S − deptha (M) : a ∈ W(I,
/ S}, where M is a finitely
I,J
e J).
generated R-module, or is a ZD-module such that M/aM ∈
/ S for all a ∈ W(I,
In section 3, we get some identities between local cohomology modules. Let
i (M) be a finite subset of Max(R) for all i < t. Then there are maximal
SuppR HI,J
ideals m1 , m2 , . . ., mk of R such that H i (M) ∼
= H i (M) ⊕ H i (M) ⊕ · · · ⊕ H i (M)
I,J
m1
m2
mk
for all i < t; see 3.7. As a consequence we conclude that, if (R, m) is a local ring,
∼ H i (M)}; see 3.10.
then inf{i : H i (M)is not Artinian} = inf{i : H i (M) 6=
I,J
2. E XTENSION
I,J
m
FUNCTORS OF LOCAL COHOMOLOGY MODULES
Recall that R is a Noetherian ring, I and J are ideals of R and M is an R-module.
Definition 2.1. A full subcategory of the category of R-modules is said to be Serre
subcategory, if it is closed under taking submodules, quotients and extensions. A
Serre subcategory S is said to be a Melkersson subcategory with respect to I, if for
REGULAR SEQUENCES AND LOCAL COHOMOLOGY MODULES...
3
any I-torsion R-module M, 0 :M I ∈ S implies that M ∈ S. A Serre subcategory is
called Melkersson subcategory when it is a Melkersson subcategory with respect to
all ideals of R.
The class of finitely generated modules and the class of weakly Laskerian modules are Serre subcategories. Aghapournahr and Melkersson [1, Lemma 2.2] proved
that if a Serre subcategory is closed under taking injective hulls, then it is a Melkersson subcategory. The class of zero modules, Artinian R-modules, modules with
finite support and the class of R-modules N with dimR N ≤ k, where k is a nonnegative integer, are Serre subcategories closed under taking injective hulls, and
hence are Melkersson subcategories; see [1, Example 2.4]. The class of I-cofinite
Artinian modules is a Melkersson subcategory with respect to I, but is not closed
under taking injective hulls; see [1, Example 2.5].
The following result is a generalization of [2, Theorem 2.1].
i
Theorem 2.2. Let N be an (I, J)-torsion R-module. If Extt−i
R (N, HI,J (M)) ∈ S for
all i ≤ t, then ExttR (N, M) ∈ S.
Proof. Let F(−) = HomR (N, −) and G(−) = ΓI,J (−). Then we have FG(M) =
HomR (N, M). By [15, Theorem 11.38], there is the Grothendieck spectral sequence
p,q
p
q
p+q
E2 := ExtR (N, HI,J (M)) ⇒ ExtR (N, M).
There is a finite filtration
0 = ϕ t+1 H t ⊆ ϕ t H t ⊆ · · · ⊆ ϕ 1 H t ⊆ ϕ 0 H t = ExttR (N, M)
such that E∞t−i,i ∼
= ϕ t−i H t /ϕ t+1−i H t for all i ≤ t. It is enough to show that ϕ 0 H t ∈ S.
By hypothesis, E t−i,i
∈ S for all j ≥ 2 and i ≤ t, and so E∞t−i,i ∈ S for all i ≤ t. The
j
sequence
0 −→ ϕ t+1−i H t −→ ϕ t−i H t −→ E∞t−i,i −→ 0
is exact for all i ≤ t. Therefore it follows that ϕ 0 H t ∈ S.
j
i (M)) ∈ S for all
Corollary 2.3. Let N be an (I, J)-torsion R-module. If ExtR (N, HI,J
i < t and all j < t − i, then ExtiR (N, M) ∈ S for all i < t.
Corollary 2.4. Suppose that S is a Melkersson subcategory with respect to I, and
e J). If
N is a finitely generated R-module with SuppR N = V (a) for some a ∈ W(I,
4
SH. PAYROVI AND M. LOTFI PARSA
j
i (M)) ∈ S for all i < t and all j < t − i, then H i (L, M) ∈ S for all i < t
ExtR (N, HI,J
a
and all finitely generated R-modules L.
Proof. The result follows by 2.3 and [1, Theorem 2.9].
The following result improves [13, Theorem 2.11].
Corollary 2.5. Suppose that S is a Melkersson subcategory with respect to I, and
e J). If
N is a finitely generated R-module with SuppR N = V(a) for some a ∈ W(I,
j
i (M)) ∈ S for all i < t and all j < t − i, then H i (M) ∈ S for all i < t.
ExtR (N, HI,J
a
Proof. In 2.4, put L = R.
Corollary 2.6. If S is a Melkersson subcategory with respect to I, then
i
e J)}.
inf{i : HI,J
(M) ∈
/ S} ≤ inf{inf{i : Hai (M) ∈
/ S} : a ∈ W(I,
As a generalization of finitely generated modules, Evans [9] introduced ZDmodules as follows.
Definition 2.7. An R-module M is said to be zero-divisor module (ZD-module), if
for any submodule N of M, the set ZR (M/N) is a finite union of prime ideals in
AssR M/N.
According to [7, Example 2.2], the class of ZD-modules contains finitely generated, Laskerian, weakly Laskerian, linearly compact and Matlis reflexive modules. Also it contains modules whose quotients have finite Goldie dimension and
modules with finite support, in particular Artinian modules. Therefore the class of
ZD-modules is much larger than that of finitely generated modules.
Definition 2.8. An element a of R is called S-regular on M, if 0 :M a ∈ S. A sequence a1 , . . . , at is an S-sequence on M, if ai is S-regular on M/(a1 , . . ., ai−1 )M for
i = 1, . . .,t. The S-sequence a1 , . . . , at is said to be maximal S-sequence on M, if
a1 , . . . , at , y is not an S-sequence on M for any y ∈ R.
When S is the class of zero modules, Artinian R-modules, modules with finite
support, and the class of R-modules N with dimR N ≤ k, where k is a non-negative
integer, then S-sequences on M are, poor M-sequences, filter-regular sequences,
REGULAR SEQUENCES AND LOCAL COHOMOLOGY MODULES...
5
generalized regular sequences, and M-sequences in dimension > k, respectively;
see [1, Example 2.8].
Let S be a Melkersson subcategory with respect to I, and M a ZD-module such
that M/IM ∈
/ S. The current authors, in [14, Theorem 2.9], proved that all maximal
S-sequences on M in I, have the same length.
Definition 2.9. Let S be a Melkersson subcategory with respect to I, and M a ZDmodule such that M/IM ∈
/ S. The common length of all maximal S-sequences on
M in I, is denoted by S − depthI (M). If M/IM ∈ S, we set S − depthI (M) = ∞.
Suppose that M is a ZD-module. When S is the class of zero modules, Artinian
R-modules, and modules with finite support, then S − depthI (M) is the same as
ordinary depthI (M), f − depthI (M) (filter-depth), and g − depthI (M) (generalized
depth), respectively.
Corollary 2.10. Let S be a Melkersson subcategory with respect to I, and M a
ZD-module. Then
i
e J)}.
inf{i : HI,J
(M) ∈
/ S} ≤ inf{S − deptha (M) : a ∈ W(I,
Proof. The result follows by 2.6 and [14, Corollary 2.12].
In the following, we study the relation between generalized local cohomology
modules and local cohomology modules with respect to a pair of ideals.
Corollary 2.11. Suppose that N is a finitely generated a-torsion R-module for some
e J). If Extt−i (N, H i (M)) ∈ S for all i ≤ t, then Hat (N, M) ∈ S.
a ∈ W(I,
I,J
R
Proof. The result follows by 2.2. Note that Γa (N) ⊆ ΓI,J (N) and by [8, Lemma 2.1]
we have Exti (N, M) ∼
= H i (N, M) for any integer i.
R
a
Corollary 2.12. Suppose that N is a finitely generated a-torsion R-module for some
e J). If Ext j (N, H i (M)) ∈ S for all i < t and all j < t −i, then Hai (N, M) ∈ S
a ∈ W(I,
R
I,J
for all i < t.
The following result is a generalization of [2, Theorem 2.3].
i (M)) ∈ S
Theorem 2.13. Let N be an (I, J)-torsion R-module. If ExtRs+t+1−i (N, HI,J
i (M)) ∈ S for all t < i < s + t, and Exts+t (N, M) ∈ S,
for all i < t, ExtRs+t−1−i (N, HI,J
R
t (M)) ∈ S.
then ExtsR (N, HI,J
6
SH. PAYROVI AND M. LOTFI PARSA
Proof. Let F(−) = HomR (N, −) and G(−) = ΓI,J (−). Then we have FG(M) =
HomR (N, M). By [15, Theorem 11.38], there is the Grothendieck spectral sequence
p,q
q
p+q
E2 := ExtRp (N, HI,J (M)) ⇒ ExtR (N, M).
There is a finite filtration
0 = ϕ s+t+1 H s+t ⊆ ϕ s+t H s+t ⊆ · · · ⊆ ϕ 1 H s+t ⊆ ϕ 0 H s+t = Exts+t
R (N, M)
such that E∞s+t−i,i ∼
= ϕ s+t−i H s+t /ϕ s+t+1−i H s+t for all i ≤ s +t. It is enough to show
that E2s,t ∈ S. We have the following exact sequences
s,t
0 −→ Ker dt+1−i
s,t
dt+1−i
s,t
−→ Et+1−i −→
s+t+1−i,i
Et+1−i
and
s−t−1+i,2t−i
s,t
s,t
−→ 0
0 −→ Im dt+1−i
−→ Kerdt+1−i
−→ Et+2−i
s+t+1−i,i
s+t−1−i,i
for all i. By hypothesis, Et+1−i
∈ S for all i < t, and E1+i−t
∈ S for all
s−t−1+i,2t−i
t < i < s + t. It follows that Et+1−i
∈ S for all t − s < i < t. Note that if
s−t−1+i,2t−i
s−t−1+i,2t−i
i ≤ t − s, then Et+1−i
= 0. Hence Et+1−i
∈ S for all i < t, and theres−t−1+i,2t−i
s,t
fore Im dt+1−i
∈ S for all i < t. Also we have Es+t+2
= E∞s,t ∈ S, because
E s,t ∼
= ϕ s H s+t /ϕ s+1 H s+t and ϕ s H s+t ⊆ ϕ 0 H s+t = Exts+t (N, M) ∈ S. Now the claim
∞
R
follows by the above exact sequences.
j−i
i (M)) ∈ S
Corollary 2.14. Let N be an (I, J)-torsion R-module. Let ExtR (N, HI,J
i (M)) ∈ S for all t < i <
for j = s + t, s + t + 1 and all i < t, and ExtRs+t−1−i (N, HI,J
s
t
s + t. Then Exts+t
R (N, M) ∈ S if and only if ExtR (N, HI,J (M)) ∈ S.
Proof. The claim follows by 2.2 and 2.13.
The following result is a generalization of [17, Theorem 2.3].
i (M)) ∈ S
Corollary 2.15. Let N be an (I, J)-torsion R-module. If Extt+1−i
(N, HI,J
R
t (M)) ∈ S.
for all i < t, and ExttR (N, M) ∈ S, then HomR (N, HI,J
Proof. In 2.13, put s = 0.
i (M)) be Artinian for all a ∈ W(I,
e J) and
Proposition 2.16. Let Extt+1−i
(R/a, HI,J
R
e J). Then Ext j (R/a, H t (M))
all i < t, and Extt (R/a, M) be Artinian for all a ∈ W(I,
R
e J) and all j ∈ N0 .
is Artinian for all a ∈ W(I,
R
I,J
REGULAR SEQUENCES AND LOCAL COHOMOLOGY MODULES...
7
t (M)) is Artinian for all a ∈ W(I,
e J). Also we know
Proof. By 2.15, HomR (R/a, HI,J
that
[
t
HI,J
(M) =
t (M) a) =
(0 :HI,J
e
a∈W(I,J)
[
t
HomR (R/a, HI,J
(M)).
e
a∈W(I,J)
Now the claim follows by [11, Theorem 5.1] and [1, Theorem 2.9].
e J) and all i < t. Then
Corollary 2.17. Let ExtiR (R/a, M) be Artinian for all a ∈ W(I,
j
e J), all i < t and all j ∈ N0 .
Ext (R/a, H i (M)) is Artinian for all a ∈ W(I,
I,J
R
j
e J) and
Proof. It follows by 2.16 that ExtR (R/a, ΓI,J (M)) is Artinian for all a ∈ W(I,
j
1 (M)) is Artinian for
all j ∈ N0 . Now again using of 2.16, implies that ExtR (R/a, HI,J
e J) and all j ∈ N0 . By continuing this process, the claim follows.
all a ∈ W(I,
i (M)) ∈ S for all i ≤ t, then H t (M) ∈ S.
Lemma 2.18. If HIt−i (HI,J
I,J
Proof. Let F(−) = ΓI (−) and G(−) = ΓI,J (−). Then FG(M) = ΓI,J (−). The rest
of the proof is similar to that of 2.2.
e J) and all i < t, then
Corollary 2.19. If ExtiR (R/a, M) is Artinian for all a ∈ W(I,
i (M) is Artinian for all i < t.
HI,J
i (M)) is Artinian for all
Proof. It follows by 2.17 and [1, Theorem 2.9] that HIj (HI,J
i < t and all j ∈ N0 . Now the claim follows by 2.18.
Proposition 2.20. Let S be the class of all R-modules N with dimR N ≤ k, where k
e J) and all i < t, and
is an integer. Let Extt+1−i (R/a, H i (M)) ∈ S for all a ∈ W(I,
I,J
R
ExttR (R/a, M) ∈ S
e J). Then H t (M) ∈ S.
for all a ∈ W(I,
I,J
t (M)) ∈ S for all a ∈ W(I,
e J). Also we know that
Proof. By 2.15, HomR (R/a, HI,J
t
HI,J
(M) =
[
e
a∈W(I,J)
t (M) a) =
(0 :HI,J
[
t
HomR (R/a, HI,J
(M)).
e
a∈W(I,J)
t (M) ∈ S.
It follows that HI,J
Corollary 2.21. Let S be the class of all R-modules N with dimR N ≤ k, where k is
e J) and all i < t, then H i (M) ∈ S
an integer. If Exti (R/a, M) ∈ S for all a ∈ W(I,
R
for all i < t.
I,J
8
SH. PAYROVI AND M. LOTFI PARSA
Proof. We know that
ΓI,J (M) =
[
[
HomR (R/a, ΓI,J (M)) =
e
a∈W(I,J)
HomR (R/a, M).
e
a∈W(I,J)
1 (M) ∈ S. By keeping this process,
Therefore ΓI,J (M) ∈ S. It follows by 2.20 that HI,J
the claim follows.
Corollary 2.22. Let S be the class of Artinian R-modules, or the class of all Re J)
modules N with dimR N ≤ k, where k is an integer. If Hai (M) ∈ S for all a ∈ W(I,
i (M) ∈ S for all i < t.
and all i < t, then HI,J
Proof. The claim follows by 2.19, 2.21 and [1, Theorem 2.9].
Corollary 2.23. Let S be the class of Artinian R-modules, or the class of all Rmodules N with dimR N ≤ k, where k is an integer. Then the following statements
are equivalent:
i (M) ∈ S for all i < t;
(i) HI,J
e J) and all i < t.
(ii) Hai (M) ∈ S for all a ∈ W(I,
Proof. The claim follows by 2.5 and 2.22.
Corollary 2.24. Let S be the class of Artinian R-modules, or the class of all Rmodules N with dimR N ≤ k, where k is an integer. Then
i
e J)}.
inf{i : HI,J
(M) ∈
/ S} = inf{inf{i : Hai (M) ∈
/ S} : a ∈ W(I,
The following result is a generalization of [14, Theorem 2.13].
Theorem 2.25. Let S be the class of Artinian R-modules, or the class of all Rmodules N with dimR N ≤ k, where k is an integer. Let M be a finitely generated
e J). Then
R-module, or be a ZD-module such that M/aM ∈
/ S for all a ∈ W(I,
i
e J)}.
inf{i : HI,J
(M) ∈
/ S} = inf{S − deptha (M) : a ∈ W(I,
Proof. The claim follows by 2.24 and [14, Theorem 2.13].
REGULAR SEQUENCES AND LOCAL COHOMOLOGY MODULES...
3. S OME
9
IDENTITIES BETWEEN LOCAL COHOMOLOGY MODULES
Suppose that
d0
di
E • (M) : 0 −→ ER0 (M) −→ ER1 (M) −→ · · · −→ ERi (M) −→ ERi+1 (M) −→ · · · (∗)
L
is a minimal injective resolution of M, where ERi (M) ∼
= p∈Spec(R) µ i (p, M)ER(R/p)
is a decomposition of ERi (M) as the direct sum of indecomposable injective Rmodules, ER (R/p) denotes the injective hull of R/p and µ i (p, M) denotes the i-th
Bass number of M with respect to p. It follows by [16, Proposition 1.11] that
M
ΓI,J (ERi (M)) ∼
=
µ i (p, M)ER (R/p).
p∈W(I,J)
Hence SuppR ΓI,J (ERi (M)) = {p ∈ W(I, J) : µ i (p, M) 6= 0}.
The above mentioned results are assumed known through this section.
Theorem 3.1. Let S be a Serre subcategory closed under taking injective hulls. The
following conditions are equivalent:
i (M) ∈ S for all i < t.
(i) HI,J
(ii) ΓI,J (ERi (M)) ∈ S for all i < t.
Proof. Since ΓI,J (ERi (M)) is injective and Ker ΓI,J (d i ) = Ker d i ∩ ΓI,J (ERi (M)), thus
ΓI,J (ERi (M)) is injective hull of Ker ΓI,J (d i ). Now the claim follows by [12, Lemma
5.4]. We note that the proof of [12, Lemma 5.4] is still valid if the class of Artinian
R-modules is replaced by a Serre subcategory that is closed under taking injective
hulls.
Corollary 3.2. The following statements are equivalent:
i (M) is a finite subset of Max(R) for all i < t;
(i) SuppR HI,J
i (M) is Artinian for all i < t.
(ii) HI,J
Corollary 3.3. If (R, m) is a local ring, then
i
i
inf{i : SuppR HI,J
(M) 6⊆ {m}} = inf{i : HI,J
(M)is not Artinian}.
Proposition 3.4. Let M be a finitely generated R-module, or be a ZD-module such
that Mp /pMp 6= 0 for all p ∈ W(I, J). Then
i
inf{i : HI,J
(M) 6= 0} = inf{depth Mp : p ∈ W(I, J)}.
10
SH. PAYROVI AND M. LOTFI PARSA
Proof. Let t = inf{depth Mp : p ∈ W(I, J)}. It follows by [14, Corollary 2.14] that
µ i (p, M) = 0 for all p ∈ W(I, J) and all i < t. So ΓI,J (ERi (M)) = 0 for all i < t,
i (M) = 0 for all i < t. Therefore t ≤ inf{i : H i (M) 6= 0}. Now
and hence HI,J
I,J
t (M) 6= 0. By assumption, there is q ∈ W(I, J) such
it is enough to show that HI,J
that t = depth Mq . It follows by [14, Corollary 2.14] that µ t (q, M) 6= 0. Therefore
t (M) 6= 0 by 3.1.
ΓI,J (ERt (M)) 6= 0, and hence HI,J
We can get a generalization of [6, Theorem 2.4].
Proposition 3.5. Let M be a finitely generated R-module, or be a ZD-module such
that Mp /pMp 6= 0 for all p ∈ W(I, J) − Max(R). Then
i
inf{i : SuppR HI,J
(M) 6⊆ Max(R)} = inf{depth Mp : p ∈ W(I, J) − Max(R)}.
i (M) 6⊆ Max(R)}. It follows by 3.1 that µ i (p, M) = 0
Proof. Let t = inf{i : SuppR HI,J
for all p ∈ W(I, J) − Max(R) and all i < t, and there is q ∈ W(I, J) − Max(R) such
that µ t (q, M) 6= 0. Now it follows by [14, Corollary 2.14] that depth Mp ≥ t for all
p ∈ W(I, J) − Max(R), and depth Mq = t. Therefore inf{depth Mp : p ∈ W(I, J) −
Max(R)} = t.
i (M) be a finite subset of Max(R) for all i < t. Then
Theorem 3.6. Let SuppR HI,J
i (M) ∼ H i
there are maximal ideals m1 , m2 , . . ., mk ∈ W(I, J) such that HI,J
= m1 m2 ···mk (M)
for all i < t.
Proof. It follows by 3.1 that SuppR ΓI,J (E i (M)) is a finite subset of Max(R) for all
i < t. Let SuppR ΓI,J (E i (M)) = {mi1 , mi2 , ..., miki }, where ki is an integer. Then
ΓI,J (E (M)) ∼
=
i
ki
M
µ i (mi j , M)ER (R/mi j )
j=1
for all i < t. Put a = ∏i, j mi j . Then V(a) = {mi j : 0 ≤ i < t, 1 ≤ j ≤ ki }, and
e J). Therefore
a ∈ W(I,
Γa (E i (M)) ∼
=
M
µ i (p, M)ER (R/p)
p∈V(a)
=
ki
M
µ i (mi j , M)ER (R/mi j ) ∼
= ΓI,J (E i (M))
j=1
i (M) ∼ H i (M) for all i < t.
for all i < t, and hence HI,J
= a
REGULAR SEQUENCES AND LOCAL COHOMOLOGY MODULES...
11
i (M) be a finite subset of Max(R) for all i < t. Then
Corollary 3.7. Let SuppR HI,J
there are maximal ideals m1 , m2 , . . . , mk ∈ W(I, J) such that
i
HI,J
(M) ∼
= Hmi 1 (M) ⊕ Hmi 2 (M) ⊕ · · · ⊕ Hmi k (M)
for all i < t.
Proof. The claim follows by 3.6 and the Mayer-Vietoris sequence [4, 3.2.3].
i (M) ⊆ {m} for all i < t,
Corollary 3.8. Let (R, m) be a local ring. If SuppR HI,J
then H i (M) ∼
= H i (M) for all i < t.
I,J
m
Corollary 3.9. If (R, m) is a local ring, then
i
i
inf{i : SuppR HI,J
(M) 6⊆ {m}} = inf{i : HI,J
(M) ∼
6 Hmi (M)}.
=
The following result is a generalization of [6, Proposition 2.5].
Corollary 3.10. If (R, m) is a local ring, then
i
i
(M) ∼
6 Hmi (M)}.
(M)is not Artinian} = inf{i : HI,J
inf{i : HI,J
=
Proof. The claim follows by 3.3 and 3.9.
i (M) ⊆ {m} for all i < t,
Corollary 3.11. Let (R, m) be a local ring. If SuppR HI,J
e J) and all i < t.
then H i (M) ∼
= Hai (M) for all a ∈ W(I,
I,J
e J) and all i < t.
Proof. It follows by 2.5 that SuppR Hai (M) ⊆ {m} for all a ∈ W(I,
e J) and all i < t, by 3.8. Now again
Therefore Hai (M) ∼
= Hmi (M) for all a ∈ W(I,
i (M) ∼ H i (M) for all i < t, and the claim follows.
using of 3.8 implies that HI,J
= m
R EFERENCES
[1]
M. Aghapournahr, L. Melkersson, Local cohomology and Serre subcategories, J. Algebra 320
(2008), 1275-1287.
[2]
M. Aghapournahr, A. J. Taherizadeh, A. Vahidi, Extension functors of local cohomology modules, Bull. Iranian Math. Soc. 37 (2011), 117-134.
[3]
M. H. Bijan-Zadeh, Torsion theories and local cohomology over commutative Noetherian
rings, J. London Math. Soc. (2) 19 (1979), 402-410.
[4]
M. P. Brodmann, R. Y. Sharp, Local Cohomology: An Algebraic Introduction with Geometric
Applications, Cambridge University Press, Cambridge, 1998.
[5]
W. Bruns, J. Herzog, Cohen-Macaulay Rings, Cambridge University Press, Cambridge, 1993.
12
[6]
SH. PAYROVI AND M. LOTFI PARSA
L. Chu, Q. Wang, Some results on local cohomology modules defined by a pair of ideals, J.
Math. Kyoto Univ. 49 (2009), 193-200.
[7]
K. Divaani-Aazar, M. A. Esmkhani, Artinianness of local cohomology modules of ZDmodules, Comm. Algebra 33 (2005), 2857-2863.
[8]
K. Divaani-Aazar, R. Sazeedeh, M. Tousi, On vanishing of generalized local cohomology
modules, Algebra Colloq. 12 (2005), 213-218.
[9]
E. G. Evans, Zero divisors in Noetherian-like rings, Trans. Amer. Math. Soc. 155 (1971),
505-512.
[10] A. Grothendieck (notes by R. Hartshorne), Local Cohomology, Springer Lecture Notes in
Math., 41, Springer-Verlag, 1966.
[11] L. Melkersson, Some applications of a criterion for Artinianness of a module, J. Pure Appl.
Algebra 101 (1995), 291-303.
[12] L. Melkersson, Modules cofinite with respect to an ideal, J. Algebra 285 (2005), 649-668.
[13] Sh. Payrovi, M. Lotfi Parsa, Finiteness of local cohomology modules defined by a pair of
ideals, Comm. Algebra 41 (2013), 627-637.
[14] Sh. Payrovi, M. Lotfi Parsa, Regular sequences and ZD-modules, Eprint arXiv: 1305.0164v1.
[15] J. Rotman, An Introduction to Homological Algebra, Academic Press, Orlando, 1979.
[16] R. Takahashi, Y. Yoshino, T. Yoshizawa, Local cohomology based on a nonclosed support
defined by a pair of ideals, J. Pure Appl. Algebra 213 (2009), 582-600.
[17] A. Tehranian, A. Pour Eshmanan Talemi, Cofiniteness of local cohomology based on a nonclosed support defined by a pair of ideals, Bull. Iranian Math. Soc. 36 (2010), 145-155.
I. K. I NTERNATIONAL U NIVERSITY, P OSTAL C ODE : 34149-1-6818 Q AZVIN - IRAN
E-mail address: [email protected]
E-mail address:
[email protected]
| 0math.AC
|
arXiv:1703.03701v2 [math.GR] 20 Mar 2017
Profinite detection of 3-manifold
decompositions
Henry Wilton∗ and Pavel Zalesskii†
March 21, 2017
Abstract
The profinite completion of the fundamental group of a closed,
orientable 3-manifold determines the Kneser–Milnor decomposition.
If M is irreducible, then the profinite completion determines the Jaco–
Shalen–Johannson decomposition of M .
When trying to distinguish two compact 3-manifolds M, N, in practice the
easiest method is often to compute some finite quotients of their fundamental
groups, and notice that there is a finite group Q which is a quotient of π1 M,
say, but not of π1 N. It would be very useful, both theoretically and in
practice, to know that this method always works. The set of finite quotients
̂ (the inverse limit of the
of a group Γ is encoded by the profinite completion Γ
system of finite quotient groups), and so one is naturally led to the following
question.
Question 0.1. Let M be a closed, orientable 3-manifold. To what extent is
π1 M determined by its profinite completion?
In particular, if M is determined among all compact, orientable 3-manifolds
̂
by π1 M, then M is said to be profinitely rigid. If there are at most finitely
̂
̂
many closed, orientable 3-manifolds N with π
1 M ≅ π1 N , then M is said
∗
Department of Pure Mathematics and Mathematical Statistics, Centre for Mathematical Sciences, Wilberforce Road, Cambridge, CB3 0WB, United Kingdom
†
Department of Mathematics, University of Brası́lia, 70910-9000 Brası́lia, Brazil
1
to have finite genus. More precise versions of Question 0.1 ask which 3manifolds are profinitely rigid, which have finite genus and whether various
̂
properties of M are determined by π
1M .
̂
The results of this paper show that the profinite completion π
1 M detects
both the Kneser–Milnor and the JSJ decompositions of M. This complê
ments our previous results showing that π
1 M detects the geometry of M
[WZ17]. The first theorem concerns the Kneser–Milnor decomposition.
Theorem A. Let M, N be closed, orientable 3-manifolds with Kneser–Milnor
decompositions M ≅ M1 # . . . #Mm #r(S1 ×S2 ) and N ≅ N1 # . . . #Nn #s(S1 ×
̂
̂
̂
S2 ). If π
1 M ≅ π1 N then m = n, r = s, and up to re-indexing, π1 M i is conjugate to π̂
1 N i for each i.
̂
In particular, π
1 M detects whether or not M is irreducible. While this
work was in progress we discovered that a similar result has also been proved
in the pro-p setting by Wilkes, using l2 -cohomology [Wil17, Proposition
6.2.4]. Our proof is different, using the continuous cohomology of the profinite completion, and naturally generalizes to our next theorem, which shows
that the profinite completion detects the JSJ decomposition of M.
Theorem B. Let M and N be closed, orientable, irreducible 3-manifolds,
̂
̂
and suppose π
1 M ≅ π1 N . Then the underlying graphs of the JSJ decompositions of π1 M, π1 N are isomorphic, and corresponding vertex groups have
isomorphic profinite completions.
See Theorem 4.3 for a more precise statement, phrased in terms of profinite Bass–Serre trees. Partial results along the lines of Theorem B have also
been obtained by Wilkes [Wil16].
In [WZ17, Theorem 8.4], it was shown that the profinite completion of
the fundamental group of a closed, orientable 3-manifold M determines the
geometry of M. As an immediate consequence of Theorem B, we can extend
this result to the case with toral boundary.
Corollary C. Let M, N be compact, orientable, irreducible 3-manifolds with
̂
̂
non-empty toral boundaries, and suppose that π
1 M ≅ π1 N. If M is geometric
then N is also geometric, with the same geometry. In particular, N is Seifert
fibred if and only if M is Seifert fibred.
Proof. Let D be the double of M and E be the double of N. Note that
̂
π̂
1 D ≅ π1 E, and also that the isomorphism respects the profinite completions
2
of the boundary subgroups of π1 M and π1 N. The result follows from the
observation that the geometry of M is reflected in properties of the double
D.
Indeed, if M is Seifert fibred then so is D. In this case, if M is homeomorphic to an interval bundle over the torus then D has Euclidean geometry;
otherwise, M and D both have H2 × R geometry. Finally, M is hyperbolic if
and only if D has non-trivial JSJ decomposition, and the boundary tori of
M are the only JSJ tori of D. Combining these facts with Theorem B and
[WZ17, Theorem 8.4], the result follows.
Theorem B shows that the problem of distinguishing 3-manifolds using
their profinite completions can be reduced to the pieces of the JSJ decomposition. The Seifert-fibred case has been resolved by Wilkes [Wilar], building on
work of Hempel [Hem14]: Seifert fibred 3-manifolds are not profinitely rigid,
but do have finite genus, and Wilkes was able to give a complete description
of when two such 3-manifold groups have isomorphic profinite completions;
he was subsequently able to extend this to a complete answer to Question 0.1
for graph manifolds [Wil16]. Sol-manifolds are also well understood: they are
not profinitely rigid [Fun13], but do have finite genus [GPS80]. A definitive
treatment of this case would be a valuable addition to the literature.
The case of hyperbolic 3-manifolds remains an important open problem.
The complement of the figure-eight knot was shown to be profinitely rigid
by Boileau–Friedl [BF15] and by Bridson–Reid [BR15]; see also [BCR16] for
analogous results for Fuchsian groups and [BRW16] for once-punctured-torus
bundles.
Both [BF15] and [BR15] rely on results showing that fibredness is a
profinite invariant in certain contexts, and the question of whether or not
fibredness is a profinite invariant is also an important problem. This is
known to hold for hyperbolic 3-manifolds M with b1 (M) = 1: a proof using
twisted Alexander polynomials was given by Boileau–Friedl [BF15]; topological proofs were given by Bridson–Reid in the case with boundary [BR15],
and by Bridson–Reid–Wilton in the closed case [BRW16]. More generally,
Boileau–Friedl proved that fibredness is an invariant of regular profinite isomorphism. Ueki also recently showed that the profinite completion of a knot
group determines the Alexander polynomial of the knot [Uek17].
The results of this paper are proved by considering profinite Poincaré
Duality groups. The main difficulty in the above theorems is to show that
profinite completions of 3-manifold groups do not admit unexpected splittings
3
which are not induced by splittings of the underlying group. It is well known
that non-splitting theorems for discrete Poincaré Duality groups follow from
the Mayer–Vietoris sequence. As a result of the work of Agol, Wise et al. on
the Virtual Haken conjecture [Ago13, Wis12], 3-manifold groups are known to
be good in the sense of Serre, meaning that the cohomology of the profinite
completion is isomorphic to the ordinary cohomology (with coefficients in
finite modules). Furthermore, a version of the Mayer–Vietoris sequence is
known for efficient decompositions of profinite completions. The main idea
of the proofs of Theorems A and B is to prove the analogues for profinite
completions of the non-splitting theorems from the discrete case.
Acknowledgements
The authors are grateful to Peter Kropholler for explaining the results of
[KR88], and to Michel Boileau and Stefan Friedl for pointing out Corollary C.
This work was completed while the first author was a participant of the Nonpositive curvature, group actions and cohomology programme at the Isaac
Newton Institute, funded by the EPSRC Grant number EP/K032208/1. The
first author is also supported by an EPSRC Standard Grant EP/L026481/1.
The second author was supported by CAPES as part of the ‘Estagio Senior’
programme, and thanks Trinity College and the Department of Pure Mathematics and Mathematical Statistics at the University of Cambridge for their
hospitality.
1
1.1
Preliminaries on profinite groups
Profinite trees
A graph Γ is a disjoint union E(Γ) ∪ V (Γ), with two maps d0 , d1 ∶ Γ → V (Γ)
that are the identity on the set of vertices V (Γ). For an element e of the set
of edges E(Γ), d0 (e) is called the initial and d1 (e) the terminal vertex of e.
Definition 1.1. A profinite graph Γ is a graph such that:
(i) Γ is a profinite space (i.e. an inverse limit of finite discrete spaces);
(ii) V (Γ) is closed; and
(iii) the maps d0 and d1 are continuous.
4
Note that E(Γ) is not necessary closed.
A morphism α ∶ Γ Ð→ ∆ of profinite graphs is a continuous map with
αdi = di α for i = 0, 1.
By [ZM88, Prop. 1.7] every profinite graph Γ is an inverse limit of finite
quotient graphs of Γ.
For a profinite space X that is the inverse limit of finite discrete spaces
̂
̂ j ], where [ZX
̂ j ] is the free Z-module
̂
Xj , [[ZX]]
is the inverse limit of [ZX
with basis Xj . For a pointed profinite space (X, ∗) that is the inverse limit
̂
of pointed finite discrete spaces (Xj , ∗), [[Z(X,
∗)]] is the inverse limit of
̂ j , ∗)], where [Z(X
̂ j , ∗)] is the free Z-module
̂
[Z(X
with basis Xj ∖ {∗} [RZ10,
Chapter 5.2].
For a profinite graph Γ define the pointed space (E ∗ (Γ), ∗) as Γ/V (Γ)
with the image of V (Γ) as a distinguished point ∗, and denote the image of
e ∈ E(Γ) by ē.
Definition 1.2. A profinite tree Γ is a profinite graph such that the sequence
δ
ǫ
̂ ∗ (Γ), ∗)]] →
̂ (Γ)]] →
̂→0
0 → [[Z(E
[[ZV
Z
is exact, where δ(ē) = d1 (e) − d0 (e) for every e ∈ E(Γ) and ǫ(v) = 1 for every
v ∈ V (Γ).
If v and w are elements of a profinite tree T , we denote by [v, w] the
smallest profinite subtree of T containing v and w and call it a geodesic.
By definition a profinite group G acts on a profinite graph Γ if we have
a continuous action of G on the profinite space Γ that commutes with the
maps d0 and d1 .
We shall need the following lemma; its proof is contained in the first eight
lines of the proof of [Zal90, Lemma 2.3].
Lemma 1.3. Suppose that a profinite group G acts on a profinite tree T and
does not fix any vertex. Then there exists an open normal subgroup U of G
that is not generated by its vertex stabilizers.
When we say that G is a finite graph of profinite groups we mean that it
contains the data of the underlying finite graph, the edge profinite groups,
the vertex profinite groups and the attaching continuous maps. More precisely, let ∆ be a connected finite graph. The data of a graph of profinite
5
groups (G, ∆) over ∆ consists of a profinite group G(m) for each m ∈ ∆, and
continuous monomorphisms ∂i ∶ G(e) Ð→ G(di (e)) for each edge e ∈ E(∆).
The definition of the profinite fundamental group of a connected profinite graph of profinite groups is quite involved (see [ZM89]). However, the
profinite fundamental group Π1 (G, Γ) of a finite graph of finitely generated
profinite groups (G, Γ) can be defined as the profinite completion of the abstract (usual) fundamental group Πabs
1 (G, Γ) (using here that every subgroup
of finite index in a finitely generated profinite group is open, [NS07, Theorem
1.1]). The fundamental profinite group Π1 (G, Γ) has the following presentation:
Π1 (G, Γ) = ⟨G(v), te ∣ rel(G(v)), ∂1 (g) = ∂0 (g)te , g ∈ G(e),
te = 1 for e ∈ T ⟩;
(1)
where T is a maximal subtree of Γ and ∂0 ∶ G(e) Ð→ G(d0 (e)), ∂1 ∶ G(e) Ð→
G(d1 (e)) are monomorphisms.
In contrast to the abstract case, the vertex groups of (G, Γ) do not always
embed in Π1 (G, Γ), i.e., Π1 (G, Γ) is not always proper. However, the edge
and vertex groups can be replaced by their images in Π1 (G, Γ) and after this
replacement Π1 (G, Γ) becomes proper.
The profinite fundamental group Π1 (G, Γ) acts on the standard profinite
tree T (defined analogously to the abstract Bass–Serre tree) associated to it,
with vertex and edge stabilizers being conjugates of vertex and edge groups,
and such that Π1 (G, Γ)/T = Γ [ZM88, Proposition 3.8]. In particular, this
applies to the cases of an amalgamated free product G = G1 ∐H G2 (Γ is an
edge with two vertices) and an HNN-extension G = G1 ∐H (Γ is a loop); if G
is proper and, in the case of an amalgamated free product, G1 ≠ H ≠ G2 , we
say that G splits over H.
Example 1.4. If G = π1 (G, Γ) is the fundamental group of a finite graph of
(abstract) groups then one has the induced graph of profinite completions of
̂ Γ) and a natural homomorphism G = π1 (G, Γ) Ð→
edge and vertex groups (G,
̂ Γ). It is an embedding if π1 (G, Γ) is residually finite. In this case
Π1 (G,
̂ Γ) is simply the profinite completion G.
̂ Moreover, if the edge groups
Π1 (G,
G(e) are separable in G then the standard tree TG naturally embeds in the
standard profinite tree T̂G (see [CB13, Proposition 2.5]). In particular this is
the case if edge groups are finitely generated and G is subgroup separable.
6
1.2
Profinite Poincaré duality groups
In this section we collect the facts about profinite groups that we will need.
The following results are all profinite analogues of well known results in the
setting of discrete groups.
Definition 1.5 ([SW00]). Let p be a prime. A profinite group G of type
p-F P∞ is called a Poincaré duality group at p of dimension n if cdp (G) = n
and
H i (G, Zp [[G]]) = 0,
if i =/ n,
H n (G, Zp [[G]]) ≅ Zp (as abelian groups).
We say that such a group G is a profinite P D n -group at p.
If G is a profinite group with cdp (G) < ∞ and U is an open subgroup
of G, then G is a profinite P D n -group at p if and only if U is a profinite
P D n -group at p.
Our proof relies on the following lemma. In the discrete case, the corresponding result is Strebel’s theorem [Str77]. In the profinite case, this is
Exercise 5(b) on p. 44 of [Ser97].
Lemma 1.6. Let G be a profinite P D n group at p and H a a closed subgroup
of G such that p∞ ∣[G ∶ H]. Then cdp (H) < n.
The following theorem is the profinite analogue of the well known fact
in the discrete setting that P D n groups cannot split over groups of much
smaller cohomological dimension [DD89, Proposition V.7.4].
Theorem 1.7. Let G be a profinite P D n group at every prime p acting on a
profinite tree T with edge stabilizers of cd(Ge ) < n − 1. Then G fixes a vertex.
Proof. By [ZM88, paragraph 2.7],
cd G ≤ sup{cd Gv , cd Ge + 1 ∣ v ∈ V (T ), e ∈ E(T )}.
Suppose that G acts on T without fixing a vertex. We now argue that there
exists p such that p∞ ∣[G ∶ Gv ] for every v ∈ V (Γ), and deduce a contradiction
from Lemma 1.6.
By [Zal90, Lemma 1.5] we may assume that the action of G on T is irreducible (i.e. does not contain proper G-invariant subtrees). If K is the kernel
of the action then G/K acts faithfully on T . Hence by [Zal90, Proposition
2.10 and Lemma 2.7] G/K contains a free pro-p subgroup acting freely on T
and therefore so does G, whence p∞ ∣[G ∶ Gv ] as claimed.
7
We will apply these results to discrete groups Γ such that the cohomology
of Γ is closely intertwined with the cohomology of the profinite completion
̂
Γ – Serre called such groups ‘good’ [Ser97, I.2.6].
Definition 1.8. A discrete group Γ is good (in the sense of Serre) if, for
any finite Γ-module M, the natural map to the profinite completion Γ → ̂
Γ
̂ M) (where the cohomology of
induces an isomorphism H ∗(Γ, M) ≅ H ∗ (Γ,
̂ is defined using the continuous Hom functor).
the profinite group Γ
It has been noticed in various places (eg. [Cav12], [AFW15]; cf. [GJZZ08])
that 3-manifold groups are good. For convenience, we record the result here.
Theorem 1.9. If M is a closed 3-manifold then π1 M is good.
Proof. Since goodness passes to finite extensions, we may assume that M
is orientable. By [WZ10, Proposition 4.3] and the usual Kneser–Milnor and
JSJ decompositions, it suffices to prove that Seifert fibred and hyperbolic
3-manifold groups are good. The Seifert-fibred case is Proposition 4.2 of the
same paper, and the case of closed hyperbolic 3-manifolds follows from the
Virtually fibred conjecture [Ago13], by [GJZZ08, Lemmas 3.2 and 3.3].
The next result is the subject of [KZ08, Theorem 4.1] for P D 3 -groups,
and for general n the proof can be repeated replacing 3 by n.
Theorem 1.10. If Γ is a good P D n group, then ̂
Γ is P D n at every p.
We immediately obtain a profinite non-splitting result for good Poincaré
duality groups.
Corollary 1.11. Let G be P D n group which is good in the sense of Serre.
̂ on a profinite tree with edge stabilizers of cohomological
Then any action of G
dimension n − 2 has a global fixed point.
̂ is a profinite P D n group at p for every p by
Proof. Since G is good, G
Theorem 1.10, so the result follows from Theorem 1.7.
Remark 1.12. The combined hypotheses of goodness and P D n apply to many
examples in dimensions 2 and 3, but are restrictive in higher dimensions.
Combining all of the above results, we obtain the following fact, which
will be extremely useful to us in what follows.
8
Corollary 1.13. If M is a closed, orientable, irreducible 3-manifold then
̂
any action of π
1 M on a profinite tree with procyclic edge stabilizers has a
global fixed point.
Proof. By the Sphere Theorem, irreducible 3-manifolds either have finite
̂
fundamental group or are aspherical. In the first case π
1 M is finite, and the
result follows from [ZM88, Theorem 2.10]. In the second case, π1 M is P D 3 ,
and the result follows from Theorem 1.9 and Corollary 1.11.
2
The Kneser–Milnor decomposition
As a warm-up, we show that the profinite completion of a 3-manifold group
detects its Kneser–Milnor decomposition. As noted above, this result can
also be obtained using methods from l2 -cohomology [Wil17]. Recall that a
closed 3-manifold M is irreducible if every embedded 2-sphere bounds a 3ball; equivalently, π1 M does not admit a non-trivial splitting over the trivial
subgroup.
Proposition 2.1. Suppose that M1 , M2 are closed, orientable 3-manifolds.
̂
̂
If π
1 M 1 ≅ π1 M 2 and M1 is irreducible then so is M2 .
Proof. If M2 were reducible then π1 M2 would act on a tree with trivial edge
̂
̂
stabilizers and without a global fixed point, and π
1 M 1 ≅ π1 M 2 would act on
a profinite tree with trivial edge stabilizers and without a global fixed point.
This contradicts Corollary 1.13.
Non-irreducible 3-manifolds admit non-trivial Kneser–Milnor decompositions. If M is a closed, oriented 3-manifold then the Kneser–Milnor decomposition decomposes M as a connect sum
M ≅ N1 # . . . #Nm #Fr
where each Ni is irreducible and Fr is a connect sum of copes of S 1 × S 2 .
The Ni are uniquely determined, in an appropriate sense. In particular, the
conjugacy classes of the subgroups π1 Ni are unique up to reordering, and the
integer r is also unique. The reader is referred to [AFW15, Theorem 1.2.1]
for details.
9
Theorem 2.2 (Profinite Kneser–Milnor). Consider closed, orientable 3manifolds with Kneser–Milnor decompositions M = N1 # . . . #Nm #Fr and
′
′
M ′ = N1′ # . . . #Nm
′ #Fr ′ , where each Ni and Nj is irreducible and Fr and Fr ′
′
′
̂
̂′
are connect sums of S 1 × S 2 ’s. If π
1 M ≅ π1 M then m = m , r = r , and up to
̂′
̂
reordering, π
1 Ni is conjugate to π1 Ni for each i.
Proof. Let S be the Bass–Serre tree of the corresponding decomposition of
̂
π1 M, and let Ŝ be the corresponding profinite tree for π
1 M on which π1 M
acts with trivial edge stabilizers. By Corollary 1.13, each profinite completion
′
̂
̂
π̂
1 Ni fixes a vertex of S, and hence is conjugate into some π1 Nj . By symmê′
try, each π̂
1 Ni is conjugate into some π1 Nj . Since profinite subgroups cannot
be conjugate to proper subgroups of themselves, it follows that m = m′ and
the profinite completions of the vertex groups are conjugate. Factoring the
normal closures of these subgroups out, we see that F̂r ≅ F̂r′ and hence r = r ′
as claimed.
3
Cusped hyperbolic 3-manifolds
An immediate consequence of Corollary 1.11 is that, for a closed 3-manifold
̂
M, π
1 M does not split over a subgroup of cohomological dimension 0 or 1
(for instance a profinite free group). In this section, we prove some profinite
non-splitting results for hyperbolic manifolds with toral boundary. In the
hyperbolic case, we will need a fact from [WZ17], describing the non-procyclic
abelian subgroups of π̂
1N .
Proposition 3.1. Let N be a finite-volume hyperbolic 3-manifold and A a
closed abelian subgroup of π̂
1 N. If A is not procyclic then A is in the closure
of a peripheral subgroup of π1 N, and this peripheral subgroup is unique up to
conjugacy.
Proof. By [WZ17, Theorem 9.3], A is conjugate into the closure of a peripheral subgroup, and by [WZ17, Lemma 4.5], the conjugacy class of the cusp
subgroup is unique.
In the classical P D n setting, one handles manifolds with boundary using
the theory of P D n pairs [Dic80]. One of the upshots of this theory is that
the fundamental group of an aspherical manifold with aspherical boundary
cannot split over a boundary subgroup, relative to the collection of boundary
10
subgroups. (This can be deduced from the results of [KR88].) No doubt the
profinite analogue of this statement can be proved by developing the theory
of profinite P D n pairs. We take a quicker route here: we prove the result in
the cusped hyperbolic case, using Dehn filling. First, we need to recall the
definition of an acylindrical splitting.
Definition 3.2. An action of a group Γ on a tree T is k-acylindrical (for
an integer k) if, for every γ ∈ Γ ∖ 1, the subtree fixed by γ is either empty
or of diameter at most k. Likewise, an action of a profinite group G on
profinite tree T̂ is k-acylindrical if the subtree fixed by γ̂ is either empty or
of diameter at most k, for every γ̂ ∈ G. Such an action is called acylindrical
if it is k-acylindrical for some k.
Acylindricity gives useful control over non-cyclic abelian subgroups, via
the following lemma. This was proved in [WZ17, Theorem 5.2]. (The discrete
version of this fact is left as an instructive exercise to the reader.)
Lemma 3.3. If A is an abelian, profinite, non-procyclic group, and A acts
acylindrically on a profinite tree T , then A fixes a vertex.
We are now ready to prove the non-splitting result for hyperbolic manifolds with cusps.
Lemma 3.4. If N is a compact, orientable, hyperbolic 3-manifold with toral
boundary and π̂
1 N acts on a profinite tree T with each edge stabilizer either
procyclic or conjugate into a peripheral subgroup, then π̂
1 N fixes a vertex.
Proof. First, note that if π̂
1 N acts on T without fixed points then, by Lemma
1.3, after passing to a proper open subgroup we may assume that π̂
1 N is not
generated by vertex stabilizers.
Let the family of peripheral subgroups of π1 N be P1 , . . . , Pn . By Thurston’s
hyperbolic Dehn filling theorem, we may choose slopes ci ∈ Pi so that the
resulting Dehn filled manifold N(c1 , . . . , cn ) is a closed, hyperbolic (in particular, aspherical) manifold. Therefore,
̂
π̂
1 N/⟪c1 , . . . , cn ⟫ ≅ π1 N(c1 , . . . , cn )
is a profinite P D 3 group. Since ⟪c1 , . . . , cn ⟫ is generated by vertex stabilizers,
π̂
1 N(c1 , . . . , cn ) acts on a profinite tree ⟪c1 , . . . , cn ⟫/T (see [ZM88, Prop. 2.5])
and still does not fix a vertex. The edge stabilizers of the latter action are
procyclic. This contradicts Corollary 1.11.
11
4
The JSJ decomposition
In this section we show that, as well as the Kneser–Milnor decomposition,
the JSJ decomposition is also detected by the profinite completion. In order
to avoid ambiguity, we start by stating the form of the JSJ decomposition
we consider. In a nutshell, it is the minimal decomposition along tori such
that the complementary pieces are geometric.
Definition 4.1. Let M be a closed, orientable, irreducible 3-manifold which
is not a torus bundle over the circle. Let T ⊆ M be an embedded disjoint
union of essential tori such that the connected components of M ∖T are each
either Seifert fibred or hyperbolic. Such a union T with the smallest number
of connected components is called the JSJ decomposition of M.
The existence of the JSJ decomposition follows from the work of Jaco–
Shalen–Johannson together with Perelman’s proof of the geometrization conjecture; see [AFW15, §1.6, §1.7] for details. The tori are unique up to isotopy.
We follow Wilkes’ elegant terminology [Wil16], and use the term minor to
denote those components of M ∖ T that are homeomorphic to the twisted interval bundle over the Klein bottle; the remaining components we call major.
If two minor components are adjacent then their union is virtually a torus
bundle over a circle, and so admits either Euclidean, Nil- or Sol-geometry,
which contradicts the hypothesis that T was minimal. Therefore, every edge
adjoins at least one major vertex.
The submanifold T induces a graph-of-spaces decomposition of M, and
hence a graph-of-groups decomposition of π1 M and a profinite graph-of̂
groups decomposition of π
1 M (see Example 1.4). The Bass–Serre trees of the
latter are denoted by TM and T̂M , respectively. Crucially, these trees turn
out to be acylindrical, in the sense of Definition 3.2.
Proposition 4.2. For M a closed, orientable, irreducible 3-manifold, the
JSJ tree TM and the profinite JSJ tree T̂M are both 4-acylindrical.
Proof. In [WZ10] the authors showed that the corresponding decomposition
of π1 M is 4-acylindrical and fits into Example 1.4. In [HWZ13], the authors showed with Hamilton that the corresponding profinite decomposition
̂
of π
1 M is proper and also 4-acyindrical (see also [WZ17, Lemma 4.5]).
We are now ready to state our main theorem,
12
Theorem 4.3. If M, M ′ are closed, orientable, irreducible 3-manifolds and
≅
̂
̂′
f ∶π
1 M → π1 M
is an isomorphism, then there is an f -equivariant isomorphism
φ ∶ T̂M → T̂M ′
of the corresponding profinite Bass–Serre trees. In particular, the underlying
graphs of the JSJ decompositions of M and M ′ are isomorphic, as are the
profinite completions of the fundamental groups of the corresponding pieces.
Consider a vertex space N of M. The next three lemmas show that π̂
1N
̂
must act with a fixed point on TM ′ . We start with the hyperbolic case.
Lemma 4.4. Consider N a compact, hyperbolic 3-manifold with (possibly
̂
empty) toral boundary. If π̂
1 N acts acylindrically on a profinite tree S with
abelian edge stabilizers then π̂
1 N fixes a unique vertex.
Proof. If N is closed then every abelian subgroup of π̂
1 N is procyclic [WZ17,
Theorem D] and the result follows from Corollary 1.13.
Suppose therefore that N has non-empty toroidal boundary. By Proposition 3.1 every edge stabilizer is either procyclic or conjugate into a peripheral
subgroup, and therefore π̂
1 N fixes a vertex by Lemma 3.4.
Uniqueness follows from [ZM88, Corollary 2.9], since π̂
1 N is non-abelian
and edge stabilizers are abelian.
We move on to the major Seifert fibred case.
Lemma 4.5. Consider N a compact, major Seifert fibred 3-manifold with
(possibly empty) toral boundary. If π̂
1 N acts acylindrically on a profinite tree
Ŝ with abelian edge groups then π̂
N
fixes a unique vertex.
1
̂2
Proof. The subgroups of π̂
1 N which are isomorphic to Z each fix a vertex
by Lemma 3.3. Thus the maximal procyclic normal subgroup C of π̂
1 N fixes
a vertex.
Suppose on the contrary π̂
1 N does not fix a vertex. By [Zal90, Lemma
̂
̂
1.5], there exists a unique minimal π̂
1 N-invariant subtree D of S , which
̂ which
is infinite. Now by [ZM88, Theorem 2.12], C acts trivially on D,
contradicts the acylindricity of the action.
Uniqueness again follows from [ZM88, Corollary 2.9].
13
The case of minor Seifert-fibred vertex follows immediately from [WZ17,
Theorem 5.2] and [ZM88, Corollary 2.9].
Lemma 4.6. Consider N a compact, minor Seifert fibred 3-manifold with
̂
toral boundary. If π̂
1 N acts acylindrically on a profinite tree S with abelian
̂
edge groups then π1 N fixes a unique vertex.
̂ 2 subgroups of π
̂
We next classify the fixed point sets of Z
1 M . First, we
need an analysis of their normalizers in vertex stabilizers. We start with the
hyperbolic case, in which case normalizers coincide with centralizers.
Lemma 4.7. Let N be a compact, orientable, hyperbolic 3-manifold with
̂2 a subgroup of π̂
(H).
(H) = Cπ̂
toral boundary, and H ≅ Z
̂
1 N. Then Nπ
1N
1N
Proof. By [WZ17, Theorem 9.3], H is conjugate into the closure of a cusp
subgroup, and by [WZ17, Lemma 4.5], that cusp subgroup is malnormal.
The result follows.
We next treat the case of a major Seifert fibred manifold.
Lemma 4.8. Let N be a compact, orientable, major Seifert-fibred 3-manifold
̂2 a subgroup of π̂
with toral boundary, and H ≅ Z
1 N conjugate to the closure of
the fundamental group of a boundary component. Then Nπ̂
(H) = Cπ̂
(H).
1N
1N
Proof. The fundamental group π1 N is torsion-free of the form
1 → Z → π1 N → π1 O → 1
where π1 O is a Fuchsian group and Z is infinite cyclic (and not necessarily
central). Since Seifert-fibred 3-manifold groups are LERF [Sco78, Sco85] we
have a corresponding short exact sequence of profinite completions.
f
̂ → π̂
̂
1→Z
1 N → π1 O → 1
̂ and centralizes it. So f (Ĉ (H)) = C ̂ (f (H))
Then Cπ̂
(H) contains Z
π1 N
π1 O
1N
and f (Nπ̂
(H)) = Nπ̂
(f (H)). Hence it suffices to show that Cπ̂
(f (H)) =
1N
1O
1O
Nπ̂
(f (H)). We may assume that H is the closure of the fundamental
1O
group of a boundary component; then f (H) is the closure of a peripheral
infinite-cyclic subgroup C of π1 O. Since Fuchsian groups are conjugacy separable [FR90] we deduce that every-finite index subgroup of π1 O is conjugacy
(f (H)) by [Min12, Corollary 12.3], and
separable. Then Cπ1 O (C) = Cπ̂
1O
Nπ1 O (C) = Nπ̂
(f (H)) by [CZ13, Lemma 2.3 combined with Theorem 2.14].
1O
But Nπ1 O (C) = Cπ1 O (C), so Cπ̂
(f (H)) = Nπ̂
(f (H)) as required.
1O
1O
14
̂2 subgroups of π
̂
Next, we classify the possible fixed subtrees for Z
1M .
̂2 of π
̂
̂
Lemma 4.9. Consider the action of a subgroup H ≅ Z
1 M on TM . One
of the following holds.
(i) The fixed point set of H is a vertex with Seifert-fibred stabilizer.
(ii) The fixed point set of H consists of exactly one edge.
(iii) The fixed subtree of H consists of exactly two edges; the central vertex has a minor Seifert fibred stabilizer, and the other two vertices are
major.
Furthermore, if the centralizer Cπ̂
(H) is properly contained in the normal1M
izer Nπ̂
(H), then we are in case (i) or case (iii).
1M
Proof. By Lemma 3.3 and Proposition 4.2, H fixes a non-empty subtree.
Recall that every edge adjoins a major vertex, and that every minor vertex
adjoins exactly two edges. If H stabilizes an edge e and an adjacent major
vertex v then, by Lemmas 4.7 and 4.8, e is the unique edge incident at v
stabilized by H. It follows that the fixed tree of H is of one of the three
claimed forms.
We now prove that, in case (ii), Nπ̂
(H) = Cπ̂
(H). Indeed, Nπ̂
(H)
1M
1M
1M
preserves the fixed subtree of Cπ̂
(H),
and
so
if
H
fixes
a
unique
edge,
1M
Nπ̂
(H)
is
contained
in
an
edge
stabilizer,
hence is abelian, and so Nπ̂
(H) =
1M
1M
(H).
Cπ̂
1M
We now have enough information to construct a map φ. To start with, it
will only be a map of abstract, unoriented graphs.
Lemma 4.10. Consider closed, orientable, irreducible 3-manifolds M, M ′ ,
̂
̂′
and let f ∶ π
1 M → π1 M be an isomorphism. Then there exists an f equivariant morphism of graphs
φ ∶ T̂M → T̂M ′ .
Note that, here, we only claim that φ is a map of abstract, non-oriented
graphs. This map may in principle send edges to either edges or vertices.
̂
̂
̂ ̂
Proof. For brevity, we write G = π
1 M and S = TM ′ , and let G act on S via
̂ 2 . Let u1 , u2 be the adjacent
f . Let e be an edge of T̂M with stabilizer Ge ≅ Z
15
vertices of T̂M . Lemmas 4.4, 4.5 and 4.6 together guarantee the existence of
unique vertices v1 , v2 of Ŝ such that Gui ⊆ Gvi for both i.
We claim that v1 and v2 are either equal or adjacent. Suppose therefore
that v1 , v2 are at distance greater than 2 (possibly infinite). Then Ge stabilizes the geodesic [v1 , v2 ] (see [ZM88, Cor 2.9]) and therefore, by Lemma 4.9,
v1 , v2 are at distance precisely two, are both adjacent to a minor vertex w,
and CG (Ge ) is properly contained in NG (Ge ). Therefore, by Lemma 4.9, e is
adjacent to a minor vertex; without loss of generality, we may assume that
u1 is major, Gu1 ⊆ Gv1 , and that u2 is minor and Gu2 ⊆ Gv2 . But Gw also
normalizes Ge , so Gw ⊆ Gu2 ⊆ Gv2 . This implies that Gw stabilizes an edge,
which is absurd because edge stabilizers are abelian. Therefore, v1 and v2
are either equal or adjacent. If they are equal to a vertex v, we set φ(e) = v.
If they are adjacent, we set φ(e) to be the image of the unique edge joining
̂ which is
them. This completes the construction of the map φ ∶ T̂M → S,
equivariant by construction.
We are now ready to complete the proof of the main theorem.
Proof of Theorem 4.3. Applying Lemma 4.10 twice, we obtain maps of graphs
φ ∶ T̂M → T̂M ′ , ψ ∶ T̂M ′ → T̂M ,
where φ is f -equivariant and ψ is f −1 -equivariant. Equivariance implies that
gψ ○ φ(x) = ψ ○ φ(gx)
̂
̂
for all g ∈ π
1 M and x ∈ TM , whence the stabilizer of x is contained in the
stabilizer of ψ ○ φ(x). Since vertex-stabilizers stabilize unique vertices, it
follows that ψ○φ is equal to the identity on vertices, and hence on the whole of
T̂M . In particular, φ and ψ induce isomorphisms of the finite quotient graphs
̂
̂
̂ ′ ̂′
π
1 M/TM and π1 M /TM ; we may therefore choose consistent orientations on
these graphs, which lift to equivariant orientations on the profinite trees T̂M
and T̂M ′ , which are respected by φ and ψ; this also implies continuity of φ
and ψ.
References
[AFW15]
Matthias Aschenbrenner, Stefan Friedl, and Henry Wilton. 3manifold groups. Series of Lecture Notes in Mathematics. European Mathematical Society (EMS), Zürich, 2015.
16
[Ago13]
Ian Agol. The virtual Haken conjecture. Doc. Math., 18:1045–
1087, 2013. With an appendix by Agol, Daniel Groves, and Jason
Manning.
[BCR16]
M. R. Bridson, M. D. E. Conder, and A. W. Reid. Determining Fuchsian groups by their finite quotients. Israel J. Math.,
214(1):1–41, 2016.
[BF15]
Michel Boileau and Stefan Friedl. The profinite completion
of 3-manifold groups, fiberedness and the Thurston norm.
arXiv:1505.07799, 2015.
[BR15]
Martin R. Bridson and Alan W. Reid. Profinite rigidity, fibering,
and the figure-eight knot. arXiv:1505.07886v1, 2015.
[BRW16]
Martin R. Bridson, Alan W. Reid, and Henry Wilton. Profinite
rigidity and surface bundles over the circle. arXiv:1610.02410,
2016.
[Cav12]
Will Cavendish. Finite-Sheeted Covering Spaces and Solenoids
over 3-manifolds. ProQuest LLC, Ann Arbor, MI, 2012. Thesis
(Ph.D.)–Princeton University.
[CB13]
Owen Cotton-Barratt. Detecting ends of residually finite groups
in profinite completions. Math. Proc. Cambridge Philos. Soc.,
155(3):379–389, 2013.
[CZ13]
S. C. Chagas and P. A. Zalesskii. Hereditary conjugacy separability of free products with amalgamation. J. Pure Appl. Algebra,
217(4):598–607, 2013.
[DD89]
Warren Dicks and M. J. Dunwoody. Groups acting on graphs,
volume 17 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 1989.
[Dic80]
Warren Dicks. Groups, trees and projective modules, volume 790
of Lecture Notes in Mathematics. Springer, Berlin, 1980.
[FR90]
Benjamin Fine and Gerhard Rosenberger. Conjugacy separability of fuchsian groups and related questions. In Combinatorial
group theory (College Park, MD, 1988), volume 109 of Contemp.
Math., pages 11–18. Amer. Math. Soc., Providence, RI, 1990.
17
[Fun13]
Louis Funar. Torus bundles not distinguished by TQFT invariants. Geom. Topol., 17(4):2289–2344, 2013. With an appendix
by Funar and Andrei Rapinchuk.
[GJZZ08]
F. Grunewald, A. Jaikin-Zapirain, and P. A. Zalesskii. Cohomological goodness and the profinite completion of Bianchi groups.
Duke Math. J., 144(1):53–72, 2008.
[GPS80]
F. J. Grunewald, P. F. Pickel, and D. Segal. Polycyclic groups
with isomorphic finite quotients. Ann. of Math. (2), 111(1):155–
195, 1980.
[Hem14]
John Hempel. Some 3-manifold groups with the same finite quotients. arXiv:1409.3509, 2014.
[HWZ13]
Emily Hamilton, Henry Wilton, and Pavel A. Zalesskii. Separability of double cosets and conjugacy classes in 3-manifold
groups. J. Lond. Math. Soc. (2), 87(1):269–288, 2013.
[KR88]
P. H. Kropholler and M. A. Roller. Splittings of Poincaré duality
groups. II. J. London Math. Soc. (2), 38(3):410–420, 1988.
[KZ08]
Dessislava H. Kochloukova and Pavel A. Zalesskii. Profinite and
pro-p completions of Poincaré duality groups of dimension 3.
Trans. Amer. Math. Soc., 360(4):1927–1949, 2008.
[Min12]
Ashot Minasyan. Hereditary conjugacy separability of rightangled Artin groups and its applications. Groups Geom. Dyn.,
6(2):335–388, 2012.
[NS07]
Nikolay Nikolov and Dan Segal. On finitely generated profinite
groups. I. Strong completeness and uniform bounds. Ann. of
Math. (2), 165(1):171–238, 2007.
[RZ10]
Luis Ribes and Pavel Zalesskii. Profinite groups, volume 40 of
Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A
Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys
in Mathematics]. Springer-Verlag, Berlin, second edition, 2010.
18
[Sco78]
Peter Scott. Subgroups of surface groups are almost geometric. Journal of the London Mathematical Society. Second Series,
17(3):555–565, 1978.
[Sco85]
Peter Scott. Correction to: “Subgroups of surface groups are
almost geometric”. J. London Math. Soc. (2), 32(2):217–220,
1985.
[Ser97]
Jean-Pierre Serre. Galois cohomology. Springer-Verlag, Berlin,
1997. Translated from the French by Patrick Ion and revised by
the author.
[Str77]
R. Strebel. A remark on subgroups of infinite index in Poincaré
duality groups. Comment. Math. Helv., 52(3):317–324, 1977.
[SW00]
Peter Symonds and Thomas Weigel. Cohomology of p-adic analytic groups. In New horizons in pro-p groups, volume 184 of
Progr. Math., pages 349–410. Birkhäuser Boston, Boston, MA,
2000.
[Uek17]
Jun Ueki. The profinite completions of knot groups determine
the Alexander polynomials. arXiv:1702.03836, 2017.
[Wil16]
Gareth Wilkes. Profinite detection of JSJ decompositions of 3manifolds. arXiv:1605.08244, 2016.
[Wil17]
Gareth Wilkes. Profinite Properties of 3-Manifold Groups. PhD
thesis, University of Oxford, 2017.
[Wilar]
Gareth Wilkes. Profinite rigidity for Seifert fibre spaces. Geom.
Dedicata, to appear.
[Wis12]
D. T. Wise. The structure of groups with a quasi-convex hierarchy. Preprint, http://goo.gl/3ctNvX, April 2012.
[WZ10]
Henry Wilton and Pavel Zalesskii. Profinite properties of graph
manifolds. Geometriae Dedicata, 147(1):29—45, 2010.
[WZ17]
Henry Wilton and Pavel Zalesskii. Distinguishing geometries
using finite quotients. Geom. Topol., 21(1):345–384, 2017.
19
[Zal89]
P. A. Zalesskii. Geometric characterization of free constructions
of profinite groups. Akademiya Nauk SSSR. Sibirskoe Otdelenie.
Sibirskii Matematicheskii Zhurnal, 30(2):73–84, 226, 1989.
[Zal90]
P. A. Zalesskii. Profinite groups, without free nonabelian
pro-p-subgroups, that act on trees. Matematicheskii Sbornik,
181(1):57–67, 1990.
[Zal92]
P. A. Zalesskiı̆. Open subgroups of free profinite products. In
Proceedings of the International Conference on Algebra, Part 1
(Novosibirsk, 1989), volume 131 of Contemp. Math., pages 473–
491. Amer. Math. Soc., Providence, RI, 1992.
[ZM88]
P. A. Zalesskii and O. V. Mel’nikov. Subgroups of profinite groups acting on trees.
Matematicheskii Sbornik,
135(177)(4):419—439, 1988.
[ZM89]
P. A. Zalesskii and O. V. Mel’nikov. Fundamental groups of
graphs of profinite groups. Algebra i Analiz, 1(4):117—135, 1989.
20
| 4math.GR
|
Generating Contradictory, Neutral, and Entailing Sentences
Yikang Shen * 1 Shawn Tan * 1 Chin-Wei Huang * 1 Aaron Courville 1
arXiv:1803.02710v1 [cs.CL] 7 Mar 2018
Abstract
Learning distributed sentence representations remains an interesting problem in the field of Natural Language Processing (NLP). We want to learn
a model that approximates the conditional latent
space over the representations of a logical antecedent of the given statement. In our paper,
we propose an approach to generating sentences,
conditioned on an input sentence and a logical
inference label. We do this by modeling the different possibilities for the output sentence as a
distribution over the latent representation, which
we train using an adversarial objective. We evaluate the model using two state-of-the-art models
for the Recognizing Textual Entailment (RTE)
task, and measure the BLEU scores against the
actual sentences as a probe for the diversity of sentences produced by our model. The experiment
results show that, given our framework, we have
clear ways to improve the quality and diversity of
generated sentences.
1. Introduction
Algorithms designed to learn distributed sentence representations have been shown to be transferable across a range of
tasks (Mou et al., 2016) and languages (Tiedemann, 2018).
For example, Guu et al. (2017) proposed to represent sentences as vectors that encode a notion similarity between
sentence pairs, and showed that vector manipulations of the
representation can result in meaningful change in semantics.
The question we would like to explore is whether the semantic relationship between sentence pairs can be modeled in a
more explicit manner. More specifically, we want to model
the logical relationship between sentences.
Controlling the logical relationship between sentences has
many direct applications. First of all, we can use it to provide a more clear definition of paraphrasing. To do so, we
*
Equal contribution 1 Department of Computation, Université
de Montréal, Montréal, Canada. Correspondence to: Yikang Shen
<[email protected]>, Shawn Tan <[email protected]>, ChinWei Huang <[email protected]>.
require two simultaneous conditions: (i) that the input sentence entails the output sentence; and (ii) that the output
sentence entails the input sentence.
(S ENTENCE 1 |= S ENTENCE 2) ∧
(S ENTENCE 2 |= S ENTENCE 1)
(1)
The first requirement ensures the output sentence cannot be
false if the input sentence is true, so that the output sentence
can be considered a fact expressed by the input sentence.
The second requirement ensures that the output contains
at least the input’s information. The two requirements together can be used to define semantic equivalence between
sentence.
Another interesting application is multi-document summarization. Traditionally, to summarize multiple documents,
one would expect the model to abstract the most important
part of the source documents, and this is usually measured
by the amount of overlap that the output document has with
the inputs. Informally, one finds the maximal amount of
information that has the highest precision with each source
document. Alternatively, if one wants to automate news
aggregation, the ideal summary would need to contain the
same number of facts as are contained in the union of all
source documents. We can think of this second objective
as requiring that the output document entail every single
sentence across all source documents.
In this paper, we propose an approach to generating sentences, conditioned on an input sentence and a logical inference label. We do this by modeling the different possibilities
for the output sentence as a distribution over the latent representation, which we train using an adversarial objective.
In particular, we differ from the usual adversarial training on text by using a differentiable global representation.
Architecture-wise, we also propose a Memory Operation
Selection Module (MOSM) for encoding a sentence into a
vector representation. Finally, we evaluate the quality and
the diversity of our samples.
The rest of the paper is organized as follows: Sec. 2 will
cover the related literature. Sec. 3 will detail the proposed
model architecture, and Sec. 4 will describe and analyze the
experiments run. Sec. 5 will then discuss the implications
of being able to solve this task well, and the future research
directions relating to this work. Finally, we conclude in Sec.
Generating Contradictory, Neutral, and Entailing Sentences
6.
η
z
φ
2. Related Work
Many natural language tasks require reasoning capabiliities.
The Recognising Textual Entailment (RTE) task requires the
system to determine if the premise and hypothesis pair are
(i) an entailment, (ii) contradicting each other or (iii) neutral
to each other. The Natural language Inference (NLI) Task
from Bowman et al. (2015a) introduces a large dataset with
labeled pairs of sentences and their corresponding logical
relationship. This dataset allows us to quantify how well
current systems are able to be trained to recognise sentences
with those relationships. Examples of the current state-ofthe-art for this task include Chen et al. (2017) and Gong
et al. (2017).
Here we are interested in generating natural language that
satisfies the given textual entailment class. Kolesnyk et al.
(2016) has attempted this using only sentences from the
entailment class, and focusing on generating a hypothesis
given the premise. Going in this direction results in removal
of information from the premise sentence. In this paper, we
focus on going in the other direciton: generating a premise
from a hypothesis. This requires adding additional details to
the premise which have to make sense in context. In order to
produce sentences with extra details and without some other
details, we suggest that a natural way to model this kind of
structure is to impose a distribution over an intermediate
distribution representing the semantic space of the premise
sentence.
In the realm of learning representations for sentences, Kiros
et al. (2015) has a popular method for learning representations called “skip-thought” vectors. These are trained by
using the encoded sentence to predict the previous and next
sentence in a passage. Conneau et al. (2017) specifically
learned sentence representations from the SNLI dataset.
They claim that using the supervised data from SNLI can
outperform “skip-thought” representations on different tasks.
There have also been several efforts towards learning a distribution over sentence embeddings. Bowman et al. (2015b)
used Variational Autoencoders (VAEs) to learn Gaussian
distributed word embeddings. Hu et al. (2017) use a combined VAE/GAN objective to produce a disentangled representation that can be used to modify some attributes like
sentiment and tense.
There have also been forays into conditional distributions
for sentences – which is what is required here. Both Gupta
et al. (2017) and Guu et al. (2017) introduce models of the
form p(x|z, x0 ), where x is a paraphrase of x0 , and z represents the variability in the output sentence. Guu et al. (2017)
introduces z as an edit vector. However, because z has to be
paired with x0 in order to generate the sentence, z serves a
`
Figure 1. The conceptual graphical model behind the formulation
of our model. The red arrow represents the inference path from φ
to z.
very different purpose, and cannot be considered a sentence
embedding in its own right. Ideally, what we want is a distribution over sentence representations, each one mapping
to a set of semantically similar sentences. This is important
if we want the distribution to model the possibilities of concepts that correspond to the right textual entailment with the
hypothesis.
3. Method
Some approaches map a sentence to a distribution in the
embedding space (Bowman et al., 2015b). The assumption
when doing this is that there is some uncertainty over the
latent space when mapping from the sentence. Some approaches, like Hu et al. (2017) attempt to disentangle factors
in the learnt latent variable space, so that modifying each
dimension in the latent representation modifies sentiment or
tense in the original sentence.
If we consider plausible premise sentences φ given a hypothesis η and an inference label `, there are many possible solutions, of varying likelihoods. We can model this
probabilistically as p(φ|η, `). In our model, we assume an
underlying latent variable z that accounts for the variation
in possible output sentences,
Z
p(φ|η, `) =
p(φ|z)p(z|η, `)dz
Another assumption we make is that given φ, z is independent of η and `. The resulting graphical model associated
with the above dependency assumptions are depicted in
Figure 1.
In our proposed model, we take inspiration from the Adversarial Autoencoder (Makhzani et al., 2015), however
our prior is conditioned on η and `. Zhang et al. (2017)
also proposed a Conditional Adversarial Autoencoder for
age progression prediction. In addition to the adversarial
discriminator, our model includes a classifier on the representation and the hypothesis and label. A similar framework
is also discussed in Salimans et al. (2016).
Generating Contradictory, Neutral, and Entailing Sentences
Prior We draw a sample, conditioned on (η, `), through
the prior, which is described using following equations:
hη1 , ..., hη|η|
=
RNNenc (xη1 , ..., xη|η| )
(7)
h̃t
=
MLP([hηt , e` , ])
(8)
ĥ1 , ..., ĥ|η|
=
RNNrefine (h̃1 , ..., h̃|η| )
(9)
z
Figure 2. The architecture of the model. The autoencoder maps
given premise φ to a sentence representation z, and reconstructs φ
from z. Samples are drawn from the prior conditioned on η and `.
The classifier takes z and η as input, and outputs probability of l.
The discriminator takes z, η and l as input, and predicts whether z
is given by the autoencoder or the prior.
= fcompress (ĥ1 , ..., ĥ|η| )
(10)
where is a random vector, i ∼ N (0, 1); e` is the label
embedding and [·, ·] represents the concatenation of input
vectors.
Classifier This outputs the probability distribution over
labels, taking as input the tuple (z, η), and is described
using the following equations:
hη1 , ..., hη|η|
=
RNNenc (xη1 , ..., xη|η| )
(11)
3.1. Architecture
ct
=
The model consists of an encoder q(z|φ), a conditional prior,
p(z|η, `), a decoder p(φ|z), and a discriminator D(z, η, `).
h̃t
=
fretrieve (z, hηt )
MLP([hηt , ct , ||hηt
ĥ1 , ..., ĥ|η|
=
RNNrefine (h̃1 , ..., h̃|η| )
(14)
ĥmax
=
Poolingmax (ĥ1 , ..., ĥ|η| )
(15)
ĥmean
=
Poolingmean (ĥ1 , ..., ĥ|η| )
(16)
Autoencoder The autoencoder comprises of two parts.
An encoder that maps the given premise φ to a sentence
representation z, and a decoder that reconstructs φ from a
given z. In our model, the encoder reads the input premise
φ
φ = (xφ
1 , ..., x|φ| ) using an RNN network:
φ
φ
φ
hφ
1 , ..., h|φ| = RNNenc (x1 , ..., x|φ| )
(2)
φ
z = fcompress (hφ
1 , ..., h|φ| )
(3)
and
where ht ∈ Rn is a hidden state at time t. z is a vector
generated from sequence of the hidden states. We will call
fcompress (·) the compression function.
The decoder is trained to predict the next word x0t given
the sentence representation z and all the previously predicted words (x01 , ..., x0t−1 ). With an RNN, the conditional
probability distribution of x0t is modeled as:
p(x0t |x01 , ..., x0t−1 , z) = g(st , ct )
(4)
and
s1 , ..., s|φ| = RNNdec (x01 , ..., x0|φ| )
(5)
ct = fretrieve (z, st )
(6)
where g(·) is a nonlinear, potentially multi-layered, function
that outputs the probability of x0t , st is the hidden state of
decoder RNN, and fretrieval takes st as the key to retrieve
related information from z. We note that other architectures
such as a CNN or a transformer (Vaswani et al., 2017) can
be used in place of the RNN. The details of the compression
function and retrieval function will be discussed in Sec. 3.2.
p(`|z, η)
(12)
− ct ||, hηt
= σ(MLP([ĥmax , ĥmean ]))
ct ])
(13)
(17)
where Pooling(·) refers to an element-wise pooling operator, and the activation function σ for output layer is the
softmax function. The architecture of the classifier is inspired by (Chen et al., 2017). Instead of doing attention over
the sequence of hidden states for the premise, we use the
retrieval function in Equation 12 to retrieve related information ct in z for hηt .
Discriminator The discriminator takes as input (z, η, `),
and tries to determine if the z in question comes from the
encoder or prior. The architecture of the discriminator is
similar to that of the classifier, with the exception that Equation 13 is replaced by:
h̃t
=
MLP([hηt , ct , e` ])
(18)
to pass label information to the discriminator. The sigmoid
function is used as the activation for the output layer.
In our model the autoencoder, prior and classifier share the
same RNNenc (·) parameters. The prior and the autoencoder
share the same fcompress (·) parameters. The classifier and
the autoencoder share the same fretrieve (·) parameters. The
discriminator does not share any parameters with the rest of
model.
3.2. Compression and Retrieval Functions
The compression (Equation 3) and retrieval (Equation 6)
functions can be modeled through many different mechanisms. Here, we introduce two different methods:
Generating Contradictory, Neutral, and Entailing Sentences
Retrieval functions use {st } as control vectors to retrieve
information from z. Since the layer generates a different
weight matrix for the feedforward path for different st , we
can output different o for the same z.
3.3. Model Learning
Like most adversarial networks, the conditional adversarial
autoencoder is trained with a gradient descent based method
in two phases: the generative phase and the discriminative
phase.
Figure 3. Memory Operation Selection Module takes a pair of vector (k, v) as input, output a vector o. k provide the control signal
for the layer to compute a weighted sum of candidate weight matrices. The obtained matrix is used as the weight matrix in a normal
feedforward layer, that takes v as input and outputs o.
Mean Pooling can be used to compress the sequence of
the hidden states:
fcompress (h1 , ..., hT ) =
T
1X
ht
T t=1
Lauxiliary =
(20)
Memory Operation Selection Module (MOSM) As an
alternative to mean pooling, we use the architecture shown
in Figure 3. A layer is defined as:
γ
W̃
=
=
softmax(Ωk)
N
W
X
γi Wi
(21)
(22)
i=1
o = σ(W̃v)
In our initial experiments, we found that the samples from
just the adversarial training alone results in wildly varied
output sentences. To ameliorate this, we propose an auxiliary loss:
(19)
and its retrieve counterpart directly returns z:
fretrieve (z, st ) = z
In the generative phase, the autoencoder is updated to minimize the reconstruction error of the premise. The classifier
and the encoder are updated to minimize the classification
error of the premise-hypothesis pair. The prior is also updated to optimize the classification error of p(`|z, η), where
z is draw from the prior. The encoder and the prior are
updated to confuse the discriminator.
min
{NLL(φ|zi )} ,
i∈(1,...,N )
zi ∼ p(z|η, `)
(26)
where N is the number of samples that are drawn from
prior. The auxiliary loss measures how far our generated
premises are from the true premise when conditioned on the
hypothesis and label. As shown in experiment the model
has better generating diversity, while more samples were
drawn during training.
One can view this auxiliary loss as a ‘hard’ version of taking the log average of the probability of N Monte-Carlo
samples,
(23)
where σ can be any activation function, v is the input vector,
k is the control vector, {Wi } are NW candidate weight
matrices. For convenience, we denote the MOSM function
as fMOSM (v, k).
− log Ep(z|η,`) [p(φ|z)]
≈ − log
N
1 X
p(φ|zi ),
N i
(27)
zi ∼ p(z|η, `)
(28)
N
Thus, we can define the MOSM compression method as:
!
T
1X
fMOSM (ht , ht )
fcompress (h1 , ..., hT ) = tanh
T t=1
(24)
The compression function uses {ht } as both control and
input vector, to write themselves into z. Because different
ht s select different combinations of candidate matrices, we
can have different mapping function each different ht at
each time step.
fretrieve (z, st ) = fMOSM (z, st )
(25)
X
1
= − log
− log
exp log p(φ|zi )
N
i
(29)
1
+ min − log p(φ|zi )
N i∈(1,...,N )
(30)
≤ − log
Since log N1 is a constant, minimizing over the Equation 30
is the same as minimizing Equation 26.
In the discriminative phase, the discriminator is updated
to tell apart the true z (generated using the prior) from the
generated samples (given by autoencoder).
Generating Contradictory, Neutral, and Entailing Sentences
4. Experiments
We use the Stanford Natural Language Inference (SNLI)
corpus (Bowman et al., 2015a) to train and evaluate our
models. From our experiments, we want to determine two
things. First, do the sentences produced by the model form
the correct textual entailment class on which it was conditioned on? Second, is there diversity among the sentences
that are generated?
Table 1. Classification accuracies for different state-of-the-art models on our samples. The row labeled R ANDOM we randomly permuted the premises of the original test set and ran them through
the classifiers to test for the models’ reliance on just the hypothesis
for classification.
4.1. Baseline Methods
For comparison, we use a normal RNN encoder-decoder as
a baseline method. The model uses a bidirectional LSTM
network as encoder. The encoder reads the input hypothesis
into a sequence of hidden states {ht }:
hη1 , .., hη|η|
zη
=
RNNenc (xη1 , .., xη|η| )
(31)
=
fcompress (hη1 , .., hη|η| )
(32)
Where fcompress (·) can be the mean method or an MOSM.
The distributed representation of label e` and zη are concatenated together to feed into a normal MLP network, which
output the sentence representation z:
z = MLP([zη , e` ])
(33)
The decoder compute the conditional probability distribution with equations:
p(x0t |x01 , ..., x0t−1 ) = g([st , z])
(34)
RNNdec (x0t−1 , st−1 )
(35)
st =
Thus, the baseline model share a similar architecture with
prior and decoder in our model, while the randomness been
toke out.
4.2. Experiment Settings
For all models, RNNenc and RNNrefine are 2-layers bidirectional LSTM (Hochreiter & Schmidhuber, 1997),
RNNdec are 2-layers uni-directional LSTM. The dimension of hidden state, embeddings and latent representation
z are 300. When training, optimization is performed with
Adam using learning rate lr = 0.001, β1 = 0, β2 = 0.999
and σ = 10−8 . We carry out gradient clipping with maximum norm 1.0. We train each model for 30 epoch. For each
iteration, we randomly choose to run the generative phase
or discriminative phase with probability 0.5 : 0.5. Since we
didn’t observe significant benefit from using Beam Search,
all premises are generated using greedy search.
4.3. Quality Evaluation
In order to evaluate the quality of the samples from our
model, we trained two state-of-the-art models for SNLI: (1)
M ODEL
DIIN
ESIM
R ANDOM
42.7%
41.1%
BASELINE (M EAN )
BASELINE (MOSM)
59.6%
62.7%
59.6%
62.6%
MOSM (N=1, - CLASSIFIER )
MOSM (- AUXILIARY LOSS )
67.2%
63.2%
67.3%
60.6%
M EAN (N=1)
M EAN (N=10)
MOSM (N=1)
MOSM (N=10)
64.4%
64.3%
76.1%
72.6%
62.4%
62.3%
75.9%
71.8%
Densely Interactive Inference Network (DIIN)1 (Gong et al.,
2017), (2) Enhanced Sequential Inference Model (ESIM)2
(Chen et al., 2017).
In our experiments, we found that it is possible to achieve
an accuracy of 68% on SNLI label prediction by training a
classifier using only the hypothesis as input. This calls into
question how much the classification models rely on just
the hypothesis for performing its task. To investigate this
phenomena further, we randomly permuted the premises of
the original test set and passed these new (random) permishypothesis pairs to the classifiers. The results are shown in
the row labelled R ANDOM in Table 1. We were satisfied
that at 42.7% and 41.1%, the classification models (both
DIIN and ESIM) were not relying entirely on the hypothesis
for prediction.
We sampled 9845 hypotheses from the test set, and produced
φ for each example with the given `. The (η, φ, `) triplet
was then passed to the classifiers and evaluated for accuracy.
Both classification models perform at ∼88% accuracy, but,
while they were not perfect, they provided a good probe for
how well our models were generating the required sentences.
Table 1 shows the accuracy of prediction on the respective
models. Both the DIIN and ESIM models give similar
results.
Our results show that using the MOSM gives an improvement over just taking the mean. Using the adversarial training also results in some gains, which suggests that training
the model with the ‘awareness’ of the distribution over the
representation space results in better quality samples. Us1
https://github.com/YichenGong/Densely-InteractiveInference-Network
2
https://github.com/lukecq1231/nli
Generating Contradictory, Neutral, and Entailing Sentences
Table 2. The confusion matrix for the samples from the best model
MOSM (N = 1)
L ABEL \P RED .
E NTAILMENT
N EUTRAL
C ONTRADICTION
E NT.
N EUT.
C ONT.
67.8%
6.6%
2.9%
20.9%
76.7%
12.8%
11.4%
16.7%
84.4%
ing the adversarial training in conjunction with the MOSM
layer gives us the model with the best performance. We also
performed ablation tests, removing certain components of
the model from the training to see how it affects the quality
of samples. The difference between our best model against
MOSM (N = 1, - CLASSIFIER ) suggests that the classifier
plays in important role in ensuring z is a representation in
the right class. In our experiment removing the auxiliary
loss, we still achieve an accuracy ∼61%. However, looking at the samples for this iteration of the model, while
having some concepts in common with the hypothesis, the
sentences in general are more nonsensical in comparison
to those trained with the auxiliary loss (See an example in
Figure 6).
The confusion matrix produced when evaluating our best
model (MOSM, N = 1) on DIIN shows us where the
classification model and our generative model agree (See
Table 2). In our R ANDOM experiments, we find that the
model has a bias towards predicting contradictions. This is
observed here as well, with contradictions being the category
with the highest agreement. We therefore cannot conclude
that contradictions are easier for our model to generate.
Also, using the original test set, the category in which DIIN
performs the best is entailment, with a precision of 89.1%
compared to 84.3% for neutral and 88.4% for contradiction.
This suggests that generating suitable premises that entail
the hypothesis is the hardest task for the model.
We also want to study how the classifier component of our
model affects the generation of good samples. As shown
in Figure 4, “Z precision” is higher then 0.9. This suggests
that the classifier provides a strong regularization signal to
the sentence representation z. Because the autoencoder is
not perfect, we do not observe the the same sample classification precision after z is decoded. However, we still
observe a synchronous improvement of both sample and
valid precision. It is therefore reasonable to expect that a
better classifier and a better autoencoder would result in
better generated premises.
4.4. Diversity Evaluation
In order to evaluate the diversity of samples given by our
model, we compute the BLEU score between to premises
Figure 4. Different classification precisions given by our classifier
in our model (MOSM, N=10) during training. Sample Precision
shows the probability that classifier predicts correct label for generated premise and related real hypothesis. Valid precision shows
the probability that classifier predicts correct label for real premise
and real hypothesis. Z precision shows the probability that the
feedforward network fclassifier (z, zη ) predicts correct label `, for
given η, ` and z drawn from prior p(z|η, `).
generated conditioned on the same hypothesis and label.
In other words, given a triple (φi , η i , `i ) from test set, we
draw two different samples (zi1 , zi2 ) from the prior distribution p(z|η i , `i ). Then the decoder generates two premises
(φi1 , φi2 ) using greedy search conditioned on (zi1 , zi2 ) respectively. The similarity score between generated premises
is then estimated by:
BLEUi =
1
(BLEU(zi1 , zi2 ) + BLEU(zi2 , zi1 )) . (36)
2
For comparison, we also compute the BLEU score between
real premise and generated premise (φi , φi1 ). The average
of diversity score between two generated premises is noted
as BLEUSS , the one between real and generated premises
is noted as BLEURS . Since it is not necessary have n-gram
match between premises, BLEU score can be inaccurate on
some data points. We employ the Smoothing technique 2
described in Chen & Cherry (2014).
As shown in Table 3, when we increase the number of
samples N in the auxiliary loss, the diversity of samples
increases for both mean pooling and MOSM. This can serve
as empirical evidence that the diversity of our model can
be controlled by choosing a different hyper-parameter N .
The higher BLEURS given by MOSM method could be
interpreted as real premise is more close to the center of
mass of prior distribution. We also observe a gap between
BLEURS and BLEUSS . The gap shows that the sampled
premise is still relatively similar between themselves. After
Generating Contradictory, Neutral, and Entailing Sentences
Table 3. BLEU score for different models
S AMPLES FROM MOSM (N=10)
H: a worker stands over a bread display .
L: Entailment
S1: a man in a blue shirt is preparing food in a kitchen .
S2: a man in a blue shirt is washing a window .
BLEURS
BLEUSS
BASELINE (M EAN )
BASELINE (MOSM)
14.4
14.7
N/A
N/A
MOSM (N=1, - CLASSIFIER )
MOSM (- AUXILIARY LOSS )
14.4
10.3
46.7
14.8
H: there is a jockey riding a horse .
L: Entailment
S1: a horse rider on a bucking horse .
S2: a jockey riding a horse in a rodeo .
M EAN (N=1)
M EAN (N=10)
MOSM (N=1)
MOSM (N=10)
11.9
11.3
14.2
13.2
27.9
17.3
38.9
22.5
H: a man sitting on the couch reading a book .
L: Contradiction
S1: a man is sitting on a bench with his hands in his pockets .
S2: a man in a blue shirt is standing in front of a store .
M ODEL
H: a baby in his stroller outside .
L: Contradiction
S1: a woman is sitting on a bench next to a baby .
S2: a woman is sitting on a bench in a park .
H: the man is being watched .
L: Neutral
S1: a man jumps from a bridge for an elderly couple at a beach .
S2: a man in a blue shirt is standing in front of a building .
H: there is a human selling hot dogs .
L: Neutral
S1: a person is standing in front of a food cart .
S2: a woman in a white shirt is standing in front of a counter selling
food .
S AMPLES FROM MOSM (- AUXILIARY LOSS )
H: a restaurant prepares for a busy day .
L: Neutral
S1: a pink teenager prepares on a tune on the roots .
S2: a UNK restaurant dryer for a canvas .
Figure 5. Visualization of the effect of auxiliary loss with multiple
samples. For a pair of (φ, η), we repeat 100 times the process
of compute auxiliary loss (N=10) in Equation 26. Blue points
represent zi selected by minimum function, green points represent
zi that are not selected. Our model (MOSM, N=10) is used for
computing z and perplexities. t-SNE is used to visualize highdimensional data (Maaten & Hinton, 2008).
removing the classifier, we observe an increase in BLEUSS .
One possible explanation is that classifier prevents the prior
from overfitting the training data. We observe an decrease in
both BLEU scores, after removing the auxiliary loss. However, Table 1 and Figure 6 shows that removing auxiliary
loss give low quality samples.
While the auxiliary loss is essential for the prior and the
decoder to learn to cooperate, using an auxiliary loss where
(N = 1) will collapse the prior distribution; instead of a
distribution, the prior will learn to ignore the random input
and deterministically predict z. As shown in Figure 5, the
auxiliary loss (N = 10) only passes gradients to the zs in
the left region of the distribution. As a result, samples drawn
from right region have a significant lower chance receive
gradient from decoder, while the entire region receives gra-
Figure 6. Example sentence generated by our model (MOSM,
N=10). H is the hypothesis, L is the label, S1 is the first sample,
and S2 is the second sample. The samples shown below the line
are drawn from a model trained without the auxiliary loss.
dients from the discriminator and classifier. Therefore, the
prior distribution can expand to more regions, but only those
regulated by discriminator and classifier. This will increase
the diversity of samples. However, we also observe that the
precision slightly decreases in Table 1. This suggests that
the discriminator and classifier in our model are not perfect
for regularizing the prior distribution.
4.5. Samples
Figure 6 shows several examples generated by our model
(MOSM, N=10). These example shows that our model
can generate a variety of different premise while keep the
correct semantic relation. Some of subjects in hypothesis
are correctly replace by synonyms (e.g. “jocky” is replaced
by “horse rider”, “human” is replaced by “person” and
“woman”). The model also get some potential logical relation correct (e.g. “reading a book” is contradicted by “with
Generating Contradictory, Neutral, and Entailing Sentences
his hands in his pockets”, “stands over a bread display” can
either means “washing a window” or “preparing food in a
kitchen”).
However, we also observe that the model tries to add “a blue
shirt” for most “man”s in the sentences, which is one of the
easiest way to add extra information into the model. The
phenomenon aligned with well-known model collapse failure case for most adversarial training based method. This
observation give an explanation for the relatively higher
BLEU between sample. The model also have some bias
while generating premise (e.g. when hypothesis mention
“a baby”, the premise automatically mention “a woman”),
which aligns with the recent discovery that visual recognition tasks model tend to output biased predictions (Zhao
et al., 2017).
5. Discussion
The broader vision of our project is to attain logical control
for language, which we believe will allow us to perform better across many natural language applications. This is most
easily achieved at the word-level, by adding or removing
specific words to a sentence, using word generation rules
based on language-specific grammars. However, just as
distributed word representations can be meaningfully combined (Mikolov et al., 2013) with good outcomes, we believe
that sentence-level representations are the way forward for
manipulation of text.
The kind of control we seek to model, specifically, is characterized by the logical relationships between sentence pairs.
Controlling semantic representation by modeling logical
relationship between the input and output sentences has
many potential use cases. Returning to the task of multidocument summarization discussed in the introduction, operating in the semantic space allows one to abstract the
information of a document. Controlling the logical relationships among sentences provides a new way to think about
what a summary is. Ideally, when multiple sources of information are given, we would like the output summary φ
generated by a machine to be entailable by the union of
inputs (∪j∈J η j ) |= φ 3 . This addresses the problem of
precision: the resulting summary now has a subset of the information available in the union of all the given hypotheses.
To address the problem of recall, we need the resulting
summary to entail each one of the individual hypotheses:
∧i (φ |= η j ) Together, these two criteria form a clear formal
definition for multi-document summarization,
{ φ : ∧i (φ |= η j ) ∧ (∪j∈J η j ) |= φ }
which represents the set of all possible φ that fit the criteria.
3
Here we assume there are no conflicting details.
In our paper, we toyed with the possibility of modeling the
set { φ : φ |= η } by training a model with a distribution
over different premises in the latent space z. A good subsequent step would be modelling the first part of our logical
description of multi-document summarisation,
{ φ : ∧i∈J (φ |= η j ) } = ∩i∈J { φ : φ |= η j }
This suggests a possible avenue for producing such a
premise is finding the intersection of the distribution over z
for two given hypotheses that are likely enough to occur.
Future work can explore the possibility of this and determining the union of the hypotheses entailing the given premise.
6. Conclusion
We have proposed a model that generates premises from
hypotheses with an intermediate latent space, which we interpret as different possible premises for a given hypothesis.
This was trained using a Conditional Adversarial Autoencoder. This paper also proposed the Memory Operation
Selection Module for encoding sentences to a distributed
representation that uses attention over different operations
in order to encode the input. The model was evaluated for
quality and diversity. In terms of quality, we used two stateof-the-art models for the RTE task on SNLI, and the samples
generated by our best model were able to achieve an accuracy of 76.1%. For diversity, we compared the BLEU scores
between the real premises and the generated premises, and
the BLEU scores between the generated premises. In this regard, while our model is able to generate different premises
for each hypothesis, there is still a gap between when compared to the similarities to the real premises. Looking at the
samples, we note that the additional details that our model
generates tend to repeat, and correspond to some type of
mode collapse.
The task of performing reasoning well with natural language still remains a challenging problem. Our experiments
demonstrate that while we can generate sentences with the
logical entailment properties we desire, there is still much
to be done in this direction. We hope with the new lens
on some NLP tasks as natural language manipulation with
logical control, new perspectives and methods will emerge
to improve the field.
References
Bowman, Samuel R., Angeli, Gabor, Potts, Christopher, and
Manning, Christopher D. A large annotated corpus for
learning natural language inference. In Proceedings of
the 2015 Conference on Empirical Methods in Natural
Language Processing (EMNLP). Association for Computational Linguistics, 2015a.
Bowman, Samuel R, Vilnis, Luke, Vinyals, Oriol, Dai, An-
Generating Contradictory, Neutral, and Entailing Sentences
drew M, Jozefowicz, Rafal, and Bengio, Samy. Generating sentences from a continuous space. arXiv preprint
arXiv:1511.06349, 2015b.
Chen, Boxing and Cherry, Colin. A systematic comparison
of smoothing techniques for sentence-level bleu. In Proceedings of the Ninth Workshop on Statistical Machine
Translation, 2014.
Chen, Qian, Zhu, Xiaodan, Ling, Zhen-Hua, Wei, Si, Jiang,
Hui, and Inkpen, Diana. Enhanced lstm for natural language inference. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, 2017.
Conneau, Alexis, Kiela, Douwe, Schwenk, Holger, Barrault, Loic, and Bordes, Antoine. Supervised learning of
universal sentence representations from natural language
inference data. arXiv preprint arXiv:1705.02364, 2017.
Gong, Yichen, Luo, Heng, and Zhang, Jian. Natural language inference over interaction space. arXiv preprint
arXiv:1709.04348, 2017.
Gupta, Ankush, Agarwal, Arvind, Singh, Prawaan, and Rai,
Piyush. A deep generative framework for paraphrase
generation. arXiv preprint arXiv:1709.05074, 2017.
Guu, Kelvin, Hashimoto, Tatsunori B, Oren, Yonatan, and
Liang, Percy. Generating sentences by editing prototypes.
arXiv preprint arXiv:1709.08878, 2017.
Hochreiter, Sepp and Schmidhuber, Jürgen. Long short-term
memory. Neural computation, 1997.
Hu, Zhiting, Yang, Zichao, Liang, Xiaodan, Salakhutdinov,
Ruslan, and Xing, Eric P. Toward controlled generation of
text. In International Conference on Machine Learning,
2017.
Kiros, Ryan, Zhu, Yukun, Salakhutdinov, Ruslan R, Zemel,
Richard, Urtasun, Raquel, Torralba, Antonio, and Fidler,
Sanja. Skip-thought vectors. In Advances in neural
information processing systems, 2015.
Kolesnyk, Vladyslav, Rocktäschel, Tim, and Riedel, Sebastian. Generating natural language inference chains. arXiv
preprint arXiv:1606.01404, 2016.
Maaten, Laurens van der and Hinton, Geoffrey. Visualizing
data using t-sne. Journal of machine learning research,
2008.
Makhzani, Alireza, Shlens, Jonathon, Jaitly, Navdeep,
Goodfellow, Ian, and Frey, Brendan. Adversarial autoencoders. arXiv preprint arXiv:1511.05644, 2015.
Mikolov, Tomas, Sutskever, Ilya, Chen, Kai, Corrado,
Greg S, and Dean, Jeff. Distributed representations of
words and phrases and their compositionality. In Advances in neural information processing systems, 2013.
Mou, Lili, Meng, Zhao, Yan, Rui, Li, Ge, Xu, Yan, Zhang,
Lu, and Jin, Zhi. How transferable are neural networks
in nlp applications? arXiv preprint arXiv:1603.06111,
2016.
Salimans, Tim, Goodfellow, Ian, Zaremba, Wojciech, Cheung, Vicki, Radford, Alec, and Chen, Xi. Improved
techniques for training gans. In Advances in Neural Information Processing Systems, 2016.
Tiedemann, Jrg. Emerging language spaces learned
from massively multilingual corpora. arXiv preprint
arXiv:1802.00273, 2018.
Vaswani, Ashish, Shazeer, Noam, Parmar, Niki, Uszkoreit,
Jakob, Jones, Llion, Gomez, Aidan N, Kaiser, Łukasz,
and Polosukhin, Illia. Attention is all you need. In Advances in Neural Information Processing Systems, 2017.
Zhang, Zhifei, Song, Yang, and Qi, Hairong. Age progression/regression by conditional adversarial autoencoder.
In The IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), volume 2, 2017.
Zhao, Jieyu, Wang, Tianlu, Yatskar, Mark, Ordonez, Vicente, and Chang, Kai-Wei. Men also like shopping:
Reducing gender bias amplification using corpus-level
constraints. arXiv preprint arXiv:1707.09457, 2017.
| 2cs.AI
|
1
Information-Theoretic Privacy for Smart Metering
Systems with a Rechargeable Battery
arXiv:1510.07170v4 [cs.IT] 15 Sep 2017
Simon Li, Ashish Khisti, and Aditya Mahajan
Abstract—Smart-metering systems report electricity usage of a
user to the utility provider on almost real-time basis. This could
leak private information about the user to the utility provider.
In this work we investigate the use of a rechargeable battery in
order to provide privacy to the user.
We assume that the user load sequence is a first-order Markov
process, the battery satisfies ideal charge conservation, and
that privacy is measured using normalized mutual information
(leakage rate) between the user load and the battery output.
We study the optimal battery charging policy that minimizes
the leakage rate among the class of battery policies that satisfy
causality and charge conservation. We propose a series reductions
on the original problem and ultimately recast it as a Markov
Decision Process (MDP) that can be solved using a dynamic
program.
In the special case of i.i.d. demand, we explicitly characterize
the optimal policy and show that the associated leakage rate can
be expressed as a single-letter mutual information expression. In
this case we show that the optimal charging policy admits an
intuitive interpretation of preserving a certain invariance property of the state. Interestingly an alternative proof of optimality
can be provided that does not rely on the MDP approach, but is
based on purely information theoretic reductions.
I. I NTRODUCTION
Smart meters are a critical part of modern power distribution
systems because they provide fine-grained power consumption
measurements to utility providers. These fine-grained measurements improve the efficiency of the power grid by enabling
services such as time-of-use pricing and demand response [1].
However, this promise of improved efficiency is accompanied
by a risk of privacy loss. It is possible for the utility provider—
or an eavesdropper—to infer private information including
load taxonomy from the fine-grained measurements provided
by smart meters [2]–[4]. Such private information could be
exploited by third parties for the purpose of targeted advertisement or surveillance. Traditional techniques in which an
intermediary anonymizes the data [5] are also prone to privacy
loss to an eavesdropper. One possible solution is to partially
obscure the load profile by using a rechargeable battery [6]. As
the cost of rechargeable batteries decreases (for example, due
to proliferation of electric vehicles), using them for improving
privacy is becoming economically viable.
Simon Li was with the University of Toronto when this work was done.
Ashish Khisti ([email protected]) is with the University of Toronto,
Toronto, ON, Canada. Aditya Mahajan ([email protected]) is with
McGill University, Montreal, QC, Canada.
Preliminary version of this work was presented at the Signal Processing for
Advanced Wireless Communications (SPAWC), Jul. 2015, Stockholm, Sweden, International Zurich Seminar (IZS), Mar. 2-4, 2016, Zurich, Switzerland
and the American Control Conference (ACC), July 6-8, 2016, Boston, MA,
USA.
In a smart metering system with a rechargeable battery, the
energy consumed from the power grid may either be less than
the user’s demand—the rest being supplied by the battery; or
may be more than the user’s demand—the excess being stored
in the battery. A rechargeable battery provides privacy because
the power consumed from the grid (rather than the user’s
demand) gets reported to the electricity utility (and potentially
observed by an eavesdropper). In this paper, we focus on
the mutual information between the user’s demand and consumption (i.e., the information leakage) as the privacy metric.
Mutual Information is a widely used metric in the literature
on information theoretic security, as it is often analytically
tractable and provides a fundamental bound on the probability
of detecting the true load sequence from the observation [7].
Our objective is to identify a battery management policy
(which determine how much energy to store or discharge from
the battery) to minimize the information leakage rate.
We briefly review the relevant literature. The use of a
rechargeable battery for providing user privacy has been
studied in several recent works, e.g., [6], [8]–[11]. Most of the
existing literature has focused on evaluating the information
leakage rate of specific battery management policies. These
include the “best-effort” policy [6], which tries to maintain a
constant consumption level, whenever possible; and battery
conditioned stochastic charging policies [8], in which the
conditional distribution on the current consumption depends
only on the current battery state (or on the current battery state
and the current demand). In [6], the information leakage rate
was estimated using Monte-Carlo simulations; in [8], it was
calculated using the BCJR algorithm [12]. The methodology
of [8] was extended by [9] to include models with energy
harvesting and allowing for a certain amount of energy waste.
Bounds on the performance of the best-effort policy and hideand-store policy for models with energy harvesting and infinite
battery capacity were obtained in [10]. The performance of
the best effort algorithm for an individual privacy metric was
considered in [11]. None of these papers address the question
of choosing the optimal battery management policy.
Rate-distortion type approaches have also been used to study
privacy-utility trade-off [13]–[15]. These models allow the user
to report a distorted version of the load to the utility provider,
subject to a certain average distortion constraint. Our setup
differs from these works as we impose a constraint on the
instantaneous energy stored in the battery due to its limited
capacity. Both our techniques and the qualitative nature of the
results are different from these papers.
Our contributions are two-fold. First, when the demand is
Markov, we show that the minimum information leakage rate
2
and optimal battery management policies can be obtained by
solving an appropriate dynamic program. These results are
similar in spirit to the dynamic programs obtained to compute
capacity of channels with memory [16]–[18]; however, the
specific details are different due to the constraint on the battery
state. Second, when the demand is i.i.d., we obtain a single
letter characterization of the minimum information leakage
rate; this expression also gives the optimal battery management
policy. We prove the single letter expression in two steps. On
the achievability side we propose a class of policies with a
specific structure that enables a considerable simplification of
the leakage-rate expression. We find a policy that minimizes
the leakage-rate within this restricted class. On the converse
side, we obtain lower bounds on the minimal leakage rate and
show that these lower bound match the performance of the best
structured policy. We provide two proofs. One is based on the
dynamic program and the other is based purely on information
theoretic arguments.
After the present work was completed, we became aware
of [19], where a similar dynamic programming framework is
presented for the infinite horizon case. However, no explicit
solutions of the dynamic program are derived in [19]. To the
best of our knowledge, the present paper is the first work that
provides an explicit characterization of the optimal leakage
rate and the associated policy for i.i.d. demand.
A. Notation
Random variables are denoted by uppercase letters (X, Y ,
etc.), their realization by corresponding lowercase letters (x, y,
etc.), and their state space by corresponding script letters (X ,
Y, etc.). PX denotes the space of probability distributions on
X ; PX|Y denotes the space of condition distributions from Y
to X . Xab is a short hand for (Xa , Xa+1 , . . . , Xb ) and X b =
X1b . For a set A, 1A (x) denotes the indicator function of the
set that equals 1 if x ∈ A and zero otherwise. If A is a
singleton set {a}, we use 1a (x) instead of 1{a} (x).
Given random variables (X, Y ) with joint distribution
PX,Y (x, y) = PX (x)q(y|x), H(X) and H(PX ) denote the
entropy of X, H(Y |X) and H(q|PX ) denote conditional
entropy of Y given X and I(X; Y ) and I(q; PX ) denote the
mutual information between X and Y .
II. P ROBLEM F ORMULATION AND M AIN R ESULTS
A. Model and problem formulation
Consider a smart metering system as shown in Fig. 1. At
each time, the energy consumed from the power grid must
equal the user’s demand plus the additional energy that is either stored in or drawn from the battery. Let {Xt }t≥1 , Xt ∈ X ,
denote the user’s demand; {Yt }t≥1 , Yt ∈ Y, denote the energy
drawn from the grid; and {St }t≥1 , St ∈ S, denote the energy
stored in the battery. All alphabets are finite. For convenience,
we assume X := {0, 1, . . . , mx }, Y := {0, 1, . . . , my }, and
S = {0, 1, . . . , ms }. Here ms corresponds to the size of
the battery. We note that such a restriction is for simplicity
of presentation; the results generalize even when X and Y
are not necessarily contiguous intervals or integer valued. To
Home
Applicances
Demand: 𝑋𝑡
Consumption: 𝑌𝑡
Smart Meter
Controller
Power
Grid
𝑌𝑡 − 𝑋 𝑡
Battery
(State 𝑆𝑡 )
Eavesdropper/
Adversory
Fig. 1. A smart metering system.
guarantee that user’s demand is always satisfied, we assume
mx ≤ my or that X ⊆ Y holds more generally.
The demand {Xt }t≥1 is a first-order time-homogeneous
Markov chain1 with transition probability Q. We assume
that Q is irreducible and aperiodic. The initial state X1 is
distributed according to probability mass function PX1 . The
initial charge S1 of the battery is independent of {Xt }t≥1 and
distributed according to probability mass function PS1 .
The battery is assumed to be ideal and has no conversion
losses or other inefficiencies. Therefore, the following conservation equation must be satisfied at all times:
St+1 = St + Yt − Xt .
(1)
Given the history of demand, battery charge, and consumption, a randomized battery charging policy q = (q1 , q2 , . . . )
determines the energy consumed from the grid. In particular,
given the histories (xt , st , y t−1 ) of demand, battery charge,
and consumption at time t, the probability that current consumption Yt equals y is qt (y | xt , st , y t−1 ). For a randomized
charging policy to be feasible, it must satisfy the conservation
equation (1). So, given the current power demand and battery
charge (xt , st ), the feasible values of grid consumption are
defined by
Y◦ (st − xt ) = {y ∈ Y : st − xt + y ∈ S}.
(2)
Thus, we require that
qt (Y◦ (st − xt ) | xt , st , y t−1 ) :=
X
qt (y | xt , st , y t−1 ) = 1.
y∈Y◦ (st −xt )
The set of all such feasible policies is denoted by QA 2 . Note
that while the charging policy qt (·) can be a function of the
entire history, the support of qt (·) only depends on the present
value of xt and st through the difference st − xt . This is
emphasized in the definition in (2).
The quality of a charging policy depends on the amount
of information leaked under that policy. There are different
notions of privacy; in this paper, we use mutual information
as a measure of privacy. Intuitively speaking, given random
variables (Y, Z), the mutual information I(Y ; Z) measures
1 In practice, the energy demand is periodic rather than time homogeneous.
We are assuming that the total demand may be split into a periodic predictable
component and a time-homogeneous stochastic component. In this paper, we
ignore the predictable component because it does not affect privacy.
2 With a slight abuse of notation, we use Q to denote the battery policy
A
for both the infinite and finite-horizon problems
3
the decrease in the uncertainty about Y given by Z (or viceversa). Therefore, given a policy q, the information about
(X T , S1 ) leaked to the utility provider or eavesdropper is
captured by I q (X T , S1 ; Y T ), where the mutual information
is evaluated according to the joint probability distribution on
(X T , S T , Y T ) induced by the distribution q as follows:
𝑥 = 0/𝑦 = 0
𝑥 = 0/𝑦 = 0
𝑥 = 0/𝑦 = 1
𝑠=0
𝑠=1
𝑥 = 1/𝑦 = 0
Pq (S T = sT , X T = xT , Y T = y T )
𝑥 = 1/𝑦 = 1
𝑥 = 1/𝑦 = 1
T
Y
= PS1 (s1 )PX1 (x1 )q1 (y1 | x1 , s1 )
1st {st−1 −xt−1 +yt−1 } Fig. 2. Binary System model. The battery can be either in s = 0 or s = 1.
The set of feasible transitions from each state are shown in the figure.
t=2
× Q(xt |xt−1 )qt (yt | xt , st , y t−1 ) .
B. Example: Binary Model
We use information leakage rate as a measure of the
We illustrate the special case when X = Y = S = {0, 1}
quality of a charging policy. For a finite planning horizon, the
in Fig. 2. The input, output, as well as the state, are all
information leakage rate of a policy q = (q1 , . . . , qT ) ∈ QA
binary valued. When the battery is in state st = 0, there
is given by
are three possible transitions. If the input xt = 1 then we
1 q T
T
(3) must select yt = 1 and the state changes to st+1 = 0. If
LT (q) := I (X , S1 ; Y ),
T
instead xt = 0, then there are two possibilities. We can select
while for an infinite horizon, the worst-case information leak- yt = 0 and have st+1 = 0 or we can select yt = 1 and
have st+1 = 1. In a similar fashion there are three possible
age rate of a policy q = (q1 , q2 , . . . ) ∈ QA is given by
transitions from the state st = 1 as shown in Fig. 2. We will
L∞ (q) := lim sup LT (q).
(4) assume that the demand (input) sequence is sampled i.i.d. from
T →∞
an equiprobable distribution, i.e., PX (0) = PX (1) = 12 .
We are interested in the following optimization problems:
Consider a simple policy that sets yt = xt and ignores the
Problem A. Given the alphabet X of the demand, the initial battery state. It is clear that such a policy will lead to maximum
distribution PX1 and the transistion matrix Q of the demand leakage LT = 1. Another feasible policy is to set yt = 1 − st .
process, the alphabet S of the battery, the initial distribution Thus whenever st = 0, we will set yt = 1 regardless of the
PS1 of the battery state, and the alphabet Y of the consump- value of xt , and likewise st = 1 will result in yt = 0. It turns
out that the leakage rate for this policy also approaches 1. To
tion:
T
1) For a finite planning horizon T , find a battery charging see this note thatT the eavesdropper having access to y also
s . Using the battery update equation (1) the
policy q = (q1 , . . . , qT ) ∈ QA that minimizes the in turn knows
T −1
sequence
x
is thus revealed to the eavesdropper, resulting
leakage rate LT (q) given by (3).
in
a
leakage
rate
of at least 1 − 1/T .
2) For an infinite planning horizon, find a battery charging
In
reference
[8],
a probabilistic battery charging policy is
policy q = (q1 , q2 , . . . ) ∈ QA that minimizes the
introduced
that
only
depends on the current state and input i.e.,
leakage rate L∞ (q) given by (4).
∆
qt (yt |xt , st ) = qt (yt |xt , st ). Furthermore the policy makes
The above optimization problem is difficult because we equiprobable decisions between the feasible transitions i.e.,
have to optimize a multi-letter mutual information expression
over the class of history dependent probability distributions.
qt (yt = 0|xt , st ) = q(yt = 1|xt , st ) = 1/2, if xt = st (5)
In the spirit of results for feedback capacity of channels
with memory [16]–[18], we show that the above optimization
problem can be reformulated as a Markov decision process and qt (yt |xt , st ) = 1xt (yt ) otherwise. The leakage rage for
where the state and action spaces are conditional probability this policy was numerically evaluated in [8] using the BCJR
distributions. Thus, the optimal policy and the optimal leakage algorithm and it was shown numerically that L∞ = 0.5. Such
rate can be computed by solving an appropriate dynamic numerical techniques seem necessary in general even for the
program. We then provide an explicit solution of the dynamic class of memoryless policies and i.i.d. inputs, as the presence
of the battery adds memory into the system.
program for the case of i.i.d. demand.
As a consequence of our main result it follows that the
Remark 1. We note that the class of policies we consider in
above policy admits a single-letter expression for the leakage
QA are randomized policies i.e., the output at any given time
rate3 L∞ = I(S ∗ − X; X), thus circumventing the need for
is governed by the conditional distribution qt (yt | xt , st , y t−1 ).
numerical techniques. Furthermore it also follows that this
The class of deterministic policies where yt = ht (xt , st , y t−1 )
leakage rate is indeed the minimum possible one among the
is a deterministic function of the past inputs, state and outputs
class of all feasible policies. Thus it is not necessary for the
is a special case of randomized policies where qt (·) is an
battery system to use more complex policies that take into
atomic distribution. As will be apparent from out results,
deterministic policies do not suffice to minimize the leakage
3 The random variable S ∗ is an equiprobable binary valued random variable,
rate and hence we focus on the class of randomized policies. independent of X. See Sec. II-E.
4
account the entire history. We note that a similar result was
shown in [20] for the case of finite horizon policies. However
the proof in [20] is specific to the binary model. In the present
paper we provide a complete single-letter solution to the case
of general i.i.d. demand, and a dynamic programming method
for the case of first-order Markovian demands, as discussed
next.
C. Main results for Markovian demand
We identify two structural simplifications for the battery
charging policies. First, we show (see Proposition 1 in Section III-A) that there is no loss of optimality in restricting
attention to charging policies of the form
qt (yt |xt , st , y t−1 ).
(6)
The intuition is that under such a policy, observing y t gives
partial information only about (xt , st ) rather than about the
whole history (xt , st ).
Next, we identify a sufficient statistic for y t−1 when the
charging policies are of the form (6). For that matter, given a
policy q and any realization y t−1 of Y t−1 , define the belief
state πt ∈ PX,S as follows: for all x ∈ X , s ∈ S,
πt (x, s) = Pq (Xt = x, St = s|Y t−1 = y t−1 ).
(7)
Then, we show (see Theorems 1 and 2 below) that there is no
loss of optimality in restricting attention to charging policies
of the form
qt (yt |xt , st , πt ).
(8)
where the function ϕ is a non-linear filtering equation defined
in Sec. III-C.
Our main results are the following:
Theorem 1. We have the following for Problem A with a finite
planning horizon T :
1) Value functions: Iteratively define value functions
Vt : PX,S → R as follows. For any π ∈ PX,S ,
VT +1 (π) = 0, and for t ∈ {T, . . . , 1},
Vt (π) = min[Ba Vt+1 ](π),
a∈A
∀(x, s) ∈ X × S . (9)
The set A is convex and compact.4
To succinctly write the dynamic program, for any a ∈ A, we
define the Bellman operator Ba : [PX,S → R] → [PX,S → R]
as follows: for any V : PX,S → R and any π ∈ PX,S ,
[Ba V ](π) = I(a; π)
X
+
π(x, s)a(y | x, s)V (ϕ(π, y, a)) (10)
x∈X ,s∈S,
y∈Y
4 If a , a ∈ A, then any linear combination a0 of them, where a0 =
1 2
λa1 + (1 − λ)a2 with λ ∈ (0, 1), also satisfies a0 (Y◦ (s − x | x, s) = 1
for all (x, s) ∈ X × S and, therefore, belongs to A. Hence A is convex.
Moreover, it is easy to see that A is closed and PY |X,s is compact. Therefore,
A is also compact.
(11)
Then, Vt (π) are is continuous and concave in π.
2) Optimal policy: Let ft∗ (π) denote the arg min of the
right hand side of (11). Then, optimal policy q∗ =
(q1∗ , . . . , qT∗ ) is given by
qt∗ (yt | xt , st , πt ) = at (yt | xt , st ), where at = ft∗ (πt ).
Thus, there is no loss of optimality in restricting attention to charging policies of the form (8).
3) Optimal leakage rate: The optimal (finite horizon) leakage rate is given by V1 (π1 )/T , where π1 (x, s) =
PX1 (x)PS1 (s).
See Section III for proof.
Theorem 2. We have the following for Problem A with an
infinite planning horizon:
1) Value function: Consider the following fixed point equation
J + v(π) = min[Ba v](π),
a∈A
Such a charging policy is Markovian in the belief state πt
and the optimal policies of such form can be searched using
a dynamic program.
To describe such a dynamic program, we assume that there
is a decision maker that observes y t−1 (or equivalently πt ) and
chooses “actions” at = qt (·|·, ·, πt ) using some decision rule
at = ft (πt ). We then identify a dynamic program to choose
the optimal decision rules.
Note that the actions at take vales in a subset A of PY |X,S
given by
A = a ∈ PY |X,S : a(Y◦ (s − x) | x, s) = 1,
∀π ∈ PX,S .
∀π ∈ PX,S ,
(12)
where J ∈ R is a constant and v : PX,S → R.
2) Optimal policy: Suppose there exists (J, v) that satisfy (12). Let f ∗ (π) denote the arg min of the right hand
side of (12). Then, the time-homogeneous optimal policy
q∗ = (q ∗ , q ∗ , . . . ) given by
q ∗ (yt | xt , st , πt ) = a(yt | xt , st ), where a = f ∗ (π)
is optimal. Thus, there is no loss of optimality in restricting attention to charging time-homogeneous policies of
the form (8).
3) Optimal leakage rate: The optimal (infinite horizon)
leakage rate is given by J.
Proof. Given the result of Theorem 1, the result of Theorem 2
follows from standard dynamic programming arguments. See,
for example, [21, Theorem 5.2.4 and Eq. (5.2.18)].
There are various conditions that guarantee the existence of
a (J, v) that satisfies (12). Most of these conditions require the
ergodicity of the process {Πt }t≥1 for every standary Markov
procily f : PX,S → A. We refer the reader to [22, Chapter 3]
and [23, Chapter 10] for more details.
The dynamic program above resembles the dynamic program for partially observable Markov decision processes
(POMDP) (see [24, Chapter 5]) with hidden state (Xt , St ), observation Yt , and action At . However, in contrast to POMDPs,
the expected per-step cost I(a; π) is not linear in π. Nonetheless, one could use computational techniques from POMDPs
5
to approximately solve the dynamic programs of Theorems 1
and 2. See Section III-D for a brief discussion.
E. Binary Model (Revisited)
Remark 2. Although the above results assume that the demand
is a first-order Markov chain, they extend naturally to the
case when the demand is higher-order Markov. In particular, suppose the demand {Xt }t≥1 is a k-th order Markov
chain. Then, we can define another process {X̃t }t≥1 where
X̃t = (Xt−k+1 , . . . , Xt ), and use X̃t in Theorems 1 and 2.
We revisit the binary model in Section II-B, where the
demand has equiprobable distribution, i.e., PX (x) = 12 for
x ∈ {0, 1}. Consider θ(0) = p, θ(1) = 1−p. Let W = S −X.
Then,
1
if w = −1
2 p,
1
P(W = w) = 2 ,
(17)
if w = 0
1
(1
−
p),
if
w
=
1.
2
D. Main result for i.i.d. demand
Thus,
Assume the following:
I(W ; X) = H(W ) − H(S) = 1 − 21 h(p)
Assumption (A). The demand {Xt }t≥1 is i.i.d. with probability distribution PX .
We provide an explicit characterization of optimal policy
and optimal leakage rate for this case.
Define an auxiliary state variable Wt = St − Xt that takes
values in W = {s − x : s ∈ S, x ∈ X }. For any w ∈ W,
define:
D(w) = {(x, s) ∈ X × S : s − x = w}.
(13)
Theorem 3. Define
θ∈PS
θ∈PS
(14)
where X and S are independent with X ∼ PX and S ∼
∗
θ.
arg min in (14). Define ξ ∗ (w) =
P Let θ denote the
∗
(x,s)∈D(w) PX (x)θ (s). Then, under (A)
1) J ∗ is the optimal (infinite horizon) leakage rate.
2) Define b∗ ∈ PY |W as follows:
(
∗
if y ∈ X ∩ Y◦ (w)
PX (y) θ ξ(y+w)
∗ (w)
∗
b (y|w) =
0
otherwise .
(15)
We call b∗ as a structured policy with respect to (θ∗ , ξ ∗ ).
Then, the memoryless charging policy q∗ = (q1∗ , q2∗ , . . . )
given by
qt∗ (y | x, s, πt ) = b∗ (y | s − x)
h(p) = −p log p − (1 − p) log(1 − p).
Thus, the value of p that minimizes I(W ; X) is p∗ = 12 and
the optimal leakage rate is I(W ; X) = 21 .
The corresponding ξ ∗ is obtained by substituting p = 12
in (17). For ease of notation, we denote b∗ (·|w) as
[b∗ (0|w), b∗ (1|w)]. Then,
b∗ (·| − 1) = [0, 1],
Then, we have the following.
J ∗ = min I(S −X; X) = min H(S − X) − H(S)
where h(p) is the binary entropy function:
(16)
is optimal and achieves the optimal (infinite horizon)
leakage rate.
Note that the optimal charging policy is time-invariant and
memoryless, i.e., the distribution on Yt does not depend on πt
(and, therefore on y t−1 ).
The proof, which is presented in Section IV, is based
on the standard arguments of showing achievability and a
converse. On the achievability side we show that the policy
in (15) belongs to a class of policies that satisfies a certain
invariance property. Using this property the multi-letter mutual
information expression can be reduced into a single-letter
expression. For the converse we provide two proofs. The
first is based on a simplification of the dynamic program of
Theorem 2. The second is based on purely probabilistic and
information theoretic arguments.
b∗ (·|0) = [ 12 , 21 ],
b∗ (·| − 1) = [1, 0].
It can be shown that this strategy is the same as (5), which
was proposed in [8]. This yields an analytical proof of the
result in [8].
F. Salient features of the result for i.i.d. demand
Theorem 3 shows that even if consumption could take larger
values than the demand, i.e., Y ⊃ X , under the optimal policy,
Yt takes values only in X . This agrees with the intuition that
a consumption larger that mx reveals that the battery has low
charge and that the power demand is high. In extreme cases,
a large consumption may completely reveal the battery and
power usage thereby increasing the information leakage.
We now show some other properties of the optimal policy.
Property 1. The mutual information I(S − X; X) is strictly
convex in the distribution θ and, therefore, θ∗ ∈ int(PS ).
See Appendix A for proof.
As a consequence, the optimal θ∗ in (14) may be obtained
using the Blahut-Arimoto algorithm [25], [26].
Property 2. Under the battery charging policy specified in
Theorem 3, the power consumption {Yt }t≥1 is i.i.d. with
marginal distribution PX . Thus, {Yt }t≥1 is statistically indistinguishable from {Xt }t≥1 .
See Remarks 3 and 5 in Section IV-B for proof.
Property 3. If the power demand has a symmetric PMF, i.e.,
for any x ∈ X , PX (x) = PX (mx − x), then the optimal θ∗
in Theorem 3 is also symmetric, i.e., for any s ∈ S, θ∗ (s) =
θ∗ (ms − s).
Proof. For θ ∈ PS , define θ̄(s) = θ(ms − s). Let X ∼ PX ,
S ∼ θ and S̄ ∼ θ̄. Then, by symmetry
I(S − X; X) = I(S̄ − X; X).
(18)
6
For any λ ∈ (0, 1), let θλ (s) = λθ(s) + (1 − λ)θ̄(s) denote the
convex combination of θ and θ̄. Let Sλ ∼ θλ . By Property 1,
if θ 6= θ̄, then
Leakage Rates for Binomial Distributed Power Demand
1
10
Jeq for mx = 5
J* for mx = 5
Jeq for mx = 10
I(Sλ − X; X) < λI(S − X; X) + (1 − λ)I(S − X; X)
J* for mx = 10
0
= I(S − X; X),
10
Leakage Rate
where the last equation uses (18).
Thus, if θ 6= θ̄, we can strictly decrease the mutual
information by using θλ . Hence, the optimal distribution must
have the property that θ∗ (s) = θ∗ (ms − s).
Jeq for mx = 20
J*
for mx = 20
−1
10
Given a distribution µ on some alphabet M, we say that the
distribution is almost symmetric and unimodal around m∗ ∈
M if
−2
10
µm∗ ≥ µm∗ +1 ≥ µm∗ −1 ≥ µm∗ +2 ≥ µm∗ −2 ≥ . . .
where we use the interpretation that for m 6∈ M, µm = 0.
Similarly, we say that the distribution is symmetric and unimodal around m∗ ∈ M if
0
5
10
15
20
25
30
Battery Size
35
40
45
50
Fig. 3. A comparison of the performance of qeq ∈ QB as defined
in (19) with the optimal leakage rate for i.i.d. Binomial distributed
demand Binomial(mx , 0.5) for mx = {5, 10, 20}.
µm∗ ≥ µm∗ +1 = µm∗ −1 ≥ µm∗ +2 = µm∗ −2 ≥ . . .
Note that a distribution can be symmetric and unimodal only
if its support is odd.
Property 4. If the power demand is symmetric and unimodal
around bmx /2c, then the optimal θ∗ in Theorem 3 is almost
symmetric and unimodal around bms /2c. In particular, if ms
is even, then
∗
θm
∗
≥
∗
θm
∗ +1
=
∗
θm
∗ −1
≥
∗
θm
∗ +2
=
∗
θm
∗ −2
≥ ...
and if ms is odd, then
∗
∗
∗
∗
∗
θm
∗ = θm∗ +1 ≥ θm∗ −1 = θm∗ +2 ≥ θm∗ −2 = . . .
where m∗ = bms /2c.
Proof. Let X̄ = −X. Then, I(S − X; S) = H(S − X) −
H(S) = H(S + X̄) − H(S). Note that PX̄ is also symmetric
and unimodal around bmx /2c.
Let S ◦ and X̄ ◦ denote the random variables S − bms /2c
and X̄ − bmx /2c. Then X̄ ◦ is also symmetric and unimodal
around the origin and
I(S − X; X) = H(S ◦ + X̄ ◦ ) − H(S ◦ ).
Now given any distribution θ◦ of S ◦ , let θ+ be a permutation of θ◦ that is almost symmetric and unimodal with a
positive bias around origin. Then by [27, Corollary III.2],
H(PX ∗ θ◦ ) ≥ H(PX ∗ θ+ ). Thus, the optimal distribution
must have the property that θ◦ = θ+ or, equivalently, θ is
almost unimodal and symmetric around bms /2c.
Combining this with the result of Property 3 gives the characterization of the distribution when ms is even or odd.
G. Numerical Example: i.i.d. demand
Suppose there are n identical devices in the house and each
is on with probability p. Thus, X ∼ Binomial(n, p). We derive
the optimal policy and optimal leakage rate for this scenario
under the assumption that Y = X . We consider two specific
examples, where we numerically solve (14).
Suppose n = 6 and p = 0.5.
1) Consider S = [0:5]. Then, by numerically solving (14),
we get that the optimal leakage rate J ∗ is is 0.4616 and
the optimal battery charge distribution θ∗ is
{0.1032, 0.1747, 0.2221, 0.2221, 0.1747, 0.1032}.
2) Consider S = [0:6]. Then, by numerically solving (14),
we get that the optimal leakage rate J ∗ is is 0.3774 and
the optimal battery charge distribution θ∗ is
{0.0773, 0.1364, 0.1847, 0.2031, 0.1847, 0.1364, 0.0773}.
Note that both results are consistent with Properties 3 and 4.
We next compare the performance with the following timehomogeneous benchmark policy qeq ∈ QB : for all y ∈ Y, w ∈
W,
1Y◦ (wt ) {yt }
qt (yt |wt ) =
.
(19)
|Y◦ (wt )|
This benchmark policy chooses all feasible values of Yt with
equal probability. For that reason we call it equi-probably
policy and denote its performance by Jeq .
In Fig. 3, we compare the performance of q∗ and qeq as a
function of battery sizes for different demand alphabets.
Under qeq , the MDP converges to a belief state that is
approximately uniform. Hence, in the low battery size regime,
Jeq is close to optimal but its performance gradually worsens
with increasing battery size.
H. Numerical Example: Continuous Valued Alphabets
Although our setup assumes discrete alphabets, it can shown
that Theorem 3 can be extended to continuous alphabets under
mild technical conditions assuming that the density function
for X, say fX (x) exists; see [28] for details. As such we
provide two achievability proofs for Theorem 3. The weak
achievability in Section IV-B assumes that the distribution
of the initial state S1 can be selected by the user. The
strong achievability in Section IV-C assumes that the initial
7
state S1 can be arbitrary. The proof of the weak achievability extends immediately to continuous valued alphabets.
Furthermore the proof of the converse based on information
theoretic arguments in Section IV-E also extends to continuous
valued alphabets. We provide a numerical example involving
a continuous valued input.
Let X = [0, 1] be continuous valued, and let fX (x) = 1 for
x ∈ [0, 1]. We assume that S = [0, B] where B denotes the
storage capacity. We will assume that B ≥ 2 for convenience.
Following Theorem 3, it suffices to take the output alphabet
to be Y = [0, 1]. Note that for W = S − X, we have that
W = [−1, B] and that the support of the output for any given
w is given by:
−1 ≤ w ≤ 0,
[−w, 1],
Y◦ (w) = [0, 1],
(20)
0 ≤ w ≤ B − 1,
[0, B − w], B − 1 ≤ w ≤ B.
For the achievability, we evaluate the leakage rate by
selecting θ? (s) to be a uniform distribution over [0, B] and
using (21). The resulting leakage rate is given by
Let θ? (·) be the density function that minimizes (14) and S ?
denote the random variable with this density, independent of
X. Let W = S ? − X and ξ ? (w) be the associated density.
Then it follows from (15) that the optimal policy is given by
and thus it follows that
?
b? (y|w) =
θ (y + w)
,
ξ ? (w)
y ∈ Y◦ (w).
(21)
Instead of computing θ? (·) numerically, which is cumbersome
due to the density functions, we provide an analytical lower
bound on the leakage rate. Note that the objective in (14) can
be expressed as:
I(S − X; X) = h(S − X) − h(S)
(22)
where h(·) is the differential entropy. Using the entropy power
inequality [29], since S and X are independent, we have for
any density fS (·) that:
22h(S−X) ≥ 22h(S) + 22h(X) ,
L+ = h(W ) − h(S) = h(ξ) − log2 (B)
where the density ξ(w) for W
1+w
B ,
ξ(w) = B1 ,
B−w
B ,
(27)
= S − X is as follows:
−1 ≤ w ≤ 0
0≤w ≤B−1
B−1≤w ≤B
and we use that when θ(s) is a uniform density over [0, B],
it follows that h(S) = log2 (B). Through straightforward
computations it can be shown that
h(ξ) =
1
+ log2 (B)
2B ln 2
L+ =
(28)
1
2B ln 2
is achievable with a uniform distribution on the state. We note
that the analytical lower bound on the leakage rate in (26)
decays as 1/B 2 for large B while the achievable rate decays as
1/B. It will be interesting in future to determine the structure
of the optimal input distribution and study the associated
leakage rate.
III. P ROOF OF T HEOREM 1
One of the difficulties in obtaining a dynamic programming
decomposition for Problem
A is that the objective function
PT
is not of the form
cost(state
t , actiont ). We show that
t=1
there is no loss of optimality to restrict attention to a class of
policies QB and for any policy in QB , the mutual information
may be written in an additive form.
(23)
where we use the fact that h(−X) = h(X). Substituting
in (22) we have
1
I(S − X; X) ≥ log2 22h(S) + 22h(X) − h(S)
2
1
= log2 1 + 22h(X)−2h(S)
2
1
= log2 1 + 2−2h(S)
(24)
2
1
≥ min log2 1 + 2−2h(S)
S 2
1
= log2 1 + 2−2{maxS h(S)}
(25)
2
1
1
= log2 1 + 2
(26)
2
B
where we use the fact that when X ∼ Unif[0, 1], we have
that h(X) = 0, see e.g., [30] in (24) and the fact that the
expression in (25) is decreasing in h(S). Finally in (26) we use
the fact that a uniform distribution maximizes the differential
entropy when S = [0, B] is fixed and the maximum value is
h(S) = log2 B.
A. Simplification of optimal charging policies
Let QB ⊂ QA denote the set of charging policies that
choose consumption xd only on the consumption history,
current demand, and battery state. Thus, for q ∈ QB , at any
time t, given history (xt , st , y t−1 ), the consumption Yt is y
with probability qt (y | xt , st , y t−1 ). Then the joint distribution
on (X T , S T , Y T ) induced by q ∈ QB is given by
Pq (S T = sT , X T = xT , Y T = y T )
T
Y
= PS1 (s1 )PX1 (x1 )q1 (y1 | x1 , s1 )
1st {st−1 −xt−1 +yt−1 }
t=2
× Q(xt |xt−1 )qt (yt | xt , st , y t−1 ) .
Proposition 1. In Problem A, there is no loss of optimality in
restricting attention to charging policies in QB . Moreover, for
any q ∈ QB , the objective function takes an additive form:
LT (q) =
T
1X q
I (Xt , St ; Yt | Y t−1 )
T t=1
8
where
I q (Xt , St ; Yt | Y t−1 )
X
=
Pq (Xt = xt , St = st , Y t = y t )
xt ∈X ,st ∈S
qt (yt | xt , st , y t−1 )
y t ∈Y t
× log q
.
P (Yt = yt | Y t−1 = y t−1 )
See Appendix B for proof. The intuition behind why policies in QB are no worse that others in QA is as follows.
For a policy in QA , observing the realization y t of Y t gives
partial information about the history (xt , st ) while for a policy
in QB , yt gives partial information only about the current
state (xt , st ). The dependence on (xt , st ) cannot be removed
because of the conservation constraint (1).
Proposition 1 shows that the total cost may be written in an
additive form. Next we use an approach inspired by [16]–[18]
and formulate an equivalent sequential optimization problem.
B. An equivalent sequential optimization problem
Consider a system with state process {Xt , St }t≥1 where
{Xt }t≥1 is an exogenous Markov process as before and
{St }t≥1 is a controlled Markov process as specified below.
At time t, a decision maker observes Y t−1 and chooses a
distribution valued action At ∈ A, where A is given by (9),
as follows:
At = ft (Y t−1 )
(29)
where f = (f1 , f2 , . . . ) is called the decision policy.
Based on this action, an auxiliary variable Yt ∈ Y is chosen
according to the conditional probability at (· | xt , st ) and the
state St+1 evolves according to (1).
At each stage, the system incurs a per-step cost given by
at (yt | xt , st )
. (30)
ct (xt , st , at , y t ; f ) := log f
P (Yt = yt | Y t−1 = y t−1 )
The objective is to choose a policy f = (f1 , . . . , fT ) to
minimize the total finite horizon cost given by
" T
#
1 f X
t
L̃T (f ) := E
ct (Xt , St , At , Y ; f )
(31)
T
t=1
where the expectation is evaluated with respect to the probability distribution Pf induced by the decision policy f .
Proposition 2. The sequential decision problem described
above is equivalent to Problem A. In particular,
1) Given q = (q1 , . . . , qT ) ∈ QB , let f = (f1 , . . . , fT ) be
ft (y t−1 ) = qt (· | ·, ·, y t−1 ).
Thus, the probability distribution on (X T , S T , Y T ) induced
by a decision policy f = (f1 , . . . , fT ) is given by
Pf (S T = sT , X T = xT , Y T = y T )
= PS1 (s1 )PX1 (x1 )q1 (y1 | x1 , s1 )
T
Y
1st {st−1 −xt−1 +yt−1 }
t=2
× Q(xt |xt−1 )at (yt |xt , st ) .
where at = ft (y t−1 ). Under the transformations described
in the Proposition, Pf and Pq are identical probability distributions. Consequently, Ef [ct (Xt , St , At , Y t ; f )] =
I q (St , Xt ; Yt | Y t−1 ). Hence, LT (q) and L̃T (f ) are equivalent.
Eq. (32) implies that {Xt , St }t≥1 is a controlled Markov
process with control action {At }t≥1 . In the next section,
we obtain a dynamic programming decomposition for this
problem. For the purpose of writing the dynamic program,
it is more convenient to write the policy (29) as
At = ft (Y t−1 , At−1 ).
(33)
Note that these two representations are equivalent. Any policy
of the form (29) is also a policy of the form (33) (that simply
ignores At−1 ); any policy of the form (33) can be written
as a policy of the form (29) by recursively substituting At in
terms of Y t−1 . Since the two forms are equivalent, in the next
section we assume that the policy is of the form (33).
C. A dynamic programming decomposition
The model described in Section III-B above is similar to a
POMDP (partially observable Markov decision process): the
system state (Xt , St ) is partially observed by a decision maker
who chooses action At . However, in contrast to the standard
cost model used in POMDPs, the per-step cost depends on
the observation history and past policy. Nonetheless, if we
consider the belief state as the information state, the problem
can be formulated as a standard MDP.
For that matter, for any realization y t−1 of past observations
and any choice at−1 of past actions, define the belief state
πt ∈ PX,S as follows: For s ∈ S and x ∈ X ,
πt (x, s) = Pf (Xt = x, St = s|Y t−1 = y t−1 , At−1 = at−1 ).
If Y t−1 and At−1 are random variables, then the belief state
is a PX,S -valued random variable.
The belief state evolves in a state-like manner as follows.
Then L̃T (f ) = LT (q).
2) Given f = (f1 , . . . , fT ), let q = (q1 , . . . , qT ) ∈ QB be
Lemma 1. For any realization yt of Yt and at of At , πt+1 is
given by
qt (yt | xt , st , y t−1 ) = at (yt | xt , st ), where at = ft (y t−1 ).
πt+1 = ϕ(πt , yt , at )
(34)
Then LT (q) = L̃T (f ).
t
where ϕ is given by
t
Proof. For any history (x , s , y
t
t
t
t−1
t
), a ∈ A, and st+1 ∈ S,
t
t
t
t
t
P(St+1 = st+1 | X = x , S = s , Y = y , A = a )
X
=
1st+1 {st + yt − xt } at (yt | xt , st )
yt ∈Y
= P(St+1 = st+1 | Xt = xt , St = st , At = at ). (32)
ϕ(π, y, a)(x0 , s0 )
P
Q(x0 |x)a(y|x, s0 − x + y)π(x, s0 − x + y)
P
.
= x∈X
(x,s)∈X ×S a(y|x, s)π(x, s)
Proof. For ease of notation, we use P(xt , st |y t−1 , at−1 ) to
denote P(Xt = xt , St = st |Y t−1 = y t−1 , At−1 = at−1 ).
9
Similar interpretations hold for other expressions as well.
Consider
πt+1 (xt+1 , st+1 ) = P(xt+1 , st+1 |y t , at )
P(xt+1 , st+1 , yt , at |y t−1 , at−1 )
=
P(yt , at |y t−1 , at−1 )
(35)
Now, consider the numerator of the right hand side.
P(xt+1 , st+1 , yt , at |y t−1 , at−1 )
= P(xt+1 , st+1 , yt , at |y t−1 , at−1 , πt )
X
=
P(xt+1 |xt )1st+1 (st + xt − yt )
x∈X ,s∈S,
y∈Y
(xt ,st )∈X ×S
× at (yt |xt , st )1at (ft (y
t−1
t−1
,a
))πt (xt , st )
Lemma 2 implies that the objective function can be expressed
in terms of the state Πt and the action At . Thus, one can use
Markov decision theory [21] to identify the optimal policy.
Since both the state space and the action space are continuous
valued, we need to verify the standard technical conditions.
Define the stochastic kernel K : PX,S × A → PX,S as
follows. For any Borel subset B of PX,S and any π ∈ PX,S
and a ∈ A,
X
π(x, s)a(y | x, s)1B {ϕ(π, y, a))}.
K(B | π, a) =
(36)
Substituting (36) in (35) (and observing that the denominator of the right hand side of (35) is the marginal of
the numerator over (xt+1 , st+1 )), we get that πt+1 can be
written in terms of πt , yt and at . Note that if the term
1at (ft (y t−1 , at−1 )) is 1, it cancels from both the numerator
and the denominator; if it is 0, we are conditioning on a null
event in (35), so we can assign any valid distribution to the
conditional probability.
Note that an immediate implication of the above result is
that πt depends only on (y t−1 , at−1 ) and not on the policy f .
This is the main reason that we are working with a policy of
the form (33) rather than (29).
Lemma 2. The cost L̃T (f ) in (31) can be written as
X
T
1
L̃T (f ) = E
I(At ; Πt )
T
t=1
where I(at ; πt ) does not depend on the policy f and is
computed according to the standard formula
X
I(at ; πt ) =
πt (x, s)at (y | x, s)
x∈X ,s∈S,
at (y|x, s)
y∈Y
× log X
.
πt (x̃, s̃)at (y | x̃, s̃)
(x̃,s̃)∈X ×S
Proof. By the law of iterated expectations, we have
X
T
1
t
t−1
t−1
L̃T (f ) = E
E[ct (Xt , St , At , Y ; f )|Y
,A ]
T
t=1
(37)
Now, from (30), each summand may be written as
Ef [ct (Xt , St , At , Y t ; f ) | Y t−1 = y t−1 , At = at ]
X
=
πt (x, s)at (y | x, s)
x∈X ,s∈S,
at (y|x, s)
y∈Y
× log X
πt (x̃, s̃)at (y | x̃, s̃)
(x̃,s̃)∈X ×S
= I(at ; πt ).
Thus, Ef [ct (Xt , St , Y t ; f | Y t−1 , At ] = I(At , Πt ). Substituting this back in (37), we get the result of the Lemma.
Proof of Theorem 1. Lemma 1 implies that {Πt }t≥1 is a
controlled Markov process with control action At . In addition,
The sequential model of Sec. III-B has the following properties.
1) Based on (32), we can write the controlled dynamics of
the state (Xt , St ) as follows:
P(Xt+1 = x+ , St+1 = s+ | Xt = x, St = s, At = a)
X
=
Q(x+ |x)a(y)1s+ {s + y − x},
y∈Y
which is continuous in a.
2) The observations are given by
P(Yt = y|Xt = x, St = s, At = a) = a(y)
which is continuous in a.
3) The observations Yt are discrete.
Therefore, from [22, Sec 4.4 and Lemma 4.1], we get the
following:
4) The stochastic kernel K(dπ+ | π, a) is weakly continuous.
In addition, the model has the following properties:
5) The action set A is compact.
6) The per-step cost I(a, π) is continuous and bounded
below. (In fact, the per-step cost is also bounded above).
Properties 4)–6) imply [21, Condition 3.3.3], which by [21,
Theorem 3.3.5], implies the measurable selection condition [21, Assumption 3.3.1]. Under the measurable selection
condtion, the “inf” in the dynamic program can be replaced
by a “min”.
The continuity of the value function in π follows from
the continuity of the per-step cost I(a, π)and the controlled
stochastic kernel K(dπ+ | π, a). The concavity of the value
functions is proved in Appendix C.
From standard results in Markov decision theorem (e.g.,
see [21, Theorem 3.2.1]), it follows that the policy given in
part 2) of the theorem is optimal for the sequential model of
Sec III-B. Proposition 2 implies that this policy is also optimal
for Problem A.
D. Remarks about numerical solution
The dynamic programs of Theorems 1 and 2, both state and
action spaces are distribution valued (and, therefore, subsets of
Euclidean space). Although, an exact solution of the dynamic
program is not possible, there are two approaches to obtain
an approximate solution. The first is to treat it as a dynamic
program of an MDP with continuous state and action spaces
10
and use approximate dynamic programming [24], [31]. The
second is to treat it as a dynamic program for a POMDP and
use point-based methods [32]. The point-based methods rely
on concavity of the value function, which was established in
Theorem 1.
IV. P ROOF OF T HEOREM 3
We follow the standard approach and show that the proposed leakage rate is optimal by showing achievability and
a converse. As a preliminary step, we first show that under
assumption (A), the objective can be rewritten in a simpler
but equivalent form. To show achievability, we show that the
proposed optimal policy belongs to a class of policies that
satisfies a certain invariance property. Using this property the
multi-letter mutual information expression can be reduced into
a single-letter expression. For the converse we provide two
proofs: the first uses dynamic programming and the second
uses purely probabilistic and information theoretic arguments.
A. Simplification of the dynamic program
Define
Then, under Assumption (A), we can simplify the belief state
πt as follows:
πt (x, s) = P(Xt = x, St = s|Y t−1 = y t−1 , At−1 = at−1 )
(a)
= P(Xt = x|St = s, Y t−1 = y t−1 , At−1 = at−1 )
× P(St = s | Y t−1 = y t−1 , At−1 = at−1 )
(b)
= PX (x)θt (s)
where (a) follows from the product rule of probability and (b)
uses Assumption (A) and the definition of θt .
Since πt (x, s) = PX (x)θt (s), in principle, we can simplify
the dynamic program of Theorem 2 by using θt as an information state. However, for reasons that will become apparent, we
provide an alternative simplification that uses an information
state ξt ∈ PW .
Recall that Wt = St − Xt which takes values in W =
{s − x : s ∈ S, x ∈ X }. For any realization (y t−1 , at−1 ) of
past observations and actions, define ξt ∈ PW as follows: for
any w ∈ W,
ξt (w) = P (Wt = w | Y
t−1
=y
t−1
t−1
,A
=a
t−1
).
If Y t−1 and At−1 are random variables, then ξt is a PW valued random variable. As was the case for πt , it can be
shown that ξt does not depend on the choice of the policy f .
Lemma 3. Under Assumption (A), θt and ξt are related as
follows:
X
ξt (w) =
PX (x)θt (s).
(38)
(x,s)∈D(w)
Proof.
ξt (w) = Pf (Wt = w | Y t−1 = y t−1 , At−1 = at−1 )
= Pf (St − Xt = w | Y t−1 = y t−1 , At−1 = at−1 )
X
=
PX (x)θt (s).
(x,s)∈D(w)
Lemma 4. Given a ∈ A and π ∈ PX,S , define the following:
P
• ξ ∈ PW as ξ(w) =
(x,s)∈D(w) π(x, s)
• b ∈ B as follows: for all y ∈ Y, w ∈ W
P
(x,s)∈D(w) a(y | x, s)π(x, s)
;
b(y | w) =
ξ(w)
•
ã ∈ A as follows: for all y ∈ Y, x ∈ X , s ∈ S
ã(y|x, s) = b(y|s − x).
θt (s) = Pf (St = s | Y t−1 = y t−1 , At−1 = at−1 ).
f
Since πt (x, s) = PX (x)θt (s), Lemma 3 shows that ξt is a
function of πt . We will show that we can simplify the dynamic
program of Theorem 2 by using ξt as the information state
instead of πt . For such a simplification to work, we would
have to use charging policies of the form qt (yt |wt , y t−1 ). We
establish that restricting attention to such policies is without
loss of optimality. For that matter, define B as follows:
B = b ∈ PY |W : b(Y◦ (w) | w) = 1, ∀w ∈ W .
(39)
Then under Assumption (A), we have
1) Invariant Transitions: for any y ∈ Y, ϕ(π, y, a) =
ϕ(π, y, ã).
2) Lower Cost: I(a; π) ≥ I(ã; π) = I(b; ξ).
Therefore, in the sequential problem of Sec. III-B, there is no
loss of optimality in restricting attention to actions b ∈ B.
Proof. 1) Suppose (X, S) ∼ π and W = S − X, S+ =
W + Y , X+ ∼ PX . We will compare P(S+ |Y ) when
Y ∼ a(·|X, S) with when Y ∼ ã(·|X, S). Given w ∈ W
and y ∈ Y,
X
Pa (W = w, Y = y) =
a(y|x, s)π(x, s)
(x,s)∈D(w)
=
X
b(y|w)π(x, s)
(x,s)∈D(w)
(a)
=
X
ã(y|x, s)π(x, s)
(x,s)∈D(w)
ã
= P (W = w, Y = y)
(40)
where (a) uses that for all (x, s) ∈ D(w), s − x = w.
Marginalizing (40) over W , we get that Pa (Y = y) =
Pã (Y = y). Since S+ = W + Y , Eq. (40) also implies
Pa (S+ = s, Y = y) = Pã (S+ = s, Y = y). Therefore,
Pa (S+ = s|Y = y) = Pã (S+ = s|Y = y).
2) Let (X, S) ∼ π and W = S − X. Then W ∼ ξ.
Therefore, we have
I(a; π) = I a (X, S; Y ) ≥ I a (W ; Y ).
where the last inequality is the data-processing inequality. Under ã, (X, S) − W − Y , therefore,
I(ã; π) = I ã (X, S; Y ) = I ã (W ; Y ).
Now, by construction, the joint distribution of (W, Y ) is
the same under a, ã, and b. Thus,
I a (W ; Y ) = I ã (W ; Y ) = I b (W ; Y ).
11
Note that I b (W ; Y ) can also be written as I(b; ξ). The
result follows by combining all the above relations.
Once attention is restricted to actions b ∈ B, the update of
ξt may be expressed in terms of b ∈ B as follows:
Lemma 5. For any realization yt of Yt and bt of Bt , ξt+1 is
given by
ξt+1 = ϕ̃(ξt , yt , bt )
(41)
where ϕ̃ is given by
ϕ̃(ξ, y, b)(w+ )
P
=
x∈X ,w∈W
PX (x)1w+ {y + w − x}b(y | w)ξ(w)
P
.
w∈W b(y | w)ξ(w)
Proof. The proof is similar to the proof of Lemma 1.
For any b ∈ B and ξ ∈ PW , let us define the Bellman
operator B̃b : [PW → R] → [PW → R] as follows: for any
Ṽ : PW → R and any ξ ∈ PW ,
X
[B̃b Ṽ ](ξ) = I(b; ξ) +
ξ(w)b(y | w)Ṽ ϕ̃(ξ, y, b) .
y∈Y,w∈W
Theorem 4. Under assumption (A), there is no loss of optimality in restricting attention to optimal policies of the form
qt (yt |wt , ξt ) in Problem A.
1) For the finite horizon case, we can identify the optimal
policy q∗ = (q1∗ , . . . , qT∗ ) by iteratively defining value
functions Ṽt : PW → R. For any ξ ∈ PW , ṼT +1 (ξ) = 0,
and for t = T, T − 1, . . . , 1,
Ṽt (ξ) = min[B̃b Ṽt+1 ](ξ).
(42)
b∈B
Let ft◦ (ξ) denote the arg min of the right hand side
of (42). Then, the optimal policy q∗ = (q1∗ , . . . , qT∗ ) is
given by
qt∗ (yt |wt , ξt )
= bt (yt |wt ), where bt =
ft◦ (ξt ).
Moreover, the optimal (finite horizon) leakage
rate
is given by Ṽ1 (ξ1 )/T , where ξ1 (w)
=
P
P
(x)P
(s).
X
S
1
(x,s)∈D(w)
2) For the infinite horizon, suppose that there exists a
constant J˜ ∈ R and a function ṽ : PS → R which
satisfies the following fixed point equation:
J˜ + ṽ(ξ) = min[B̃b ṽ](ξ), ∀ξ ∈ PW .
b∈B
(43)
◦
Let f (ξ) denote the arg min of the right hand side
of (43). Then, the time-homogeneous policy q∗ =
(q ∗ , q ∗ , . . . ) given by
q ∗ (yt |wt , ξt ) = bt (yt |wt ), where bt = f ◦ (ξt )
is optimal. Moreover, the optimal (infinite horizon) leak˜
age rate is given by J.
Proof. Lemma 5 implies that {ξt }t≥1 is a controlled Markov
process with control action bt . Lemma 4, part 2), implies that
the per-step cost can be written as
X
T
1
E
I(bt ; ξt ) .
T
t=1
Thus, by standard results in Markov decision theory [21], the
optimal solution is given by the dynamic program described
above.
B. Weak achievability
To simplify the analysis, we assume that we are free to
choose the initial distribution of the state of the battery, which
could be done by, for example, initially charging the battery
to a random value according to that distribution. In principle,
such an assumption could lead to a lower achievable leakage
rate. For this reason, we call it weak achievability. In the next
section, we will show achievability starting from an arbitrary
initial distribution, which we call strong achievability.
Definition 1 (Constant-distribution policy). A timehomogeneous policy f ◦ = (f ◦ , f ◦ , . . . ) is
called a
constant-distribution policy if for all ξ ∈ PW , f ◦ (ξ) is a
constant. If f ◦ (ξ) = b◦ , then with a slight abuse of notation,
we refer to b◦ = (b◦ , b◦ , . . . ) as a constant-distribution
policy.
Recall that under a constant-distribution policy b ∈ B, for
any realization y t of Y t , θt and ξt are given as follows:
θt (s) = P(St = s | Y t−1 = y t−1 , B t−1 = bt−1 )
ξt (w) = P(Wt = w | Y t−1 = y t−1 , B t−1 = bt−1 ).
1) Invariant Policies: We next impose an invariance property on the class of policies. Under this restriction the leakage
rate expression will simplify substantially. Subsequently we
will show that the optimal policy belongs to this restricted
class.
Definition 2 (Invariance Property). For a given distribution θ1
of the initial battery state, a constant-distribution policy b ∈ B
is called an invariant policy if for all t, θt = θ1 and ξt = ξ1 ,
where ξ1 and θ1 satisfy (38).
Remark 3. An immediate implication of the above definition is that under any invariant policy b, the conditional
distribution Pb (Xt , St , Yt |Y t−1 ) is the same as the joint
distribution Pb (X1 , S1 , Y1 ). Marginalizing over (X, S) we get
that {Yt }t≥1 is an i.i.d. sequence.
Lemma 6. If the system starts with an initial distribution θ of
the battery state, and ξ and θ satisfy (38), then an invariant
policy b = (b, b, . . . ) corresponding to (θ, ξ) achieves a
leakage rate
LT (b) = I(W1 ; Y1 ) = I(b; ξ)
for any horizon T .
Proof. The proof is a simple corollary of the invariance property in Definition 2. Recall from the dynamic program of Theorem 4, that the performance of any policy b = (b1 , b2 , . . . )
such that bt ∈ B, is given by
X
T
1
I(bt ; ξt ) .
LT (b) = E
T
t=1
12
Now, we start with an initial distribution ξ1 = ξ and follow
the constant-distribution structured policy b = (b, b, . . . ).
Therefore, by Definition 2, ξt = ξ for all t. Hence, the leakage
rate under policy b is
LT (b) = I(b; ξ).
LT (b) =
1
T
I b (St , Xt ; Yt |Y t−1 ).
Pb (S2 = s2 , Y1 = y1 ) = Pb (S2 = s2 , Y1 = y1 , W1 = s2 − y1 )
= ξ1 (s2 − y1 )b(y1 |s2 − y1 )
Remark 4. Note that Lemma 6 can be easily derived independently of Theorem 4. From Proposition 1, we have that:
T
X
of S2 and y1 of Y1 . This means that w1 = s2 − y1 . Since Y1 is
chosen according to b(·|w1 ), it must be that y1 ∈ X ∩ Y◦ (w1 ).
Therefore,
(44)
t=1
From Remark 3, we have that I b (St , Xt ; Yt |Y t−1 ) =
I b (S1 , X1 ; Y1 ) = I b (W1 ; Y1 ), which immediately results in
Lemma 6.
For invariant policies we can express the leakage rate in the
following fashion, which is useful in the proof of optimality.
Lemma 7. For any invariant policy b,
I b (W1 ; Y1 ) = I b (W1 ; X1 ).
Proof. Consider the following sequence of simplifications:
I b (W1 ; Y1 ) = H b (W1 ) − H b (W1 |Y1 )
= H b (W1 ) − H b (W1 + Y1 |Y1 )
(a)
= H b (W1 ) − H b (S2 |Y1 )
(b)
= H b (W1 ) − H b (S1 )
(c)
= H b (W1 ) − H b (S1 |X1 )
(d)
= H b (W1 ) − H b (W1 |X1 )
= I b (W1 ; X1 ).
where (a) is due to the battery update equation (1); (b) is
because b is an invariant ; (c) is because S1 and X1 are
independent; and (d) is because S1 = W1 + X1 .
2) Structured Policy: We now introduce a class of policies
that satisfy the invariance property in Def. 2. This will be then
used in the proof of Theorem 3.
Definition 3 (Structured Policy). Given θ ∈ PS and ξ ∈ PW ,
a constant-distribution policy b = (b, b, . . . ) is called a
structured policy with respect to (θ, ξ) if:
(
PX (y) θ(y+w)
ξ(w) , y ∈ X ∩ Y◦ (w)
b(y|w) =
0,
otherwise.
Note that it is easy to verify that the distribution b defined
above is a valid conditional probability distribution.
Lemma 8. For any θ ∈ PS and ξ ∈ PW , the structured policy
b = (b, b, . . . ) given in Def. 3 is an invariant policy.
Proof. Since (θt , ξt ) are related according to Lemma 3, in
order to check whether a policy is invariant it is sufficient to
check that θt = θ1 for all t. Furthermore, to check if a timehomogeneous policy is an invariant policy, it is sufficient to
check that either θ2 = θ1 .
Let the initial distributions (θ1 , ξ1 ) = (θ, ξ) and the system
variables be defined as usual. Now consider a realization s2
= PX (y1 )θ1 (s2 ),
(45)
where in the last equality we use the fact that y1 ∈ X ∩
Y◦ (s2 − y1 ). Note that if y1 6∈ X ∩ Y◦ (s2 − y1 ), then Pb (S2 =
s2 , Y1 = y1 ) = 0. Marginalizing over s2 , we get Pb (Y1 =
y1 ) = PX (y1 ).
Consequently, θ2 (s2 ) = Pb (S2 = s2 |Y1 = y1 ) = θ1 (s2 ).
Hence, b is invariant as required.
Remark 5. As argued in Remark 3, under any invariant policy,
{Yt }t≥1 is an i.i.d. sequence. As argued in the proof of
Lemma 8, for a structured policy the marginal distribution of
Yt is PX . Thus, an eavesdropper cannot statistically distinguish
between {Xt }t≥1 and {Yt }t≥1 .
Proposition 3. Let θ∗ , ξ ∗ , and b∗ be as defined in Theorem 3.
Then,
1) (θ∗ , ξ ∗ ) satisfy (38);
2) b∗ is a structured policy with respect to (θ∗ , ξ ∗ ).
3) If the system starts in the initial battery state θ∗ and follows the constant-distribution policy b∗ = (b∗ , b∗ , . . . ),
the leakage rate is given by J ∗ .
Thus, the performance J ∗ is achievable.
Proof. The proofs of parts 1) and 2) follows from the definitions. The proof of part 3) follows from Lemmas 6 and 7.
This completes the proof of the achievability of Theorem 3.
C. Strong achievability
Lemma 9. Assume that for any x ∈ X , PX (x) > 0. Let
(θ◦ , ξ ◦ ) be a pair satisfying (38) and b◦ = (b◦ , b◦ , . . . ) be
the corresponding structured policy.
Assume that θ◦ ∈ int(PS ) or equivalently, for any w ∈ W
and y ∈ X ∩ Y◦ (w), b◦ (y|w) > 0. Suppose the system starts
in the initial state (θ1 , ξ1 ) and follows policy b◦ . Then:
1) the process {θt }≥1 converges weakly to θ◦ ;
2) the process {ξt }≥1 converges weakly to ξ ◦ ;
3) for any continuous function c : PW → R,
T
1X
E[c(ξt )] = c(ξ ◦ ).
T →∞ T
t=1
lim
(46)
4) Consequently, the infinite horizon leakage rate under
b◦ is
L∞ (b◦ ) = I(b◦ , ξ ◦ ).
Proof. The proof of parts 1) and 2) is presented in Appendix D. From 2), limt→∞ E[c(ξt )] = c(ξ ◦ ), which implies (46). Part 4) follows from part 3) by setting c(ξt ) =
I(b◦ , ξt ).
Proposition 1 implies that θ∗ defined in Theorem 3 lies in
int(PS ). Then, by Lemma 9, the constant-distribution policy
b∗ = (b∗ , b∗ , . . . ) (where b∗ is given by Theorem 3), achieves
13
the leakage rate I(b∗ , ξ ∗ ). By Lemma 7, I(b∗ , ξ ∗ ) is same as
J ∗ defined in Theorem 3. Thus, J ∗ is achievable starting from
any initial state (θ1 , ξ1 ).
Now consider
I(Wt ; Yt |Y t−1 ) = H(Wt |Y t−1 ) − H(Wt |Y t )
= H(Wt |Y t−1 ) − H(Wt + Yt |Y t )
(b)
= H(Wt |Y t−1 ) − H(St+1 |Y t )
D. Dynamic programming converse
(c)
We provide two converses. One is based on the dynamic
program of Theorem 4, which is presented in this section;
the other is based purely on information theoretic arguments,
which is presented in the next section.
In the dynamic programming converse, we show that for J ∗
given in Theorem 3, v ∗ (ξ) = H(ξ), and any b ∈ B,
J ∗ + v ∗ (ξ) ≤ [B̃b v ∗ ](ξ),
∀ξ ∈ PW ,
(47)
Since H(ξ) is bounded, from [21, Lemma 5.2.5(b)], we get
that J ∗ is a lower bound of the optimal leakage rate.
To prove (47), pick any ξ ∈ PW and b ∈ B. Suppose W1 ∼
ξ, Y1 ∼ b(·|W1 ), S2 = Y1 + W1 , X2 is independent of W1
and X2 ∼ PX and W2 = S2 − X2 . Then,
X
[B̃b v ∗ ](ξ) = I(b; ξ) +
ξ(w1 )b(y1 |w1 )v ∗ (ϕ̂(ξ, y1 , b))
= H(Wt |Y t−1 ) − H(St+1 |Y t , Xt+1 )
= H(Wt |Y t−1 ) − H(St+1 − Xt+1 |Y t , Xt+1 )
(d)
= H(Wt |Y t−1 ) − H(Wt+1 |Y t , Xt+1 )
where (b) follows from (1); (c) follows because of assumption (A); and (d) also follows from (1).
Substituting (51) in (50) (but expanding the last term as
H(WT |Y T −1 ) − H(WT |Y T ), we get
I(S1 , X T ; Y T ) ≥
=
T
X
t=2
T
X
(54)
I(Wt ; Xt |Y t−1 ) − H(WT |Y T ).
S̃2 ∼ θ2
(49)
where (a) uses S2 = Y1 + W1 and W2 = S2 − X2; (b) uses
the fact that H(A
1 |B) − H(A1 − A2 |B) ≥ minPA1 H(A1 ) −
H(A1 − A2 ) for any joint distribution on (A1 , A2 , B).
The equality in (49) occurs when b is an invariant policy
and θ2 is same as θ∗ defined in Theorem 3. For ξ that are not
equivalent to θ∗ via (38), the inequality in (49) is strict.
We have shown that Eq. (47) is true. Consequently, J ∗ is a
˜
lower bound on the optimal leakage rate J.
Now, we take the limit T → ∞ to obtain a lower bound to
the leakage rate:
1
L∞ (q) = lim sup I(S1 , X T ; Y T )
T →∞ T
T
X
1
t−1
T
≥ lim sup
H(W1 ) +
I(Wt ; Xt |Y
) − H(WT |Y )
T →∞ T
t=2
" T
#
1 X
(a)
t−1
= lim supT →∞
I(Wt ; Xt |Y
)
T t=2
(b)
≥ min I(S − X; X) = J ∗
PS ∈PS
where (a) is because the entropy of any discrete random
variable is bounded and (b) follows from the fact that each
term in the summation satisfies:
I(Wt ; Xt |Y t−1 ) ≥ min I(S − X; X),
PS ∈PS
E. Information theoretic converse
Consider the following inequalities: for any admissible
policy q ∈ QB , we have
I(S1 , X T ; Y T ) =
(a)
≥
t=1
T
X
(56)
which can be justified as follows. First note that for any policy
q we have that:
I(Wt ; Xt |Y t−1 ) = I(St − Xt ; Xt |Y t−1 )
depends on the joint distribution PSqt ,Xt ,Y t−1 (st , xt , y t−1 )
which factors as:
I(St , Xt ; Yt |Y t−1 )
I(Wt ; Yt |Y t−1 )
(55)
t=2
(a)
T
X
(53)
− H(Wt |Y t−1 , Xt ) + H(Wt |Y t−1 )
− H(WT |Y T )
= H(W1 ) +
= H(W2 |Y1 ) − H(W1 + Y1 |Y1 )
=J
(52)
H(Wt |Y t−1 ) − H(Wt+1 |Y t , Xt+1 )
= H(W1 ) +
(48)
[B̃b v ](ξ) − v (ξ) = H(W2 |Y1 ) − H(W1 |Y1 )
θ2 ∈PS
∗
T
−1
X
+H(WT |Y T −1 ) − H(WT |Y T )
∗
= H(S2 − X2 |Y1 ) − H(S2 |Y1 )
(b)
≥ min H(S̃2 − X2 ) − H(S̃2 ) ,
I(Wt ; Yt |Y t−1 )
t=1
where the second equality is due to the definition of conditional entropy. Consequently,
∗
T
X
t=1
(w1 ,y1 )∈W×Y
= I(W1 ; Y1 ) + H(W2 |Y1 )
(51)
(50)
t=1
where (a) follows from the fact that Wt = Xt − St is a deterministic function of (Xt , St ) and that the mutual information
is non-negative.
PSqt ,Xt ,Y t−1 (st , xt , y t−1 ) = PX (xt )PSqt ,Y t−1 (st , y t−1 ) (57)
as Xt is sampled i.i.d. from the distribution PX (·) and from
the state update equation (1), we have that St is a function of
(X t−1 , Y t−1 ), and thus independent of Xt . Now note that:
I(Wt ; Xt |Y t−1 )
14
=
X
2
;X)
> 0.
I(W ; X) is strictly convex on PS we require d I(W
dα2
Due to independence of X and S, I(W ; X) = H(W )−H(S).
Therefore,
I(St − Xt ; Xt |Y t−1 = y t−1 )p(y t−1 )
y t−1 ∈Y t−1
≥
≥
min
y t−1 ∈Y t−1
I(St − Xt ; Xt |Y t−1 = y t−1 )
min
min
y t−1 ∈Y t−1 pS
t−1 (·|y
t ,Xt |Y
≥
min
y t−1 ∈Y t−1 ,pS
t
|Y t−1 (·|y
t−1 )
t−1 )
I(St − Xt ; Xt |Y t−1 = y t−1 )
I(St − Xt ; Xt |Y t−1 = y t−1 )
(58)
where (58) follows from (57) so that Xt is independent of
(St , Y t−1 ) and distributed according to PX (·) Since (58) is
simply equivalent to minimizing over the distribution PS (·),
the relation in (56) holds. This completes the proof. This shows
that J ∗ is a lower bound to the minimum (infinite horizon)
leakage rate.
V. C ONCLUSIONS AND D ISCUSSION
In this paper, we study a smart metering system that uses
a rechargeable battery to partially obscure the user’s power
demand. Through a series of reductions, we show that the
problem of finding the best battery charging policy can be
recast as a Markov decision process. Consequently, the optimal
charging policies and the minimum information leakage rate
are given by the solution of an appropriate dynamic program.
For the case of i.i.d. demand, we provide an explicit
characterization of the optimal battery policy and the leakage
rate. In this special case it suffices to choose a memoryless
policy where the distribution of Yt depends only on Wt . Our
achievability results rely on restricting attention to a class of
invariant policies. Under an invariant policy, the consumption
{Yt }t≥1 is i.i.d. and the leakage rate is characterized by a
single-letter mutual information expression. We then further
restrict attention to what we call structured policies under
which the marginal distribution of {Yt }t≥1 is PX . Thus, under
the structured policies, an eavesdropper cannot statistically
distinguish between {Xt }t≥1 and {Yt }t≥1 . We provide two
converses; one is based on the dynamic programming argument while the other is based on a purely information theoretic
argument.
Extending of our MDP formulation to incorporate an
additive cost, such as the price of consumption, is rather
immediate. However, the approach presented in this work
for explicitly characterizing the optimal leakage rate in the
i.i.d. case may not immediately extend to such general cost
functions. In another direction one can allow for a certain
controlled wastage of energy drawn from the grid to increase
privacy. It would be interesting to see how the leakage rate
decreases with the wasted energy. The study of such problems,
as well as finer implementation details of the proposed system,
remains an interesting future direction.
d [−H(S) + H(W )]
dI(W ; X)
=
dα
dαX
X
δ(s̃) ln θα (s̃) −
PX (s − w)δ(s) ln PW (w)
=
s̃
w∈W,s∈S
X
d I(W ; X) X δ(s)2
=
−
2
dα
θα (s)
s
2
P
s̃∈S
w∈W
2
PX (s̃ − w)δ(s̃)
.
PW (w)
q
p
Let aw (s) = δ(s) PXθα(s−w)
and bw (s) = θα(s) PX (s − w).
(s)
Using the Cauchy-Schwarz inequality, we can show that
2
P
X
d2 I(W ; X) X δ(s)2
s̃∈S aw (s̃)bw (s̃)
=
−
dα2
θα (s)
PW (w)
s
w∈W
P
P
2
2
X δ(s)2
X
s̃∈S aw (s̃)
ŝ∈S bw (ŝ)
>
−
θα (s)
PW (w)
s
w∈W
!
X X
X δ(s)2
−
aw (s̃)2 = 0.
=
θα (s)
s
w∈W
s̃∈S
The strict inequality is because a and b cannot be linearly
δ(s)
dependent. To see this, observe that a(s)
b(s) = θ(s)+αδ(s) cannot
be equal to a constant for all s ∈ S since δ must contain
negative as well as positive elements.
A PPENDIX B
P ROOF OF P ROPOSITION 1
The proof of Proposition 1 relies on the following intermediate results (which are proved later):
Lemma B.1. For any q ∈ QA ,
q
T
T
I (S1 , X ; Y ) ≥
T
X
I q (Xt , St ; Yt |Y t−1 )
t=1
with equality if and only if q ∈ QB .
Lemma B.2. For any qa ∈ QA , there exists a qb ∈ QB , such
that
T
X
t=1
I qa (Xt , St ; Yt |Y t−1 ) =
T
X
I qb (Xt , St ; Yt |Y t−1 ).
t=1
Combining Lemmas B.1 and B.2, we get that for any qa ∈
QA , there exists a qb ∈ QB such that
A PPENDIX A
P ROOF OF P ROPERTY 1
I qa (S1 , X T ; Y T ) ≥ I qb (S1 , X T ; Y T ).
PFor any θ ∈ int(PS ) and δ(s) : S → R such that
s∈S δ(s) = 0. Let θα (s) := θ(s) + αδ(s). Then for small
enough α, θα ∈ PS . Given such a θα , let PW,X (w, x) =
PW |X (w|x)PX (x) = θα (w + x)PX (x). Then to show that
Therefore there is no loss of optimality in restricting attention
to charging policies in QB . Furthermore, Lemma B.1 shows
that for any q ∈ QB , LT (q) takes the additive form as given
in the statement of the proposition.
15
Proof of Lemma B.1. For any q ∈ QA , we have
(a)
n
X
(b)
t=1
n
X
I q (S1 , X n ; Y n ) =
=
(c)
≥
t=1
n
X
A PPENDIX C
P ROOF OF CONCAVITY OF THE VALUE FUNCTION
I q (S1 , X t ; Yt |Y t−1 )
q
t
t
I (X , S ; Yt |Y
q
I (Xt , St ; Yt |Y
t−1
t−1
To prove the result, we show the following:
Lemma C.1. For any action a ∈ A, if V : PX,S → R is
concave, then Ba V is concave on PX,S .
)
The concavity of the value functions follows from backward
induction. VT +1 is a constant and, therefore, also concave.
Lemma C.1 implies that VT , VT −1 , . . . , V1 are concave.
)
t=1
where (a) uses the chain rule of mutual information and the
fact that (S t−2 , Y t−1 ) → Xt−1 → Xt ;5 (b) uses the fact that
the battery process S t is a deterministic function of S1 , X t−1 ,
and Y t−1 given by (1); and (c) uses the fact that removing
terms does not reduce the mutual information.
Proof of Lemma B.2. For any qa = (q1a , q2a , . . . , qTa ) ∈ QA ,
construct a qb = (q1b , q2b , . . . , qTb ) ∈ QB as follows: for any t
and realization (xt , st , y t ) of (X t , S t , Y t ) let
Proof of Lemma C.1. The first term I(a; π) of [Ba V ](π) is
a concave function of π. We show the same for the second
term.
Note that if a function V is concave, then it’s perspective g(u, t) := tV (u/t) is concave in the domain
{(u, t) : u/t ∈ Dom(V ), t > 0}. The second term in the definition of the Bellman operator (10)
X X
a(y|x, s)π(x, s) V (ϕ(π, y, a))
y∈Y
qtb (yt |xt , st , y t−1 ) = PqYa|Xt ,St ,Y t−1 (yt |xt , st , y t−1 ).
t
(59)
To prove the Lemma, we show that for any t,
PqXat ,St ,Y t = PqXbt ,St ,Y t .
(60)
By definition of qb given by (59), to prove (60), it is sufficient
to show that
PqXat ,St ,Y t−1 = PqXbt ,St ,Y t−1 .
(xt ,st )∈X ×S
=
X
Q(xt+1 |xt )1st+1 {st − xt + yt }qa (yt |xt , st , y t−1 )
(xt ,st )∈X ×S
qa
× P (xt , st , y t−1 )
(a) X
=
Q(xt+1 |xt )1st+1 {st − xt + yt }qb (yt |xt , st , y t−1 )
(xt ,st )∈X ×S
qb
× P (xt , st , y t−1 )
qb
has this form because the
Pnumerator of ϕ(π, y, a) is linear in π
and the denominator is x,s a(y|x, s)π(x, s) (and corresponds
to t in the definition of perspective). Thus, for each y, the
summand is concave in π, and the sum of concave functions
is concave. Hence, the second term of the Bellman operator
is concave in π. Thus we conclude that concavity is preserved
under Ba .
(61)
We do so using induction.
For t = 1, PqXa1 ,S1 (x, s) = PX1 (x)PS1 (s) = PqXb1 ,S1 (x, s).
This forms the basis of induction. Now assume that (61) hold
for t.
In the rest of the proof, for ease of notation, we denote
PqXat+1 ,St+1 ,Y t (xt+1 , st+1 , y t ) simply by Pqa (xt+1 , st+1 , y t ).
For t + 1, we have
X
Pqa (xt+1 , st+1 , y t ) =
Pqa (xt+1 , xt , st+1 , st , y t )
A PPENDIX D
P ROOF OF L EMMA 9
The proof of the convergence of {ξt }t≥1 relies on a result
on the convergence of partially observed Markov chains due
to Kaijser [33] that we restate below.
Definition 4. A square matrix D is called subrectangular if for
every pair of indices (i1 , j1 ) and (i2 , j2 ) such that Di1 ,j1 6= 0
and Di2 ,j2 6= 0, we have that Di2 ,j1 6= 0 and Di1 ,j2 6= 0.
Theorem 5 (Kaijser [33]). Let {Ut }t≥1 , Ut ∈ U, be a finite
state Markov chain with transition matrix P u . The initial state
U1 is distributed according to probability mass function PU1 .
Given a finite set Z and an observation function g : U → Z,
define the following:
• The process {Zt }t≥1 , Zt ∈ Z, given by
Zt = g(Ut ).
•
t
= P (xt+1 , st+1 , y )
where (a) uses (59) and the induction hypothesis. Thus, (61)
holds for t + 1 and, by the principle of induction, holds for
all t. Hence (60) holds and, therefore, I qa (Xt , St ; Yt |Y t−1 ) =
I qb (Xt , St ; Yt |Y t−1 ). The statement in the Lemma follows by
adding over t.
5 The notation A → B → C is used to indicate that A is conditionally
independent of C given B.
(x,s)∈X ×S
The process {ψt }t≥1 , ψt ∈ PU , given by
ψt (u) = P(Ut = u | Z t ).
•
A square matrix M (z), z ∈ Z, given by
(
Piju if g(j) = z
[M (z)]i,j =
0
otherwise
i, j ∈ U.
Qm
If there exists a finite sequence z1m such that t=1 M (zt )
is subrectangular, then {ψt }t≥1 converges in distribution to a
limit that is independent of the initial distribution PU1 .
16
We will use the above theorem to prove that under policy
b , {ξt }t≥1 converges to a limit. For that matter, let U =
S × Y, Z = Y, Ut = (St , Yt−1 ) and g(St , Yt−1 ) = Yt−1 .
First, we show that {Ut }t≥1 is a Markov chain. In particular,
for any realization (st+1 , y t ) of (S t+1 , Y t ), we have that
Since the sequnce of demands xm has a positive probability,
◦
◦
Pb (Ut+1 = (st+1 , yt ) | U t = (st , y t−1 ))
X
=
P (Ut+1 = (st+1 , yt ), Xt = x̃t | U t = (st , y t−1 ))
=
Y ms −1 = (111 . . . 1)) > 0.
Therefore,
P(Sms = ms | U1 = (s, y), Y ms −1 = (111 . . . 1)) > 0
which completes the proof.
x̃t ∈X
X
P(Sms = ms , X ms = xms | U1 = (s, y),
∗
1st+1 {yt + st − x̃t }b (yt |st − x̃t )PX (x̃t )
x̃t ∈X
b◦
= P (Ut+1 = (st+1 , yt ) | Ut = (st , yt−1 )).
Next, let m = 2ms and consider
z m = |111{z
· · · 1} 000
· · · 0} .
| {z
ms times
ms times
We will show that this z m satisfies the subrectangularity
condition of Theorem 5. The basic idea is the following.
Consider any initial state u1 = (s, y) and any final state
um = (s0 , 0). We will show that
P(Sms = ms | U1 = (s, y), Z ms = (111 . . . 1)) > 0,
(62)
and
ms
P(S2ms = s0 | Ums = (sm , 1), Zm
= (000 . . . 0)) > 0.
s +1
(63)
Eqs. (62) and (63) show that given the observation sequence
z m , for any initial state (s, y) there is a positive probability of observing any final state (s0 , 0).6 Hence, the matrix
Q
m
t=1 M (z) is subrectangular. Consequently, by Theorem 5,
the process {ψt }t≥1 converges in distribution to a limit that
is independent of the initial distribution
PU1 .
P
Now observe that θt (s) =
ψ
y∈Y t (s, y) and (θt , ξt )
are related according to Lemma 3. Since {ψt }t≥1 converges
weakly independent of the initial condition, so do {θt }t≥1 and
{ξt }t≥1 .
Let θ̄ and ξ¯ denote the limit of {θt }t≥1 and {ξt }t≥1 .
Suppose the initial condition is (θ◦ , ξ ◦ ). Since b◦ is an
invariant policy, (θt , ξt ) = (θ◦ , ξ ◦ ) for all t. Therefore, the
¯ = (θ◦ , ξ ◦ ).
limits (θ̄, ξ)
Proof of Eq. (62). Given the initial state (s, y), define s̄ =
ms − s, and consider the sequence
· · · 1} .
xms = |000{z
· · · 0} |111{z
s̄ times
s times
Under this sequence of demands, consider the sequence of
consumption y ms −1 = (11 . . . 1), which is feasible because
the state of the battery increases by 1 for the first s̄ steps (at
which time it reaches ms ) and then remains constant for the
remaining s steps. Therefore,
P(Sms = ms | U1 = (s, y),
Y ms −1 = (111 . . . 1), X ms = xms ) > 0.
6 Note that given the observation sequence z m , the final state must be of
the form (s0 , 0).
Proof of Eq. (63). The proof is similar to the Proof of (62).
Given the final state (s0 , 0), define s̄0 = ms − s0 and consider
the sequence
s
x2m
111{z
· · · 1} 000
· · · 0} .
ms +1 = |
| {z
s̄0 times
s0 times
Under this sequnce of demains and the sequence of consump2ms −1
= (00 . . . 0), the state of the battery
tion given by ym
s
decreases by 1 for the first s̄0 steps (at which time it reaches
s0 ) and then remains constant for the remaining s0 steps. Since
s
x2m
ms +1 has positive probability, we can complete the proof by
following an argument similar to that in the proof of (62).
R EFERENCES
[1] A. Ipakchi and F. Albuyeh, “Grid of the future,” IEEE Power Energy
Mag., vol. 7, no. 2, pp. 52–62, 2009.
[2] A. Prudenzi, “A neuron nets based procedure for identifying domestic
appliances pattern-of-use from energy recordings at meter panel,” in
Proc. Power Engineering Society Winter Meeting, vol. 2. IEEE, 2002,
pp. 941–946.
[3] G. W. Hart, “Nonintrusive appliance load monitoring,” Proc. IEEE,
vol. 80, no. 12, pp. 1870–1891, 1992.
[4] H. Y. Lam, G. Fung, and W. Lee, “A novel method to construct taxonomy
electrical appliances based on load signatures,” IEEE Trans. Consum.
Electron., vol. 53, no. 2, 2007.
[5] C. Efthymiou and G. Kalogridis, “Smart grid privacy via anonymization
of smart metering data,” in Proc. Smart Grid Communications. IEEE,
2010, pp. 238–243.
[6] G. Kalogridis, C. Efthymiou, S. Z. Denic, T. A. Lewis, and R. Cepeda,
“Privacy for smart meters: Towards undetectable appliance load signatures,” in Proc. Smart Grid Communications. IEEE, 2010, pp. 232–237.
[7] I. Csiszar and J. Körner, Information theory: coding theorems for
discrete memoryless systems. Cambridge University Press, 2011.
[8] D. Varodayan and A. Khisti, “Smart meter privacy using a rechargeable
battery: Minimizing the rate of information leakage,” in Proc. Int. Conf.
on Acoustics, Speech and Signal Processing. IEEE, 2011, pp. 1932–
1935.
[9] O. Tan, D. Gunduz, and H. V. Poor, “Increasing smart meter privacy
through energy harvesting and storage devices,” IEEE J. Sel. Areas
Commun., vol. 31, no. 7, pp. 1331–1341, 2013.
[10] G. Giaconi, D. Gündüz, and H. V. Poor, “Smart meter privacy with an
energy harvesting device and instantaneous power constraints,” in IEEE
International Conference on Communications (ICC). IEEE, 2015, pp.
7216–7221.
[11] S. Han, U. Topcu, and G. J. Pappas, “Event-based information-theoretic
privacy: A case study of smart meters,” in Proc. American Control
Conference. IEEE, 2016, pp. 2074–2079.
[12] L. Bahl, J. Cocke, F. Jelinek, and J. Raviv, “Optimal decoding of
linear codes for minimizing symbol error rate,” IEEE Trans. Inf. Theory,
vol. 20, no. 2, pp. 284–287, 1974.
[13] L. Sankar, S. R. Rajagopalan, and H. V. Poor, “Utility-privacy tradeoffs
in databases: An information-theoretic approach,” IEEE Trans. Inf.
Theory, vol. 8, no. 6, pp. 838–852, 2013.
[14] L. Sankar, S. R. Rajagopalan, and S. Mohajer, “Smart meter privacy:
A theoretical framework,” IEEE Trans. Smart Grid, vol. 4, no. 2, pp.
837–846, 2013.
[15] O. Tan, D. Gunduz, and H. V. Poor, “Increasing smart meter privacy
through energy harvesting and storage devices,” IEEE J. Sel. Areas
Commun., vol. 31, no. 7, pp. 1331–1341, 2013.
17
[16] S. Tatikonda and S. Mitter, “The capacity of channels with feedback,”
IEEE Trans. Inf. Theory, vol. 55, no. 1, pp. 323–349, 2009.
[17] A. J. Goldsmith and P. P. Varaiya, “Capacity, mutual information,
and coding for finite-state Markov channels,” IEEE Trans. Inf. Theory,
vol. 42, no. 3, pp. 868–886, 1996.
[18] H. Permuter, P. Cuff, B. Van Roy, and T. Weissman, “Capacity of the
trapdoor channel with feedback,” IEEE Trans. Inf. Theory, vol. 54, no. 7,
pp. 3150–3165, 2008.
[19] J. Yao and P. Venkitasubramaniam, “The privacy analysis of battery
control mechanisms in demand response: Revealing state approach and
rate distortion bounds,” IEEE Trans. Smart Grid, vol. 6, no. 5, pp. 2417–
2425, 2015.
[20] S. Li, A. Khisti, and A. Mahajan, “Structure of optimal privacypreserving policies in smart-metered systems with a rechargeable battery,” in Proc. Int. Workshop on Signal Processing Advances in Wireless
Communications. IEEE, 2015, pp. 375–379.
[21] O. Hernández-Lerma and J. Lasserre, Discrete-Time Markov Control
Processes. Springer-Verlag, 1996.
[22] O. Hernández-Lerma, Adaptive Markov Control Processes. SpringerVerlag, 1989.
[23] ——, Further Topics on Discrete-Time Markov Control Processes.
Springer-Verlag, 1991.
[24] D. P. Bertsekas, Dynamic programming and optimal control. Athena
scientific Belmont, MA, 1995, vol. 1, no. 2.
[25] R. Blahut, “Computation of channel capacity and rate-distortion functions,” IEEE Trans. Inf. Theory, vol. 18, no. 4, pp. 460–473, 1972.
[26] S. Arimoto, “An algorithm for computing the capacity of arbitrary
discrete memoryless channels,” IEEE Trans. Inf. Theory, vol. 18, no. 1,
pp. 14–20, 1972.
[27] L. Wang and M. Madiman, “Beyond the entropy power inequality, via
rearrangements,” IEEE Trans. Inf. Theory, vol. 60, no. 9, pp. 5116–5137,
2014.
[28] S. Li, “Information theoretic privacy in smart metering systems using
rechargeable batteries,” Master’s thesis, University of Toronto (Canada),
2016.
[29] A. Dembo, T. M. Cover, and J. A. Thomas, “Information theoretic
inequalities,” IEEE Trans. Inf. Theory, vol. 37, no. 6, pp. 1501–1518,
1991.
[30] T. M. Cover and J. A. Thomas, Elements of information theory. John
Wiley & Sons, 2012.
[31] W. B. Powell, Approximate Dynamic Programming: Solving the curses
of dimensionality. John Wiley & Sons, 2007, vol. 703.
[32] G. Shani, J. Pineau, and R. Kaplow, “A survey of point-based pomdp
solvers,” Autonomous Agents and Multi-Agent Systems, pp. 1–51, 2013.
[33] T. Kaijser, “A limit theorem for partially observed Markov chains,” The
Annals of Probability, pp. 677–696, 1975.
| 7cs.IT
|
On the Evaluation of Plug-in Electric Vehicle Data
of a Campus Charging Network
arXiv:1602.02943v1 [math.OC] 9 Feb 2016
Islam Safak Bayram
Qatar Environment and Energy Research Institute
and College of Science and Engineering
Hamad Bin Khalifa University
Doha, Qatar
[email protected]
Abstract—The mass adoption of plug-in electric vehicles
(PEVs) requires the deployment of public charging stations.
Such facilities are expected to employ distributed generation and
storage units to reduce the stress on the grid and boost sustainable
transportation. While prior work has made considerable progress
in deriving insights for understanding the adverse impacts of
PEV chargings and how to alleviate them, a critical issue that
affects the accuracy is the lack of real world PEV data. As
the dynamics and pertinent design of such charging stations
heavily depend on actual customer demand profile, in this paper
we present and evaluate the data obtained from a 17 node
charging network equipped with Level 2 chargers at a major
North American University campus. The data is recorded for
166 weeks starting from late 2011. The result indicates that
the majority of the customers use charging lots to extend their
driving ranges. Also, the demand profile shows that there is a
tremendous opportunity to employ solar generation to fuel the
vehicles as there is a correlation between the peak customer
demand and solar irradiation. Also, we provided a more detailed
data analysis and show how to use this information in designing
future sustainable charging facilities.
I. I NTRODUCTION
There has been a surge of interest in plug-in electric vehicles
(PEVs), as electric miles are cheaper and environmentally
friendly when compared to their gas-powered counterparts.
This motivation is further boosted by government policies
and regulations to support energy security and independence.
For instance, the State of California has set a goal of putting
1.5 million zero-emission vehicles on the road by year 2025.
Therefore, PEV sales have risen tremendously [1]. In the last
five years more than 300, 000 vehicles have been sold in the
U.S. and almost 40% of the PEVs reside in California [2],
[3]. Therefore, such adoption rates have spurred the need to
deploy a robust network of public charging facilities.
Typically, PEV owners use garage charging with Level
1, 120 AC chargers during nighttime when there is unused
generation capacity. On the other hand, PEV owners living in
metropolitan areas may not have access to this option. Also,
the need for longer driving ranges requires the deployment of
public charging facilities. During the weekdays, PEV owners
can spend several hours parked at the workplace, school, or
while shopping. Hence, charging service at parking lots is
gaining popularity. A charging lot can be equipped with either
Vahraz Zamani, Ryan Hanna, Jan Kleissl
Department of Mechanical and Aerospace Engineering
University of California at San Diego,
La Jolla, USA
[email protected], [email protected]
[email protected]
Level 2 single/three phase or fast DC charging technology [4].
On the other hand, incorporating these technologies translates
to major phase, additions to the existing power grid. For instance under Level 1 charging conditions each PEV represents
a load equivalent to 50% of a house while under Level 2
conditions PEV represents close to 2.5 fold the equivalent
load.
The concern about the rapid adoption of PEVs by the
utilities is that they could have disruptive impacts on the power
grid. Excessive instantaneous demand can overload circuits,
increase distribution transformer losses, lead to harmonics distortion, voltage deviations, and thermal loading on distribution
system [5], [6]. Obviously, the extent of PEV impact depends
on the degree and regional density of their penetration, service
requirements and the time of the day they are charged.
For PEV adoption to occur seamlessly, the design and
operation of charging facilities are of paramount importance.
Corollary, there has been a growing body of literature on the
design of charging facilities [4], [7], [8] and efficient load management at them [9]–[12]. The station designs often consider
renewable generation (mostly solar power), energy storage,
and efficient charge rate control. For instance, the works in
[11] and [12] consider a parking lot charging facility equipped
with energy storage systems to smoothen the stochastic customer demand. The studies in [4] and [8] provide a capacity
planning framework for large-scale charging facilities. Also,
utilities have started to deploy solar-powered charging stations
[13] and [14]. Moreover, the works in [9] and [10] propose
optimization frameworks for power flow.
However, the lack of real world data limits the applicability
of such studies in the US, while some work has been done
in Europe [15], [16]. Most of the studies have to rely on
simplifying assumptions, which may not be true. For instance,
the common assumption on PEV service/parking time is either
Gaussian or Exponential distribution, however, our results
indicate that real customer behavior is far from this two
assumptions. Also critical design problems such as resource
provisioning and sizing of storage and PV panels heavily
rely on the customer statistics, e.g., peak and spatiotemporal
demand, seasonal and diurnal variations, charging duration,
etc. Therefore, in this study we evaluate the data obtained from
One Year Average Solar Power at UC San Diego
250
Level 2 chargers at a major North American University campus
for a period of 166 weeks. To the best of our knowledge, this
paper presents one of the most comprehensive data in this field.
The main thrusts of this paper are to: (1) provide basic design
and capacity planning principles for charging stations equipped
with distributed generation and storage units; (2) analyze and
present real-world PEV charging data on PEV charging; (3)
provide guidelines for how to use the measured information
to build the future sustainable charging facilities.
II. PEV C HARGING S TATION D ESIGNS
The design and architecture of sustainable charging facilities
should carefully address the potential negative impacts of the
concurrent PEV charging. The primary goal is to manage the
customer demand both in time and space so that the power
grid is not strained beyond its limits, and additional peaks
are prevented. An overview of the systems components is
presented in Fig.1. Next, we enumerate the design parameters
that enable effective demand control at PEV parking lots.
A. Renewable Energy
The recent studies conducted by Lawrence Berkeley National Laboratory and National Renewable Energy Laboratory
reveal that the price of solar photovoltaic (PV) was reduced by
12% to 19% in 2013 and the trend continues to fall 3%-12%
more by the end of 2014 in the United States [17]. Moreover,
Department of Energy SunShot Initiative aims to reduce PV
integration cost 75% by 2020 [18]. As the cost of photovoltaic
panels steadily drops and governments provide incentives,
solar generation can be used at parking lots to fuel PEVs. The
main advantage of using solar generation at Level 2 parking
lots is that customers are typically parked and stationary for
3-5 hours during their daily activities. Due to long parking
durations, this method of charging has enough flexibility to
exploit solar generation, e.g., the effect of intermittencies
introduced by clouds, dust, etc. could be alleviated. Also,
since the electricity is generated and consumed locally, with
proper control techniques it eliminates the need for transferring
electricity from the far-off generator. This way the stress
on the grid is also reduced. Furthermore, solar electricity is
typically cheaper and cleaner than utility electricity power. For
instance, according to [13], the average cost per mile for solar
powered vehicles is $0.04, whereas the average cost for utility
First Quartile
Third Quartile
200
150
100
50
0
Fig. 1: Illustration of PEV charging lot.
Mean
0:00
1:30
3:00
4:30
6:00
7:30
9:00
10:30
12:00
13:30
15:00
16:30
18:00
19:30
21:00
22:30
Solar PV Power Output (kW)
300
Fig. 2: Solar power generation at the University campus
parking deck (capacity of 300kW).
electricity powered and gas powered are $0.06 and $0.20,
respectively.
The correlation between the seasonal customer demand
and the renewable generation should be carefully addressed.
For instance, one average solar PV power output of one
generation facility at UC San Diego is shown in Fig. 2. This
PV generation has 300 kW-AC nominal capacity and is located
on one of the University parking deck. As it is clear, solar
generation has significant variation by presenting the first
and third quartiles, i.e., 25th percentile and 75th percentile.
This profile shows that stand-alone solar charging may not be
practical. To aid the uncertainty, storage units and grid power
are used.
B. Energy Storage Systems (ESS)
Energy storage units are critical components as they aid in
smoothing customer demand and the integration of renewables.
Moreover, due to stochasticities involved, the local generation
may not perfectly align with customer demand. Storage units
can act as energy buffers and decouple the time of production
and demand by storing cheaper and cleaner off-peak hour
electricity and delivering it during the peak load periods to
handle the residual demand.
Considering their high acquisition, operation, and maintenance costs, the optimal sizing of storage units is an important
design issue. Over-provisioning ESS size leads to underutilizing costly assets and under-provisioning it taxes operation
lifetime. Sizing storage units for DC fast charging stations is
addressed in [11] and [12], where authors model the charging
facility with a multi-dimensional Markov chain and solve
the storage size problem for the peak hour. Furthermore, in
[19] sizing problem is solved by Monte Carlo simulations.
Apparently, such studies could be improved by using data
of temporal customer load profile. For instance, the sizing
problem can be formulated in the following way. The station
can draw a constant power to serve some base load, and
the storage size could be determined to accommodate the
difference between the remaining customer demand and the
renewable generation. Another important aspect of storage
planning is to determine the required power rating parameter,
which depends on the statistics of customer charging rate,
energy demand, and the length of stay. For instance, a fastresponding storage unit (e.g., supercapacitors, flywheels, etc.)
may be desired to overcome the problems introduced by shortterm variations in the solar generation.
System designers should also consider the physical size of
storage units as it may grow tremendously depending on the
application. To illustrate this point, assume that a storage unit
of size 64kWh of Maxwell BMOD0083P048 modules (each
has 26.6Wh capacity) are chosen. Then the total volume of
the ESS would be 24m3 for over 2400 modules that would be
required for this application. This simple example depicts how
physical space can be problematic if the station is located in
regions with expensive real estate.
C. Capacity Planning
Capacity provisioning problem –determining how much
power to draw from the grid –is an important decision problem for sustainable charging facilities. The proper resource
provisioning depends on the customer demand profile and its
variability. For example, if the station resides over a relatively
well-confined region such as a university campus, then the
variability is probably lower than a larger region such as
stations located at shopping malls or interstate exchanges.
Hence, resource allocation problem of the first case is relatively simpler than the second one. For instance, the work
presented in [8] proposes a capacity planning framework
for a campus size network composed of different charger
technologies (Level 1, Level 2, DC fast etc.). They propose a
capacity planning framework to guarantee a certain quality
of service levels. However, since there is no measurement
study, this study assumes a sinusoidal demand function for
illustrative examples. Moreover, in most capacity planning
studies, resource provisioning is computed respect to peak
demand. On the other hand, the shape of the customer demand
and the peak-to-average demand ratio can affect important
design questions such as storage type, renewable portfolio,
etc. Hence, it is crucial to know the daily, weekly, and the
monthly demand profile.
Another key aspect of station planning is to determine
the number of physical chargers to be deployed at charging
facilities. The cost per Level 2 charger hovers between $600 to
$1000, hence over-provisioning may be a costly option. The
works presented in [4], [8], [12] and [20] employ queuing
theory to model the charging stations. In such studies, the
number of physical chargers represents “servers” and the
operation principle investigates the case when some customers
do not get the charging service immediately if all of the
chargers are occupied. Hence, one of the goals is to investigate
the relation between waiting time and the number of chargers.
Then the design question is to find the optimal number of
chargers such that customers are provided with statistical
guarantees (no customer waits more than a target waiting
time). Queuing-based studies heavily rely on the assumptions
made on customer arrivals, departures, energy demand and the
population of PEVs. In the next section, we provide detailed
results of our data set.
TABLE I: Summary of PEV station design
Parameter Challenges
Required Statistics
-Decide PV size
Solar
-Forecast generation
-Power and energy rating
ESS
-Decide physical space
-Decide # of Chargers
Capacity
-Capacity planning
-Temporal customer load
-Charging current ranges
-Customer demand profile
-PV Output
-Daily load profile
-Arrival/Service stats.
D. Overview PEV Demand Control Techniques
As the population of PEVs increase, station operators may
need to apply control techniques to manage PEV demand.
The customer demand can be shaped by optimally setting;
(1) charging start time; (2) charging service location; and
(3) charging current. An overview of techniques is presented
in [5]. The impact and applicability of the control technique
depend on customer types. For individual customers who are
interested in maximizing their benefits, incentive-based frameworks are applied to manage the spatiotemporal demand. On
the other hand, companies may have PEV fleets, who adhere to
the decisions of a central decision maker. Such entities can up
to a large extent decide charging current, service location and
start time. This type of load is also considered as managable
load. For instance, the work presented in [11] manages a
hybrid population of both customer types to balance the load
among neighboring stations. The work in [12] offers incentives
to individual customers to convince them to receive the service
in a less congested station An overview is provided in Table
I.
III. PEV C HARGING M EASUREMENT
A. Charging Station Information
The chargers are physically located in parking lots at a
major North American University campus. All of the parking
lots employ Level 2, 240 volt AC input chargers with SAE
J1772 connectors. Chargers can operate up to 30 amperes and
can deliver 6.5kW of electric power [21]. Energy and demand
metering employs ANSI 12.20 and IEC standards and the
communication network uses existing cellular networks.
In total, there are 17 Level 2 chargers deployed at 9 different
parking lots. The charging nodes are located within a few
miles away from each other, and exact driving distances are
calculated with Google Maps and presented in Table II. It is
noteworthy that in most cases the driving distance is further
than the actual physical distance. From the design point of
view, this information is valuable because neighboring charging stations may be fed by the same distributed generation
units (e.g., solar arrays, wind), share the same storage device,
or connected to the same distribution transformer. Then the
chargers of charging stations reside in small well-confined
regions can be abstracted collectively as a super-station. For
instance, this may apply to Stations 4 − 6, as the physical
distance is relatively short.
Among these 9 places, 2 of them, Stations 2 and 8 are
dedicated to University vehicles. Hence, they continue to park
TABLE II: Driving distance (in miles) between the chargers.
TABLE III: Detailed station information.
From/To St. 1 St. 2 St. 3 St. 4 St. 5 St. 6 St. 7 St. 8 St. 9
St.
St.
St.
St.
St.
St.
St.
St.
St.
1
2
3
4
5
6
7
8
9
0.4
0.6
1
1.5
1
1.6
0.9
1.3
0.4
0.3
1.4
1.3
0.7
1.3
0.2
0.2
0.6
0.3
1.3
1.5
1
1
0.5
0.7
1
1.4
1.3
0.3
0.3
2.3
1
1.4
1.5
1.3
1.5
0.3
0.5
2.5
1.1
1.6
1
0.7
1
0.3
0.5
2
0.6
1
1.6
1.3
1
2.3
2.5
2
1.4
1.5
0.9
0.2
0.5
1
1.1
0.6
1.4
0.5
1.3
0.2
0.7
1.4
1.6
1
1.5
0.5
-
even after completion of the charge service. We refer to these
vehicles as PEV fleet. The remaining 7 locations serve the
public, and we refer them as individual customers. Detailed
charging station information is presented in Table III.
B. Data Set
Our data set contains more than 6800 charge events collected for 166 weeks, starting from 14 Nov. 2011 to 8 Jan.
2015. The collected data contains the following information
[22].
• Location name and the address (street name, city, state,
zip code) of the charger. The chargers are at 9 different
locations.
• Serial number of the charger. This allows us to know the
exactly which charger is in use. As a future work, this
information can be used to conduct a study at a finer
granularity.
• Member guest is the unique identification number assigned to each customer. We identify 1144 unique visits,
however, if the customer is not a member of the charger
network, a new ID number will be assigned for each visit.
• Connection and disconnection time shows the arrival
and departure of each customer. The format is in
dd/mm/yyyy hh : mm : ss.
• Cumulative energy shows the amount of energy (in kWh)
stored at each charge service. This information is important to learn the demand statistics of vehicles and it can be
used in simulation-based studies for larger scale charging
networks.
• Fee refers the cost for the charge service in $/kWh. The
fee depends on the location and the tariffs of the utility.
For instance, the rate of Level 2 charging is $0.49 per
kWh in Southern California. However, the charging rate
in Seattle is 10 cents cheaper. Furthermore, there can be
an additional cost in downtown areas related to parking
fees.
IV. A NALYSIS & R ESULTS
We start our analysis by differentiating two different PEV
types individuals and PEV fleet. General public such as
students, faculty, employees, and visitors constitute the first
group, whereas the second group is composed of company
vehicles (shuttles, postal cars etc.). The rationale behind this
St.
St.
St.
St.
St.
St.
St.
St.
St.
1
2
3
4
5
6
7
8
9
# of
Chargers
# of Service
Completion
Dates
Type
3
1
2
1
1
1
2
2
4
1093
196
1182
838
147
121
1294
838
1111
12/19/12-1/8/15
5/18/13-1/7/15
9/14/12-1/8/15
12/2/11-1/8/15
7/4/13-1/8/15
5/21/13-1/6/15
1/3/13-1/8/15
12/2/11-1/8/15
11/15/11-1/8/15
Public
Public
Public
Fleet
Public
Public
Public
Fleet
Public
categorization is that since PEV fleets usually have certain
tasks, the demand for such vehicles can assumed to be manageable [11]. This can be done by optimally adjusting the charging
current and service duration, with respect to varying network
conditions. Thus, in most of our results are categorized as
individuals, fleets, and aggregated results.
Next, we present the distribution of energy demand of single
charging session for each customer case. The results depicted
in Fig. 3 shows that more than half of the customer demand
is within 0 − 8kWh and 90.1% of the total service request
is between 0 − 16kWh. Considering the fact that mainstream
PEV packs range between 16 − 85kWh, we conclude that the
vast majority of the customers use parking lots to extend their
driving ranges. Furthermore, almost 70% of the total energy
is transferred to individual PEVs. The findings reveal that
the system operator can manage the remaining 30% of the
demand.
Obviously, these results may not be meaningful without
knowing the temporal customer behavior. Thus, in Fig. 4 we
evaluate the 24 hours customer behavior. The results in Fig.
4(a) shows that the individual demand peak occurs between
10am to 12pm, whereas the demand peak for company fleets
is around 1pm. It is also noteworthy that the peak PEV fleets
demand may coincide with the peak demand of other electrical
loads. If this trend continues, fleet operators may need to shift
their usage to night time. From station design standpoint, this
information can be used to decide on how many physical
chargers to deploy in order to provide good Level of QoS,
as the number of vehicles increases the charging stations may
act as waiting systems, and station operators could guarantee
to finish service within a target deadline. Moreover, charging
facilities are expected to employ solar panels and energy
storage units to aid the power grid operations. As given in
Section II-A, solar generation depends on daily and seasonal
variations. It can be seen that there is a correlation between
PEV demand and solar generation (depicted in Fig.2). This
means that solar energy has a great potential to provide
electricity during such hours. Another important parameter
is the seasonal customer demand; hence, we show the PEV
demand for each month of the year in Fig.5.
Next, we evaluate the daily customer demand. The results
depicted in Fig.6 shows that the service requests during the
weekdays constitute almost a flat profile. From the capacity
Individual Customers
Managable Load: PEV Fleet
1200
35.41%
# of PEV Service
1500
1000
800
10.2%
600
400
4
12
20
2.19%
200
1.95%
0
Percentage respect to total demand
6.2%
500
0
28 36 44 52
Energy Demand (kWh)
(a) Energy demand of individual customers.
53.54%
3000
1000
# of PEV Service
Percentage respect to total demand
26.36%
2000
All Customers: Individual and PEV Fleet
3500
18.13%
# of PEV Service
2500
4
12
2500
36.56%
2000
1500
1000
8.39%
500
20
28
36
44
Energy Demand (kWh)
1.05%
0
52
4
(b) Energy demand of PEV fleet.
12
20
0.46%
28 36 44 52
Energy Demand (kWh)
(c) Aggregated energy demand.
Fig. 3: Energy demand distribution of single charging session. The average energy transfer is 8.53kWh and the standard
deviation is 6.49 kWh
# of Customer Arrivals
600
300
200
400
200
200
100
0
0am
1am
2am
3am
4am
5am
6am
7am
8am
9am
10am
11am
12pm
1pm
2pm
3pm
4pm
5pm
6pm
7pm
8pm
9pm
10pm
11pm
100
All Customers: Individual and PEV Fleet
800
# of Customer Arrivals
300
0
Managable Load: PEV Fleet
400
# of Customer Arrivals
400
500
0
0am
1am
2am
3am
4am
5am
6am
7am
8am
9am
10am
11am
12pm
1pm
2pm
3pm
4pm
5pm
6pm
7pm
8pm
9pm
10pm
11pm
Individual Customers
0am
1am
2am
3am
4am
5am
6am
7am
8am
9am
10am
11am
12pm
1pm
2pm
3pm
4pm
5pm
6pm
7pm
8pm
9pm
10pm
11pm
500
Hour of Day
Hour of Day
(a) Hourly individual customer demand.
Hour of Day
(b) Hourly company PEV fleet demand.
(c) Hourly aggregated demand.
Fig. 4: PEV hourly demand
600
Individual Customers
200
Managable Load: PEV Fleet
800
All Customers: Individual and PEV Fleet
400
300
200
150
# of PEV Charging
# of PEV Charging
# of PEV Charging
500
100
50
600
400
200
100
0
0
0
t t
l
Jan Feb Mar Apr May Jun Ju Aug Sep Oc Nov Dec
(a) Monthly individual customer demand.
t t
l
Jan Feb Mar Apr May Jun Ju Aug Sep Oc Nov Dec
(b) Monthly company PEV fleet demand.
t t
l
Jan Feb Mar Apr May Jun Ju Aug Sep Oc Nov Dec
(c) Monthly aggregated demand.
Fig. 5: PEV Monthly demand
planning standpoint, this is a desired behavior because less uncertainty in customer demand will translate to higher resource
utilization, i.e., storage and solar usage. It is also noteworthy
that the customer demand during weekends is considerable
less than the weekdays. Station operator can store and sell the
energy back to the power grid. Also, some portion of the PEV
fleet load on Mondays could be served on Sundays to allow
more individual customers.
2013 and the results depicted in Fig. 7 shows that as the
PEV adoption increases the need for charging service surges
accordingly. In fact, 4 out of 9 charging locations are opened
after year 2013. This data is extremely important to forecast
the customer demand and determine the required number of
charging nodes.
Our final evaluation is on customer demand over time. We
compute the number of PEV services from Nov. 2011 to Nov.
In this paper, we have analyzed and presented the data
obtained from a Level 2 charging network located in major
V. C ONCLUSIONS & F UTURE W ORK
600
400
200
400
300
200
100
1000
500
Su
nd
ay
M
on
da
y
Tu
es
da
W
y
ed
ne
sd
ay
Th
ur
sd
ay
Fr
id
ay
Sa
tu
rd
ay
0
Su
nd
ay
M
on
da
y
Tu
es
da
W
y
ed
ne
sd
ay
Th
ur
sd
ay
Fr
id
ay
Sa
tu
rd
ay
0
All Customers: Individual and PEV Fleet
Su
nd
ay
M
on
da
y
Tu
es
da
W
y
ed
ne
sd
ay
Th
ur
sd
ay
Fr
id
ay
Sa
tu
rd
ay
0
1500
# of Customer Arrivals
500
# of Customer Arrivals
# of Customer Arrivals
Managable Load: PEV Fleet
Individual Customers
800
(a) Daily individual customer demand.
(b) Daily company PEV fleet demand.
(c) Daily aggregated demand.
Day of Week
Day of Week
Day of Week
Fig. 6: PEV daily demand
140
[8] I. Bayram, A. Tajer, M. Abdallah, and K. Qaraqe, “Capacity planning
frameworks for electric vehicle charging stations with multiclass customers,” IEEE Trans on Smart Grid, vol. PP, no. 99, pp. 1–1, 2015.
[9] C. Hutson, G. K. Venayagamoorthy, and K. A. Corzine, “Intelligent
100
y = 6e−05*x3 − 0.0053*x2 + 0.33*x + 8.4
scheduling of hybrid and electric vehicle storage capacity in a parking
80
lot for profit maximization in grid power transactions,” in Proc. IEEE
Holiday: last
week of the year
Energy 2030 Conference, 2008, pp. 1–8.
60
[10] W. Su and M.-Y. Chow, “Performance evaluation of an eda-based largescale plug-in hybrid electric vehicle charging algorithm,” IEEE Trans.
40
on Smart Grid, vol. 3, no. 1, pp. 308–315, March 2012.
20
[11] I. Bayram, G. Michailidis, M. Devetsikiotis, and F. Granelli, “Electric
power allocation in a network of fast charging stations,” IEEE Journal
0
0
20
40
60
80
100
120
140
on Selected Areas in Communications, vol. 31, no. 7, pp. 1235–1246,
Week # from Nov., 2011 to Nov., 2014
July 2013.
Fig. 7: Weekly charging data. The number of service request [12] I. Bayram, G. Michailidis, and M. Devetsikiotis, “Unsplittable load
balancing in a network of charging stations under qos guarantees,” IEEE
increases as the rise in PEV sale.
Trans. on Smart Grid, vol. 6, no. 3, pp. 1292–1302, May 2015.
[13] Sullivan Solar Power Inc. [Accessed: April, 2015]. [Online]. Available:
http://www.sullivansolarpower.com/solar-power/electric-vehicles
North American University campus. As the future sustainable
charging facilities are expected to employ renewable genera- [14] Revision Energy. [Accessed: April, 2015]. [Online]. Available:
http://www.revisionenergy.com/solar-electric-vehicle-charging-maine.php
tion and storage units, along with smart energy management [15] J. Quiros-Tortos, L. Ochoa, and B. Lees, “A statistical analysis of ev
charging behavior in the uk,” in Innovative Smart Grid Technologies
systems, we have provided detailed guidelines for charging
Latin America (ISGT LATAM), 2015 IEEE PES, Oct 2015, pp. 445–
station design. As a future work, we are aiming to investigate
449.
the potential usage of solar powered stations using the obtained [16] P. Richardson, M. Moran, J. Taylor, A. Maitra, and A. Keane, “Impact
of electric vehicle charging on residential distribution networks: An irish
data.
R EFERENCES
demonstration initiative,” in Electricity Distribution (CIRED 2013), 22nd
International Conference and Exhibition on, June 2013, pp. 1–4.
[1] Light duty electric drive vehicles monthly sales updates. Argonne
[17] P. Denholm, R. Margolis, B. Palmintier, C. Barrows, E. Ibanez, L. Bird,
National Laboratory. [Accessed: April, 2015]. [Online]. Available:
and J. Zuboy, Methods for Analyzing the Benefits and Costs of Dishttp://www.anl.gov/energy-systems/project/light-duty-electric-drive-vehicles-monthly-sales-updates
trubuted Photovoltaic Generation to the US Electric Utility System.
[2] Improving
electric
vehicle
sales
may
require
solving
National Renewable Energy Laboratory, 2014.
unique
chicken
and
egg
problem.
National
Science
[18] D. Feldman, G. Barbose, R. Margolis, N. Darghouth, T. James,
Foundation.
[Accessed:
April, 2015]. [Online].
Available:
S. Weaver, A. Goodrich, and R. Wiser, “Photovoltaic system prichttp://www.nsf.gov/discoveries/disc summ.jsp?cntn id=133947
ing trends: Historical, recent, and near-term projections–2013 edition,”
[3] Plug-in Electric Vehicle Collaborative. [Accessed: April, 2015].
Golden, CO: National Renewable Energy Laboratory, PR-6A20-60207.
[Online]. Available: http://www.pevcollaborative.org/
Accessed December, vol. 9, p. 2013, 2013.
[4] I. Bayram, G. Michailidis, and M. Devetsikiotis, “Electric power re[19] S. Bai, D. Yu, and S. Lukic, “Optimum design of an ev/phev charging
source provisioning for large scale public EV charging facilities,” in
station with DC bus and storage system,” in Proc. IEEE Energy
Proc. IEEE International Conference on Smart Grid Communications,
Conversion Congress and Exposition, 2010, pp. 1178–1184.
Vancouver, CA, Oct. 2013, pp. 133–138.
[20] P. Fan, B. Sainbayar, and S. Ren, “Operation analysis of fast charging
[5] I. S. Bayram and I. Papapanagiotou, “A survey on
stations with energy demand control of electric vehicles,” IEEE Trans.
communication technologies and requirements for internet of
on Smart Grid, vol. PP, no. 99, pp. 1–1, 2015.
electric vehicles,” EURASIP Journal on Wireless Communications
[21] M. Falvo, D. Sbordone, I. Bayram, and M. Devetsikiotis, “Ev charging
and Networking, vol. 2014, no. 1, 2014. [Online]. Available:
stations and modes: International standards,” in Power Electronics, Elechttp://dx.doi.org/10.1186/1687-1499-2014-223
trical Drives, Automation and Motion (SPEEDAM), 2014 International
[6] K. Clement-Nyns, E. Haesen, and J. Driesen, “The impact of charging
Symposium on, June 2014, pp. 1134–1139.
plug-in hybrid electric vehicles on a residential distribution grid,” IEEE [22] ETEC, “Electric vehicle charging infrastructure deployment guidelines
Trans. on Power Systems, vol. 25, no. 1, pp. 371–380, 2010.
for the greater san diego area,” Electric Transportation Engineering
[7] B. Sanzhong, D. Yu, and S. Lukic, “Optimum design of an ev/phev
Corporation, Tech. Rep., May 201o.
charging station with dc bus and storage system,” in Proc. IEEE Energy
Conversion Congress and Exposition, Atlanta, GA, Sept. 2010, pp.
1178–1184.
PEV Charging per Week
120
Weekly EV Service
cubic fitting
| 3cs.SY
|
Closed loop interactions between spiking neural network and robotic
arXiv:1604.04764v1 [cs.NE] 16 Apr 2016
simulators based on MUSIC and ROS
Philipp Weidel1∗ , Mikael Djurfeldt2,1 , Renato Duarte1,3,4 , Abigail Morrison1,4,5,6
1
Institute for Advanced Simulation (IAS-6),
Theoretical Neuroscience &
Institute of Neuroscience and Medicine (INM-6),
Computational and Systems Neuroscience &
JARA BRAIN Institute I
Jülich Research Center and JARA
Jülich, Germany
2
PDC Center for High Performance Computing
KTH
Stockholm, Sweden
3
Faculty of Biology
Albert-Ludwig University of Freiburg
Freiburg im Breisgau, Germany
4
Bernstein Center Freiburg
Albert-Ludwig University of Freiburg
Freiburg im Breisgau, Germany
1
5
Simulation Laboratory Neuroscience –
Bernstein Facility for Simulation and Database Technology
Institute for Advanced Simulation
Jülich Aachen Research Alliance
Jülich Research Center
Jülich, Germany
6
Institute of Cognitive Neuroscience
Faculty of Psychology
Ruhr-University Bochum
Bochum, Germany
∗
Forschungszentrum Jülich GmbH
Institute of Neuroscience and Medicine (INM-6)
52425 Jülich, Germany
e-mail: [email protected]
Abstract
In order to properly assess the function and computational properties of simulated neural systems, it is necessary
to account for the nature of the stimuli that drive the system. However, providing stimuli that are rich and yet both
reproducible and amenable to experimental manipulations is technically challenging, and even more so if a closedloop scenario is required. In this work, we present a novel approach to solve this problem, connecting robotics and
neural network simulators. We implement a middleware solution that bridges the Robotic Operating System (ROS)
to the Multi-Simulator Coordinator (MUSIC). This enables any robotic and neural simulators that implement the
corresponding interfaces to be efficiently coupled, allowing real-time performance for a wide range of configurations.
This work extends the toolset available for researchers in both neurorobotics and computational neuroscience, and
creates the opportunity to perform closed-loop experiments of arbitrary complexity to address questions in multiple
areas, including embodiment, agency, and reinforcement learning.
1
Introduction
Studying a functional, biologically plausible neural network that performs a particular task is highly relevant for progress in
both neuroscience and robotics. The major focus on this topic in the field of robotics consists of using of neural networks
of varying degrees of complexity for controlling robots. So far, the majority of research has focused on non-spiking,
artificial neural networks for this task (Quiñonez et al., 2015; Antonelo et al., 2007; Dasgupta et al., 2013), but there is
considerable interest in investigating the capacities of spiking neural networks.
In computational neuroscience, the study of simulated neural networks is paramount to gain a better understanding of
the processes underlying learning/adaptation to complex environments and global behavior (e.g. Chorley & Seth, 2008).
However, neural network simulators do not typically include functionality for representing environments and sensory input.
As a consequence, most tasks used to test the function of a simulated neural network are hardcoded to represent a highly
specific task. This has the disadvantage that virtual experiments are complex and time consuming to develop and adapt.
More importantly, tasks defined in this way are rather artificial (Jitsev et al., 2012; Frémaux et al., 2013; Potjans et al.,
2011; Legenstein & Maass, 2014; Friedrich & Lengyel, 2015). Whereas there is certainly value in investigating very
simplified tasks and sensory representations, it is also vital to be able to check that proposed neural architectures are
capable of handling richer, noisier and more complex scenarios.
However, design of environments, representation of sensory states and conversion of motor commands into movements
are primary features of robotic simulators. Therefore, it would be of great value to both research fields if the powerful
tools developed for robotic simulation and spiking neural network simulation could be made to work together. This
would allow researchers from both fields to perform simulated closed loop experiments with flexible experiment design,
rich sensory input, complex neuronal processing, and motor output. The major challenge lies in the fact that robotic
simulators communicate via continuous data streams, while neural network simulators communicate with spike events
(so as their biological counterparts). Thus, a principled approach for bi-directional conversion is required.
So far there have been few attempts to address this. Moren et al. (2015) describe a technical setup for a specific
use case, in which a robot is connected to a neural network model of early saccade movement simulated with the Neural
Simulation Tool (NEST: Gewaltig & Diesmann, 2007). As the robot is not in the same location as the cluster where the
neural simulation is running, the connection is realized via the internet using an SSH tunnel and the Multi-Simulation
Coordinator (MUSIC: Djurfeldt et al., 2010), a middleware facilitating communication between different neural simulators
in one simulation. Using a minimal model and optimal number of cores, the NEST simulation ran a factor of two slower
than real time and achieved an output response of 110 − 140 ms. Although this study provides a proof of concept for
interaction between robotics and neural simulators, it does not represent a general solution, in part due to the limited
3
implementation detail provided.
A more general solution was presented by Hinkel et al. (2015), who describe an interface between the Robotic
Operating System (ROS: Quigley et al., 2009) and NEST. ROS is the most popular middleware for the robotic community,
with interfaces to many robotic simulators and also robotic hardware. It allows the user to create their own simulated
robots in great detail using the Unified Robot Description Format. Similarly, NEST is one of the most commonly used
neural simulators for spiking point neuron models. However, the proposed interface is on the basis of Python: the neural
network must be simulated in brief periods in an external loop, reading out and communicating the spikes to the robotic
simulator at each iteration. Although no performance data are provided, the overheads inherent in repeatedly stopping
and starting NEST imply severe performance limitations for this approach.
In this paper we present an alternative approach based on ROS and MUSIC that is both general and efficient, enabling
real-time performance for up to hundreds of thousands of neurons on our hardware. Our approach provides roboticists the
opportunity to extend their work on neural control into the realm of spiking neural networks using any neural simulator
implementing a MUSIC interface, including NEST or NEURON (Hines & Carnevale, 2001). Conversely, our toolchain
frees the researcher from the constrictions of the hardcoding approach described above, by enabling neural simulators
to be connected to any of the myriad robotic simulators implementing a ROS interface (including Gazebo1 , Morse2 or
Webots3 ). To demonstrate the capabilities of the toolchain, we implement a Braitenberg Vehicle (Braitenberg, 1986) in
which the agent is simulated in Gazebo and the neural controller in NEST.
1.1
Description of the toolchain
The main purpose of the present toolchain is to capture a broad set of use cases by connecting robotic with neural
simulators in a generic way, obtained by interfacing the two middlewares (ROS and MUSIC). The interface is licensed
under GPLv3 and available on Github4 . In addition to our own repository, we plan to distribute the toolchain as a plug-in
for MUSIC5 .
In this work, we propose a set of possible solutions to overcome the problem of converting between continuous
and spiking signals. In particular, we investigate the performance of three different kinds of encoders (NEF, regular
rate coding, Poisson rate coding) and a linear readout decoder. As the correct encoding mechanism is debatable and
dependent on the scientific question being addressed, in addition to our own mechanisms, we provide extensibility such
that researchers can implement their own custom encoders, decoders and adapters in Python or C++.
1 http://gazebosim.org/
2 https://www.openrobots.org/wiki/morse/
3 https://www.cyberbotics.com
4 https://github.com/weidel-p/ros_music_adapter
5 https://github.com/INCF/MUSIC
4
Gazebo
WeBots
Hardware
Publisher
sensor
Adapter
Subscriber
sensor
motor
Publisher
MUSIC Event
In Proxy
Neural
Simulator
MUSIC
ROS
Subscriber
Encoder
Decoder
motor
MUSIC Event
Out Proxy
Figure 1: Information flow in the ROS-MUSIC toolchain. The boxes with red outlines depict our extensions to MUSIC
that facilitate the bidirectional conversion of continuous and spiking signals (solid and dashed arrows, respectively)
To achieve these ends, our interface to ROS extends MUSIC by three different kind of binaries: adapters, encoders
and decoders (see Figure 1). Encoders and decoders are used to translate between spiking data and continuous data
while adapters can be used for pre-processing data and connect to ROS. All binaries run in their own process and solve
only one specific purpose. This way, the interface is highly modular and the implementation of custom adapters, encoders
or decoders is as simple as possible.
2
Materials and Methods
2.1
The MUSIC configuration file
The adapters, encoders, decoders and simulators introduced in the toolchain are specified, connected and parameterized
in the MUSIC configuration file. Detailed information about the MUSIC configuration file can be found in the MUSIC
manual6 and examples of complete configuration files for this toolchain can be found in the Github repository.
2.2
Encoding, decoding and adaptation
The choice of encoder and decoder to convert continuous to spiking signals is non-trivial, and potentially dependent on
the scientific question or task. Here, we describe three possible encoding mechanisms and one decoding mechanism.
We also describe a simple adapter to map the dimenstionality of the input signal to the dimensionality of the receiving
network.
6 http://software.incf.org/software/music
5
2.2.1
Regular rate coding
The simplest way to encode continuous data into spiking data is regular rate coding. i. e. mapping the input signal
to a time-dependent firing rate variable. Each input dimension is encoded by a single neuron by transforming the input
signal, which is a continuous signal in the range [−1, 1], into a time-dependent variable representing the spike density of
the encoding neuron. This is achieved by calculating the interspike interval of each encoding neuron at time t by:
ISIn (t) =
1
vmin + (vmax − vmin ) 1+I2n (t)
(1)
where In (t) is the input which is mapped to neuron n. The firing rate is scaled between [vmin , vmax ] which are free
parameters of the regular rate encoder. This encoding mechanism leads to very regular spike patterns but has the clear
advantage of being computationally very efficient. Note that this approach does not take into consideration biological
neuron properties such as refractoriness, but these could be included without much impact on the computational efficacy.
2.2.2
Poisson rate coding
In this encoding scheme, we follow a similar implementation to that described in the previous section, but introduce
a stochastic component in the generated spike trains, by transforming the input signal into a time-dependent variable
representing the intensity of a stochastic point process with Poissonian statistics. This can be achieved by using the inverse
of the expression (Equation 1) for interspike interval in the previous section as the rate parameter for an exponential ISI
distribution:
ISIn (t) = −
1
vmin + (vmax − vmin ) 1+I2n (t)
ln(r)
(2)
where r is a random number drawn from a uniform distribution in [0, 1) and In (t) is the input mapped to neuron n.
The firing rate is scaled between [vmin , vmax ] which are free parameters of the Poisson rate-encoder. This encoding
mechanism produces spike trains with Poissonian statistics, similar to those typically assumed to occur in central neurons
in the mammalian brain (see, for example, Calvin & Stevens (1967); Dayan & Abbott (2001) or Averbeck (2009) for a
counter-argument), and therefore are thought to be a more realistic approximation than regular spike trains. As Section
3 shows, the computational complexity of this approach is comparable to that of regular rate encoding.
6
2.2.3
Neural Engineering Framework
The Neural Engineering Framework (NEF: Eliasmith & Anderson, 2004) is a generic formalism to represent stimuli
with neural ensembles that allows a wide variety of functions to be realised by deriving optimal projections between
neural populations. In the simplest formulation, NEF encodes a vector of continuous data I as an input current to an
integrate-and-fire neuron whose activity is defined by
an (I) = Gn [αn · I + Inbias ]
(3)
where G represents a generic non-linear functional which is determined by the neuron dynamics, αn is the neuron’s
response preference, or tuning curve, and Inbias a constant bias current setting the base activation of neuron n. It is
worth noting that this encoding does not depend on the specific form of G and any neuron model can be used for this
operation. The original signal I can be adequately reconstructed (depending on the properties of G and alpha) by linearly
combining the activity a of the encoded representations
Iˆ =
X
an (I)φn
(4)
n
where the decoding weights φ are obtained by linear regression in order to minimize the reconstruction error.
2.2.4
Linear decoder
Having described possible options explored in this work to convert continuous signals into discrete spike event trains, we
now explore the reverse operation. To transform the spiking activity of a given neural ensemble to a continuous signal,
which can be used, for example, to provide motor commands to the simulated robot, we first perform a low-pass filter of
each spike train by convolving with a causal (i.e. only defined for t ≥ 0) exponential kernel k(t) = exp (−t/τdec ) with
time constant τdec
an (t) =
X
k(t − ti,n )
(5)
i
where ti,n represents the i-th spike time of neuron n. The resulting activities at time t can be linearly combined to obtain
a continuous output signal
zk =
X
an (t)φnk
n
7
(6)
where weights φnk define the contribution of ensemble neuron n to readout unit k.
2.2.5
Signal adapter
For the experiments carried out below we implement a simple adapter that maps the m dimensions of the input signal
to the n dimensions of the receiving neurons. Note that this is only necessary for regular and Poisson rate encoding; the
NEF encoder already incorporates this functionality by construction. For the performance measurements, each receiving
neuron receives all dimensions of the input signal. For the Braitenberg vehicle, the input signal is split up into two
hemispheres, each mapped to one of the two controlling neurons. For more complex experiments, a more sophisticated
adapter can be implemented.
2.3
Performance Measurements
In order to be as close as possible to a real use-case, we simulate the whole toolchain while measuring the performance
capabilities of the different parts. The agent used for the performance measurements is a four wheeled, mobile robot
simulated in Gazebo with an attached virtual laser scanner. The laser scanner has 100 beams with an update rate of
20 Hz and a maximal range of 5 m. We use the SkidSteerDrivePlugin provided by Gazebo, which allows us to steer
the robot with a ROS::Twist message, updated with a rate of 20 Hz. During the measurements involving the NEF
encoder, we keep the parameters of the integrate-and-fire (IAF) neurons unchanged and simulate these neurons with a
resolution of 1 ms using exact integration (Rotter & Diesmann, 1999; Morrison et al., 2007; Hanuschkin et al., 2010). For
decoding, we always use a linear readout where the spiking activity is filtered with an exponential kernel with a resolution
of 1 ms. The MUSIC adapters run with an update rate equal to the update rate of the sensory input and motor command
output. The measurements were executed on one node of a transtec CALLEO 551 cluster and were averaged over five
trials with a simulation-time of ten seconds. Measurements involving neural simulators were carried out with NEST 2.8
(Eppler et al., 2015) and NEURON v6.2.
2.3.1
Real-time factor
The sensory input, the motor output and the processing of the sensory data are all updated in parallel but asynchronously.
For robotic applications it is crucial that the execution time for processing input data in each time step is less than or
equal to the wall-clock time of each time step, meaning the process is running in real-time.
The real-time factor (RT F ) is calculated by dividing the simulated time, tsim , by the wallclock time required to run
the simulation, trun .
8
RT F =
tsim
trun
(7)
The total time the toolchain needs to perform a simulation can be divided into different components
ttotal = tbuild + trun +
(8)
where tbuild is the time the toolchain needs for initialization (allocate memory etc), trun is the actual time the toolchain
needs to perform the simulation and is the time which is not covered by tbuild and trun (i.e. garbage collection, freeing
memory, etc.). In our measurements we clearly separate the buildup phase from the run-time phase by synchronizing the
processes after the initialization of the software using an MPI barrier. This allows us to measure trun consistently over
all processes.
We use this measurement to investigate different aspects of the toolchain’s performance, as described below.
Dimensionality of the input
The sensory input to a robot can be very complex and high dimensional. The number
of neurons needed to encode a multidimensional input strongly depends on the encoding mechanism. In NEF, about
100 IAF neurons per dimension are used when encoding a stimulus in order to be able to decode that stimulus with
a root-mean-square error about 1% (Eliasmith & Anderson, 2004). Using rate coding, the number of neurons can be
freely chosen. In the example of the Braitenberg Vehicle (Braitenberg, 1986), we use only two neurons for encoding the
complete sensory input (see Section 3.1).
To examine the scalability of the sensory input in our toolchain, we measure how many encoding neurons for different
encoders we can simulate until the real-time performance breaks down. We use a binary search to find these limits in
an efficient way. When measuring the performance with NEST and NEURON, each neuron instantiated by the encoder
is matched by a corresponding neuron in the neuronal simulator, which receives the input spike train and repeats it to
the decoder as an identical spike train, thus implementing a minimal processing network. In NEST this is realized by
the parrot_neuron, a neuron model that emits a spike for each spike received. In NEURON the repeater neurons are
of type IntFire1 equipped with very high input synaptic weights and zero refractory time, thus enforcing the neuron to
spike after each input spike. For the rate and Poisson encoder, the encoder neurons have a low firing rate between 1 and
2 Hz. Except for NEST (7 processes) and NEURON (40 processes), each process in the toolchain runs on one dedicated
process.
9
Bandwidth Apart from the computational limitations of simulating neurons, communication is another potential bottleneck in this toolchain. Parts of the toolchain are communicating action potentials (events) between different processes
(i.e. from encoder to NEST). In order to investigate the influence of the communication on the real-time performance,
we use a rate based encoder and measure the real-time factor as a function of the firing rate of the encoding neurons.
We choose the amount of neurons close to the border of real-time capability, so that we can see an immediate impact of
the firing-rate on the performance of the toolchain.
Latency The latency between sensory input and motor output can be crucial for the robotic application. A fast reaction
time is needed for many applications (for example, catching a ball). In order to measure the latency (or reaction time),
we artificially disturb the sensory signal of the robotic simulator, switching discontinuously from the minimum to the
maximum sensor range. A very simple encoder responds to this change by beginning to produce spikes, which the neural
simulator receives over its MUSIC Event In Proxy and repeats its the MUSIC Event Out Proxy (see Figure 1). A decoder
responds to the arrival of this spike train by producing a motor command, which is conveyed over ROS to the robotic
simulator. The latency of the tool chain is therefore the wall-clock time between the change in the sensory signal and
the reception of the motor signal.
Overhead For measuring the input scalability as described above, we use a MUSIC time step equal to the ROS update
rate. This minimizes the communication overhead as only new data is communicated. As latency depends on the MUSIC
time step, we additionally investigate the communication overhead of the toolchain. To do this, we determine the realtime factor of the toolchain, using a simple rate encoding mechanism, whilst systematically varying the MUSIC time step
and the number of simulated neurons.
3
Results
Figure 2 shows the effect of the computational load of the different encoding mechanisms on the real-time factor, with
and without neural simulators involved in the toolchain. For the rate or Poissonian encoding mechanisms (Figure 2a,b),
simulations of up to 150, 000 encoding neurons are possible in real-time when using NEST or when the toolchain does
not include a neural simulator. When using NEURON, the real-time factor breaks down at about 1000 neurons. For the
NEF encoding mechanism (Figure 2c), simulations of up to 20, 000 encoding neurons are possible in real-time when using
NEST or no neural simulator in the toolchain. In this case too, the performance when using NEURON breaks down at
about 1000 neurons.
10
a
b
c
realtime factor
1.0
0.9
0.8
0.7
0.6
102 103 104 105 106 102 103 104 105 106 102 103 104 105 106
# neurons
# neurons
# neurons
Figure 2: Scalability of the toolchain. Real-time factor as a function of the number of neurons used for encoding using
different encoders: a) rate encoding mechanism; b) Poissonian encoding mechanism; c) NEF encoding mechanism. In
each panel, the black curve shows the real-time performance in the absence of a neuronal simulator, the red curve shows
the performance using NEST and the blue curve the performance using NEURON.
11
b
1.0
0.30
0.25
0.9
latency (s)
realtime factor
a
0.20
0.8
0.10
0.7
0.6
0.15
0.05
20
40
60
80
firing rate (spikes/sec)
100
0.00
0.01
0.02
0.03
0.04
MUSIC time step (s)
0.05
Figure 3: Bandwidth limitations of the toolchain. a) Real-time factor as a function of neuronal firing rate per neuron
for a regular rate encoder of 50, 000 neurons without a neural simulator (black curve) and with NEST (red curve). b)
Latency as a function of MUSIC time step without a neural simulator (black curve), with NEST (red curve) and with
NEURON (blue curve).
These results demonstrate that the NEF encoder is computationally more expensive than either the rate or Poissonian
encoders, and that NEURON is computationally expensive in this context, limiting the real-time performance to 1000
neurons regardless of which encoding mechanism is used.
The breakdown of the real-time performance of the regular rate and Poissonian encoders happens at about the
same amount of neurons, which leads to the question what the actual bottleneck of the toolchain is in these cases.
Computational complexity is only one limiting factor in this toolchain; another candidate is the communication.
But as Figure 3a shows, the real-time factor breaks down at about 40 Hz using 50, 000 encoding neurons. In other
words, the toolchain is able to communicate about 2, 000, 000 spikes per second, which is more than was required for
the simulations measured in Figure 2. Thus we can conclude firstly that communication was not the bottleneck for the
previous experiment, and secondly that not just the number of the encoding neurons but also their firing rate can be a
limiting factor for the real-time performance.
Figure 3b shows the latency of the toolchain, i.e. the difference in time between a change in sensory input and
12
MUSIC time step (ms)
5000
30000 55000 80000 105000 130000 155000 180000
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
realtime factor
1
5
10
15
20
25
30
35
40
45
50
#neurons
Figure 4: Real-time factor as a function of the MUSIC time step and number of neurons.
receiving a motor command evoked by that change (see Section 2.3.1). The latency increases linearly with the MUSIC
time step; a MUSIC time step of 1 ms, results in a latency of about 70 ms, growing to around 350 ms for a time step of
50 ms when using a neural simulator. In general, the presence of a neural simulator in the toolchain increases the latency
slightly, independently of which simulator is chosen.
Figure 4 depicts the dependency of real-time capability of the toolchain on the MUSIC time step. The border of
real-time simulation capability increases linearly with the MUSIC time step, from 10, 000 neurons for a time step of 1 ms
to 185, 000 neurons for a 50 ms time step. However, as demonstrated in Figure 3b, the latency also increases with the
MUSIC time step. From these results, we conclude that the latency and the dimensionality of the input are two conflicting
properties, which have to be balanced for the specific use case.
13
3.1
Implementation of a Braitenberg Vehicle
Figure 5: A Braitenberg Vehicle simulated in Gazebo, controlled by simulated neurons in NEST with the use of the
ROS-MUSIC toolchain. The inset shows the spiking activity of the two controlling neurons.
As an example for the usage of the toolchain, we created a Braitenberg Vehicle III “Explorer” (Braitenberg, 1986) which
is simulated in the robotic environment Gazebo (see Figure 5). The Braitenberg Vehicle is implemented as a fourwheeled mobile robot with an attached laser scanner for sensory input. With the use of the ROS-MUSIC toolchain, two
neurons, simulated in NEST, control this vehicle to avoid obstacles. A video of the demonstration can be found in the
supplementary material and the source code is available on GitHub7 .
4
Discussion
We have described our plug-in for MUSIC, which allows any neuronal network simulator implementing a MUSIC interface
to communicate with any robotics simulators implementing a ROS interface. The plug-in converts continuous signals
7 https://github.com/weidel-p/ros_music_adapter
14
from a robotics simulator into spike train signals for a neuronal network simulator and vice versa. We showed that the
toolchain allows real-time performance for a wide range of configurations and provided a simple working example. In the
following, we discuss limits and perspectives for this approach.
Performance
Dependent on the choice of encoder, our toolchain allows the simulation of around 20, 000 neurons with a NEF encoder
and up to 150, 000 neurons with a regular rate or Poisson encoder and achieve real-time performance. Using the NEF
encoder, it is recommended to have at least 100 neurons per input dimension in order to encode and decode stimuli
with root-mean-square errors of less than 1% (Eliasmith & Anderson, 2004), meaning that our toolchain can encode a
200-dimensional input with NEF in real time on a single process of our hardware.
If a higher dimensional input is required, there are two solutions at hand. First, the NEF encoder could be parallelized
and run in more than one process. Second, if a less sophisticated but computationally cheaper way of encoding is
acceptable for the scientific question at hand, we also provide regular rate and Poissonian encoding mechanisms. A
previous neurorobotic interface implemented the communication between ROS and the neural simulators on the Python
level (Hinkel et al., 2015). Although no performance data were published in that study, it is clear that the performance
would be strongly limited by this approach.
Another critical characteristic of the toolchain is the latency between sensory input and motor command output of the
toolchain, or in other words the reaction time. To minimize the overhead and avoid repeated communication of the same
data, it makes sense to set the MUSIC time step equal to the sensory update or motor command rate. However, Figure
3 demonstrates that this leads to a rather high latency. The architecture of communicating processes currently carries
an unavoidable source of latency, since data is buffered in MUSIC for every pair of communicating processes. In any
chain of communicating processes, the latency due to buffering will thus grow in proportion to the length of the chain.
In future work, this can be tackled by combining multiple adapter, encoder and decoder steps in a single process, thus
simultaneously minimizing communication and reaching a lower latency, and/or by introducing non-buffered inter-process
communication. Combining adapters, encoders and decoders can be realised using a plug-in adapter API which maintains
independence between the software components.
Complexity of neuronal simulations
In our measurements of the toolchain incorporating NEST and NEURON, we used only minimal networks solving no particular task. However, neural network models can rapidly become computationally complex, especially when incorporating
15
synaptic plasticity, large network sizes, or multi-compartment neuron models. Such models would be impossible to run in
real-time using the current approach. In some cases this issue could be solved by increasing the computational resources,
using a cluster or supercomputer (Helias et al., 2012; Kunkel et al., 2014). However, this does not guarantee real-time
execution, as neuronal network simulators do not have perfect scaling, due to serial components in their algorithms and
communication overhead. Even if the problem can theoretically be addressed in real-time by increasing resources, it
might not be feasible to access the quantity required. An alternative prospect is offered by neuromorphic hardware, i.e.
hardware purpose built for simulation/emulation of neuronal networks. Two examples that are being developed within
the framework of the Human Brain Project are SpiNNaker (Furber et al., 2013) and NM-PM1 (Schemmel et al., 2010),
which have the potential to speed up the neuronal simulation massively. In particular SpiNNaker has great potential
for neurorobotic applications and an interface between SpiNNaker and MUSIC is already in the prototype phase. This
development enhances the value of the toolchain we describe for the neurorobotic community.
Whereas the real-time property is important for robotic control, it is not nearly so important for addressing questions
in the field of computational neuroscience. Here, the advantage of the toolchain is that a neuronal network simulation
can be provided with rich sensory input from an agent interacting with an environment that is easy for the researcher to
configure. In this case, arbitrarily computationally demanding networks can be coupled with robotics simulators simply
by slowing down the latter to compensate - gazebo, for example, provides a parameter to conveniently control the time
scaling.
Applications
The toolchain we describe gives researchers in computational neuroscience the possibility to test their hypothesis and
models under more realistic conditions of noisy sensory input, and researchers in neurorobotics the opportunity to investigate more realistic neurally based controllers. One area of potential interest is the ability to construct interactions
between robotic simulators and hybrid neuronal simulations on multiple scales, e.g. a network of point neuron models
into which detailed biophysical models are embedded, simulated by NEST and NEURON respectively. This demonstrates
a further advantage of the MUSIC-based interaction over pairings of particular simulators or the Python-based interaction
presented by Hinkel et al. (2015).
Moreover, our toolchain is particularly well suited for studying closed-loop scenarios, where the neural network receives
stimuli from a complex environment and produces an output, which in turn causes the robotic agent to perform actions
within that environment. For example, a robotic agent can be placed in a classic experimental set-up like a T-maze
and the behaviour of the robot adapted by a neurally implemented reinforcement learner (Jitsev et al., 2012; Potjans
et al., 2011; Frémaux et al., 2013; Friedrich & Lengyel, 2015). Here, there is a clear advantage over studying such
16
questions just using neural simulators, as the representation of an external environment as a collection of neural recorders
and stimulators is complex, and difficult to either generalize or customize. By separating the concerns of environmental,
motor, and sensory representation from those of neural processing, our toolchain provides a highly flexible and performant
research approach.
Acknowledgments
We thank Karolína Korvasová, Tom Tetzlaff and Jenia Jitsev for helpful discussions, and Jenia also for his comments on an
earlier version of this manuscript. We acknowledge partial support by the German Federal Ministry of Education through
our German-Japanese Computational Neuroscience Project (BMBF Grant 01GQ1343), EuroSPIN, the Helmholtz Alliance
through the Initiative and Networking Fund of the Helmholtz Association and the Helmholtz Portfolio theme “Supercomputing and Modeling for the Human Brain” and the European Union Seventh Framework Programme (FP7/2007-2013)
under grant agreement no. 604102 (HBP). We also thank the INCF for travel support enabling work meetings.
References
Antonelo, E. A., Schrauwen, B., Dutoit, X., Stroobandt, D., & Nuttin, M. (2007). Event Detection and Localization in
Mobile Robot Navigation Using Reservoir Computing. In J. M. de Sá, L. A. Alexandre, W. Duch, & D. Mandic (Eds.),
Artificial Neural Networks – ICANN 2007, Volume 4669 of Lecture Notes in Computer Science, Berlin, Heidelberg, pp.
660–669. Springer Berlin Heidelberg.
Averbeck, B. B. (2009). Poisson or not Poisson: differences in spike train statistics between parietal cortical areas.
Neuron 62 (3), 310–1.
Braitenberg, V. (1986). Vehicles: Experiments in synthetic psychology. MIT press.
Calvin, W. H., & Stevens, C. F. (1967). Synaptic noise as a source of variability in the interval between action potentials.
Science (New York, N.Y.) 155 (3764), 842–4.
Chorley, P., & Seth, A. K. (2008). Closing the sensory-motor loop on dopamine signalled reinforcement learning. In From
Animals to Animats 10, pp. 280–290. Springer.
Dasgupta, S., Wörgötter, F., & Manoonpong, P. (2013). Information dynamics based self-adaptive reservoir for delay
temporal memory tasks. Evolving Systems 4 (4), 235–249.
17
Dayan, P., & Abbott, L. F. (2001). Theoretical Neuroscience: Computational and Mathematical Modeling of Neural
Systems.
Djurfeldt, M., Hjorth, J., Eppler, J. M., Dudani, N., Helias, M., Potjans, T. C., Bhalla, U. S., Diesmann, M., Kotaleski,
J. H., & Ekeberg, Ö. (2010). Run-time interoperability between neuronal network simulators based on the music
framework. Neuroinformatics 8 (1), 43–60.
Eliasmith, C., & Anderson, C. H. (2004). Neural engineering: Computation, representation, and dynamics in neurobiological systems. MIT press.
Eppler, J. M., Pauli, R., Peyser, A., Ippen, T., Morrison, A., Senk, J., Schenck, W., Bos, H., Helias, M., Schmidt, M.,
Kunkel, S., Jordan, J., Gewaltig, M.-O., Bachmann, C., Schuecker, J., Albada, S., Zito, T., Deger, M., Michler, F.,
Hagen, E., Setareh, H., Riquelme, L., Shirvani, A., Duarte, R., Deepu, R., & Plesser, H. E. (2015). Nest 2.8.0.
Frémaux, N., Sprekeler, H., & Gerstner, W. (2013). Reinforcement Learning Using a Continuous Time Actor-Critic
Framework with Spiking Neurons. PLoS Computational Biology 9 (4), e1003024.
Frémaux, N., Sprekeler, H., & Gerstner, W. (2013). Reinforcement learning using a continuous time actor-critic framework
with spiking neurons. PLoS Comput Biol 9 (4), e1003024.
Friedrich, J., & Lengyel, M. (2015). Goal-directed decision making with spiking neurons.
Furber, S. B., Lester, D. R., Plana, L. A., Garside, J. D., Painkras, E., Temple, S., & Brown, A. D. (2013). Overview of
the spinnaker system architecture. Computers, IEEE Transactions on 62 (12), 2454–2467.
Gewaltig, M.-O., & Diesmann, M. (2007). Nest (neural simulation tool). Scholarpedia 2 (4), 1430.
Hanuschkin, A., Kunkel, S., Helias, M., Morrison, A., & Diesmann, M. (2010). A general and efficient method for
incorporating precise spike times in globally time-driven simulations. Frontiers in neuroinformatics 4 (October), 113.
Helias, M., Kunkel, S., Masumoto, G., Igarashi, J., Eppler, J. M., Ishii, S., Fukai, T., Morrison, A., & Diesmann, M.
(2012). Supercomputers ready for use as discovery machines for neuroscience. Front Neuroinform 6 (26), 2.
Hines, M., & Carnevale, N. T. (2001). Neuron: a tool for neuroscientists. The Neuroscientist 7 (2), 123–135.
Hinkel, G., Groenda, H., Vannucci, L., Denninger, O., Cauli, N., & Ulbrich, S. (2015). A Domain-Specific Language
(DSL) for Integrating Neuronal Networks in Robot Control. In Proceedings of the 2015 Joint MORSE/VAO Workshop
on Model-Driven Robot Software Engineering and View-based Software-Engineering, MORSE/VAO ’15, New York,
NY, USA, pp. 9–15. ACM.
18
Jitsev, J., Morrison, A., & Tittgemeyer, M. (2012). Learning from positive and negative rewards in a spiking neural
network model of basal ganglia. In Neural Networks (IJCNN), The 2012 International Joint Conference on, pp. 1–8.
IEEE.
Kunkel, S., Schmidt, M., Eppler, J. M., Plesser, H. E., Masumoto, G., Igarashi, J., Ishii, S., Fukai, T., Morrison, A.,
Diesmann, M., et al. (2014). Spiking network simulation code for petascale computers. Frontiers in neuroinformatics 8 (78).
Legenstein, R., & Maass, W. (2014). Ensembles of spiking neurons with noise support optimal probabilistic inference in
a dynamically changing environment. PLoS Comput Biol 10 (10), e1003859.
Moren, J., Sugimoto, N., & Doya, K. (2015). Real-time utilization of system-scale neuroscience models. Journal of the
Japanese Neural Network Society 22 (3), 125–132.
Morrison, A., Aertsen, A., & Diesmann, M. (2007). Spike-timing-dependent plasticity in balanced random networks.
Neural computation 19 (6), 1437–67.
Potjans, W., Diesmann, M., & Morrison, A. (2011). An imperfect dopaminergic error signal can drive temporal-difference
learning. PLoS Comput Biol 7 (5), e1001133.
Quiñonez, Y., Ramirez, M., Lizarraga, C., Tostado, I., & Bekios, J. (2015). Autonomous Robot Navigation Based on
Pattern Recognition Techniques and Artificial Neural Networks. In J. M. F. Vicente, J. R. Álvarez-Sánchez, F. d. l. P.
López, F. J. Toledo-Moreo, & H. Adeli (Eds.), Bioinspired Computation in Artificial Systems, Number 9108 in Lecture
Notes in Computer Science, pp. 320–329. Springer International Publishing.
Quigley, M., Conley, K., Gerkey, B., Faust, J., Foote, T., Leibs, J., Wheeler, R., & Ng, A. Y. (2009). Ros: an open-source
robot operating system. In ICRA workshop on open source software, Volume 3, pp. 5.
Rotter, S., & Diesmann, M. (1999). Exact digital simulation of time-invariant linear systems with applications to neuronal
modeling. Biological cybernetics 81 (5-6), 381–402.
Schemmel, J., Bruderle, D., Grubl, A., Hock, M., Meier, K., & Millner, S. (2010). A wafer-scale neuromorphic hardware
system for large-scale neural modeling. In Circuits and systems (ISCAS), proceedings of 2010 IEEE international
symposium on, pp. 1947–1950. IEEE.
19
| 9cs.NE
|
1
Sensor Fault Detection, Isolation and Identification
Using Multiple Model-based Hybrid Kalman Filter
for Gas Turbine Engines*
arXiv:1505.02063v2 [cs.SY] 27 Aug 2015
Bahareh Pourbabaee1 , Nader Meskin2 and Khashayar Khorasani1
Abstract
In this paper, a novel sensor fault detection, isolation and identification (FDII) strategy is proposed
by using the multiple model (MM) approach. The scheme is based on multiple hybrid Kalman filters (HKF) which represents an integration of a nonlinear mathematical model of the system with a
number of piecewise linear (PWL) models. The proposed fault detection and isolation (FDI) scheme is
capable of detecting and isolating sensor faults during the entire operational regime of the system by
interpolating the PWL models using a Bayesian approach. Moreover, the proposed multiple HKF-based
FDI scheme is extended to identify the magnitude of a sensor fault by using a modified generalized
likelihood ratio (GLR) method which relies on the healthy operational mode of the system. To illustrate
the capabilities of our proposed FDII methodology, extensive simulation studies are conducted for a
nonlinear gas turbine engine. Various single and concurrent sensor fault scenarios are considered to
demonstrate the effectiveness of our proposed on-line hierarchical multiple HKF-based FDII scheme
under different flight modes. Finally, our proposed HKF-based FDI approach is compared with various
filtering methods such as the linear, extended, unscented and cubature Kalman filters (LKF, EKF, UKF
and CKF, respectively) corresponding to both interacting and non-interacting multiple model (MM)
based schemes. Our comparative studies confirm the superiority of our proposed HKF method in terms
of promptness of the fault detection, lower false alarm rates, as well as robustness with respect to the
engine health parameters degradations.
I. INTRODUCTION
An effective fault detection, isolation and identification (FDII) technology can play a crucial
role in improving the system availability, safety and reliability as well as reducing the maintenance costs and risks of catastrophic failures. Over the past few years, many researchers have
focused on proposing sophisticated fault detection and isolation (FDI) schemes constituting as
a significant component of an FDII solution [1] and [2]. The general methodology for an FDII
consists of several important steps including: (a) generation of residuals as indicators of faults,
(b) isolation of the faulty actuator, or sensor, or component element, and (c) identification and
estimation of the severity of the fault parameter.
The FDII approaches can be categorized into three distinct groups including the modelbased, data-based and hybrid methods [3] and [4]. The model-based FDI approaches have been
introduced in survey papers [1], [2] and [5]–[7] and can be grouped into basic approaches,
namely (i) Kalman filter and unknown input observer based approaches, (ii) parity relations,
(iii) optimization-based algorithms, and (iv) parameter estimation and identification techniques.
*This publication was made possible by NPRP grant No. 4 - 195 - 2 - 065 from the Qatar National Research Fund (a member
of Qatar Foundation). The statements made herein are solely the responsibility of the authors.
1
B. Pourbabaee and K. Khorasani are with the Department of Electrical and Computer Engineering, Concordia University,
Montreal, Canada. b [email protected] and [email protected].
2
N. Meskin is with the Department of Electrical Engineering, Qatar University, Doha, Qatar. [email protected].
August 28, 2015
DRAFT
2
Multiple model approach is a popular structure for FDII. It enables one not only to detect and
isolate different faults but it also provides one with information on the magnitude of the fault
and its identification information [8]–[13].
Since early research on FDII, gas turbines have been one of the challenging application areas
which have received much attention. The main approach in gas turbines FDII is based on the
Gas Path Analysis (GPA), which enables one to accomplish the actuator, sensor and component
diagnosis by observing the engine’s parameters such as the rotor speed, temperature and pressure
at different stages and the fuel flow rates [14]. Various fault diagnosis techniques have been
developed for gas turbines based on GPA ranging from Kalman filters [15]–[18], neural networks
[19], fuzzy logic [20], genetic algorithms [21], sliding mode observer [22], component adaptation
approach [23] - [24], and hybrid diagnosis [4]. Some of the linear approaches have been extended
to fully nonlinear models of the engine that have led to nonlinear extensions of the Kalman filters
to extended Kalman filters (EKF) and unscented Kalman filters (UKF) [12] and [25].
In this paper, a modular and a hierarchical non-interacting MM-based approach is proposed
for both single and concurrent permanent sensors fault detection and isolation in a single spool
jet engine. In our proposed hierarchical approach, multiple levels of detection filters are used to
detect and isolate concurrent faults where at each time only one level of the scheme is active
according to the engine sensors health status.
We also propose a hybrid Kalman filter (HKF) [17] scheme that consists of a single nonlinear
on-board engine model (OBEM) augmented with piecewise linear (PWL) models constituting
as the MM-based estimators to cover the entire engine operating regime. Also, the discrete-time
derivation of the HKF is formally shown in this paper. Therefore, multiple HKFs are constructed
to ensure that the FDII algorithm works effectively in a wide range of operating conditions by
decomposing the engine operating range into sub-regions each represented by a PWL model.
We then apply a Bayesian approach to generate a general combined model based on the PWL
models normalized weights. This provides us with a soft transition or interpolation among the
PWL models. At any given operating point, the combined model represents the nonlinear system
dynamics with less error than a single PWL model.
For interpolating the PWL models, various techniques are available in the literature including
Takagi-Sugeno fuzzy models [26], the functional state approach using discrete representations
such as automaton for describing the transitions between regions [27], Markov mixtures of
experts to define the transitions [28], etc. Unlike the functional state approach, the Bayesian
and fuzzy approaches enable soft transitions among the PWL models as the operating condition
changes between various points. Moreover, the Bayesian approach uses the statistical method to
infer the applicable and most appropriate operating regime at each time instant.
Sensor fault severity identification is another important part of our proposed FDII framework
which may either be utilized for reconstructing the correct sensor measurements to design an
output feedback control strategy or to remove the faulty sensor from the measurement set. The
reconstructed fault-free system outputs are essential for a normal closed-loop system operation
with a faulty sensor in order to maintain the system stability and set-point tracking under faulty
situations [29]. Towards this end, we have integrated the generalized likelihood ratio (GLR)based method [30]–[32] with our MM-based scheme to estimate the sensor fault severity under
various single and concurrent fault scenarios. The MM-based scheme does provide us with the
time and location of a sensor fault that subsequently simplifies the use of the GLR method
without requiring to solve an optimization problem for determining the fault detection time as
part of its fault estimation process.
August 28, 2015
DRAFT
3
One of the major non-fault related factors leading to false alarms in a given FDI scheme arises
due to the normal aging effects of the engine components such as compressor fouling, turbine
erosion and wear or buckling of a component. The above phenomena are associated with gradual
degradation of the health parameters from their healthy reference baselines, which is a type
of modeling uncertainty. Different approaches, including the robust generation and evaluation
methods, are introduced in [33] to improve the robustness of the on-line FDI algorithm with
respect to modeling uncertainties. In [34], modeling uncertainty is assumed as the structured input
with known distribution matrix and certain rank conditions in order to design a decoupling filter
to achieve the robustness. However, in most applications including ours, usually it is challenging
to model the uncertainty in a structured format. Therefore, robust residual evaluation methods in
which the adaptive threshold is designed based on the upper bound of the unstructured modeling
uncertainty that use the historical data have been introduced. Nevertheless, it is required to
design an individual threshold for each fault type. In addition, in few works such as [4], a
periodic updating mechanism is used for the on-board engine model in order to enhance the
robustness and the fault sensitivity.
In our paper, it is assumed that the reference baselines for the engine health parameters, namely
the efficiencies and the mass flow rates of the turbine and the compressor, can be estimated after
a certain number of flights by using an off-line health monitoring module [35] for determining
the percentages of degradations. The applied method can be utilized off-line in a ground station
given that the health parameters are degraded more slowly as compared to an abrupt fault.
Consequently, in order to ensure the reliability of our proposed FDII scheme during the engine
life cycle, the reference baselines are periodically updated for the OBEM health parameters.
This procedure will be useful in preventing occurrence of false alarms due to the engine health
parameters degradations. Despite this updating process, there is always a mismatch between the
values of the real engine health parameters and the ones that are used in the OBEM due to the
off-line estimation errors of the health monitoring method. Therefore, we will also investigate
the robustness of our proposed FDII scheme with respect to various estimation errors through
performing Monte Carlo simulations.
To summarize the main contributions of this paper can be stated as follows:
1) The explicit derivations of the discrete-time HKF scheme is formally provided in Appendix
A to complement the continuous-time HKF scheme derivations provided in [17].
2) A modular and hierarchical non-interacting MM-based HKF structure is developed to detect
and isolate single and concurrent sensor faults during the entire engine operating regime
(flight profile) having lower fault detection time and better robustness towards the engine
health parameters degradations as compared with the other linear and nonlinear filtering
methods in the literature. Our method is capable of operating during the entire engine life
cycle through periodically updating the reference baselines of the engine health parameters
for the nonlinear OBEM. In addition, the computational time of our proposed method is
lower than the investigated linear and nonlinear methods.
3) Through performing extensive simulation studies and measuring the mean of residual
signals, it is shown that the HKF is capable of estimating the engine outputs more
accurately with less number of operating points and false alarm rates as compared to
multiple linear Kalman filters (MLKF) based FDI method.
4) Our proposed MM-based HKF structure is integrated with the GLR scheme to estimate
the fault severity. This is accomplished by specifically eliminating the estimation of the
fault detection time which is a necessary step in the standard GLR scheme.
August 28, 2015
DRAFT
4
5) Our proposed multiple HKF-based FDI scheme is compared with the MM-based schemes
that utilize various linear and nonlinear filtering approaches such as the linear Kalman filter
(LKF), extended Kalman filter (EKF), unscented Kalman filter (UKF) and the cubature
Kalman filter (CKF) in terms of the promptness of the fault detection, false alarm and
incorrect fault detection rates, robustness with respect to the engine degradations and
computational time.
This paper is the extended version of [36] from both theoretical and case-based simulation
scenarios perspectives. Theoretically, we formally show how the HKF is derived and how
the previously proposed sensor FDI algorithm in [36] is integrated with the GLR method to
investigate the sensor fault identification problem. The case study scenarios are also extended
from the following perspectives, namely (a) different single and concurrent sensor faults are
considered, (b) the reductions in the fault detection times, lower false alarms and incorrect fault
detection rates under various parameter uncertainty levels are investigated, and (c) our proposed
FDI scheme is compared with linear and nonlinear observer-based methods in the literature
(LKF, EKF, UKF, and CKF) in terms of fault detection time, robustness towards the engine
health degradations and the computational time.
The remainder of the paper is organized as follows. Section II provides the design of the
HKF which includes the OBEM and the PWL models. The interpolation of the PWL models is
provided in Section III, and Section IV describes the overall hierarchical MM-based FDI scheme
applicable to both single and concurrent sensor faults. The modified GLR-based approach is
developed and described in Section V for estimating the severity of a sensor bias fault. A number
of Monte Carlo simulation scenarios and cases associated with various faults under multiple
operational conditions are presented. Finally, comparisons with various linear and nonlinear
filtering methods (LKF, EKF, UKF and CKF) are included in Section VI, and the paper is
concluded in Section VII.
II. HYBRID KALMAN FILTER (HKF) DESIGN
An actual aircraft gas turbine engine, used for deriving the on-board engine model (OBEM),
and which is used for on-line diagnostic analysis can be described according to the following
representation:
Ẋ(t) = Fc (X(t), H(t), U (t), ζ(t)),
Y (t) = Gc (X(t), H(t)) + v(t),
(1)
where X(t) ∈ Rn , H(t) ∈ Rr , Y (t) ∈ Rq , U (t) ∈ Rp , ζ(t) ∈ R2 and v(t) ∈ Rq denote
the engine state variables, health parameters, sensor measurements, input signals, and Gaussian
zero-mean process and measurement noise at time t, respectively. The engine dynamics is also
an implicit function of the ambient condition parameters including the ambient temperature and
pressure. The ambient parameters are defined in terms of the environmental parameters including
the altitude and the Mach number. The engine health parameters in (1) will become degraded
from their healthy reference baselines during the entire engine life cycle. Moreover, it is assumed
that the inputs and outputs of an actual engine are discretized with sufficiently small sampling
period for performing simulation and implementation of our proposed on-line fault detection,
isolation and identification (FDII) scheme. The discrete-time representation of the variables in
(1) are denoted by X(k), Y (k), H(k), U (k), ζ(k) and v(k) with the discrete dynamic functions
F and G replacing Fc and Gc , respectively.
August 28, 2015
DRAFT
5
The hybrid Kalman filter (HKF) consists of two main blocks that include a nonlinear on-board
engine model (OBEM) and multiple piecewise linear (PWL) models derived at different operating
points to cover the entire engine operating range. The continuous-time fault-free representation
of the OBEM which can also be derived based on thermodynamics laws is now given as follows:
ẊOBEM (t) = fc (XOBEM (t), HOBEM (t), U (t)),
YOBEM (t) = gc (XOBEM (t), HOBEM (t)),
(2)
where XOBEM (t) ∈ Rn and YOBEM (t) ∈ Rq denote the OBEM state variables and outputs. Both
the actual engine and the OBEM operate in parallel under the same flight conditions. Moreover,
HOBEM (t) ∈ Rr denotes the OBEM health parameters that can be represented by HOBEM (t) =
λT h(XOBEM (t)), in which λ defines the OBEM health parameters degradation factors (reference
baselines). This parameter is an all-ones vector for the OBEM that represents a healthy or nondegraded engine. It is also assumed that the OBEM health parameters can be periodically updated
and λ is considered as fixed in between the updating intervals. In addition, h(XOBEM (t)) denotes a
state-dependent smooth function that corresponds to the compressor and the turbine performance
maps and is modeled as a polynomial function for our developed gas turbine engine model. The
updating process for the OBEM health parameters will be described below in detail. Moreover,
for performing simulation and implementations, the continuous-time OBEM model is assumed
to be discretized with sufficiently small sampling period. The discrete-time representation of the
variables in (2) are denoted by XOBEM (k), YOBEM (k), HOBEM (k) and U (k) with the discrete
dynamic functions f and g replacing fc and gc , respectively.
The continuous-time OBEM model is linearized and discretized at multiple operating points
(corresponding to engine steady-state values) that are denoted by (Xssi , Ussi , Yssi ) with sufficiently
small sampling period. The constructed multiple linear discrete-time state-space models is now
given by:
∆Xi (k + 1) = A|Xssi ∆Xi (k) + B|Ussi ∆Ui (k),
∆Yi (k) = C|Xssi ∆Xi (k),
(3)
where i ∈ 1, . . . , L (L is the number of the operating points), A|Xssi , B|Ussi and C|Xssi denote the
state-space matrices associated with the ith operating point, and ∆Xi (k) = XOBEM (k) − Xssi ,
∆Yi (k) = YOBEM (k) − Yssi and ∆Ui (k) = U (k) − Ussi . In the Appendix A, the linearization and
discretization process of the continuous-time OBEM model is described in more detail. In this
paper, it is assumed that the OBEM linearized models are obtained with the health parameters that
are set to their healthy reference baselines. Also, the health parameter effects due to changes in
the engine state variables have been implicitly incorporated into the matrices A and C, although
the deviations from the healthy reference baselines are not incorporated in the linear models.
For each linearized model, an off-line linear Kalman filter is designed to estimate both the
actual engine states and sensor outputs as follows:
∆X̂i (k + 1) = A|Xssi ∆X̂i (k) + B|Ussi ∆Ui (k) + Kssi (Y (k) − Ŷi (k)),
∆Ŷi (k) = C|Xssi ∆X̂i (k),
(4)
where ∆X̂i (k) = X̂(k) − Xssi , ∆Ŷi (k) = Ŷ (k) − Yssi and Kssi denotes the steady-state Kalman
filter gain matrix. For the purpose of constructing the multiple model HKFs, the steady-state
Kalman gain matrices as well as the matrices A|Xssi and C|Xssi that are constructed associated
August 28, 2015
DRAFT
6
with multiple operating points are stored in a look-up table. According to (4), the linear Kalman
filter does not take into account the effects of the health parameter degradations from their healthy
reference baselines since the state-space matrices have already been determined for an all-ones
λ. Consequently, it does not have the required level of robustness for handling a vast number
of health degradations that occur during the entire engine life cycle. Therefore, it is essential
to update the health parameter reference baselines of the OBEM to maintain the reliability and
accuracy of the state estimates and the performance of the FDII scheme through out the entire
engine life cycle operation. For this purpose, the health parameter reference baselines can be
estimated by an off-line health monitoring system and then periodically updated in the discretetime OBEM model. Therefore, the on-line FDII scheme is integrated with an off-line health
monitoring system.
The frequency of the health monitoring system updates is significantly lower than that of
the on-line FDII algorithm, since the health degradation process is slow and gradual during
one flight, although their accumulated effects after a number of flights may generate a large
discrepancy between the OBEM and the actual engine. The off-line health monitoring module
can either be a single augmented Kalman filter that estimates the health parameters based on
the collected data during several flights or a nonlinear approximation method such as a neural
network that receives the engine historical data. It must be noted that both the health monitoring
and updating mechanism can be performed on-line having a sufficiently large sampling interval.
However, the health parameter estimation process needs to be terminated whenever a fault occurs
in the engine to avoid generating incorrect estimates of the health parameters.
The off-line health monitoring module uses the collected input and the measured output data
to estimate the health parameter reference baselines, that is λ̂. The estimated λ̂ is updated
periodically in the OBEM model within a determined time interval T that is usually a given
number of flights or days. The update process feeds λ̂ into the OBEM model so that the online FDII scheme can operate within the neighborhood of the degraded actual engine condition.
Consequently, the OBEM model can now be re-written as follows:
XOBEM (k + 1) = f (XOBEM (k), λ̂h(XOBEM (k)), U (k)),
YOBEM (k) = g(XOBEM (k), λ̂h(XOBEM (k))).
(5)
It must be noted that the OBEM model that is provided in (2) is used only once for performing
the linearization process with λ set to an all-ones vector, although the OBEM model in (5), with
its reference baselines periodically updated, is the one that is utilized in our proposed HKF
structure as well as in our on-line FDII scheme. For sake of notational simplicity, we use the
same notations for the above two versions of the OBEM model.
To construct the HKF representation we modify (4) where the steady-state variables are
replaced by the OBEM states and outputs that are obtained from (5) and also by using the
previously stored steady-state Kalman filter gain and state-space matrices as follows:
X̂i (k + 1) − XOBEM (k + 1) = A|Xssi (X̂i (k) − XOBEM (k)) + Kssi (Y (k) − Ŷi (k)),
Ŷi (k) = C|Xssi (X̂i (k) − XOBEM (k)) + YOBEM (k).
(6)
In the above model the effects of the input and the B matrix are eliminated from the HKF
formulation given the fact that these have already been accounted for by the OBEM model. The
procedure for derivation of the discrete-time HKF is formally shown in the Appendix A.
August 28, 2015
DRAFT
7
Consequently, multiple HKFs are designed by using (6) for multiple operating points. After
updating the OBEM health parameter reference baselines, there is no longer a need to recalculate
the A|Xssi , C|Xssi and Kssi matrices for each operating point, since the effects of the health
parameter degradations have been incorporated into the XOBEM and YOBEM as given by (5).
One of the factors that can affect the efficiency of our proposed FDII algorithm is the
estimation error of the off-line health monitoring module, that leads to mismatches between the
actual engine and the OBEM outputs. Larger estimation errors may increase the fault detection
time, and lead to occurrence of false alarms and incorrect fault detection rates. The acceptable
ranges of the estimation errors that do not lead to false alarms and incorrect fault detection
rates will be specified subsequently in Section VI for different health parameters under various
healthy and faulty scenarios. Moreover, the reliability of our proposed on-line FDII strategy as a
function of different mismatching factors between the actual operational engine and the OBEM
(as represented by the reference baseline estimation errors (RBEE = |λ−λ λ̂| )) and the magnitude
of the process and measurement noise signals will be investigated subsequently in Section VI
by means of a confusion matrix analysis.
In order to develop our proposed HKF-based scheme as an FDII strategy, multiple piecewise
linear (PWL) models need to be generated for each fault hypothesis at various operating points.
Moreover, the PWL models will be integrated and fused to cover the entire operational regime
of an engine. The detail description of this process is provided below.
III. P IECEWISE L INEAR M ODELS (PWL) I NTERPOLATION
Any given linear model of a nonlinear system has a limited operating range in which it remains
valid. Nevertheless, our ultimate goal is to obtain a globally valid model which is valid for the
entire operating regime. Therefore, the full operating range is divided into several sub-regions
where each is defined around an operating point for which a piecewise linear (PWL) model
can be derived [37]. The PWL models can then be integrated in order to construct a parametervarying general model whose parameters are the PWL models weights that are obtained through
an on-line Bayesian approach. This will provide one with a soft interpolation among the PWL
models as opposed to a hard switching among them. In this paper, the engine inputs including
the fuel flow rate and ambient variables are used to partition the engine operational regime into
multiple operating points for which the PWL models are constructed. These operating points are
associated with different flight conditions such as take-off, climbing, cruise and landing modes.
The selected number of operating points depends on (a) the required HKF state estimation
accuracy, (b) the FDII strategy reliability on correct decisions, and (c) false alarm rates within
the range of the applied health parameter degradations. Hence, if there are no concerns on the
memory utilization, the number of the operating points can be selected to be as high as possible
to enhance the HKF estimation accuracy and the FDII scheme valid decision rates and also to
decrease the false alarm rates.
One of the important advantages of our proposed HKF scheme is in requiring a smaller
number of operating points as compared to standard linear Kalman filters for covering an entire
operational regime of the engine. This is facilitated and made possible due to substitutions of
the steady-state variables in (6) by the OBEM state and output variables. The operating range
of a PWL model in (4) is only limited to the neighborhood of a corresponding operating point,
although this can be extended to a larger range in (6) given that XOBEM and YOBEM are changed
according to the engine operating condition, which enable the PWL model to be valid in a wider
range.
August 28, 2015
DRAFT
8
Since, the sensor fault is injected into the actual gas turbine engine and not the OBEM, we
have for the faulty engine
X(k + 1) = F(X(k), H(k), U (k), ζ(k)),
Y (k) = G(X(k), H(k)) +
q
X
bs zs δ(k − kf s ) + v(k),
(7)
s=1
where F and G represent the discrete-time dynamic equations of the actual gas turbine engine, q
is the number of sensors, bs represents the sth sensor bias fault magnitude and zs represents the
fault location vector that has a unit value for the sth element while the other elements are set to
zero, and δ(k − kf s ) denotes a unit step function that occurs at the sample kf s corresponding to
the sth fault occurrence time. The bias is set to zero for the healthy sensor scenario. Therefore,
there are a total of q + 1 sensor modes (corresponding to one healthy and q faulty sensor modes).
The PWL models constructed for multiple operating points are now used to compute the
corresponding Kssi for various sensor modes of the (7). The matrices A|Xssi , C|Xssi and Kssi are
finally stored in a look-up table and are used to construct multiple HKFs (MHKFs) as given by
(6) for all the sensor modes. Therefore, the HKF for the j th sensor mode at the ith operating
point is designed as follows:
X̂ (i,j) (k + 1) − XOBEM (k + 1) = Ai (X̂ (i,j) (k) − XOBEM (k)) + Kssi (Y (k) − Ŷ (i,j) (k)),
Ŷ (i,j) (k) = C i (X̂ (i,j) (k) − XOBEM (k)) + YOBEM (k) + bdj aj δj (k),
(8)
where i = 1, . . . , L, j = 1, . . . , (q + 1), bdj denotes the pre-determined sensor bias fault that can
be different from the actual sensor fault bs that is injected into (7), aj denotes the q-dimensional
vector and is one of q + 1 modes of a which is the fault parameter vector. For the healthy mode
or j = 1, the fault parameter vector, a, is set to a zero vector whereas for j = 2, . . . , q + 1, aj
has a unit value for the (j − 1)th element and all the other elements are set to zero. Moreover,
Ai = A|Xssi , C i = C|Xssi and Kssi are the previously stored state-space and Kalman gain matrices
that depend on the ith operating point. Therefore, L × (q + 1) HKFs are constructed covering the
entire engine operating range corresponding to different sensor modes. It must be noted that the
state-space matrices depend only on the operating points but the Kalman gain matrices depend
on both the operating points and the noise covariance matrices that are the same for all the
sensor fault modes, given that the OBEM does not take into account the effects of sensor faults.
Although the operating range of a PWL model in the HKF scheme has been increased as
compared to that of the one that uses the standard linear Kalman filters by replacing the steadystate values with the OBEM variables, still none of the PWL models are solely valid over the
entire operating range of an engine. Therefore, corresponding to each PWL model one can
associate a validity function that is based on its normalized weight as obtained by means of the
Bayes formula. For this purpose, the residual vectors γ (i,j) and the covariance matrices S (i,j)
that are generated by the multiple HKFs are used to compute the likelihood function f (i,j) for
the j th sensor mode at the ith operating region as follows:
γ (i,j) (k) = Y (k) − Ŷ (i,j) (k), S (i,j) (k) = cov(γ (i,j) (k)),
−1
1
p
× exp[ (γ (i,j) (k))T (S (i,j) (k))(−1) (γ (i,j) (k))],
f (i,j) (γ (i,j) (k)) =
2
(2π)q/2 |S (i,j) (k)|
August 28, 2015
(9)
DRAFT
9
where it is assumed that the innovation sequence generated by the hybrid Kalman filter, γ (i,j) (k),
is a Gaussian white noise process with zero mean and covariance matrix S (i,j) (k) that is calculated
numerically. The normalized weights for the j th sensor mode are updated recursively by using
the Bayes formula as follows:
f (i,j) (γ (i,j) (k))w(i,j) (k − 1)
.
w(i,j) (k) = PL
(i,j) (γ (i,j) (k))w (i,j) (k − 1)
i=1 f
(10)
The weights computed above should also remain outside a narrow bound to avoid becoming
close to zero as:
if w(i,j) (k) > ρ then w(i,j) (k) = w(i,j) (k),
if w(i,j) (k) ≤ ρ then w(i,j) (k) = ρ,
(11)
where ρ is a design parameter that is determined by trial and error and it invokes the PWL
models that have very small weights and it avoids them from being removed from the set of L
models. It is also useful for numerical robustness of the recursive weight algorithm [38].
Following the computation of the normalized weights, multiple time-varying models are now
constructed that cover the entire engine operating range subject to various sensor modes. For each
sensor mode, the corresponding model state-space matrices as well as the weighted innovation
vector and the covariance matrix are obtained by using the PWL model state-space matrices and
their associated normalized weights as follows:
Ajc (k)
=
L
X
w
(i,j)
i
(k)A ,
Ccj (k)
=
i=1
γcj (k) =
L
X
i=1
Ajc (k)
L
X
w(i,j) (k)C i ,
i=1
w(i,j) (k)γ (i,j) (k),
Scj (k) =
L
X
(w(i,j) (k))2 S (i,j) (k),
(12)
i=1
Ccj (k)
th
denote the weighted state-space matrices of a linear time-varying model
and
where
associated with the j sensor mode. Also, γcj (k) and Scj (k) denote the weighted innovation
vector and the covariance matrix of the j th sensor mode, respectively. The above procedure is
now designated as the PWL models interpolation. Consequently, (q + 1) weighted innovation
vectors and covariance matrices that operate through out the entire engine operational regime
are used in the next two sections to develop our proposed FDII scheme.
IV. SENSOR FDI VIA MULTIPLE-MODEL-BASED SCHEME
In this section, the overall structure of our proposed MM-based FDI scheme is presented.
It is assumed that the fault parameter vector can take on only one of (q + 1) sensor modes
as aj . Therefore, at each operating point there are (q + 1) PWL models; one for the healthy
sensors scenario and q corresponding to various faulty sensor scenarios, that have been designed
and integrated with only one OBEM for constructing the multiple HKF (MHKF)-based scheme
as formulated in (8). The innovation vectors and the covariance matrices that are generated
by the MHKFs are fused and weighted as given by (12). Finally, there are (q + 1) weighted
innovation vectors and covariance matrices that are used in the MM-based FDI scheme, where
they are operating under different healthy and faulty sensor scenarios during the entire engine
operational regime.
August 28, 2015
DRAFT
10
Y(k)
U(k)
OBEM
𝑿
++
𝒀𝑶𝑩𝑬𝑴 (𝒌)
(𝒌 + 𝟏)
(𝒊,𝒋)
𝑿
𝒁−𝟏
𝜸 𝟏,𝒋 (𝒌), 𝑺 𝟏,𝒋 (𝒌)
(𝒌)
+-
𝜸𝒄 𝟏 (𝒌), 𝑺𝒄 𝟏 (𝒌)
𝒘(𝟏,𝒋)
𝑳
++
𝑪𝒊
-+
𝜸 𝟐,𝒋 (𝒌), 𝑺 𝟐,𝒋 (𝒌)
𝒘 𝒊,𝒋
𝒘(𝟐,𝒋)
𝒊=𝟏
+
Ai
𝑲𝒔𝒔 𝒊
Offline PWL Kalman Filter
𝜸
𝒊,𝒋
(𝒌)
𝑺
𝒊,𝒋
PWL(i , j) Model from
Look - up Table with
dimension (L , q+1)
𝒊,𝒋
𝒊,𝒋
(𝒌) 𝜸 (𝒌), 𝑺 (𝒌)
𝒘(𝒊,𝒋)
𝜸 𝑳,𝒋 (𝒌), 𝑺 𝑳,𝒋 (𝒌)
𝒘(𝑳,𝒋)
+
++
Combination of
𝒌 𝜸 𝒊,𝒋 (𝒌)
L linear Models
for jth Sensor
Mode
𝒋
𝑳
(𝒘 𝒊,𝒋 (𝒌))𝟐 𝑺 𝒊,𝒋 (𝒌)
𝒊=𝟏
𝜸𝒄 𝟐 (𝒌), 𝑺𝒄 𝟐 (𝒌)
𝜸𝒄 𝒋 (𝒌), 𝑺𝒄 𝒋 (𝒌)
𝒋
𝜸𝒄 𝒌 , 𝑺𝒄 (𝒌)
𝑷𝟏
𝑷
𝟐
Conditional
Hypothesis
Probability 𝑷
𝒋
Calculator
Equations
(13) and (14) 𝑷𝒒+𝟏
MAX
𝝀
With Updating
Interval T
(𝒊,𝒋)
Bayesian Weight Estimation
𝑿𝑶𝑩𝑬𝑴 (𝒌)
Fault
Detection
&
Isolation
𝜸𝒄 (𝒒+𝟏) (𝒌), 𝑺𝒄 (𝒒+𝟏) (𝒌)
Fig. 1: The MM-based FDI scheme based on the HKF corresponding to multiple operating
regime.
A. Single Fault Detection and Isolation (FDI) Scheme
In the MM-based approach [10]–[12], the hypothesis conditional probability Pj (k) is defined
as the probability that the fault parameter a assumes the mode aj , j = 1, . . . , q+1, conditioned on
the observed measurement history up to the k th sample, that is: Pj (k) = Pr[a = aj |Y(k) = Yk ],
where Y(k) is the random vector measurement history with Y (1), Y (2), . . . , Y (k) partitions displaying the available measurements up to the k th sample time. Similarly, Yk is the measurement
history vector realization that has the partitions of Y1 , Y2 , . . . , Yk . Therefore, the conditional
probability can be computed recursively as follows:
fY (k)|a,Y(k−1) (Yj |aj , Yk−1 )Pj (k − 1)
,
Pj (k) = Pq+1
~=1 fY (k)|a,Y(k−1) (Yj |a~ , Yk−1 )P~ (k − 1)
(13)
where fY (k)|a,Y(k−1) (Yj |aj , Yk−1 ) denotes the Gaussian density function for the current measurement given by:
fY (k)|a,Y(k−1) (Yj |aj , Yk−1 ) =
(2π)q/2
1
q
× exp[
Scj (k)
−1 j
(γ (k))T (Scj (k))(−1) (γcj (k))],
2 c
(14)
where γcj (k) and Scj (k) are given by (12).
If the j th sensor mode occurs, the probability associated with the j th model will be larger
than that of the others since its corresponding innovation vector and also the determinant of the
covariance matrix will be much smaller than those that are predicted by the other filters and
which are mismatched with the assumed fault scenario. Hence, the condition of the system and
the location of a single faulty sensor can be detected and isolated based on evaluating Pj (k) and
determining its maximum value. Consequently, our proposed MM-based approach is capable of
detecting and isolating sensor faults. Fig. 1 shows the structure of the MM-based FDI scheme
that employs the MHKFs for the entire engine operational regime.
Our contribution here is the modification of our previously developed MM-based structure
in [10] and [11] that utilizes only multiple standard linear Kalman filters and is designed for a
single operating point into a general strategy that is applicable to the entire engine operational
regime by means of the multiple HKF scheme and fusion and integration of the corresponding
innovation vectors and covariance matrices.
As we will describe in more detail in Section VI, in this paper five sensor faults are considered.
Therefore, the total number of modes corresponding to each operating point of the jet engine
is six, where mode #1 (P1 ) corresponds to the healthy engine sensors and modes #2 to #6 (P2
August 28, 2015
DRAFT
11
TABLE I: Operational modes corresponding to various possible two concurrent sensor faults.
Level
First
#1
Healthy
P2
P3
Second
P4
P5
P6
Operational Modes
#2 #3 #4
P2
P3
P4
P2
P2
P2
P2
P3
P4
P3
P3
P3
P2
P3
P4
P4
P4
P4
P2
P3
P4
P5
P5
P5
P2
P3
P4
P6
P6
P6
P2
P3
P4
#5
P5
P2
P5
P3
P5
P4
P5
P5
P5
P6
P5
#6
P6
P2
P6
P3
P6
P4
P6
P5
P6
P6
P6
to P6 ) correspond to a 3% sensor bias fault injected into sensors measuring the compressor exit
temperature (P2 ) and pressure (P3 ), shaft rotational speed (P4 ), the turbine exit temperature (P5 )
and pressure (P6 ), respectively.
B. Concurrent Fault Detection and Isolation (FDI) Scheme
For detection and isolation of two concurrent faults in the gas turbine engine, a hierarchical
MM approach is proposed in [9] and [11] as shown in TABLE I. In this scheme, it is assumed that
the engine starts operating with healthy sensors when the first level of filters are active and the
FDI scheme observes the sensors condition for occurrence of one of the five faulty modes. The
active bank of filters operate based on the weighted innovation vectors and covariance matrices
that are obtained in Section III.
Once the first fault is detected and isolated according to the maximum probability criterion,
the FDI scheme will activate the second level of the filters as shown in TABLE I for detection
of the second concurrent faults. In our hierarchical MM scheme, it is assumed that the sensor
faults do not occur simultaneously and there exists always a minimum time interval between
the occurrence of two sensor faults that are called concurrent faults. TABLE I indicates all
the possible configurations for the second bank of filters that are considered in the simulation
results in Section VI. The first filter in the second level always corresponds to the detected faulty
mode in the first level. For example, if the first filter detects a 3% bias in the compressor exit
temperature sensor (P2 ), then the first and the second filters in the second level respectively
correspond to P2 and the possibility of a larger bias fault within the compressor exit temperature
sensor. Moreover, the third filter corresponds to the concurrent bias faults in the compressor exit
temperature and pressure sensors (P2 and P3 ), the forth filter corresponds to P2 and P4 , etc. This
structure can easily be extended to the third and higher levels that correspond to occurrence of
multiple concurrent sensor faults. It must be noted that when the new bank of filters is activated
in the second level, the first bank will be disabled in order to avoid adding any unnecessary
computational burden. Therefore, at any given time only six filters are operating on-line.
Remark 1. In this paper, the value of the pre-determined sensor bias fault severity, bdj as
given by (8), is considered to be the same for all the filters and only one level of fault severity is
considered for the hierarchical MM scheme, although various pre-determined sensor bias faults
can easily be incorporated into our strategy by correspondingly increasing the number of filters
that are used in this scheme.
August 28, 2015
DRAFT
12
V. SENSOR FAULT IDENTIFICATION/ESTIMATION
One of the important requirements of a general control system is the capability to integrate the
FDI scheme with a fault identification or estimation module in order to estimate the severity of
a fault that has occurred in different components of the system such as sensors and actuators. In
this paper, a modified version of the generalized likelihood ratio (GLR) scheme is developed to
estimate the severity of a sensor bias fault. The GLR was initially proposed by Willsky in [30]
and was subsequently modified in [31], [32], [39] and [40]. This method is capable of estimating
the time, location and severity of an occurred fault using a selected threshold. However, in this
paper the occurrence time and location of a fault have already been determined by means of our
proposed MHKF-based approach as presented in Section IV. Hence, the fault severity will be
estimated through the development of a modified version of the GLR scheme. The GLR method
is a detection-estimation scheme in which the time of a fault is an a priori information input
to the estimation stage. In this paper, the GLR method is simplified by removing the detection
process through integrating the GLR method with our proposed MM-based FDI scheme, and
hence there is no longer a need to select a threshold for our FDII scheme.
Let us assume that a bias fault has occurred in the sth sensor with the severity bs at the sample
kf s and it is detected and isolated at the sample kds when the P1 (k) mode probability intersects
with the P(s+1) (k) mode probability. Also, for k ≥ kds , P(s+1) (k) has the highest value among
the other mode probabilities as shown mathematically by (s + 1) = argj max Pj (k). Therefore,
the effects of the sth sensor fault still remain in the residual of the MHKF that is associated
with the healthy mode. There are L PWL models corresponding to the healthy mode that will
be used subsequently to estimate the fault severity as follows:
i
X̂ (i,1) (k + 1) − XOBEM (k + 1) = Ai (X̂ (i,1) (k) − XOBEM (k)) + Kss
(Y (k) − Ŷ (i,1) (k)),
Ŷ (i,1) (k) = C i (X̂ (i,1) (k) − XOBEM (k)) + YOBEM (k),
(15)
in which the j th index and bdj in (8) are equal to one and zero, respectively, that are designated
to the healthy sensor mode. Therefore, the residual vector of the MHKF that is designed for the
healthy sensor mode at any subsequent time can be expressed as follows:
γ (i,1) (k) = bs Gis (k, kds )zs + v(k)
for k ≥ kds ,
(16)
where zs is defined in (7) and Gis (k, kds ) denotes the failure signature matrix that provides one
with the information on the failure propagation through the filter, and which also depends on
both kds and the sample time k at which the set of L innovation vectors are computed for the
MHKFs associated with the healthy mode. The signature matrices for a sensor bias fault can be
computed by using the recursive relations for the healthy mode MHKFs as follows [39]:
Gis (k, kds ) = I − C i Ai Jsi (k − 1, kds ),
i
Jsi (k, kds ) = Ai Jsi (k − 1, kds ) + Kss
Gis (k, kds ),
(17)
where Jsi (k, kds ) denotes the signature matrix for the state correction. The signature matrices
Jsi (k, kds ) and Gis (k, kds ) are defined separately for the L operating points over a window of
duration [kds , kds + N ], where N denotes the data samples window length. Also, it is assumed
that Jsi (kds − 1, kds ) = 0. The weighted signature matrices for a detected fault that cover the
entire operating range of the gas turbine engine can be computed as follows:
August 28, 2015
DRAFT
13
Gs (k, kds ) =
L
X
w(i,1) (k)Gis (k, kds ),
i=1
Js (k, kds ) =
L
X
w(i,1) (k)Jsi (k, kds ),
(18)
i=1
where w(i,1) (k) denotes the assigned weight to the ith PWL model that is designed for the
healthy sensor mode. Therefore, the modified GLR test which consists of a maximum likelihood
estimation (MLE) of bs when kds is known is used. For performing the maximization process,
normally the log likelihood function is chosen as the GLR criterion as follows:
J =
kds
+N
X
T
γc1 (k)(Sc1 (k))(−1) γc1 (k)
−
kds
+N
X
[γc1 (k) − bs Gs (k, kds )zs ]T (Sc1 (k))(−1) [γc1 (k)
k=kds
k=kds
− bs Gs (k, kds )zs ],
(19)
where γc1 (k) and Sc1 (k) denote the weighted innovation vector and covariance matrix for the
healthy sensor mode that are computed from (12). The maximization of the log likelihood
function is associated with the minimization of the second summation in (19). It can easily be
shown that the optimal solution to the MLE problem above is the estimated fault b̂s = dcss , that
is the unbiased estimate of bs having the minimum variance as follows:
ds =
zsT
kds
+N
X
1
(−1) 1
GT
γc (k),
s (k, kds )(Sc (k))
k=kds
cs =
zsT
kds
+N
X
1
(−1)
GT
Gs (k, kds )zs .
s (k, kds )(Sc (k))
(20)
k=kds
The above procedure can also be employed for estimating the severities of multiple concurrent
faults that have already been detected and isolated by means of our proposed hierarchical MMbased FDI approach.
Remark 2. In this paper, it is assumed that there is no feedback and information sent from
the fault identification module to the FDI scheme; otherwise, the FDI scheme needs to have
a variable structure for updating the pre-determined sensor bias faults based on the estimated
faults severities.
VI. SIMULATION RESULTS
In this section, simulation results and performance evaluation of our proposed sensor FDII
scheme as applied to several fault scenarios are presented for a nonlinear mathematical model of
a commercial single spool jet engine previously developed in [12]. This model is generated based
on the rotor, volume and heat transfer dynamic behavior and is also validated by the commercially
available standard software GSP 10 [41]. The responses corresponding to our mathematical model
and the GSP match each other within an acceptable error tolerance (below 5%). For this study,
our model is extended to the entire flight profile and is simulated in the SIMULINK for use
as both the actual engine and the OBEM model. The actual engine operates at a given health
condition subject to the effects of process and measurement noise signals, whereas the OBEM
health parameters reference baselines are periodically updated to their recently estimated values
August 28, 2015
DRAFT
14
that are assumed to be generated by an off-line health monitoring module as described in Section
II.
Both the actual engine and the OBEM models operate in almost the same ambient conditions.
The set of nonlinear state-space equations that are used for the engine model are given as follows:
PCC
[(cp TC λṁC ṁC + ηCC Hu ṁf − cp TCC λṁT ṁT ) − cv TCC (λṁC ṁC + ṁf
TCC cv ṁCC
γRTCC
− λṁT ṁT )] +
(λṁC ṁC + ṁf − λṁT ṁT ),
VCC
ηm λṁT ṁT cp (TCC − TT ) − λṁC ṁC cp (TC − Td )
,
Ṅ =
π 2
JN ( 30
)
1
[(cp TC λṁC ṁC + ηCC Hu ṁf − cp TCC λṁT ṁT ) − cv TCC (λṁC ṁC + ṁf − λṁT ṁT )],
ṪCC =
cv mCC
RTM
β
ṖT =
(λṁT ṁT +
λṁ ṁC − ṁn ),
(21)
VM
β+1 C
ṖCC =
where X = [PCC , N, TCC , PT ]T denotes the state variable vector that includes the combustion
chamber pressure PCC and temperature TCC , the rotational speed N and the turbine pressure PT . There is a single actuator that supplies the fuel flow
well asi five
h (u = ṁf ) asγ−1
PCC
1
T
sensors measuring Y = [TC , PC , N, TT , PT ] , where TC = Td 1 + λη ηC [( Pd ) γ − 1] and
C
h
i
γ−1
T
TT = TCC 1 − ληT ηT [1 − ( PPCC
) γ ] denote the compressor and the turbine temperatures,
respectively. The variables PC and PT denote the pressures of the compressor and turbine,
respectively. Moreover, H = [ηC , ηT , ṁC , ṁT ]T denotes the health parameter vector, where
ηC and ηT denote the compressor and the turbine efficiencies, and ṁC and ṁT denote their
mass flow rates, respectively. The health parameters are also multiplied by their corresponding
health parameter degradation factors or the reference baselines λ = [ληC , ληT , λṁC , λṁT ]T .
The sensors are affected by the Gaussian measurement noise with the standard deviations of
SDv = [0.23, 0.164, 0.051, 0.097, 0.164]T corresponding to the percentages of output vector at
the cruise condition as defined in [12]. More details on other variables and constants in (21) are
provided in [12].
The flight condition is defined by two environmental variables, namely the altitude and the
Mach number. The ambient temperature and pressure can be computed according to Tamb =
Alt
Ts − 6.5Alt
and Pamb = Ps exp( −gM
), where Ts = 288 ◦ K and Ps = 1.01325 bar are set to the
1000
288R
standard condition, R = 8.31447 is the gas constant, g is the gravitational acceleration, and M
and Alt denote the Mach number and the altitude, respectively. Moreover, the ambient variables
are affected by the Gaussian process noise with the standard deviations of SDζ = [0.01, 0.01]T
corresponding to the percentages of standard conditions of ambient variables. The noise free
values of the ambient parameters are applied to the OBEM model. Also, the same control input
is applied to both the actual engine and the OBEM models. The system is simulated for 520
sec with the sampling rate of 0.01 sec. The profiles of the altitude, Mach number and the fuel
flow rate are shown in Fig. 2.
According to the overall structure of our proposed scheme, as shown in Fig. 1, six PWL
models are constructed (q = 5) corresponding to each operating point for the five faulty sensor
modes as well as the one healthy mode with the pre-determined sensor bias faults set to 3%
of the engine steady-state outputs under the cruise condition. The states and sensor outputs
August 28, 2015
DRAFT
15
2
0.3
0.25
0.2
0.15
x 10
4
1
1.8
0.9
1.6
0.8
1.4
0.7
Mach Number
0.35
Altitude (ft)
Fuel Flow Rate (Kg/m2)
0.4
1.2
1
0.8
0.6
0.1
0.4
0.05
0
0
200
300
Time (Sec)
(a)
400
500
600
0
0
0.5
0.4
0.3
0.2
0.2
100
0.6
0.1
100
200
300
Time (Sec)
400
500
600
0
0
100
200
300
400
500
600
Time (Sec)
(c)
(b)
Fig. 2: Profiles of (a) the fuel flow rate, (b) altitude, and (c) the Mach number during a flight
mission.
TABLE II: The operating point specifications corresponding to the designed PWL models, where
ṁf and Alt are measured in Kg/m2 and ft, respectively.
Models Corresponding
to Flight Conditions
Operating Point 1 (Climbing)
Operating Point 2 (Climbing)
Operating Point 3 (Cruise)
Operating Point 4 (Landing)
Operating Point 5 (Landing)
ṁf
M
Alt
0.38
0.38
0.25
0.3
0.3
0.2109
0.6585
0.85
0.5402
0.1203
4070.538
12708.33
16404.2
10424.87
2322.835
are estimated by means of our proposed MHKFs. Finally, six weighted matrices are calculated
according to (12) and are applied in the MM-based scheme for detecting and isolating the sensor
bias faults that occur at different points of the flight profile.
For our simulations, the measurement and process noise covariance matrices are set to 0.01I
and 0.1I, respectively, where I denotes an identity matrix. In this paper, the FDII scheme
is implemented for the entire flight profile including the climbing, cruise and landing modes.
Therefore, we assume and consider five operating points handling the entire flight profile (L = 5).
This is the minimum number of the operating points that are obtained for each faulty mode that
prevents the occurrence of a false alarm in the range of the applied health degradations. The
number of the operating points depends on the dynamics of the engine as well as the range of
the applied fuel flow rate and the environmental parameters. We simulated our proposed FDII
scheme with different number of operating points and also derived a confusion matrix for each
case in order to analyze the false alarm rates and to decide on the minimum required number of
the operating points that does not lead to any false alarms for the injected health degradations.
These details are not included here due to space limitations.
TABLE II shows the corresponding fuel flow rates as well as the flight conditions corresponding to all the applied PWL models that are designed for each sensor mode. In order to
efficiently track the variations of the system input during the climbing and landing conditions,
two operating points are selected for each of these two modes; whereas only one operating point
is selected for the cruise condition given the presence of a constant input.
During the fault detection and isolation process, a mode probability Pj (k) is generated for
each weighted model using (13) and by determining the maximum Pj (k), j = 1, . . . , q + 1, the
sensor fault is detected and isolated.
August 28, 2015
DRAFT
16
TABLE III: Fault detection time (FDT) corresponding to the maximum tolerable percentage of
the RBEE (∆λ) at different stages of the flight profile.
Faulty Scenario
Fault on TC
Fault on PC
Fault on N
Fault on TT
Fault on PT
Maximum Tolerable RBEE%
Compressor
Turbine
Compressor
Turbine
Compressor
Turbine
Compressor
Turbine
Compressor
Turbine
∆λṁC
∆λṁT
∆λṁC
∆λṁT
∆λṁC
∆λṁT
∆λṁC
∆λṁT
∆λṁC
∆λṁT
= 2.52
= 2.52
= 1.01
= 1.01
= 1.21
= 1.01
= 0.2
= 0.1
= 0.3
= 0.3
∆ληC
∆ληT
∆ληC
∆ληT
∆ληC
∆ληT
∆ληC
∆ληT
∆ληC
∆ληT
= 2.52
= 1.51
= 1.01
= 0.30
= 2.02
= 0.81
= 0.2
= 0.1
= 0.3
= 0.1
kf = 50
3.7
3.5
7.8
3.4
6.5
6.5
7.8
6
7
8
Sensor FDT (sec)
kf = 250
4.1
5.1
5.9
2.7
2.6
2.5
3
2.2
2.6
2.2
kf = 450
6.3
5.9
8
2.5
3.5
2.9
4.7
2.3
3.4
2.2
As described in Section I, the health parameter degradations due to the engine aging is one
of the non-fault related factors that should be considered for a reliable and an accurate FDII
strategy. Therefore, the performance of the FDII strategy is evaluated in presence of compressor
or turbine health parameters degradations from their healthy reference baselines. Generally, there
is a difference between the actual and the estimated health parameter reference baselines due to
off-line health monitoring estimation errors and also due to effects of noise and disturbances.
Therefore, the robustness of our proposed FDII scheme will also be evaluated subsequently in
this section with respect to the percentage of the reference baselines estimation errors.
A. Case 1: False Alarms Evaluation
Many factors such as (i) the dynamic mismatch between the OBEM model and the actual
engine model, (ii) the large estimation errors for an off-line health monitoring module, and
(iii) the process and measurement noise, may lead to false alarm flags. In order to evaluate the
reliability and the efficiency of our proposed FDII scheme in terms of avoiding false alarms, our
proposed scheme is simulated over the entire flight profile as shown in Fig. 2. The estimated
reference baselines (λ̂) may differ form the actual degraded engine (λ) and this can play a
source of uncertainty in our FDII scheme in addition to the process and measurement noise. It
should be noted that the dynamic mismatch between the actual engine and the OBEM models
is also another source of uncertainty, but this is not investigated in this paper. Therefore, the
robustness of the FDII scheme against the percentage of the reference baselines estimation error
RBEE (RBEE% = 100 × ∆λ
, where ∆λ = |λ − λ̂| and λ ∈ [0, 1]) is investigated in this section.
λ
Based on our simulation studies for the healthy sensors scenario, the FDII algorithm is robust
to the maximum percentage of reference baselines estimation error of 3% for the compressor
health parameters and a maximum of 2% for the turbine health parameters. Our proposed FDII
algorithm has declared no false alarms in the range of the above RBEEs, however the possibility
of false alarms will increase if the RBEEs are increased beyond these upper limits. In case that
one is within these limits, the mode probability of the healthy sensors is near one whereas the
other probabilities corresponding to the faulty sensor scenarios are all almost zero during the
entire flight profile. The values for the probabilities depend on the design parameter ρ. In general
the MM-based FDI is independent of ρ as long as it is selected as sufficiently small.
August 28, 2015
DRAFT
17
P1
P2
P3
P4
P5
P6
1
Mode Probability
Mode Probability
1
0.5
0
0
100
200
300
400
Time (Sec)
(a)
500
600
P
1
P
2
P
3
P
4
P
5
1.2
P
P
6
0.5
0
0
1
P
2
P
P
3
4
P
5
P
6
1
Mode Probability
1.2
1.2
100
200
300
400
Time (Sec)
(b)
500
600
0.5
0
0
100
200
300
400
Time (Sec)
500
600
(c)
Fig. 3: Mode probabilities for 3% bias fault applied at (a) kf = 50 sec to the TC sensor, (b)
kf = 250 sec to the PC sensor, and (c) kf = 450 sec to the N sensor in presence of the obtained
maximum tolerable percentage of the estimation errors in TABLE III for the compressor health
parameter reference baselines.
B. Case 2: A 3% Sensor Bias Fault Detection and Isolation
In this section, the performance of the sensor FDI scheme with respect to the fault detection
time and the robustness of the algorithm with respect to the percentage of reference baselines
estimation error (RBEE) are evaluated. This evaluation is performed during the entire flight
profile which lasts for 520 sec, when a pre-determined bias fault with the severity of 3% of the
engine steady-state output values under cruise condition occurs for a single sensor. TABLE III
shows the fault detection times for each single sensor fault scenario at different stages of the flight
profile including the climbing, cruise and landing modes. It also shows the maximum tolerable
percentages of the RBEE of engine health parameters that is denoted by ∆λ. It is assumed that
λ = 0.99. The limits are obtained through preserving two performance requirements for our
proposed FDI scheme, namely: 1) the sensor fault detection time (FDT) should be less than 8
sec, and 2) no false alarms should be generated. This implies that our proposed FDI scheme can
tolerate the reference baselines estimation error as long as the above performance requirements
are fulfilled.
In addition, Fig. 3 depicts the mode probabilities for three selected fault scenarios when the
bias fault occurs at different instants of the flight profile. The value of the injected fault is set
to 3% of the engine steady-state output values, while the percentage of the estimation errors for
the compressor health parameters reference baselines (RBEE) are set at their maximum tolerable
levels that are indicated in TABLE III.
By comparing the results in TABLE III, it can be concluded that the sensor fault detection
times during the cruise mode are much less than that of the other flight modes since there is
less variation of thrust and ambient conditions. In spite of a large input and ambient condition
variations during the climbing and the landing modes, it is still possible to detect a sensor fault
by applying our proposed FDI scheme. In order to show this capability, the sensor faults occur at
kf = 50 sec during the climbing mode, at kf = 250 sec during the cruise mode and at kf = 450
sec during the landing mode. The fault detection times are all indicated in TABLE III for various
fault scenarios.
To quantify the effectiveness and reliability of our proposed FDI scheme in presence of the
mismatch between the OBEM model and the actual engine, confusion matrices are obtained for
both the healthy and the faulty scenarios subject to different health parameter degradations and
August 28, 2015
DRAFT
18
TABLE IV: The confusion matrix for 3% RBEE of the compressor health parameters.
TC
PC
N
TT
PT
No Fault
TC
50
1
0
12
17
0
PC
0
49
0
0
0
0
N
0
0
50
0
0
0
TT
0
0
0
27
0
0
PT
0
0
0
10
29
0
No Fault
0
0
0
1
4
50
TABLE V: The confusion matrix for 3% RBEE of the turbine health parameters.
TC
PC
N
TT
PT
No Fault
TC
50
0
1
2
0
0
PC
0
50
0
0
1
1
N
0
0
48
0
0
0
TT
0
0
0
41
0
0
PT
0
0
0
0
44
0
No Fault
0
0
1
7
5
49
measurement noise. The rows in the confusion matrix are the fault conditions and the columns
are the actual isolated faults. The element in the ith row and j th column (CMij ) shows the rate
that fault j is isolated when fault i occurs. Ideally, the confusion matrix should be a purely
diagonal matrix. To obtain the confusion matrix for our sensor FDI scheme, 50 Monte Carlo
simulations are performed in which the percentage of the RBEE exceeds the maximum tolerable
limit that is reported in TABLE III. Tables IV and V show the confusion matrices for the healthy
and 3% faulty scenarios as applied to the gas turbine engine sensors during the cruise mode
subjected to 3% RBEE of both the compressor and the turbine health parameters, respectively.
In addition to the percentage of the RBEE, the measurement noise is another factor that
can affect the performance of our proposed FDI scheme. Consequently, TABLE VI depicts the
confusion matrix that is obtained for investigating the effects of measurement noise signals,
when their original standard deviations (SDv ) are multiplied by a factor of 20.
According to our simulations, the initial impact of any increase in either the standard deviation
of the noise measurements or the percentage of the RBEE is to delay the fault detection time.
Moreover, the occurrence of false alarms and incorrect fault detections are the other consequences
due to increases in the above uncertainty sources. Different performance indices can be defined
to quantify the robustness of our proposed sensor FDI algorithm with respect to the levels of
uncertainty sources [42], namely:
P5
P5 P5
P6
j=1
i=1 CMij (i 6= j)
j=1 CM6j
i=1 CMii
FPR = P6
, ACC = P6 P6
, IFDR =
, (22)
P6 P5
CM
CM
CM
6j
ij
ij
j=1
j=1
i=1
j=1
i=1
where FPR, ACC and IFDR denote the false positive (false alarm) rate, accuracy and incorrect
fault detection rate, respectively, that are calculated in TABLE VII to investigate the effects of
various uncertainty sources.
According to TABLE VII, increasing the levels of the uncertainty sources results in decrease
of ACC, but increase in FPR and IFDR. Notwithstanding these observations, our proposed sensor
FDI scheme still works sufficiently robust against a high level of measurement noise as well
as discrepancies between the OBEM model and the actual engine health parameters reference
August 28, 2015
DRAFT
19
TABLE VI: The confusion matrix for the measurement noise signals where the SDs are multiplied
by a factor of 20.
TC
PC
N
TT
PT
No Fault
TC
48
0
0
0
0
0
PC
0
46
0
0
0
0
N
0
0
49
0
0
0
TT
0
0
0
48
1
0
PT
0
0
0
0
47
1
No Fault
2
4
1
2
2
49
TABLE VII: Sensor FDI algorithm performance indices corresponding to different levels of
uncertainties.
Scenarios
3% Compressor RBEE
4% Compressor RBEE
3% Turbine RBEE
4% Turbine RBEE
Noise SD×20
Noise SD×25
FPR
0
0.12
0.02
0.04
0.02
0.02
ACC
0.85
0.66
0.94
0.91
0.956
0.91
IFDR
0.16
0.376
0.016
0.076
0.0004
0.012
baselines.
C. Case 3: Sensor Fault Detection and Isolation for Different Fault Severities
In real applications, there is no guarantee that the sensor bias fault severity always matches
the 3% pre-determined fault for which the MM structure is designed. Therefore, it is essential
to investigate the performance of our proposed MM-based FDI scheme for an applied sensor
bias fault having different severities starting from the minimum detectable bias. TABLE VIII
shows the average detection times for all the faulty modes as a function of the fault severities,
when the applied faults occur at different stages of the flight profile with also the maximum
tolerable RBEE% for the compressor health parameters that are indicated in TABLE III. It can be
observed from TABLE VIII that the higher the fault severity with respect to the pre-determined
3% fault, the later the detection time, given that the actual fault becomes further different from
the pre-determined bias fault. Note that the minimum detectable sensor bias fault is 2% that
requires larger time to be detected as compared to the higher fault severities, especially during
the climbing and landing flight modes.
It should be pointed out that our proposed MM-based structure is designed for a 3% sensor
bias fault and is not capable of detecting and isolating sensor bias faults that have far greater
TABLE VIII: The average sensor fault detection times for all fault modes as a function of the
fault severity at different stages of the flight profile, under the maximum tolerable RBEE% for
the compressor health parameters.
Fault Time
kf = 50 sec
kf = 250 sec
kf = 450 sec
August 28, 2015
2%
14.18
8.34
17.66
3%
6.56
3.64
5.18
4%
6.74
3.98
5.32
5%
7.42
4.76
6.66
6%
8.6
5.7
7.52
DRAFT
20
TABLE IX: The average of the maximum detectable sensor bias fault (as percentage of the
engine steady-state outputs) over the entire flight profile by using our FDI scheme when the
corresponding maximum tolerable RBEE% reported in TABLE III are applied to the compressor
and turbine health parameters.
Fault Modes
TC
PC
N
TT
PT
Max. Tolerable
Compressor RBEE%
25%
20%
12%
10%
70%
Max. Tolerable
Turbine RBEE%
70%
50%
100%
14%
75%
severities. For this purpose, TABLE IX shows the maximum detectable sensor bias faults by
using our proposed FDI scheme by also considering the maximum tolerable RBEE% for the
compressor and the turbine health parameters. The reported detectable fault severities in TABLE
IX correspond to the average fault severities for all the flight modes. However, if an applied sensor
bias fault increases beyond the corresponding maximum detectable fault as indicated in TABLE
IX, an incorrect fault may be detected. Therefore, it is recommended that one incorporates more
models within the MM-based structure corresponding to higher pre-determined sensor faults to
become capable of detecting and isolating faults with higher magnitudes in shorter durations of
time.
D. Case 4: Sensor Fault Severity Identification
One of the advantages of our proposed MM-based FDI scheme is its capability in providing
information on the occurred fault severity. However, this method is not precise for fault severity
estimation, since it provides one with only information on the level of an occurred fault instead of
the exact fault severity. Therefore, the modified GLR method introduced in Section V is integrated
with our MM-based FDI scheme to estimate the sensor bias fault magnitude and severity. In
order to investigate the performance of our proposed fault severity estimation method, the notion
of the percentage of weighted mean square normalized residual errors (WMSNE) are obtained
for various detected fault scenarios as follows:
WMSNE% =
PK
(i,s+1)
(i,s+1)
L
(k)( (Y (k)−YŶ (k) (k)) )2
1 X k=kds w
Pκ
× 100,
(i,s+1) (k)
L i=1
k=kds w
Ŷ (i,s+1) (k) = C i (X̂ (i,s+1) (k) − XOBEM (k)) + YOBEM (k) + b̂s zs δ(k − kds ),
s = 1, . . . , q (23)
where K is the total simulation samples and Ŷ (i,s+1) (k) is the numerically constructed output
associated with the (s + 1)th operational mode corresponding to the sth fault, in which the fault
vector effect is generated based on the estimated fault magnitude b̂s and the fault occurrence
detected time kds . It must be noted that there is no feedback and information sent from the fault
estimation module to the filters for updating their pre-determined bias faults. TABLE X shows
the average percentage of the WMSNE that is measured for different stages of the flight profile
as a function of the bias fault severity as applied to various sensors.
August 28, 2015
DRAFT
21
TABLE X: The average WMSNE% for all flight modes as a function of the bias fault severity
applied to various sensors without considering any estimation error for the health parameter
reference baselines.
Fault
TC
PC
N
TT
PT
2%
0.0085
0.1360
0.0111
0.0137
0.1313
3%
0.0067
0.1106
0.0109
0.0066
0.1269
4%
0.0073
0.1293
0.0107
0.0128
0.1359
5%
0.0124
0.1689
0.0105
0.0312
0.1503
6%
0.0195
0.2537
0.0104
0.0944
0.1615
According to TABLE X, the average percentage of WMSNE is less than 0.5% for various
faults that are limited within the operational range of our FDI scheme as was defined in the
preceding case. It can also be observed that the percentage of the WMSNE is lower for the
cruise flight mode as compared to the climbing and landing modes due to the high variations of
the engine thrust.
E. Case 5: Concurrent Fault Detection, Isolation and Identification
In previous simulations only a single sensor fault is applied to the gas turbine engine. In this
case study, our proposed FDII scheme is used for concurrent fault scenarios that occur at different
stages of the flight profile. For this purpose, the hierarchical scheme described in Section IV-B
is used to diagnose the concurrent sensor faults. The fault detection times and the percentages
of the WMSNE as given by (23) are obtained for two selected scenarios with concurrent sensor
faults. The first scenario simulates the effects of a 6% bias fault applied to the TC sensor at
kf = 50 sec during the climbing mode and the concurrent 5% bias fault applied to the N sensor
at kf = 250 sec during the cruise mode. The estimated concurrent faults are depicted in Fig. 4.
The concurrent faults for the first scenario are detected and isolated after 1.7 and 2.4 seconds, respectively from their occurrence instants, whereas the percentages of the WMSNEs
are 0.0208% and 0.106%, respectively. The second scenario simulates the effects of a 4%
bias fault that is applied to the TT sensor at kf = 250 sec during the cruise mode and the
concurrent 6% bias fault that is applied to the PT sensor at kf = 450 sec during the landing
mode. The concurrent faults for the second scenario are detected and isolated after 0.9 and 0.4
seconds, respectively from their occurrence instants, whereas the percentages of the WMSNEs
are 0.0125% and 0.1702%, respectively.
F. Case 6: Comparison
In this section, the efficiency of our proposed MM-based sensor FDI scheme which utilizes
the MHKF as a detection filter is compared with different filtering methods including the linear
Kalman filter (LKF), extended Kalman filter (EKF), unscented Kalman filter (UKF), and the
cubature Kalman filter (CKF) [43]. It must be noted that in the following conducted experiments,
multiple combined residual vectors and covariance matrices associated with multiple operating
points are computed for both the HKF and LKF filtering methods to construct the multiple
HKF (MHKF) and multiple LKF (MLKF)-based FDI schemes, respectively. Moreover, our MMbased FDI scheme is compared with the interacting multiple model (IMM) approach [44], [45].
Different experiments are conducted in this section to compare the promptness of the fault
August 28, 2015
DRAFT
Sensor N Bias Fault
Sensor TC Bias Fault
22
40
20
Estimated Fault
Real Fault
0
-20
0
100
200
300
Time (sec)
400
500
600
(a)
600
400
Estimated Fault
Real Fault
200
0
-200
0
100
200
300
Time (sec)
400
500
600
(b)
Fig. 4: Estimated two concurrent sensor fault severities (a) a 6% bias fault that is applied to the
TC sensor at kf = 50 sec, and (b) a 5% bias fault that is applied to the N sensor at kf = 250
sec without considering any health parameter degradations.
detection and isolation scheme as well as the degree of robustness towards the health parameters
degradations among the above filtering methods. In all the experiments, similar process and
measurement noise signals are applied to the gas turbine engine and similar noise covariance
matrices are used to construct the above filters. The sampling rates for simulating the actual
engine and the nonlinear filtering approaches are 0.01 sec and 0.1 sec, respectively. Also, 2n + 1
sigma points, where n is the dimension of the state variable vector, are generated in the UKF
method with the tuning parameters α = 10−3 , β = 2 and κ = 0. For the CKF method, 2n
cubature points are √
generated through intersecting the unit sphere with the Cartesian axes and
are then scaled by n. Due to space limitations, the details regarding the formulations of the
above applied filtering methods are not included as they can be found in the literature.
Experiment 1: The 3% single bias fault is applied to different sensors at kf = 250 sec during
the cruise mode with no health parameter degradations. All the filters are capable of detecting
and isolating the injected fault correctly with different fault detection times. The sensor FDTs
are shown in TABLE XII for the detection filters that are utilized in the MM-based FDI scheme.
It can be concluded from TABLE XII that our proposed approach is capable of detecting and
isolating the sensor faults faster than other methods, particularly the MLKF method, given that the
OBEM can capture the nonlinear behaviour of the engine more effectively than others. Although,
August 28, 2015
DRAFT
23
1.2
1.2
P1
P2
P3
P4
P5
P
P6
Mode Probability
Mode Probability
0.5
0
0
100
200
300
400
Time (Sec)
500
P1
P2
P3
P4
P5
200
3
P
4
P
5
P
6
100
200
300
400
Time (Sec)
(d)
500
300
400
Time (Sec)
500
600
(c)
P
1
0.5
100
P
1.2
P6
Mode Probability
Mode Probability
2
(b)
1
0
0
P
0.5
0
0
600
(a)
1.2
1
1
1
600
1
P
2
P
3
P
4
P
5
P
6
0.5
0
0
100
200
300
400
Time(Sec)
500
600
(e)
Fig. 5: Mode probabilities for a 3% bias fault that is applied at kf = 250 sec to the TC sensor
by applying multiple (a) HKF, (b) LKF, (c) EKF, (d) UKF and (e) CKF based FDI approaches
with 2.5% degradations applied to the compressor health parameters of the actual engine.
TABLE XI: Absolute mean of residual signals associated with various types of nonlinear filters
corresponding to the healthy mode in the MM-based scheme for the healthy engine sensors
without health degradation.
Sensor
TC
PC
N
TT
PT
Absolute Mean of Residual Signal
MHKF
EKF
UKF
CKF
8.8329 × 10−4 0.00063 0.0058 0.0039
0.0044
0.0186 0.0127 0.0105
−4
4.0454 × 10
0.0032 0.0011 0.0009
8.1498 × 10−4
0.0081 0.0085 0.0069
0.0014
0.0028 0.0082 0.0074
the EKF, UKF, and CKF are well-known nonlinear filtering methods, however they detect and
isolate the injected sensor faults with more delay than our proposed method. According to our
observations that are included in TABLE XI, the HKF has smaller estimation error as compared
to the EKF, UKF, and CKF, which leads to faster FDI performance. It is also expected that
in case the OBEM is subjected to certain unmodeled dynamics with respect to a real engine,
the FDTs that are obtained with multiple EKF, UKF, and CKF based FDI schemes are more
comparable with our proposed MHKF-based FDI scheme.
Experiment 2: The robustness of our developed MHKF-based FDI approach is compared
with the other filtering methods here. The 3% bias fault is applied to the TC sensor at kf = 250
sec during the cruise mode with 2.5% degradations that are applied to the compressor health
parameters. In this experiment, it is assumed that the OBEM health parameters are not updated
August 28, 2015
DRAFT
24
and the 2.5% degradation is only applied to the compressor health parameters of the actual engine.
Therefore, the reference baselines associated with the compressor health parameters are one for
both the OBEM model and the nonlinear model that is used for deriving the other nonlinear
filtering methods. The mode probabilities that are generated in this scenario are depicted in Fig.
5. It can be concluded that our method outperforms the other approaches in terms of robustness
with respect to the health parameters degradations. It is also shown in Fig. 5 that the applied
fault can be detected by using the EKF and MLKF methods in addition to our method, but false
alarms are generated by these methods before the occurrence of the fault.
Experiment 3: Our proposed MHKF method is similar to the EKF in the sense of the
linearization process at certain operating points instead of the entire flight profile. Therefore,
one may expect that the EKF method can yield similar results to that of the MHKF method in
case the health parameters are also updated for the nonlinear engine model used for deriving
the EKF. In this experiment, the same engine model where its health parameters are updated is
used for the MM-based structures that utilize the MHKF and EKF schemes. The 3% bias fault is
applied to the TC sensor at kf = 250 sec during the cruise mode with λ = 0.99 and λ̂ = 0.975.
The estimated health parameters are updated similarly for both the EKF and MHKF schemes.
Fig. 6 shows the mode probabilities that are generated in this scenario. According to Fig. 6, the
EKF is also capable of correctly detecting and isolating the applied fault by using the health
parameters updating mechanism. However, the fault detection time and the computational time
of the EKF is far greater than that of the MHKF. The MM-based FDI method is capable of
detecting and isolating the applied fault in 2.7 sec and 20.8 sec by using the MHKF and EKF
methods, respectively. The same results are also obtained for the other fault scenarios, but are
not included here due to the space limitations.
Experiment 4: In this experiment, the performance of the FDI algorithm that utilizes our
proposed MHKF as well as the other filtering methods are compared in the interacting multiple
model (IMM) structure. For the IMM-based FDI structure, the sensor FDTs are computed for
the 3% single bias fault that is applied to the sensors with no health parameter degradations
as represented in the TABLE XIII. According to the simulation results, the fault is detected
sooner by using the IMM structure that utilizes all the detection filters except the MHKF. It
appears that the interactions among the multiple models do not improve the MHKF-based FDI
scheme performance. For simulating the IMM scheme which utilizes our proposed MHKF, the
off-diagonal elements of the transition probability matrix should be selected less than 0.001;
otherwise, the probability of the healthy mode monotonically decreases to reach almost 0.8
before the occurrence of a sensor fault which can enhance the risk of a false alarm. Therefore,
it is not recommended to use the IMM scheme for the MHKF method.
Experiment 5: In this experiment, the computational time of all the above filtering methods
are obtained and compared with our proposed MHKF approach in TABLE XIV. This comparison
confirms the advantages of implementing the MHKF approach in real-time applications. The
computational time is measured for only the filters corresponding to the healthy mode by the
use of “tic-toc” command in Matlab. It must be noted that the computational time of the MHKF
includes the time for not only simulating the healthy mode but also for running the OBEM
model. Therefore, based on the information provided in TABLE XIV, the MHKF scheme is
faster and more suitable than the other filtering methods to be utilized during the entire flight
August 28, 2015
DRAFT
25
TABLE XII: Sensor fault detection times using multiple non-interacting HKF, LKF, EKF, UKF
and CKF based FDI approaches without considering health parameter degradation.
MM-Based
FDI Approach
HKF
LKF
EKF
UKF
CKF
Detection Time (FDT) (sec)
PC
N
TT
PT
0.2
0.2
0.2
0.3
38.2 16.9 12.1
12.4
1
0.9
0.8
0.9
0.7
0.6
0.7
2.7
0.6
0.7
0.7
2.3
1.2
1.2
P
1
P
2
P
3
P
P
4
5
P
6
1
Mode Probability
Mode Probability
Fault
TC
0.2
8.3
1
0.6
0.6
0.5
0
0
100
200
300
400
Time (Sec)
500
600
(a)
1
P
1
P
2
P
3
P
4
P
5
P
6
0.5
0
0
100
200 300 400
Time (Sec)
500
600
(b)
Fig. 6: Mode probabilities for a 3% bias fault applied at kf = 250 sec to the TC sensor using
(a) the MHKF and (b) the EKF-based FDI methods, with λ = 0.99 and λ̂ = 0.975.
profile. Computing the Jacobian functions in each time step for the EKF method and also the
cubature and sigma points with their corresponding mean and covariance weight matrices for
the CKF and UKF methods, respectively are time consuming operations that lead to the higher
computational time compared to our proposed MHKF scheme. Moreover, the MLKF approach
computes the Kalman gain matrices on-line, whereas the MHKF utilizes the previously stored
gain matrices. However, if the same steady-state Kalman gain matrices are computed off-line
TABLE XIII: Sensor fault detection times using multiple interacting HKF, LKF, EKF, UKF and
CKF based FDI approaches without considering any health parameter degradation.
MM-Based
FDI Approach
HKF
LKF
EKF
UKF
CKF
August 28, 2015
Fault
TC
0.3
6.3
0.9
0.5
0.5
Detection Time (FDT) (sec)
PC
N
TT
PT
0.4
0.3
0.3
0.5
35.8 15.7 11.6
11.8
0.9
0.7
0.8
0.8
0.5
0.5
0.6
2.3
0.5
0.5
0.5
1.9
DRAFT
26
TABLE XIV: Computational time as measured in seconds for different filtering methods.
MHKF
131.21
MLKF
526.81
EKF
1184.2
UKF
2856.4
CKF
2673.44
and stored for use in the MLKF approach, the average computational time associated with the
healthy mode filters in the MLKF scheme is about 80 seconds, which is less than that of the
MHKF method. Under this situation, the extra computational time required by the MHKF is
associated with the OBEM operations. Therefore, the only disadvantage of the MHKF method
is the amount of memory that is needed for storing the look-up tables that include the matrices
Ai , C i and Kssi at different operating points.
Finally, in the following experiment results, it is shown that the number of required operating
points, and therefore the amount of required memory for the MHKF, is less than that of the
MLKF-based FDI method.
Experiment 6: In this experiment, the number of required operating points is investigated
by measuring the mean of the combined residual signals in presence of the health parameter
degradations for the MLKF and MHKF methods. To cover the entire flight profile, five operating
points are generated for the MHKF method as described in TABLE II. Note that the MLKF
method still needs more operating points since it is not capable of tracking rapid variations of
the thrust during the climbing and landing modes, and therefore it generates false alarms due to
the higher estimation error as compared to the MHKF scheme. TABLE XV compares the mean
of the MHKF and MLKF combined residual signals that are associated with the healthy mode
as a function of the number of operating points. Fig. 7 displays the mode probabilities that are
generated for the MHKF and MLKF methods for different number of operating points. In this
experiment, the engine has no sensor fault, but a 3% degradation is applied to the actual engine
compressor health parameters. As stated in Section VI-A, the MHKF is robust with respect to the
applied health degradations even without updating the OBEM model health parameters, although
the MHKF needs to be updated for larger degradation levels. Based on the results obtained, it
can be concluded that the mean of the combined residual signal associated with the healthy mode
can be reduced by increasing the number of the operating points. It is expected that the MLKF
estimation error decreases further by increasing the number of the operating points. However,
we are not capable of perfectly avoiding generation of a false alarm by using the MLKF method
corresponding to our particular flight profile by increasing the number of the operating points,
since the MLKF cannot track high variations of the thrust in our application. Therefore, it may
be feasible to apply the MLKF with the combined PWL models if the engine thrust increases
with the slower speed than ours since the PWL model weights calculation is a time-consuming
process which may diminish the ability of the MLKF to track the rapid variations of the thrust.
Therefore, it can be concluded that the MM-based FDI scheme which utilizes our proposed
MHKF approach is capable of promptly detecting and isolating various sensor bias faults during
the entire flight profile by updating the OBEM health parameters reference baselines and even
without updating the OBEM model for certain levels of health parameter degradations. Our
scheme also estimates the engine state and output variables with more accuracy and smaller
number of operating points as compared to that by the MLKF-based scheme.
August 28, 2015
DRAFT
27
1.2
P1
P2
P3
P4
P5
1.2
P6
P
1
2
P
3
P
4
P
5
P
1.2
100
200
300
400
Time (Sec)
500
600
1
P
2
P
3
P
4
P
5
P
6
1
0.5
0
0
P
6
Mode Probability
0.5
0
0
P
1
Mode Probability
Mode Probability
1
100
200
(a)
300
400
Time (Sec)
500
0.5
0
0
600
100
200
300
400
Time (Sec)
500
600
500
600
(c)
(b)
1.2
P
1
P
2
P
3
P
4
P
5
P
6
Mode Probability
1
0.5
0
0
(d)
(e)
100
200
300
400
Time (Sec)
(f)
Fig. 7: Mode probabilities that are generated for the MM-based FDI method by utilizing (a-c)
the MHKF and (d-f) the MLKF with one, three and five operating points.
TABLE XV: Mean of the combined residual signals (γ 1c ) associated with the healthy mode for
both the MHKF and MLKF methods in terms of the number of the operating points, whereas
3% degradation is applied to the compressor health parameters.
Sensor
TC
PC
N
TT
PT
1 operating point
HKF
LKF
0.109
0.2305
0.2286
0.943
0.0255 0.0123
0.0813 0.3339
0.0152 0.0477
3 operating points
MHKF MLKF
0.077
0.1091
0.0771
0.1523
0.0167
0.0053
0.0745
0.2372
0.0116
0.036
5 operating points
MHKF MLKF
0.0119
0.0857
0.0216
0.1315
0.0003
0.0014
0.0197
0.1008
0.0006
0.0049
VII. CONCLUSIONS
In this paper, a novel approach for both single and concurrent sensor fault detection, isolation
and identification/estimation (FDII) for gas turbine engines is proposed. Our methodology is
based on the developed Hybrid Kalman Filter (HKF) as the detection filter of a hierarchical
multiple-model based structure. Despite the use of linear Kalman filters, the HKF is capable
of capturing the nonlinearities of the system by integrating a nonlinear on-board engine model
(OBEM) with piecewise linear (PWL) models to cover the entire operating range of the engine.
Compared to the multiple linear Kalman filter (MLKF), our proposed approach requires fewer
number of operating points although each corresponds to a larger operating range.
Another important contribution of this work is in the inclusion of the effects of health parameter
degradations in our proposed sensor FDII scheme through updating the OBEM health parameters
reference baselines that enables one to prevent false alarms and incorrect fault detections.
August 28, 2015
DRAFT
28
Therefore, unlike most of the previous work in the literature that have not considered the
influence of health parameters variations on the performance of the developed FDII approaches,
our proposed sensor FDII scheme is sufficiently more effective for use in the entire flight profile.
It is furthermore suitable during the engine life cycle by updating the OBEM health parameters
reference baselines, and by replacing the steady-state values with the OBEM states and outputs
and integration of the designed PWL models. Moreover, our proposed MM-based sensor FDI
scheme is integrated with the modified GLR method to estimate a sensor fault severity. Finally,
it is demonstrated through extensive simulation studies that the accuracy and robustness of
our proposed MHKF-based FDI approach against the engine health parameter degradations are
significantly superior to that of those generated by other investigated methods such as the MLKF,
EKF, UKF, and CKF.
A PPENDIX A
In Section II, for sake of notational simplicity the same notations were used for both versions
of the OBEM. However, here we use XNOBEM and YNOBEM to designate the nominal OBEM
(λ = 1) that is equivalent to (2), and XOBEM and YOBEM to designate the OBEM with the
updated health parameters (λ̂ 6= 1) that is equivalent to (5).
Let the nominal continuous-time OBEM model (NOBEM) be represented by:
ẊNOBEM (t) = fc (XNOBEM (t), HNOBEM (t), U (t)),
YNOBEM (t) = gc (XNOBEM (t), HNOBEM (t)).
To construct the HKF, the nominal continuous-time OBEM is initially linearized at a certain
operating point (Xss , Uss , Yss ) without considering degradations as follows:
∂fc
∂fc
|U (U (t) − Uss ) + ∆fc ,
|Xss (XNOBEM (t) − Xss ) +
∂XNOBEM
∂U ss
∂gc
YNOBEM (t) = gc (Xss , Uss ) +
|X (XNOBEM (t) − Xss ) + ∆gc ,
∂XNOBEM ss
ẊNOBEM (t) = fc (Xss , Uss ) +
where fc (Xss , Uss ) ≡ 0 and gc (Xss , Uss ) ≡ Yss . The health parameters are the state-dependent
functions whose effects are incorporated in the above partial derivatives. The linear continuoustime model is now discretized by using a sufficiently small sampling period to yield the corresponding discrete-time linear model:
XNOBEM (k + 1) − Xss = A|Xss (XNOBEM (k) − Xss ) + B|Uss (U (k) − Uss ) + ∆F,
YNOBEM (k) = Yss + C|Xss (XNOBEM (k) − Xss ) + ∆G,
RT
where A|Xss = exp(Ac |Xss Ts ), B|Uss = ( 0 s exp(Ac |Xss τ )dτ )Bc |Uss , and C|Xss = Cc |Xss are
obtained in terms of the state-space matrices that are associated with the continuous-time linear
∂gc
∂fc
c
|Xss , Bc |Uss = ∂f
| , and Cc |Xss = ∂XNOBEM
|Xss , and the sampling
model Ac |Xss = ∂XNOBEM
∂U Uss
period is denoted by Ts . Also, ∆fc , ∆gc , ∆F and ∆G represent the higher order terms in
the linearization process. To construct the HKF, it is necessary to update the OBEM health
parameters. Therefore, the nominal model is only used once to derive the state-space matrices
and the steady-state values.
Assumption 1. The linearization of the nominal OBEM with λ = 1 and the updated OBEM
with λ̂ 6= 1 generate approximately the same matrices and steady-state values.
August 28, 2015
DRAFT
29
Therefore, given that Assumption 1 holds, the linearized version of the updated OBEM can
be derived as follows:
XOBEM (k + 1) − Xss ≈ A|Xss (XOBEM (k) − Xss ) + B|Uss (U (k) − Uss ) + ∆F,
YOBEM (k) ≈ Yss + C|Xss (XOBEM (k) − Xss ) + ∆G.
(24)
Moreover, the relationship between the instantaneous values of XOBEM (k) and YOBEM (k) and
the steady-state values associated with each linearized model can be given by:
XOBEM (k) = Xss + Xl (k),
YOBEM (k) = Yss + Yl (k),
(25)
where Xl (k) and Yl (k) represent the perturbations from the steady-state values as well as
variations of the OBEM state variables and outputs due to updating the health parameters
reference baselines. The linear Kalman filter can be designed as follows:
X̂(k + 1) − Xss = A|Xss (X̂(k) − Xss ) + B|Uss (U (k) − Uss ) + Kss (Y (k) − Ŷ (k)),
Ŷ (k) = Yss + C|Xss (X̂(k) − Xss ),
which can be rewritten by using (25) as follows:
X̂(k + 1) − XOBEM (k) + Xl (k) = A|Xss (X̂(k) − XOBEM (k)) + Kss (Y (k) − Ŷ (k)) + A|Xss Xl (k)
+ B|Uss (U (k) − Uss ),
Ŷ (k) = YOBEM (k) + C|Xss (X̂(k) − XOBEM (k)) + C|Xss Xl (k) − Yl (k).
(26)
Using (24), the last two terms in the R.H.S of (26) can be rewritten as follows:
A|Xss (XOBEM (k) − Xss ) + B|Uss (U (k) − Uss ) ≈ XOBEM (k + 1) − Xss − ∆F,
C|Xss (XOBEM (k) − Xss ) − Yl (k) ≈ YOBEM (k) − Yss − Yl (k) − ∆G = −∆G.
Therefore, the discrete-time form of the HKF can be obtained as follows:
X̂(k + 1) − XOBEM (k + 1) ≈ A|Xss (X̂(k) − XOBEM (k)) + Kss (Y (k) − Ŷ (k)) − ∆F,
Ŷ (k) ≈ YOBEM (k) + C|Xss (X̂(k) − XOBEM (k)) − ∆G,
which will lead to (6) for the particular operating point if the higher order terms ∆F and ∆G
as well as the approximation error due to Assumption 1 are neglected.
R EFERENCES
[1] R. Isermann and P. Balle, “Trends in the application of model-based fault detection and diagnosis of technical processes,”
Control Engineering Practice, vol. 5, no. 5, pp. 709–719, 1997.
[2] P. M. Frank, “Fault diagnosis in dynamic systems using analytical and knowledge-based redundancy: A survey and some
new results,” Automatica, vol. 26, no. 3, pp. 459–474, 1990.
[3] W.-M. Lin and T.-C. Ou, “Unbalanced distribution network fault analysis with hybrid compensation,” Generation,
Transmission & Distribution, IET, vol. 5, no. 1, pp. 92–100, 2011.
[4] X. Zhang, L. Tang, and J. Decastro, “Robust fault diagnosis of aircraft engines: a nonlinear adaptive estimation-based
approach,” IEEE Trans. on Control Systems Technology, vol. 21, no. 3, pp. 861–868, 2013.
[5] V. Venkatasubramanian, R. Rengaswamy, K. Yin, and S. N. Kavuri, “A review of process fault detection and diagnosis:
Part I: Quantitative model-based methods,” Computers & Chemical Engineering, vol. 27, no. 3, pp. 293–311, 2003.
August 28, 2015
DRAFT
30
[6] I. Hwang, S. Kim, Y. Kim, and C. E. Seah, “A survey of fault detection, isolation, and reconfiguration methods,” IEEE
Trans. on Control Systems Technology, vol. 18, no. 3, pp. 636–653, 2010.
[7] J. C. d. Silva, A. Saxena, E. Balaban, and K. Goebel, “A knowledge-based system approach for sensor fault modeling,
detection and mitigation,” Expert Systems with Applications, vol. 39, no. 12, pp. 10977–10989, 2012.
[8] T. E. Menke and P. S. Maybeck, “Sensor/actuator failure detection in the vista F-16 by multiple model adaptive estimation,”
IEEE Trans. on Aerospace and Elec. Sys., vol. 31, no. 4, pp. 1218–1229, 1995.
[9] P. S. Maybeck, “Multiple model adaptive algorithms for detecting and compensating sensor and actuator/surface failures
in aircraft flight control systems,” Int. J. of Robust and Nonlinear Control, vol. 9, no. 14, pp. 1051–1070, 1999.
[10] N. Meskin, E. Naderi, and K. Khorasani, “Fault diagnosis of jet engines by using a multiple model-based approach,” ASME
Turbo Expo, no. GT2010-23442, pp. 319–329, 2010.
[11] N. Meskin, E. Naderi, and K. Khorasani, “A multiple model-based approach for fault diagnosis of jet engines,” IEEE
Trans. on Control Sys. Tech., pp. 254–262, 2013.
[12] E. Naderi, N. Meskin, and K. Khorasani, “Nonlinear fault diagnosis of jet engines by using a multiple model-based
approach,” J. of Eng. for Gas Turbines and Power, vol. 134, no. 1, p. 011602, 2012.
[13] B. Pourbabaee, N. Meskin, and K. Khorasani, “Sensor fault detection and isolation using multiple robust filters for linear
systems with time-varying parameter uncertainty and error variance constraints,” in IEEE Conf. on Control Applications
(CCA), pp. 382–389, 2014.
[14] A. J. Volponi, “Foundation of gas path analysis (part I and II),” Gas Turbine Condition Monitoring and Fault Diagnosis,
vol. 1, 2003.
[15] T. Kobayashi and D. L. Simon, “Evaluation of an enhanced bank of Kalman filters for in-flight aircraft engine sensor fault
diagnostics,” J. of Eng. for Gas Turbines and Power, vol. 127, p. 497, 2005.
[16] T. Kobayashi and D. L. Simon, “Application of a bank of Kalman filters for aircraft engine fault diagnostics,” in ASME
Turbo Expo, collocated with the International Joint Power Generation Conference, pp. 461–470, 2003.
[17] T. Kobayashi and D. L. Simon, “Hybrid Kalman filter approach for aircraft engine in-flight diagnosis: Sensor fault detection
case,” J. of Eng. for Gas Turbines and Power, vol. 129, no. 3, pp. 746–754, 2007.
[18] S. Simani, C. Fantuzzi, and S. Beghelli, “Diagnosis techniques for sensor faults of industrial processes,” IEEE Trans. on
Control Systems Technology, vol. 8, no. 5, pp. 848–855, 2000.
[19] A. J. Volponi, H. DePold, R. Ganguli, and C. Daguang, “The use of Kalman filter and neural network methodologies in
gas turbine performance diagnostics: a comparative study,” J. of Eng. for Gas Turbines and Power, vol. 125, pp. 917–924,
2003.
[20] R. Ganguli, “Application of fuzzy logic for fault isolation of jet engines,” J. of Eng. for Gas Turbines and Power, vol. 125,
pp. 617–623, 2003.
[21] S. Sampath and R. Singh, “An integrated fault diagnostics model using genetic algorithm and neural networks,” J. of Eng.
for Gas Turbines and Power, vol. 128, pp. 49–56, 2006.
[22] S. Rahme and N. Meskin, “Adaptive sliding mode observer for sensor fault diagnosis of an industrial gas turbine,” Control
Engineering Practice, vol. 38, pp. 57–74, 2015.
[23] E. Tsoutsanis, N. Meskin, M. Benammar, and K. Khorasani, “Transient gas turbine performance diagnostics through
nonlinear adaptation of compressor and turbine maps,” J. of Engineering for Gas Turbines and Power, vol. 137, no. 9,
p. 091201, 2015.
[24] E. Tsoutsanis, N. Meskin, M. Benammar, and K. Khorasani, “A component map tuning method for performance prediction
and diagnostics of gas turbine compressors,” Applied Energy, vol. 135, pp. 572–585, 2014.
[25] L. Shang and G. Liu, “Sensor and actuator fault detection and isolation for a high performance aircraft engine bleed air
temperature control system,” IEEE Trans. on Control Systems Technology, vol. 19, no. 5, pp. 1260–1268, 2011.
[26] R. Babuska and H. B. Verbruggen, “Fuzzy set methods for local modeling and identification,” 1997.
[27] X. C. Zhang, A. Visala, A. Halme, and P. Linko, “Functional state modelling approach for bioprocesses: local models for
aerobic yeast growth processes,” J. of Process Control, vol. 4, no. 3, pp. 127–134, 1994.
[28] M. Meila and M. I. Jordan, “Markov mixtures of experts,” Multiple Model Approaches to Modeling and Control, pp. 145–
166, 1997.
[29] M. M. Seron, J. A. De Dona, and J. H. Richter, “Integrated sensor and actuator fault-tolerant control,” Int. J. of Control,
vol. 86, no. 4, pp. 689–708, 2013.
[30] A. S. Willsky and H. L. Jones, “A generalized likelihood ratio approach to the detection and estimation of jumps in linear
systems,” IEEE Trans. on Automatic Control, vol. 21, no. 1, pp. 108–112, 1976.
[31] J. Prakash, S. Narasimhan, and S. C. Patwardhan, “Integrating model based fault diagnosis with model predictive control,”
Industrial & Engineering Chemistry Research, vol. 44, no. 12, pp. 4344–4360, 2005.
[32] S. C. Patwardhan, S. Manuja, S. Narasimhan, and S. L. Shah, “From data to diagnosis and control using generalized
orthonormal basis filters. part II: Model predictive and fault tolerant control,” J. of Process Control, vol. 16, no. 2,
pp. 157–175, 2006.
[33] P. M. Frank and X. Ding, “Survey of robust residual generation and evaluation methods in observer-based fault detection
systems,” J. of process control, vol. 7, no. 6, pp. 403–424, 1997.
August 28, 2015
DRAFT
31
[34] R. Seliger and P. M. Frank, “Fault-diagnosis by disturbance decoupled nonlinear observers,” in Proceedings of the 30th
IEEE Conf. on Decision and Control, pp. 2248–2253, 1991.
[35] D. Simon, “A comparison of filtering approaches for aircraft engine health estimation,” J. of Aerospace Science and
Technology, vol. 12, no. 4, pp. 276–284, 2008.
[36] B. Pourbabaee, N. Meskin, and K. Khorasani, “Multiple-model based sensor fault diagnosis using hybrid Kalman filter
approach for nonlinear gas turbine engines,” in American Control Conference (ACC), pp. 4717–4723, 2013.
[37] R. Murray-Smith and T. Johansen, Multiple Model Approaches to Nonlinear Modelling and Control. CRC press, 1997.
[38] X. R. Li and Y. Zhang, “Numerically robust implementation of multiple-model algorithms,” IEEE Trans. on Aerospace
and Electronic Systems, vol. 36, no. 1, pp. 266–278, 2000.
[39] J. Prakash, S. C. Patwardhan, and S. Narasimhan, “A supervisory approach to fault-tolerant control of linear multivariable
systems,” Industrial & Engineering Chemistry Research, vol. 41, no. 9, pp. 2270–2281, 2002.
[40] A. P. Deshpande and S. C. Patwardhan, “Online fault diagnosis in nonlinear systems using the multiple operating approach,”
Ind. Eng. Chem. Res., vol. 47, no. 17, pp. 6711–6726, 2008.
[41] N. L. R. Group, “GSP software,” http://www.gspteam.com, 2011.
[42] T. Fawcett, “An introduction to ROC analysis,” Pattern recognition letters, vol. 27, no. 8, pp. 861–874, 2006.
[43] I. Arasaratnam and S. Haykin, “Cubature Kalman filters,” IEEE Trans. on Automatic Control, vol. 54, no. 6, pp. 1254–1269,
2009.
[44] Y. Zhan and J. Jiang, “An interacting multiple-model based fault detection, diagnosis and fault-tolerant control approach,”
in Proceedings of the 38th IEEE Conf. on Decision and Control, vol. 4, pp. 3593–3598, 1999.
[45] J. Ru and X. R. Li, “Variable-structure multiple-model approach to fault detection, identification, and estimation,” IEEE
Trans. on Control Systems Technology, vol. 16, no. 5, pp. 1029–1038, 2008.
August 28, 2015
DRAFT
| 3cs.SY
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.